Eighteenth International Conference on

Fifteenth International Conference on

Artificial
ArtificialIntelligence
Intelligence and
and Law
Law

San Diego, June 8 - 12, 2015

São Paulo, Brasil, June 21-25, 2021
University of San Diego School of Law
Law school, University of São Paulo

Proceedings
Proceedingsofofthe
theConference
Conference
Sponsored by:
The International Association for Artificial Intelligence and Law
Sponsored
Thomson by:
Reuters
InternationalUniversity
Association forDiego
of San Artificial
CenterIntelligence
for IP Law & and Law
Markets
Davis Polk & WardwellJusbrasil
LLP
Albert Einstein Israeli Hospital
TrademarkNow
Lawgorithm
Legal Robot
LegalCode
Pires e Gonçalves Advogados 
Opice Blum Advogados 
In cooperation
OASIS with:
Open
Association for the Advancement of Artificial Intelligence
Urbano Vitalino Advogados (AAAI)
ACM SIGART

In cooperation with:
Association for the Advancement of Artificial Intelligence (AAAI)
ACM SIGAI
 The Association for Computing Machinery
1601 Broadway, 10th Floor
New York, New York 10019, USA

ACM COPYRIGHT NOTICE. Copyright © 2021 by the Association for Computing Ma-
chinery, Inc. Permission to make digital or hard copies of part or all of this work for
personal or classroom use is granted without fee provided that copies are not made
or distributed for profit or commercial advantage and that copies bear this notice and
the full citation on the first page. Copyrights for components of this work owned by
others than ACM must be honored. Abstracting with credit is permitted. To copy
otherwise, to republish, to post on servers, or to redistribute to lists, requires prior
specific permission and/or a fee. Request permissions from Publications Dept., ACM,
Inc., fax +1 (212) 869-0481, or permissions@acm.org.

For other copying of articles that carry a code at the bottom of the first or last page,
copying is permitted provided that the per-copy fee indicated in the code is paid
through the Copyright Clearance Center, 222 Rosewood Drive, Danvers, MA 01923,
+1-978-750-8400, +1-978-750-4470 (fax).

ACM ISBN: 978-1-4503-8526-8

ii
Conference Organization
Program Chair
Adam Zachary Wyner (Swansea University, United Kingdom)

Conference Chair
Juliano Maranhão (University of São Paulo, Brazil)

Tutorials/Workshops & Proceedings Chair

Adam Zachary Wyner (Swansea University, United Kingdom)

Secretary / Treasurer
Michał Araszkiewicz (Jagiellonian University, Poland)

Local Organising Committee

Renata Wassermann (University of São Paulo, Brazil)
Samuel Barbosa (University of São Paulo, Brazil)
Alexandre Zavaglia (FGV-SP, Brazil)
Aline Macohin (Federal University of Paraná, Brazil)
Marcelo Lopes (University of São Paulo, Brazil)
Roberto Bornhausen (University of São Paulo, Brazil)

Doctoral Consortium & Mentoring Programme Chair

Michał Araszkiewicz (Jagiellonian University)

Industry Chair
Fabio Cozman (University of Sao Paulo, Brazil)

Program committee
Tommaso Agnoloni (CNR, Italy)
Thomas Ågotnes (University of Bergen, Norway)
Laura Alonso Alemany (Universidad Nacional de Córdoba, Spain)
Francisco Andrade (Universidade do Minho, Portugal)
Michał Araszkiewicz (Uniwersytet Jagiellonski w Krakowie, Poland)
Kevin Ashley (University of Pittsburgh, United States)
Katie Atkinson (University of Liverpool, United Kingdom)
Trevor Bench-Capon (University of Liverpool, United Kingdom)
Floris Bex (Utrecht University, Netherlands)
Luther Branting (The MITRE Corporation, United States)

iii
Scott Brewer (Harvard Law School, United States)
Pompeu Casanovas (La Trobe, Australia)
Jack G. Conrad (Thomson Reuters, United States)
Claudia d’Amato (University of Bari, Italy)
Luigi Di Caro (University of Turin, Italy)
Rossana Ducato (University of Aberdeen, Italy)
Jenny Eriksson Lundström (Uppsala University, Sweden)
Enrico Francesconi (European Parliament, Luxembourg)
Fernando Galindo Ayuda (Universidad de Zaragoza,Spain)
Kripabandhu Ghosh (Indian Institute of Science Education and Research, India)
Saptarshi Ghosh (Indian Institute of Technology Kharagpur, India)
Randy Goebel (University of Alberta, Canada)
Thomas Gordon (Germany)
Guido Governatori (CSIRO, Australia)
Matthias Grabmair (Technical University of Munich, Germany)
Davide Grossi (University of Groningen and University of Amsterdam, Netherlands)
Maura R. Grossman (University of Waterloo, Canada)
Mustafa Hashmi (Data61, CSIRO Australia)
Bruce Hedin (H5, United States)
Hans Henseler (University of Applied Sciences Leiden, Netherlands)
Rinke Hoekstra (Elsevier, Netherlands)
Joris Hulstijn (Tilburg University, Netherlands)
John Joergensen (Rutgers Law School, United States)
Yoshinobu Kano (Shizuoka University, Japan)
Daniel Katz (Illinois Tech & Bucerius Law School, United States)
Jeroen Keppens (King’s College London, United Kingdom)
Marc Lauritsen (Capstone Practice Systems, United States)
Rūta Liepin, a (Maastricht University, Netherlands)
Arno R. Lodder (Vrije Universiteit Amsterdam, Netherlands)
Prasenjit Majumder (DAIICT, India)
Juliano Maranhao (University of São Paulo, Brazil)
L. Thorne McCarty (Rutgers University, United States)
Parth Mehta (Parmonic AI, India)
Raquel Mochales (Cerence, Belgium)
Ashutosh Modi (Indian Institute of Technology Kanpur, India)
Katsumi Nitta (Tokyo Institute of Technology, Japan)
Merel Noorman (Tilburg University, Netherlands)
Paulo Novais (Universidade do Minho, Portugal)
Gordon Pace (University of Malta, Malta)
Ugo Pagallo (University of Turin, Italy)
Arindam Pal (Data61, CSIRO, Australia)
Monica Palmirani (University of Bologna, Italy)
Girish Palshikar (Tata Consultancy Services Ltd., India)
Sachin Pawar (Tata Consultancy Services, India)
Wim Peters (University of Aberdeen, Netherlands)
Henry Prakken (Utrecht University & University of Groningen, Netherlands)

iv
Paulo Quaresma (University of Evora, Portugal)
Edwina Rissland (University of Massachusetts/Amherst, United States)
Livio Robaldo (Swansea University, United Kingdom)
Anna Ronkainen (University of Helsinki, Finland)
Antonino Rotolo (University of Bologna, Italy)
Giovanni Sartor (University of Bologna, Italy)
Ken Satoh (NII, Japan)
Burkhard Schafer (University of Edinburgh, United Kingdom)
Fernando Schapachnik (Universidad de Buenos Aires, Argentina)
Uri Schild (Bar Ilan University, Israel)
Frank Schilder (Thomson Reuters, United States)
Marijn Schraagen (Utrecht University, Netherlands)
Erich Schweighofer (University of Vienna, Austria)
Giovanni Sileno (University of Amsterdam, Netherlands)
Munindar Singh (North Carolina State University, United States)
Clara Patricia Smith (Universidad Nacional de La Plata (UNLP), Argentina)
Katsuhiko Toyama (Nagoya University, Japan)
Thomas Vacek (Thomson Reuters, United States)
Leon van der Torre (Luxembourg)
Marc van Opijnen (Ministry of the Interior and Kingdom Relations of the Nether-
lands, Netherlands)
Bart Verheij (University of Groningen, Netherlands)
Serena Villata (Universite Cote d’Azur, CNRS, France)
Vern R. Walker (Hofstra University, United States)
Radboud Winkels (University of Amsterdam, Netherlands)
Masaharu Yoshioka (Hokkaido University, Japan)
Haozhen Zhao (Ankura, United States)
Tomasz Zurek (Maria Curie-Sklodowska University, Poland)

v
Preface
I am pleased to share with you the proceedings of the 18th International Conference
on Artificial Intelligence and Law (ICAIL 2021). Since 1987, the International Asso-
ciation for Artificial Intelligence and Law (IAAIL) has biennially organised ICAIL to
present and discuss research and applications as well as to stimulate interdisciplinary
and international collaboration. The ICAIL series can lay claim to substantial influ-
ence on the recent growth in AI in legal services. This year’s ICAIL upholds and
extends the mission of the IAAIL. ICAIL 2021 runs the week of June 21-25. For the
first time, the conference and workshops are presented 100% online and cost free,
leading to over 1500 registrations in 65 countries!
The conference had 89 submissions; 17 were selected for publication as full pa-
pers (~19%), 17 as short papers (~19%), 8 as extended research abstracts (~9%), 2 as
demonstration papers (~2%), and 3 as COLIEE papers (~3%). ICAIL strives to max-
imise the opportunities for researchers to present their work. In addition, ICAIL will
hold a Doctoral Consortium, helping emerging researchers to engage with the ICAIL
community. There will be 11 collocated workshops on focused topics.
Research in AI & Law is highly interdisciplinary. A range of AI theories and tech-
niques may apply to diverse legal information, processes, or topics. As well, there are
important considerations about how the Law applies to AI. The relation between AI
and Law is, then, many-to-many. While machine learning techniques perform well, it
may be crucial to explain results in legal contexts. Moreover, as AI continues to make
inroads into legal services, other success criteria must be addressed such as: account-
ability, accessibility, portability, linking, consistency, and resource sharing. These are
matters for further research.
The interdisciplinarity of AI & Law shows in our invited speakers. Prof. Stuart
Russell, an internationally recognised researcher and AI educator, will speak on “Prov-
ably Beneficial AI”, which will be discussed by a panel, deepening our understanding
of the relation between AI and Law. Joe Cohen of Dentons law firm will talk about
advances in automation, highlighting the real world impact of AI and Law. Finally,
IAAIL president Enrico Francesconi will outline the evolution and the perspectives of
AI research in relation to ICAIL.
Finally, many people worked hard over months to make ICAIL 2021 excellent.
Many thanks to the following. Conference chair Juliano Maranhão and his team took
on the task to put ICAIL 2021 online and in very trying circumstances. The IAAIL
secretary Michał Araszkiewicz, along with Anne Gardner, addressed administrative
and management matters. The most substantive contributors were the authors who
submitted papers and the reviewers who took the time to assess and discuss the sub-
missions; they have shaped the content that advances the field. The organisers of and
presenters at the workshops and Doctoral Consortium all extended the discussion and
supported emerging researchers. Our sponsors provided essential recognition and
support. And finally, we are most appreciative of the IAAIL Executive Committee,
which promotes AI & Law research through the ICAIL conferences.

Adam Zachary Wyner, ICAIL 2021 Program Chair

vi
ICAIL 2021 Program and Schedule of Events
All times in GMT. Events marked REC. are streamed events (replay with live discus-
sion with pre-registered questions).

Monday, 21 June – Workshops

Various-13:30 Workshops
Schedule independently defined by the organizers.
6 hours - 1st half
LEGAL AIIA AI and Intelligent Assistance for Legal Professionals in the
Digital Workplace
JUL.IA Artificial Intelligence in Jurisdictional Logistics
COLIEE Competition on Legal Information Extraction/Entailment
3 hours
BEFAIR Bias, Ethics and Fairness in Artificial Intelligence: Representa-
tion and Reasoning
13:30-15:00 Great Opening and Industry sessions
Time
13:30-14:00 Minister Luís Roberto Barroso (Brazilian Supreme Court)
14:00-14:30 Fabio Cozman (Industry Chair)
14:30-15:00 Sponsors and discussion
15:00-15:30 Awards ceremony
15:30-Various Workshops
Schedule independently defined by the organizers.
6 hours - 2nd half
LEGAL AIIA AI and Intelligent Assistance for Legal Professionals in the
Digital Workplace
JUL.IA Artificial Intelligence in Jurisdictional Logistics
COLIEE Competition on Legal Information Extraction/Entailment
3 hours
COPYRIGHT Copyright Regulation of Inputs and Outputs of AI Systems
XAILA EXplainable & Responsible AI in Law

Tuesday, 22 June – Main conference

9:00-10:30 Session 1
Time Page
09:00 On Semantics-based Minimal Revision for Legal Reasoning 50
Fungwacharakorn, Wachara; Tsushima, Kanae; Satoh, Ken
09:30 Converting Copyright Legislation into Machine-Executable Code: In- 139
terpretation, Coding Validation and Legal Alignment
Witt, Alice; Huggins, Anna; Governatori, Guido; Buckley, Joshua
10:00 Unravel Legal References in Defeasible Deontic Logic 69
Governatori, Guido; Olivieri, Francesco

vii
10:30-11:00 Short break / Networking space
11:00- Session 2
Time Page
11:00 Hardness of Case-Based Decisions: a Formal Theory 149
Zheng, Heng; Grossi, Davide; Verheij, Bart
11:30 Precedential Constraint: The Role of Issues 12
Bench-Capon, Trevor; Atkinson, Katie
12:00 Incorporating Domain Knowledge for Extractive Summarization of Le- 22
gal Case Documents
Bhattacharya, Paheli; Poddar, Soham; Rudra, Koustav; Ghosh, Kri-
pabandhu; Ghosh, Saptarshi
12:30:13:30 Break
13:30-14:30 Session 3
Time Page
13:30 A dynamic model for balancing values 89
Maranhão, Juliano; Souza, Edelcio; Sartor, Giovanni
14:00 On Semantics-based Minimal Revision for Legal Reasoning 50
REC. Fungwacharakorn, Wachara; Tsushima, Kanae; Satoh, Ken
14:30-15:30 Keynote Speaker – iRobot: how to use Robotic Process Automation to
automate certain legal work
15:30-16:00 Short break / Networking space
16:00-16:55 Session 4
Time Page
16:00 Incorporating Domain Knowledge for Extractive Summarization of Le- 22
gal Case Documents
REC. Bhattacharya, Paheli; Poddar, Soham; Rudra, Koustav; Ghosh, Kri-
pabandhu; Ghosh, Saptarshi
16:30 To Tune or Not To Tune? Zero-shot Models for Legal Case Entailment 295
Rosa, Guilherme Moraes; Rodrigues, Ruan Chaves; Lotufo, Roberto;
Nogueira, Rodrigo
16:45 Interactive System for Arranging Issues based on PROLEG in Civil Lit- 273
igation
Satoh, Ken; Takahashi, Kazuko; Kawasaki, Tatsuki
16:50 Live Demonstration of a Working Collaborative e-Negotiaton System 275
(Smartsettle Infinity)
Ross, Graham Laurence; Thiessen, Ernest
16:55-17:30 Break

18:30-20:00 Session 5

viii
 Time Page
18:30 Precedential Constraint: The Role of Issues 12
REC. Bench-Capon, Trevor; Atkinson, Katie
19:00 BERT-based Ensemble Methods with Data Augmentation for Legal Tex- 278
tual Entailment in COLIEE Statute Law Task
Yoshioka, Masaharu; Aoki, Tasuhiro; Suzuki, Youta
19:15 Legal Norm Retrieval with Variations of the BERT Model Combined 285
with TF-IDF Vectorization
Wehnert, Sabine Sarah; Sudhi, Viju; Dureja, Shipra; Kutty, Libin
Johnny; Shahania, Saijal; De Luca, Ernesto William
19:30 Toward Summarizing Case Decisions via Extracting Argument Issues, 250
Reasons, and Conclusions
Xu, Huihui; Savelka, Jaromir; Ashley, Kevin
19:45 Practical Tools from Formal Models: The ECHR as a Case Study 170
Atkinson, Katie; Collenette, Joe; Bench-Capon, Trevor; Dzehtsiarou,
Kanstantsin
20:30-21:45 Session 6
Time Page
20:30 Hardness of Case-Based Decisions: a Formal Theory 149
REC. Zheng, Heng; Grossi, Davide; Verheij, Bart
21:00 When Does Pretraining Help? Assessing Self-Supervised Learning for 159
Law and the CaseHOLD Dataset of 53,000+ Legal Holdings
Guha, Neel; Zheng, Lucia; Anderson, Brandon Ray; Henderson, Pe-
ter; Ho, Daniel En-Wenn
21:30 Modelling Legal Procedures 220
Rotolo, Antonino; Smith, Clara
21:45 Towards compliance checking in reified I/O logic via SHACL 215
Robaldo, Livio

Wednesday, 23 June – Main conference

09:00-10:30 Session 1
Time Page
09:00 A Combined Rule-Based and Machine Learning Approach for Auto- 40
mated GDPR Compliance Checking
El Hamdani, Rajaa; Mustapha, Majd; Restrepo Amariles, David;
Troussel, Aurore; Meeus, Sébastien; Krasnashchok, Katsiaryna
09:30 The Burden of Persuasion in Structured Argumentation 180
Calegari, Roberta; Riveret, Regis; Sartor, Giovanni
09:45 Discovering the Rationale of Decisions: Towards a Method for Aligning 235
Learning and Reasoning
Steging, Cornelis Cor; Renooij, Silja; Verheij, Bart
10:00 A dynamic model for balancing values 89
REC. Maranhão, Juliano; Souza, Edelcio; Sartor, Giovanni
10:30-11:00 Short break / Networking space

ix
11:00-11:40 Session 2
Time Page
11:00 Applying Decision Tree Analysis to Family Court Decisions: Factors 258
Determining Child Custody in Taiwan
Huang, Sieh-Chuen; Shao, Hsuan-Lei; Leflar, Robert B
11:05 Constraint Answer Set Programming as a Tool to Improve Legislative 262
Drafting: A Rules as Code Experiment
Morris, Jason Patrick
11:10 CriminelBART: A French Canadian Legal Language Model Specialized 256
in Criminal Law
Garneau, Nicolas; Gaumond, Eve; Lamontagne, Luc; Déziel, Pierre-
Luc
11:15 Sentence Classification for Contract Law Cases: A Natural Language 260
Processing Approach
Mok, Wai Yin; Mok, Jonathan R.; Mok, Rachel V.
11:20 Labels distribution matters in performance achieved in legal judgment 268
prediction task
Salaün, Olivier; Langlais, Philippe; Benyekhlef, Karim
11:25 Pathways to Legal Dynamics in Robotics 266
Rotolo, Antonino; Tamargo, Luciano H.; Martìnez, Diego C.
11:30 A simple mathematical model for the legal concept of balancing of in- 270
terests
Zufall, Frederike; Kimura, Rampei; Peng, Linyu
11:35 Predicting Legal Proceedings Status: Approaches Based on Sequential 264
Text Data
Polo, Felipe Maia; Ciochetti, Itamar; Bertolo, Emerson
11:40-13:00 Break

13:00-14:30 Session 3
Time Page
13:00 Automatic Extraction of Amendments from Polish Statutory Law 225
Smywiński-Pohl, Aleksander; Piech, Mateusz; Kaleta, Zbigniew;
Wróbel, Krzysztof
13:15 Enhancing a Recidivism Prediction Tool With Machine Learning: Ef- 210
fectiveness and Algorithmic Fairness
Karimi-Haghighi, Marzieh; Castillo, Carlos
13:30 Converting Copyright Legislation into Machine-Executable Code: In- 139
terpretation, Coding Validation and Legal Alignment
REC. Witt, Alice; Huggins, Anna; Governatori, Guido; Buckley, Joshua
14:00 Unravel Legal References in Defeasible Deontic Logic 69
REC. Governatori, Guido; Olivieri, Francesco
14:30-15:30 Keynote Speaker – Provably Beneficial Artificial Intelligence

15:30-16:00 Short break / Networking space

x
16:00-17:00 Session 4
Time Page
16:00 A Combined Rule-Based and Machine Learning Approach for Auto- 40
mated GDPR Compliance Checking
El Hamdani, Rajaa; Mustapha, Majd; Restrepo Amariles, David;
Troussel, Aurore; Meeus, Sébastien; Krasnashchok, Katsiaryna
16:30 When Does Pretraining Help? Assessing Self-Supervised Learning for 159
Law and the CaseHOLD Dataset of 53,000+ Legal Holdings
REC. Guha, Neel; Zheng, Lucia; Anderson, Brandon Ray; Henderson, Pe-
ter; Ho, Daniel En-Wenn
17:00-18:30 Break
18:30-20:00 Session 5
Time Page
18:30 Context-Aware Legal Citation Recommendation using Deep Learning 79
Huang, Zihan; Low, Charles; Teng, Mengqiu; Zhang, Hongyi; Ho,
Daniel E.; Krass, Mark; Grabmair, Matthias
19:00 From Data to Information: Automating Data Science to Explore the U.S. 119
Court System
Li Zhao, Andong L.; Pack, Harper; Servantez, Sergio; Adler, Rachel
F.; Sterbentz, Marko; Pah, Adam; Schwartz, David; Barrie, Cameron;
Einarsson, Alexander; Hammond, Kristian
19:30 Case-level Prediction of Motion Outcomes in Civil Litigation 99
McConnell, Devin J.; Zhu, James; Pandya, Sachin S.; Aguiar, Derek
Cole
20:00-20:30 Short break / Networking space
20:30-21:30 Session 6
Time Page
20:30 Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdic- 129
tions, and Legal Domains
Savelka, Jaromir; Westermann, Hannes; Benyekhlef, Karim; Alexan-
der, Charlotte S.; Grant, Jayla C.; Amariles, David Restrepo; El-
Hamdani, Rajaa; Meeus, Sebastien; Troussel, Aurore; Araszkiewicz,
Michal; Ashley, Kevin D.; Ashley, Alexandra; Branting, Karl L.; Fal-
duti, Mattia; Grabmair, Matthias; Harasta, Jakub; Novotna, Tereza;
Tippett, Elizabeth; Johnson, Shiwanni
21:00 Plum2Text: A French Plumitifs–Descriptions Data-to-Text Dataset for 200
Natural Language Generation
Garneau, Nicolas; Gaumond, Eve; Lamontagne, Luc; Déziel, Pierre-
Luc
21:15 Process Mining-Enabled Jurimetrics: Analysis of a Brazilian Court’s 240
Judicial Performance in the Business Law Processing
Unger, Adriana Jacoto; dos Santos Neto, Jose Francisco; Trecenti,
Julio; Hirota, Renata; Fantinato, Marcelo; Peres, Sarajane Marques

xi
Thursday, 24 June – Main conference
09:00-10:30 Session 1
Time Page
09:00 Explainable Artificial Intelligence, lawyer’s perspective 60
Górski, Łukasz; Ramakrishna, Shashishekar
09:30 Evaluating Document Representations for Content-based Legal Litera- 109
ture Recommendations
Ostendorff, Malte; Ash, Elliott; Ruas, Terry; Gipp, Bela; Moreno-
Schneider, Julian; Rehm, Georg
10:00 Structural Text Segmentation of Legal Documents 2
Aumiller, Dennis; Almasian, Satya; Lackner, Sebastian; Gertz,
Michael
10:30-11:00 Short break / Networking space
11:00-12:00 Session 2
Time Page
11:00 AI systems and product liability 32
Borges, Georg
11:30 A Dataset for Evaluating Legal Question Answering on Private Inter- 230
national Law
Sovrano, Francesco; Palmirani, Monica; Distefano, Biagio; Sapienza,
Salvatore; Vitali, Fabio
11:45 Making Intelligent Online Dispute Resolution Tools available to Self- 195
Represented Litigants in the Public Justice System
Esteban de la Rosa, Fernando; Zeleznikow, John
12:00-12:30 Break
12:30-13:30 IAAIL General Meeting
13:30-14:30 Session 3
Time Page
13:30 Structural Text Segmentation of Legal Documents 2
REC. Aumiller, Dennis; Almasian, Satya; Lackner, Sebastian; Gertz,
Michael
14:00 Anonymization of German Legal Court Rulings 205
Glaser, Ingo; Schamberger, Tom; Matthes, Florian
14:15 Regulating Artificial Intelligence: A Technology Regulator’s Perspective 190
Ellul, Joshua; McCarthy, Stephen; Sammut, Trevor; Brockdorff,
Juanita; Scerri, Matthew; Pace, Gordon J.
14:30-15:30 Presidential Address – The Winter, The Summer and The Summer Dream
of AI in LAW

15:30-16:00 Short break / Networking space

xii
16:00-17:15 Session 4
Time Page
16:00 Using Transformers to Improve Answer Retrieval for Legal Questions 245
Vold, Andrew; Conrad, Jack G
16:15 From Data to Information: Automating Data Science to Explore the U.S. 119
Court System
REC. Li Zhao, Andong L.; Pack, Harper; Servantez, Sergio; Adler, Rachel
F.; Sterbentz, Marko; Pah, Adam; Schwartz, David; Barrie, Cameron;
Einarsson, Alexander; Hammond, Kristian
16:45 Case-level Prediction of Motion Outcomes in Civil Litigation 99
REC. McConnell, Devin J.; Zhu, James; Pandya, Sachin S.; Aguiar, Derek
Cole
17:15-18:30 Break
18:30-20:00 Session 5
Time Page
18:30 Context-Aware Legal Citation Recommendation using Deep Learning 79
REC. Huang, Zihan; Low, Charles; Teng, Mengqiu; Zhang, Hongyi; Ho,
Daniel E.; Krass, Mark; Grabmair, Matthias
19:00 Explainable Artificial Intelligence, lawyer’s perspective 60
REC. Górski, Łukasz; Ramakrishna, Shashishekar
19:30 On the relevance of algorithmic decision predictors for judicial decision 175
making
Bex, Floris; Prakken, Henry
19:45 Prediction of monetary penalties for data protection cases in multiple 185
languages
Ceross, Aaron William Karl; Zhu, Tingting
20:00-20:30 Short break / Networking space
20:30-22:00 Session 6
Time Page
20:30 Evaluating Document Representations for Content-based Legal Litera- 109
ture Recommendations
REC. Ostendorff, Malte; Ash, Elliott; Ruas, Terry; Gipp, Bela; Moreno-
Schneider, Julian; Rehm, Georg
21:00 AI systems and product liability 32
REC. Borges, Georg
21:30 Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdic- 129
tions, and Legal Domains
REC. Savelka, Jaromir; Westermann, Hannes; Benyekhlef, Karim; Alexan-
der, Charlotte S.; Grant, Jayla C.; Amariles, David Restrepo; El-
Hamdani, Rajaa; Meeus, Sebastien; Troussel, Aurore; Araszkiewicz,
Michal; Ashley, Kevin D.; Ashley, Alexandra; Branting, Karl L.; Fal-
duti, Mattia; Grabmair, Matthias; Harasta, Jakub; Novotna, Tereza;
Tippett, Elizabeth; Johnson, Shiwanni
22:00-22:15 Closing comments conduced by the program chair

xiii
Friday, 25 June – Workshops, Doctoral Consortium, Closing
Various-14:30 Workshops
Time schedule to be defined by the organizer of each workshop, finishing activities
before 14:30.
6 hours
ASAIL Automated Detection, Extraction and Analysis of Semantic In-
formation in Legal Texts
MWAIL Multilingual Workshop on AI & Law
PATENTS Artificial Intelligence and Patents
3 hours
AILBIZ (Legal International Workshop on A.I. for Understanding the Legal
Business) Business
RELATED RELATED – Relations in the Legal Domain
07:30-09:15 1st Panel
07:30 Opening speeches of the Doctoral Consortium
07:45 Beyond persons and things: the legal status of artificial intel-
ligence
Diana Mocanu
08:15 The Digital Administrative Act
Alexander Stepanov
08:45 Constitutional limits to the use of artificial intelligence in court
proceedings
Elisabeth Paar
09:15-09:30 Short break / Networking space
09:30-11:00 2nd Panel
09:30 An African Perspective on Answering the Ethics Question:
Who Should Make the Rules on Self-Driving Cars?
Okechukwu Effoduh
10:00 Judged by Machines? How do algorithms, now used in crimi-
nal justice, impact on the legitimacy of the system?
Cari Hyde-Vaamonde
10:30 Transactions on Privacy and the Tools that Assist – An Inter-
disciplinary Analysis
Kartik Chawla
11:00-11:30 Long break / Networking space

xiv
11:30-13:00 3rd Panel
11:30 ALGORITHMS (DIS)SERVING JUSTICE: Risk Assessment
Tools in Pre-trial Process
Mina Ilhan
12:00 HONto: A Knowledge Base from Textbooks for Legal Text Re-
trieval and Recommendations
Sabine Wehnert
12:30 ESRA: An End-to-End System for Re-Identification and
Anonymization of Swiss Court Decisions
Joel Niklaus
13:00-13:15 Short break / Networking space
13:15-14:30 4th Panel
13:15 Measurement of Consistency in Judicial Decisions
Aline Macohin
13:45 Artifact design for design-science research on process mining
for legal compliance
Adriana Jacoto Unger
14:15 Doctoral Consortium Best Paper award & ending
14:30-15:30 Closing speeches

xv
Invited Speakers
Provably Beneficial Artificial Intelligence
Professor Stuart Russell
University of California Berkeley, US

Topic Prof. Russell discusses an approach to ethical AI based on the idea that AI
systems should be beneficial to humans, with the key caveat that what counts as
"beneficial" is very unlikely to be fully specified. Issues include social aggregation
and social choice, laws, equity, sadism, pride, envy, and mental integrity.

About Stuart Russell is a Professor of Computer Science at the University of Cali-

fornia at Berkeley, holder of the Smith-Zadeh Chair in Engineering, and Director of
the Center for Human-Compatible AI. He is a recipient of the IJCAI Computers and
Thought Award and from 2012 to 2014 held the Chaire Blaise Pascal in Paris. He is
an Honorary Fellow of Wadham College, Oxford, an Andrew Carnegie Fellow, and
a Fellow of the American Association for Artificial Intelligence, the Association for
Computing Machinery, and the American Association for the Advancement of Sci-
ence. His book "Artificial Intelligence: A Modern Approach" (with Peter Norvig) is
the standard text in AI, used in 1500 universities in 135 countries. His research covers
a wide range of topics in artificial intelligence, with an emphasis on the long-term
future of artificial intelligence and its relation to humanity.

iRobot: how to use Robotic Process Automation to automate cer-

tain legal work
Joseph Cohen
Innovation Lead, Dentons UK & Middle East

Topic Robotic Process Automation (RPA) is usually thought of with respect to back-
office functions such as Finance and HR. Recently, however, Dentons have been putting
this technology into the hands of our junior lawyers and asking them to not only sug-
gest some legal tasks for the software to automate, but to actually do the automation
as well. In this talk Joe will cover: How RPA fits into the wider legaltech landscape;
How RPA actually works; How to choose RPA use cases, and when not to use it; How
best to empower lawyers with the right tools.

About Joe leads the Innovation team for the Dentons UK, Ireland and Middle East
region. This includes responsibility for legal technology pilots and implementations,
as well as other ’innovation culture’ initiatives such as innovation training and design
thinking. Prior to this, as a non-lawyer, Joe held innovation positions at Linklaters
and Slaughter and May, and before that was a technology consultant for Deloitte. Joe
was shortlisted for the recent Law.com Innovation Trailblazer of the Year award.

xvi
Contents

Conference Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iii

Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vi

ICAIL 2021 Program and Schedule of Events . . . . . . . . . . . . . . . . . vii

Invited Speakers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xvi

I Full Papers 1
Structural Text Segmentation of Legal Documents . . . . . . . . . . . . . 2
Dennis Aumiller, Satya Almasian, Sebastian Lackner and Michael Gertz

Precedential Constraint: The Role of Issues . . . . . . . . . . . . . . . . . 12

Trevor Bench-Capon and Katie Atkinson

Incorporating Domain Knowledge for Extractive Summarization of Le-

gal Case Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh and
Saptarshi Ghosh

AI Systems and Product Liability . . . . . . . . . . . . . . . . . . . . . . . . 32

Georg Borges

A Combined Rule-Based and Machine Learning Approach for Automated

GDPR Compliance Checking . . . . . . . . . . . . . . . . . . . . . . . . 40
Rajaa EL HAMDANI, Majd Mustapha, David Restrepo Amariles, Aurore Trous-
sel, Sébastien Meeùs and Katsiaryna Krasnashchok

On Semantics-based Minimal Revision for Legal Reasoning . . . . . . . 50

Wachara Fungwacharakorn, Kanae Tsushima and Ken Satoh

Explainable Artificial Intelligence, Lawyer’s Perspective . . . . . . . . . 60

Łukasz Górski and Shashishekar Ramakrishna

xvii
Unravel Legal References in Defeasible Deontic Logic . . . . . . . . . . . 69
Guido Governatori and Francesco Olivieri

Context-Aware Legal Citation Recommendation using Deep Learning . 79

Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark
S. Krass and Matthias Grabmair

A dynamic model for balancing values . . . . . . . . . . . . . . . . . . . . 89

Juliano Maranhão, Edelcio G. de Souza and Giovanni Sartor

Case-level Prediction of Motion Outcomes in Civil Litigation . . . . . . 99

Devin J. McConnell, James Zhu, Sachin Pandya and Derek Aguiar

Evaluating Document Representations for Content-based Legal Litera-

ture Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
Malte Ostendorff, Elliott Ash, Terry Ruas, Bela Gipp, Julian Moreno-Schneider
and Georg Rehm

From Data to Information: Automating Data Science to Explore the U.S.

Court System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
Andrew Paley, Andong L. Li Zhao, Harper Pack, Sergio Servantez, Rachel F.
Adler, Marko Sterbentz, Adam Pah, David Schwartz, Cameron Barrie, Alexan-
der Einarsson and Kristian Hammond

Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdic-

tions, and Legal Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Jaromir Savelka, Hannes Westermann, Karim Benyekhlef, Charlotte S. Alexan-
der, Jayla C. Grant, David Restrepo Amariles, Rajaa El Hamdani, Sébastien
Meeùs, Aurore Troussel, Michał Araszkiewicz, Kevin D. Ashley, Alexandra
Ashley, Karl Branting, Mattia Falduti, Matthias Grabmair, Jakub Harašta,
Tereza Novotná, Elizabeth Tippett and Shiwanni Johnson

Converting Copyright Legislation into Machine-Executable Code: In-

terpretation, Coding Validation and Legal Alignment . . . . . . . . . 139
Alice Witt, Anna Huggins, Guido Governatori and Joshua Buckley

Hardness of Case-Based Decisions: a Formal Theory . . . . . . . . . . . . 149

Heng Zheng, Davide Grossi and Bart Verheij

When Does Pretraining Help? Assessing Self-Supervised Learning for

Law and the CaseHOLD Dataset of 53,000+ Legal Holdings . . . . . . 159
Lucia Zheng, Neel Guha, Brandon R. Anderson, Peter Henderson and Daniel E.
Ho

xviii
II Short Papers 169
Practical Tools from Formal Models: The ECHR as a Case Study . . . . . 170
Katie Atkinson, Joe Collenette, Trevor Bench-Capon and Kanstantsin Dzeht-
siarou

On the relevance of algorithmic decision predictors for judicial decision

making . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
Floris Bex and Henry Prakken

The Burden of Persuasion in Structured Argumentation . . . . . . . . . 180

Roberta Calegari, Régis Riveret and Giovanni Sartor

Prediction of monetary penalties for data protection cases in multiple

languages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
Aaron Ceross and Tingting Zhu

Regulating Artificial Intelligence: A Technology Regulator’s Perspective 190

Joshua Ellul, Gordon Pace, Stephen McCarthy, Trevor Sammut, Juanita Brock-
dorff and Matthew Scerri

Making Intelligent Online Dispute Resolution Tools available to Self-

Represented Litigants in the Public Justice System . . . . . . . . . . . 195
Fernando Esteban de la Rosa and John Zeleznikow

Plum2Text: A French Plumitifs–Descriptions Data-to-Text Dataset for

Natural Language Generation . . . . . . . . . . . . . . . . . . . . . . . . 200
Nicolas Garneau, Eve Gaumond, Luc Lamontagne and Pierre-Luc Déziel

Anonymization of German Legal Court Rulings . . . . . . . . . . . . . . . 205

Ingo Glaser, Tom Schamberger and Florian Matthes

Enhancing a Recidivism Prediction Tool With Machine Learning: Ef-

fectiveness and Algorithmic Fairness . . . . . . . . . . . . . . . . . . . 210
Marzieh Karimi-Haghighi and Carlos Castillo

Towards compliance checking in reified I/O logic via SHACL . . . . . . . 215

Livio Robaldo

Modelling Legal Procedures . . . . . . . . . . . . . . . . . . . . . . . . . . . 220

Antonino Rotolo and Clara Smith

Automatic Extraction of Amendments from Polish Statutory Law . . . 225

Aleksander Smywiński-Pohl, Mateusz Piech, Zbigniew Kaleta and Krzysztof
Wróbel

xix
A Dataset for Evaluating Legal Question Answering on Private Interna-
tional Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 230
Francesco Sovrano, Monica Palmirani, Biagio Distefano, Salvatore Sapienza
and Fabio Vitali

Discovering the Rationale of Decisions: Towards a Method for Aligning

Learning and Reasoning . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
Cornelis Cor Steging, Silja Renooij and Bart Verheij

Process Mining-Enabled Jurimetrics: Analysis of a Brazilian Court’s Ju-

dicial Performance in the Business Law Processing . . . . . . . . . . . 240
Adriana Jacoto Unger, Jose Francisco dos Santos Neto, Julio Trecenti, Renata
Hirota, Marcelo Fantinato and Sarajane Marques Peres

Using Transformers to Improve Answer Retrieval for Legal Questions . 245

Andrew Vold and Jack G. Conrad

Toward Summarizing Case Decisions via Extracting Argument Issues,

Reasons, and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 250
Huihui Xu, Jaromir Savelka and Kevin D. Ashley

III Extended Abstracts 255

CriminelBART: A French Canadian Legal Language Model Specialized
in Criminal Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
Nicolas Garneau, Eve Gaumond, Luc Lamontagne and Pierre-Luc Déziel

Applying Decision Tree Analysis to Family Court Decisions: Factors De-

termining Child Custody in Taiwan . . . . . . . . . . . . . . . . . . . . 258
Sieh-Chuen Huang, Hsuan-Lei Shao and Robert B Leflar

Sentence Classification for Contract Law Cases: A Natural Language

Processing Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
Jonathan R. Mok, Wai Yin Mok and Rachel V. Mok

Constraint Answer Set Programming as a Tool to Improve Legislative

Drafting: A Rules as Code Experiment . . . . . . . . . . . . . . . . . . 262
Jason Morris

Predicting Legal Proceedings Status: Approaches Based on Sequential

Text Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
Felipe Maia Polo, Itamar Ciochetti and Emerson Bertolo

Pathways to Legal Dynamics in Robotics . . . . . . . . . . . . . . . . . . . 266

Antonino Rotolo, Luciano H. Tamargo and Diego C. Martínez

xx
Labels distribution matters in performance achieved in legal judgment
prediction task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268
Olivier Salaün, Philippe Langlais and Karim Benyekhlef

A simple mathematical model for the legal concept of balancing of in-

terests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
Frederike Zufall, Rampei Kimura and Linyu Peng

IV Demonstrations 272
Interactive System for Arranging Issues based on PROLEG in Civil Lit-
igation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
Ken Satoh, Kazuko Takahashi and Tatsuki Kawasaki

Live Demonstration of a Working Collaborative eNegotiaton System

(Smartsettle Infinity) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
Ernest Thiessen and Graham Ross

V COLIEE Papers 277

BERT-based Ensemble Methods with Data Augmentation for Legal Tex-
tual Entailment in COLIEE Statute Law Task . . . . . . . . . . . . . . 278
Masaharu Yoshioka, Yasuhiro Aoki and Youta Suzuki

Legal Norm Retrieval with Variations of the BERT Model Combined

with TF-IDF Vectorization . . . . . . . . . . . . . . . . . . . . . . . . . . 285
Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania and
Ernesto W. De Luca

To Tune or Not To Tune? Zero-shot Models for Legal Case Entailment . 295
Guilherme Moraes Rosa, Ruan Chaves Rodrigues, Roberto de Alencar Lotufo
and Rodrigo Nogueira

Index of Authors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301

xxi
 Part I

Full Papers
 Structural Text Segmentation of Legal Documents
Dennis Aumiller∗
Institute of Computer Science, Heidelberg University
Heidelberg, Germany
aumiller@informatik.uni-heidelberg.de
Sebastian Lackner
Institute of Computer Science, Heidelberg University
Heidelberg, Germany
lackner@informatik.uni-heidelberg.de
ABSTRACT
The growing complexity of legal cases has lead to an increasing
interest in legal information retrieval systems that can effectively
satisfy user-specific information needs. However, such downstream
systems typically require documents to be properly formatted
and segmented, which is often done with relatively simple pre-
processing steps, disregarding topical coherence of segments. Sys-
tems generally rely on representations of individual sentences or
paragraphs, which may lack crucial context, or document-level
representations, which are too long for meaningful search results.
To address this issue, we propose a segmentation system that can
predict topical coherence of sequential text segments spanning sev-
eral paragraphs, effectively segmenting a document and providing
a more balanced representation for downstream applications. We
build our model on top of popular transformer networks and for-
mulate structural text segmentation as topical change detection, by
performing a series of independent classifications that allow for
efficient fine-tuning on task-specific data. We crawl a novel dataset
consisting of roughly 74,000 online Terms-of-Service documents,
including hierarchical topic annotations, which we use for training.
Results show that our proposed system significantly outperforms
baselines, and adapts well to structural peculiarities of legal doc-
uments. We release both data and trained models to the research
community for future work.1
CCS CONCEPTS
• Applied computing → Law; Annotation; • Information sys-
tems → Document structure.
KEYWORDS
Document Understanding, Outline Generation, Text Segmentation
∗ Both authors contributed equally to this research.
1 https://github.com/dennlinger/TopicalChange
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466085
2
Satya Almasian∗
Institute of Computer Science, Heidelberg University
Heidelberg, Germany
almasian@informatik.uni-heidelberg.de
Michael Gertz
Institute of Computer Science, Heidelberg University
Heidelberg, Germany
gertz@informatik.uni-heidelberg.de
ACM Reference Format:
Dennis Aumiller, Satya Almasian, Sebastian Lackner, and Michael Gertz.
2021. Structural Text Segmentation of Legal Documents . In Eighteenth
International Conference for Artificial Intelligence and Law (ICAIL’21), June
21–25, 2021, Sāo Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https:
//doi.org/10.1145/3462757.3466085
1 INTRODUCTION
Written texts are often a sequence of semantically coherent seg-
ments, designed to create a smooth transition between various
subtopics discussed in a single document. Usually, the information
needs of a user are satisfied by retrieving only the relevant subtopic,
and retrieving the whole document is unwieldy and may result in
information overload [29, 43]. However, the context of a single
subtopic frequently spans multiple sentences and contains local-
ized context, which is crucial for proper understanding. Despite the
clear relevance of segmentation to downstream performance, many
(legal) retrieval systems choose structurally rigid representations of
only a single text element (generally either the full document [8, 28],
or a single paragraph/sentence [34, 42]), disregarding the semantic
coherence. Especially in legal documents, which can be extremely
lengthy and contain domain-specific complexities in their topics,
it is important to suitably represent entire topics in a single cohe-
sive unit. Furthermore, aside from semi-structured legal texts, such
as laws, other documents do not necessarily contain uniform and
easily separable segments, to begin with. Especially input formats
such as PDF or scans frequently lack any sort of meta descriptors
for hierarchical information about the document contents, which
makes this a challenging task. To find a fitting representation that
captures the precise topical context in the text, a robust and flexible
framework to obtain such a structural segmentation is required.
We therefore propose a new approach for the estimation of topic
boundaries to generate more suitable document representations
for the mentioned downstream applications, by considering the
topical coherence of paragraphs. We define a coherent section in a
document as a unit consisting of potentially one or multiple para-
graphs, which together share a common topic. Section boundaries
often coincide with a change in topic and can thus be assumed to
generate candidates for the later segmentation.
Despite their importance, many previous works for structural text
segmentation ignore the notion of paragraphs and focus only on
the granularity of sentences. This is contrary to the nature of writ-
ten text, where paragraphs represent a semantically cohesive unit,
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
Figure 1: Visual cues such as paragraphs often give away a no-
tion of semantic coherence, which is disregarded in sentence-
level models.
which is already available and represents a coarser and more mean-
ingful structure than sentences. In this work, we assume that topic
boundaries generally do not appear in the middle of a paragraph,
and, consequently, operating on paragraph level can reduce the
risk of false-positive segmentations and lower the computation
cost of per sentence prediction. Figure 1 shows how paragraphs
group sentences and divide a text into coherent parts and how by
overlooking this valuable information the structure in the text is
lost to the model.
Focusing on the field of text segmentation in the context of the
Natural Language Processing (NLP), we find an already large body
of existing work. Because no large labeled dataset existed, early
approaches to text segmentation were mainly unsupervised, us-
ing heuristics to identify whether two sentences belong to the
same topic or not. Such approaches either exploit the fact that
topically related words tend to appear in semantically coherent
segments [6, 18, 22, 26, 39], or focus on the representation of text in
terms of latent-topic vectors using topic modeling methods, such
as Latent Dirichlet Allocation [2, 29, 30, 37]. Recently, with the
availability of annotated data, text segmentation has also been for-
mulated as a supervised learning problem. Most methods utilize a
hierarchical neural model, where the lower-level network creates
sentence representations, and a secondary network models the de-
pendencies between the embedded sentences [14, 21]. These models
use sentence dependencies to predict the potential segment bound-
aries. One drawback of these approaches is their sentence-level
granularity, which disregards the paragraph coherence previously
mentioned. This problem is partially solved by hierarchical neu-
ral models, where the dependency between sentences is modeled
in a hierarchical structure, to combine sentence into bigger units.
However, training such models are, due to the different document
lengths, computationally expensive. Moreover, these models fail
to take advantage of larger pre-trained language representations,
such as BERT [9] and RoBERTa [23], which have recently proven
to be valuable feature generators with a low cost of fine-tuning,
advancing the state-of-the-art in several disciplines.
In this paper, we aim to tackle the task of structural text segmenta-
tion using transformer-based models, and introduce a novel dataset
of Terms-of-Service documents, containing annotated paragraphs
belonging to the same topic. We focus on the text segmentation
alone and leave the use of segmentation for retrieval enhancement
to future work. We formulate topical coherence as a special case of
binary classification of Same Topic Prediction (STP) and fine-tune
our transformer-based models to detect paragraphs belonging to
3
Aumiller and Almasian, et al.
sections with similar topics. We consider sections as the top-level
hierarchy in our dataset and do not consider subsections. We as-
sume topical independence between consecutive paragraphs and
show that it does not lead to a deterioration of performance while
avoiding the costly computation of hierarchical models. The hy-
pothesis is that by fine-tuning the paragraph embedding for topic
similarity we can generate segment features that detect coherent
topical structures in a document. We evaluate our models against
the traditional embedding baselines and compare them to super-
vised and unsupervised approaches for text segmentation and find
significant improvements by our method.
Contributions. The contributions of this paper are as follows:
(𝑖) We present the task of structural text segmentation on coarser
cohesive text units (paragraphs/sections). (𝑖𝑖) We investigate the
performance of transformer-based models for topical change detec-
tion, and (𝑖𝑖𝑖) frame the task as a collection of independent binary
predictions, reducing overhead for hierarchical training and sim-
plified training sample generation. (𝑖𝑣) We present a new dataset
consisting of online Terms-of-Service documents partitioned into
hierarchical sections, and make the data available for future re-
search. (𝑣) We evaluate our model against classical baselines for
text segmentation, and (𝑣𝑖) show the effectiveness of our gener-
ated embeddings for structural segmentations to obtain superior
performance to other text segmentation techniques.
2 RELATED WORK
Our work is closely tied to the broader fields of legal document
understanding, topic analysis, text segmentation, and transformer
language models, and we briefly review related work in each of
these areas in this section.
2.1 Legal Document Understanding
The area of legal document understanding and legal information
retrieval has a long-standing history. A great overview is presented
by Moens [31], who details more of the applications that were
mentioned in the introduction. While there is existing work on
the topic of legal document segmentation [24, 25, 27], they are
generally concerned with a specific information extraction inter-
est. In the case of Mencía, they are concerned with metadata of
French law documents [27]. Specifically, they also make use of an
existing HTML/XML structure in their input documents, but do
not generalize to arbitrary text inputs without structural features.
Lu et al., on the other hand, are utilizing clustering techniques to
identify sub-topics in legal documents. However, these topics are
irrespective of actual section boundaries within the document [24].
They further include additional metadata, such as citations, head-
notes, and key numbers, in their task setup, which are suitable to
their specific application in case law. Similarly, Conrad et al. [8]
have previously attempted to employ clustering for heterogeneous
document collections, again focusing on hierarchical representa-
tions in the form of short topic descriptors and not focusing on
the actual textual content of the documents themselves. Lyte and
Branting [25] classify metadata labels based on a CRF (Conditional
Random Field) model, building on prior work by Branting [3] in the
same direction, but focus mostly on element classification. Slightly
Structural Text Segmentation of Legal Documents
longer segments in the form of entire sentences are both used by
Poudyal et al. [34], who mine arguments from European case-law
decision, and Westermann et al. [42], where a system for efficient
similarity search based on sentence embeddings is presented.
2.2 Topic Analysis
Detection and analysis of topical change are grounded in topic mod-
eling approaches. Earlier work such as LDA [2] treat documents as
bag-of-words, where each document is assigned to a topic distribu-
tion, and each topic is a distribution over all words. More recent
work has adopted a more sophisticated representation than bag-of-
words and generally models Markovian topic or state transitions to
capture dependencies between words in a document [15, 41]. With
the rise of distributed word representation, the focus has shifted
to the combination of LDA and word embeddings [11, 32]. Since
we are interested in a primary segmentation without necessarily
predicting topics, we put a stronger focus on the related work of
segmentation methods, as discussed in the following section.
2.3 Text Segmentation
Text Segmentation is the task of dividing a document into a multi-
paragraph discourse unit that is topically coherent, with the cut-off
point usually indicating a change in topic [17, 39]. Although the
task itself dates back to 1994 [17], most existing text segmentation
datasets are small and limit their scope to sentences (predicting
whether two sentences discuss the same topic or not). The most
common dataset is by Choi [7], containing only 920 synthesized pas-
sages from the Brown corpus. Choi’s method (C99) is a probabilistic
algorithm measuring similarity via term overlap. GraphSeg [13] is
an unsupervised graph method that segments documents using a
semantic relatedness graph of a document. GraphSeg is also evalu-
ated on a small set of 5 manually-segmented political manifestos
from the Manifesto project2 . Another class of methods are topic-
based document segmentations, which are statistical models that
find latent topic assignments reflecting the underlying structure
of the document [1, 4, 5, 10, 29, 38]. TopicTiling [38] performs best
among this family of methods and uses LDA to detect topic shifts,
with computing similarities between adjacent blocks based on their
term frequency vectors. Brants et al. [4] follow a similar approach
but employ PLSA [19] to compute the estimated word distributions.
Another noteworthy approach based on Bayesian topic models is
by Chen et al. [5], where they constrain latent topic assignments to
reflect the underlying organization of document topics. They also
publish a test dataset with 218 Wikipedia articles about cities and
chemical elements.
All mentioned methods are unsupervised learning approaches, and
small annotated datasets are only used for evaluation and, hence,
are not directly comparable to our approach. Instead, we focus on
supervised learning of topics and introduce a new dataset with
43,056 automatically labeled documents.
The only two comparable supervised approaches are from Koshorek
et al. [21] and Glavas et al. [14]. Koshorek et al. [21] propose a
hierarchical LSTM architecture for learning sentence representation
and their dependencies. They train their hierarchical model on a
dataset of cleaned Wikipedia articles, called Wiki-727k. Glavas
2 https://manifestoproject.wzb.eu
4
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
et al. [14] introduce Coherence-Aware Text Segmentation, which
encodes a sentence sequence using two hierarchically connected
transformer networks. The two latter models are closest to our
work in terms of data size and problem formulation. However, they
rely solely on per sentence predictions, which is incomparable to
our paragraph-based method. The model by Glavas et al. is similar
to our approach in that it is based on a transformer architecture, yet,
they do not take advantage of transfer learning from pre-trained
language models and learn all the features from scratch. Finally,
Zhang et al. [45] extend text segmentation by outline generation
and trained an end-to-end LSTM-model for identifying sections
and generating corresponding headings for Wikipedia documents.
2.4 Transformer Language Models
The transformer architecture, much like recurrent neural networks,
aims to solve sequence-to-sequence tasks, relying entirely on self-
attention to compute representations of its input and output [40].
Transformers have made a significant step in bringing transfer
learning to the NLP community, which allows the easy adaptation
of a generically pre-trained model for specific tasks. Pre-trained
models such as BERT, GPT-2, and RoBERTa [9, 23, 35] use lan-
guage modeling for pre-training on large corpora. These models
are powerful feature generators, which with minimal task-specific
fine-tuning achieve state-of-the-art performance on a wide variety
of tasks. Although at the core of all these models lies the idea of
transformers and attention mechanisms, many have been modi-
fied and optimized to fit various downstream applications. One
variation based on BERT is Sentence-BERT [36], which combines
two BERT-based models in Siamese fashion to derive semantically
meaningful sentence embeddings. By its design, Sentence-BERT
also allows for longer input sequences for pairwise training tasks
and outperforms BERT on semantic textual similarity tasks, mak-
ing it a suitable choice for embedding paragraphs. Another notable
variant of BERT is RoBERTa, a retraining of BERT with improved
training methodology and more training data, it achieves slightly
better results than BERT on some natural understanding tasks. Due
to the advantages of RoBERTa, we chose RoBERTa and Sentence-
RoBERTa from the Sentence-BERT variant for the setup in our
approach.
3 SAME TOPIC PREDICTION
We formulate structural text segmentation as a supervised learning
task of the same topic prediction. Our model consists of two steps:
(𝑖) Independent and Identically Distributed Same Topic Prediction
(IID STP) and (𝑖𝑖) Sequential inference over a full document. As
mentioned previously, sections are the considered level of hierarchy
in our model and the structure of sub-sections is ignored in this
study. However, the model is easily adaptable to any granularity, and
our dataset contains information for all the levels. In the first step,
we fine-tune transformer-based models to detect topical change
for both paragraphs and entire sections. Given two paragraphs or
sections, the classifier should correctly identify if they discuss the
same subject or not. We assume that the topic of each paragraph
or section is independent of the text before and after, meaning
that the topic of one paragraph does not affect the likelihood of
the next paragraph belonging to the same topic. We later prove
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
that this assumption yields good performance without a costly
training of hierarchical models. In the second step, we use the
fine-tuned transformer-based classifiers for sequential inference
on entire documents, where the segment boundaries are defined
by topical change. In the following, we discuss these steps in more
detail.
3.1 IID Same Topic Prediction (STP)
A document 𝑑 ∈ 𝐷 is represented as a sequence of 𝑁 sections
𝑆𝑑 = (𝑠 1, ..., 𝑠 𝑁 ), where each section is assigned one of 𝑀 topics
𝑇 = (𝑡 1, ..., 𝑡𝑀 ), and each section contains up to 𝐾 paragraphs
𝑃𝑛 = (𝑝 1, ....𝑝𝐾 ). We assume topical consistency within a paragraph
and argue that the results for classification do not change based
on the position of the paragraph in the document, since the most
relevant part for our inference is the intra-section information.
Therefore, all paragraphs in a section belong to the same topic. If
the topic assignment is defined by the function 𝑇𝑜𝑝𝑖𝑐, we have:
𝑠𝑛 = (𝑝 1, .., 𝑝𝑘 ) ∧ 𝑇𝑜𝑝𝑖𝑐 (𝑠𝑛 ) = 𝑡 1 =⇒
𝑇𝑜𝑝𝑖𝑐 (𝑝 1 ) = 𝑡 1 = ... = 𝑇𝑜𝑝𝑖𝑐 (𝑝𝑘 ) = 𝑡 1 (1)
If we define 𝐶 as a chunk of text corresponding to either a section
or paragraph, the topic prediction task is defined for section and
paragraph granularity as follows: Given two chunks of text of the
same type (both paragraphs or both sections) (𝑐 1, 𝑐 2 ) and labels 𝑦 ∈
{0, 1}, indicating whether the two chunks belong to the same topic,
topical change detection can be formulated as a binary classification
problem. The positive class indicates that both chunks have the
same topic, whereas the negative class indicates a change in topic
and potentially the beginning of a new segment in text. Note that we
only consider chunks of the same type, namely, either only sections
or only paragraphs, in each model. By formulating the problem
as a binary classification, detecting the topic consistency between
two chunks of text can now be solved with any type of classifier.
In this work, we train two types of transformer-based classifiers
for this task, one from the pre-trained language models [23] and
another Siamese network [36] variation, which is more suitable for
encoding pairwise similarity. Subsequently, the two variations are
discussed.
3.1.1 RoBERTa is a replication study of BERT pre-training with
optimized hyper-parameters that applies minor adjustments to the
BERT language model to achieve better performance [23]. BERT
and RoBERTa both belong to the family of pre-trained transformer-
based language models. The transformer is an architecture for
shaping one sequence into another one with the help of the self-
attention mechanism, which helps the model to extract features
from each word relative to all the other words in the sequence. The
encoder stacks in BERT and RoBERTa consist of one or multiple
self-attention blocks followed by a feed-forward network. During
pre-training, two sentences are taken as input, and models are
trained on two tasks of language modeling, by predicting masked
words in the input and next sentence prediction, and by classi-
fying whether the two sentences are sequential. By these means,
the models learn task-independent features from a vast amount of
unlabelled text that can then be used in a fine-tuning stage for vari-
ous natural language understanding tasks. Since the performance
difference between most transformer-based language models is
5
Aumiller and Almasian, et al.
𝑃1
𝑇 𝐹 (𝑃1, 𝑃2 ) = 1
𝑃2
𝑇 𝐹 (𝑃2, 𝑃3 ) = 1
𝑃3
𝑃4 𝑇 𝐹 (𝑃3, 𝑃4 ) = 0
Figure 2: Demonstration of how the transformer classifier is
used during inference, by comparing consecutive paragraphs
to detect section boundaries.
negligible, we choose RoBERTa as the representative of this fam-
ily. In the fine-tuning process, the model receives two chunks as
input and learns to predict whether they belong to the same topic
or not. To distinguish between two chunks in training a [CLS]
token is inserted at the beginning of the first chunk and a [SEP]
token is inserted at the end of both the first and second chunk. The
embedding of the [CLS] is what is used for pre-training the next
sentence prediction task and contains RoBERTa’s understanding
at the sentence-level. This token is used by a simple classification
layer, learned during fine-tuning, for the same topic prediction task.
Since the input size for both chunks combined is limited to a maxi-
mum of 512 tokens, shorter than many sections and paragraphs in
our dataset, any longer chunk of text has to be truncated to fit.
3.1.2 Sentence-Transformers (SRoBERTa) aims to enhance the sen-
tence embeddings by modification of RoBERTa using a Siamese
architecture to derive semantically meaningful sentence embed-
dings [36]. Their method is available for several transformer models.
We choose a RoBERTa-based variant to make the results compara-
ble to the first approach. SRoBERTa enables RoBERTa to be used
for certain new tasks, such as large-scale semantic similarity com-
parison. Their modifications result in faster inference and better
representation for sentence-pair tasks. Moreover, because of the
Siamese structure and coupling of two RoBERTa networks, the in-
put size doubles, which allows for longer sequences and thus more
context. In this setup, each sentence is passed through a separate
RoBERTa network with an input limit of 512 tokens. The sentence
embeddings are derived from a pooling operation over the output
of two models with tied weights. Sentence-Transformers introduce
several learning objectives, out of which we use the classification
objective function with binary cross-entropy loss to classify the
chunks into the same topics.
3.2 Sequential Inference
For inference, we use the classifiers of the previous step as topic
change detectors for text segmentation. We read each paragraph
of the document sequentially and classify the adjacent paragraphs
for topical mutuality. More concretely, given a document 𝑑 ∈ 𝐷
divided into consecutive paragraphs 𝑃 = (𝑝 1, ....𝑝𝑘 ), section breaks
are marked as where the paragraph’s topic changes. Considering a
transformer 𝑇 𝐹 as our classifier and two consecutive paragraphs
Structural Text Segmentation of Legal Documents
as our input, the classifier outputs the probability of the two para-
graphs belonging to the same topic, independent of their surround-
ing context, e.g., 𝑇 𝐹 (𝑝 1, 𝑝 2 ) = 𝑃 (𝑇𝑜𝑝𝑖𝑐 (𝑝 1 ) = 𝑇𝑜𝑝𝑖𝑐 (𝑝 2 )). There-
fore, given sequences of paragraphs 𝑝 1, ....𝑝𝑘 , and the corresponding
predicted labels 𝑦 = (𝑦1, ..., 𝑦𝑘−1 ), a segmentation of the document
is given by 𝑘 − 1 predictions of 𝑇 𝐹 , where 𝑦𝑖 = 0 denotes the end
of a segment by 𝑝𝑖 . It is worth noting that regardless of the chunk
type used during the training of the classifiers (section or para-
graph inputs) the segmentation module operates on paragraphs
only. Figure 2 shows the inference on a sample document with four
paragraphs and two sections, where the paragraph colors show
the topics. The 𝑇 𝐹 classifier is applied on a paragraph pair and
can ideally recognize the topic change from 𝑃3 to 𝑃4 and mark the
beginning of the new section.
3.3 Legal Applications
To put the presented segmentation into a legal context, we focus
on three main application areas: (𝑖) As mentioned, a section-based
semantic segmentation can be used as a pre-processing step for a
passage retrieval context. This, however, would require additional
data with relevance annotations for both sentence- and paragraph-
level relevance to compare the specific benefits of our approach,
which we leave to future work in this area. (𝑖𝑖) However, seman-
tically coherent sections can also be used as a basis for similarity
search. This is especially helpful when looking for, e.g., related
sections in existing contracts [42]. Here, we focus on Terms-of-
Service documents that are widely available, and contain sections
that follow a general pattern of similar topics. (𝑖𝑖𝑖) Lastly, the sec-
tion separation can be used for generating outlines of documents,
which has previously been shown to work well on other domains
such as Wikipedia [45]. During our document crawl, we also en-
countered several documents not including any sectional headings,
which makes it especially hard to understand the legal contexts for
laymen users.
4 TERMS-OF-SERVICE (TOS) DATASET
Due to data governance policies in many countries, it is generally
mandated that commercial websites contain the necessary legal
information for site users. Specifically, these must be easily reach-
able via the landing page, which makes it comparatively easy to
be crawled. For each Terms-of-Service document, we automati-
cally extract the content divided into paragraphs and respective
hierarchical section headings. Further, ToS documents allow us to
experiment with a large-scale dataset that comes with a shared
set of topics, while still maintaining a heterogeneous set of topics
due to the different types of websites. In the following, we will
discuss the detailed mining process, and implicate limitations of
this approach.
4.1 Crawling
As seeds to our crawler, we use the Alexa 1M URL dataset.3 For
each URL in the dataset, we try to access the website both with
and without the www prefix. First, the landing page is downloaded
and parsed using the Beautiful Soup Python package. We then
3 Available at: http://s3.amazonaws.com/alexa-static/top-
1m.csv.zip
6
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
search for hyperlinks with texts Terms of Service, Terms of Use,
Terms and Conditions, and Conditions of Use, and follow them to
get to the respective terms-of-service pages. Levenshtein distance
with a threshold of 0.75 is used to allow for spelling mistakes and
different wording. The raw Hypertext Markup Language (HTML)
content of the Terms-of-Service page is downloaded and stored
for further processing. In case of an error, e.g., if the website is
temporarily unreachable, we retry the same website 2 additional
times before skipping it. The unprocessed dataset contains HTML
code for roughly 74,000 websites. Note that due to limitations of the
current crawler implementation, websites that rely on JavaScript
to display content are not supported.
4.2 Section Extraction
Despite the fact that HTML is a structured format, it is a non-trivial
task to extract text and hierarchies. The main reasons are that Web
pages often contain a lot of boilerplate (e.g., navigational elements,
advertisements, etc.), generally have heterogeneous appearences
and implementations, and that they simply do not always conform
to the HTML standard. Here, only a rough outline of the pipeline
is given. For further reference, please refer to the implementation
in our repository.
Boilerplate Removal. For boilerplate removal, we use the
boilerpipe package by Kohlschütter et al. [20], which is based on
shallow text features for classifying the text elements on a Web
page. The result is an HTML page with all navigational elements,
advertisements, and template code removed. Importantly, relevant
hierarchical information is retained past this step.
HTML Cleanup. To deal with websites that do not conform to HTML
standards, we perform several cleanup steps. This includes, for ex-
ample, fixing mistakes such as text appearing without a correspond-
ing paragraph (<p> tag), or incorrectly nested tags (e.g., section
headings within a <p> tag). We fix such mistakes by adding missing
tags and adjusting nested tags similar to how a web-browser would
interpret the code.
Language Detection. Since the Alexa dataset also contains many
non-English websites, we reject extracted terms-of-service, where
the majority of text most likely has a language different from Eng-
lish. We use the langid Python package for detecting the language
of each individual paragraph (<p> tag).
Extracting Hierarchy. To obtain the hierarchy, we split the docu-
ment into smaller chunks. Splits are done in the following order:
first we split on each section heading (<h1>-<h6> tags), then on
bold text (<b> tag) starting with an enumeration pattern, then on
enumerations (<li> tags), then on underline text (<u> tag) starting
with an enumeration pattern, and lastly on regular text (<p> tag)
starting with an enumeration pattern. To prevent spurious splits,
each criterion is only used if there are at least 5 occurrences within
the document. Each time the document is split, we save the corre-
sponding headings, which then form the hierarchy. As enumeration
patterns we recognize Latin numbers, roman numerals, and letters,
optionally prefixed with Part, Section, or Article. The majority of
documents contain at most two levels of section hierarchy.
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil Aumiller and Almasian, et al.

Table 1: Top 10 section topics by document frequency. Addi- Doc 1 Doc 1 Doc 1
tionally, the number of associated paragraphs is given.

Document Paragraph
Topic Label
Frequency Frequency
limitation of liability 21,317 68,517
indemnification 16,698 25,683
law and jurisdiction 15,113 29,790
links to other websites 13,752 24,727
termination 12,855 33,978
Doc 2 Doc 2 Doc 2
warranty 9,926 41,403
privacy 8,958 25,022
disclaimer 8,575 29,265
general terms 7,936 54,693

4.3 Data Set Statistics

In addition to the full dataset, we provide a cleaned subset for
which we manually grouped sections into distinct topics based on
similar spelling or meaning. We manually merged 554 section titles,
which corresponds to all titles with at least 250 distinct occurrences
in the corpus. After merging, 82 topics were obtained, and only
sections that have at least one of these aliases as a heading were kept.
The dataset contains different levels of section hierarchy. For our
work, we group document content into top-level sections only, any
further hierarchies are discarded, but are present in the raw data and
available for future work. After removing documents without any
valid sections, belonging to the predefined 82 topics, we are left with
approximately 43,000 documents for the same section prediction
task, and around 40,000 documents for our paragraph-level setup.
We randomly split the data 80/10/10 into train, validation, and test
set. The average number of sections per document is 6.56, and each
document consists of 22.32 paragraphs on average, which results
in a mean of 2.92 paragraphs per section. Table 1 shows the top 10
section topic labels. The average number of paragraphs per section
varies between different topics.
5 EVALUATION
We demonstrate the capabilities of transformer-based architectures
for topical change detection using dataset consisting of online
Terms-of-Service (ToS) documents, which was discussed above.
Results are compared for the introduced IID STP task (see Sec-
tion 3.1) as well as a downstream comparison of text segmentation
results to a range of baselines and existing methods. Results show
a great improvement in the performance for all transformer-based
models.
5.1 Evaluation of Models
We compare our methods against a range of baselines, including
averaging over Global Vectors (GLVavg) [33], tf-idf vectors (tf-
idf ), and Bag of Words (BoW ) [16]. For transformer language mod-
els, we evaluate the standard [CLS] sequence classification with
roberta-base (Ro-CLS). For Sentence-Transformers [36] we use
the Siamese transformer setup with a variant of roberta-base
(ST-Ro) and an additional model that has been pre-trained on NLI
Section (S) Random Paragraph (RP) Consecutive Paragraph (CP)
Figure 3: Visualization of three distinct setups for same topic
classification, where the example chunk of text is depicted
in dashed purple, a positive sample in line of green, and a
negative sample in dotted red. Each vertical line is a sen-
tence and the grouping of them shows a paragraph, different
sections in the document are shown with different colors,
denoting the same topic across all paragraphs of the same
section. From left to right: Same Section prediction, Random
Paragraph, and Consecutive Paragraph sampling.
sentence similarity tasks (ST-Ro-NLI ) to investigate performance of
further pretraining.
Transformer models are trained using the popular HuggingFace
transformer library [44] for the [CLS] models, as well as the
sentence-transformers package from [36] for Siamese variants. We
use two Nvidia Titan RTX GPUs for training, and each model vari-
ant has been trained with five different random seeds. Details for
the training parameters can be found in our public repository. Due
to the length limitation of 512 tokens, we employ an iterative trun-
cation strategy for two-sentence inputs. Due to the coupling of two
transformers for the Sentence-Transformers the input size doubles,
accepting a total input of 1024 tokens.
5.2 Prediction Tasks
As previously introduced, we train models with an independent
classification setup, which is generally much faster than more com-
plicated hierarchical sequential models. Specifically, we highlight
the differences in the setup for the same section prediction task,
compared to the two paragraph-based methods. We point out that
results depicted in Table 2, for the prediction accuracy are not
directly comparable between sampling methods, as they gener-
ate different development and test sets based on the employed
sampling strategy. We show in the subsequent section, however,
that downstream performance for the text segmentation is in line
with results on topic prediction. Figure 3 visualizes the different

7
Structural Text Segmentation of Legal Documents
Table 2: Prediction accuracy for the independent topic prediction tasks, Same Topic Prediction (STP), Random Paragraph (RP),
Consecutive Paragraph (CP) with different sampling strategies. Standard deviation is reported over 5 runs and the best model
on each respective set is depicted in bold.
GLV𝑎𝑣𝑔 tf-idf BoW
Dev 89.70 ±.07 82.10 ±.05 50.94 ±.33
STP
Test 90.01 ±.06 82.54 ±.07 51.05 ±.51
Dev 76.63 ±.04 70.94 ±.07 50.34 ±.04
RP
Test 76.16 ±.06 70.41 ±.09 50.31 ±.37
Dev 77.64 ±6.6 74.94 ±.11 56.34 ±.83
CP
Test 78.63 ±6.8 76.17 ±.07 56.58 ±1.1
sampling strategies. In the following, we describe each strategy
in detail and highlight their difference. Across all strategies, we
added three positive and three negative samples for each individual
section/paragraph.
5.2.1 Section (S) Topic Prediction. In this setup, we use sections
as input chunks to the transformer classifiers. The section task
showcases how different levels of granularity can affect outcomes
in the prediction results. Specifically, the extremely long input
sequences test the limits of what transformers can predict from
partial observations since the majority of inputs will be heavily
truncated. To ensure an equal distribution of samples from within
the same and different sections, we match each section with three
samples from the same topic, and three from different topics. The
positive and negative sections can be sampled from a different
document. The important point is that the positive samples should
come from the same topic and negative samples from different ones.
The first column of Figure 3 visualizes the section sampling, where
the first section of 𝐷𝑜𝑐 1 is paired with the second section of 𝐷𝑜𝑐 2
to form a positive sample and the first section of 𝐷𝑜𝑐 2 to form a
negative sample, respectively. The same strategy is employed for
the generation of the development and test set.
Despite the constraints with respect to the input length, we find that
all transformers perform on a near-perfect level, compare Table 2.
Comparing these results to already very well-performing baselines,
we suspect that certain keywords give away similar sections, but
highlight the fact that the explicit representation of different topics
is not given during training in the binary classification task, which
makes this a suitable method for dealing with imbalanced topics.
5.2.2 Random Paragraph (RP) Topic Prediction. In contrast to the
section-level task, we revert to a more fine-grained distinction of
paragraphs in a text. In the Random Paragraph setting, we still
generate samples similarly, meaning we include three paragraphs
from a random document with the same topic and three negative
samples from random paragraphs with different topics. The main
difference between the Section prediction and Random Paragraph
is in the level of granularity and not how the samples are chosen.
The second column of Figure 3 highlights this difference, where the
samples are paragraphs inside the sections rather than the entire
section. Paragraph-based sampling is closer to our inference setup,
where each input document is considered one paragraph at a time.
8
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
Ro-CLS ST-Ro ST-Ro-N
96.42 ±.52 96.38 ±.03 96.39 ±.03
96.58 ±.52 96.45 ±.06 96.46 ±.02
57.63 ±10.4 87.50 ±.13 87.39 ±.08
57.48 ±10.2 87.19 ±.64 86.88 ±.11
89.63 ±.12 91.17 ±.05 91.12 ±.04
90.34 ±.08 91.17 ±.04 91.69 ±.02
However, results show a sharp drop in the performance, which can
come from a much narrower context of the paragraphs, as well as
a differing selection of test samples compared to the section task.
Solely the BoW model seems to be largely unaffected, which is
simply due to its low performance in either setting.
5.2.3 Consecutive Paragraph (CP) Topic Prediction. To boost
performance and account for the coherent structures in the text,
we employ a sampling strategy inspired by Ein Dor et al. [12]. For
their triplet loss, samples are generated inside the same document
only, which can be translated into sampling from intra-document
paragraphs. Note that this strategy also no longer requires any
merging and annotation of topics across documents, as all relevant
information is now contained within a single document. This fact
opens up much larger generation of training data, which we omit
in our current work for the sake of comparability with the RP
model. To generate samples, we look at all paragraphs of a section
and pair them as positive samples. Negative samples are picked
from paragraphs of different sections in the same document. The
third column of Figure 3 depicts the consecutive paragraph setup,
where the samples are limited to paragraphs of 𝐷𝑜𝑐1. Note that
despite their similar setup, results of RP and CP runs in Table 2 are
not evaluated on the same test set and thus are not comparable,
since the test sets are each generated with the respective sampling
strategies (RP or CP) as well. However, we are able to compare
their downstream performance on the subsequently introduced
text segmentation task (see Section 5.3 and Table 3).
The result of different sampling strategies along with the perfor-
mance of the baselines is shown in Table 2, where the transformer-
based models all outperform the baselines by a significant margin.
Among the baselines BoW has the worst performance overall, with
the accuracy close to random, showing that distinct word occur-
rences are not a sufficient indicator. Average GloVe has the best
performance of all baselines, but is still behind the transformers
by a large margin. Despite the NLI-pretrained SRoBERTa model
(ST-Ro-N) achieving better scores than the base model (ST-Ro) for
most setups, the difference is insignificant, indicating that the pre-
training on sentence similarity tasks does not directly influence our
topic prediction setup.
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil Aumiller and Almasian, et al.

Table 3: Boundary error rate 𝑃𝑘 for compared models (lower is Specifically, we compare the paragraph-based training methods
better), based on sampling strategies Random Paragraph (RP), CP and RP. As an evaluation metric, we follow related literature
Consecutive Paragraph (CP) and their Ensemble variates, and adopt the 𝑃𝑘 metric introduced by Beeferman et al. [1], which
RP𝐸𝑛𝑠 and CP𝐸𝑛𝑠 , respectively. Ensemble (”Ens”) predictions is the error rate of two segments at 𝑘 sentences apart being
are obtained by majority voting over model runs. classified incorrectly. We use the default window size of half the
document length for our evaluation, again following related work.
RP CP RP𝐸𝑛𝑠 CP𝐸𝑛𝑠 Furthermore, we count the number of explicit misclassifications,
and use the accuracy 𝑎𝑐𝑐𝑘 of “up to 𝑘 mistakes per document” as
GLV𝑎𝑣𝑔 29.97 ±.09 26.23 ±6.2 29.55 23.06 an evaluation metric. Due to the coarser nature of paragraphs and
tf-idf 39.87 ±.24 29.70 ±.28 39.36 28.60 the lower number of predictions per document compared to the
BoW 45.76 ±.67 43.46 ±1.5 46.20 41.80 sentence-level segmentation, this is a more illustrative metric. This
Random Oracle 35.08 ±.15 - 31.88 - also relates to the “exact match” metric EM𝑜𝑢𝑡𝑙𝑖𝑛𝑒 employed by
Zhang et al. [45], where 𝑎𝑐𝑐 0 = EM𝑜𝑢𝑡𝑙𝑖𝑛𝑒 .
GraphSeg - 32.48 ±.46 - 32.28
Here, we also include the performance of related works where
WikiSeg - 48.29 ±.30 - 48.29
public and up-to-date code repositories are available. Specifically,
Ro-CLS 37.26 ±4.8 15.15 ±.00 41.15 15.15 we compare to the unsupervised segmentation algorithm Graph-
ST-Ro 15.72 ±.11 14.06 ±.14 14.62 13.14 Seg [13], and the supervised model by Koshorek et al. [21], which
ST-Ro-N 15.97 ±.14 13.97 ±.19 14.81 12.95 we dub “WikiSeg”. Both approaches are trained on a sentence-level
approach, though, and predictions have to be translated back to a
Ens consec - - - 12.50 paragraph level for comparison of results. We train each model with
the suggested parameters in their publicly available repositories.
For an additional pseudo-sequential baseline, we use an informed
random oracle that has a-priori information on the number of
topics in the document, and samples from a distribution with
adjusted probability 𝑃 (“next section”) = #𝑠𝑒𝑐𝑡𝑖𝑜𝑛𝑠/#𝑝𝑎𝑟𝑎𝑔𝑟𝑎𝑝ℎ𝑠.
Note that no additional parameters are learned for any model, and
predictions are binarized with a simple 0.5 threshold over the same
topic predictions. We provide ensembling results for the majority
voting decisions by the five seed runs of each model variant (Ens),
which provides further improvements. Best results are obtained
by ensembling all consecutive transformer-based methods (Ens
consec).
Table 3 shows the results of the evaluation, where one can see
that results in the sequential segmentation are directly linked
to the performance on the independent classification task seen
in Table 2. To verify our initial assumption of cross-document
comparability of content from similar sections, we make the
following observations: (𝑖) Evaluation performance for the STP
setup is consistent for both training strategies (RP and CP) when
using Sentence-Transformer models (see Table 2). (𝑖𝑖) Similarly,
both CP and RP-trained Sentence-Transformer segmentations
achieve results within 2 percentage points of the respective 𝑃𝑘
Figure 4: Mistake rate of per-model ensembles, where the scores. (𝑖𝑖𝑖) In general, CP training setup yields slightly better
suffix CP indicates the consecutive paragraph sampling and 𝑃𝑘 scores, likely because the intra-document dependencies are
RP the random paragraph sampling for each model. The captured better with this sampling strategy, which is a more
baseline is Rand Oracle (Random Oracle), GLV𝑎𝑣𝑔 (average appropriate sampling for the segmentation task. (𝑖𝑣) We find
GloVe vectors), tf-idf, and BoW (Bag-of-Words). Ro-CLS is convergence problems for RP training with the [CLS] models, as
the fine-tuned CLS token for Roberta and Sentence trans- well as the tf-idf model. Due to the size of our general training
former models are ST-Ro and ST-Ro-N, where the latter is corups, we therefore conclude that it is realistic to expect topical
pre-trained on NLI task. The best performing model is the similarity within a section, even across documents. However, due
Ens All (Ensemble of all models). to the seemingly inconsistent convergence of RP models, we
caution against blindly using this strategy, especially when dealing
with more heterogeneous corpora. The oracle baseline performs
5.3 Text Segmentation unexpectedly better than both tf-idf and BoW, indicating that
By generating a text segmentation over the paragraphs of a additional information about the sections of a document can
full document, the independent prediction results from the greatly boost task performance, which might be relevant for future
previous section can now be compared across several approaches. work. Additional pre-training of ST models (ST-Ro-N) does not

9
Structural Text Segmentation of Legal Documents
show any significant improvement over the standard ST-Ro models.
To our surprise, sentence-based implementations (GraphSeg and
WikiSeg) show significantly lower performance, and fall even
behind the simpler baselines. For GraphSeg, an unsupervised
segmentation approach, the lack of explicit training on the different
granularity seems to significantly prohibit correct predictions
on longer segments. WikiSeg heavily preprocesses the data and
discards many samples, thus significantly shrinking the training set.
Since performance on the reduced training set is still decent, this
indicates that training a network from scratch is not suitable with
the smaller training set of a reduced corpus and tends to overfit.
We expect a significant increase in performance if the training
would instead be performed without such strict preprocessing
criteria, or continuing fine-tuning on pre-trained weights from a
paragraph-level WikiSeg model. For either baseline model, it is
also important to note that these models predict on the entirety
of the sequence, which theoretically allows information sharing
between different sections in the current sample. However, they
show no improvement over our binary prediction setup which
does not share this information. It would be of interest to compare
results to sequential transformer-based architectures, such as they
are used by Glavas et al. [14]. However, their model again requires
training from scratch, which has proven to be inconsistent in our
experiments with WikiSeg.
Lastly, the plots for 𝑎𝑐𝑐𝑘 for various models in Figure 4 indicate
a correlation between the 𝑎𝑐𝑐𝑘 and 𝑃𝑘 measures, which does not
apply to sentence-level segmentations. Overall, the best-performing
ensembles classify around 25% of documents without any mistake
(𝑎𝑐𝑐 0 ), and around 70% with less than three mistakes (𝑎𝑐𝑐 2 ) over
the entire document. We therefore suggest 𝑎𝑐𝑐𝑘 as an interpretable
addition to the classic evaluation of segmentation approaches when
dealing with paragraph-level segmentations.
6 CONCLUSION AND FUTURE WORK
Despite a multitude of previous works, structural text segmenta-
tion methods have always focused on very finely segmented text
chunks in the form of sentences. In this work, we have shown that
a relaxation of this problem to coarser text structures reduces the
complexity of the problem, while still allowing for semantic segmen-
tation. Further, we reformulate the oftentimes expensive-to-train
sequential setup of text segmentation as a supervised Same Topic
Prediction task, which reduces training time, while allowing for a
near-trivial generation of samples from automatically crawled text
documents. To show the applicability of our method, we present a
new domain-specific and large corpus of online Terms-of-Service
documents, and train transformer-based models that vastly outper-
form a number of text segmentation baselines.
We are currently investigating the setup for deeper hierarchical
sections, which our dataset already contains annotations for, to
see whether such notions can also be picked up by an independent
classifier and benefit a legal retrieval system. Also, the findings from
our Consecutive Paragraph model already indicate that training
requires no further information than the ground truth segmentation,
which can generally be inferred from structured input formats, such
10
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil
as HTML or XML, making this an attractive option for a larger-scale
study of cross-domain document collections. Finally, an interface
build on top of our framework, enabling the users to judge the
usefulness of segmentation for legal use cases, such as a collection
of documents from mergers and acquisitions, could be used to
determine the efficacy of our improved segmentation.
ACKNOWLEDGEMENTS
We thank the anonymous reviewers for their insightful comments.
REFERENCES
[1] Doug Beeferman, Adam L. Berger, and John D. Lafferty. 1999. Statistical Models
for Text Segmentation. Mach. Learn. 34, 1-3 (1999), 177–210.
[2] David M. Blei, Andrew Y. Ng, and Michael I. Jordan. 2003. Latent Dirichlet
Allocation. J. Mach. Learn. Res. 3 (2003), 993–1022.
[3] Luther Karl Branting. 2017. Automating Judicial Document Analysis. In Pro-
ceedings of the Second Workshop on Automated Semantic Analysis of Information
in Legal Texts co-located with the 16th International Conference on Artificial In-
telligence and Law (ICAIL 2017), London, UK, June 16, 2017 (CEUR Workshop
Proceedings, Vol. 2143), Kevin D. Ashley, Katie Atkinson, Luther Karl Branting,
Enrico Francesconi, Matthias Grabmair, Marc Lauritsen, Vern R. Walker, and
Adam Zachary Wyner (Eds.). CEUR-WS.org. http://ceur-ws.org/Vol-
2143/paper2.pdf
[4] Thorsten Brants, Francine Chen, and Ioannis Tsochantaridis. 2002. Topic-based
Document Segmentation with Probabilistic Latent Semantic Analysis. In Proceed-
ings of the 2002 ACM CIKM International Conference on Information and Knowledge
Management, McLean, VA, USA. ACM, 211–218.
[5] Harr Chen, S. R. K. Branavan, Regina Barzilay, and David R. Karger. 2009. Global
Models of Document Structure using Latent Permutations. In Human Language
Technologies: Conference of the North American Chapter of the Association of
Computational Linguistics, Proceedings, 2009, Boulder, Colorado, USA. 371–379.
[6] Freddy Y. Y. Choi. 2000. Advances in Domain Independent Linear Text Segmen-
tation. In Proceedings of the 1st North American Chapter of the Association for
Computational Linguistics Conference (Seattle, Washington) (NAACL 2000). ACL,
USA, 26–33.
[7] Freddy Y. Y. Choi. 2000. Advances in Domain Independent Linear Text Seg-
mentation. In 6th Applied Natural Language Processing Conference, ANLP, Seattle,
Washington, USA, 2000. ACL, 26–33.
[8] Jack G. Conrad, Khalid Al-Kofahi, Ying Zhao, and George Karypis. 2005. Effective
Document Clustering for Large Heterogeneous Law Firm Collections. In The
Tenth International Conference on Artificial Intelligence and Law, Proceedings of the
Conference, June 6-11, 2005, Bologna, Italy, Giovanni Sartor (Ed.). ACM, 177–187.
https://doi.org/10.1145/1165485.1165513
[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proceedings of the 2019 Conference of the North American Chapter of the Associa-
tion for Computational Linguistics: Human Language Technologies, NAACL-HLT ,
Minneapolis, USA, 2019, Volume 1. 4171–4186.
[10] Satya Dharanipragada, Martin Franz, J. Scott McCarley, Salim Roukos, and Todd
Ward. 1999. Story Segmentation and Topic Detection for Recognized Speech.
In Sixth European Conference on Speech Communication and Technology, EU-
ROSPEECH 1999, Budapest, Hungary. ISCA.
[11] Adji B. Dieng, Francisco J. R. Ruiz, and David M. Blei. 2019. Topic Modeling in
Embedding Spaces. CoRR abs/1907.04907 (2019). arXiv:1907.04907
[12] Liat Ein Dor, Yosi Mass, Alon Halfon, Elad Venezian, Ilya Shnayderman, Ranit
Aharonov, and Noam Slonim. 2018. Learning Thematic Similarity Metric from
Article Sections Using Triplet Networks. In Proceedings of the 56th Annual Meeting
of the Association for Computational Linguistics (Volume 2: Short Papers). ACL,
Melbourne, Australia, 49–54.
[13] Goran Glavas, Federico Nanni, and Simone Paolo Ponzetto. 2016. Unsupervised
Text Segmentation Using Semantic Relatedness Graphs. In Proceedings of the Fifth
Joint Conference on Lexical and Computational Semantics, *SEM@ACL, Berlin,
Germany, 2016. The *SEM 2016 Organizing Committee.
[14] Goran Glavas and Swapna Somasundaran. 2020. Two-Level Transformer and Aux-
iliary Coherence Modeling for Improved Text Segmentation. CoRR abs/2001.00891
(2020). arXiv:2001.00891
[15] Thomas L. Griffiths, Mark Steyvers, David M. Blei, and Joshua B. Tenenbaum.
2004. Integrating Topics and Syntax. In Advances in Neural Information Process-
ing Systems 17 [Neural Information Processing Systems, 2004, British Columbia,
Canada]. 537–544.
[16] Zellig S. Harris. 1954. Distributional Structure. WORD 10, 2-3 (1954), 146–162.
[17] Marti A. Hearst. 1994. Multi-Paragraph Segmentation of Expository Text. In 32nd
Annual Meeting of the Association for Computational Linguistics, 1994, Las Cruces,
ICAIL’21, June 21–25, 2021, Sāo Paulo, Brazil Aumiller and Almasian, et al.

New Mexico, USA, Proceedings. ACL, 9–16. Methods in Natural Language Processing and the 9th International Joint Conference
[18] Marti A. Hearst. 1997. TextTiling: Segmenting Text into Multi-Paragraph Subtopic on Natural Language Processing, EMNLP-IJCNLP, Hong Kong, China. ACL, 3980–
Passages. Comput. Linguist. 23, 1 (March 1997), 33–64. 3990.
[19] Thomas Hofmann. 2017. Probabilistic Latent Semantic Indexing. SIGIR Forum 51, [37] Martin Riedl and Chris Biemann. 2012. How Text Segmentation Algorithms Gain
2 (2017), 211–218. from Topic Models. In Human Language Technologies: Conference of the North
[20] Christian Kohlschütter, Peter Fankhauser, and Wolfgang Nejdl. 2010. Boilerplate American Chapter of the Association of Computational Linguistics, Proceedings,
Detection Using Shallow Text Features. In Proceedings of the Third International 2012, Montréal, Canada. ACL, 553–557.
Conference on Web Search and Web Data Mining, WSDM, New York, USA 2010. [38] Martin Riedl and Chris Biemann. 2012. TopicTiling: A Text Segmentation Al-
ACM, 441–450. gorithm based on LDA. In Proceedings of the Student Research Workshop of the
[21] Omri Koshorek, Adir Cohen, Noam Mor, Michael Rotman, and Jonathan Berant. 50th Meeting of the Association for Computational Linguistics. Republic of Korea,
2018. Text Segmentation as a Supervised Learning Task. In Proceedings of the 2018 37–42.
Conference of the North American Chapter of the Association for Computational [39] Masao Utiyama and Hitoshi Isahara. 2001. A Statistical Model for Domain-
Linguistics: Human Language Technologies, NAACL-HLT, New Orleans, Louisiana, Independent Text Segmentation. In Association for Computational Linguistic, 39th
USA, 2018, Volume 2 (Short Papers). ACL, 469–473. Annual Meeting and 10th Conference of the European Chapter, Proceedings of the
[22] Hideki Kozima. 1993. Text Segmentation Based on Similarity between Words. In Conference, 2001, Toulouse, France. ACL, 491–498.
31st Annual Meeting of the Association for Computational Linguistics, 1993, Ohio [40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
State University, Columbus, Ohio, USA, Proceedings. ACL, 286–288. Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All
[23] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer you Need. In Advances in Neural Information Processing Systems 30: Annual Con-
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A ference on Neural Information Processing Systems 2017, Long Beach, CA, USA.
Robustly Optimized BERT Pretraining Approach. CoRR abs/1907.11692 (2019). 5998–6008.
arXiv:1907.11692 [41] Hanna M. Wallach. 2006. Topic modeling: Beyond Bag-of-Words. In Machine
[24] Qiang Lu, Jack G. Conrad, Khalid Al-Kofahi, and William Keenan. 2011. Legal Learning, Proceedings of the Twenty-Third International Conference (ICML), Pitts-
document clustering with built-in topic segmentation. In Proceedings of the 20th burgh, Pennsylvania, USA, 2006 (ACM International Conference Proceeding Series,
ACM Conference on Information and Knowledge Management, CIKM 2011, Glasgow, Vol. 148). 977–984.
United Kingdom, October 24-28, 2011, Craig Macdonald, Iadh Ounis, and Ian [42] Hannes Westermann, Jaromír Savelka, Vern R. Walker, Kevin D. Ashley, and
Ruthven (Eds.). ACM, 383–392. https://doi.org/10.1145/2063576. Karim Benyekhlef. 2020. Sentence Embeddings and High-Speed Similarity Search
2063636 for Fast Computer Assisted Annotation of Legal Documents. In Legal Knowledge
[25] Alex Lyte and Karl Branting. 2019. Document Segmentation Labeling Techniques and Information Systems - JURIX 2020: The Thirty-third Annual Conference, Brno,
for Court Filings. In Proceedings of the Third Workshop on Automated Seman- Czech Republic, December 9-11, 2020 (Frontiers in Artificial Intelligence and Ap-
tic Analysis of Information in Legal Texts co-located with the 17th International plications, Vol. 334), Villata Serena, Jakub Harasta, and Petr Kremen (Eds.). IOS
Conference on Artificial Intelligence and Law (ICAIL 2019), Montreal, QC, Canada, Press, 164–173. https://doi.org/10.3233/FAIA200860
June 21, 2019 (CEUR Workshop Proceedings, Vol. 2385), Kevin D. Ashley, Katie [43] Ross Wilkinson. 1994. Effective Retrieval of Structured Documents. In Proceedings
Atkinson, Luther Karl Branting, Enrico Francesconi, Matthias Grabmair, Bern- of the 17th Annual International ACM-SIGIR Conference on Research and Devel-
hard Waltl, Vern R. Walker, and Adam Zachary Wyner (Eds.). CEUR-WS.org. opment in Information Retrieval. Dublin, Ireland, 1994 (Special Issue of the SIGIR
http://ceur-ws.org/Vol-2385/paper5.pdf Forum). ACM/Springer, 311–317.
[26] Igor Malioutov and Regina Barzilay. 2006. Minimum Cut Model for Spoken [44] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,
Lecture Segmentation. In ACL, 21st International Conference on Computational Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie
Linguistics and 44th Annual Meeting of the Association for Computational Linguis- Brew. 2019. HuggingFace’s Transformers: State-of-the-art Natural Language
tics, Proceedings of the Conference, Sydney, Australia, 2006. ACL. Processing. ArXiv abs/1910.03771 (2019).
[27] Eneldo Loza Mencía. 2009. Segmentation of legal documents. In The 12th Interna- [45] Ruqing Zhang, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi Cheng. 2019.
tional Conference on Artificial Intelligence and Law, Proceedings of the Conference, Outline Generation: Understanding the Inherent Content Structure of Documents.
June 8-12, 2009, Barcelona, Spain. ACM, 88–97. https://doi.org/10.1145/ In Proceedings of the 42nd International ACM SIGIR Conference on Research and
1568234.1568245 Development in Information Retrieval, SIGIR, Paris, France, 2019. 745–754.
[28] Nada Mimouni. 2013. Modeling Legal Documents as Typed Linked Data for
Relational Querying. In Proceedings of the First JURIX Doctoral Consortium and
Poster Sessions in conjunction with the 26th International Conference on Legal
Knowledge and Information Systems, JURIX 2013, Bologna, Italy, December 11-13,
2013 (CEUR Workshop Proceedings, Vol. 1105), Monica Palmirani and Giovanni
Sartor (Eds.). CEUR-WS.org. http://ceur-ws.org/Vol-1105/paper6.
pdf
[29] Hemant Misra, François Yvon, Olivier Cappé, and Joemon M. Jose. 2011. Text
Segmentation: A Topic Modeling Perspective. Inf. Process. Manag. 47, 4 (2011),
528–544.
[30] Hemant Misra, François Yvon, Joemon M. Jose, and Olivier Cappé. 2009. Text
segmentation via Topic Modeling: an Analytical Study. In Proceedings of the 18th
ACM Conference on Information and Knowledge Management, CIKM , Hong Kong,
China, 2009. 1553–1556.
[31] Marie-Francine Moens. 2001. Innovative techniques for legal text retrieval. Artif.
Intell. Law 9, 1 (2001), 29–57.
[32] Christopher E. Moody. 2016. Mixing Dirichlet Topic Models and Word Embed-
dings to Make lda2vec. CoRR abs/1605.02019 (2016). arXiv:1605.02019
[33] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. Glove:
Global Vectors for Word Representation. In Proceedings of the 2014 Conference on
Empirical Methods in Natural Language Processing, EMNLP, 2014, Doha, Qatar, A
meeting of SIGDAT, a Special Interest Group of the ACL. 1532–1543.
[34] Prakash Poudyal, Teresa Gonçalves, and Paulo Quaresma. 2019. Using Clustering
Techniques to Identify Arguments in Legal Documents. In Proceedings of the Third
Workshop on Automated Semantic Analysis of Information in Legal Texts co-located
with the 17th International Conference on Artificial Intelligence and Law (ICAIL
2019), Montreal, QC, Canada, June 21, 2019 (CEUR Workshop Proceedings, Vol. 2385),
Kevin D. Ashley, Katie Atkinson, Luther Karl Branting, Enrico Francesconi,
Matthias Grabmair, Bernhard Waltl, Vern R. Walker, and Adam Zachary Wyner
(Eds.). CEUR-WS.org. http://ceur-ws.org/Vol-2385/paper2.pdf
[35] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya
Sutskever. 2018. Language Models are Unsupervised Multitask Learners. (2018).
[36] Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings
using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Empirical

11
 Precedential Constraint: The Role of Issues
Trevor Bench-Capon and Katie Atkinson
Department of Computer Science, University of Liverpool
Liverpool, United Kingdom
{tbc,katie}@liverpool.ac.uk
ABSTRACT
Horty, Rigoni and Prakken have developed formal characterisa-
tions of precedential constraint based on dimensions and factors as
introduced in HYPO and CATO. We discuss the relation between
dimensions and factors and also describe the current models of
precedential constraint based on factors, along with some criti-
cisms of them. We argue that problems arise from ignoring the
structure of legal cases that is provided by the notion of issues,
and that seeing precedential constraint in terms of issues rather
than whole cases provides a more effective approach and better
reflects legal practice. The advantages of the issue based approach
are illustrated with a concrete example. We then discuss how dimen-
sions should be accommodated, suggesting that this is best done
by seeing reasoning with legal cases as a two stage process: first
factors are ascribed to cases and then factor based reasoning can
be used to arrive at a decision. Thus precedential constraint can be
described in terms of factors, dimensions being handled at the first
stage. Both stages are constrained, in different ways, by precedents:
we identify three types of precedent: framework precedents which
structure cases into issues, preference precedents which resolve
conflicts between opposing sets of factors within these issues, and
ascription precedents which constrain the mapping from facts to
factors.
CCS CONCEPTS
• Applied computing → Law.
KEYWORDS
reasoning with precedents, factors, dimensions, issues
ACM Reference Format:
Trevor Bench-Capon and Katie Atkinson. 2021. Precedential Constraint: The
Role of Issues. In Eighteenth International Conference for Artificial Intelligence
and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York,
NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466062
1 INTRODUCTION
Reasoning with precedent cases has been a central concern of AI
and Law since the very beginning. At least three questions can
be posed in relation to reasoning with precedent cases: (1) how do
people reason with precedents? (2) can we use precedents to predict
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466062
12
the outcome of new cases? and (3) can we formally characterise how
precedents constrain future cases?
An early project addressing precedential reasoning was the
HYPO project of Rissland and Ashley, introduced at the first ICAIL
[43] and most fully described in [6]. HYPO modelled reasoning with
precedents in the US Trade Secrets domain. It influenced a great
deal of research by a number of different researchers, as discussed in
[10], including the CATO system of Ashley and Aleven, introduced
in [4] and most fully described in [3]. CATO also addressed US
Trade Secrets. Both HYPO and CATO were concerned with the first
of our questions: their goal was to show how arguments concerning
new cases can be constructed on the basis of precedent cases, and
how such arguments can be challenged by distinguishing the cited
precedents. These systems presented arguments for and against
particular decisions, but did not attempt to choose between them:
that was left to the judgement of the user.
In contrast, systems based on rules, whether based on expert
knowledge [49], or on a formalisation of legislation [47], or a com-
bination of the two [8], were able to predict the outcome of a new
case entered into the system, answering our second question. It was
therefore a natural development to adapt systems such as CATO
to offer predictions based on reasoning with precedents. This was
done in the Issue Based Prediction system (IBP) [21], in which ar-
guments generated from CATO were organised and evaluated so as
to predict an outcome. Subsequently Grabmair further developed
this approach to accommodate his value judgement formalism [24].
Predictions based on precedent continue to be implemented in both
symbolic systems [2] and machine learning (ML) systems such as
[35] which base their predictions on large collections of case deci-
sions. Factor based reasoning is acquiring an important new role
in explaining the predictions of ML systems (e.g. [19] and [38]).
The reasoning in HYPO and CATO was embodied in algorithms
rather than expressed declaratively and so was not readily amenable
to formalisation to address the third question. This situation was
changed when Prakken and Sartor provided a means of expressing
a case base of precedents as a set of rules and priorities between
them [39]. The resulting rule base could then be deployed to predict
the outcome of a new case. Further, this laid the foundations for
the provision of a formal account of precedential constraint1 . The
work was begun by Horty [26], using a factor based representation
taken from [6] and [3]. His approach was developed in [30] and
extended by Rigoni in [41]. However, it became recognised that
factors were not sufficient to capture all the necessary nuances of
precedents: some aspects of cases can favour a party to different
extents. The need to address dimensions was argued in [15] and
1 These formal accounts consider that a decision is constrained if any other decision
would be inconsistent with past decisions. In practice this constraint may not be
respected in given judicial settings. For a jursiprudential discussion see [46].
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
addressed by Horty in [27], [28] and [29] and by Rigoni in [42]. A
comparison of the approaches of Horty and Rigoni is given in [37].
In this paper we will address the question of how precedents
constrain decisions in new cases, and in particular identify how
domain knowledge can complement the purely formal characterisa-
tions. Section 2 reviews the use of dimensions and factors in HYPO
and CATO to clarify their different roles: whereas dimensions iden-
tify the aspects of cases which must be considered, factors record
their legal significance in the particular case by identifying the
party favoured by that aspect. Section 3 gives an overview of the
formalisations of precedential constraint using factors. In Section 4
we show how these approaches can be improved by exploiting the
structure found in legal cases. Section 5 considers how to accommo-
date dimensions, by considering precedential reasoning as a two
stage process. First factors are ascribed on the basis of dimensional
facts and then these factors supply the reasons to resolve the is-
sues, and hence constrain the overall decision. Different precedents
are relevant at each stage: some constrain the ascription of factors
while others constrain the preferences between sets of factors.
The contributions of the paper are: improvement in the formal
characterisation of precedential constraint, both in terms of effec-
tiveness and in reflecting actual decisions, by applying it to issues
rather than whole cases; clarification of the role of dimensions by
articulating the reasoning process into two distinct stages; and iden-
tifying the need to recognise that precedents operate differently at
the two stages. Throughout the paper we use examples from US
Trade Secrets cases, the most widely discussed domain for reason-
ing with precedents in AI and Law: as well as HYPO and CATO it
has been used in [21], [22], [2], [24], [13], [36], [50] and [38], among
many others.
2 DIMENSIONS AND FACTORS
To relate formal work on precedential constraint to actual legal
cases, it is important to have a clear understanding of factors and
dimensions and the relationship between them. The terms have
been used in different ways, but we will consider dimensions as
used in HYPO and factors as used in CATO, discussed by Rissland
and Ashley in [44]. This is the most common use, and HYPO and
CATO were explicitly identified by Horty in [26] and [30] as the
source of the factors used in his formal account of precedential
constraint, which is the starting point for subsequent discussions
of this topic. Moreover both HYPO dimensions and CATO factors
resulted from thorough domain analyses. Most of the many systems
addressing US Trade Secrets have taken both the analysis of the
domain and the ascription of factors to cases from CATO [3].
In HYPO cases are represented as collections of facts (see Appen-
dix B of [6]). There are thirteen implemented dimensions (Appendix
F of [6]) which may be applicable to a case on the basis of these
facts. In general a dimension can take a range of values, but in fact
ten of the thirteen were two-valued. A list of HYPO’s dimensions,
summarising Appendix F of [6], is given in Table 1.
Dimensions identify the aspects of cases which need to be con-
sidered to see if they are applicable:
Each dimension has prerequisites that must be satis-
fied in order for the dimension to be applicable. For
example, the dimension Secrets-Voluntarily-Disclosed
13
Bench-Capon and Atkinson
has as one of its prerequisites that the plaintiff made
disclosures of confidential information to outsiders.
([44], p 67)).
Although disclosures to outsiders may be a reason to find for the
defendant, the lack of disclosures was never found to be used as a
reason to find for the plaintiff in the analysed cases [44]: because the
plaintiff is expected to take measures to protect the secret, simply
refraining from disclosure seems not to strengthen the plaintiff’s
case. Thus this dimension is not applicable if no disclosures were
made. Typically only a few dimensions will be applicable in any
given case: in HYPO four or five is typical.
2.1 From Dimensions to Factors
Even if applicable, the value on the dimension may be such that
it does not favour either party; the dimension may be neutral in
the particular case. Applicable dimensions must be assessed for
their legal significance for the particular case, that is whether they
favour a party, and if so, which one. This significance is shown by
ascribing a factor to the case. A factor is present if the case lies
within a range on a dimension which favours a particular side. At
one end the dimension will either be inapplicable because it does
not affect the strength of a side’s case, or it will favour a particular
side. Moving along the dimension we may enter a neutral area
favouring neither side, and then an area which favours the other
side. In practice many dimensions have only two points and either
favour a particular side or are inapplicable. For a many-valued
dimension, such as D3d, if sufficient disclosures to provide a reason
for the defendant were made, then the corresponding factor (F10d)
applies. It may be, however, that too few disclosures were made
to favour the defendant (e.g. Emery v Marcan “Even though parts
drawings may on occasion have been shown to a limited number of
outsiders for a particular purpose, this did not in itself necessarily
destroy the secrecy which protected them.”). Here no factor will
apply (although the dimension remains applicable if any disclosures
were made, because whether the factor should be ascribed needs to
be considered). The point about neutrality is made in [44]
Note that CATO does not automatically treat the fact
that a factor does not apply to a case as a strength for
the opponent. ([44], pp 68-9).
As can be seen from Table 1, only one dimension, D13b, Security
Measures, was seen as capable of favouring both sides.
[...] the Security-Measures dimension was broken into
two factors: Security-Measures [F6p], favoring the
plaintiff, and No-Security-Measures [F19d], favoring
the defendant. This was done because judges explic-
itly said that the fact that plaintiff had taken no secu-
rity measures was a positive strength for the oppo-
nent. By contrast, Ashley and Aleven did not create a
“No-Secrets-Disclosed-Outsiders” factor because they
found no cases where judges had said that the absence
of any disclosures to outsiders was a positive strength
for the plaintiff. ([44], p 69).
Thus the security measures dimension is always applicable, al-
though it is possible that neither F6p nor F19d is present: the plaintiff
may have taken sufficient measures to prevent the lack of concern
Precedential Constraint: The Role of Issues ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 1: Dimensions in HYPO and their corresponding CATO factors. Dimension and Factor IDs are D or F for dimension or
factor (factor numbers are those in CATO) followed by p, d, or b to indicate whether it can favour plaintiff, defendant or both.

Plaintiff Defendant
ID Dimension Values Number of Values Factors Factors
in CATO in CATO
Computed from
D1p Competitive Advantage Gained Many F8p
development time and cost
D2d Vertical Knowledge Vertical or technical 2 F11d
D3d Secrets Voluntarily Disclosed Number of Disclosures Many F10d F27d
D4d Disclosures Subject to Restriction Yes or No 2 F12p
D5p Agreement Supported by Consideration Something or Nothing 2
D6p Common Employee Paid to Change Employers Something or Nothing 2 F2d
D7p Exists Express Noncompetition Agreement Yes or No 2 F13p
D8p Common Employee Transferred Product Tools Something or Nothing 2 F7p
D9p Non-Disclosure Agreement Re Defendant Access Yes or No 2 F4p
D10d Common Employee Sole Developer Yes or No 2 F3d
D11d Non-disclosure Agreement Specific Yes or No 2 F5d
D12d Disclosure in Negotiations with Defendant Yes or No 2 F1d
D13b Security Measures Range of possible measures 8 F6p F19d

being a strength for the defendant, but without sufficient rigour
to be a reason to find for the plaintiff. Thus, although it is always
relevant to consider the security measures taken, in many cases
there will be no legal significance. Indeed many cases in CATO
[3] do not have either F6p or F19d. Of the thirteen dimensions
in HYPO, ten, the two-valued dimensions, are either inapplicable
or favour a particular side. Of the three multi-valued dimensions,
two are considered, if applicable, to be either neutral or capable of
favouring only one party (defendant for disclosures, and plaintiff
for competitive advantage). Only security measures is capable of
favouring either side, or being neutral.
Note, however, that in one case, disclosures to outsiders, there are
two pro-defendant factors associated with the dimension. As well
as F10d, SecretsDisclosedOutsiders, we also have F27d DisclosureIn-
PublicForum. This is because F27d provides a much stronger reason
for the defendant than F10d, so that it might be that a plaintiff
factor such as F12p OutsiderDisclosuresRestricted would defeat F10d
but not F27d. Thus a dimension may give rise to multiple factors
favouring the same side.
This understanding of dimensions and factors shows why it is
a mistake to speak of the “negations” of base level factors, as in
some recent formally oriented approaches (e.g. [50], [37]). CATO
used two distinct factors for the rare case where a dimension could
favour either side. Moreover, if a factor is absent, a different factor
favouring that side may be present, as with disclosures. Thus the
absence of F10d might mean that no disclosures had been made, so
that the dimension was inapplicable; that too few disclosures had
been made, meaning that the dimension was not legally significant
in this case; or that disclosures had been made in a public forum,
so that the stronger F27d was present. There seems little sense in
wrapping these three quite different notions under the “negation”
of F10d. Nor is negation needed to distinguish cases where a base
level factor is known absent from those where there is no infor-
mation about that factor. If a base level factor is not mentioned in
the opinion, it played no role in the decision, and hence can safely
be considered absent. The absence of a base level factor does not
provide a reason for the other side, and so its absence will be men-
tioned only when its presence was considered but rejected because
the case fell into a neutral area on an applicable dimension.
In addition to the fourteen factors derived from the HYPO di-
mensions in Table 1, CATO introduced another twelve factors. This
is because CATO analysed considerably more cases than HYPO
and seems to have included more cases questioning whether the
information was a trade secret rather than whether there was a
confidential relationship. In Table 2 we have related these additional
factors to dimensions in the manner of Table 1. These additional
factors can be accommodated in seven dimensions, only one of
which, D14b, has both plaintiff and defendant factors. Three have
multiple factors for the same side. Four are two-valued. The mix is
similar to that found in HYPO, and so may be considered typical.
That the additional cases analysed by CATO led to additional
factors and dimensions is an indication of how precedent cases are
the source of dimensions and factors. The opinions in precedents
show what aspects of the cases judges thought relevant, and what,
if any, significance they accorded to them in that case. Any given
case will only have a few applicable dimensions, and so will only
contain a small subset of possible factors. Therefore as we analyse
more cases we are likely to encounter more dimensions and more
distinctions and hence more factors.
2.2 Arguing with Factors
HYPO and CATO were not concerned with determining or pre-
dicting outcomes, but rather the identification of arguments for
the two parties. These arguments were organised in the “three ply”
structure common in law (e.g. US Supreme Court Oral Argument
and witness testimony which follows the initial questions with a
cross examination and a redirect). In this structure an outcome is

14
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 2: Factors Introduced in CATO Organised into Dimensions. See Table 3 for factor names.
ID Dimension Values
D14b Use Of Available Information Various types of use
D15p Similarity Of Products Degrees of similarity
D16d Availability Of Information Various forms of availability
D17p Invasive Techniques Yes or No
D18p Obtained by Deception Yes or No
D19d Confidentiality Waived Yes or No
D20p Knew Confidential Yes or No
proposed, a response made by the other side followed by a rebut-
tal from the original side. For reasoning with precedents with the
proponent arguing for the plaintiff, these three plies in CATO are:
(1) Cite the precedent case with a decision for the desired side
which has the most factors in common and fewest distin-
guishing factors compared with the current case. The side
favoured by the factors does not matter.
(2) The opponent may distinguish the cited case. Typically the
new case will not contain exactly the same factors as the
precedent. Some of these differences will make the case
stronger for the plaintiff: plaintiff factors in the current case
but not the precedent, and defendant factors in the precedent
but not the current case. The defence will be wise to remain
silent as to these differences. If, however, the precedent con-
tains plaintiff factors not in the current case, or defendant
factors in the current case but not the precedent, the differ-
ences may be significant and so provide an argument not to
follow the cited precedent2 .
(3) The proponent may now attempt a rebuttal: downplaying
distinctions by citing factors favouring the plaintiff (the dif-
ferences the defendant could not use in the second ply).
Assuming that the opponent was able to make some distinctions
in the second ply, it is now up to the user to decide whether, given
the rebuttal, the distinctions are of sufficient weight to merit an
outcome different from the precedent case.
This method of arguing with precedents is the basis of the formal
characterisations of precedential cases discussed in the next section.
Precedents are converted into sets of rules with conjunctions of
factors as antecedents. These rules constrain a new case if there is a
rule applicable to the new case which finds for a particular side (ply
1) which cannot be distinguished (ply 2), and which is preferred to
any applicable rule favouring the other side (ply 3).
3 MODELS OF PRECEDENTIAL CONSTRAINT
HYPO and CATO were realised as programs, with the knowledge
represented as particular data structures (e.g. case frames in HYPO),
and the operation of the reasoning defined in terms of algorithms
manipulating these structures (the algorithms for CATO are given
in Appendix 3 of [3]). As such, reasoning with cases was not readily
amenable to logical analysis until Prakken and Sartor provided a
2 In HYPO and CATO the opponent can also cite counter examples in this ply, but we
will not discuss counter examples in this paper.
15
Bench-Capon and Atkinson
Plaintiff Defendant
Number of Values Factors Factors
in CATO in CATO
Many F14p F16d F25d F17d
Many F15p F18p
Many F20d F24d
2 F22p
2 F26p
2 F23d
2 F21p
means of expressing precedent cases as a sets of rules [39]. Since
the factors for the plaintiff provide a reason to find for the plaintiff
and the factors favouring the defendant a reason to find for the
defendant, the decision in the case can be seen as expressing a pref-
erence for one of these reasons. The conjunction of all the factors
for a side is the strongest reason for that side, so the precedent can
be modelled as a set of three rules expressing that the strongest
reason for the winner was preferred to the strongest reason for the
loser. Where the case comprises a set of factors 𝑃 ∪ 𝐷 where 𝑃 is
the set of plaintiff factors and 𝐷 the set of defendant factors, the
three rules are:
r1: 𝑃 → plaintiff; r2: 𝐷 → defendant;
r3: 𝑟 2 ≺ 𝑟 1 if the decision was for the plaintiff and 𝑟 1 ≺ 𝑟 2 if
the decision was for the defendant.
If we represent all the precedents in the domain using this tech-
nique we can build a logcal theory representing our case base of
precedents. If we are given a new case, we can see whether the rules
apply to it, and if so whether an outcome is determined by the cur-
rent theory. A distinction will mean that the winner’s rule does not
apply or that the loser may have a stronger rule. This representation
was used in [9] in which the possible sets of plaintiff factors were
represented as a partial order, the possible sets of defendant factors
were represented as a partial order, and the precedents as ordering
relations between these two partial orders. The nodes contain all
the possible antecedents for plaintiff and defendant rules and the
arcs show the priorities between particular rules. The example from
[9] is shown in Figure 1. Deciding a new case is now a matter of
adding an arc between the two relevant nodes representing the
factors in the new case and deciding which way the arrow should
point. The constraint is that the arrow should not introduce a cycle
since this would introduce an inconsistency to the case base. Thus
a case which could introduce a cycle is constrained, but if no cycle
can result, the judge is free to decide either way.
This idea was refined and presented in a more rigorous way by
Horty in [26] and further refined in [30]. Horty was interested in
modelling two different accounts of precedential constraint from
the jurisprudence literature. One is a very strict version, for which
Horty cites [5]. Here any distinction between the precedent and the
current case is enough to allow the judge to come to a different deci-
sion. This version, which corresponds to Figure 1, is now normally
termed the results model in AI and Law [37]. This model encodes
precedents as rules in the same way as [39] and [9]. Any weakening
Precedential Constraint: The Role of Issues
Figure 1: Plaintiff and Defendant lattices with two prece-
dents from [9]. Precedent 1 was [A,B,C,D,E] found for
plaintiff and Precedent 2 was [A,D,E] found for defendant.
of the plaintiff’s case or strengthening of the defendant’s case is
enough to prevent a precedent for the plaintiff from being followed
(and vice versa). Although this model provides an unchallengeable
constraint when the rule is applicable, the standard required for
the rule to be applicable is rather high. Even with only 3 factors for
each side we have 64 possible comparisons, and if we have the 13
factors for each side from CATO we have 226 (67,108,864) compar-
isons of sets of plaintiff factors with sets of defendant factors [1].
Of course, an example of every comparison is not needed, but still,
given the number of possible distinctions, expecting enough prece-
dents to provide significant guidance is unrealistic. This suggests
that the result model does not offer us enough constraints from
the precedents: rarely will the match be exact enough to constrain
the decision. Horty therefore used an alternative model, which he
termed the reason model and attributed to [31]. Horty noted that
although the full set of factors for the plaintiff represented the best
reason to decide for the plaintiff, a subset of these factors may well
have been sufficient to defeat the defendant reason without the
additional factors. Thus instead of representing the winner’s rule
by the full set of factors favouring the winner, the winner could
be represented by a subset of these factors, which can be seen as
the reason for the case. Thus in the first precedent of Figure 1, we
could see the reason for deciding the first case for the plaintiff
as 𝐵 → plaintiff 3 , so that future cases containing [𝐴, 𝐵, 𝐷, 𝐸] and
[𝐵, 𝐶, 𝐷, 𝐸] will also be constrained for the plaintiff, without being
distinguished by the absence of 𝐶 and 𝐴 respectively as was the
case with the results model shown in Figure 1. Horty formalised
this reason model in [26]. For a formal comparison of the result
and reason models of factor based precedential constraint see [37].
3.1 Criticisms of the Reason Model of
Precedential Constraint with Factors
Horty’s formalisation of the reason model provides an effective way
of characterising precedential constraint using factors. But there
have been several questions raised about it.
Rigoni advanced two criticisms in [41]. First, he pointed to cases
with multiple rationales:
3 Of
course, identifying the reason in practice is not a straightforward task: it requires
close analysis of the decision and there my be different interpretations of the decision.
The problem is avoided by the more conservative results model, at the expense of
constraining far fewer cases
16
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The first problematic sort of cases are those in which
the court makes a decision on the basis of multiple le-
gal rules, each of which would be sufficient for the de-
cision. These are sometimes known as cases with alter-
native holdings” [34] ... Consider Newport Yacht Basin
Ass’n ... the defendant prevailed and was awarded
attorney’s fees “based upon a prevailing party provi-
sion of a purchase and sale agreement, a contractual
indemnity provision, and principles of equitable in-
demnity” (NYBA 2012, p. 75). According to the trial
court judge, each of these was sufficient to justify the
awarding of attorney’s fees. ([41], p 141).
In such cases it is impossible to choose between the alternative
holdings to give the rule to represent the case in the theory. This,
however, is not an insurmountable problem: we can simply include
all of these rules in our theory, each with a preference over the rule
for the other side in the case.
Rigoni’s second observation is that not all precedents can be seen
as expressing a preference between reasons. Some cases instead
lay out a method for considering cases of a particular type. Rigoni
termed these framework precedents, and used Lemon v. Kurtzman
(1971) as his example:
In that case the US Supreme Court addressed the ques-
tion of whether Pennsylvania’s and Rhode Island’s
statutes that provided money to religious primary
schools subject to state oversight violated the Estab-
lishment Clause of the First Amendment. The court
introduced a three-pronged test and ultimately
ruled that both programs did violate the Establish-
ment Clause. ([41], p 142. Emphasis ours).
In Lemon it is not the balance of factors that led to the decision that
are instructive, but the tests themselves. These tests will be applied
in future cases, providing a set of issues that give a framework for
deciding future cases. Framework precedents did not arise in CATO
because the Restatements of Torts4 provides just such a framework
of issues, which formed the basis of the factor hierarchy in CATO,
rendering framework precedents unnecessary.
Another problem arises from the possibility of there being several
factors for the same side with different strengths on a dimension.
Consider the following example:
Example 1. Consider a case in which the plaintiff had a unique
product (F15p), but made disclosures to outsiders (F10d), which was
found for the defendant. This gives the three rules:
r1: F15p → plaintiff; r2: F10d → defendant
r3: r1 ≺ r2
Now consider a case where the plaintiff had made the disclosures
in a public forum, so that F27d applies rather than F10d. Now r2
does not match, and so the theory does not constrain the new case.
But since F27d is stronger on the dimension than F10d, the new
case should, a fortiori, be decided for the defendant on the basis of
the precedent.
This problem has a simple pragmatic fix: where a factor is in-
cluded in a case, include also all weaker factors on that dimension.
4 The relevant section, section 757, Liability for disclosure or use of another’s Trade
Secret, can be found at https://www.lrdc.pitt.edu/ashley/restatem.htm.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
So the new case becomes [F15p, F10d, F27d] and r2 will match. This
will work from a logical perspective, although care must be taken
to use the factor actually present in any explanation.
The fourth criticism is simply that even the reason model does
not constrain enough cases. This is because no account is taken
of whether the distinction is sufficient to overturn the rule. The
reason model as formalised above accepts any distinction, whereas
in CATO distinctions can be rejected through downplaying through
substitution and cancellation [40] of factors. The point is clear in
the following example.
Example 2. Here we have two cases, both found for the plaintiff.
In case 1 the plaintiff took security measures (F6p) and, although
the defendant claimed the information was reverse engineerable
(F16d), the plaintiff won5 . In the second case we have the plaintiff
making disclosures (F10d), but restricting these disclosures (F12p).
Here the preference for the plaintiff is clear. We therefore have the
following rules:
r1: F6p → plaintiff r4 F12p → plaintiff
r2: F16d → defendant r5: F10d → defendant
r3: r2 ≺ r1 r6: r5 ≺ r4
Now consider a case with all four factors: [F6p, F12p. F10d, F16d].
It seems this should clearly be found for the plaintiff: however, we
cannot apply r1 because of the distinction F10d and we cannot
apply r4 because F16d distinguishes. Although neither distinction
would be significant in CATO, being cancelled for both precedents
by the other factors available, the reason model gives the case as
unconstrained. We will address this problem in the next section.
4 THE IMPORTANCE OF ISSUES
The fourth problem discussed in the previous section arises because
the reason model considers cases as unstructured bundles of factors,
so that a difference which should not be considered significant
prevents us from applying the rule which should constrain the
case. We can use knowledge of the domain structure to solve this
problem. That we do not exploit the full power of our precedents if
we consider whole cases was noticed by Branting in [17]:
combining portions of multiple precedents can permit
new cases to be resolved that would be indeterminate
if new cases could only be compared to entire prece-
dents. ([17], Abstract).
When Brüninghaus and Ashley adapted CATO to predict cases in
IBP [21] they structured the cases around issues, as did Grabmair in
his prediction system VJAP [24]. Grabmair reports an improvement
over IBP through the use of values, but values raise several addi-
tional questions, such as the extent to which they are promoted by
different factors and whether value preferences are global, or local
to issues. Since precedential constraint with values has not yet been
given a formal characterisation, we will restrict our consideration in
this paper to factors. Issues are a well known concept in law: many
law schools teach the Issue-Rule-Application-Conclusion (IRAC)
method (or some variant) as a way of analysing legal cases. The
5 This preference for F6p was used in Mason v. Jack Daniel Distillery, 518 So.2d 130
(Ala.Civ.App.1987): “courts have protected information as a trade secret despite ev-
idence that such information could be easily duplicated by others competent in the
given field. KFC Corp. v. Marion-Kay Co., 620 F. Supp. 1160 (S.D.Ind. 1985); Sperry
Rand Corp. v. Rothlein, 241 F. Supp. 549 (D.Conn. 1964)”.
17
Bench-Capon and Atkinson
IRAC method was applied to the explanation of factor based rea-
soning in [11]. A key point about IRAC is that the rule (the reason
in the sense of the model above) relates to an issue, not to the case
as a whole. The issues in US Trade Secrets Law, taken from the
Restatement of Torts, were used to group related factors together
for IBP [21] and VJAP [24]. These issues were also the basis for the
factor hierarchy in [3] and the Abstract Dialectical Framework in
[2]. A similar structure is given by Rigoni’s framework precedents,
and we would argue that the role of framework precedents is to
identify the issues in a domain. Other systems, such as CABARET
[48] derive their framework of issues from statutes. Unlike factors,
issues can be seen as in a logical relation to the outcome. To find for
the plaintiff, it must be shown both that the information was a trade
secret and that it was misappropriated. A trade secret must both be
valuable and the information not generally known. To be misap-
propriated it must be either that improper means were used, or that
the information was used in breach of a confidential relationship.
We can express this as the following non-defeasible rules:
ROT1: TradeSecret ∧ and Misappropriated ↔ plaintiff
ROT2: InfoValuable ∧ SecrecyMaintained ↔ TradeSecret
ROT3: ImproperMeans ∨ (InfoUsed ∧ ConfidentialRelationship)
↔ Misappropriated
In [7] these issues are used to group CATO’s factors as shown
in Table 36 . Even though some factors appear under two issues,
the issues contain only five to seven factors, greatly reducing the
possible combinations of relevant factors. That using issues rather
than whole cases to constrain decisions will enable us to decide
more cases is evidenced by [21]. The issue based IBP was able
to reach a prediction in 99.5% of cases, as opposed to the 73.1%
achieved by a system considering cases as a whole.
This suggests that instead of describing cases simply as a set of
factors, we should distribute these factors across the issues they
relate to. Note also that structuring into issues is an implicit feature
of rule based systems such as [47] and [2]. We can now apply the
methods of precedential constraint developed in [27] and [37] not
at the case level, but at the issue level. To see the difference this
makes, we will consider a set of cases7 , taken from [3] and used in
[22], shown in Table 4. We have not re-analysed the decisions: the
factors for each case are taken from Table II of [22].
Notice that, in all these cases, some issues are uncontested. It
seems that we can regard the information as a trade secret, unless
argued otherwise, and that there is a presumption that the informa-
tion was used. On the other hand the plaintiff needs to establish that
improper means were used or that a confidential relationship ex-
isted. To find the issue and rule in the case we look at the contested
issues, and which factors led to the outcome. This is the method
used to identify the rule and resolve the issue when applying the
IRAC methodology in [11].
In the next sections we will illustrate the use of the standard
reason model followed by the use of the proposed issue based
6 In [24] Grabmair associated factors to issues a little differently. This does not, however
affect any of the factors in our example below, and so we follow [7] here.
7 National Instrument Labs, Inc. v. Hycel, Inc., 478 F.Supp. 1179 (D.Del.1979), M. Bryce
& Associates, Inc. v. Gladstone, 107 Wis.2d 241, 319 N.W.2d 907 (Wis.App.1982), K & G
Oil Tool & Service Co. v. G & G Fishing Tool Serv., 314 S.W.2d 782 (1958), Televation
Telecommunication Systems, Inc. v. Saindon, 522 N.E.2d 1359 (Ill.App. 2 Dist. 1988),
Mason v. Jack Daniel Distillery, 518 So.2d 130 (Ala.Civ.App.1987) and The Boeing
Company v. Sierracin Corporation, 108 Wash.2d 38, 738 P.2d 665 (1987).
Precedential Constraint: The Role of Issues ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 3: CATO factors grouped by Issues [7]

Issue Plaintiff Factors Defendant Factors

F16d Info Reverse Engineerable
F8p Competitive Advantage F20d Info Known to Competitors
InfoValuable
F15p Unique Product F24d Info Obtainable Elsewhere
F27d Disclosure In Public Forum
F4p Agreed Not To Disclose
F10d Secrets Disclosed Outsiders
SecrecyMaintained F6p Security Measures
F19d No Security Measures
F12p Outsider Disclosures Restricted
F2p Bribe Employee
F7p Brought Tools
F17d Info Independently Generated
ImproperMeans F14p Restricted Materials Used
F25d Info Reverse Engineered
F22p Invasive Techniques
F26p Deception
F7p Brought Tools
F8p Competitive Advantage F17d Info Independently Generated
InfoUsed
F14p Restricted Materials Used F25d Info Reverse Engineered
F18p Identical Products
F4p Agreed Not To Disclose
F1d Disclosure In Negotiations
ConfidentialRelationship F13p Noncompetition Agreement
F23d Waiver of Confidentiality
F21p Knew Info Confidential

Table 4: Cases in the Example with Factors Grouped By Issue

Case InfoValuable SecrecyMaintained ImproperMeans InfoUsed ConfidentialRelationship

National Instruments P F18p F1d F21p
Bryce P F6p F18p F1d F4p F21p
K and G P F15p F16d F6p F14p 25d F14p F18p F25d F21p
Televation P F15p F16d F6p F10d F12p F18p F21p
Mason P F15p F16d F6p F1d F21p
Boeing P F6p F10d F12p F14p F14p F1d F4p F21p

reason model. We will see that when using issues, more cases are
constrained because distinctions relating to issues unrelated to that
governed by a rule no longer distinguish that rule and are relevant
only if they constrain that other issue so as to lead to a different
outcome.
4.1 Using the reason model
Suppose our first case8 is National Instruments. As can be seen
from Table 4, the case turned on whether there was a confiden-
tial relationship, given that the plaintiff had made disclosures in
negotiations (F1d). The defendant, however, did know that the infor-
mation was confidential (F21p), and the court found for the plaintiff.
The reason model then gives the three rules:
NatInstP: F21p → plaintiff NatInstD: F1d → defendant
NatInstO: NatInstD ≺ NatInstP
If the next case is Bryce, we can see that it is constrained by
these rules: the additional factors are not distinctions because both
favour the plaintiff, and so do not give the defendant anything
better than NatInstD, which is defeated by NatInstP, which also
applies to Bryce. Bryce thus adds no new rules.
8 The sequencing of the cases used here is for the purposes of illustrating our approach,
and is not the actual sequence.
Now suppose we are presented with K and G. Here the plain-
tiff argues that improper means were used, because the defendant
used restricted materials (F14p). The defendant counters this by a
claim to have reverse engineered the information (F25d). Moreover
the defendant argues that the information is not a trade secret be-
cause it was reverse engineerable (F16d)9 . This is in turn countered
by the claim that the uniqueness of the product (F15p) suggests
that the information was not readily reverse engineerable. In the
judgement both the issues were decided in favour of the plaintiff,
since the reverse engineering had made use of restricted materials.
Note that there was no need to decide the breach of confidence
issue: improper means suffice to establish misappropriation. Al-
though NatInstP applies, it cannot be used because the defendant
has stronger rules than NatInstD. Thus the reason in K and G must
cover two different issues, InfoValuable and ImproperMeans, and
so we get rules spanning both these issues:
KGP: F15p and F14p → plaintiff
KGD: F16d and F25d → defendant
9 Both F25d and F16d were introduced in CATO and relate to the same dimension, D14b.
However, F25d, that the information was actually reverse engineered, relates to the
issues of whether the whether the information was used and whether improper means
were used, whereas F16d, the possibility of reverse engineering, relates to whether the
information was valuable and hence a trade secret.

18
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
KGO: KGD ≺ KGP
Note, however, that this means that the reasons extracted from
KGP conflate the two issues in dispute in the case. The next case is
Televation. The plaintiff had made disclosures (F10d), distinguishing
all the previous cases, although these were restricted (F12p). These
restrictions were held sufficient to find that the plaintiff had made
efforts to maintain secrecy. Again it was found that the uniqueness
of the product argued against it being readily reverse engineerable.
The misappropriation consisted of using the information, shown
by F18p, and the confidential relationship, shown by F21p. Because,
however, misappropriation was not contested, these do not form
part of the reason. The reason model rules from Televation also
cover two issues, InfoValuable and SecrecyMaintained:
TelevationP: F15p and F12p → plaintiff
TelevationD: F16d and F10d → defendant
TelevationO: TelevationD ≺ TelevationP
We now reach Mason. With respect to the existence of a trade
secret Mason is identical to K and G. However, Mason disclosed
the information in negotiations (F1d), which distinguishes it from
this case, and so KGP is not applicable. Similarly although it can
instantiate NatInstP from National Instruments, that case is distin-
guished by F16d, and so the defendant has a reason stronger than
NatInstD. So the reason for Mason must cover both trade secret and
confidential relationship, and when Mason is found for the plaintiff,
we get the rules:
MasonP: F15p and F21p → plaintiff
MasonD: F16d and F1d → defendant
MasonO: MasonD ≺ MasonP
Finally consider Boeing. We have two contested issues: whether
secrecy was maintained, with the same factors as Televation, and
whether there was a confidential relationship with the same factors
as Bryce. But we cannot use these to find for the plaintiff: we are not
constrained by Televation without F15p, and we are not constrained
by NatInstP as used in Bryce with F10d present.
In Bryce, Mason and Boeing, the decision seems clear given the
preceding cases, but only Bryce is constrained under the reason
model. Now we will consider the difference made by exploiting our
understanding of the domain and structuring the cases into issues.
4.2 Using the Issue Based Reason Model
In this section we look at how the sequence of cases in 4.1 works
out if we associate our reasons with issues, and rely on the frame-
work provided by the Restatement of Torts (rules ROT1-3 above)
to combine these partial findings. National Instruments concerned
only one issue, Confidential Relationship, so our rules will be:
NatInstCRP: F21p → ConfidentialRelationship
NatInstCRD: F1d → Not ConfidentialRelationship
NatInstCR: NatInstCRD ≺ NatInstCRP
In Bryce, we have the same issue, and we can use these rules: the
additional factor also supports a confidential relationship. K and
G, however, concerns two different issues, and so gives us distinct
sets of rules for each of these two issues.
KGIVP: F15p → InformationValuable
KGIVD: F16d → Not InformationValuable
KGOIV: KGIVP ≺ KGIVD
19
Bench-Capon and Atkinson
KGIMP: F15p → ImproperMeans
KGIMD: F25d → Not ImproperMeans
KGOIM: KGIMD ≺ KGIMP
Now when we come to Televation, K and G constrains Informa-
tionValuable, but we get rules for the new issue, SecrecyMaintained.
TelevationSMP: F12p → SecrecyMaintained
TelevationSMD: F10d → Not SecrecyMaintained
TelevationSM: TelevationSMD ≺ TelevationSMP
By using issues, when we come to Mason, the decision is con-
strained. We can use the rules from K and G to constrain Infor-
mationValuable and those from National Instruments to constrain
ConfidentialRelationship. Under the result model above, F1d dis-
tinguished K and G, and F16d distinguished National Instruments.
Similarly Boeing is constrained: by Televation with respect to Se-
crecyMaintained and by National Instruments (like Bryce the F4p
factor is not needed) with respect to ConfidentialRelationship. F15p
would have distinguished Televation and F10d National Instruments
under the results model. Thus by focussing on issues, we are able
to constrain cases which required new rules under the “whole case”
reason model. We also argue that this better fits with legal prac-
tice: not only does it follow the IRAC methodology taught in law
schools, but it also reflects how precedents are used in decisions.
Consider this extract from Mason, finding that the information
could be regarded as a trade secret10 :
We note that absolute secrecy is not required ... “a
substantial element of secrecy is all that is necessary
to provide trade secret protection.” Drill Parts, 439
So.2d at 49. ... we note that courts have protected
information as a trade secret despite evidence that
such information could be easily duplicated by others
competent in the given field. KFC Corp. v. Marion-
Kay Co., 620 F. Supp. 1160 (S.D.Ind. 1985); Sperry Rand
Corp. v. Rothlein, 241 F. Supp. 549 (D.Conn. 1964).
Three precedents are used to justify the finding in Mason: for none
of the three precedents was there any consideration of aspects of
these cases not germane to the specific issue being resolved. Thus
the focus provided by issues means that we are able to constrain
more cases, since we can ignore differences which do not relate to
the issue on which the case was decided. Issues would be even more
useful in the results model, since they would allow us to ignore
irrelevant pro-plaintiff as well as irrelevant pro-defendant factors.
We now look at how we can handle dimensional features of cases.
5 RELATING TO DIMENSIONS
In Bryce the Restatement of Torts was explicitly cited as the frame-
work for considering the case:
Some factors to be considered in determining whether
given information is one’s trade secret are: (1) the
extent to which the information is known outside of
his business; (2) the extent to which it is known by
employees and others involved in his business; (3) the
10 Actuallythis suggests that the court used F6p rather than F15p to defeat F16d. This
suggests that perhaps F6p should be added to the factors related to InfoValuable and
used as the reason in KGIVP. Alternatively, TradeSecret could be considered as a single
issue subsuming InfoValuable and MaintainSecrecy. This, however, is a matter of legal
analysis, and for the purposes of this paper we adopt the analysis of [7].
Precedential Constraint: The Role of Issues
extent of measures taken by him to guard the secrecy
of the information; (4) the value of the information
to him and to his competitors; (5) the amount of
effort or money expended by him in developing the
information; (6) the ease or difficulty with which the
information could be properly acquired or duplicated
by others. Emphasis ours.
These points are all reflected in CATO’s factor hierarchy. But as
the emphasised terms indicate, ascribing these factors is not simple,
but requires a judgement as to whether the extent is sufficient for
the factor to apply. This point was addressed by Horty in [27] and
[28]. The issue was further discussed by Rigoni in [42]. Horty has
modified his approach in [29], and a formal comparison of Horty
and Rigoni’s approaches is given in [37].
Horty’s main example in [27] is taken from [39] and concerns
change of fiscal domicile, decided on the basis of several consid-
erations including length of absence and percentage of income
earned abroad. In [39] absence was modelled as two factors, which
we will call shortStay and longStay 11 , favouring no change and
change respectively. But this raises the question of how we deter-
mine whether a particular length of absence, say 24 months, is a
shortStay, a longStay, or somewhere in between, and so neutral.
Horty responds by introducing the notion of a factor with mag-
nitude, (i.e. a factor deriving from a dimension with more than
two values) based on the dimensional fact of length of stay. The
ascription of factors on the basis of dimensional facts can also be
found in [40]. Ascription of factors is constrained by precedents:
in a previous case a judge may have found for change on the basis
of an absence of 18 months, showing on the result model that any
absence of at least 18 months must be considered a longStay. But
the judge may have spoken of an absence of greater than one year,
so that on the reason model that any absence over 12 months is to
be considered longStay. Rigoni’s suggestion was to see precedents
as fixing “switching points”, which determine which (if any) factor
applies for various values of the dimensional fact. Rigoni also notes
that a dimension may encompass multiple factors for a given side
(as with disclosures (D3d) in CATO).
Note here that the precedents which impose bounds on the
ranges occupied by factors are a different kind of precedent from
those which resolve factor conflicts as discussed above: they express
no preferences. Thus what is required to accommodate dimensional
facts and factors with magnitude is not a different way of represent-
ing precedential constraint, but to recognise that we are looking
at a two stage process, with each stage using different types of
precedents. The need for two stages was observed in [40]:
Once the facts of a case have been established - and
this is rarely straightforward since the move from
evidence to facts is often itself the subject of debate -
legal reasoning can be seen, following Ross [45] and
Lindhal and Odelstad [33], as a two stage process, first
from the established facts to intermediate predicates,
and then from these intermediate predicates to legal
11 In[39] long duration and not long duration were used, but for reasons explained in
section 2.1, negating factors is problematic and we follow CATO and [44] and use
two distinct factors when the dimension can favour both sides. This also permits the
possibility of a moderate duration being neutral.
20
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
consequences. CATO has been explicitly identified
with the second of these steps (e.g. [20]). ([40], p 22).
This can be seen clearly in [7] where factors – the intermedi-
ate predicates – were ascribed to cases by the machine learning
program SMILE, before being passed to IBP to predict the legal con-
sequences. More recently this two-stage approach has been used
by Branting in [19] and [18]. Thus before we can consider whether
a case is constrained, which can be done in terms of factors using
the issue based reason model described above, we must first as-
sign the factors. For some factors, those derived from many-valued
dimensions, this will involve ascribing the factors on that dimen-
sion respecting ranges identified in precedent cases. This can be
done using either the reason or the result model, or using Rigoni’s
switching points. For a discussion of mapping a dimensional fact
(age) into ranges through precedents see [25].
Thus the conclusion is that attempting to model precedential
constraint in terms of cases represented as sets of dimensions rather
than sets of factors as in [37] conflates two distinct steps in the
process of reasoning with legal cases. Cases are not represented
as sets of dimensions: cases are represented as facts in HYPO and
where they are represented as sets of points on dimensions, as in
[40] and [14], these are dimensional facts, the legal significance
of which is unknown until they are mapped into factors. If we
represent cases as sets of dimensional facts (including dimensions
with two values) as in [40], we can derive the factors applicable
to the case. Or we can get our factors through machine learning
as in [19]. We then organise these factors into issues and apply
precedential constraint in terms of the factors associated with each
issue as described in section 4.2. While some precedents will supply
the plaintiff, defendant and priority rules as described in section 3.1,
others will supply rules to move from dimensional facts to factors.
To give an example from the fiscal domicile domain, such a rule
would be something like: longStay ← absence(A) ∧ A ≥ 12.
One issue in the ascription of factors is that in some cases they
do not seem to be independent. Thus in the fiscal domicile case it
is possible that there is a trade off between length of absence and
amount of income earned, so that whether the percentage of income
is considered to be “substantial” is relative to the length of absence.
The question of balancing factors has been discussed in [32] and
[23], and an equation representing the trade off was used in [12].
In that paper a single factor (e.g. SufficientIncomeGivenAbsence) is
ascribed on the basis of the two dimensional facts. How factors are
ascribed on the basis of facts relates to the first stage and the focus
of this paper is on the second stage, namely determining how the
precedents, when described in terms of factors, constrain the deci-
sion. Therefore we will not discuss the important and interesting
questions relating to balancing and trade-offs further in this paper.
6 CONCLUDING REMARKS
A number of conclusions can be drawn from the above discussion:
• Reasoning with cases is a two stage process: first factors are
ascribed on the basis of (often dimensional) facts, and then
the cases are compared with precedents using factors.
• Precendential constraint should be considered in terms of
factors, even if we wish to represent cases in terms of dimen-
sional facts. Applicable dimensions show which aspects must
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
be considered, while factors show which side is favoured in
the particular case.
• Comparison (for both the results and the reason models)
should be at the level of issues, to ignore irrelevant distinc-
tions, and to reflect legal practice better.
• Precedents do not always have the same role:
– Framework precedents (e.g. Lemon v. Kurtzman) identify
the issues and set out the logical framework in which they
are considered;
– Preference precedents (the standard use in CATO) say how
conflicting factors within an issue should be resolved;
– Ascription precedents (e.g. National Instruments states its
reasons for withholding F16d at some length), give reasons
to determine if a factor should be ascribed to a case or not.
Here we have used the issues from IBP [7]. But we could have
used coarser grained issues, perhaps merging the conjoined issues
in IBP, or finer grained issues, using the abstract factors of [3] as
issues, or even the nodes of the 2-regular hierarchy of [1]. The finer
the granularity, the more decisions are constrained. Experiments
to investigate the impact of different granularities on predictive ac-
curacy would be interesting. It would also be interesting to explore
the use of values rather than factors as the elements over which
preferences are expressed as in [16] and [24]. The possibility of
using multiple granularities could also be explored, with some ar-
guments being in terms of issues, some in terms of abstract factors,
others in terms of values, and others considering whole cases.
Perhaps the best way to deploy machine learning is for the first
stage, factor ascription, as in [7] and [19]. Moreover, if we wish to
address the second stage with machine learning, perhaps it would be
better to predict issues rather than whole cases, and then combine
the results using a logical framework to get the overall decision.
REFERENCES
[1] Latifa Al-Abdulkarim, Katie Atkinson, and Trevor Bench-Capon. 2015. Factors,
issues and values: Revisiting reasoning with cases. In Proceedings of the 15th
International Conference on AI and Law. 3–12.
[2] Latifa Al-Abdulkarim, Katie Atkinson, and Trevor Bench-Capon. 2016. A method-
ology for designing systems to reason with legal cases using ADFs. AI and Law
24, 1 (2016), 1–49.
[3] Vincent Aleven. 1997. Teaching case-based argumentation through a model and
examples. Ph.D. thesis. University of Pittsburgh.
[4] Vincent Aleven and Kevin D Ashley. 1995. Doing things with factors. In Proceed-
ings of the 5th International Conference on AI and Law. 31–41.
[5] Larry Alexander. 1989. Constrained by precedent. Southern California Law Review
63 (1989), 1–64.
[6] Kevin D Ashley. 1990. Modeling legal arguments: Reasoning with cases and hypo-
theticals. MIT press, Cambridge, Mass.
[7] Kevin D Ashley and Stefanie Brüninghaus. 2009. Automatically classifying case
texts and predicting outcomes. AI and Law 17, 2 (2009), 125–165.
[8] Trevor Bench-Capon. 1991. Practical legal expert systems: the relation between
a formalisation of legislation and expert knowledge. In Law, Computer Science
and Artificial Intelligence, M Bennun and A Narayanan (Eds.). Ablex, 191–201.
[9] Trevor Bench-Capon. 1999. Some observations on modelling case based reasoning
with formal argument models. In Proceedings of the 7th International Conference
on AI and Law. 36–42.
[10] Trevor Bench-Capon. 2017. HYPO’s legacy: introduction to the virtual special
issue. AI and Law 25, 2 (2017), 205–250.
[11] Trevor Bench-Capon. 2020. Explaining Legal Decisions Using IRAC. In Proceed-
ings of CMNA 2020. CEUR Workshop Proceedings 2669, 74–83.
[12] Trevor Bench-Capon and Katie Atkinson. 2017. Dimensions and Values for Legal
CBR. In Proceeding of JURIX 2017. 27–32.
[13] Trevor Bench-Capon and Katie Atkinson. 2018. Lessons from Implementing
Factors with Magnitude. In Proceedings of JURIX 2018. 11–20.
[14] Trevor Bench-Capon and Floris Bex. 2015. Cases and Stories, Dimensions and
Scripts.. In Proceedings of JURIX 2015. 11–20.
21
Bench-Capon and Atkinson
[15] Trevor Bench-Capon and Edwina L Rissland. 2001. Back to the future: Dimensions
revisited. In Proceedings of JURIX 2001. IOS Press, 41–52.
[16] Trevor Bench-Capon and Giovanni Sartor. 2003. A model of legal reasoning with
cases incorporating theories and values. Artificial Intelligence 150, 1-2 (2003),
97–143.
[17] L Karl Branting. 1991. Reasoning with portions of precedents. In Proceedings of
the 3rd International Conference on AI and Law. 145–154.
[18] L Karl Branting. 2020. Explanation in Hybrid, Two-Stage Models of Legal Predic-
tion. In The 3rd XAILA Workshop at JURIX 2020.
[19] L Karl Branting, Craig Pfeifer, Bradford Brown, Lisa Ferro, John Aberdeen, Brandy
Weiss, Mark Pfaff, and Bill Liao. 2020. Scalable and explainable legal prediction.
AI and Law (2020), 1–26.
[20] Stefanie Brüninghaus and Kevin Ashley. 2003. A predictive role for intermediate
legal concepts. In Proceedings of Jurix 2003. 153–62.
[21] Stefanie Brüninghaus and Kevin D Ashley. 2003. Predicting outcomes of case
based legal arguments. In Proceedings of the 9th International Conference on AI
and Law. 233–242.
[22] Alison Chorley and Trevor Bench-Capon. 2005. An empirical investigation of
reasoning with legal cases through theory construction and application. AI and
Law 13, 3 (2005), 323–371.
[23] Thomas F Gordon and Douglas Walton. 2016. Formalizing Balancing Arguments..
In Proceedings of COMMA 2016. 327–338.
[24] Matthias Grabmair. 2017. Predicting trade secret case outcomes using argument
schemes and learned quantitative value effect tradeoffs. In Proceedings of the 16th
International Conference on AI and Law. 89–98.
[25] John Henderson and Trevor Bench-Capon. 2019. Describing the development
of case law. In Proceedings of the 17th International Conference on AI and Law.
32–41.
[26] John F Horty. 2011. Reasons and precedent. In Proceedings of the 13th International
Conference on AI and Law. 41–50.
[27] John F Horty. 2017. Reasoning with dimensions and magnitudes. In Proceedings
of the 16th the International Conference on Articial Intelligence and Law. 109–118.
[28] John F Horty. 2019. Reasoning with dimensions and magnitudes. AI and Law 27,
3 (2019), 309–345.
[29] John F Horty. 2021. Modifying the Reason Model. AI and Law (2021), On Line.
[30] John F Horty and Trevor Bench-Capon. 2012. A factor-based definition of prece-
dential constraint. AI and Law 20, 2 (2012), 181–214.
[31] Grant Lamond. 2005. Do precedents create rules. Legal Theory 11 (2005), 1–26.
[32] Marc Lauritsen. 2015. On balance. AI and Law 23, 1 (2015), 23–42.
[33] Lars Lindahl and Jan Odelstad. 2006. Open and closed intermediaries in normative
systems. In Proceedings of JURIX 2006. IOS Press, 91–99.
[34] Jo Desha Lucas. 1983. The direct and collateral estoppel effects of alternative
holdings. The University of Chicago Law Review 50, 2 (1983), 701–730.
[35] Masha Medvedeva, Michel Vols, and Martijn Wieling. 2019. Using machine
learning to predict decisions of the European Court of Human Rights. AI and
Law (2019), 1–30.
[36] Henry Prakken. 2019. Modelling accrual of arguments in ASPIC+. In Proceedings
of the 17th International Conference on AI and Law. 103–112.
[37] Henry Prakken. 2021. A A formal analysis of some factor- and precedent-based
accounts of precedential constraint. AI and Law (2021), Available On–Line.
[38] Henry Prakken and Ratsma Rosa. 2021. A top-level model of case-based ar-
gumentation for explanation: formalisation and experiments. Argument and
Computation (2021), Available On–line.
[39] Henry Prakken and Giovanni Sartor. 1998. Modelling reasoning with precedents
in a formal dialogue game. AI and Law 6, 3-4 (1998), 231–87.
[40] Henry Prakken, Adam Wyner, Trevor Bench-Capon, and Katie Atkinson. 2015.
A formalization of argumentation schemes for legal case-based reasoning in
ASPIC+. Journal of Logic and Computation 25, 5 (2015), 1141–1166.
[41] Adam Rigoni. 2015. An improved factor based approach to precedential constraint.
AI and Law 23, 2 (2015), 133–160.
[42] Adam Rigoni. 2018. Representing dimensions within the reason model of prece-
dent. AI and Law 26, 1 (2018), 1–22.
[43] Edwina L Rissland and Kevin D Ashley. 1987. A case-based system for Trade
Secrets law. In Proceedings of the 1st International Conference on AI and Law.
60–66.
[44] Edwina L Rissland and Kevin D Ashley. 2002. A note on dimensions and factors.
AI and Law 10, 1-3 (2002), 65–77.
[45] Alf Ross. 1957. Tû-tû. Harvard Law Review (1957), 812–825.
[46] Frederick Schauer. 1987. Precedent. Stanford Law Review (1987), 571–605.
[47] Marek Sergot, Fariba Sadri, Robert Kowalski, Frank Kriwaczek, Peter Hammond,
and Therese H Cory. 1986. The British Nationality Act as a logic program.
Commun. ACM 29, 5 (1986), 370–386.
[48] David B Skalak and Edwina L Rissland. 1992. Arguments and cases: An inevitable
intertwining. AI and Law 1, 1 (1992), 3–44.
[49] Richard E Susskind. 1989. The Latent Damage system: A jurisprudential analysis.
In Proceedings of the 2nd International Conference on AI and Law. 23–32.
[50] Heng Zheng, Davide Grossi, and Bart Verheij. 2020. Case-Based Reasoning with
Precedent Models: Preliminary Report. In Proceedings of COMMA 2020. 443–450.
Incorporating Domain Knowledge for Extractive Summarization
of Legal Case Documents
Paheli Bhattacharya Soham Poddar
Department of CSE, IIT Kharagpur Department of CSE, IIT Kharagpur
India India
Kripabandhu Ghosh
Department of CDS, IISER, Kolkata
India
ABSTRACT
Automatic summarization of legal case documents is an important
and practical challenge. Apart from many domain-independent text
summarization algorithms that can be used for this purpose, sev-
eral algorithms have been developed specifically for summarizing
legal case documents. However, most of the existing algorithms
do not systematically incorporate domain knowledge that spec-
ifies what information should ideally be present in a legal case
document summary. To address this gap, we propose an unsuper-
vised summarization algorithm DELSumm which is designed to
systematically incorporate guidelines from legal experts into an
optimization setup. We conduct detailed experiments over case
documents from the Indian Supreme Court. The experiments show
that our proposed unsupervised method outperforms several strong
baselines in terms of ROUGE scores, including both general sum-
marization algorithms and legal-specific ones. In fact, though our
proposed algorithm is unsupervised, it outperforms several super-
vised summarization models that are trained over thousands of
document-summary pairs.
KEYWORDS
Legal document summarization; Integer Linear Programming
ACM Reference Format:
Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh,
and Saptarshi Ghosh. 2021. Incorporating Domain Knowledge for Extrac-
tive Summarization of Legal Case Documents. In Eighteenth International
Conference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021,
São Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https://doi.org/10.
1145/3462757.3466092
1 INTRODUCTION
In a Common Law system (followed in most countries including
USA, UK, India, Australia), law practitioners have to go through
hundreds of legal case documents to understand how the Court
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466092
22
Koustav Rudra
L3S Research Center, Leibniz
University, Hannover
Germany
Saptarshi Ghosh
Department of CSE, IIT Kharagpur
India
behaved in different legal situations. Case documents usually span
several tens to hundreds of pages and contain complex text, which
makes comprehending them difficult even for legal experts. Hence,
summaries of case documents often prove beneficial. Existing le-
gal IR systems employ legal attorneys or para-legals to write case
summaries, which is an expensive process. Hence, automatic sum-
marization of case documents is a practical and well-known chal-
lenge [8, 14, 18, 20, 25].
Document summarization is broadly of two types – (i) extractive,
where important sentences are extracted from the source document
and included in the summary [7, 14, 16], and (ii) abstractive, where
the model attempts to generate a summary using suitable vocabu-
lary [15, 17]. This paper focuses on extractive summarization since
it is more common for legal case documents.
Law practitioners in various countries have a set of guidelines
for how to summarize court cases, e.g., see [9, 11, 12] for guidelines
on summarizing case documents of the USA and the UK. All these
guidelines basically advise to consider different rhetorical/thematic
segments that are usually present in case documents (e.g., facts of the
case, legal issues being discussed, final judgment, etc.) and then to
include certain important parts of each segment into the summary.
We got similar advice from law experts in India – two senior law
students and a faculty from the Rajiv Gandhi School of Intellectual
Property Law, India – regarding Indian court case documents (see
details in Section 5.1). Apart from focusing on summarizing the
case document as a whole, it is also necessary that each of these
segments is summarized well and has representations in the final
summary. Another advantage of including representations of each
segment in the summary is that, often law practitioners wish to
read only the summary of particular segments (e.g., only the facts
of a case, or the legal issues involved in a case).
There are several summarization algorithms specifically meant
for legal court case documents [8, 14, 18, 20, 25] (details in Section 2).
However, most of these algorithms do not take into account the
guidelines from law practitioners. Though a few legal domain-
specific algorithms recognize the presence of various rhetorical
segments in case documents [8, 20], they also do not consider expert
guidelines in deciding what to include in the summary from which
segment. As a result, these algorithms cannot represent some of
the important segments well in the summary. For instance, a recent
work [2] showed that most existing algorithms do not represent
the final judgement well in the summary, though this segment is
considered very important by law practitioners.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The above limitations of existing methods motivate us to develop
a summarization algorithm that will consider various rhetorical seg-
ments in a case document, and then decide which parts from each
segment to include in the summary, based on guidelines from law
practitioners. Supervised methods which are expected to learn these
guidelines automatically would require a large number of expert
written summaries for lengthy case judgements, which may not
be available for all jurisdictions. Hence we go for an unsupervised
summarization algorithm in this work.
We propose DELSumm (Domain-adaptive Extractive Legal
Summarizer ), an unsupervised extractive summarization algo-
rithm for legal case documents. We formulate the task of summariz-
ing legal case documents as maximizing an Integer Linear Program-
ming (ILP) objective function that aims to maximize the inclusion of
the most informative sentences in the summary, and has balanced
representations from all the thematic segments, while reducing
redundancy. We demonstrate how DELSumm can be tuned to oper-
ationalize summarization guidelines obtained from law experts over
Indian Supreme Court case documents. Comparison with several
baseline methods – including five legal domain-specific methods
and two state-of-the-art deep learning models trained over thou-
sands of document-summary pairs – suggest that our proposed
approach outperforms most existing methods, especially in summa-
rizing the different rhetorical segments present in a case document.
To summarize, the contributions of this paper are:
(1) We propose DELSumm that incorporates domain knowledge pro-
vided by legal experts for summarizing lengthy case documents.1
(2) We perform extensive experiments on documents from the
Supreme Court of India and compare our proposed method with
eleven baseline approaches. We find that DELSumm outperforms
a large number of legal-specific as well as general summarisation
methods, including supervised neural models trained over thou-
sands of document-summary pairs. Especially, DELSumm achieves
much better summarization of the individual rhetorical segments in
a case document, compared to most prior methods.
(3) We also show that our proposed approach is robust to inaccurate
rhetorical labels generated algorithmically. There is a negligible
drop in performance when DELSumm uses rhetorical sentence
labels generated by a rhetorical segmentation method (which is
more a practical setup), instead of using expert-annotated labels.
To the best of our knowledge, this is the first systematic attempt
to computationally model and incorporate legal domain knowledge
for summarization of legal case documents. Through comparison
with 11 methods, including 2 state-of-the-art deep learning methods,
we show that an unsupervised algorithm developed by intelligently
including domain expertise, can surpass the performance of su-
pervised learning models even when the latter are trained over
large training data (which is anyway expensive to obtain in an
expert-driven domain such as Law).
2 RELATED WORK
Extractive text summarization aims to detect important sentences
from the full document and include them in the summary. Existing
methods for extractive summarization that can be applied for legal
document summarization, can be broadly classified into four classes:
1 Implementation publicly available at https://github.com/Law-AI/DELSumm
Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh, and Saptarshi Ghosh
(i) Unsupervised domain-independent, (ii) Unsupervised domain-
specific, (iii) Supervised domain-independent, and (iv) Supervised
domain-specific. We briefly describe some popular methods from
each of the classes in this section.
2.1 Unsupervised domain-independent
methods
Popular, unsupervised extractive summarization algorithms iden-
tify important sentences either by using Frequency-based methods
(e.g. Luhn [16]) or Graph-based methods (e.g. LexRank [7]). The
summary is the top ranked sentences. There are algorithms from
the family of matrix-factorization, such as LSA [10].
Summarization has also been treated as an Integer Linear Pro-
gramming (ILP) optimization problem. Such methods have been
applied for summarizing news documents [1] and social media
posts [19]. Our proposed approach in this work is also based on
ILP-based optimization.
2.2 Unsupervised domain-specific methods
There are several unsupervised methods for extractive summariza-
tion specifically designed for legal case documents. One of the ear-
liest methods for unsupervised legal document summarization that
takes into account the rhetorical structure of a case document was
LetSum [8]. They consider certain cue-phrases to assign rhetor-
ical/semantic roles to a sentence in the source document. They
specifically consider the roles – Introduction, Context, Juridical
Analysis and Conclusion. Sentences are ranked based on their TF-
IDF values for estimating their importance. The final summary is
generated by taking 10% from the Introduction, 25% from Context,
60% from Juridical Analysis and 5% from the Conclusion segments.
While LetSum considers TF-IDF to rank sentences, Saravanan
et.al. [20] use a K-Mixture Model to rank sentences for deciding
which sentences to include in the summary. We refer to this work
as KMM in the rest of the paper. Note that this work also identifies
rhetorical roles of sentences (using a graphical model). However,
this is a post-summarization step that is mainly used for displaying
the summary in a structured way; the rhetorical roles are not used
for generating the summary.
Another method CaseSummarizer [18] finds the importance of
a sentence based on several factors including its TF-IDF value, the
number of dates present in the sentence, number of named entities
and whether the sentence is at the start of a section. Sentences are
then ranked in order to generate the summary comprising of the
top-ranked sentences.
Zhong et.al. [25] creates a template based summary for Board
of Veteran Appeals (BVA) decisions from the U.S. Department of
Veteran Affairs. The summary contains (i) one sentence from the
procedural history (ii) one sentence from issue (iii) one sentence
from the service history of the veteran (iv) variable number of
Reasoning & Evidential Support sentences selected using Maximum
Margin Relevance (v) one sentence from the conclusion. We refer
to this method as MMR in this paper.
Limitations of the methods: CaseSummarizer does not assume
the presence of any rhetorical role. The other methods – LetSum,
KMM, and MMR – consider their presence but do not include
23
Incorporating Domain Knowledge for Extractive Summarization of Legal Case Documents
them while estimating the importance of sentences for generat-
ing the summary. Specifically, LetSum and KMM use generic term-
distributional models (TF-IDF and K-mixture models respectively)
and MMR uses Maximum Margin Relevance. As evident, rhetorical
roles do not play any role in finally selecting the sentences for
inclusion in the summary.
We believe that it is more plausible to measure the importance
of sentences in each rhetorical segment separately using domain
knowledge. In fact, for each segment, the parameters measuring
sentence importance can vary. As an example, consider the segment
‘Statute’. One can say that sentences from this segment that actually
contain a reference to a Statute/Act name are important. In contrast,
this reasoning will not hold for the segment ‘Facts’ and one may
derive a different measure based on the presence of certain Part-of-
Speech tags e.g., Nouns (names of people, location, etc.). Hence, in
the present work, we especially include such domain knowledge
along with segment representations in a more systematic way for
generating summaries.
2.3 Supervised domain-independent methods
Supervised neural (Deep Learning-based) methods for extractive
text summarization treat the task as a binary classification prob-
lem, where sentence representations are learnt using a hierarchical
encoder. A survey of the existing approaches can be found in [6].
Two popular extractive methods are NeuralSum [4] and Sum-
maRunner [17]. These methods use RNN (Recurrent Neural Net-
work) encoders to learn the sentence representations from scratch.
Sentence selection into summary is based on – content, salience,
novelty and absolute and relative position importance. These pa-
rameters are also learned in the end-to-end models.
Recently, pretrained encoders (especially transformer-based mod-
els such as BERT [5]) have gained much popularity. These encoders
have already been pretrained using a large amount of open do-
main data. Given a sentence, they can directly output its sentence
representation. These models can be used in a supervised summa-
rization by fine-tuning the last few layers on domain-specific data.
BERTSUM [15] is a BERT-based extractive summarization method.
Unlike SummaRuNNer, here sentence selection into summary is
based on trigram overlap with the currently generated summary.
Limitations of the neural methods: These Deep Learning archi-
tectures have been evaluated mostly in the news summarization
domain, and news documents are much shorter and contain simpler
language than legal documents. Recent works [22, 23] show that
these methods do not perform well in summarizing scientific arti-
cles from PubMed and arxiv, which are longer sequences. However,
it has not been explored how well these neural models would work
in summarizing legal case documents; we explore this question for
two popular neural summarization models in this paper.
2.4 Supervised domain-specific methods
Liu et.al. [14] have recently developed a supervised learning method
for extractive summarization of legal documents. Sentences are
represented using handcrafted features like number of words in
a sentence, position of the sentence in a document etc. Machine
Learning (ML) classifiers (e.g., Decision Tree, MLP, LSTM) are then
24
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
applied for obtaining the final summary. We refer to this method
as Gist [14].
Limitations of the method: Although this method was developed
and applied for Chinese legal case documents, domain-specific
attributes such as rhetorical labels were not considered.
2.5 Rhetorical roles in a legal case document
A legal case document can be structured into thematic segments,
where each sentence can be labelled with a rhetorical role. Note that
case documents often do not implicitly specify these rhetorical roles;
there exist algorithms that assign rhetorical roles to the sentences [3,
21, 24]. Different prior works have considered different sets of
rhetorical roles [3, 8, 20, 25]. It was shown in our prior work [2]
that a mapping between the different sets of rhetorical roles is
possible.
In this work, we consider a set of eight rhetorical roles suggested
in our prior work [3]. Briefly, the rhetorical roles are as follows –
(i) Facts: the events that led to filing the case, (ii) Issue: legal ques-
tions/points being discussed, (iii) Ruling by Lower Court: case
documents from higher courts (e.g., Supreme Court) can contain de-
cisions delivered at the lower courts (e.g. Tribunal), (iv) Precedent:
citations to relevant prior cases, (v) Statute: citations to statutory
laws that are applicable to the case (e.g., Dowry Prohibition Act,
Indian Penal Code), (vi) Arguments delivered by the contending
parties, (vii) Ratio: rationale based on which the judgment is given,
and (viii) Final judgement of the present court.
In this work, we attempt to generate a summary that includes
representations of all rhetorical segments in the source document
(full text). We assume that the rhetorical labels of every sentence
in the source document is already labeled either by legal experts or
by applying the method proposed in [3].
3 DATASET
For evaluation of summarization algorithms, we need a set of source
documents (full text) and their gold standard summaries. Addition-
ally, since we propose to use the rhetorical labels of sentences in
a source document for generating the summary, we need rhetori-
cal label annotations of the source documents. Additionally, since
the supervised methods Gist, SumaRuNNer and BERTSUM require
training over large number of document-summary pairs, we also
need such a training set. In this section, we describe the training
set and the evaluation set (that is actually used for performance
evaluation of summarization algorithms).
Evaluation set: Our evaluation dataset consists of a set of 50 In-
dian Supreme Court case documents, where each sentence is tagged
with a rhetorical/semantic label by law experts (out of the rhetori-
cal roles described in Section 2.5). This dataset is made available
by our prior work [3]. We asked two senior law students (from
the Rajiv Gandhi School of Intellectual Property Law, one of the
most reputed law schools in India) to write summaries for each of
these 50 document. They preferred to write extractive summaries of
approximately one-third of the length of the documents. We asked
the experts to summarize each rhetorical segment separately (so
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh, and Saptarshi Ghosh

that we can evaluate how well various models summarize the indi- Notation Meaning
vidual sections); only, they preferred to summarize the rhetorical 𝐿 Desired summary length (number of words)
segments ‘Ratio’ and ‘Precedent’ together. 𝑛 Number of sentences in the document
All the summarization methods have been used to generate 𝑔 Number of segments in the document
summaries for these 50 documents. These summaries were then 𝑚 Number of distinct content words in the document
uniformly evaluated against the two gold standard summaries writ- 𝑖 Index for sentence (𝑖 = [1 . . . 𝑛])
ten by the law students, using the standard ROUGE scores (details 𝑗 Index for content word (𝑗 = [1 . . . 𝑚])
given in later sections). We report the average ROUGE scores over 𝑘 Index for segment (𝑘 = [1 . . . 𝑔])
the two sets of gold standard summaries. 𝑥𝑖 Indicator variable for sentence 𝑖 (1 if sentence 𝑖 is to
be included in summary, 0 otherwise)
Training set : For training supervised methods (Gist, SumaRuNNer
𝑦𝑗 Indicator variable for content word 𝑗 (1 if 𝑗 is to be
and BERTSUM), we need an additional training dataset consisting
included in summary, 0 otherwise)
of document-summary pairs. To this end, we crawled 7, 100 Indian
𝑝𝑖 Indicator variable for sentence 𝑖 citing a prior-case
Supreme Court case documents and their headnotes (short abstrac- (1 if there is a citation, 0 otherwise)
tive summaries) from http://www.liiofindia.org/in/cases/cen/INSC/
𝑎𝑖 Indicator variable for sentence 𝑖 citing a statute (1 if
which is an archive of Indian court cases. These document-summary there is a citation, 0 otherwise)
pairs were used to train the supervised summarization models (de- 𝐿 (𝑖) Number of words in sentence 𝑖
tails in later sections). We ensured that there was no overlap between
𝐼 (𝑖) Informativeness of sentence 𝑖
this training set and the evaluation set of 50 documents.
𝐶 (𝑖) Set of content words in sentence 𝑖
Note that the headnotes described above are not considered to
𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛 (𝑖) Position of sentence 𝑖 in the document
be summaries of sufficiently good quality by our law experts, and
𝑆𝑐𝑜𝑟𝑒 ( 𝑗) Score (a measure of importance) of content word 𝑗
this is why the experts preferred to write their own summaries
𝑇𝑗 Set of sentences where content word 𝑗 is present
as gold standard for the documents in the evaluation set. It can
𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘) Weight (a measure of importance) of segment 𝑘
be argued that it is unfair to train the supervised summarization
models using summaries that are known to be of poorer quality 𝑆𝑘 Set of sentences belonging to segment 𝑘
than the target summaries. However, the supervised summarization 𝑁 𝑂𝑆𝑘 Minimum number of sentences to be selected from
segment 𝑘 in the summary
models require thousands of document-summary pairs for training,
and it is practically impossible to obtain so many summaries of Table 1: Notations used in the DELSumm algorithm
the target quality. Hence our only option is to use the available
headnotes for training the supervised summarization models. This included in the summary. We assume that domain experts will de-
situation can be thought of as a trade-off between quantity and fine how to estimate the informativeness of a sentence.
quality of training data – if a method requires large amounts of (ii) Content words: Content words signify domain-specific vo-
training data, then that data may be of poorer quality than the cabulary (e.g., terms from a legal dictionary, or names of statutes,
target quality. etc) and noun-phrases. The importance of each content word 𝑗
is given by 𝑆𝑐𝑜𝑟𝑒 ( 𝑗). We assume that domain experts will define
what words/terms should be considered as ‘content words’, and
4 PROPOSED APPROACH: DELSUMM
how important the content words are.
In this section, we describe our proposed algorithm DELSumm . A summary of length 𝐿 words, consisting of the most informa-
Our algorithm uses an optimization framework to incorporate legal tive sentences and content words, is achieved by maximizing the
domain knowledge (similar to what is stated in [9, 11, 12]) into an following objective function:
objective function with constraints. The objective function is then
Õ𝑛 Õ
𝑚
maximized using Integer Linear Programming (ILP). The symbols 𝑚𝑎𝑥 ( 𝐼 (𝑖) · 𝑥𝑖 + 𝑆𝑐𝑜𝑟𝑒 ( 𝑗) · 𝑦 𝑗 ) (1)
used to explain the algorithm are stated in Table 1. 𝑖=1 𝑗=1
We consider that a case document has a set of 𝑔 rhetorical seg-
subject to constraints
ments (e.g., the 𝑔 = 8 rhetorical segments stated in Section 2.5),
and the summary is supposed to contain a representation of each Õ
𝑛
segment (as indicated in [9, 11, 12]). The algorithm takes as input 𝑥𝑖 · 𝐿(𝑖) ≤ 𝐿 (2)
– (i) a case document where each sentence has a label signifying
Õ
𝑖=1
its rhetorical segment, and (ii) the desired number of words 𝐿 in
𝑦 𝑗 >= |𝐶 (𝑖)| · 𝑥𝑖 , 𝑗 = [1 . . . 𝑚] (3)
the summary. The algorithm then outputs a summary of at most 𝐿
𝑗 ∈𝐶 (𝑖)
words containing a representation of each segment. Õ
𝑥𝑖 ≥ 𝑦 𝑗 , 𝑖 = [1 . . . 𝑛] (4)
The optimization framework: We formulate the summarization
Õ
𝑖 ∈𝑇 𝑗
problem using an optimization framework. The ILP formulation
maximizes the following factors: 𝑥𝑖 ≥ 𝑁𝑂𝑆𝑘 , 𝑘 = [1 . . . 𝑔] (5)
(i) Informativeness of a sentence: The informativeness 𝐼 (𝑖) of 𝑖 ∈𝑆𝑘
a sentence 𝑖 defines the importance of a sentence in terms of its The objective function in Eqn. 1 tries to maximize the inclusion
information content. More informative sentences are likely to be of informative sentences (through the 𝑥𝑖 indicator variables) and

25
Incorporating Domain Knowledge for Extractive Summarization of Legal Case Documents
the number of important content words (through the 𝑦 𝑗 indicator
variables). Here 𝑥𝑖 (respectively, 𝑦 𝑗 ) is set to 1 if the algorithm
decides that sentence 𝑖 (respectively, content word 𝑗) should be
included in the summary. Eqn. 2 constraints that the summary
length is at most 𝐿 words (note from Table 1 that 𝐿(𝑖) is the number
of words in sentence 𝑖). Eqn. 3 implies that if a particular sentence
𝑖 is selected for inclusion in the summary (i.e., if 𝑥𝑖 = 1), then the
all the content words contained in that sentence (𝐶 (𝑖)) are also
selected. Eqn. 4 suggests that if a content word 𝑗 is selected for
inclusion in the summary (i.e., if 𝑦 𝑗 = 1), then at least one sentence
where that content word in present is also selected.
Eqn. 5 ensures from each segment 𝑘 a minimum number of
sentences 𝑁𝑂𝑆𝑘 should be selected in the summary. This is a key
step that ensures representation of all rhetorical segments in the
generated summary. We assume that suitable values of 𝑁𝑂𝑆𝑘 will
be obtained from domain knowledge.
Any ILP solver can be used to solve the ILP; we specifically used
the GUROBI optimizer (http://www.gurobi.com/). Finally, those
sentences for which 𝑥𝑖 is set to 1 will be included in the summary
generated by DELSumm .
Handling Redundancy: DELSumm implicitly handles redundancy
in the summary through the 𝑦 𝑗 variables that indicate inclusion of
content words (refer to Eqn. 3). According to Eqn. 3, if a sentence is
included in the summary, then the content words in that sentence
are also selected. Consider that two sentences 𝑖 and 𝑖 ′ are similar,
so that including both in the summary would be redundant. The
two sentences are expected to have the same content words. Hence,
adding both 𝑖 and 𝑖 ′ (which have the same content words) will not
help in maximizing the objective function. Hence the ILP method
is expected to refrain from adding both sentences, thus prohibiting
redundancy in the summary.
Applying DELSumm to case documents of a specific juris-
diction: It can be noted that this section has given only a general
description of DELSumm , where it has been assumed that many
details will be derived from domain knowledge (e.g., the informa-
tiveness or importance of a sentence or content word, or values of
𝑁𝑂𝑆𝑘 ). In the next section, we specify how these details are derived
from the guidelines given by domain experts from India.
5 APPLYING DELSUMM ON INDIAN CASE
DOCUMENTS
In this section, we describe how DELSumm was adopted to summa-
rize Indian case documents. We first describe the summarization
guidelines stated by law experts from India, and then discuss how
to adopt those guidelines into DELSumm . We then discuss com-
parative results of the proposed algorithm and the baselines on the
India dataset.
5.1 Guidelines from Law experts
We consulted law experts – senior law students and faculty mem-
bers from the Rajiv Gandhi School of Intellectual Property Law (a
reputed law school in India) – to know how Indian case documents
should be summarized. Based on the rhetorical segments described
in Section 2.5, the law experts suggested the following guidelines.
• (G1) In general, the summary should contain representations
26
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
from all segments of a case document, except the segment ‘Ruling
by Lower Court’ which may be omitted. The relative importance
of segments in a summary should be: Final judgment > Issue >
Fact > (Statute, Precedent, Ratio) > Argument.
• (G2) The segments ‘Final judgement’ and ‘Issue’ are excep-
tionally important for the summary. The segments are usually
very short in the documents, and so they can be included completely
in the summary.
• (G3) The important sentences in various rhetorical segments
(Fact, Statute, Precedent, Ratio) should be decided as follows – (a)
Fact: sentences that appear at the beginning; (b) Statute : sentences
that contain citations to an Act; (c) Precedent : sentences that con-
tain citation to a prior-cases; (d) Ratio : sentences appearing at
the end of the document and sentences that contain citation to an
act/law/prior-case.
• (G4) The summary must contain sentences that give important
details of the case, including the Acts and sections of Acts that were
referred, names of people, places etc. that concern the case, and so
on. Also sentences containing specific legal keywords are usually
important.
5.2 Operationalizing the guidelines
Table 2 shows how the above-mentioned guidelines from law ex-
perts in India have been operationalized in DELSumm .
• Operationalizing G1: We experiment with two ways of assign-
ing weights to segments – linearly decreasing and exponentially
decreasing – based on guideline G1 stated above. We finally decided
to go with the exponentially decreasing weights. Specifically, we
assign weight 27 to the ‘Final Judgement’ segment (that is judged
to be most important by experts), followed by 26 for ‘Issue’, 25
for ‘Fact’, 23 for ‘Statute’, ‘Ratio & ‘Precedent’ and finally 21 for
‘Argument’.
• Operationalizing G2: As per guideline G2, 𝑁𝑂𝑆𝑘 (the minimum
number of sentences from a segment 𝑘) is set to ensure a minimum
representation of every segment in the summary – the ‘final judge-
ment’ and ‘issue’ rhetorical segments are to be included fully, and
at least 2 sentences from every other segment are to be included.
• Operationalizing G3: The informativeness of a sentence 𝑖 de-
fines the importance of a sentence in terms of its information con-
tent. More informative sentences are likely to be included in the
summary. To find the informativeness 𝐼 (𝑖) of a sentence 𝑖, we use the
guideline G3 stated above. 𝐼 (𝑖) therefore depends on the rhetorical
segment 𝑘 (Fact, Issue, etc.) that contains a particular sentence 𝑖. For
instance, G3(a) dictates that sentences appearing at the beginning
of the rhetorical segment ‘Fact’ are important. We incorporate this
guideline by weighing the sentences within ‘Fact’ by the inverse of
its position in the document.
According to G3(b) and G3(c), sentences that contains mentions
of Statutes/Act names (e.g., Section 302 of the Indian Penal Code,
Article 15 of the Constitution, Dowry Prohibition Act 1961, etc.) and
prior-cases/precedents are important. We incorporate this guideline
through a Boolean/indicator variable 𝑎𝑖 which is 1 if the sentence
contains a mention of a Statute. We use regular expression patterns
to detect Statute/Act name mentions. Similarly for detecting if a
sentence contains a reference to a prior case, we use the Boolean
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh, and Saptarshi Ghosh

Guideline Operationalization in DELSumm

(G1) Segment weights Highest 𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘) assigned to Final judgement and Issue; weight decreasing exponentially /
linearly for other segments: 𝐹𝑖𝑛𝑎𝑙 𝐽𝑢𝑑𝑔𝑒𝑚𝑒𝑛𝑡 > 𝐼𝑠𝑠𝑢𝑒 > 𝐹𝑎𝑐𝑡 > 𝑆𝑡𝑎𝑡𝑢𝑡𝑒, 𝑅𝑎𝑡𝑖𝑜, 𝑃𝑟𝑒𝑐𝑒𝑑𝑒𝑛𝑡 >
(
𝐴𝑟𝑔𝑢𝑚𝑒𝑛𝑡
(G2) Segment |𝑆𝑘 | , if 𝑘 = Final judgement, Issue
𝑁𝑂𝑆𝑘 =
representation 𝑚𝑖𝑛(2, |𝑆𝑘 |), otherwise




𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘) · (1/𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛(𝑖)), if 𝑘 = Fact



 𝑒𝑖𝑔ℎ𝑡 (𝑘) · 𝑎𝑖 , if 𝑘 = Statute


𝑊
(G3) Sentence
𝐼 (𝑖) = 𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘) · 𝑝𝑖 , if 𝑘 = Precedent
Informativeness 


𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘) · 𝑝𝑜𝑠𝑖𝑡𝑖𝑜𝑛(𝑖) · (𝑝𝑖 OR 𝑎𝑖 ), if 𝑘 = Ratio



𝑊 𝑒𝑖𝑔ℎ𝑡 (𝑘), otherwise



(G4) Content Words & 


5, if j = Act/sections of Acts (e.g.,Indian Penal Code, 1860; Article 15, etc.)
𝑆𝑐𝑜𝑟𝑒 ( 𝑗) = 3, if j = Keywords from a legal dictionary
their weights 

 1, if j = Noun Phrases = 1

Table 2: Parameter settings in DELSumm for operationalizing the guidelines provided by law experts for Indian case docu-
ments (stated in Section 5.1). Parameter settings such as weights of various segments, how to compute importance of sentences
in various segments, etc. are decided based on the segment 𝑘. The symbols are explained in Table 1.

variable 𝑝𝑖 . If a regular expression pattern of the form Party 1 vs.
Party 2 is found in a sentence 𝑖, 𝑝𝑖 is set to 1.
For understanding important Ratio sentences, we lookup guide-
line G3(d), which mandates that sentences containing references
to either a Statute/Act name or a prior-case are important to be in-
cluded in the summary. We therefore use the two Boolean/indicator
variables 𝑎𝑖 and 𝑝𝑖 for detecting the presence of a Statute/Act name
or a prior-case. If any one of them occurs in the sentence, we con-
sider that sentence from segment Ratio to be informative.
• Operationalizing G4: Based on guideline G4 stated above, we
need to identify some important ‘content words’ whose presence
would indicate the sentences that contain especially important
details of the case. To this end, we consider three types of content
words – mentions of Acts/sections of Acts, keywords from a legal
dictionary, and noun phrases (since names of people, places, etc. are
important). For identifying legal keywords, we use a legal dictionary
from the website advocatekhoj.com. For identifying mentions of
Acts and statutes, we use a comprehensive list of Acts in the Indian
judiciary, obtained from Westlaw India (westlawindia.com). As
content word scores (𝑆𝑐𝑜𝑟𝑒 ( 𝑗)), we assign a weight of 5 to statute
mentions, 3 to legal phrases and 1 to noun phrases; these scores are
based on suggestions by the legal experts on the relative importance
of sentences containing different types of content words.
6 BASELINES AND EXPERIMENTAL SETUP
This section describes the baseline summarization methods that
we consider for comparison with the proposed method. Also the
experimental setup used to compare all methods is discussed.
6.1 Baseline summarization methods
As described in Section 2, there are four classes of methods that can
be applied for extractive summarization of legal case documents. We
consider some representative methods from each class as baselines.
Unsupervised domain-independent methods: We consider as
baselines the following popular algorithms from this class: (1) Luhn [16],
(2) LexRank [7], (3) Reduction [13], and (4) LSA [10] (see Section 2.1
for brief descriptions of these methods).
Unsupervised domain-specific methods: Section 2.2 gave brief
descriptions of these methods. Among such methods, we consider
the following four as baselines – (1) LetSum [8], (2) KMM [20],
(3) CaseSummarizer [18] and (4) a simplified version of MMR [25]
– in the absence of identical datasets, we adopt only the Maximum
Margin Relevance module and use it to summarize a document.
Supervised domain-specific methods: From this class of algo-
rithms, we consider Gist [14], a recent method that applies general
ML algorithms to the task. The best performance was observed by
using Gradient Boosted Decision Tree as the ML classifier, which
we report (see Section 2.4 for details of the method).
Supervised domain-independent methods: Among these meth-
ods, we consider the neural method SummaRuNNer [17] (imple-
mentation available at https://github.com/hpzhao/SummaRuNNer).
Similar to Gist, they consider the task of extractive summarization
as binary classification problem. The classifier returns a ranked
list of sentences based on their prediction/confidence probabilities
about their inclusion in the summary. We include sentences from
this list in decreasing order of their predicted probabilities, until
the desired summary length is reached.
We also apply a recent BERT-based summarization method BERT-
SUM [15] (implementation available at https://github.com/nlpyang/
PreSumm). In BERTSUM, the sentence selection function is a simple
binary classification task (whether or not to include a sentence in
the summary).2 Similar to SummaRuNNer, we use the ranked list of
sentences based on the confidence probabilities of their inclusion
in the summary. We include sentences one-by-one into the final
summary, until the desired length is reached.
2 The original BERTSUM model uses a post-processing step called Trigram Blocking
that excludes a candidate sentence if it has a significant amount of trigram overlap with
the already generated summary (to minimize redundancy in the summary). However,
we observed that this step leads to summaries that are too short, as also observed
in [22]. Hence we ignore this step.

27
Incorporating Domain Knowledge for Extractive Summarization of Legal Case Documents ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Implementations of the baselines: For Luhn, LexRank, Reduc- ROUGE 2 ROUGE-L

Algorithm
tion and LSA, we use the implementations from the Python sumy R F R F
module. The implementations of SummaRuNNEr and BERTSUM Unsupervised Domain Independent
are also available publicly, at the URLs stated above. We imple- (i) LexRank 0.3662 0.3719 0.5068 0.5392
mented the legal domain-specific summarization methods – Cas- (ii) LSA 0.3644 0.3646 0.5632 0.5483
eSummarizer, LetSum, GraphicalModel, MMR and Gist – following
(iii) Luhn 0.3907 0.3873 0.5424 0.5521
the descriptions in the corresponding papers as closely as possible
(the same parameter values were used as specified in the papers). (iv) Reduction 0.3778 0.3814 0.5194 0.542
Unsupervised Domain Specific
Converting abstractive summaries to extractive for training
(v) LetSum 0.403 0.4137 0.5898 0.5846*
supervised summarization models: The supervised methods
(vi) KMM 0.354 0.3546 0.5407 0.5385
Gist, SumaRuNNer and BERTSUM require labelled data for train-
ing, where every sentence in the source document must be labeled (vii) CaseSummarizer 0.3699 0.3765 0.4925 0.5349
as 1 if this sentence is suitable for inclusion in the summary, and (viii) MMR 0.3733 0.3729 0.6064 0.568
labeled as 0 otherwise. The publicly available headnotes in our DELSumm (proposed) 0.4323 0.4217 0.6831 0.6017
training datasets (see Section 3) are abstractive in nature and can- Supervised Domain Independent
not be directly used to train such methods. Hence we convert the (ix) SummaRuNNer 0.4104* 0.4149* 0.5835 0.5821
abstractive summaries to extractive for training these methods. (x) BERTSUM 0.4044 0.4048 0.56 0.5529
To this end, we adopt the technique mentioned in [17]. Briefly,
Supervised Domain Specific
we assign label 1 (suitable for inclusion in the summary) to those
sentences from the source document (full text), that greedily max- (xi) Gist 0.3567 0.3593 0.5712 0.5501
imizes the ROUGE 2-overlap with the abstractive summary. The Table 3: Performance of the summarization methods as per
rest of the sentences in the source document are assigned label 0. ROUGE metrics (R = recall, F = Fscore). All values are aver-
aged over the 50 documents in the evaluation set. The best
6.2 Experimental setup value for each measure is in bold and the second best in
Now we discuss the setup for our experiments to compare the underlined. Entries with asterisk(*) indicates DELSumm is
performances of various summarization models. not statistically significant than baselines using 95% confi-
dence interval by Student’s T-Test.
Target length of summaries: Summarization models need to be
told the target length 𝐿 (in number of words) of the summaries that
they need to generated. For every document in the India dataset, we various methods in summarizing the Indian Supreme Court case
have two gold standard summaries written by two domain experts. documents. In both tables, ROUGE scores are averaged over the 50
For a particular document, we consider the target summary length documents in the India evaluation set.
𝐿 to be the average of the number of words in the summaries (for
that document) written by the two experts.
7.1 Evaluation of the full summaries
For each document, every summarization model is made to gen-
erate a summary of length at most 𝐿 words, which ensures a fair Table 3 shows the performances of summarizing the full documents
comparison among the different models. in terms of ROUGE-2 and ROUGE-L Recall and F-scores. Consider-
ing all the approaches, DELSumm achieves the best performance
Evaluation of summary quality: We use the standard ROUGE across all the measures. The second-best metric values are achieved
metrics for evaluation of summary quality (computed using https: by LetSum (For ROUGE-L F-score), MMR (for ROUGE-L Recall),
//pypi.org/project/rouge/ v. 1.0.0). For a particular document, we and SummaRuNNer (for ROUGE-2 Recall and F-score). To check
have gold standard summaries written by two legal experts (as if their performances are statistically significantly different from
was described in Section 3). We compute the ROUGE scores of those achieved by DELSumm , we perform Student’s T-Test at 95%
the algorithmic summary and each of the gold standard summary, (𝑝 < 0.05) with these three methods. Entries that are not statistically
and take the average of the two. We remove a common set of significant are marked with asterisk (*) in Table 3.
English stopwords from both the gold standard summaries and the It can be noted that, out of all the domain-specific methods in
algorithm-generated summaries before measuring ROUGE scores. Table 3 (both supervised and unsupervised), DELSumm performs
We report ROUGE-2 Recall and F-score, and ROUGE-L Recall and the best for all the measures. While LetSum is the only prior work
F-scores. The scores are averaged over all 50 documents in the that considers rhetorical segments while generating the summary
evaluation set. (see Section 2), none of these methods take into account domain-
specific guidelines on what to include in the summary. DELSumm
7 RESULTS AND ANALYSES on the other hand models the legal knowledge about what to include
This section compares the performances of the proposed DELSumm in a summary, thus being able to perform better.
and all the baseline algorithms. We compare the performances Importantly, DELSumm outperforms even supervised neural
broadly in two ways – (i) how good the summaries for the full models such as SummaRuNNer and BERTSUM (though the differ-
documents are, and (ii) how well each individual rhetorical segment ence in the performance of SummaRuNNer with respect to DEL-
is summarized. Table 3 and Table 4 show the performances of Summ is not statistically significant in terms of ROUGE-2 scores).

28
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh, and Saptarshi Ghosh

Final
Issue Facts Statute
Precedent
Argument Similar to [2], we also observe that many of the baseline meth-
Algorithm judgement
(1.2%) (23.9%) (7.1%)
+Ratio
(8.6%) ods could not represent the ‘Final judgement’ and ‘Issue’ well in
(2.9%) (53.1%)
their summaries. This flaw is especially critical since these two
Unsupervised, Domain Independent
segments are the most important in the summary according to
LexRank 0.0619 0.3469 0.4550 0.2661 0.3658 0.4284
the law experts (see Section 5.1). In contrast, DELSumm achieves
LSA 0.0275 0.2529 0.5217 0.2268 0.3527 0.3705
Luhn 0.0358 0.2754 0.5408 0.2662 0.2927 0.3781
a very high performance on these two important segments. This
Reduction 0.0352 0.3153 0.5064 0.2579 0.3059 0.4390
difference in performance is possibly because these two segments
Unsupervised, Domain Specific
are also the shortest segments (constitutes only 2.9% and 1.2% of
LetSum 0.0423 0.3926 0.6246 0.3469 0.3853 0.2830 the whole document, as stated in the first row of Table 4), and
KMM 0.3254 0.2979 0.4124 0.3415 0.4450 0.416 hence are missed by other methods which do not know of their
CaseSummarizer 0.2474 0.3537 0.4500 0.2255 0.4461 0.4184 domain-specific importance. This observation show the necessity
MMR 0.4378 0.3548 0.4442 0.2763 0.4647 0.3705 of an informed algorithm that can incorporate domain knowledge
DELSumm 0.7929 0.6635 0.5539 0.4030 0.4305 0.4370 from experts.
Supervised, Domain Independent DELSumm also represents the ‘Statute’ segment better than all
SummaRuNNer 0.4451 0.2990 0.5231 0.1636 0.5215 0.3090 other methods. For the ‘Statute’ segment, the algorithm was for-
BERTSUM 0.0662 0.3544 0.6376 0.2535 0.3121 0.3262 mulated in such a way (through the 𝑎𝑖 variable) that it is able to
Supervised, Domain Specific incorporate sentences that contain mention of an Act/law. Other
Gist 0.5844 0.3856 0.4621 0.2759 0.4537 0.2132 methods did not perform well in this aspect. The performance of
Table 4: Segment-wise performance (ROUGE-L F-scores) of DELSumm for the ‘Argument’ segment is second-best (0.4370) after
the methods. All values are averaged over the 50 documents that of Reduction (0.4390). These values are very close, and the
in the evaluation set. Values in the column headings are the difference is not statistically significant.
% of sentences in the full document that belong to a segment. Our method is unable to perform as well for the ‘Precedent +
Values < 0.3 highlighted in red-underlined. The best value Ratio’ and ‘Facts’ segments as some other methods. Note that, these
for each segment in green-bold. segments accounts for maximum number of sentences in a document
Note that DELSumm is a completely unsupervised method while and also forms a large part of the summary (see Table 4, first row).
SummaRuNNer and BERTSUM are deep-learning based supervised Hence neural methods (e.g., BERTSUM) and methods relying of
methods trained over 7, 100 document-summary pairs. The fact TF-IDF measures (e.g., LetSum) obtained relatively large amounts
that DELSumm still outperforms these supervised models is be- of training data for these segments, and hence performed well for
cause DELSumm intelligently utilizes domain knowledge while these segments.
generating the summaries. Finally, although LetSum is a segment-aware algorithm, its sum-
marization mechanism is not strong enough to understand the
importance of sentences in all segments; for instance, it performs
7.2 Evaluation of segment-wise summarization very poorly for the ‘Final judgement’ segment. DL methods fail to
Overall ROUGE scores are not the best metrics for evaluating case perform well for the smaller segments (e.g., Final judgement, Issue)
document summaries. Law experts opine that, even methods that for which lesser amounts of training data is available.
achieve high overall ROUGE scores may not represent every seg- Overall, it is to note that DELSumm achieves a much more
ment well in the summary [2]. A segment-wise performance evalu- balanced representation of segments in the summary, com-
ation is practically important since law practitioners often intend pared to all the baseline methods. Contrary to the baseline methods,
to read the summary of a particular segment, and not always the DELSumm shows decent performance for all the segments. This
full summary. Hence we perform a segment-wise performance has been taken care mainly through the constraint in Eqn. 5. Hence,
evaluation of the algorithms. even when the most important segments (e.g., Final judgement) are
To this end, we proceed as follows. For each rhetorical segment optimized in the summary, the less important ones (e.g., Arguments)
(e.g., Facts, Issues), we extract the portions of an algorithmic sum- are not missed out.
mary and the gold standard summary that represent the given
segment, using the gold standard rhetorical labels of individual
sentences (which are present in our evaluation set, as detailed in
Section 3). Then we compute the ROUGE scores on those specific
text portions only. For instance, to compute the ROUGE score on the 7.3 Discussion on the comparative results
‘Fact’ segment of a particular document, we only consider those sen- The results stated above suggest that our proposed method per-
tences in the gold standard summary and the algorithmic summary, forms better than many existing domain-independent as well as
which have the rhetorical label ‘Fact’ in the said document. We legal domain-specific algorithms. The primary reason for this su-
report the average ROUGE score for a particular segment, averaged perior performance of DELSumm is that it is able to incorporate
over all 50 documents in the evaluation set. legal domain knowledge much more efficiently through a theoret-
Table 4 shows the segment-wise ROUGE-L F-scores (averaged ically grounded approach. By doing this, it is able to surpass the
over all 50 documents in the evaluation set). The values < 0.3 are performance of deep learning and machine learning approaches
underlined and highlighted in red color, while the best value for such as SummaRuNNer, BERTSUM and Gist (that are trained over
each segment is highlighted in boldface and green color. 7, 100 document-summary pairs). These results show that if legal

29
Incorporating Domain Knowledge for Extractive Summarization of Legal Case Documents
domain knowledge can be used intelligently, even simple unsuper-
vised summarization algorithms can compete with state-of-the-art
supervised learning approaches.
It can be argued that 7, 100 document-summary pairs are too
less to train a deep learning summarization model; additionally,
as stated in Section 3, the summaries (headnotes) in the training
set may not be of the same quality as those in the evaluation set.
However, in domains such as legal, it is difficult and expensive to
obtain rich quality training data in huge amounts. In such scenarios,
DELSumm , being an unsupervised method, can work efficiently
by incorporating domain knowledge.
Another advantage of DELSumm is that it is easily customiz-
able – if the requirement is that a particular segment should have
more representation in the summary than the others, the segment
weights and representations can be tuned accordingly. If the sen-
tence importance measures are different, that too can be adjusted.
Other domain-specific and supervised learning methods do not
exhibit such flexibility.
8 USING DELSUMM WITH ALGORITHMIC
RHETORICAL LABELS
DELSumm requires sentences of the source document (full text) to
be labelled with a rhetorical role prior to summarization. To this
end, till now we have been using expert-annotated labels. However,
in a real scenario, it is not possible to get expert-annotations for
rhetorical role labelling for every document. A more practical ap-
proach is to first computationally generate these rhetorical labels
through some algorithm that detects rhetorical roles of sentences
in a legal document, and then consider the labelled document as
input to DELSumm for summarization.
Note that algorithmically derived rhetorical labels are likely to
be noisy (i.e., not 100% accurate). In this section, we analyse how
the performance of DELSumm is affected by such noisy algorithmic
labels when compared to gold standard labels. To this end, we
perform the following experiment (which is inspired by the idea of 𝑘-
fold cross validation that is popular in Machine Learning literature).
Experimental Design: Assume that we have a dataset 𝐷 consist-
ing of 𝑛 documents, each having sentence-wise rhetorical label
annotations as well as gold-standard summaries (e.g., the Indian
dataset we are using in this paper, with 𝑛 = 50). We divide the
dataset into 𝑘 folds (e.g., 𝑘 = 5). In each fold 𝑓 , there are 𝑛/𝑘 docu-
ments for evaluation, which we call as the evaluation set 𝐸 𝑓 . The
remaining 𝑛 − (𝑛/𝑘) documents are considered for training a model
for rhetorical role labeling, termed as training set 𝑇𝑓 . We also have
the expert annotated labels for the documents in fold 𝑓 , and their
gold standard summaries. For each fold, we use 𝑇𝑓 for training the
state-of-the-art rhetorical role identification model Hier-BiLSTM-
CRF from our prior work [3].3
Using the trained model for that fold, we infer the rhetorical
labels of the documents in the evaluation set 𝐸 𝑓 for that fold. Now,
we have two sets of rhetorical labels for the sentences in the doc-
uments in 𝐸 𝑓 – (i) gold standard labels given by law experts, and
(ii) algorithmic labels assigned by the model [3]. We next summarize
every document in 𝐸 𝑓 with DELSumm twice, once using the gold
3 Implementation
available at https://github.com/Law-AI/semantic-segmentation.
30
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 1: Effect on the performance of DELSumm with al-
gorithmic labels. There is minimal degradation of ROUGE
scores when using algorithmic labels
standard sentence labels, and again using the algorithmic sentence
labels. We perform three evaluations:
(i) Accuracy of rhetorical labelling: we check what fraction of
the algorithmic labels match with the gold standard labels (over
the documents in 𝐸 𝑓 ). This measure gives an idea of the accuracy
of the rhetorical labeling model [3].
(ii) ROUGE scores with Gold Std. Labels: We compute the ROUGE
scores between the gold standard summaries and the summaries
generated by DELSumm considering the gold standard labels. We
average the ROUGE scores over all documents in the set 𝐸 𝑓 (this is
very similar to what we did in Section 5).
(iii) ROUGE scores with Algorithmic Labels: We compute the
ROUGE scores between the gold standard summaries and sum-
maries generated by DELSumm with the algorithmic labels, i.e.,
labels inferred by the model trained over 𝑇𝑓 . Again we average the
ROUGE scores over all documents in the set 𝐸 𝑓 .
The differences between these two sets of ROUGE scores can be
used to quantify the degradation in performance of DELSumm
while using noisy algorithmic labels instead of gold standard labels.
Results: We perform the above experiment with the India dataset
where 𝑛 = 50, and we consider 𝑘 = 5 folds. Hence in each fold,
we use a training set of 40 documents and an evaluation set of 10
documents. Figure 1 shows the results of the above experiment.
The horizontal axis shows the rhetorical labeling accuracy across
the 𝑘 = 5 folds; we see that the model [3] achieves accuracies in the
range [0.825, 0.851], i.e., about 82%–85% of the algorithmic labels
are correct. The figure shows the two types of ROUGE-L F-scores
as described above – one set using the algorithmic labels (denoted
by red diamonds), and the other set using the gold-standard labels
(denoted by black circles) – for each of the 5 folds. Each value
is averaged over the 10 documents in the evaluation set of the
corresponding fold. We find that the two sets of ROUGE scores are
very close. In other words, even when noisy/inaccurate labels are
used in DELSumm , the drop in performance is very low, indicating
the robustness of DELSumm .
Finally, we compute the ROUGE scores averaged over all 𝑛 = 50
documents across all 𝑘 = 5 folds. Table 5 shows the average ROUGE
scores achieved by DELSumm using the algorithmic labels (denoted
‘DELSumm AL’). For ease of comparison, we repeat some rows
from Table 3 – the performance of DELSumm with gold standard
rhetorical labels (denoted ‘DELSumm GSL’), and the performance
of the closest competitors LetSum, MMR and SummaRuNNer. We
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paheli Bhattacharya, Soham Poddar, Koustav Rudra, Kripabandhu Ghosh, and Saptarshi Ghosh
ROUGE-2 ROUGE-L
Algorithm
R F R F
LetSum 0.4030 0.4137 0.5898 0.5846
MMR 0.3733 0.3729 0.6064 0.5680
SummaRuNNer 0.4104 0.4149 0.5835 0.5821
DELSumm (GSL) 0.4323 0.4217 0.6831 0.6017
DELSumm (AL) 0.4193 0.4075 0.6645 0.5897
Table 5: Performance of DELSumm with Gold Standard la-
bels (GSL) and Algorithmic labels (AL) over the 50 docu-
ments. Apart from the last row, all other rows are repeated
from Table 3. While DELSumm (AL) shows slightly degraded
performance than DELSumm (GSL), it still outperforms the
other methods according to all ROUGE measures except
ROUGE-2 F-score.
observe that, even when DELSumm is used with automatically
generated algorithmic labels (with about 15% noise), it still per-
forms better than LetSum, MMR and SummaRuNNer (which were
its closest competitors) according to all ROUGE measures except
ROUGE-2 Fscore. This robustness of DELSumm is because, through
the sentence informativeness and content words measures in the
ILP formulation, DELSumm can capture useful information even if
the rhetorical labels may be a little inaccurate.
9 CONCLUSION
We propose DELSumm , an unsupervised algorithm that systemati-
cally incorporates domain knowledge for extractive summarization
of legal case documents. Extensive experiments and comparison
with as many as eleven baselines, including deep learning-based
approaches as well as domain-specific approaches, show the utility
of our approach. The strengths our approach are: (i) DELSumm
systematically encodes domain knowledge necessary for legal doc-
ument summarization into a computational approach, (ii) although
an unsupervised approach, it performs at par with supervised learn-
ing models trained over huge amounts of training data, (iii) it is
able to provide a summary that has a balanced representation from
all the rhetorical segments, which is highly lacking in prior ap-
proaches (iv) inaccuracy in labels do not degrade the performance
of DELSumm much, thus showing its robustness and rich informa-
tion identification capabilities, and (v) the method is flexible and
generalizable to summarize documents from other jurisdictions;
all that is needed are the expert guidelines for what to include
the summary, and how to identify important content words. The
objective function and constraints can be adjusted as per the re-
quirements of different jurisdictions (e.g., giving more weight to
certain segments). The implementation of DELSumm is publicly
available at https://github.com/Law-AI/DELSumm.
In future, we plan to apply DELSumm to documents from other
jurisdictions as well as to generate different types of summaries
(e.g., for different stakeholders) and analyse the performance.
Acknowledgements: The authors thank the Law experts from the
Rajiv Gandhi School of Intellectual Property Law, India who helped
in developing the gold standard data and provided the guidelines for
summarization. The research is partially supported by the TCG Cen-
tres for Research and Education in Science and Technology (CREST)
through the project titled ‘Smart Legal Consultant: AI-based Legal
31
Analytics’. This work is also supported in part by the European
Union’s Horizon 2020 research and innovation programme under
grant agreement No 832921. P. Bhattacharya is supported by a
Fellowship from Tata Consultancy Services.
REFERENCES
[1] Siddhartha Banerjee, Prasenjit Mitra, and Kazunari Sugiyama. 2015. Multi-
document abstractive summarization using ILP based multi-sentence compres-
sion. In Proc. International Conference on Artificial Intelligence.
[2] Paheli Bhattacharya, Kaustubh Hiware, Subham Rajgaria, Nilay Pochhi, Kripa-
bandhu Ghosh, and Saptarshi Ghosh. 2019. A comparative study of summariza-
tion algorithms applied to legal case judgments. In Proc. European Conference on
Information Retrieval.
[3] Paheli Bhattacharya, Shounak Paul, Kripabandhu Ghosh, Saptarshi Ghosh, and
Adam Wyner. 2019. Identification of Rhetorical Roles of Sentences in Indian
Legal Judgments. Proc. Legal knowledge and information systems (JURIX) (2019).
[4] Jianpeng Cheng and Mirella Lapata. 2016. Neural Summarization by Extracting
Sentences and Words. In Proc. Annual Meeting of the Association for Computational
Linguistics.
[5] J. Devlin, M. Chang, K. Lee, and K. Toutanova. 2019. BERT: Pre-training of Deep
Bidirectional Transformers for Language Understanding. In Proc. NAACL-HLT.
[6] Yue Dong. 2018. A Survey on Neural Network-Based Summarization Methods.
CoRR abs/1804.04589 (2018). arXiv:1804.04589 http://arxiv.org/abs/1804.04589
[7] Günes Erkan and Dragomir R. Radev. 2004. LexRank: Graph-based Lexical
Centrality As Salience in Text Summarization. J. Artif. Int. Res. 22, 1 (2004).
[8] Atefeh Farzindar and Guy Lapalme. 2004. Letsum, an automatic legal text sum-
marizing system. Proc. Legal knowledge and information systems (JURIX) (2004).
[9] Jessica Giles. 2015. Writing Case Notes and Case Comments.
http://law-school.open.ac.uk/sites/law-school.open.ac.uk/files/files/PILARS-
Writing-case-comments.pdf.
[10] Yihong Gong and Xin Liu. 2001. Generic Text Summarization Using Relevance
Measure and Latent Semantic Analysis. In Proc. International conference on Re-
search and development in information retrieval (SIGIR).
[11] how-to-brief-case-cuny 2017. How to brief a case. https://www.lib.jjay.cuny.edu/
how-to/brief-a-case.
[12] intro-case-briefing-northwestern [n.d.]. Introduction to Case Briefing.
http://www.law.northwestern.edu/law-school-life/studentservices/orientation/
documents/Orientation-Reading-Introduction-to-Case-Briefing.pdf.
[13] Hongyan Jing. 2000. Sentence reduction for automatic text summarization. In
Proc. Applied Natural Language Processing Conference.
[14] Chao-Lin Liu and Kuan-Chun Chen. 2019. Extracting the Gist of Chinese Judg-
ments of the Supreme Court. In Proc. International Conference on Artificial Intelli-
gence and Law (ICAIL).
[15] Yang Liu and Mirella Lapata. 2019. Text Summarization with Pretrained Encoders.
In Proc. EMNLP-IJCNLP.
[16] H.P. Luhn. 1958. The automatic creation of literature abstracts. IBM Journal of
Research Development,2(2) (1958).
[17] Ramesh Nallapati, Feifei Zhai, and Bowen Zhou. 2017. SummaRuNNer: A Recur-
rent Neural Network Based Sequence Model for Extractive Summarization of
Documents. In Proc. AAAI Conference on Artificial Intelligence.
[18] Seth Polsley, Pooja Jhunjhunwala, and Ruihong Huang. 2016. CaseSummarizer:
A System for Automated Summarization of Legal Texts. In Proc. Iinternational
conference on Computational Linguistics (COLING) System Demonstrations.
[19] Koustav Rudra, Pawan Goyal, Niloy Ganguly, Prasenjit Mitra, and Muhammad
Imran. 2018. Identifying Sub-Events and Summarizing Disaster-Related Informa-
tion from Microblogs. In Proc. International Conference on Research & Development
in Information Retrieval (SIGIR).
[20] M Saravanan, B Ravindran, and S Raman. 2006. Improving Legal Document
Summarization Using Graphical Models. In Legal knowledge and information
systems, JURIX.
[21] M. Saravanan, B. Ravindran, and S. Raman. 2008. Automatic Identification of
Rhetorical Roles using Conditional Random Fields for Legal Document Summa-
rization. In Proc. International Joint Conference on Natural Language Processing.
[22] Sajad Sotudeh, Arman Cohan, and Nazli Goharian. 2021. On Generating Extended
Summaries of Long Documents. In The AAAI-21 Workshop on Scientific Document
Understanding (SDU 2021).
[23] Wen Xiao and Giuseppe Carenini. 2019. Extractive Summarization of Long
Documents by Combining Global and Local Context. In Proc. EMNLP-IJCNLP.
[24] Hiroaki Yamada, Simone Teufel, and Takenobu Tokunaga. 2019. Neural Network
Based Rhetorical Status Classification for Japanese Judgment Documents.. In
Proc. Legal knowledge and information systems (JURIX).
[25] Linwu Zhong, Ziyi Zhong, Zinian Zhao, Siyuan Wang, Kevin D Ashley, and
Matthias Grabmair. 2019. Automatic Summarization of Legal Decisions Using
Iterative Masking of Predictive Sentences. In Proc. International Conference on
Artificial Intelligence and Law (ICAIL).
 AI Systems and Product Liability
Prof. Dr. Georg Borges
Institute of Legal Informatics
Saarland University
Saarbrücken, Germany
georg.borges@uni-saarland.de
ABSTRACT
The article examines whether the current product liability law
pro-vides an appropriate regulation for AI systems. This question,
which is discussed at the example of the European Product Liabil-
ity Directive, is of great practical importance in the current legal
policy discussion on liability for AI systems.
This article demonstrates that in principle the liability require-
ments are also applicable to AI systems. If the conduct of an AI
system is carefully distinguished from its properties, excessive li-
ability can be avoided. To reverse the burden of proof in favour of
the injured party in the case of faulty behaviour enables a liability
regime that is fair to the interests at stake.
However, product liability law only applies if AI systems lead di-
rectly to personal injury or damage to property. Product liability
law is not applicable insofar as AI systems indirectly lead to con-
siderable disadvantages for the person concerned, in particular
through assessments of persons. Protection against discrimination
or otherwise unfair assessments by AI systems shall be effected by
other legal instruments.
CCS CONCEPTS
• Applied computing → Law
KEYWORDS
Product liability, AI systems, Product Liability Directive, strict li-
ability, burden of proof
ACM Reference format:
Georg Borges. 2021. AI systems and product liability. In Proceedings of
ICAIL 2021, 21-25 June 2021, São Paulo, Brazil, 8 pages.
https://doi.org/10.1145/3462757.3466099
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distrib-
uted for profit or commercial advantage and that copies bear this notice and the
full citation on the first page. Copyrights for components of this work owned by
others than the author(s) must be honored. Abstracting with credit is permitted.
To copy otherwise, or republish, to post on servers or to redistribute to lists, re-
quires prior specific permission and/or a fee. Request permissions from permis-
sions@acm.org.
ICAIL'21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright is held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06…$15.00
https://doi.org/10.1145/3462757.3466099
32
1 The role of liability within the legal frame-
work for AI systems
High expectations are associated with artificial intelligence. At
the same time, the risks related to this technology are also a
highly debated topic.
This naturally applies in particular to the legal discussion, the
aim of which is to reconcile the interest in raising the benefits
associated with the introduction of new technologies with the
interest in protecting against the associated risks.
Artificial intelligence especially the production and use of AI-
equipped systems, often referred to as autonomous systems or
AI systems has been a challenge not least for the law, which has
the task of balancing the interest in enhancing the benefits asso-
ciated with the introduction of new technologies with the inter-
est in pro-tecting against the risks associated with such technol-
ogies.
1.1 Liability for AI systems as a pillar of the le-
gal framework for AI
It goes without saying that artificial intelligence and AI systems
raise questions in almost every field of law. An aspect which has
gained much attention recently, is what is often referred to as
“algorithmic fairness” [1]. In the US, the Federal Trade Commis-
sion in the USA recently published some guidance on the truth-
ful, fair and equitable use of AI [2]. Another aspect of highest
relevance certainly is liability for damage caused by AI systems.
The importance of a legal framework for AI systems and the
role of liability can be seen clearly by looking at the current dis-
cussion in Europe.
In its 2018 strategy “Artificial Intelligence for Europe” [3], the
EU Commission already explicitly focused on ethical and legal
issues in addition to the technical and economic aspects. The EU
Commission established several expert groups who support the
commission in these tasks. In the field of ethical aspects of AI,
the High Level Expert Group on AI [4] developed ethical guide-
lines for trustworthy AI which were published in June 2018 [5].
Regarding the regulation on liability and security, the Expert
Group on liability and new technologies consisting of a Product
Liability Directive formation and a “New Technologies” for-
mation was established [6].
 ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
Recently, in April 2021, the Commission presented a proposal
for an Artificial Intelligence Act (‘AI Act’) [7], which refers to
the security of so-called AI systems. The concept of AI systems is
defined very broadly in the draft and is likely to cover a large
part of modern software and software-equipped products. The
Act essentially distinguishes between three groups of AI sys-
tems. Firstly, the proposal provides a list of AI systems whose
operation is to be prohibited (Art. 5 of the proposal). The second
group of so-called high-risk AI systems will be subject to manda-
tory risk management (Art. 9 of the proposal), while a third
group of certain AI systems will be subject to transparency re-
quirements only (Art. 52 of the proposal).
As the second pillar of the legal framework for AI systems,
civil liability, i.e. the duty to compensate for damage, is of great
importance.
In its final report, the New Technologies formation of the EU
Expert Group liability for new technologies suggested amend-
ments of the existing liability law regarding AI and autonomous
systems [8].
In November 2020, the European Parliament issued a resolu-
tion on the civil liability regime for AI systems [9] which even
contains a proposal for a new regulation on liability for systems
equipped with artificial intelligence [10].
1.2 Risks Associated with the Use of AI
The risks associated with the use of AI systems are very differ-
ent. Legal protection against these risks depends, among other
things, on what protected interests are affected and whether the
AI system in question can cause damage directly or whether the
respective AI system only has an indirect effect through use of
its activity by third parties. These differences are expressed in
two groups of cases that are currently intensively discussed and
investigated in the German "ExamAI" project, among others [11].
One case group concerns damage to persons and property di-
rectly caused by AI systems, also called cyber-physical systems
[12]. The Uber vehicle accident of 2019, in which a highly auto-
mated vehicle hit and fatally injured a pedestrian [13], vividly
demonstrated this danger, hence the awareness of the risks in-
duced by AI systems raised significantly.
Another group of cases concerns assessments of general in-
formation on individuals provided by an AI system and used by
third parties for decisions which interfere with the rights of the
affected individuals. In this group of cases, there is no direct in-
terference with health or property and, most importantly, the in-
terference is not directly caused by the AI system itself.
For the latter case group, connected with the keyword “bias
in the data”, the COMPAS system can be mentioned, as it illus-
trates the risks arising from ‘biased’ algorithms even though the
related questions touch criminal law rather than liability law.
COMPAS generates an evaluation of a human being (here: prob-
ability of recidivism of offenders) that has been used by third
parties (here: judges) to make a decision that was highly relevant
for human beings (here: offenders). The evaluation generated by
the system was criticized for discriminating against certain
groups of people [14]. although this criticism is not without con-
troversy [15].
33
G. Borges
In the current discussion, systems for pre-selecting job appli-
cants for jobs are often mentioned [16].
An essential characteristic of damage caused directly or indi-
rectly by AI systems is that the damage is not caused by a physi-
cal feature of the system, such as its material properties, but by
its – own – behaviour, i.e., behaviour that is not directly con-
trolled by a human being but by the AI immanent in the system.
It is not the human driver who steers the car against the victim,
but the automated vehicle itself. The assessment of the criminal
or job seeker is not carried out by a human, but by an AI system.
This special feature of AI systems is of decisive importance
for liability law, since the damage cannot be directly traced back
to the behaviour of a particular natural person which is why the
justification of and the responsibility for the damage must be
discussed.
The importance of liability law is based on the dual function
of civil liability, which grants the injured party a claim for dam-
ages against the addressee of liability. The rules on civil liability
traditionally aim to compensate the victim for damage suffered
to protected legal assets [17, p. 259f.; 18, p. 166, 565]. At the same
time, the threat of liability also serves to steer behaviour by
providing an incentive to avoid damage [17, p. 187; 18, p. 169].
However, the risk of liability can also trigger undesirable strate-
gies of liability avoidance. In the current debate, for example, a
possible chilling effect of liability risks to the development and
market launch of products equipped with artificial intelligence is
often mentioned.
In the current legal policy discussion, a tendency is emerging
to assign responsibility to the operator of the AI system, at least
in the sense of civil liability for damages. In Europe, Art. 4 of the
proposed regulation on liability for the operation of AI systems
[10] imposes strict liability on the operator of AI systems classi-
fied as dangerous for damages resulting from the operation of
the AI system. Art. 11 of the proposed regulation [10] states that
manufacturers of AI systems are deliberately not covered by the
directive.
The proposed regulation mirrors the current discussion in
Europe regarding liability for highly automated vehicles. Some
authors, e.g., Buck-Heeb and Dieckmann [19] or Kreutz [20], re-
fer to the existing strict liability of the owner of a motor vehicle
and, with regard to the manufacturer, to product liability law.
The latter raises the question of whether product liability law
provides an adequate liability regime for AI systems.
The term product liability refers to the non-contractual, strict
liability of the producer for damage caused by a defective prod-
uct.
Product liability was introduced in many countries in the sec-
ond half of the 20th century, starting in the U.S. which is consid-
ered the “birthplace” of product liability [21]. The transition
from fault-based liability to strict liability of the manufacturer
for defective products was heavily influenced by the 1963 land-
mark decision of the California Supreme Court in Greenman v
Yuba Power Products [22]. This approach found its way into §
402A of the Restatement Second of Torts from 1966 and was
later accepted in most states of the US [23, p.19; 24, p. 251].
 AI Systems and Product Liability
In Europe, product liability was mainly introduced by way of
statutory law based on the European Product Liability Directive
of 1985 [25].
This issue of whether product liability is able to offer ade-
quate compensation and incentives is discussed in this article on
the basis of the European Product Liability Directive, which is of
particular interest insofar as it is currently subject to evaluation
and possible revision, especially with regard to AI systems [26].
2 The liability concept of the Product Liability
Directive
The 1985 European Product Liability Directive introduced strict
liability for defective products in European tort law [27]. Con-
sumers who suffer damage caused by a product have a claim for
damages against the producer of the defective product.
Unlike in general tort law, liability is strict, i.e., it requires
neither intent nor negligence on the part of the producer. How-
ever, it is not strict liability in the sense of causal liability, since
liability only applies when the product has a defect and this was
the cause of the damage [28]. Since a defect exists if the product
does not provide the safety that could reasonably be expected,
the result is that the conditions for liability are closely approxi-
mated to negligence [29].
Liability not only applies to end products, but also to basic
materials and component parts, and therefore addresses not only
the manufacturer of the finished product, but also the manufac-
turers of parts and basic materials (infra 3.1). Thus, the directive
deliberately covers all stages of the value chain [29, no. 13; 30,
no. 28].
Liability is limited in several respects to avoid excessive lia-
bility risks for the manufacturer [31, p. 52f.]. First of all, accord-
ing to Art. 9 lt. b) of the Product Liability Directive [25], only pri-
vate users of the final product are beneficiaries, not, for example,
commercial users of the product [30, no. 15].
Liability only applies to damage to health or property, in par-
ticular it does not apply to mere financial losses [32; 33, p. 468].
The list of protected rights in Art. 1 of the Product Liability Di-
rective, [25] is exhaustive [29, no. 19; 34, no. 2]. Negative evalua-
tions of a person, which at most might infringe the persons’ per-
sonality rights, are therefore not covered by product liability
law.
Finally, liability is limited in amount to ensure insurability. It
thus has the following prerequisites: Damage to a protected right
(health or property) must have been caused by a product. The
product must have a defect that was the cause of the damage.
The burden of proof for all prerequisites is on the injured party.
The claim for damages is directed against the producer of the
finished product as well as the manufacturer of the defective
component part or basic material.
This liability concept gives rise to specific challenges in the
area of AI systems.
3 Questions of Application of Product Liability
in the Case of AI
34
ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
3.1 AI as a Product
An essential precondition of product liability is that the damage
was caused by a product. According to the definition of Art. 2 of
the Product Liability Directive, a product is any movable thing,
even if it forms part of another movable thing or of an immova-
ble thing, as well as electricity. Cyber-physical systems, such as
highly automated vehicles, are without doubt a product, [35, p.
549; 36, p. 316; 37, p. 3538] as is ultimately any system equipped
with AI.
An essential question is whether the software that produces
this AI is in itself a product. Above all, this question becomes rel-
evant if the programme is distributed as such without being
bound to a data carrier, or if it is part of the system, but the ques-
tion of the software manufacturer's liability arises.
Art. 2 of the Product Liability Directive restricts the concept
of product to things thus defining it as a tangible object.
The programme as such is not a tangible object, [16, p. 46f.]
but, as the WIPO defined it in § 1 of the Model Law for the Pro-
tection of Computer Programs, a sequence of commands that can
be executed by a machine. The same definition is used in a deci-
sion by the German Federal Supreme Court [38, p. 1047]. Despite
this fact, there is unanimity in current German literature that
software is subject to product liability [39, no. 830; 40, no. 172;
41, p. 706; 42, p. 714f.], and according to the prevailing view,
even if it is delivered via the internet, for example by download
from a website [40, no. 173f.; 43, no. 43]. The same seems to ap-
ply to the Product Liability Directive in general [44, p. 245]. In
the upcoming revision of the Product Liability Directive, this re-
sult - the application to software - should be clarified.
Since AI is generated by computer software, product liability
therefore applies to any form of AI.
3.2 Producers
According to Art. 1 of the Product Liability Directive, the oppo-
nent of the claim for damages and thus the addressee of liability
is the producer of the product. According to Art. 3(1) of the
Product Liability Directive, the producer in this sense is both the
producer of the finished product and the producer of the compo-
nent parts or raw material.
The manufacturer of the overall system is the primary liabil-
ity addressee as the manufacturer of the finished product. Soft-
ware that is integrated into the system is to be qualified as a
component part, the supplier of this software is consequently
subject to liability as the manufacturer of the component part.
Insofar as (artificial) neural networks are a component of an
AI system, a differentiation must be made: The untrained neural
network does not have functional properties but acquires them
only through the process of machine learning. Considering this,
the untrained network is not a component part but a basic mate-
rial. Since the manufacturer of basic products is also included in
the liability, this differentiation as such is not decisive in the re-
sult. However, it points the way to the classification of the train-
ing. Insofar as the trained network is integrated into an AI sys-
tem as software, it is a component part, the manufacturer of the
trained network is consequently a component part manufacturer.
 ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
The status of being a component part producer is thus acquired
through the training of the network. This aspect can be general-
ised: whoever controls the learning process in the context of ma-
chine learning and thus determines the properties of the learning
system is the producer of the respective system.
3.3 Limitation of Damage to Health and Prop-
erty
Art. 9 of the Product Liability Directive [25] explicitly limits lia-
bility to damage to health and property. This is intended to ex-
clude, in particular, mere pecuniary damages, which occur, for
example, in the event of a business interruption.
There is thus a "liability gap" in relation to AI systems that
generate assessments or evaluations of persons. Whether it is an
assessment of the likelihood of recidivism, creditworthiness or
performance for a job or a place at university: any damage
caused by incorrect assessments is not covered by this concept.
Even if one were to assume a violation of personality rights of
the person concerned, these rights would not be subject to the
scope of product liability law.
This is not to say that product liability law should be ex-
tended to cover such damages. However, it is important to note
that these damages are currently not covered by product liability
law and therefore the legislator must be encouraged to achieve
sufficient protection against erroneous assessments by other
means.
4 Faulty Decisions and Product Liability
4.1 Conduct as a Defect of a Product?
A major challenge of product liability law in relation to AI is the
requirement of a defect of the product. The characteristic of an
AI system is that it reacts to situations independently, i.e., with-
out direct control by a human being, thus exhibiting an inde-
pendent and not previously determined behaviour [45, p. 43f.; 46,
p. 7; 47, p. 4]. However, the behaviour of a system in a particular
situation as such is not a property of the system. Even in cases
where the actor is human, it cannot be directly inferred that a
certain behaviour equals a certain characteristic.
The bridge between behaviour and property can be closed by
defining property as the ability to behave in a certain way in a
certain situation or not to show a certain behaviour. The prop-
erty of a highly automated vehicle is thus the suitability or abil-
ity to drive "correctly" in a certain situation.
Yet, this makes clear the error that exists in a simple equation
of behaviour and property. If one were to conclude a (product)
error, i.e., a certain (insufficient) property of this AI system, di-
rectly from a faulty behaviour, for example a driving error, the
required ability would be directed at behaving correctly in every
traffic situation.
This task can be very complex. For example, a traffic situation
consists of a multitude of facts which must be converted into in-
formation by the vehicle through sensors and then be inter-
preted correctly in several steps. Then the vehicle must derive a
correct decision on the driving strategy (trajectory planning)
35
G. Borges
based on the perceived situation, i.e., the interpreted data. In do-
ing so, all information available in the traffic situation must be
taken into account.
The decisive factor is that situations occurring in traffic are
unpredictably diverse. Accordingly, the concrete behaviour in
such a unique, future traffic situation is also unpredictable. If one
equates a faulty behaviour of the system with a product defect,
liability arises for behaviour in an unforeseeable multitude of fu-
ture traffic situations. This is a conceivable liability concept that
can be achieved in particular by specific legislation providing for
a causal liability of the producer, as some authors suggest in the
case of highly automated vehicles [48, p. 277; 49, p. 574].
However, this is not the liability concept of the Product Lia-
bility Directive. Product liability law focusses, as can be derived
from Art. 6 of the Product Liability Directive, on the existence of
properties of the product when it is placed on the market, prop-
erties, which must correspond to justified safety expectations.
The expectation of being able to control unforeseeable future sit-
uations is certainly not a justified safety expectation.
If the legislator wanted to introduce such liability, this could
be done e.g., by imposing causal liability on the producer, similar
to the operator's liability under § 7 StVG (German Road Traffic
Act). According to this norm, the owner of a motor vehicle is lia-
ble for accidents that are not due to force majeure. Another ex-
ample of such causal liability in German law is the liability of the
pet owner for so-called "luxury animals" (§ 833 BGB).
A similar result could be achieved by applying the construct
of vicarious liability to AI systems and equating them to humans
in this respect. This concept is not a part of tort law in Germany
or in most legal systems. German law contains such an attribu-
tion of fault in § 278 of the German Civil Code (BGB) for agents.
It is also argued that this provision should be applied analo-
gously to AI systems [50, p. 211f.]. This liability only applies to
contractual obligations and does not apply to non-contractual li-
ability.
As an interim result, this shows that product liability law
does not provide for liability for conduct, not even for the prop-
erty of behaving correctly in future, unforeseeable situations.
Therefore, only the ability to behave correctly in a spectrum of
foreseeable situations can be described as a property of an AI
system in the sense of product liability.
4.2 Reference Point of the Defectiveness of AI
Systems
With the interim result that product liability only refers to prop-
erties of the AI system, but not to its behaviour in a specific situ-
ation, follow-up questions necessarily arise.
The starting point is that the ability of the AI system to be-
have according to a certain expectation in a certain situation is
to be regarded as a property of the AI system and that the lack of
this ability can trigger product liability.
First of all, it is important to establish a reference point for
this ability, in particular to analyse whether and to what extent
human abilities are to be taken into account.
An intuitively obvious reference point would be the human
behaviour that is to be replaced by the activity of the AI system,
 AI Systems and Product Liability
in the case of automated driving, for example, the behaviour of a
human driver [48, p. 276; 51, p. 77]. For the determination of the
defectiveness of an AI system, reference would therefore have to
be made to human capabilities.
However, a simple reference to human capabilities as a yard-
stick for errors of AI systems would not be convincing. For ex-
ample, a human driver is granted a certain reaction time; the
ability to react immediately is not expected. The granting of a re-
action time to be measured against human limitations, however,
is obviously not necessary for AI systems; on the contrary, it
would be highly counterproductive. The improved safety of
highly automated vehicles, for example, is based not least on the
fast reaction of the system compared to the human driver [42, p.
734f.]. Therefore, an independent standard must be developed, as
is generally the case in product liability law [42, p. 733f.].
Admittedly, the reference to human capabilities is to some
extent predetermined. Insofar as AI systems are used to provide
"human performance", i.e., performance which was previously
provided by humans, the expected behaviour of the machine is
based on human behaviour. Thus, requirements for its behaviour
in road traffic are entirely oriented towards the abilities of a hu-
man being. Consequently, insofar as AI systems are used, a be-
haviour comparable to that of a human being is required [48, p.
275f.].
This orientation towards human behaviour leads to an orien-
tation towards human capabilities without being limited to them.
Accordingly, the safety expectations within the framework of
product liability law are also not to be limited by the abilities of
humans [42, p. 735f.].
This leads to a huge challenge for product liability law: the
requirements for the capabilities of AI systems are currently
very unclear, and clear, legally secure requirements have not yet
emerged in many areas.
Structurally, none of this is new. Requirements often have to
be concretised in the respective case, binding ultimately by
judges. It is therefore particularly relevant to define safety re-
quirements for new technologies in the form of technical stand-
ards.
4.3 Requirements for Accuracy of AI Systems
In the context of concretising the safety requirements, it is of
great importance whether the accuracy of the system with re-
gard to its behaviour can be expected as a property of an AI sys-
tem, or whether fault tolerance is permitted.
As already shown, an error-free behaviour for every future
situation cannot be part of the justified safety expectation, so
that a certain degree of error tolerance is required. In this regard,
the special features of AI systems compared to traditional, hu-
man-controlled machines become particularly clear.
Insofar as the required behaviour is specified, negligence is
assumed in the case of a human acting otherwise. If a human
driver does not drive correctly, there is fault. However, the pro-
ducer of the AI system is not its operator like a driver would be.
If one wanted to make the producer liable for driving errors of
the system itself, this would be a new type of causal liability,
which is rightly demanded (supra 4.1).
36
ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
Product liability is aimed at the justified expectations of
safety, i.e. the safety that can reasonably be achieved. This stand-
ard also applies to the behaviour of the system: what is required
is the degree of safety of correct behaviour that can reasonably
be ensured by the manufacturer. In this respect, the technical
possibilities at the time of placing the product on the market are
decisive.
To avoid misunderstandings, it should be said that liability
law must be distinguished from market approval rules which ap-
ply to numerous machines, such as vehicles, and that the re-
quirements are not identical. The obligation may be stricter than
in the approval procedure.
However, if the safety requirements are tied to what is rea-
sonable, exceptions can be considered; product liability law con-
tains a fault tolerance for the behaviour of AI systems.
The extent of this fault tolerance can ultimately only be clari-
fied in individual cases. However, one can ask the fundamental
question of whether it is tolerable for an AI system to behave
worse than the "average human" in a certain situation. In princi-
ple, this question can be answered in the affirmative, since AI
systems act completely differently from humans and show
strong weaknesses especially in unusual situations. A basic prob-
lem of machine learning, as it has been practised until recently,
is the lack of consideration dedicated to the confidence in the
classification. In many cases, doubts about the classification
(such as of an object perceived by sensors) were not or not suffi-
ciently taken into account for planning the behaviour of the AI
system
In addition, AI systems are highly susceptible to manipula-
tion, as has been demonstrated in numerous studies on IT secu-
rity. It is therefore inherent in the liability concept of product li-
ability law that machines make mistakes that would not be toler-
ated if made by a human.
This necessarily creates a liability gap in relation to the pro-
ducer. This gap must be filled outside of product liability law, for
example through specific causal liability on the part of the man-
ufacturer, or, as is currently the trend, through the liability of
other parties involved, such as the operator of the AI system.
However, product liability law offers another important con-
necting factor for the safety of AI systems. Part of the expecta-
tion of safety in a product is that the function of the product has
been properly verified. This general aspect is crucial in the case
of AI systems: since the behaviour of AI systems based on ma-
chine learning cannot be verified by code analysis but can at
least be partially verified by testing, the tests on the behaviour of
the AI system are a crucial aspect of the safety expectation.
Proper testing of the behaviour of AI systems thus becomes the
anchor of product liability: if there is no proper test, the product
is defective. If the damage can be traced back to a behaviour
which would have been detectable if proper tests had been car-
ried out, this error is also causal for the damage.
This shows a major advantage of AI systems compared to hu-
man activity: the behaviour of AI systems is in principle repro-
ducible and can therefore be checked by tests in many cases -
specifically: in complex situations with a correspondingly large
 ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
amount of different input information - not fully but much better
than with humans.
4.4 Burden of Proof
Another aspect of product liability that is often decisive in prac-
tice concerns the burden of proof. According to Art. 4 of the Eu-
ropean Product Liability Directive, the injured party bears the
burden of proof for the existence of the product defect, i.e. he
must present the defectiveness of the product and prove it in the
event of a dispute.
In this respect, the circumstances of the procedural law sys-
tems, which are subject to the law of the EU member states, have
a decisive effect. In the United States in particular, the instru-
ment of pretrial discovery provides the plaintiff with a strong in-
strument to strengthen the factual basis of his claim [52, p.
115f.], which is not known in most European legal systems.
In German law, the requirements for the presentation of the
facts on which a claim is based are particularly strict. In princi-
ple, the plaintiff is required to provide a complete statement of
the facts in such a way that a decision of the dispute can be
made based on the plaintiff's submissions [29, no. 92].
These requirements also apply in product liability, there also
with regard to the defect of the product. As a rule, the injured
party cannot provide such evidence, as he does not know any de-
tails about the design or manufacture of the product. Normally,
he will only be able to present circumstantial evidence from
which it may be possible to conclude that the product is defec-
tive.
In view of this procedural starting position, presumptions of
the existence of a defect, which lead to the reversal of the burden
of proof, play a central role. In fact, case law in Germany, for ex-
ample, has assumed a presumption in a number of cases: For ex-
ample, the lack of compliance with statutory safety regulations is
supposed to give rise to a presumption that products are defec-
tive.
Such presumptions are of crucial importance, especially for
AI systems. Without them, it will hardly be possible for the in-
jured party to prove a fault of the AI system, as he has no access
to the relevant information. He does not know the design of the
system, in machine learning cases he lacks any knowledge about
the learning process, also the injured party does not know any-
thing about the tests of the AI system [53, p. 285f.].
Against this background, the Export Group Liability for New
Technologies - New Technology Formation has recommended a
shift of the burden of proof in certain cases regarding AI and
other new technologies [8, p. 42]. The European Parliament has
taken up this idea in its resolution for a civil liability regime for
AI [9]. The proposal for a regulation on liability for AI systems
[10] contains a different distribution of risk for so-called "AI sys-
tems with high risk", for which strict liability of the operator is
to apply, and for "other AI systems" for which fault-based liabil-
ity is to remain. However, according to Art. 8 Para. 2 of the pro-
posed regulation, the operator must prove that personal injury
or damage to property was caused through no fault of his own.
This raises the crucial question of the conditions under which
a reversal of the burden of proof for the existence of a fault
37
G. Borges
should apply to AI systems. In this respect, great caution is re-
quired, as a premature presumption, which may not be rebutta-
ble, would transform product liability into a causal liability of the
manufacturer, which is precisely what the legislator wanted to
avoid.
In the cases of interest here, the starting point of the damage
is a faulty behaviour of the AI system. This behaviour appears
externally and is usually more recognisable for the injured party
than for the producer. In this respect, the assignment of the bur-
den of proof to the injured party is appropriate. Since the injured
party's lack of evidence concerns the facts relating to the design
and manufacture underlying the conduct, including the testing
of the AI system, the burden of proof should start at precisely
this point. Therefore, there should be a presumption of the exist-
ence of a defect in the AI system if it has behaved "incorrectly".
It is then incumbent on the manufacturer to prove that the fault
could not have been detected with proper design and testing.
This proof can be provided by the manufacturer presenting and,
if necessary, proving the existence of a sufficient test. Since he
has access to the relevant data and since he can use external help
if necessary, he can reasonably be expected to provide this evi-
dence.
With this easing of the burden of proof, there is probably an
appropriate balance of interests. It is often difficult to prove a
misconduct. In the example of the highly automated vehicle - in
the absence of causal liability of the operator - a driving error
would have to be proven, which can be difficult if the course of
the accident cannot be determined. In this respect, however, the
burden of proof on the injured party is correct, since the manu-
facturer has no better information in this respect. In this respect,
however, the black box, which is to become obligatory for highly
and fully automated vehicles, should provide a remedy, for ex-
ample in the case of automated driving. The inspection of the
recorded data can be ordered by the court, e.g., within the frame-
work of independent evidence proceedings prior to the filing of a
lawsuit.
5 Conclusion
As an overall result, product liability law can be said to have a
certain performance capacity also in relation to AI systems.
Product liability is comprehensively applicable to AI systems
and their components, in particular to neural networks and any
other software that creates AI. The manufacturers of the system
as well as of the software as a partial product are the addressees
of product liability law and can be held liable for damages.
A significant gap arises in the case of AI systems that cannot
directly lead to personal injury or damage to property, but which
set an essential prerequisite for decisions by third parties, as is
the case with AI systems for the assessment of persons.
The central challenge of product liability law in relation to AI
systems is the determination of a product defect. Since the con-
cept of defect is based on properties of the product that must be
present when the product is placed on the market, the behaviour
of the AI system cannot be directly linked to the defect. How-
ever, one property of the AI system is the ability to behave in a
 AI Systems and Product Liability
certain way. The more the behaviour depends on a situation and
the more diverse the possible situations are - in the case of auto-
mated driving, for example, represented by the unforeseeable
multitude of traffic situations - the less the behaviour of the AI
system can be anticipated by the manufacturer and secured in
advance. The justified expectation of safety as a measure of
faultiness is therefore not directed towards fault-free behaviour
in all situations, but only towards such behaviour in situations
which can be reasonably foreseen.
The sufficient degree of safety is, as described, primarily con-
cretised by the reasonableness of the measures. Technical stand-
ards will play a special role in this, as they emerge.
In the case of AI systems, an important starting point for
product liability will probably be the testing of AI systems, since
the existence of proper tests of AI systems is to be regarded as a
component of the justified expectation of safety. If this is lack-
ing, the AI system is defective.
In practice, the burden of proof will play a central role. In this
respect, a reversal of the burden of proof for the existence of a
fault is to be assumed for AI systems if the system behaves incor-
rectly. The burden of proof for the faulty behaviour lies with the
injured party.
With these requirements, product liability law can contribute
to an appropriate balancing of interests with regard to damage
caused by AI systems. Nevertheless, the legal framework of AI
systems needs to be supplemented, as otherwise liability risks
and false incentives exist.
REFERENCES
[1] Virginia Foggo, John Villasenor, Pratyush Garg. 2021. Algorithms and Fairness,
Ohio State Technology Law Journal 17(1), 123–188.
[2] Federal Trade Commission (FTC). 2021. Aiming for truth, fairness, and equity
in your company’s use of AI, retrieved from https://www.ftc.gov/news-
events/blogs/business-blog/2021/04/aiming-truth-fairness-equity-your-compa-
nys-use-ai (accessed 17.5.2021).
[3] Communication from the Commission to the European Parliament, the Euro-
pean Council, the Council, the European Economic and Social Committee and
the Committee of the Regions, Artificial Intelligence for Europe, COM(2018) 237
final, 25.4.2018, https://eur-lex.europa.eu/legal-con-
tent/EN/TXT/PDF/?uri=CELEX:52018DC0237 (accessed 17.5.2021).
[4] For detailed information on the group see https://ec.europa.eu/transparency/re-
gexpert/index.cfm?do=groupDetail.groupDetail&groupID=3591&news=1 (ac-
cessed 17.5.2021).
[5] High-Level Expert Group on AI, Ethics Guidelines for Trustworthy Artificial
Intelligence, retrieved form https://digital-strategy.ec.europa.eu/en/library/eth-
ics-guidelines-trustworthy-ai (accessed 17.5.2021).
[6] For detailed information on the group see https://ec.europa.eu/transparency/re-
gexpert/index.cfm?do=groupDetail.groupDetail&groupID=359 (accessed
17.5.2021).
[7] Proposal for a Regulation laying down harmonized rules on artificial intelli-
gence (Artificial Intelligence Act), 21.4.2021, https://ec.europa.eu/news-
room/dae/redirection/document/75788 (accessed 17.5.2021).
[8] Report from the Expert Group on Liability and New Technologies – New Tech-
nologies Formation, Executive Summary, 3 f.; https://ec.europa.eu/transpar-
ency/regexpert/index.cfm?do=groupDetail.groupMeetingDoc&docid=36608
(accessed 17.5.2021).
[9] Civil liability regime for artificial intelligence, European Parliament resolution
of 20 October 2020 with recommendations to the Commission on a civil liability
regime for artificial intelligence (2020/2014(INL)), retrieved from
https://www.europarl.europa.eu/doceo/document/TA-9-2020-0276_EN.pdf (ac-
cessed 17.5.2021).
[10] Proposal for a Regulation of the European Parliament and of the Council on
liability for the operation of Artificial Intelligence-systems, (2020/2014(INL)),
retrieved from https://www.europarl.europa.eu/doceo/document/TA-9-2020-
0276_EN.pdf (accessed 17.5.2021).
[11] ExamAI – Testing and Auditing of AI systems; retrieved from https://testing-
ai.gi.de (accessed 17.5.2021).
38
ICAIL 2021, 21-25 June 2021, São Paulo, Brazil
[12] Rasmus Adler, Jens Heidrich, Lisa Jöckel, Michael Kläs. 2020. Anwendungssze-
narien: KI-Systeme in der Produktionsautomatisierung; retrieved from
https://testing-ai.gi.de/fileadmin/GI/Projekte/KI_Testing_Auditing/Exa-
mAI_Publikation_Anwendungsszenarien_KI_Industrie.pdf (accessed
17.5.2021).
[13] National Transportation Safety Board (NTSB). 2018. Collision Between Vehicle
Controlled by Developmental Automated Driving System and Pedestrian
Tempe, Arizona, March 18, 2018 accident report, retrieved from
https://www.ntsb.gov/investigations/AccidentReports/Reports/HAR1903.pdf
(accessed 17.5.2021).
[14] Julia Angwin, Jeff Larson, Surya Mattu and Lauren Kirchner. 2016. Machine
Bias, ProPublica, May 23, 2016, retrieved from <https://www.propublica.org/ar-
ticle/machine-bias-risk-assessments-in-criminal-sentencing> (accessed
17.5.2021).
[15] Anthony W. Flores, Kristin Bechtel, Christopher T. Lowenkamp. 2016. False
Positives, False Negatives and False Anaylses: A Rejoinder to “Machine Bias:
There’s Software Used Across the Country to Predict Future Criminals. And It’s
Biased Against Blacks.”, Federal Probation 80(2), 38–46.
[16] Katharina Zweig, Marc Hauer, Franziska Raudonat. 2020. Anwendungsszenar-
ien: KI-Systeme im Personal- und Talentmanagement, retrieved from
https://testing-ai.gi.de/fileadmin/PR/Testing-AI/ExamAI_Publikation_Anwen-
dungsszenarien_KI_HR.pdf (accessed 17.5.2021).
[17] Steven Shavell. 2004. Foundations of Economic Analysis of Law. Cambridge, MA:
Belknap Press.
[18] Hans-Bernd Schäfer, Claus Ott. 2012. Lehrbuch der ökonomischen Analyse des
Zivilrechts, Springer.
[19] Petra Buck-Heeb, Andreas Dieckmann. 2020. in: Oppermann/Stender-Vor-
wachs (eds.), Autonomes Fahren, C.H. Beck, 2nd ed., Chap. 3.1.1.
[20] Peter Kreutz. 2020. in: Oppermann/Stender-Vorwachs (eds.), Autonomes Fahren,
C.H. Beck , 2nd ed., Ch. 3.1.2.
[21] Geraint Howells. 1993. Comparative Product Liability, Dartmouth Publishing.
[22] Greenman v. Yuba Power Products, Inc. (1963), 59 Cal. 2d 57.
[23] Fairgrieve, Howells, Møgelvang-Hansen, Straetmans, Verhoevens, Machnikow-
ski, Janssen, Schulze. 2016. Product Liability Directive. In Pjotr Machnikowski
(Ed.), European Product Liability. An Analysis of the State of the Art in the Era of
New Technologies, 17–111.
[24] Matthias Reimann, M. 2015. Product liability. In M. Bussani & A. Sebok (eds.),
Comparative Tort Law, 250 – 279, Edward Elgar Publishing.
[25] Council Directive of 25 July 1985 on the approximation of the laws, regulations
and administrative provisions of the Member States concerning liability for de-
fective products (85/374/EEC), O.J. 1985 No. L 210/9.
[26] Astrid Seehafer, Joel Kohler. 2020. Künstliche Intelligenz: Updates für das Pro-
dukthaftungsrecht? Mögliche Anpassungen der europäischen Produkthaf-
tungsrichtlinie für neue Technologien, EuZW 2020, 213–218.
[27] Thomas Riehm. 2010. 25 Jahre Produkthaftungsrichtlinie – Ein Lehrstück zur
Vollharmonisierung, EuZW 2010, 567–571.
[28] Georg Borges. 2018. Rechtliche Rahmenbedingungen für autonome Systeme,
NJW 2018, 977–982.
[29] Georg Borges. 2021. in: Borges/Hilber (Eds.), Beck’scher Online-Kommentar IT-
Recht, § 1 ProdHaftG.
[30] Gerhard Wagner. 2020. in: Münchener Kommentar zum BGB, 8th ed., § 4 Pro-
dHaftG.
[31] Georg Borges. 2020. Kann ein Gegenstand nicht Sache und doch Sache sein?.
Computerprogramme im Privatrecht, in: Omlor (ed.), Weltbürgerliches Recht,
Festschrift für Michael Martinek zum 70. Geburtstag, 45–58, C.H. Beck.
[32] OLG Hamm NJW-RR 2012, 355.
[33] Erwin Deutsch. 1989. Der Schutzbereich der Produzentenhaftung nach dem
BGB und dem PHG, JZ 1989, 465–470.
[34] Gerhard Wagner. 2020. in: Münchener Kommentar zum BGB, 8th ed., § 1 Pro-
dHaftG.
[35] Ulrich Berz, Eva Dedy, Claudia Granich. 2000. Haftungsfragen bei dem Einsatz
von Telematik-Systemen im Straßenverkehr, DAR 2000, 545–554.
[36] Volker Jänich, Paul Schrader, Vivian Reck. 2015. Rechtsprobleme des autono-
men Fahrens, NZV 2015, 313–318.
[37] Paul Schrader. 2015. Haftungsrechtlicher Begriff des Fahrzeugführers bei
zunehmender Automatisierung von Kraftfahrzeugen, NJW 2015, 3537–3542.
[38] Bundesgerichtshof [German Federal Supreme Court], GRUR 1985, 1041.
[39] Helmut Redeker. 2017. IT-Recht, 6th ed., C.H.Beck.
[40] Philipp Reusch. 2020. In Kaulartz/Braegelmann (eds.), Rechtshandbuch Artificial
Intelligence und Machine Learning, Chap. 4.1, C.H.Beck.
[41] Friedrich-Wilhelm Engel. 1986. Produzentenhaftung für Software, CR 1986,
702–708.
[42] Gerhard Wagner. 2017. Produkthaftung für autonome Systeme, AcP 2017(6),
707–765.
[43] Georg Borges. 2021. in: Borges/Hilber (eds.), Beck’scher Online-Kommentar IT-
Recht, § 2 ProdHaftG.
[44] Ulrich Magnus. 2017. In Machnikowski (ed.), European Product Liability. In-
tersentia.
 ICAIL 2021, 21-25 June 2021, São Paulo, Brazil G. Borges

[45] Thomas Schulz. 2015. Verantwortlichkeit bei autonom agierenden Systemen. No-
mos.
[46] Hans Steege. 2021. Auswirkungen von künstlicher Intelligenz auf die Produzen-
tenhaftung in Verkehr und Mobilität. Zum Thema des Plenarvortrags auf dem
59. Deutschen Verkehrsgerichtstag, NZV 2021, 6–13.
[47] Chris Reed, Elizabeth Kennedy, Sara Silva. 2016. Responsibility, Autonomy and
Accountability: Legal Liability for Machine Learning, Queen Mary University of
London Legal Studies Research Paper No. 243/2016.
[48] Georg Borges. 2016. Haftung für selbstfahrende Autos. Warum eine Kausalhaf-
tung für selbstfahrende Autos gesetzlich geregelt werden sollte, CR 2016, 272–
280.
[49] Sabine Gless, Ruth Janal. 2016. Hochautomatisiertes und autonomes Autofahren
– Risiko und rechtliche Verantwortung, JR 2016, 561–575.,
[50] Herbert Zech. 2019. Künstliche Intelligenz und Haftungsfragen, ZfPW 2019,
198–219.
[51] Christian Gomille. 2016. Herstellerhaftung für automatisierte Fahrzeuge, JZ
2016, 76–82.
[52] Geoffrey Hazard, Michele Taruffo.1993. American Civil Procedure. Yale Univer-
sity Press.
[53] Mario Martini. 2019. Blackbox Algorithmus – Grundfragen einer Regulierung
Künstlicher Intelligenz. Springer.

39
 A Combined Rule-Based and Machine Learning Approach for
Automated GDPR Compliance Checking
Rajaa EL HAMDANI Majd Mustapha
HEC Paris EURA NOVA
France Belgium
Aurore Troussel Sébastien Meeùs
Steptoe & Johnson LLP HEC Paris
Belgium France
ABSTRACT
The General Data Protection Regulation (GDPR) requires data con-
trollers to implement end-to-end compliance. Controllers must
therefore ensure that the terms agreed with the data subject and
their own obligations under GDPR are respected in the data flows
from data subject to controllers, processors and sub processors (i.e.
data supply chain). This paper seeks to contribute to bridge both
ends of compliance checking through a two-pronged study. First,
we conceptualize a framework to implement a document-centric
approach to compliance checking in the data supply chain. Second,
we develop specific methods to automate compliance checking of
privacy policies. We test a two-modules system, where the first
module relies on NLP to extract data practices from privacy poli-
cies. The second module encodes GDPR rules to check the presence
of mandatory information. The results show that the text-to-text
approach outperforms local classifiers and enables the extraction
of both coarse-grained and fine-grained information with only one
model. We implement full evaluation of our system on a dataset of
30 privacy policies annotated by legal experts. We conclude that
this approach could be generalized to other documents in the data
supply as a means to improve end-to-end compliance.
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Information extraction; Machine learning; • Social and pro-
fessional topics → Privacy policies; • Security and privacy →
Usability in security and privacy.
ACM Reference Format:
Rajaa EL HAMDANI, Majd Mustapha, David Restrepo Amariles, Aurore
Troussel, Sébastien Meeùs, and Katsiaryna Krasnashchok. 2021. A Com-
bined Rule-Based and Machine Learning Approach for Automated GDPR
Compliance Checking. In Eighteenth International Conference for Artificial
Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM,
New York, NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466081
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466081
40
David Restrepo Amariles
HEC Paris
France
Katsiaryna Krasnashchok
EURA NOVA
Belgium
1 INTRODUCTION
It has become widely acknowledged that complying with data pro-
tection laws, particularly the European General Data Protection
Regulation (GDPR), is the most difficult compliance challenge or-
ganizations face today across industries [1]. Moreover, technology
became the most important compliance cost for organizations, as
they turned to specialized technologies to carry out compliance
tasks such as document review, regulatory checks and operational
audits [1]. Despite the increasing interest in technology as a com-
pliance tool in data protection, there is not yet a comprehensive
conceptual framework characterizing the tasks required to verify
compliance in the entire data supply chain (i.e., end-to-end compli-
ance), and how current methods can contribute to address them.
This paper intends to contribute to both of these issues. First, we
lay down a framework to implement and monitor GDPR compliance
in the data supply chain through a document-centric approach.
We define three key tasks of compliance checking based on the
function and content of the document: (1) document to regulation,
(2) document to document, and (3) document to operations. We
apply this framework to analyze the compliance function of privacy
policies in the data supply chain and define the tasks required to
determine if a privacy policy is compliant with the GDPR.
Second, we develop and test several methods to verify compli-
ance of privacy policies to the GDPR by leveraging the advantages
of both machine learning and rule-based approaches. In particu-
lar, we build a two-modules system to verify the completeness of
privacy policies with regards to mandatory information. The first
module automatically extracts coarse-grained and fine-grained data
practices, while the second module analyzes extracted data prac-
tices and checks the presence of mandatory information according
to the provisions of the GDPR. We make use of the OPP-115 dataset
[48] for training and evaluation of our models. We treat the extrac-
tion of data practices as a Hierarchical Multi-label Classification
(HMTC) task and experiment with two different approaches: local
classifiers and text-to-text. Our proposed text-to-text method has
several advantages over local classifiers, including extraction of
additional information and better scalability.
Our contributions are the following:
• we present a theoretical framework to implement automated
GDPR compliance in the data supply chain;
• we provide formal and substantive approach to verify com-
pliance of privacy policies with the GDPR;
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
• we develop a combined rule-based and machine learning
approach to experiment with automated formal compliance
checking of privacy policies with the GDPR;
• we propose a text-to-text approach for HMTC using transfer
learning and multi-task learning ;
• additionally, we extract the span of text from a privacy pol-
icy corresponding to the fine-grained data practices, which
provides better explainability of the classification results;
• we further annotate a dataset of 30 privacy policies with
the presence of mandatory information, in order to fully
evaluate our two-modules system of compliance checking;
• finally, we release a public repository of the dataset, imple-
mentations and the fine-tuned T5-11B model on OPP-1151 .
This paper is organized as follows. Section 2 provides an overview
of the related work. Section 3 lays down a general compliance
framework to automate GDPR compliance and applies it to privacy
policies. Section 4 describes the development of data practices ex-
traction algorithms and its evaluation results. Section 5 presents
the rule-based system, which verifies the presence of mandatory
information, and its evaluation on an annotated dataset.
2 RELATED WORK
The Information Commissioner’s Office (ICO), the Data Protection
Authority of the United Kingdom, observes that overall compli-
ance with the GDPR is jeopardized by the "opaque nature" of the
data supply chain, which is poorly documented, and generally fails
to comply with the GDPR’s accountability principle [2]. The ICO
further points out that controllers and processors should not limit
compliance to entering into contract and producing legal docu-
ments. They must monitor processing activities, and conduct audit
to ensure that appropriate technical and organizational measures
are in place throughout the data supply chain [2].
Most research in AI and law studying GDPR compliance has
focused on the relations between data subject and data controllers,
rather than on the compliance challenges in the data supply chain
[3, 6, 11, 54]. More recently, the BPR4GDPR project started working
on a compliance ontology specification that supports end-to-end
compliance [37] and which can contribute to address some of the
operational challenges raised by the ICO. A compliance framework
is provided in [40] for specific documents in the supply chain such
as the data protection impact assessment (DPIA).
2.1 Model-based Compliance Checking
The AI and law research community has developed model-based
methods for automated compliance checking, such as legal ontolo-
gies that support legal reasoning via logic programming [9, 12, 13,
15, 16, 19, 30, 44, 45]. However, even though most legal rules can
be described in logic programming, these methods face two main
challenges when applied to real-life cases:
• Knowledge acquisition bottleneck [17]: Logic programming
requires the encoding of facts into predicate form, but such
an encoding would be very cumbersome for each data pro-
tection documents in the data supply chain.
1 https://github.com/smartlawhub/Automated-GDPR-Compliance-Checking
41
EL HAMDANI, et al.
• Open texture of legal language [46]: Privacy concepts can be
quite abstract, and their evaluation arduous, as it is impossi-
ble to define a finite set of rules for all possible applications.
For example, the storage limitation principle states that per-
sonal data must not be kept "longer than is necessary for
the purposes for which the personal data are processed."(Art.
5.1(e)). This principle does not define specific time limits that
can be evaluated easily.
To solve the aforementioned problems, we employ natural lan-
guage processing (NLP) techniques to automatically extract infor-
mation from the data supply chain documents. Likewise, authors of
[44] suggest using NLP to extract information from legal documents
according to their UML model of the GDPR, to automatically con-
struct their model-based representation. A large and growing body
of literature uses machine learning and NLP algorithms to extract
privacy information from legal documents, however, it only con-
siders privacy policies and does not yet deal with other documents
of the data supply chain, such as DPA and DPIA.
2.2 Information extraction from privacy
policies
Privacy policies are long documents that are difficult to read for
data subjects. Empirical studies have been conducted to study and
measure the ambiguity and vagueness of privacy policies [5, 20, 24,
36]. Efforts have been made to decrease these deficiencies [42, 51],
by using classification techniques to extract data practices from
the text and represent them in a user-friendly interface. Other
approaches focus on the extraction of one specific type of data
practice, such as "opt-out choice" in [38].
More and more methods are developed to analyze the compliance
of privacy policies with the GDPR. A variety of NLP tools such as
word embeddings are used in [7, 43] to verify the completeness of
privacy policies according to the rules set out by the GDPR. The
CLAUDETTE project [6] extracts clauses that are problematic with
respect to the GDPR. In a different project [23], privacy policies were
analyzed in a large-scale setting to study the effect of the GDPR on
their provisions. For instance, the comparison between pre-GDPR
and post-GDPR versions of 6,278 English privacy policies showed
that the GDPR caused textual changes of the privacy policies, such
that their appearance improved, their length increased, and that
they cover more data practice categories.
A great deal of previous research on privacy policies relies on
supervised machine learning methods, which require datasets of
annotated privacy policies. However, there are very few such pub-
licly released datasets. In our work, we use OPP-115 [48] – a corpus
of 115 privacy policies annotated with both coarse-grained and
fine-grained data practices. Several works used OPP-115 to train
machine learning algorithms on the task of extracting data prac-
tices [14, 25, 28, 29, 32]. PrivacyQA [35] is another publicly available
dataset of 1,750 questions about the privacy policies of mobile ap-
plications, which is used to train question-answering algorithms.
A serious limitation of the publicly available datasets is that their
annotation schemes contain few concepts aligned with the GDPR.
A connection between the data practices identified in OPP-115 and
the GDPR was presented in [31], which revealed that the principle
of accountability of Article 5 is absent from OPP-115 concepts.
A Combined Rule-Based and Machine Learning Approach for Automated GDPR Compliance Checking
2.3 Transfer learning in NLP
Until very recently, in computer science research, general NLP
tasks such as text classification were commonly handled with ar-
chitectures based on word embeddings [27], Convolutional Neural
Networks (CNN) [21] and Recurrent Neural Networks (RNN) [18].
RNN-based solutions, which achieved state-of-the-art results at the
time, are, however, limited when it comes to dealing with long text,
due to their sequential nature. Moreover, this nature stands in the
way of most parallel training methods, therefore limiting the ability
to decrease training time. Encoder-decoder architectures – com-
monly used for sequence to sequence problems – have inspired the
attention-based transformers [47], and these architectures managed
to go beyond the limitations of sequential networks.
The use of transformers is gradually becoming the state of the
art of NLP. Indeed, they are effective in the variety of tasks, includ-
ing masked token prediction, next sentence prediction, question
answering, machine translation, summarization, sentiment analysis
and classification [10, 33]. The fact that transformers are trained in
an unsupervised way, which reduces the reliance on labeled data
and allows the use of a larger pool of text, explains the increased
performance. Transformers are very effective in transfer learn-
ing, allowing researchers to pretrain them with large amounts of
general-purpose texts and then to fine-tune them for their specific
tasks with good results, less effort, and less labeled data. Several
works explored the potential of fine-tuning transformers on OPP-
115 to extract coarse-grained data practices [28, 29].
Similarly to [14], we extract both coarse-grained and fine-grained
data practices from privacy policies. And similarly to [28, 29], we
use variations of the transformer architecture as base for fine-tuning
on OPP-115. Furthermore, our best-performing model is able to
extract the fine-grained data practices with corresponding spans of
text from the policy, thus improving explainability of the results.
3 A FRAMEWORK FOR COMPLIANCE
CHECKING IN THE DATA SUPPLY CHAIN
New data protection and privacy regulations around the world have
empowered data subjects vis-à-vis data controllers (EU fundamental
rights, California Privacy laws, Canada PIPEDA, Brazil, etc.). As
mentioned earlier, the data supply chain (i.e., the way data flows
from data subject to controllers, processors and sub processors)
is opaque and remains hard to monitor, ultimately hindering the
effectiveness of data-subject rights.
At the operation level, the data supply chain is characterized by
the high volume of data flows across processors and jurisdictions
for tasks as varied as storage, pre-processing, producing analyt-
ics, implementing AI-methods, generating visualizations, etc. In
addition to regulations, these flows are regulated by fragmented
legal artifacts such as contracts, data protection addendum (DPA),
technical and operational measures (TOMs), etc., which are often
of restricted access, lack interoperability and have low operational
efficiency, ultimately obstructing compliance with the GDPR.
To help data controllers use technological tools to improve com-
pliance with the GDPR, we draw a general conceptual framework
to automate compliance checking in the data supply chain based
on a document-centric approach and lay down the tasks required
to achieve end-to-end compliance.
42
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
3.1 Compliance Checking Framework: A
Document-centric Approach
The starting point of our compliance framework is the set of doc-
uments establishing the conditions and processing activities that
data controllers and processors are intended to and effectively carry
out (hereafter compliance documents), such as privacy policies, con-
tracts, DPAs, TOMs, DPIAs, etc. This is a reasonable starting point
as the GDPR establishes a document-centric approach to compli-
ance, entrusting documents with different functions in relation
to the obligations of data controllers and processors. First, it re-
quires data controllers to properly demonstrate compliance with
the Regulation ("principle of accountability", Art. 5.2). Second, data
controllers shall inform data subjects about the processing activities
they carry out (Art. 12, 13, and 14) and document instructions they
give to data processors (Art. 29). Third, sub-processors must obtain
a written authorization from controllers to further subcontract a
processing activity, and must also document their instructions (Art.
28.2 and 28.3). Lastly, controllers and processors must maintain a
record of processing activities (ROPA) (Art. 30).
Hence, we define GDPR compliance checking as the assessment
of the provisions of a compliance document in relation to: (1) a
regulation, (2) another document in the supply chain, and (3) an op-
eration. These dimensions define the three key tasks of compliance
checking in the data supply chain under the GDPR.
(1) Document to Regulation Compliance: The provisions of a
compliance document are assessed against the regulation.
For example, according to the GDPR, privacy policies must
provide certain mandatory information to data subjects (Art.
12, 13 and 14). We further divide this compliance task into
two sub-tasks:
• Formal Compliance Checking: whether the document ful-
fils its informative function, i.e., is the mandatory informa-
tion included in the document. For example, this sub-task
could consist of the verification of the existence of infor-
mation about data retention (Art. 13.2(a)).
• Substantive Compliance Checking: whether the document
fulfils its accountability function, i.e., does the information
provided comply with the GDPR. For example, this sub-
task could consist of the verification that the data retention
period is lawful, i.e., it does not exceed the necessary time
as required by Article 5.1(e).
(2) Document Chain Compliance: The supply chain documents
are assessed against the main contractual standards or against
documents with a higher hierarchy in the chain. Using the
previous example on data retention, two assessments could
be done on (i) the contract between the data controllers and
the data processors to verify if the agreed data retention
period corresponds to the one provided in the privacy pol-
icy, and on (ii) the ROPAs of both the data controllers and
processors that contain the effective date of erasure (Art.
30.1).
(3) Operational Compliance: This task consists of assessing the
adequacy between operations, documents, and regulations.
For the previous example, this task would imply verifying if
the data has been effectively deleted from the servers of data
processors and controllers when the retention period ends.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
3.2 The compliance of privacy policies
Privacy policies are the compliance documents that appear at the
top of the data supply chain. Hence, we first apply our framework
to analyze privacy policies’ compliance in the data supply chain.
This paper seeks to define the tasks required to determine a privacy
policy’s formal and substantive compliance with the GDPR.
The GDPR does not explicitly mention privacy policies, but data
controllers widely use them. Their main function is to provide
mandatory information to data subjects according to Articles 13
and 14. The absence of any part of this information renders the
privacy policy non-compliant. Moreover, the GDPR requires pri-
vacy policies to use plain and clear language so individuals can
understand how their personal data are processed, provide their
consent and exercise their rights. Consequently, the tasks to verify
the formal compliance of privacy policies are:
• Check the presence of each mandatory information accord-
ing to Articles 13 and 14.
• Check the readability and clarity of the language used in the
privacy policy.
The substantive compliance checking of privacy policies consist of
verifying that the data processing complies with the data protection
rules (e.g, fair and transparent processing of Art. 5, and lawfulness
of the processing of Art. 6). For example, a privacy policy must
specify the legal basis of the processing according to Article 13.1(c),
but it must also demonstrate that this legal basis complies with
Article 6 requirements on the lawfulness of the processing (e.g., if
the legal basis is consent, the controller must ensure that consent
has been given for one or more specific purposes.).
In this study, we automate the first task of formal compliance
checking of privacy policies. In the following sections, we describe
how we combined rules and machine learning to check the presence
of mandatory information in privacy policies automatically. The
end-users of such a system would be lawyers or data protection
officers who review large numbers of privacy policies to check their
compliance with the GDPR. Another type of end-user would be
project managers in small companies who lack the legal knowledge
to ensure privacy policies’ compliance.
4 EXTRACTION OF DATA PRACTICES FROM
PRIVACY POLICIES
Ensuring both the compliance of documents and data processing ac-
tivities is becoming more burdensome to companies due to several
challenges. We focus on the challenge posed by the large number
of documents that data protection officers need to review to guar-
antee compliance. We suggest using natural language processing
technologies to assist data protection officers in performing the
compliance checking tasks. NLP could help to extract compliance
information from unstructured compliance documents and save it
into structured formats such as XML or RDF to unlock use cases
such as automated compliance checking.
In this paper, we describe our experiment in automating formal
compliance of privacy policies. We first train a machine learning
algorithm to extract from privacy policies information describing
the company’s data practices. We then use the extracted information
as input to a rule-based system that encode Articles 13 and 14.
43
EL HAMDANI, et al.
4.1 OPP-115 : Training Dataset of Online
Privacy Policies
For our task we make use of the Usable Privacy Policy Project’s
Online Privacy Policies (OPP-115) corpus, introduced by [48] which
contains detailed annotations made by Subject Matter Experts
(SMEs) for the data practices described in a set of 115 website
privacy policies.
At a high level, annotations fall into one of ten data practice
categories:
(1) 1𝑠𝑡 Party Collection/Use: What, why and how information
is collected by the service provider.
(2) 3𝑟𝑑 Party Sharing/Collection: What, why and how informa-
tion shared with or collected by third parties.
(3) User Choice/Control: Control options available to users.
(4) User Access, Edit, & Deletion: If/how users can access, edit
or delete information.
(5) Data Retention: How long user information will be stored.
(6) Data Security: Protection measures for user information.
(7) Policy Change: Informing users if policy information has
been changed.
(8) Do Not Track: If and how DNT signals for online tracking
and advertising are honored.
(9) International & Specific Audiences: Practices pertaining to a
specific group of users.
(10) Other: General text, contact information or practices not
covered by the other categories.
According to the dataset creators, the best agreement between SMEs
was achieved on Do Not Track class with Fleiss’ Kappa equal to 91%,
whereas the most controversial class was Other, with only 49% of
agreement [48]. We further decompose the latter category into its
attributes – “Introductory/Generic”, ”Privacy Contact Information”
and “Practice Not Covered” – resulting in 12 categories.
Figure 1 depicts a fragment of OPP-115 taxonomy: for each class
(grey shaded blocks), a set of lower-level privacy attributes is as-
signed (20 in total, dark blue shaded blocks), with specific values
corresponding to each attribute. For example, the attribute “Per-
sonal Information Type” designates the different types of personal
information mentioned in the text, as can be seen from the annota-
tions in Figure 2 from the IMDb policy 2 , annotated with “1𝑠𝑡 Party
Collection/Use” category.
OPP-115 comprises 3,792 segments, each segment labeled with
one or more classes out of 12. The SMEs produced a total of 23K
annotations of categories. In aggregate, these categories were asso-
ciated with 128K values for attributes and 103K selected spans of
policy text. To the extent of our knowledge, this is the first effort to
leverage these spans to extract information from privacy policies.
We split the OPP-115 dataset on a policy-document level into
3 sets: 65 policies are used for training, 35 for validation and 30
policies are kept as a testing set.
4.2 Problem formulation
The taxonomy of data practices is organized in a class hierarchy that
we model as a Directed Acyclic Graph (DAG) shown in Figure 3.
2 To retrieve the exact source used:<https://web.archive.org/web/20200526092253if_
/https://www.imdb.com/privacy#auto>(“Automatic Information” sub-section.)
A Combined Rule-Based and Machine Learning Approach for Automated GDPR Compliance Checking
Figure 1: The privacy taxonomy of [48]. The top level of the hierarchy (grey shaded blocks) defines coarse-grained data practices or privacy
categories. The lower level defines a set of privacy attributes (blue shaded blocks), each assuming a set of values. We show a subset of the
taxonomy for clarity and space considerations.
Figure 2: Annotated excerpt from IMDb privacy notice
We choose a DAG structure instead of a tree structure because
some attributes are associated with more than one category, i.e.,
have more than one parent. For example, the attribute “Personal
information type” belongs to both “1𝑠𝑡 Party Collection/Use” and
“3𝑟𝑑 Party Collection/Use” categories.
The training dataset is a corpus C of N privacy policies 𝑃𝑖 : Y =
{𝑃1, ..., 𝑃 𝑁 }. Each privacy policy 𝑃𝑖 is a set of annotated segments:
𝑃𝑖 = {(𝑠𝑖1, 𝑦𝑖1 ), ..., (𝑠𝑖𝑛 , 𝑦𝑖𝑛 )} where 𝑦𝑖 ⊆ Y = {𝜆1, ..., 𝜆𝐿 } such that
𝜆𝑖 is a path in the DAG that starts from category 𝑐 and ends at leaf
node 𝑣.
Figure 3: The DAG structure of the OPP-115 taxonomy
44
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
We treat predicting the categories of data practices and the val-
ues of each attribute as an HMTC task. There are three methods
to solve hierarchical text classification tasks: flat, local, or global
methods [39]. The flat method behaves like traditional classifica-
tion algorithms by ignoring the labels’ hierarchy and predicting
only classes at the leaf nodes. Local methods take into account the
hierarchy by training independent local classifiers. Global methods
train a single classifier for all classes. In this paper we conduct two
experiments, where we first build a local multi-label classifier, and
then cast the HMTC task to two text-to-text tasks.
4.3 Local classifiers approach
This approach is inspired by Polisis [14], where authors build a local
multi-label classifier for the higher level categories, and one local
multi-label classifier per attribute to predict their values. Predictions
are made in a top-down order: once the categories of a segment are
inferred, the second step predicts the values of attributes – children
of the predicted categories. For example, if the first-level classifier
predicts the “Data Retention” and “Data Security” categories, only
the local classifiers corresponding to the attributes “Retention Pe-
riod”, “Retention Purpose”, “Personal Information Type”, “Security
Measure” are chosen in the second step.
Authors of [14] are using the same base classifier for all the multi-
label classifiers. In this paper we reproduce their work by using
CNN as the base classifier. We use the same architecture of CNN
and hyperparameters. The CNN classifier is composed of one convo-
lutional layer with a ReLU activation, followed by a dense layer and
a ReLU activation. The last layer is a dense layer with a sigmoid ac-
tivation. We tokenize segments using PENN Treebank tokenization
in NLTK [41]. Tokens are mapped into a 300-dimensional space via
an embedding Layer. We used FastText to train Word embeddings
on 130,326 privacy policies [54].
Recently, the state-of-the-art results have been achieved by trans-
formers. We reproduce the framework of Polisis, using XLNet [49]
instead of CNN as a base classifier. XLNet is a transformer language
model, which extends Transformer-XL [8]. It is an auto-regressive
ICAIL’21, June 21–25, 2021, São Paulo, Brazil EL HAMDANI, et al.

language model, pretrained on all the permutations of the input Table 1: Results of categories prediction by the local classi-
sequence. We fine-tune the XLNet3 on 21 tasks – one task for pre- fiers approach
dicting the categories, and the rest – for each attribute’s values.
CNN XLNet
4.4 Text-to-Text approach Categories
P R F1 P R F1
In this section we explain how we use T5 [33] to solve HMTC. T5 is Introductory/Generic 74 40 52 76 54 63
a pretrained language model based on the transformer architecture.
Policy Change 85 60 71 73 65 69
T5 has two main differences in comparison to XLNet. First, it is
Specific audiences 90 77 83 85 80 82
pretrained on a multi-task mixture of unsupervised and supervised
Privacy Contact Info 87 52 65 84 75 79
tasks. Second, each task is converted into a text-to-text format. We
1𝑠𝑡 Party Collection 67 87 76 84 81 83
adopt T5 both for its top results on NLP benchmarks and for its
Data Retention 52 39 45 58 36 44
text-to-text nature.
3𝑟𝑑 party sharing/collection 71 85 78 76 87 81
The local classifiers approach has two main drawbacks. First, it
User Choice/Control 45 79 58 66 69 67
trains the set of local classifiers independently. Second, the number
Practice Not Covered 39 39 38 40 44 42
of local classifiers grows linearly with the size of the label hierarchy.
Data Security 79 48 60 77 68 72
These limitations motivate this second approach, where we convert
Access, Edit, Deletion 87 35 50 75 72 74
the HMTC task into two text-to-text tasks – one for each level of
Do Not Track 100 29 45 93 100 96
the label hierarchy – to better capture the dependencies of labels
Macro-Average 72 55 62 76 71 73
belonging to the same level. Moreover, by training one unique
algorithm for each level, we ensure that the number of classifiers
scales linearly with the hierarchy’s depth. 𝑇 𝑃𝑗 𝑇 𝑃𝑗 2 ∗ 𝑃𝑗 ∗ 𝑅𝑗
Thanks to the text-to-text nature of T5, we can simplify HMTC 𝑃𝑗 = 𝑅𝑗 = 𝐹 1𝑗 =
𝑇 𝑃𝑗 + 𝐹𝑃𝑗 𝑇 𝑃𝑗 + 𝐹 𝑁𝑗 𝑃𝑗 + 𝑅𝑗
into two text-to-text tasks shown in Figure 4. To prepare the task of
𝑇 𝑃 𝑗 , 𝐹 𝑃 𝑗 , 𝑇 𝑁 𝑗 and 𝐹 𝑁 𝑗 are the number of true positive, false pos-
categories prediction, we prepend the "categories prediction: " prefix
itive, true negative and false negative test examples with respect
to the text of segments and generate one sequence of categories
to class label 𝑦 𝑗 . To measure the global performance over a set of
separated by "; " as shown in Figure 5. The lists of categories were
labels we compute the macro-average of each metric by averaging
sorted in alphabetical order so that they have the same order across
on the set of labels4 .
training examples, as advised by the authors of [33].
For the task of values prediction, we only report the result of
The second task’s objective is to predict the values of the at-
attributes and values that we use to automate formal compliance
tributes of a category from an input segment, as well as to generate
checking of privacy policies, such that the reported metrics are the
the spans of texts related to the predicted values. This task is similar
macro-average over the values necessary for compliance checking.
to a reading comprehension task [34], where the question is "what
is the value of the attribute?", and the context paragraph is the pair
(segment, category). So we format it into a text-to-text task, similar Evaluation of span extraction: We use the F1-score, as in the
to how the authors of T5 formatted the reading comprehension SQuaD dataset [34], to evaluate the extraction of spans associated
dataset SQuAD (see Figure 6). with the values by comparing the ground truth target to the gener-
Once we format the tasks, we fine-tune the largest available T5 of ated target.
11B parameters on these tasks. We try two fine-tuning methods: the
first method (advised by [33]) is to fine-tune on each task indepen- 4.6 Results and Discussion
dently, and the second method is to fine-tune in a multi-task setting Local classifiers approach: In Table 1, we report the results of
on a mixture of both tasks to capture the global labels hierarchy. the evaluation of the CNN and XLNet experiments on categories
We fixed the hyperparameters of input sequence length and output prediction. We present the results of CNN and the XLNet for the task
sequence length to 512 and the batch size to 16, and performed a of values prediction in Table 2. XLNet has superior performance
grid-search over the learning rate. The model was fine-tuned each comparing to CNN for the task of categories predictions. However,
time for 25,000 steps. Interestingly, the best-performing learning it has significantly lower performance than CNN for the second
rate is the same (4e-3) for all the models and tasks. task, because there are enough examples to fine-tune XLNet for
categories prediction but not enough for values prediction.
4.5 Evaluation measures
Text-to-text approach: We report the results of evaluating the
Evaluation of multi-label classification: We use precision, re-
two tasks on the test dataset in Table 3, Table 4, and Table 5. We
call, and F1-score metric to evaluate the extraction of both coarse-
observe that the individual fine-tuning and multi-task fine-tuning
grained and fine-grained data practices from privacy policies seg-
have a close recall for both tasks, but they differ significantly in
ments. Since we are in a multi-label classification setting, we adapt
precision. By fine-tuning each task separately, we obtain a 2.6%
the traditional single-label metrics to this setting by using the label-
precision improvement for the task of categories prediction and
based metrics [52]: precision, recall, and F1-score for the j-th class
label 𝑦 𝑗 are defined as follows: 4 It is worth noting that we don’t use the same precision, recall and F1-score as in [14]
where they use the macro-average of each metric predicting the presence and absence
3 We used the pretrained model available at Hugging Face models hub. of the label.

45
A Combined Rule-Based and Machine Learning Approach for Automated GDPR Compliance Checking ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Figure 4: The hierarchical multi-label classification of OPP-115 is converted into two text-to-text tasks. The first task is to predict categories
of data-practice from an input segment. We then retrieve the attributes of each predicted category to feed them and their category as input to
the second task. The second task is to predict the values of the input attributes and to generate the corresponding span of text (highlighted).

Table 3: Results of categories prediction by T5

Multi-task Fine-tuning Task 1 Fine-tuning

Category
P R F1 P R F1
Introductory/Generic 68 61 64 69 58 63
Policy Change 82 67 74 82 69 75
Specific audiences 94 84 88 91 85 87
Privacy Contact Info 81 82 82 79 69 74
Figure 5: Input example to train T5 for categories prediction 1𝑠𝑡 Party Collection 88 87 87 90 85 87
Data Retention 65 47 55 77 52 62
3𝑟𝑑 party sharing/collection 83 86 85 85 85 85
User Choice/Control 65 60 62 66 64 65
Practice Not Covered 46 45 46 49 47 48
Data Security 73 66 69 78 70 74
Access, Edit, Deletion 82 84 83 78 76 77
Do Not Track 100 75 85 100 62 76
Macro-Average 77 70 73 79 69 73

Figure 6: Example of an input to train T5 for prediction of Table 4: Results of values prediction for attributes used in
attributes’ values and the corresponding spans of text compliance checking by T5. We report the macro-average
over the values of each attribute.
Table 2: Results of values prediction by CNN and XLNet
for attributes used in compliance checking. We report the Multi-task Fine-tuning Task 2 Fine-tuning
Attribute
macro-average over the values of each attribute. P R F1 P R F1
action first-party 55 56 55 62 58 60
CNN XLNet does/does not 90 78 83 91 83 87
Attribute personal information type 72 61 66 73 63 68
P R F1 P R F1
purpose 72 65 69 74 67 70
action first-party 44 47 45 16 33 21 retention period 53 47 50 50 25 33
does/does not 93 77 84 84 82 83 access type 62 58 60 71 70 70
personal information type 73 58 64 56 49 52 Macro-average 67 61 63 70 61 65
purpose 74 56 64 74 64 69
retention period 79 62 69 50 11 18
access type 61 51 55 30 26 27
Macro-average 72 60 65 51 44 47 where separate models trained on each task outperforms the multi-
task model is coherent with previous findings [4, 26, 33].
The performance of span extraction, presented in Table 5 is
4.5% precision improvement for values prediction. This behavior low in comparison with the performance of transformers models

46
ICAIL’21, June 21–25, 2021, São Paulo, Brazil EL HAMDANI, et al.

on similar tasks such as reading comprehension or named entity
recognition, which might be due to the relatively small number of
training examples given to T5.
Table 5: Results of evaluation of span extraction by T5.
communicated to data subjects (language complexity, length of sen-
tences, etc.). Legal experts manually converted rules from Articles
13 and 14 into code using the OPP-115 taxonomy. As the OPP-115
taxonomy does not cover all the concepts of the GDPR, we only
encoded the articles presented in the second column of Table 7.

Table 7: List of mandatory information from articles 13 and

Multi-task Fine-tuning Task 2 Fine-tuning
14 of the GDPR encoded by us using the OPP-115 taxonomy.
P R F1 P R F1
64 57 52 65 59 54 Mandatory information Article Reference
Identity of the controller 13 1.a; 14 1.a
Contact details of the controller 13 1.a; 14 1.a
Purpose of the processing of personal data 13 1.c; 14 1.c
Local classifiers approach vs. text-to-text approach: We present Right to data portability 13 2.b; 14 2.c
in Table 6 the macro-average of the different experiments for both Right to erasure 13 2.b; 14 2.c
Right to rectification 13 2.b; 14 2.c
tasks – categories and values prediction. Fine-tuning separate mod- Right to access 13 2.b; 14 2.c
els of T5 for each task achieves the highest F1 score. Both XLNet Data retention period or the criteria of retention period 13 2.a; 14 2.a
and T5 of the transformers family significantly outperform CNN The recipients or categories of recipients of the personal data 13 1.e; 14 1.e
Categories of personal data 14 1.d
on the first task. However, the performance of transformers on the
task of values prediction is at its best close to CNN’s performance.
The local classifiers approach requires fine-tuning separate mod- We plan to build a GUI for legal experts to convert compliance
els for each attribute. Consequently, it decreases the number of rules into code. To do so, we choose JsonLogic to serialize obtained
training examples seen by each model, explaining the significant rules as a JSON file. JsonLogic provides a simple mechanism to share
performance gap between task 1 and task 2. This performance gap rules between the front-end and back-end of a GUI. It comes with a
is more important for XLNet than CNN, which could result from parser that we use to build a first-order logic inference engine with
the high number of parameters of XLNet we need to fine-tune in Python. In Figure 7 we present an example of a rule encoded with
comparison to the CNN architecture. T5 has even more parameters JsonLogic and OPP-115 taxonomy. We consider that the purpose of
than XLNet, but its performance does not drop as significantly as processing personal data is mentioned in the privacy policy if there
XLNet. We can explain this difference by the nature of the text-to- is at least one data practice whose category is “1𝑠𝑡 Party Collection”
text approach where we fine-tune one model of T5 on the prediction and the attribute "Purpose" has a value different from “Unspecified”.
of the values of all the attributes instead of individual fine-tuning
for each attribute. Hence, T5 sees much more examples than XLNet. {"some":[{"var":"data_practices"},
{"and":[{"==":[{"var":"category"},"first_party"]},
{"!=":[{"var":"attributes.purpose"},"unspecified"
]}]}]}
Table 6: Macro-Average of precision, recall and F1-score for
categories prediction (Task 1) and values prediction (Task 2)
by the four models. Figure 7: Example of encoding a GDPR rule with JsonLogic.
The rule states the obligation of mentioning the purpose of the
processing of personal data.
Task 1 Task 2
Approach
P R F1 P R F1
CNN 72 55 62 72 60 65 5.2 Evaluation
Local classifiers
XLNet 76 71 73 51 44 47
The OPP-115 taxonomy was created before the entry into force of
Multi-task FT 77 70 73 67 61 63 the GDPR. To evaluate its capacity to capture GDPR concepts, we
Text-to-Text: T5
Individual FT 79 69 73 70 61 65 create a dataset of 30 privacy policies where legal experts indicate
each mandatory information’s presence. We use this dataset as
ground truth of mandatory information listed in Table 7.
5 RULE BASED COMPLIANCE CHECKING The ground truth dataset contains two types of privacy policies:
15 privacy policies are from the OPP-115 dataset and 15 post-GDPR
5.1 From rules to code privacy policies are from the corpus released by [23]. We extract
This section describes our rule-based approach to automate the data practices from privacy policies and feed them to the inference
formal compliance checking of privacy policies. Privacy policies engine to check for the presence of each mandatory information.
must comply with Articles 12, 13, and 14 of the GDPR. We limit For the first type of privacy policy, we use the ground truth data
our experiments to Articles 13 and 14, listing the mandatory in- practices extracted manually by legal experts from [48]. For the
formation that privacy policies must contain, for which we can second type, we use data practices predicted by T5. Therefore, any
use information extraction algorithms, described in the previous error on the first type of privacy policies will not be due to machine
section. To verify compliance with Article 12, we will need to de- learning errors but due to the OPP-115 taxonomy used to encode
velop other algorithms to assess how the mandatory information is mandatory information.

47
A Combined Rule-Based and Machine Learning Approach for Automated GDPR Compliance Checking
We report metrics of both the absence and presence of informa-
tion in Table 8. Our objective is to detect non-compliance and send
the documents for review by experts, so we need to maximize the
number of absent mandatory information we can detect.
Table 8: Results of mandatory information detection from
both OPP-115 and post-GDPR privacy policies.
Dataset P R
OPP-115: absence 93
OPP-115: presence 93
Post-GDPR: absence 78
Post-GDPR: presence 91
Most of the errors on the OPP-115 dataset (see Figure 8) are
caused by the difficulty of aligning OPP-115 concepts with GDPR
concepts. For example, to encode "Data retention period" we used
the "Retention period" attribute. However, even when the anno-
tators select a value for the "Retention period" in [48] it does not
always concern all the collected personal data. In contrast, the
GDPR requires that the retention period is stated for all of the data.
The majority of errors on the post-GDPR policies are when the
algorithm does not detect the right to data portability. This type
of error is expected: because data portability has not been widely
adopted before the GDPR (Art. 20), and no privacy policy from the
OPP-115 dataset mentions the right to data portability.
Although our rules did not capture the right to data portability,
T5 correctly predicted the category "data portability" from this sen-
tence: "Data portability, that is to say the possibility of receiving
these data in a structured format that is readable by an automatic de-
vice and of sending them to another processing owner without any
impediments." T5 could predict new categories of data practices due
to its language understanding capabilities that enable few and zero-
shot inference [22, 50, 53]. Other new categories from the GDPR
articles, predicted for post-GDPR policies are: "data minimisation"
(Art. 5.1(c)), "data accuracy" (Art. 5.1(d)), "legitimate interests" (Art.
6.1(f)), "lawful basis" (Art. 6), and "right to object" (Art. 21).
OPP-115
Normalized count of errors
0.4 Post-GDPR
0.3
0.2
0.1
0.0
iod ss ure ion ata
n per acce eras ificat nal d nal d rtab cipie
a r e tentio Right to Right to ht to rect of persoser persoto data poories of re
Dat s
Rig egorie g of u Right categ
Cat rocessin ts or
e of p ipien
r p o s Th e rec
Pu
Information type
Figure 8: The distribution of errors over types of mandatory
information for OPP-115 and post-GDPR privacy policies
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
6 CONCLUSION
This study designed a theoretical framework to implement and
monitor GDPR compliance in the data supply chain through a
document-based approach, for which we defined three key tasks.
We proposed a formal and substantive approach to verify GDPR-
compliance of privacy policies. It is worth highlighting that, as
a potential next step, our framework could be adapted to other
compliance documents like DPIAs and/or ROPAs. More broadly,
F1 research is also needed to implement this framework in a multi-
document setting, where data processing activities are described in
88 90
multiple documents.
97 95
Our second significant contribution is the experimentation on
78 78
the automation of formal compliance checking of privacy policies,
91 91
which could be generalized to other documents in the data supply
chain as a means to improve end-to-end compliance. We build a
system combining machine learning and rules to detect the presence
of information required by the GDPR. We fine-tuned the T5 model
in a multi-task setting and achieved good performance predicting
both coarse-grained and fine-grained data practices with only one
model. The T5 model also extracts the spans of text corresponding
to the fine-grained data practices. These spans of text could be used
to explain the predicted values.
We used the OPP-115 taxonomy to encode 10 GDPR rules from
Articles 13 and 14 concerning the information a privacy policy
should contain. We evaluated the system on a corpus of 30 privacy
policies, where legal experts indicated the presence of mandatory
information. Although OPP-115 taxonomy is pre-GDPR, it proved
capable of capturing some mandatory information in both pre-
GDPR and post-GDPR policies. Currently, it is one of the most
valuable resources in our research community. Still, it is not enough
to encode both GDPR rules and data protection activities defined in
compliance documents. Thus, there is a need for a new corpus of
data protection documents from the data supply chain, to automate
compliance checking tasks, which we leave for the future work.
Additionally, we pointed that T5 was able to predict new cate-
gories such as data portability. This capacity of zero-shot prediction
can be leveraged to assist law and privacy scholars in creating a
GDPR taxonomy compatible with the variety of compliance docu-
ments in the data supply chain.
REFERENCES
[1] 2017. The True Cost of Compliance with Data Protection Regulations. Technical
Report. Ponemon Institute LLC.
[2] 2019. ICO Guidance: Update report into adtech and real time bidding. Technical
Report. Information Commissioner’s Office. 19–21 pages.
[3] David Restrepo Amariles, Aurore Clément Troussel, and Rajaa El Hamdani. 2020.
Compliance Generation for Privacy Documents under GDPR: A Roadmap for
Implementing Automation and Machine Learning. arXiv preprint arXiv:2012.12718
(2020).
ata ility nts [4] Naveen Arivazhagan, Ankur Bapna, Orhan Firat, Dmitry Lepikhin, Melvin John-
son, Maxim Krikun, Mia Xu Chen, Yuan Cao, George Foster, Colin Cherry, et al.
2019. Massively multilingual neural machine translation in the wild: Findings
and challenges. arXiv preprint arXiv:1907.05019 (2019).
[5] Jaspreet Bhatia, Travis D Breaux, Joel R Reidenberg, and Thomas B Norton. 2016.
A theory of vagueness and privacy risk perception. In 2016 IEEE 24th International
Requirements Engineering Conference (RE). IEEE, 26–35.
[6] Giuseppe Contissa, Koen Docter, Francesca Lagioia, Marco Lippi, Hans-W Mick-
litz, Przemysław Pałka, Giovanni Sartor, and Paolo Torroni. 2018. Claudette meets
gdpr: Automating the evaluation of privacy policies using artificial intelligence.
Available at SSRN 3208596 (2018).
[7] Elisa Costante, Yuanhao Sun, Milan Petković, and Jerry Den Hartog. 2012. A
machine learning solution to assess privacy policy completeness: (short paper).
48
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
In Proceedings of the 2012 ACM Workshop on Privacy in the Electronic Society.
91–96.
[8] Zihang Dai, Zhilin Yang, Yiming Yang, Jaime Carbonell, Quoc V Le, and Ruslan
Salakhutdinov. 2019. Transformer-xl: Attentive language models beyond a fixed-
length context. arXiv preprint arXiv:1901.02860 (2019).
[9] Marina De Vos, Sabrina Kirrane, Julian Padget, and Ken Satoh. 2019. ODRL policy
modelling and compliance checking. In International Joint Conference on Rules
and Reasoning. Springer, 36–51.
[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
Pre-training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805 (2018).
[11] Olha Drozd and Sabrina Kirrane. 2020. Privacy CURE: Consent Comprehension
Made Easy. In IFIP International Conference on ICT Systems Security and Privacy
Protection. Springer, 124–139.
[12] María Teresa Gómez-López, Luisa Parody, Rafael M Gasca, and Stefanie Rinderle-
Ma. 2014. Prognosing the compliance of declarative business processes using
event trace robustness. In OTM Confederated International Conferences" On the
Move to Meaningful Internet Systems". Springer, 327–344.
[13] Guido Governatori and Sidney Shek. 2012. Rule Based Business Process Compli-
ance.. In RuleML (2). Citeseer.
[14] Hamza Harkous, Kassem Fawaz, Rémi Lebret, Florian Schaub, Kang G Shin,
and Karl Aberer. 2018. Polisis: Automated analysis and presentation of privacy
policies using deep learning. In 27th {USENIX } Security Symposium ( {USENIX }
Security 18). 531–548.
[15] Mustafa Hashmi, Guido Governatori, and Moe Thandar Wynn. 2012. Business
process data compliance. In International Workshop on Rules and Rule Markup
Languages for the Semantic Web. Springer, 32–46.
[16] Mustafa Hashmi, Guido Governatori, and Moe Thandar Wynn. 2016. Norma-
tive requirements for regulatory compliance: An abstract formal framework.
Information Systems Frontiers 18, 3 (2016), 429–455.
[17] Martin Hepp. 2008. Ontologies: State of the art, business potential, and grand
challenges. Ontology Management (2008), 3–22.
[18] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
[19] Katsiaryna Krasnashchok, Majd Mustapha, Anas Al Bassit, and Sabri Skhiri. 2020.
Towards Privacy Policy Conceptual Modeling. In International Conference on
Conceptual Modeling. Springer, 429–438.
[20] Logan Lebanoff and Fei Liu. 2018. Automatic detection of vague words and
sentences in privacy policies. arXiv preprint arXiv:1808.06219 (2018).
[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-
based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278–
2324.
[22] Omer Levy, Minjoon Seo, Eunsol Choi, and Luke Zettlemoyer. 2017. Zero-shot
relation extraction via reading comprehension. arXiv preprint arXiv:1706.04115
(2017).
[23] Thomas Linden, Rishabh Khandelwal, Hamza Harkous, and Kassem Fawaz. 2020.
The privacy policy landscape after the GDPR. Proceedings on Privacy Enhancing
Technologies 2020, 1 (2020), 47–64.
[24] Fei Liu, Nicole Lee Fella, and Kexin Liao. 2018. Modeling language vagueness
in privacy policies using deep neural networks. arXiv preprint arXiv:1805.10393
(2018).
[25] Frederick Liu, Shomir Wilson, Peter Story, Sebastian Zimmeck, and Norman
Sadeh. 2018. Towards automatic classification of privacy policy text. School of
Computer Science Carnegie Mellon University, Pittsburgh, PA, Tech. Rep. CMU-ISR-
17-118R and CMULTI-17-010 (2018).
[26] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. 2018.
The natural language decathlon: Multitask learning as question answering. arXiv
preprint arXiv:1806.08730 (2018).
[27] Tomas Mikolov, Kai Chen, G. S. Corrado, and J. Dean. 2013. Efficient Estimation
of Word Representations in Vector Space. In ICLR.
[28] Majd Mustapha, Katsiaryna Krasnashchok, Anas Al Bassit, and Sabri Skhiri.
2020. Privacy Policy Classification with XLNet (Short Paper). In Data Privacy
Management, Cryptocurrencies and Blockchain Technology. Springer, 250–257.
[29] Najmeh Mousavi Nejad, Pablo Jabat, Rostislav Nedelchev, Simon Scerri, and
Damien Graux. 2020. Establishing a strong baseline for privacy policy classi-
fication. In IFIP International Conference on ICT Systems Security and Privacy
Protection. Springer, 370–383.
[30] Monica Palmirani, Michele Martoni, Arianna Rossi, Cesare Bartolini, and Livio
Robaldo. 2018. PrOnto: Privacy ontology for legal reasoning. In International
Conference on Electronic Government and the Information Systems Perspective.
Springer, 139–152.
[31] Ellen Poplavska, Thomas B Norton, Shomir Wilson, and Norman Sadeh. 2020.
From Prescription to Description: Mapping the GDPR to a Privacy Policy Corpus
Annotation Scheme. In 33rd International Conference on Legal Knowledge and
Information Systems, JURIX 2020. IOS Press BV, 243–246.
[32] Wenjun Qiu and David Lie. 2020. Deep Active Learning with Crowdsourcing
Data for Privacy Policy Classification. arXiv preprint arXiv:2008.02954 (2020).
49
EL HAMDANI, et al.
[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang,
Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. 2019. Exploring the lim-
its of transfer learning with a unified text-to-text transformer. arXiv preprint
arXiv:1910.10683 (2019).
[34] Pranav Rajpurkar, Robin Jia, and Percy Liang. 2018. Know what you don’t know:
Unanswerable questions for SQuAD. arXiv preprint arXiv:1806.03822 (2018).
[35] Abhilasha Ravichander, Alan W Black, Shomir Wilson, Thomas Norton, and
Norman Sadeh. 2019. Question answering for privacy policies: Combining com-
putational and legal perspectives. arXiv preprint arXiv:1911.00841 (2019).
[36] Joel R Reidenberg, Jaspreet Bhatia, Travis D Breaux, and Thomas B Norton. 2016.
Ambiguity in privacy policies and the impact of regulation. The Journal of Legal
Studies 45, S2 (2016), S163–S190.
[37] Community Research and Development Information Service. 2021. Business
Process Re-engineering and functional toolkit for GDPR compliance. https:
//cordis.europa.eu/project/id/787149/results. Accessed: 2021-02-28.
[38] Kanthashree Mysore Sathyendra, Shomir Wilson, Florian Schaub, Sebastian
Zimmeck, and Norman Sadeh. 2017. Identifying the provision of choices in
privacy policy text. In Proceedings of the 2017 Conference on Empirical Methods in
Natural Language Processing. 2774–2779.
[39] Carlos N Silla and Alex A Freitas. 2011. A survey of hierarchical classification
across different application domains. Data Mining and Knowledge Discovery 22, 1
(2011), 31–72.
[40] Laurens Sion, Pierre Dewitte, Dimitri Van Landuyt, Kim Wuyts, Peggy Valcke,
and Wouter Joosen. 2020. DPMF: A Modeling Framework for Data Protection by
Design. Enterprise Modelling and Information Systems Architectures (EMISAJ) 15
(2020), 10–1.
[41] Duyu Tang, Furu Wei, Nan Yang, Ming Zhou, Ting Liu, and Bing Qin. 2014.
Learning sentiment-specific word embedding for twitter sentiment classification.
In Proceedings of the 52nd Annual Meeting of the Association for Computational
Linguistics (Volume 1: Long Papers). 1555–1565.
[42] Welderufael B Tesfay, Peter Hofmann, Toru Nakamura, Shinsaku Kiyomoto,
and Jetzabel Serna. 2018. PrivacyGuide: towards an implementation of the EU
GDPR on internet privacy policy evaluation. In Proceedings of the Fourth ACM
International Workshop on Security and Privacy Analytics. 15–21.
[43] Damiano Torre, Sallam Abualhaija, Mehrdad Sabetzadeh, Lionel Briand, Katrien
Baetens, Peter Goes, and Sylvie Forastier. 2020. An ai-assisted approach for
checking the completeness of privacy policies against gdpr. In 2020 IEEE 28th
International Requirements Engineering Conference (RE). IEEE, 136–146.
[44] Damiano Torre, Ghanem Soltana, Mehrdad Sabetzadeh, Lionel C Briand, Yuri
Auffinger, and Peter Goes. 2019. Using models to enable compliance checking
against the GDPR: an experience report. In 2019 ACM/IEEE 22nd International
Conference on Model Driven Engineering Languages and Systems (MODELS). IEEE,
1–11.
[45] Silvano Colombo Tosatto, Guido Governatori, Nick van Beest, and Francesco
Olivieri. 2019. Efficient Full Compliance Checking of Concurrent Components
for business Process Models. FLAP 6, 5 (2019), 963–998.
[46] Sebastian Urbina. 2002. Legal method and the rule of law. Vol. 59. Springer Science
& Business Media.
[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all
you need. arXiv preprint arXiv:1706.03762 (2017).
[48] Shomir Wilson, Florian Schaub, Aswarth Abhilash Dara, Frederick Liu, Sushain
Cherivirala, Pedro Giovanni Leon, Mads Schaarup Andersen, Sebastian Zimmeck,
Kanthashree Mysore Sathyendra, N Cameron Russell, et al. 2016. The creation
and analysis of a website privacy policy corpus. In Proceedings of the 54th Annual
Meeting of the Association for Computational Linguistics (Volume 1: Long Papers).
1330–1340.
[49] Z. Yang, Zihang Dai, Yiming Yang, J. Carbonell, R. Salakhutdinov, and Quoc V.
Le. 2019. XLNet: Generalized Autoregressive Pretraining for Language Under-
standing. In NeurIPS.
[50] Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking zero-shot text
classification: Datasets, evaluation and entailment approach. arXiv preprint
arXiv:1909.00161 (2019).
[51] Razieh Nokhbeh Zaeem, Rachel L German, and K Suzanne Barber. 2018. Privacy-
check: Automatic summarization of privacy policies using data mining. ACM
Transactions on Internet Technology (TOIT) 18, 4 (2018), 1–18.
[52] Min-Ling Zhang and Zhi-Hua Zhou. 2013. A review on multi-label learning
algorithms. IEEE transactions on knowledge and data engineering 26, 8 (2013),
1819–1837.
[53] Ben Zhou, Daniel Khashabi, Chen-Tse Tsai, and Dan Roth. 2019. Zero-shot open
entity typing as type-compatible grounding. arXiv preprint arXiv:1907.03228
(2019).
[54] Sebastian Zimmeck, Peter Story, Daniel Smullen, Abhilasha Ravichander, Ziqi
Wang, Joel Reidenberg, N Cameron Russell, and Norman Sadeh. 2019. Maps:
Scaling privacy compliance analysis to a million apps. Proceedings on Privacy
Enhancing Technologies 2019, 3 (2019), 66–86.
 On Semantics-based Minimal Revision for Legal Reasoning
Wachara Fungwacharakorn Kanae Tsushima
National Institute of Informatics and National Institute of Informatics and
SOKENDAI SOKENDAI
Chiyoda, Tokyo, Japan Chiyoda, Tokyo, Japan
wacharaf@nii.ac.jp k_tsushima@nii.ac.jp
ABSTRACT
When literal interpretation of statutes leads to counterintuitive
consequences, judges, especially in high courts, may identify coun-
terintuitive consequences and revise interpretation of statutes. Re-
searchers have studied revisions for computational legal represen-
tation. Generally, studies on revision usually consider minimal revi-
sion to reflect limitation of judges’ legislative power. However, those
studies tend to minimize the number of operations used for chang-
ing rules rather than minimize the changes of semantics (the set of
conclusions obtained from the program), which vary among cases.
In this paper, we consider minimizing the changes of semantics of
a rule-base written in a normal logic program. We consider that
each possible fact-base (the representation of a case) has its corre-
sponding semantics and corresponding dominant rule-base, which
is a set of Horn clauses obtained from the subset of rule-base that
is specific to the considered fact-base. Hence, we present a new sub
type of semantics-based minimal revision called a dominant-based
minimal revision. Furthermore, we present one guidance to obtain
one dominant-based minimal revision by using legal debugging and
Closed World Specification. We also compare the dominant-based
minimal revision with the syntax-based minimal revision in Theory
Distance Metric. As the syntax-based minimal revision minimizes
the number of operations used for changing rules, the comparison
shows that the syntax-based minimal revision may cause extra se-
mantics changes compared to the dominant-based minimal revision,
especially when the rule-base contains multiple rules for the same
consequence. We discuss that such extra semantics changes can be
considered as unintentional changes caused by the syntax-based
minimal revision. Hence, legal reasoning systems can check with
the user such extra semantics changes to confirm the user intention
of changes.
CCS CONCEPTS
• Computing methodologies → Nonmonotonic, default rea-
soning and belief revision; • Applied computing → Law.
KEYWORDS
legal reasoning, legal representation, theory revision
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466075
50
Ken Satoh
National Institute of Informatics and
SOKENDAI
Chiyoda, Tokyo, Japan
ksatoh@nii.ac.jp
ACM Reference Format:
Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh. 2021. On
Semantics-based Minimal Revision for Legal Reasoning. In Eighteenth In-
ternational Conference for Artificial Intelligence and Law (ICAIL’21), June
21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https:
//doi.org/10.1145/3462757.3466075
1 INTRODUCTION
When literal interpretation of the law leads to counterintuitive
consequences, judges, especially in high courts, may have a legal
judgement that also revises interpretation of statutes. Such a legal
judgement is called a case law, which suggests more appropriate
interpretation of statutes for a present case and also later cases.
Such revisions are usually described as discoveries of essential con-
ditions, which are revealed in a real-life case [27]. This is similar to
a qualification problem [42] in artificial intelligence, which discusses
the necessity of learning essential conditions in real-life since we
cannot know them all in the first place.
Computational law researchers have been long interested in
building legal representations for interpretations of statutes and
modelling revisions for legal representation. Such legal representa-
tions are, for example, Institutional models [15], Defeasible logic
[21], or PROLEG [37], with revision models in Institutional model
[28], in Defeasible logic [22], or in PROLEG [18]. In addition, some
studies focus on modelling revisions that are not specific for one
legal representation such as [4]. One theme that some studies (e.g.
[4, 28, 35]) have in common is that a revision should be a minimal
revision of legal rules. This reflects a principle that judges usually
limit themselves from legislative issues. However, since legal rea-
soning is a hybrid between reasoning by rules and reasoning by
cases [10, 43], a revision should not concern only effects on rules
but also effects on interpretations applying in each case. Although
effects of changes on each case and its corresponding semantics has
been considered in hybrid legal reasoning systems between rule-
based and case-based [32, 46], effects of such changes on revisions
of rules in pure rule-based legal reasoning systems have not been
fully investigated.
Therefore, in this paper, we present one semantics-based min-
imal revision relying on a rule-based structure called a dominant
rule-base. Firstly, we introduce semantics of a rule-base written in
a normal logic program and the difference of semantics. Semantics
of a rule-base is defined by a set of all possible fact-bases and corre-
sponding interpretations of statutes in pairs whereas the difference
of semantics is a set of all possible fact-bases with corresponding
differences of interpretations. Then, we define the semantics-based
minimal revision, which means a revision that causes the smallest
semantics changes. However, since the definition of semantics-
based minimal revision is considered to be too strong, we relax the
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
definition by defining a partial semantics-based minimal revision.
Then, we describe a sub type of a partial semantics-based minimal
revision based on a dominant rule-base, which is a set of Horn
clauses obtained from the subset of rule-base specific to the consid-
ered fact-base. Since the judicial revision of legal interpretation is
initiated by an exceptional case, we present one guidance to obtain
the dominant-based minimal revision based on Legal Debugging
[17] and Closed World Specification [6] by putting all facts in an
exceptional case as a condition of a new rule. We show that this
revision would affect only the dominant rule-base of the fact-bases
that are larger or equal to the exceptional case. Then, to optimize
the revision, we remove redundant conditions in the body as long
as the minimality of the revision remains.
In spite of definitions for the dominant-based minimal revision
and the guidance to obtain it, we also compare the dominant-based
minimal revision with the syntax-based minimal revision in The-
ory Distance Metric [47], which states that a distance of revision
is the minimal number of program edit operations (i.e. deleting
a rule, adding a rule with an empty body, adding a condition to
a rule or deleting a condition from a rule) that are used for re-
vising an original program to a new program. The definition has
been applied in many rule-based revisions (e.g. [12, 28]) due to its
simplicity. From this comparison, we show that the syntax-based
minimal revision may cause extra semantics changes compared to
the dominant-based minimal revision, especially when the original
rule-base contains multiple rules for the same consequence. We
discuss that such extra semantics changes can be considered as
changes unintentionally caused by the syntax-based minimal revi-
sion and the legal reasoning system can use such extra semantics
changes for check with the user the intention of changes.
This paper is structured as follows. Section 2 reviews related
work. Section 3 provides preliminary definitions of a normal logic
program for legal reasoning and a revision of a normal logic pro-
gram based on Legal Debugging and Closed World Specification.
Section 4 presents the definitions of semantics of a rule-base, the
difference of semantics, the semantics-based minimal revision and
the dominant-based minimal revision. Section 5 describes one guid-
ance to obtain a dominant-based minimal revision based on Legal
Debugging and Closed World Specification. Section 6 compares the
dominant-based minimal revision with the syntax-based minimal
revision in Theory Distance Metric with the example of the judicial
revision of Japanese Civil Code Article 612. Section 7 discusses the
results obtained from the comparison between the dominant-based
minimal revision and the syntax-based minimal revision. Finally,
Section 8 provides the conclusion and future works.
2 RELATED WORK
AI and Law researchers have been long interested in legal change
[2] especially in case-based legal reasoning systems since such legal
change can be detected through a temporal order of precedent cases
[8, 34]. The role of legal context in legal change has been explored
in [23] and classified into three aspects, which are teleological rela-
tions, temporal relations, and procedural postures. As AI and Law
researchers become interested in building case law theories (for
instance, building case law theories [11] from a case-base in HYPO
[33] or building case law theories [5] from a case-base in AA-CBR
51
Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh
[13]), recent studies become widely interested in revisions of such
theories. For example, Henderson and Bench-Capon [24] have en-
couraged contrastive revision of a theory by not only supporting
accepted consequences but also attacking the objected ones. Rotolo
and Roversi [35] have classified theory revision criteria and mini-
mal change is one of such criteria. This criterion comes from the
standard revision theory, which intuitively states that we should
adjust theories as little as possible. To minimally revise rule-based
theories, they present two strategies:
(1) keep the set of rules as close as possible to the original one
(this strategy is independent of the facts of the case)
(2) minimize the changes of the set of conclusions obtained from
the theory (the strategy is dependent of the facts of the case
since different facts give different set of conclusions)
The first strategy could be traced back to the early studies of syntax-
based (or formula-based) minimal revision e.g. [29, 47]. On the other
hand, the second strategy could be traced back to the early studies
of semantics-based (or model-based) minimal revision e.g. [14, 36]
(for an intensive survey on both types of minimal revision, see [16],
for an intensive survey specific to semantics-based revision, see
[26]).
Explaining legal change can also be seen in recent extensions
of HYPO [7, 31]. For example, Horty Bench-Capon theory [25],
one extension of HYPO, states that a new legal judgement could
be made as long as it preserves the precedential constraint. This
guides a judge to introduce new factors in a new case to preserve
the precedential constraint if the new case is exceptional. Horty
Bench-Capon theory has been extended into Rigoni theory [32]
for supporting hybrid legal reasoning systems between rule-based
and case-based. Moreover, recently presented Verheij’s case models
[44–46] have become new explanations of hybrid legal reasoning
systems by using the preference of cases. In our study, we explain
the legal change using a dominant rule-base. Hence, our study is
different from Rigoni theory and Verheij’s case models since their
approaches focus on the effects of legal change on cases such as the
precedential constraint or the preference of cases but our approach
focuses on the effects of legal change on revision of rules.
3 PRELIMINARIES
In this paper, we consider a normal logic program (also known
as Prolog program) for a computational legal representation as in
[38, 39, 41]. We use notations in our paper as follows.
Definition 3.1 (Normal Logic Program). A normal logic program
(hereafter, a program) is a set of rules of the form
ℎ ← 𝑏 1, . . . , 𝑏𝑚 , 𝑛𝑜𝑡 𝑏𝑚+1, . . . , 𝑛𝑜𝑡 𝑏𝑛 . (1)
where ℎ, 𝑏 1, . . . , 𝑏𝑛 are propositions. Let 𝑅 be a rule of the form
(1), we have
• ℎ as a head of a rule or a consequence of a rule denoted by
ℎ𝑒𝑎𝑑 (𝑅),
• {𝑏 1, . . . , 𝑏𝑚 } as a positive body of a rule denoted by 𝑝𝑜𝑠 (𝑅)
(each element of a positive body is called a requisite),
• {𝑏𝑚+1, . . . , 𝑏𝑛 } as a negative body of a rule denoted by 𝑛𝑒𝑔(𝑅)
(each element of a negative body is called an exception),
On Semantics-based Minimal Revision for Legal Reasoning
• {𝑏 1, . . . , 𝑏𝑚 , 𝑛𝑜𝑡 𝑏𝑚+1, . . . , 𝑛𝑜𝑡 𝑏𝑛 } as a body of a rule denoted
by 𝑏𝑜𝑑𝑦 (𝑅) (each element of a body is called a condition).
Sometimes, we express the rule in the form ℎ ← 𝐵. where 𝐵 is a
body of rule. We express ℎ. (called a fact) if the body of the rule is
empty. A rule is called a Horn clause if the negative body of the rule
is empty.
Since a program representing statutes is generally a non-recursive
and stratified program, we also hold this presumption in this paper.
A definition of non-recursive and stratified program is adopted
from [3] defined as follows.
Definition 3.2 (A non-recursive and stratified program). A pro-
gram 𝑇 is non-recursive and stratified if there is a partition 𝑇 =
𝑇0 ∪ 𝑇1 ∪ . . . ∪ 𝑇𝑛 (𝑇𝑖 and 𝑇 𝑗 disjoint for all 𝑖 ≠ 𝑗) such that, if a
proposition 𝑝 occurs in a body of rule in 𝑇𝑖 then a rule with 𝑝 in
the head is only contained within 𝑇0 ∪ 𝑇1 ∪ . . . 𝑇 𝑗 where 𝑗 < 𝑖.
In civil litigation, a judge would make correspondence between
factual situations in a case and factual concepts in statutes. Then, the
judge would conclude a legal decision based on related statutes. To
reflect this civil litigation, we determine a proposition occurring in
a head of a rule as a rule proposition and a proposition not occurring
in a head of a rule as a fact proposition. By this determination, we
denote a set of all fact propositions by F called a fact-domain and
we denote all fact propositions occurring in a program 𝑇 by 𝑓 (𝑇 )
hence, 𝑓 (𝑇 ) ⊆ F . We call a program 𝑅𝐵 a rule-base if 𝑅𝐵 has no
propositions in F occurring in a head of a rule, and all propositions
in 𝑅𝐵 that do not occurring in a head of a rule, are in F .
As a judge makes correspondence between factual situations in
a case and factual concepts, those factual concepts are represented
by fact propositions. Hence, we call a set of facts (rules with empty
bodies) constructed from a subset of F a fact-base. A fact-base then
represents a case. Then, the semantics is the set of propositions
(including rule propositions and fact propositions) which can be
concluded when compiling a fact-base 𝐹 𝐵 with a rule-base 𝑅𝐵
(denoted by 𝐹 𝐵 ∪ 𝑅𝐵). In this paper, we apply the stable model
semantics [20] defined as follows.
Definition 3.3 (Stable Model Semantics). Let 𝑇 be a normal logic
program and 𝑀 be a set of propositions. Let 𝑡𝑟𝑖𝑚(𝑇 ) be a trimming
function defined as follows: {ℎ𝑒𝑎𝑑 (𝑅) ← 𝑝𝑜𝑠 (𝑅)|𝑅 ∈ 𝑇 } and 𝑇 𝑀 =
𝑡𝑟𝑖𝑚({𝑅|𝑅 ∈ 𝑇 and 𝑛𝑒𝑔(𝑅) ∩𝑀 = ∅}). 𝑀 is a stable model semantics
of𝑇 if and only if 𝑀 is the minimum set (in the sense of set inclusion)
such that 𝑀 satisfies every rule 𝑅 ′ ∈ 𝑇 𝑀 , that is 𝑝𝑜𝑠 (𝑅 ′ ) ⊆ 𝑀
implies ℎ𝑒𝑎𝑑 (𝑅 ′ ) ∈ 𝑀.
The semantics of 𝐹 𝐵 ∪ 𝑅𝐵 represents the literal interpretation
of statute (represented by 𝑅𝐵) when applying in a particular case
(represented by 𝐹 𝐵). We denote a proposition 𝑝 is in the answer set
of 𝐹 𝐵 ∪ 𝑅𝐵 by 𝐹 𝐵 ∪ 𝑅𝐵 ⊢ 𝑝. Since we presume a non-recursive and
stratified program, 𝐹 𝐵 ∪ 𝑅𝐵 has a unique semantics [20]. This also
reflects a constraint that judges need one unique judgement from
legal rules.
When a judge applies the literal interpretation of statutes in
a particular case and it leads to counterintuitive consequences,
the judges may revise interpretation of statutes. We call such a
case an exceptional case. To distinguish the present case as an
exceptional case, judges would introduce a new factual concept in
52
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the present case. Then, judges use the introduced factual concept
to revise the statute so that the counterintuitive consequence is
resolved. To formalize this procedure, we firstly define agreement
and disagreement as follows.
Definition 3.4 (Agreement and Disagreement). Let 𝐹 𝐵 1, 𝐹 𝐵 2 be
fact-bases, 𝑅𝐵 1, 𝑅𝐵 2 be rule-bases, and 𝑝 be a proposition. We say
𝐹 𝐵 1 ∪ 𝑅𝐵 1 agrees with 𝐹 𝐵 2 ∪ 𝑅𝐵 2 on 𝑝 if
• 𝐹 𝐵 1 ∪ 𝑅𝐵 1 ⊢ 𝑝 and 𝐹 𝐵 2 ∪ 𝑅𝐵 2 ⊢ 𝑝 or
• 𝐹 𝐵 1 ∪ 𝑅𝐵 1 ⊬ 𝑝 and 𝐹 𝐵 2 ∪ 𝑅𝐵 2 ⊬ 𝑝
Otherwise, we say 𝐹 𝐵 1 ∪ 𝑅𝐵 1 disagrees with 𝐹 𝐵 2 ∪ 𝑅𝐵 2 on 𝑝.
Formalization of the procedure goes as follows. Firstly, we have a
proposition 𝑝 representing a counterintuitive consequence and 𝐹 𝐵𝑒
representing an exceptional case, which contains at least one fact
proposition not occurring in an original rule-base (hereafter, we
also refer to 𝐹 𝐵𝑒 as an exceptional case). Then, we revise an original
rule-base 𝑅𝐵 to a new rule-base 𝑅𝐵 ′ , Then 𝑅𝐵 ′ is a correct revision
of 𝑅𝐵 ′ with respect to 𝐹 𝐵𝑒 and 𝑝 if it can resolve the counterintuitive
consequence 𝑝. We call such a task a counterintuitive consequence
resolution task (CCR task) formally defined as follows.
Definition 3.5 (Counterintuitive Consequence Resolution (CCR)
Task). A counterintuitive consequence resolution (CCR) task is a
tuple ⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩ where 𝑅𝐵 is a rule-base representing statutes,
𝐹 𝐵𝑒 is a fact-base representing an exceptional case (𝑓 (𝐹 𝐵𝑒 ) ⊈
𝑓 (𝑅𝐵)), and 𝑝 is a considered counterintuitive consequence. A rule-
base 𝑅𝐵 ′ is a resolution to the CCR task ⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩ if 𝐹 𝐵𝑒 ∪ 𝑅𝐵 ′
disagrees with 𝐹 𝐵𝑒 ∪ 𝑅𝐵 on 𝑝 (hence, it implies that the considered
counterintuitive consequence is resolved).
Example 3.6. Let a fact-domain F = {𝑎, 𝑏, 𝑐}, and a rule-base
𝑅𝐵 1 = {𝑝 ← 𝑞. 𝑞 ← 𝑎. 𝑞 ← 𝑏.}. Suppose 𝑝 is a counterintuitive
consequence from applying 𝑅𝐵 1 in an exception case represented by
𝐹 𝐵 1 = {𝑎. 𝑐.}. Then, 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.}
and 𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑐.} are both
resolutions to the CCR task ⟨𝑅𝐵, 𝐹 𝐵 1, 𝑝⟩ since 𝐹 𝐵 1 ∪ 𝑅𝐵 1 disagrees
with 𝐹 𝐵 1 ∪ 𝑅𝐵 2 and 𝐹 𝐵 1 ∪ 𝑅𝐵 3 on 𝑝 which means 𝑝 is resolved in
both revisions.
From the previous example, we can see that there are possibly
more than one rules that we can put an exception (𝑛𝑜𝑡 𝑟 ) so that
the counterintuitive consequence 𝑝 is resolved. To let a user specify
which rule should an exception be put in, one can consider using
Legal Debugging [17], which extends from a common algorithmic
debugging [40] for legal reasoning. Legal Debugging considers a
user as an oracle query of an unknown set of intended interpretation
and the counterintuitive consequence is the symmetric difference
of the literal interpretation and the intended interpretation. Legal
debugging iterates to ask a user whether related consequences are
counterintuitive until it can no longer find any counterintuitive
consequences related. The last counterintuitive consequence found
is called a culprit, which is defined as follows.
Definition 3.7 (Culprit). A proposition 𝑝 is a culprit with respect
to an intended interpretation 𝐼𝑀 and a program 𝑇 if
• 𝑝 ∉ 𝐼𝑀 but there is a rule 𝑅 ∈ 𝑇 (called a supporting rule of 𝑝)
such that 𝑝𝑜𝑠 (𝑅) ⊆ 𝐼𝑀, 𝑛𝑒𝑔(𝑅) ∩ 𝐼𝑀 = ∅, and ℎ𝑒𝑎𝑑 (𝑅) = 𝑝
(we call such 𝑝 an incorrect culprit) or
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
• 𝑝 ∈ 𝐼𝑀 but there is no such supporting rule of 𝑝 in 𝑇 (we
call such 𝑝 an incomplete culprit)
We follow the culprit resolution from [19]. To resolve an incor-
rect culprit, we put exceptions in all supporting rules of the culprit
in the same manner of Closed World Specification [6]. To resolve
an incomplete culprit, we just put a new rule with the culprit as a
head. Following this instruction as illustrated in Algorithm 1, we
can reduce a CCR task to a task finding appropriate conditions to
a body of a new rule. Hence, we define a culprit resolution to CCR
task as follows.
Definition 3.8 (Culprit Resolution to CCR task). Given a CCR task
f 𝐻 obtained by Algorithm 1 where 𝑅𝐵
⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩, 𝑅𝐵, f is a prelim-
inary revision, and 𝐻 = {ℎ 1, . . . , ℎ𝑛 } as a set of rule propositions
that would become heads of new rules, and a set of Horn clauses
𝑅𝐵𝑛 = {ℎ 1 ← 𝐵 1 ., . . . , ℎ𝑛 ← 𝐵𝑛 .} where 𝐵𝑖 ⊆ 𝑓 (𝐹 𝐵𝑒 ) for all
1 ≤ 𝑖 ≤ 𝑛. A rule-base 𝑅𝐵 ′ = 𝑅𝐵 f ∪ {ℎ 1 ← 𝐵 1 ., . . . , ℎ𝑛 ← 𝐵𝑛 .} is
called a culprit resolution to the CCR task ⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩.
Algorithm 1 Preparation Phrase of Culprit Resolution
Given A CCR task ⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩
f = 𝑅𝐵 and 𝐻 = ∅
Let 𝑅𝐵
for all culprits 𝑝𝑐 detected from 𝑝 do
if 𝑝𝑐 is an incorrect culprit then
for all supporting rule 𝑅 of 𝑝𝑐 do
Let 𝑝𝑒 be a new rule proposition
Add 𝑝𝑒 to 𝐻
f
Add 𝑛𝑜𝑡 𝑝𝑒 to the body of 𝑅 in 𝑅𝐵
end for
else ⊲ when 𝑝𝑐 is an incomplete culprit
Add 𝑝𝑐 to 𝐻
end if
end for
Due to the property of culprits and Closed World Specification,
we get that a culprit resolution to a CCR task can resolve the con-
sidered counterintuitive consequence in the CCR task. Hence, a
culprit resolution is also a kind of resolution.
Example 3.9. Continuing from Example 3.6, if a culprit is 𝑝 then
g1 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ←
we put 𝑛𝑜𝑡 𝑟 in the first rule then 𝑅𝐵
𝑏.} and 𝐻 = {𝑟 }. If a culprit is 𝑞 then we put 𝑛𝑜𝑡 𝑟 in the second
g1 = {𝑝 ← 𝑞. 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏.} and 𝐻 = {𝑟 }.
rule then 𝑅𝐵
We get that 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.} and
𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑐.} in Example 3.6
are also culprit resolutions to the CCR task ⟨𝑅𝐵, 𝐹 𝐵 1, 𝑝⟩ since both
rule-bases include a new Horn clause 𝑟 ← 𝑐. and {𝑐} ⊆ 𝐹 𝐵 1 .
4 SEMANTICS-BASED MINIMAL REVISION
Since legal reasoning is a hybrid between reasoning by rules and
reasoning by cases [10, 43], a revision should consider semantics of
a rule-base, which vary among cases. Hence, we define semantics
of a rule-base as follows.
Definition 4.1 (Semantics of a rule-base). Let F be a fact-domain
and 𝑅𝐵 be a rule-base, The semantics of 𝑅𝐵 is a set of pairs defined
53
Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh
as follows: {⟨𝐹, 𝑀⟩|𝐹 ⊆ F and 𝑀 is the stable model semantics of
𝐹 𝐵 ∪ 𝑅𝐵 where 𝐹 𝐵 is a fact-base such that 𝑓 (𝐹 𝐵) = 𝐹 }. We denote
this set as 𝑠𝑒𝑚(𝑅𝐵).
We illustrate with the same setting in Example 3.6 throughout
this section. Let a fact-domain, F be {𝑎, 𝑏, 𝑐}, and a rule-base 𝑅𝐵 1 =
{𝑝 ← 𝑞. 𝑞 ← 𝑎. 𝑞 ← 𝑏.}. Then 𝑠𝑒𝑚(𝑅𝐵 1 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, {𝑎, 𝑝, 𝑞}⟩, ⟨{𝑏}, {𝑏, 𝑝, 𝑞}⟩, ⟨{𝑐}, {𝑐}⟩,
⟨{𝑎, 𝑏}, {𝑎, 𝑏, 𝑝, 𝑞}⟩, ⟨{𝑎, 𝑐}, {𝑎, 𝑐, 𝑝, 𝑞}⟩, ⟨{𝑏, 𝑐}, {𝑏, 𝑐, 𝑝, 𝑞}⟩,
⟨{𝑎, 𝑏, 𝑐}, {𝑎, 𝑏, 𝑐, 𝑝, 𝑞}⟩}
Then, we define the difference of semantics as follows.
Definition 4.2 (Difference of Semantics). Let F be a fact-domain
and 𝑅𝐵 1, 𝑅𝐵 2 be two rule-bases. The difference of semantics be-
tween 𝑅𝐵 1 and 𝑅𝐵 2 denoted by 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) is a set defined as
follows: {⟨𝐹, 𝐷⟩|𝐹 ⊆ F and 𝐷 is a symmetric difference between
the stable model semantics of 𝐹 𝐵 ∪ 𝑅𝐵 1 and 𝐹 𝐵 ∪ 𝑅𝐵 2 where 𝐹 𝐵 is
a fact-base such that 𝑓 (𝐹 𝐵) = 𝐹 }.
For example, let 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.}.
We get that, 𝑠𝑒𝑚(𝑅𝐵 2 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, {𝑎, 𝑝, 𝑞}⟩, ⟨{𝑏}, {𝑏, 𝑝, 𝑞}⟩, ⟨{𝑐}, {𝑐, 𝑟 }⟩,
⟨{𝑎, 𝑏}, {𝑎, 𝑏, 𝑝, 𝑞}⟩, ⟨{𝑎, 𝑐}, {𝑎, 𝑐, 𝑟 }⟩, ⟨{𝑏, 𝑐}, {𝑏, 𝑐, 𝑝, 𝑞, 𝑟 }⟩,
⟨{𝑎, 𝑏, 𝑐}, {𝑎, 𝑏, 𝑐, 𝑝, 𝑞, 𝑟 }⟩}
and 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, ∅⟩, ⟨{𝑏}, ∅⟩, ⟨{𝑐}, {𝑟 }⟩,
⟨{𝑎, 𝑏}, ∅⟩, ⟨{𝑎, 𝑐}, {𝑝, 𝑞, 𝑟 }⟩, ⟨{𝑏, 𝑐}, {𝑟 }⟩, ⟨{𝑎, 𝑏, 𝑐}, {𝑟 }⟩}
Let 𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑐.}. We get that,
𝑠𝑒𝑚(𝑅𝐵 3 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, {𝑎, 𝑝, 𝑞}⟩, ⟨{𝑏}, {𝑏, 𝑝, 𝑞}⟩, ⟨{𝑐}, {𝑐, 𝑟 }⟩,
⟨{𝑎, 𝑏}, {𝑎, 𝑏, 𝑝, 𝑞}⟩, ⟨{𝑎, 𝑐}, {𝑎, 𝑐, 𝑞, 𝑟 }⟩, ⟨{𝑏, 𝑐}, {𝑏, 𝑐, 𝑞, 𝑟 }⟩,
⟨{𝑎, 𝑏, 𝑐}, {𝑎, 𝑏, 𝑐, 𝑞, 𝑟 }⟩}
and 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, ∅⟩, ⟨{𝑏}, ∅⟩, ⟨{𝑐}, {𝑟 }⟩,
⟨{𝑎, 𝑏}, ∅⟩, ⟨{𝑎, 𝑐}, {𝑝, 𝑟 }⟩, ⟨{𝑏, 𝑐}, {𝑝, 𝑟 }⟩, ⟨{𝑎, 𝑏, 𝑐}, {𝑝, 𝑟 }⟩}
The difference of semantics reflects changes on a consequence
(both adding and removing) of two rule-bases. Now, we can define
a minimal revision of this framework as follows.
Definition 4.3 (Semantics-based Minimal Revision). Let 𝑅𝐵 1 , 𝑅𝐵 2 ,
𝑅𝐵 3 be three rule-bases. We say that 𝑅𝐵 2 has a smaller change than
𝑅𝐵 3 from 𝑅𝐵 1 denoted as 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) if for
every ⟨𝐹, 𝐷 2 ⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) and ⟨𝐹, 𝐷 3 ⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ),
then 𝐷 2 ⊆ 𝐷 3 . We define 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) < 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) if
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) but 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≰ 𝐷𝐼 𝐹 𝐹
(𝑅𝐵 1, 𝑅𝐵 2 ). We call 𝑅𝐵 2 a semantics-based minimal revision of 𝑅𝐵 1
if 𝑅𝐵 2 is a revision of 𝑅𝐵 1 and there is no revision 𝑅𝐵 ′ of 𝑅𝐵 1 such
that 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 ′ ) < 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ).
However, we consider Definition 4.3 is too strong, since it is hard
for comparing between two revisions from different schemes. For
example, 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) and 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) are incomparable
since
⟨{𝑎, 𝑐}, {𝑝, 𝑞, 𝑟 }⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 )
and ⟨{𝑎, 𝑐}, {𝑝, 𝑟 }⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 )
hence 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≰ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 )
and
On Semantics-based Minimal Revision for Legal Reasoning
⟨{𝑎, 𝑏, 𝑐}, {𝑝, 𝑟 }⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 )
and ⟨{𝑎, 𝑏, 𝑐}, {𝑟 }⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 )
hence 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≰ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ).
Therefore, we relax Definition 4.3 by considering partial seman-
tics defined as follows.
Definition 4.4 (Partial Semantics-based Minimal Revision). Let
𝑅𝐵 1 , 𝑅𝐵 2 , 𝑅𝐵 3 be three rule-bases and 𝑆 be a set of propositions. We
say that 𝑅𝐵 2 has a smaller change than 𝑅𝐵 3 from 𝑅𝐵 1 with respect
to 𝑆 denoted as 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) if for every
⟨𝐹, 𝐷 2 ⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) and ⟨𝐹, 𝐷 3 ⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ), then
𝐷 2 ∩ 𝑆 ⊆ 𝐷 3 ∩ 𝑆. We define 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) <𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 )
if 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) but 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≰𝑆
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ). We call 𝑅𝐵 2 a partial semantics-based minimal
revision of 𝑅𝐵 1 with respect to 𝑆 if 𝑅𝐵 2 is a revision of 𝑅𝐵 1 and there
is no revision 𝑅𝐵 ′ of 𝑅𝐵 1 such that 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 ′ ) <𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1,
𝑅𝐵 2 ).
From the previous example, if 𝑆 = {𝑝} or 𝑆 = {𝑝, 𝑟 }, we get that
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) but
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≰𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 )
hence 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) <𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ).
However, if 𝑆 = {𝑞}, we get that
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≤𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) but
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≰𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 )
hence 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) <𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ).
We define a partial difference of semantics as follows.
Definition 4.5 (Partial Difference of Semantics). Let F be a fact-
domain, 𝑅𝐵 1, 𝑅𝐵 2 be two rule-bases, and 𝑆 be a set of propositions. A
partial difference of semantics between 𝑅𝐵 1 and 𝑅𝐵 2 with respect to
𝑆 denoted by 𝐷𝐼 𝐹 𝐹𝑆 (𝑅𝐵 1, 𝑅𝐵 2 ) is a set defined as follows: {⟨𝐹, 𝐷 ∩
𝑆⟩|⟨𝐵, 𝐷⟩ ∈ 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 )}.
By this way, given any three rule-bases 𝑅𝐵 1 , 𝑅𝐵 2 , 𝑅𝐵 3 ,
𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤𝑆 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 3 ) ≡
𝐷𝐼 𝐹 𝐹𝑆 (𝑅𝐵 1, 𝑅𝐵 2 ) ≤ 𝐷𝐼 𝐹 𝐹𝑆 (𝑅𝐵 1, 𝑅𝐵 3 ).
Continuing from the previous example, 𝐷𝐼 𝐹 𝐹 {𝑝 } (𝑅𝐵 1, 𝑅𝐵 2 ) =
{⟨∅, ∅⟩, ⟨{𝑎}, ∅⟩, ⟨{𝑏}, ∅⟩, ⟨{𝑐}, ∅⟩,
⟨{𝑎, 𝑏}, ∅⟩, ⟨{𝑎, 𝑐}, {𝑝}⟩, ⟨{𝑏, 𝑐}, ∅⟩, ⟨{𝑎, 𝑏, 𝑐}, ∅⟩}
When a fact domain is {𝑎, 𝑏, 𝑐}, we get that 𝑅𝐵 2 is a partial
semantics-based minimal revision with respect to {𝑝}. This is be-
cause to find a rule-base 𝑅𝐵 ′ such that 𝐷𝐼 𝐹 𝐹 (𝑅𝐵 1, 𝑅𝐵 ′ ) <𝑆 𝐷𝐼 𝐹 𝐹
(𝑅𝐵 1, 𝑅𝐵 2 ), 𝐷𝐼 𝐹 𝐹 {𝑝 } (𝑅𝐵 1, 𝑅𝐵 ′ ) must be {⟨𝐹, ∅⟩|𝐹 ⊆ F }, then 𝑅𝐵 ′
is not a resolution to the CCR task since the exceptional case is
𝐹 𝐵 1 = {𝑎. 𝑐.} but 𝐹 𝐵 1 ∪ 𝑅𝐵 agrees with 𝐹 𝐵 1 ∪ 𝑅𝐵 ′ on 𝑝 (the coun-
terintuitive consequence is not resolved). However, if we extend a
fact domain to F = {𝑎, 𝑏, 𝑐, 𝑑 }, 𝐷𝐼 𝐹 𝐹 {𝑝 } (𝑅𝐵 1, 𝑅𝐵 2 ) would become:
{⟨∅, ∅⟩, ⟨{𝑎}, ∅⟩, ⟨{𝑏}, ∅⟩, ⟨{𝑐}, ∅⟩, ⟨{𝑑 }, ∅⟩,
⟨{𝑎, 𝑏}, ∅⟩, ⟨{𝑎, 𝑐}, {𝑝}⟩, ⟨{𝑎, 𝑑 }, ∅⟩, ⟨{𝑏, 𝑐}, ∅⟩,
. and 𝐴𝐷 1 ⊆ 𝐴𝐷 3 . If ⟨𝐴𝐷 1, 𝐴𝐷 2 ⟩ ∈ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) then
⟨{𝑏, 𝑑 }, ∅⟩, ⟨{𝑐, 𝑑 }, ∅⟩, ⟨{𝑎, 𝑏, 𝑐}, ∅⟩, ⟨{𝑎, 𝑏, 𝑑 }, ∅⟩,
⟨{𝑎, 𝑐, 𝑑 }, {𝑝}⟩, ⟨{𝑏, 𝑐, 𝑑 }, ∅⟩, ⟨{𝑎, 𝑏, 𝑐, 𝑑 }, ∅⟩}
Thus, 𝑅𝐵 2 is not a partial semantics-based minimal revision of
𝑅𝐵 1 with respect to {𝑝} when a fact-domain is {𝑎, 𝑏, 𝑐, 𝑑 } because
there exists a revision 𝑅𝐵 ′ such that 𝐷𝐼 𝐹 𝐹 {𝑝 } (𝑅𝐵 1, 𝑅𝐵 ′ ) = {⟨𝐹, 𝐷⟩|
𝐹 ⊆ F if 𝐹 = {𝑎, 𝑐} then 𝐷 = {𝑝} otherwise 𝐷 = ∅}. However,
we consider a revision 𝑅𝐵 ′ is too specific since it requires only
54
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
semantics changes for {𝑎, 𝑐}. Therefore, to describe that 𝑅𝐵 2 is a
suitable semantics-based minimal revision, we say each fact-base
has its corresponding specific rule-base and dominant rule-base
with respect to the considered consequence. We adopt the definition
of a specific rule-base from [30] and define a dominant rule-base as
a trimmed version of a specific rule-base as follows.
Definition 4.6 (Specific Rule-base and Dominant Rule-base). Let
𝑅𝐵 be a rule-base, 𝐹 𝐵 be a fact-base, and 𝑝 be a proposition. We say
a rule-base 𝑆𝑅 ⊆ 𝑅𝐵 is specific to 𝐹 𝐵 with respect to 𝑅𝐵 and 𝑝 if 𝑆𝑅
is a minimal set of rules (in the sense of set inclusion) such that
• 𝐹 𝐵 ∪ 𝑆𝑅 agrees with 𝐹 𝐵 ∪ 𝑅𝐵 on 𝑝, and
• no rule-base 𝑆𝑅 ′ such that 𝑆𝑅 ⊊ 𝑆𝑅 ′ ⊊ 𝑅𝐵 and 𝐹 𝐵 ∪ 𝑆𝑅 ′
disagrees with 𝐹 𝐵 ∪ 𝑆𝑅 on 𝑝.
We call 𝐷𝑅 = 𝑡𝑟𝑖𝑚(𝑆𝑅) a dominant rule-base of 𝐹 𝐵 with respect to
𝑅𝐵 and 𝑝 (𝑡𝑟𝑖𝑚 is defined in Definition 3.3).
Now, we define dominant-based semantics as follows.
Definition 4.7 (Dominant-based Semantics). Let F be a fact-domain,
𝑅𝐵 be a rule-base, and 𝑝 be a proposition. The dominant-based se-
mantics of 𝑅𝐵 with respect to 𝑝 is a set of pairs defined as follows:
{⟨𝐹 𝐵, 𝐴𝐷⟩|𝐹 𝐵 is a fact-base constructed from a subset of F and 𝐴𝐷
is a set of all dominant rule-bases of 𝐹 𝐵 with respect to 𝑅𝐵 and 𝑝}.
We denote this set as 𝑑𝑜𝑚𝑝 (𝑅𝐵).
Then, we define the difference of dominant-based semantics as
follows.
Definition 4.8 (Difference of Dominant-based Semantics). Let F be
a fact-domain, 𝑅𝐵 1, 𝑅𝐵 2 be two rule-bases, and 𝑝 be a proposition.
The difference of dominant-based semantics between 𝑅𝐵 1 and 𝑅𝐵 2
with respect to 𝑝 denoted by 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) is a set defined
as follows: {⟨𝐴𝐷 1, 𝐴𝐷 2 ⟩|⟨𝐹 𝐵, 𝐴𝐷 1 ⟩ ∈ 𝑑𝑜𝑚𝑝 (𝑅𝐵 1 ) and
⟨𝐹 𝐵, 𝐴𝐷 2 ⟩ ∈ 𝑑𝑜𝑚𝑝 (𝑅𝐵 2 ) and 𝐴𝐷 1 ≠ 𝐴𝐷 2 }.
The difference in this definition represents patterns of fact-bases
that move their dominant rule-bases after revising a program, re-
gardless how many fact-bases move such patterns. From Table 1
continuing from the previous example, 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) =
{⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}}, {{𝑟 ← 𝑐.}}⟩,
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, {{𝑝 ← 𝑞. 𝑞 ← 𝑏.}}⟩}
From Table 2, 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 ) =
{⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}}, {{𝑟 ← 𝑐.}}⟩,
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, {{𝑟 ← 𝑐.}}⟩,
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, {{𝑟 ← 𝑐.}}⟩}
The difference of dominant-based semantics does not change
when we extend the fact domain. Due to the ordering of dominants,
we get the following lemma.
Lemma 4.9. Let 𝑅𝐵 1 , 𝑅𝐵 2 be two rule-bases, 𝑝 be a proposition
and ⟨𝐹 𝐵 1, 𝐴𝐷 1 ⟩, ⟨𝐹 𝐵 3, 𝐴𝐷 3 ⟩ ∈ 𝑑𝑜𝑚𝑝 (𝑅𝐵 1 ) such that 𝐹 𝐵 1 ⊆ 𝐹 𝐵 3
⟨𝐴𝐷 3, 𝐴𝐷 4 ⟩ ∈ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ).
This lemma intuitively means if a revision affects a set 𝐴𝐷 1 of
all dominant rule-bases of 𝐹 𝐵 1 , the revision also affects every set
𝐴𝐷 3 of all dominant rule-bases of 𝐹 𝐵 3 such that 𝐹 𝐵 3 is a super
set of 𝐹 𝐵 1 and 𝐴𝐷 3 is a super set of 𝐴𝐷 1 . Now, we can define the
dominant-based minimal revision as follows.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh

Table 1: Difference of dominant-based semantics between 𝑅𝐵 1 and 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.} in the example

Possible fact-base 𝐹 𝐵 Dominant rule-base(s) of 𝐹 𝐵 (before revision) Dominant rule-base(s) of 𝐹 𝐵 (after revision)
∅ ∅ ∅
{𝑎.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}
{𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.}
{𝑐.} ∅ ∅
{𝑎. 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}
{𝑎. 𝑐.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.} {r ← c.}
{𝑏. 𝑐.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.}
{𝑎. 𝑏. 𝑐} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.}

Table 2: Difference of dominant-based semantics between 𝑅𝐵 1 and 𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑐.} in the example

Possible fact-base 𝐹 𝐵 Dominant rule-base(s) of 𝐹 𝐵 (before revision) Dominant rule-base(s) of 𝐹 𝐵 (after revision)
∅ ∅ ∅
{𝑎.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}
{𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.}
{𝑐.} ∅ ∅
{𝑎. 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}
{𝑎. 𝑐.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.} {r ← c.}
{𝑏. 𝑐.} {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {r ← c.}
{𝑎. 𝑏. 𝑐.} {𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.} {r ← c.}

Definition 4.10 (Dominant-based Minimal Revision). Let 𝑅𝐵 1 , 𝑅𝐵 2 ,
𝑅𝐵 3 be three rule-bases and 𝑝 be a proposition. We say that 𝑅𝐵 2
has a smaller dominant change than 𝑅𝐵 3 from 𝑅𝐵 1 with respect
to 𝑝 denoted as 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) ⪯ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 )
when this following condition is satisfied.
If ⟨𝐴𝐷 1, 𝐴𝐷 2 ⟩ ∈ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ),
then ⟨𝐴𝐷 1, 𝐴𝐷 3 ⟩ ∈ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 ).
We define 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) ≺ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 ) if
𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) ⪯ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 ) but
𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 3 ) ⪯̸ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ).
We call 𝑅𝐵 2 a dominant-based minimal revision of 𝑅𝐵 1 with respect
to 𝑝 if 𝑅𝐵 2 is a revision of 𝑅𝐵 1 and there is no revision 𝑅𝐵 ′ of 𝑅𝐵 1
such that 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 ′ ) ≺ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ).
From the previous example, 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ) ≺ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝
(𝑅𝐵 1, 𝑅𝐵 3 ) so 𝑅𝐵 3 is not the dominant-based minimal revision of
𝑅𝐵 1 respect to 𝑝. We also get that 𝑅𝐵 2 is the dominant-based
minimal revision respect to 𝑝. This because to find a revision
𝑅𝐵 ′ such that 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 ′ ) ≺ 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 2 ),
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, 𝐴𝐷 ′ ⟩ must be not in
𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 ′ ). However, since a revision 𝑅𝐵 ′ must affect
{𝑝 ← 𝑞. 𝑞 ← 𝑎.} as it is the dominant rule-base of {𝑎. 𝑐.} (the
exceptional case), ⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, 𝐴𝐷 ′ ⟩
must be in 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 ′ ) according to Lemma 4.9. Hence,
such revision 𝑅𝐵 ′ does not exist and 𝑅𝐵 2 is the dominant-based
minimal revision of 𝑅𝐵 1 .
5 OBTAINING MINIMAL REVISION
One question is that, given a CCR task, can we always obtain a
dominant-based minimal revision with respect to the considered
counterintuitive consequence ? The answer to this question is yes.
One way is to follow a culprit resolution in Algorithm 1 and in-
troduce all fact propositions occurring in an exceptional case as
a body of each new rule so that all new rules are applicable only
to the fact-bases that cover all facts in the considered exceptional
case.
From the example in previous sections, let 𝑅𝐵 1 = {𝑝 ← 𝑞. 𝑞 ←
𝑎. 𝑞 ← 𝑏.}. If a culprit is 𝑝 and we get a preliminary revision
g1 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏.} and 𝐻 = {𝑟 }. If we add all
𝑅𝐵
fact propositions in 𝐹 𝐵 1 = {𝑎. 𝑐.} to a body of a new rule, we get a
culprit resolution 𝑅𝐵 4 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑎, 𝑐.},
which is a dominant-based minimal revision of 𝑅𝐵 1 with respect to
𝑝 since 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 4 ) =
{⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}}, {{𝑟 ← 𝑎, 𝑐.}}⟩,
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, {{𝑟 ← 𝑎, 𝑐.}}⟩}
and there is no revision 𝑅𝐵 ′ such that 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 ′ ) ≺
𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 4 ) in the same manner of 𝑅𝐵 2 . This also ap-
plies to a scenario when a culprit is 𝑞, 𝑅𝐵 5 = {𝑝 ← 𝑞. 𝑞 ←
𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑎, 𝑐.} is also a dominant-based minimal revi-
sion of 𝑅𝐵 1 since 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝑅𝐵 1, 𝑅𝐵 5 ) =
{⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}}, {{𝑟 ← 𝑎, 𝑐.}}⟩,
⟨{{𝑝 ← 𝑞. 𝑞 ← 𝑎.}, {𝑝 ← 𝑞. 𝑞 ← 𝑏.}}, {{𝑝 ← 𝑞. 𝑞 ← 𝑏.}}⟩}.
In addition, we can optimize 𝑅𝐵 5 into 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ←
𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.} which is also a dominant-based minimal
revision. To optimize the revision like we do from 𝑅𝐵 5 to 𝑅𝐵 2 , we
need to consider specific rule-bases with respect to the original rule-
base and the considering consequence as the following theorem.
Theorem 5.1. Given a fact-domain F , a CCR task ⟨𝑅𝐵, 𝐹 𝐵𝑒 , 𝑝⟩,
f 𝐻 obtained by Algorithm 1 where 𝑅𝐵
𝑅𝐵, f is a preliminary revision, and

55
On Semantics-based Minimal Revision for Legal Reasoning ICAIL’21, June 21–25, 2021, São Paulo, Brazil

𝐻 = {ℎ 1, . . . , ℎ𝑛 } as a set of rule propositions that would become heads 𝑏. 𝑟 ← 𝑐.}. Although 𝑅𝐵 3 is the syntax-based minimal revision
of new rules, and a set of Horn clauses 𝑅𝐵𝑛 = {ℎ 1 ← 𝐵 1 ., . . . , ℎ𝑛 ← when we require to add a condition to a new rule, 𝑅𝐵 3 is not a
𝐵𝑛 .} where 𝐵𝑖 ⊆ 𝑓 (𝐹 𝐵𝑒 ) for all 1 ≤ 𝑖 ≤ 𝑛. If 𝐵𝑖 satisfies the following dominant-based minimal revision as we illustrated in the previous
condition for all 1 ≤ 𝑖 ≤ 𝑛: section.
For every fact-base 𝐹 𝐵 constructed from a subset of F Generally, we get that a syntax-based minimal revision is not
such that 𝐵 ⊆ 𝐹 𝐵, if there is a rule 𝑅 in a specific rule- always a dominant-based minimal revision especially when the
base of 𝐹 𝐵 with respect to 𝑅𝐵 and 𝑝, and ℎ𝑖 occurs in 𝑅 rule-base contains multiple rules for the same consequence. To
then 𝐹 𝐵𝑒 ⊆ 𝐹 𝐵, illustrate more on this constraint, we use an example case adopted
then, a culprit resolution 𝑅𝐵 ′ = 𝑅𝐵 f ∪ {ℎ 1 ← 𝐵 1 ., . . . , ℎ𝑛 ← 𝐵𝑛 .} is from a Supreme Court Case as in [4, 17, 38] relating to Japanese
a dominant-based minimal revision of 𝑅𝐵 with respect to 𝑝. Civil Code Article 612, which states
Phrase 1: A Lessee may not assign the lessee’s rights
Theorem 5.1 intuitively implies a method for finding dominant-
or sublease a leased thing without obtaining the ap-
based minimal revision. Let 𝐹 𝐵𝑒 be an exceptional case, the method
proval of the lessor.
is to find a minimal set 𝐵 ⊆ 𝑓 (𝐹 𝐵𝑒 ) which affects specific rule-
Phrase 2: If the lessee allows any third party to make
bases only of 𝐹 𝐵 ⊇ 𝐹 𝐵𝑒 . Thus, when 𝐵 = 𝑓 (𝐹 𝐵𝑒 ), the revision is
use of or take profits from a leased thing in violation of
a dominant-based minimal revision. However, if 𝐵 ⊊ 𝑓 (𝐹 𝐵𝑒 ), a
the provisions of the preceding paragraph, the lessor
revision affects specific rule-bases only of 𝐹 𝐵 ⊇ 𝐹 𝐵𝑒 if there is no
may cancel the contract.
rule 𝑅 in a specific rule-base of 𝐹 𝐵 such that 𝐹 𝐵 ⊇ 𝐵, 𝐹 𝐵 ⊉ 𝐹 𝐵𝑒
and 𝑝𝑒 occurs in 𝑅. For example, regarding 𝑅𝐵 2 = {𝑝 ← 𝑞. 𝑞 ← Japanese Civil Code can be represented with the rule-base as
𝑎, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑏. 𝑟 ← 𝑐.}, there is no 𝐹 𝐵 such that 𝐹 𝐵 ⊇ {𝑐.}, follows1 . Hereafter, this rule-base is denoted by 𝐽 𝑅𝐵 1 .
𝐹 𝐵 ⊉ {𝑎. 𝑐.}, and 𝑞 ← 𝑎, 𝑛𝑜𝑡 𝑟 . is in the specific rule of 𝐹 𝐵. However, 1 cancellation_due_to_sublease :-
regarding 𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏. 𝑟 ← 𝑐.}, there is 2 effective_lease_contract,
𝐹 𝐵 = {𝑏. 𝑐.} such that 𝐹 𝐵 ⊇ {𝑐.}, 𝐹 𝐵 ⊉ {𝑎. 𝑐.}, and 𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . is 3 effective_sublease_contract,
in the specific rule of 𝐹 𝐵. Therefore, 𝑅𝐵 5 can be optimized to 𝑅𝐵 2 4 using_leased_thing,
but 𝑅𝐵 4 cannot be optimized to 𝑅𝐵 3 . 5 manifestation_cancellation,
6 not approval_of_sublease.
6 COMPARISON AND EXAMPLE 7 effective_lease_contract :-
In this section, we compare the dominant-based minimal revision 8 agreement_of_lease_contract,
with the syntax-based minimal revision in Theory Distance Metric 9 handover_to_lessee.
[47], which is one common minimal revision used for describing 10 effective_sublease_contract :-
minimal revision in legislation (e.g. [12, 28]). The definition is for- 11 agreement_of_sublease_contract,
mally described in our context as follows. 12 handover_to_sublessee.
13 effective_sublease_contract :-
Definition 6.1 (Syntax-based Minimal Revision). Let 𝑅𝐵 and 𝑅𝐵 ′
14 implicit_sublease.
be rule-bases. A revision transformation 𝑟 is such that 𝑟 (𝑅𝐵) =
15 approval_of_sublease :-
𝑅𝐵 ′ , and 𝑅𝐵 ′ is obtained from 𝑅𝐵 by program edit operations as
16 approval_before_the_day.
follows: deleting a rule, creating a rule with an empty body, adding
a condition to a rule in 𝑅𝐵 or deleting a condition from a rule in 𝑅𝐵. This representation illustrates that to prove the contract was
𝑅𝐵 ′ is a revision of 𝑅𝐵 with distance 𝑐 (𝑅𝐵, 𝑅𝐵 ′ ) = 𝑛 if and only if ended due to sublease (represented as cancellation_due_to_
𝑅𝐵 ′ = 𝑟 𝑛 (𝑅𝐵) and there is no 𝑚 < 𝑛 such that 𝑅𝐵 ′ = 𝑟 𝑛 (𝑅𝐵) [47]. sublease), we must prove four requisites (lines 2-5)
Consequently, it is very simple to find a culprit resolution that is (1) the lease contract was effective
also a syntax-based minimal revision. Actually, a culprit resolution (represented as effective_lease_contract)
with an empty condition for each new rule is a syntax-based mini- (2) the sublease contract was effective
mal revision since it requires no additional program edit operation, (represented as effective_sublease_contract)
but that kind of culprit resolution can be considered as too general. (3) the third party was using the leased thing
If we require to add some condition to a new rule, we can just add (represented as using_leased_thing)
an extra fact proposition in the exceptional case as a condition of a (4) the plaintiff manifested the intention of cancellation of the
new rule (an extra fact proposition means a fact proposition that contract (represented as manifestation_cancellation)
does not occur in the rule-base but occurs in the exceptional case). And there is one exception, approval_of_sublease in line 5,
This requires only one program edit operation for each new rule so which is explicitly stated in the Civil Code. To prove the exception,
the culprit resolution is definitely the syntax-based minimal revi- we must prove that the approval before the cancellation (repre-
sion (under the constraint that we require to add some condition sented as approval_before_the_day).
to each new rule). To prove that the lease contract was effective (effective_lease
From Example 3.9, we have 𝑅𝐵 1 = {𝑝 ← 𝑞. 𝑞 ← 𝑎. 𝑞 ← 𝑏.}, _contract), we must prove two requisites (lines 8-9).
if a culprit is 𝑝 and we get a preliminary revision 𝑅𝐵 g1 = {𝑝 ←
𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← 𝑏.}. If we add an extra fact proposition 𝑐 to 1 This representation is adopted for ease of exposition. The implicit sublease contract
a body of a new rule, we get 𝑅𝐵 3 = {𝑝 ← 𝑞, 𝑛𝑜𝑡 𝑟 . 𝑞 ← 𝑎. 𝑞 ← is a fictitious condition to illustrate multiple conditions. We use :- instead of ←.

56
 ICAIL’21, June 21–25, 2021, São Paulo, Brazil Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh

(1) the lease contract was established 16 approval_of_sublease :-

(represented as agreement_of_lease_contract) 17 approval_before_the_day.
(2) the leased thing was handed over to the lessee (represented 18 new_exception :-
as handover_to_lessee) 19 agreement_of_sublease_contract,
There are two ways to prove that sublease contract was effec- 20 handover_to_sublessee,
tive (effective_sublease_contract), one way is to prove two 21 non_destruction_of_confidence.
requisites (lines 11-12). On the other hand, the syntax-based minimal revision is 𝐽 𝑅𝐵 3 =
(1) the sublease contract was established 1 cancellation_due_to_sublease :-
(represented as agreement_of_sublease_contract) 2 effective_lease_contract,
(2) the leased thing was handed over to the sublessee (repre- 3 effective_sublease_contract,
sented as handover_to_sublessee) 4 using_leased_thing,
Another way is to show this sublease is implicit (represented as 5 manifestation_cancellation,
implicit_sublease in line 14). 6 not approval_of_sublease,
When an exceptional case related to Japanese Civil Code Article 7 not new_exception.
612 went to Japanese Supreme Court, the judges have once revised 8 effective_lease_contract :-
this article as stated [1, 4]. 9 agreement_of_lease_contract,
[Japanese Civil Code Article 612] Phrase 2 is not ap- 10 handover_to_lessee.
plicable in exceptional situations where the sublease 11 effective_sublease_contract :-
does not harm the confidence between a lessee and a 12 agreement_of_sublease_contract,
lessor, and therefore the lessor cannot cancel the con- 13 handover_to_sublessee.
tract unless the lessor proves the lessee’s destruction 14 effective_sublease_contract :-
of confidence. 15 implicit_sublease.
16 approval_of_sublease :-
In this court decision, the judge introduced a factual concept of
17 approval_before_the_day.
non-destruction of confidence (represented as non_destruction_
18 new_exception :-
of_confidence) as an exception of Phrase 2 to prevent the coun-
19 non_destruction_of_confidence.
terintuitive consequence from the literal interpretation of Japanese
Civil Code Article 612. The case that goes to the Supreme Court From this example, we get that 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝐽 𝑅𝐵 1, 𝐽 𝑅𝐵 2 ) ≺
can be represented as the following set of fact-base 𝐽 𝐹 𝐵 1 = 𝐷𝑂𝑀𝐷𝐼 𝐹 𝐹𝑝 (𝐽 𝑅𝐵 1, 𝐽 𝑅𝐵 3 ) where 𝑝 is cancellation_due_to_
sublease. We get that the syntax-based minimal revision affects
{agreement_of_lease_contract.
the specific rule-base of 𝐽 𝐹 𝐵 2 =
handover_to_lessee.
agreement_of_sublease_contract. {agreement_of_lease_contract,
handover_to_sublessee. handover_to_lessee,
using_leased_thing. implicit_sublease,
manifestation_cancellation. using_leased_thing,
non_destruction_of_confidence.} manifestation_cancellation,
non_destruction_of_confidence}
Suppose we revise the rule-base 𝐽 𝑅𝐵 1 with Legal Debugging
and Closed World Specification as described in Section 3 and it In this way, we can consider such effects may be unintentionally
concludes that cancellation_due_to_sublease is a culprit. Then, caused by the syntax-based minimal revision and the system may
a dominant-based minimal revision according to our method in confirm the intention of such effects with a user.
Section 5 is 𝐽 𝑅𝐵 2 =
1 cancellation_due_to_sublease :- 7 DISCUSSION
2 effective_lease_contract, In this paper, we compare our minimal revision to the syntax-based
3 effective_sublease_contract, minimal revision in Theory Distance Metric [47]. In their study, it
4 using_leased_thing, has been discussed that two programs may be close to one another
5 manifestation_cancellation, syntactically but have entirely different semantics. Our study also
6 not approval_of_sublease, shows in line with this discussion. In this paper, we consider a
7 not new_exception. revision task as a task to find appropriate conditions to a body of a
8 effective_lease_contract :- new rule after applying Legal Debugging [17] and Closed World
9 agreement_of_lease_contract, Specification [6] as described in Section 3. Figure 1 illustrates a
10 handover_to_lessee. comparison between each type of revision discussed in this paper.
11 effective_sublease_contract :- Let 𝐵 be a set of conditions to a body of a new rule and 𝐹 𝐵𝑒 be an
12 agreement_of_sublease_contract, exceptional case. Regarding Theorem 5.1, if 𝐵 = 𝑓 (𝐹 𝐵𝑒 ), that is all
13 handover_to_sublessee. fact propositions in the exceptional case are conditions to a body of
14 effective_sublease_contract :- a new rule, then such a revision has the smallest dominant changes
15 implicit_sublease. but it requires the largest number of program edit operation (Point

57
On Semantics-based Minimal Revision for Legal Reasoning
Figure 1: A comparison between each type of revision dis-
cussed in this paper
(a) in the figure). In Section 5, we discuss briefly that we can find a
dominant-based minimal revision with an optimization of program
edit operations by finding a minimal set (in the sense of size) that
satisfies the condition in Theorem 5.1 (Point (b) in the figure, may
be the same as Point (a) for some rule-bases). If 𝐵 goes smaller
than that point, the revision will have larger dominant changes but
requires less program edit operations so line (b)-(c) in the figure
shows a decrement function. Regarding the syntax-based minimal
revision, if it requires a new condition to a new rule, we can obtain
the syntax-based minimal revision by making 𝐵 a singleton set
of extra fact proposition in 𝐹 𝐵𝑒 (Point (c) in the figure) since it
requires only one program edit operation. Otherwise, we can do no
revision after applying the Closed World Specification algorithm
then it requires zero operation but it definitely has larger dominant
changes than introducing an extra fact proposition (Point (d) in the
figure).
Our paper follows the standard revision theory since we attempt
to adjust rule-based theories as little as possible. As discussed in
[35], there are two strategies to minimally revise rule-based theo-
ries, one is keeping the revision as close as possible to the original
and another is minimizing the changes of semantics (the set of
conclusions obtained from the theory), which vary among cases.
Although the dominant-based minimal revision follows the second
strategy, we also present a guidance to obtain a dominant-based
minimal revision but keep the revision as close as possible to the
original so the guidance takes both strategies into account. How-
ever, judges may intentionally make a minimal change based on
the first strategy rather than the second strategy. In that scenario,
we think that our study of the semantics-based revision is benefi-
cial for reminding unintentional effects that may be missed when
considering revisions by merely the first strategy. This may lead to
Impact Analysis [9] (the study of analyzing the impacts of changes,
typically ones not known in advance ) in legal formalisms and may
contribute to improving coherence of laws [35].
8 CONCLUSION AND FUTURE WORK
In this paper, we investigate definitions of semantics-based minimal
revision for legal reasoning. We present definitions of semantics
of a rulebase, the difference of semantics and the semantics-based
58
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
minimal revision. These definitions were adopted for the context
of legal reasoning in normal logic programs and legal change by
judges when the literal interpretation of statutes leads to counter-
intuitive consequences. Since we consider such definitions are too
strong, we have defined a partial semantics-based minimal revision
and a dominant-based minimal revision. The dominant-based mini-
mal revision intuitively means a revision that affects the specific
rule-bases of the fact-bases only that are larger or equal to the con-
sidered exceptional case. Then, we present one guidance to obtain a
dominant-based minimal revision and also optimize the number of
edit operations. The comparison shows that a syntax-based minimal
revision may affect more specific rule-bases than a dominant-based
minimal revision hence a syntax-based minimal revision is not a
dominant-based minimal revision in general. Since this paper only
gives preliminary definitions for semantics-based and dominant-
based minimal revision, future research might investigate, for ex-
ample, an algorithm effectively finding a dominant-based minimal
revision, an extension of definitions for other types of revisions or
other legal representations, or the application for cases where the
revision impacts several rules, which may lead to impact analysis
in legal formalisms and may contribute to improving coherence of
laws.
ACKNOWLEDGMENTS
We would like to thank anonymous reviewers for their extensive
comments and suggestions. This work was supported by JSPS KAK-
ENHI Grant Numbers, JP17H06103 and JP19H05470 and JST, AIP
Trilateral AI Research, Grant Number JPMJCR20G4, Japan
REFERENCES
[1] Tokyo High Court 1994 (O) 693. 1996. Case to seek removal of a building and
surrender of lands. http://www.courts.go.jp/app/hanrei_en/detail?id=273
[2] Latifa Al-Abdulkarim, Katie Atkinson, and Trevor Bench-Capon. 2016. Accom-
modating change. Artificial Intelligence and Law 24, 4 (2016), 409–427.
[3] Krzysztof R Apt, Howard A Blair, and Adrian Walker. 1988. Towards a the-
ory of declarative knowledge. In Foundations of deductive databases and logic
programming. Morgan Kaufmann, Burlington, MA,USA, 89–148.
[4] Ryuta Arisaka. 2015. A Belief Revision Technique to Model Civil Code Updates.
In JSAI International Symposium on Artificial Intelligence. Springer International
Publishing, Cham, 204–216.
[5] Duangtida Athakravi, Ken Satoh, Mark Law, Krysia Broda, and Alessandra Russo.
2015. Automated inference of rules with exception from past legal cases using ASP.
In International Conference on Logic Programming and Nonmonotonic Reasoning.
Springer International Publishing, Cham, 83–96.
[6] Michael Bain and Stephen Muggleton. 1992. Non-monotonic learning. Inductive
logic programming 38 (1992), 145153.
[7] Trevor JM Bench-Capon. 2017. Hypo’s legacy: introduction to the virtual special
issue. Artificial Intelligence and Law 25, 2 (2017), 205–250.
[8] Donald H. Berman and Carole D. Hafner. 1995. Understanding Precedents in
a Temporal Context of Evolving Legal Doctrine. In Proceedings of the 5th Inter-
national Conference on Artificial Intelligence and Law (College Park, Maryland,
USA) (ICAIL ’95). Association for Computing Machinery, New York, NY, USA,
42–51. https://doi.org/10.1145/222092.222116
[9] Bohner. 1996. Impact analysis in the software change process: a year 2000 per-
spective. In 1996 Proceedings of International Conference on Software Maintenance.
IEEE, USA, 42–51.
[10] L Karl Branting. 1991. Building explanations from rules and structured cases.
International journal of man-machine studies 34, 6 (1991), 797–837.
[11] Alison Chorley and Trevor Bench-Capon. 2005. AGATHA: Using heuristic search
to automate the construction of case law theories. Artificial Intelligence and Law
13, 1 (2005), 9–51.
[12] Domenico Corapi, Alessandra Russo, Marina De Vos, Julian Padget, and Ken
Satoh. 2011. Normative design using inductive learning. Theory and Practice of
Logic Programming 11, 4-5 (2011), 783–799.
[13] Kristijonas Cyras, Ken Satoh, and Francesca Toni. 2016. Abstract argumentation
for case-based reasoning. In Fifteenth International Conference on the Principles of
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Knowledge Representation and Reasoning. AAAI Press, CA, USA, 243–254.
[14] Mukesh Dalal. 1988. Investigations into a theory of knowledge base revision:
preliminary report. In Proceedings of the Seventh National Conference on Artificial
Intelligence, Vol. 2. AAAI Press, CA, USA, 475–479.
[15] Marina De Vos, Julian Padget, and Ken Satoh. 2010. Legal modelling and reason-
ing using institutions. In JSAI International Symposium on Artificial Intelligence.
Springer Berlin Heidelberg, Berlin, Heidelberg, 129–140.
[16] Alvaro Del Val. 1997. Non monotonic reasoning and belief revision: syntactic,
semantic, foundational and coherence approaches. Journal of Applied Non-
Classical Logics 7, 1-2 (1997), 213–240.
[17] Wachara Fungwacharakorn and Ken Satoh. 2018. Legal Debugging in Proposi-
tional Legal Representation. In JSAI International Symposium on Artificial Intelli-
gence. Springer International Publishing, Cham, 146–159.
[18] Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh. 2020. On the legal
revision in PROLEG program. In 34th Proceedings of the Annual Conference of JSAI,
Vol. 2020. Japan Society of Artificial Intelligence, Japan, 3G5ES104–3G5ES104.
[19] Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh. 2021. Resolving
counterintuitive consequences in law using legal debugging. Artificial Intelligence
and Law (2021), 1–17.
[20] Michael Gelfond and Vladimir Lifschitz. 1988. The stable model semantics for
logic programming. In Proceedings of International Logic Programming Conference
and Symposium, Robert Kowalski, Bowen, and Kenneth (Eds.), Vol. 88. MIT Press,
Cambridge, MA, USA, 1070–1080.
[21] Guido Governatori, Michael J Maher, Grigoris Antoniou, and David Billington.
2004. Argumentation semantics for defeasible logic. Journal of Logic and Compu-
tation 14, 5 (2004), 675–702.
[22] Guido Governatori, Monica Palmirani, Regis Riveret, Antonio Rotolo, and Gio-
vanni Sartor. 2005. Norm Modifications in Defeasible Logic. In Proceedings of the
2005 Conference on Legal Knowledge and Information Systems: JURIX 2005: The
Eighteenth Annual Conference. IOS Press, Amsterdam, The Netherlands, 13–22.
[23] Carole D Hafner and Donald H Berman. 2002. The role of context in case-based
legal reasoning: teleological, temporal, and procedural. Artificial Intelligence and
Law 10, 1 (2002), 19–64.
[24] John Henderson and Trevor Bench-Capon. 2019. Describing the development of
case law. In Proceedings of the Seventeenth International Conference on Artificial
Intelligence and Law (ICAIL ’19). Association for Computing Machinery, New
York, NY, USA, 32–41.
[25] John F Horty and Trevor JM Bench-Capon. 2012. A factor-based definition of
precedential constraint. Artificial intelligence and Law 20, 2 (2012), 181–214.
[26] Hirofumi Katsuno and Alberto O Mendelzon. 1991. Propositional knowledge
base revision and minimal change. Artificial Intelligence 52, 3 (1991), 263–294.
[27] Edward H Levi. 2013. An introduction to legal reasoning. University of Chicago
Press, Chicago, USA.
[28] Tingting Li, Tina Balke, Marina De Vos, Julian Padget, and Ken Satoh. 2013. A
model-based approach to the automatic revision of secondary legislation. In
Proceedings of the Fourteenth International Conference on Artificial Intelligence
and Law (Rome, Italy) (ICAIL ’13). Association for Computing Machinery, New
York, NY, USA, 202–206.
[29] Bernhard Nebel et al. 1992. Syntax-based approaches to belief revision. Belief
revision 29 (1992), 52–88.
[30] Henry Prakken. 1991. A tool in modelling disagreement in law: preferring the
most specific argument. In Proceedings of the 3rd international conference on
Artificial intelligence and law. Association for Computing Machinery, New York,
NY, USA, 165–174.
59
Wachara Fungwacharakorn, Kanae Tsushima, and Ken Satoh
[31] Henry Prakken. 2021. A formal analysis of some factor-and precedent-based
accounts of precedential constraint. Artificial Intelligence and Law (2021), 1–27.
[32] Adam Rigoni. 2015. An improved factor based approach to precedential constraint.
Artificial Intelligence and Law 23, 2 (2015), 133–160.
[33] Edwina L Rissland and Kevin D Ashley. 1987. A case-based system for trade secrets
law. In Proceedings of the 1st international conference on Artificial intelligence and
law. Association for Computing Machinery, New York, NY, USA, 60–66.
[34] Edwina L Rissland and M Timur Friedman. 1995. Detecting change in legal
concepts. In Proceedings of the 5th international conference on Artificial intelligence
and law. Association for Computing Machinery, New York, NY, USA, 127–136.
[35] Antonino Rotolo and Corrado Roversi. 2013. Constitutive rules and coherence in
legal argumentation: The case of extensive and restrictive interpretation. In Legal
Argumentation Theory: Cross-Disciplinary Perspectives. Springer Netherlands,
Dordrecht, The Netherlands, 163–188.
[36] Ken Satoh. 1988. Nonmonotonic reasoning by minimal belief revision. Institute for
New Generation Computer Technology, Japan.
[37] Ken Satoh, Kento Asai, Takamune Kogawa, Masahiro Kubota, Megumi Naka-
mura, Yoshiaki Nishigai, Kei Shirakawa, and Chiaki Takano. 2011. PROLEG: An
Implementation of the Presupposed Ultimate Fact Theory of Japanese Civil Code
by PROLOG Technology. In New Frontiers in Artificial Intelligence (Lecture Notes
in Computer Science). Springer Berlin Heidelberg, Berlin, Heidelberg, 153–164.
[38] Ken Satoh, Masahiro Kubota, Yoshiaki Nishigai, and Chiaki Takano. 2009. Trans-
lating the Japanese Presupposed Ultimate Fact Theory into Logic Programming.
In Proceedings of the 2009 Conference on Legal Knowledge and Information Systems:
JURIX 2009: The Twenty-Second Annual Conference. IOS Press, Amsterdam, The
Netherlands, 162–171.
[39] Marek J. Sergot, Fariba Sadri, Robert A. Kowalski, Frank Kriwaczek, Peter Ham-
mond, and H Terese Cory. 1986. The British Nationality Act as a logic program.
Commun. ACM 29, 5 (April 1986), 370–386.
[40] Ehud Y. Shapiro. 1983. Algorithmic Program DeBugging. MIT Press, Cambridge,
MA, USA.
[41] David M Sherman. 1987. A Prolog model of the income tax act of Canada. In
Proceedings of the 1st international conference on Artificial intelligence and law.
Association for Computing Machinery, New York, NY, USA, 127–136.
[42] Michael Thielscher. 2001. The qualification problem: A solution to the problem
of anomalous models. Artificial Intelligence 131, 1-2 (2001), 1–37.
[43] Bart Verheij. 2008. About the Logical Relations between Cases and Rules. In
Proceedings of the 2008 Conference on Legal Knowledge and Information Systems:
JURIX 2008: The Twenty-First Annual Conference. IOS Press, Amsterdam, The
Netherlands, 21–32.
[44] Bart Verheij. 2016. Correct grounded reasoning with presumptive arguments.
In European Conference on Logics in Artificial Intelligence. Springer International
Publishing, Cham, 481–496.
[45] Bart Verheij. 2016. Formalizing value-guided argumentation for ethical systems
design. Artificial Intelligence and Law 24, 4 (2016), 387–407.
[46] Bart Verheij. 2017. Formalizing Arguments, Rules and Cases. In Proceedings of
the 16th Edition of the International Conference on Articial Intelligence and Law
(London, United Kingdom) (ICAIL ’17). Association for Computing Machinery,
New York, NY, USA, 199–208.
[47] James Wogulis and Michael J Pazzani. 1993. A methodology for evaluating
theory revision systems: Results with Audrey II. In Proceedings of the Thirteenth
International Joint Conference on Artificial Intelligence. Morgan Kaufmann, San
Francisco,CA,USA, 1128–1134.
 Explainable Artificial Intelligence, Lawyer’s Perspective
Łukasz Górski
lgorski@icm.edu.pl
Interdisciplinary Centre for Mathematical and
Computational Modelling
Warsaw, Poland
ABSTRACT
Explainable artificial intelligence (XAI) is a research direction that
was already put under scrutiny, in particular in the AI&Law commu-
nity. Whilst there were notable developments in the area of (general,
not necessarily legal) XAI, user experience studies regarding such
methods, as well as more general studies pertaining to the concept
of explainability among the users are still lagging behind. This paper
firstly, assesses the performance of different explainability methods
(Grad-CAM, LIME, SHAP), in explaining the predictions for a legal
text classification problem; those explanations were then judged
by legal professionals according to their accuracy. Secondly, the
same respondents were asked to give their opinion on the desired
qualities of (explainable) artificial intelligence (AI) legal decision
system and to present their general understanding of the term XAI.
This part was treated as a pilot study for a more pronounced one
regarding the lawyer’s position on AI, and XAI in particular.
CCS CONCEPTS
• Computing methodologies → Neural networks.
KEYWORDS
explainable artificial intelligence, SHAP, LIME, Grad-CAM, survey,
XAI
ACM Reference Format:
Łukasz Górski and Shashishekar Ramakrishna. 2021. Explainable Artificial
Intelligence, Lawyer’s Perspective. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/3462757.3466145
1 INTRODUCTION
Explainable artificial intelligence (XAI) is a research field that aims
to make AI systems results more understandable to humans [1]. Other
domains notwithstanding, law is a domain that needs focus regard-
ing explainability. Due process of law and fair trial requirements
make an explanation of AI-based decision models a must. However,
the legal and technological discourses pertaining to the creation
of explainable AI seem to be separated. In particular, the methods
used to create explainable systems seem to favour the needs of
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466145
60
Shashishekar Ramakrishna
shashi792@gmail.com
Freie Universität Berlin
Berlin, Germany
AI-engineers. This work was thought as the first step towards the
identification of requirements of explainable AI-based systems that
would involve legal perspective to a greater extent.
For the purpose of this paper, a two-way investigation was per-
formed. Firstly, to assess the performance of different explainability
methods (Grad-CAM, LIME, SHAP), we have used a convolutional
neural network (CNN) for text classification and used those meth-
ods to explain the predictions; those explanations were then judged
by legal professionals according to their accuracy. Secondly, the
same respondents were asked to give their opinion on the desired
qualities of (explainable) artificial intelligence (AI) legal decision
system and to present their general understanding of the term XAI.
This part was treated as a pilot study for a more pronounced one
regarding the lawyer’s position on AI, and XAI in particular. Our re-
sults can be treated as a stepping stone towards a more pronounced
survey-based research.
This work contributes by:
• Presenting a comparison of different explainability methods
when applied to legal classification neural network.
• Giving an assessment of explainable artificial intelligence as
understood by lawyers.
2 RELATED WORK
XAI is a research direction that was already put under scrutiny,
in particular in the AI&Law community. The general consensus
is that the opening of black-box models is a conditio sine qua non
for assuring their trustworthiness and compliance with normative
background [12]. Moreover, in [33] it was noted that the develop-
ment of explainable technical solutions is necessarily connected
with the regulatory frameworks. Importantly, AI and legal specialists
need to engage in a dialogue when tackling explainability questions in
legal AI [33]. However, currently, we are unaware of any study that
aimed to investigate the understanding of XAI and elicit system’s
requirements from the lawyers. Yet, obviously, legal perspective
was brought into the discourse with the advent of legislation such
as European GDPR or envisioned ePrivacy or Digital Services Act,
other areas of law, like tort, notwithstanding [9]. As far as the legal
perspective is concerned, in [7] it was noted that the judiciary will
play an important role in the production of different forms of XAI,
especially because judicial reasoning, using case-by-case considera-
tion of the facts to produce nuanced decisions, is a pragmatic way to
develop rules for XAI [7]. Whilst this way of thinking sounds very
attractive and reminds how seminal works on case-based-reasoning
in AI originated in the AI&Law community [24], on the other hand,
the cited passage clearly alludes to common law judge mode of
thinking and civil law judge may be more concerned with general
rules.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
This train of thought should also remind of historical develop-
ments regarding the explainability in AI&Law. Historically, case-
based reasoners and rule-based reasoners (symbolic AI) are in-
herently interpretable [3]. Yet, producing viable explanations for
deep learning algorithms (sub-symbolic AI) remains a challenge.
Atkinson et al. cite Robbins, who remarked that [a]ny decision re-
quiring an explanation should not be made by machine learning
(ML) algorithms. Automation is still an option; however, this should
be restricted to the old-fashioned kind of automation whereby the
considerations are hard-coded into the algorithm [3] [25]. Conse-
quently, historical machine learning-based systems (like SMILE)
were capable of achieving good performance, though introspection
of their domain models repeatedly shows that those were often
incomplete or faulty [3]. Development of twin systems, with tra-
ditional rule-based or case-based reasoner explaining the results
of neural network was also a subject of scrutiny. Research in this
respect continues to test different newly developed methods to
achieve good explanations. Usage of Attention Network, accompa-
nied with attention-weight-based text highlighting was explored
in [4], with the negative conclusion regarding this approach’s fea-
sibility. In general, Competition on Legal Information Extraction
and Entailment (COLIEE) is a venue for applying state-of-the-art
research regarding the domain-specific information selection and
justification generation methods [21][20].
The work [4] is exemplary in including user evaluation study.
However, whilst there were notable developments in the area of
(general, not necessarily legal) XAI, user experience studies regard-
ing such methods, as well as more general studies pertaining to
the concept of explainability among the users are still lagging be-
hind [31] (with, for example, <1% of papers in the area of case-based
reasoning including user evaluations in the first place [12] [11]).
The results presented herein can be considered as facilitating such
dialogue.
On the other hand, surveys are commonly used to assess the
usability of computer-aided methods for lawyers. Whilst the study
of law is commonly associated rather with logical and linguistic
studies of texts, interdisciplinary studies of law do include methods
like survey-based study. Works that employed such methodology
include [4] (aforementioned attention-weight-based text highlight-
ing) or [5](for building the ontology of basic legal knowledge). As
recognized in [5], usability studies are an important part in the
human-centered design process, where they are used for, inter alia,
requirements specification with the use of questionnaire-based
methods. Such methodology was employed, for example, in [32],
where A/B study was conducted to grade the understandability
of student loan decision letter created by an automated decision
system. However, such system was rule-based and explainability
is mostly connected with deep-learning systems, which remain
inherently non-interpretable.
Contemporary XAI models view human-machine interaction as
a dynamic process, in which user understanding of the system is
continuously impacted by the explanation they receive [12]. User’s
mental model of the system can further be decomposed into three
parts: user model of the domain - which pertains to their under-
standing of the area that the system is working in, for example, legal
decision making; user model of the AI system - pertaining to the
knowledge of how the system is implemented, for example using
61
Ł. Górski, S. Ramakrishna
CNN or rule-based classifier; user model of the explanation - knowl-
edge how explanation strategy works with the system [12]. This
tripartite division offers a useful conceptual framework through
which the results of this paper can be analysed. While our respon-
dents have deep domain knowledge, their knowledge of AI-based
systems or explanation techniques is much shallower, if any.
In regard to the concrete explanation methods we’ve tested (Grad-
CAM, LIME, SHAP), they are of different origins. Grad-CAM is a
method used for the detection of the most important input data for
a CNN’s prediction. While it was originally used with image data, it
was already proven to be of use in the case of textual input, including
legal texts [8]. However, the aforementioned work does not include
an end-user feasibility study and can be supplemented with the
comparison of Grad-CAM with other explainability methods. They
were, to our knowledge, studied in isolation. For example, there are
already first studies regarding the usability of LIME for text data [18]
(however, with conclusions pertaining to decision trees and linear
models). Available analyses for SHAP, on the other hand, used it
with CNN-based text classification models [36]. Nevertheless, it
should be noted that post-hoc explanation methods were already
subject of criticism, due to them representing correlations and not
being faithful to the underlying model [30].
3 METHODOLOGY
3.1 Explanations generation
Herein, the effectiveness of different XAI methods when applied to
the law is studied. By using a plug-in classification system based
on CNN [13][8], we created a number of predictions based on the
well-known PTSD dataset [34]. Those were used as a basis for gen-
erating explanations for CNN predictions. In general, our pipeline
(based on previous work by other authors [29][13] with significant
extensions by us) is composed of embedder, classifications CNN,
visualizer, XAI module, and metric-based evaluator; we have also
developed a simple pre-processing module that uses some industry
de facto standard text processing libraries for spelling correction,
sentence detection, irregular character removal, etc., enhanced with
our own implementations which make them better-suited for legal
texts (though it was not used in this research). Our embedding
module houses a plug-in system to handle different variants of
embeddings, in particular BERT and word2vec. The classification
module houses simple 1D CNN which facilitates explainability
methods. The XAI module integrates the SHAP, Grad-CAM and
LIME models (based on library solutions, as developed, respectively
by [16][29][22]). It connects the output from the CNN to arrive at
explanations based on the model used. For LIME, the following pa-
rameters were customized: kernel_width=20, num_features=150,
num_samples=400.
For the purpose of this work, as an embedding layer, we use
DistilBERT [26] (a small, fast, cheap and light Transformer model
trained by distilling BERT base) encoding input sentences into vec-
tors. Huggingface’s DistilBERT implementation was found to work
relatively well with XAI libraries that we have used; unfortunately,
some of them even rely on Huggingface’s implementation details,
which can be found a limiting factor when plugin architecture is
considered.
Explainable Artificial Intelligence, Lawyer’s Perspective
The CNN used in the pipeline was trained for classification. We
use The Post-Traumatic Stress Disorder (PTSD) [19] dataset [35], for
CNN training (as well as testing). It annotates a set of sentences orig-
inating from 50 decisions issued by the Board of Veterans’ Appeals
("BVA") of the U.S. Department of Veterans Affairs according to their
function in the decision [35] [34] [27]. The classification consists of
six elements: Finding Sentence, Evidence Sentence, (Evidence-based)
Reasoning Sentence, Legal-Rule Sentence, Citation Sentence, Other
Sentence. This was found to be a relatively good candidate for neu-
ral network training, as the classes distinguished in this dataset
are of relatively similar sizes. The output from each of the three
explainer models had different value ranges, including negative
values. Each value inside the value range indicated the association
of the token to a specific class. For this experiment of understanding
the importance of a word in the downstream task of classification,
we round-off all negative explainer values to 0 and then normalize
the remaining value ranges to fit a range between 0 - 1.
The software stack used for the development of this system was
instrumented under Anaconda 4.8.3 (with Python 3.8.3). Tensor-
flow v. 2.2.0 was used for CNN instrumentation and Grad-CAMs
calculations (with the code itself expanding prior implementation
available at [29]). DistilBERT implementation and supporting codes
were sourced from Huggingface libraries: transformers v. 3.1.0, tok-
enizers v. 0.8.1rc2, nlp v. 0.4.0. For XAI methods implementations,
SHAP v. 0.37 and LIME v. 0.1.1.18 were facilitated. The code used
for this paper is available on GitHub1 . GPU cluster was used for
calculations (with 4x Nvidia Tesla V100 32GB GPUs).
3.2 Explainability methods
The system described hereinbefore, for the purpose of this work,
was used to generate three types of explanations. Those can be
broadly classified into three different types [37]:
(1) Gradient based-diagnostics e.g. Grad-CAM
(2) Shapley values-based diagnostics e.g. SHAP and
(3) Surrogate model-based diagnostics e.g. LIME
In general, gradient-based methods look inside the box to obtain
the gradient and feature values for attributing contributions. The
other two methods explain a given model from outside of the box
by looking at what is fed into the model and what is produced. The
more detailed description of concrete XAI methods chosen for this
paper is as follows:
3.2.1 Grad-CAM. Grad-CAM is an explainability method origi-
nating from computer vision [28]. It is a well established post-hoc
explainability technique when CNNs are concerned. Moreover, the
Grad-CAM method passed independent sanity checks [2]. Whilst
it is mainly connected with the explanations of deep learning net-
works used with image data, it has already been adapted for other
areas of application. In particular, CNN architecture for text clas-
sification was described in [13], and there exists at least one im-
plementation which extends this work with Grad-CAM support
for explainability [29]. Grad-CAMs were already used in the NLP
domain, for (non-legal) document retrieval [6]. With Grad-CAM
technique it is possible to produce a class activation map (heatmap)
for a given input sentence and predicted class. Each element of
1 https://github.com/lukeg/xai_lawyers_perspective
62
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the class activation map corresponds to one token and indicates
its importance in terms of the score of the particular (usually the
predicted) class. The class activation map gives information on
how strongly the particular tokens present in the input sentence
influence the prediction of the CNN.
3.2.2 SHAP. SHAP [SHapley Additive exPlanations] [17] is an ex-
plainability method based on the shapely concept. SHAP values are
used for identifying the contribution of each concept/words/features
in a given sentence. Shapley value is the average of the marginal
contributions across all permutations, i.e. average of all the permu-
tations for each concept/feature to get each entity’s contribution.
In terms of local interpretability, each observation gets its own set
of SHAP values. For our problem here we consider the global inter-
pretability, where Collective SHAP values can show how much each
predictor contributes, either positively or negatively, towards influ-
encing predictions of the CNN. SHAP has shown to provide good
consistency in attributing importance scores to each feature [15].
3.2.3 LIME. LIME [Local Interpretable Model (Agnostic) Expla-
nations] [10][23] is another explainability method that provides
explanations to the predictions of any classifier by learning inter-
pretable or simpler models locally around the prediction. While
the Shapley values consider all possible permutations, a united ap-
proach to provide global and local consistency, the LIME approach
builds sparse linear models around individual predictions in its
local vicinity.
3.3 Comparison metrics
The XAI methods described hereinbefore are used to visualize (high-
light) the parts of the input texts that are of most importance for the
network’s predictions. Such generated heatmaps (saliency maps)
are in turn presented to a legal professional for (qualitative) grading.
On the other hand, heatmaps are not easily susceptible to quantita-
tive analysis and can be supported by more formal metrics in this
respect [14][8]. Quantitative analysis and comparison of heatmaps
generated by different XAI methods are thus facilitated by recalling
the following metrics [8]:
(1) Fraction of elements above relative threshold 𝑡 (F (𝑣, 𝑡))
(2) Intersection over union with relative thresholds 𝑡 1 and 𝑡 2
(I (𝑣 1, 𝑣 2, 𝑡 1, 𝑡 2 ))
The first metric, F (𝑡), aims to quantify what portion of input
data is taken into account by CNN when making a prediction. It
can be defined as
| {𝑥 ∈𝑣 |𝑥 >𝑡 ×𝑚𝑎𝑥 (𝑣) } |
.
𝑙𝑒𝑛 (𝑣)
The second metric, I (𝑣 1, 𝑣 2, 𝑡 1, 𝑡 2 ), is used to compare two saliency
maps, 𝑣 1 and 𝑣 2 and can be used to compare whether the same parts
of input sentence affect the CNN’s prediction. It takes as arguments
two heatmaps (𝑣 1 and 𝑣 2 ), binarizes them using relative thresholds
|𝑣 ∩𝑣 |
(𝑡 1 and 𝑡 2 ) and finally calculates |𝑣1 ∪𝑣2 | . It quantifies the relative
1 2
overlap of words considered important for the prediction by each
of two models.
In both cases, 0.15 was chosen as a threshold value. This was
selected based on suggestions in prior work [28]. In addition to
that, for the current study, a value of 0.5 was tested as well, to filter
out the lower-ranking words.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
3.4 Pilot User Study
To study prospective user’s opinions on explainability in AI as
well as on the particular method employed in this study, a survey
was prepared. It consisted of two parts. The generic one contained
three questions, pertaining to the general knowledge of XAI by the
respondents. Those were as follows:
(1) Have you ever encountered the term “explainable artificial
intelligence” (explainable AI, XAI)?
(2) According to you how can “explainable artificial intelligence”/
explainability in decision systems be characterized?
(a) Explains the decision-making process and why we arrived
at this result.
(b) Conclusions are explained without a need for you to un-
derstand the inner-workings of the system.
(c) It’s more useful for a software developer rather than to a
lawyer.
(d) AI-based systems are of little use and their explainability
is thus irrelevant
(e) Explanations given by a computer system usually are not
sufficient and need to be supplemented with background
legal knowledge.
(3) What are your expectations regarding the justifications given
by the artificial intelligence-based automated decision-making
systems?
While the list of questions could have been non-exhaustive, the
goal of these above questions was only to assess the depth of their
knowledge in XAI. Answers to these questions help us to under-
stand respondent’s degree of confidence in the domain of XAI to
evaluate/score the results generated from different XAI models.
The second part of the study was aimed to elicit the lawyer’s
assessment of the three XAI methods that can be used to explain
CNN’s predictions. In total, six correctly classified sentences from
the test set were chosen as a basis for the study. Words composing
those sentences were highlighted with colors of different intensities,
according to their importance in the final prediction. All the words
were subject to highlighting, as even stopwords can be of impor-
tance in case of legal interpretation. Respondents were asked to
grade each visualization on the scale from 0 (worst) to 10 (best). We
have decided to use only the correctly classified sentences, as the
visualizations for incorrect ones might have looked peculiar for the
respondents and could have confused them when compared with
the visualizations for the correct ones. Moreover, understanding
this part of the study was already not easy for the respondents, with
one commenting that he skipped this part as he did not understand
what the colorful words mean.
In the end, from the classes distinguished in the dataset, three
were chosen for the visualization: citation sentence, reasoning sen-
tence (or evidence-based reasoning sentence) and legal rule sentence.
The classes were chosen so that each class stands out when com-
pared with the others. Original dataset includes two additional
classes, finding sentence and evidence sentence. However, even its
authors admit that it may be difficult to tell apart different sentences
when all the classes are compared (e.g. fact-finding ones vs. legal
rule ones are easy to conflate). As our respondents do not neces-
sarily have experience with cases under the cognition of Board
63
Ł. Górski, S. Ramakrishna
of Veterans’ Appeals, it was decided that using only a subset of
sentences’ classes was a good compromise.
The following sentences were chosen for grading:
(1) Evidence-based-reasoning sentences
(a) Further, as discussed below, none of the medical evidence
indicates that a psychiatric disorder had its onset during
service, and psychiatric disorders are complex matters
requiring medical evidence for diagnosis; they are not the
kind of disorders that subject to lay observation.
(b) Given the inconsistencies between the Veteran’s reports
and the objective evidence of record, the Veteran’s credi-
bility is diminished.
(2) Legal Rule Sentence
(a) Service connection for PTSD requires medical evidence
diagnosing the condition in accordance with 38 C.F.R.
4.125(a); a link, established by medical evidence, between
current symptoms and an in-service stressor; and credible
supporting evidence that the in-service stressor occurred.
(b) There must be 1) medical evidence diagnosing PTSD; 2)
a link, established by medical evidence, between current
symptoms of PTSD and an in-service stressor; and 3) cred-
ible supporting evidence that the claimed in-service stres-
sor occurred.
(c) The Federal Circuit has held that 38 U.S.C.A. 105 and 1110
preclude compensation for primary alcohol abuse disabil-
ities and secondary disabilities that result from primary
alcohol abuse.
(3) Citation sentence
(a) See also Mittleider v. West, 11 Vet. App. 181, 182 (1998)
(in the absence of medical evidence that does so, VA is
precluded from differentiating between symptomatology
attributed to a nonservice-connected disability and symp-
tomatology attributed to a service-connected disability).
4 RESULTS
4.1 Neural network training
Firstly, for classification, CNN was trained for 10 epochs, with a
batch size of 1000 elements and a learning rate of 0.001. 80% of the
PTSD dataset was used for training, with the remaining 20% left
out as a test set. This allowed to achieve accuracy of ca. 83% on
the test set. PTSD dataset was found to be a good candidate for
training, as it is relatively balanced, with classes’ size varying from
1941 elements (Evidence sentence) to 389 (Finding sentence) in case
of the training set. Relations between the sizes of classes in the case
of the test set were similar. For details on training effectiveness,
Table 1 can be consulted.
4.2 Study results. Lawyers’ conceptualization of
XAI
For our pilot study, we have collected 21 surveys. Respondents
represented a variety of legal professions, including university pro-
fessors, attorneys and junior legal advisors. Results for questions 1
(prior exposure to the term "XAI") and questions 2 (closed questions
regarding the respondents’ understanding of XAI) are summarised
in Table 2, given as a percentage of respective answers. In Table
Explainable Artificial Intelligence, Lawyer’s Perspective ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 1: Classification report for CNN trained for the purpose computer scientists (76%); three persons disagreed, with only one
of this paper of them seeing the need to understand the system’s inner workings
as a precondition for understanding of its explanation (as evidenced
Class Precision Recall F1-score Support by the previous question).
Artificial intelligence & law curriculum seems to have penetrated
Other Sentence 0.96 0.57 0.71 83
the ranks of lawyers, as 90% of respondents see the prospects of
Reasoning Sentence 0.59 0.41 0.48 148
their use (and, by extension of XAI). Remaining 10% had no opinion.
Finding Sentence 0.63 0.51 0.57 101
71% of respondents see the need of supplementing machine-based
Legal Rule Sentence 0.76 0.97 0.85 207
explanations with further background legal knowledge, 19% think
Citation Sentence 0.99 1.00 0.99 213
otherwise.
Evidence Sentence 0.88 0.95 0.92 479
In conclusion, lawyers seem to be affirmative regarding the use-
fulness of AI and XAI. Yet, more work should be devoted to prepare
Table 2: Respondents’ answers to questions 1 and 2 the explanations that do not involve prior knowledge of how a
given system works. If XAI was to be deployed and aimed at non-
yes no no opinion professionals, clarity and completeness of the explanations should
be the focal point, so that the decisions could be understood even
Question 1 without deeper background legal knowledge.
62% 38% – When we asked each respondent on their list of expectations
regarding the justifications given by an AI-based decision-system
Question 2
(question 3), the majority of their expectations are in line with any
a 90% 5% 5% general requirements necessary for such a system. Users’ expecta-
b 33% 52% 14% tions can be divided into three broad categories. These categories
c 14% 76% 10% are also in-line with the user’s mental model of a AI system (cf.
d 0% 90% 10% Sec. 2). General requirements, coming from users’ deep knowledge
e 71% 19% 10% of the law and its principles, like due process, contain expectations
such as: fairness, lack of bias, transparency in the decision process,
Table 3: Respondents’ answers to questions 2, broken down consequence awareness. Secondly, users’ model of AI-based system
according to their prior exposure to XAI tells them how it can increase the effectiveness of their work. Here,
respondents remarked that such system should decrease their man-
ual effort, lessen time consumption and introduce automation. Finally,
Prior exposure No prior exposure the user’s knowledge of XAI implementation is concerned with
yes no no opinion yes no no opinion supporting system’s conclusions with citations and provision of facts
a 100% 0% 0% 75% 13% 13% relevant to system’s decision. It should be noted that those require-
b 46% 46% 8% 13% 63% 25% ments apply not just to XAI systems, but also to any legal software
c 23% 69% 8% 0% 88% 13% system in general, AI-based or not. Thus, the mental model of an
d 0% 85% 15% 0% 100% 0% AI-based system in the case of prospective users is not fully devel-
e 77% 15% 8% 63% 25% 13% oped. Fig. 1, depicts the key vocabulary set used by the respondents
while providing their list of expectations.

the answers to question 2 are further broken down by the status of

prior exposure to XAI.
The majority of respondents already encountered the term "XAI".
Elicitation of XAI requirements through closed question 2 shows
that almost all lawyers agree that XAI is used to explain the decision-
making process and the way we arrived at the result (with one
person having "no opinion" and one disagreeing; both persons were
unfamiliar with the term XAI, as noted in their answer to q. 1).
Therefore we can conclude that - even in case of no prior exposure
to this term - lawyers intuitively conceptualize the meaning of this
term. Yet, certain divisions can be seen when results for question
2 are analyzed. 52% of respondents agreed that the XAI systems Figure 1: User expectations regarding the justifications given
give explanations and the user does not need to understand its by an AI-based decision-system
inner workings. On the other hand, 33% of respondents concluded
otherwise. Therefore, the significant minority pointed out that the
user of such system has to possess a deeper system’s model. Despite While a majority of the requirements described above were re-
this division, almost everyone agreed that XAI in legal decision- peated by multiple respondents, a few respondents provided re-
making systems would be of more importance to the lawyer than to quirements that were interesting from an AI-driven automated legal

64
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
decision support system’s perspective. They are interesting not be-
cause of their novelty, but because only a few domain experts/end
users thought about it. Some such requirements are as shown below:
(1) “It should have the option of forgetting the decisions taken in
past. However, it should also have a certain standard form of
code of conduct.”
(2) “There should be scope for generating contrasting explanations
based on various objectives/functions/factors.”
(3) “Basis on which AI has neglected the other information.”
(4) “Legally correct explanations.”
(5) “Option for collaborative (Man-Machine) explanation of the
law, context, and linguistic.”
(6) “Understandable explanations and Better Interfaces.”
4.3 Study results. XAI methods’ assessment
Tables 4, 5, 6 can be consulted for comparions of different XAI
methods when used with our sample sentences in terms of user
score as well comparison metrics (please note that the matrix for
I metric results is symmetric, therefore repeating results were
removed from the table). Below are some observation made based
on these results:
• Users’ expectations seem to have differed to a large extent,
which can be seen from the divisions of respondent’s grades
of SHAP’s explanation of citation sentence (Table 6 can be
consulted for visualization). SHAP marked all the words as
very important, while - for example, Grad-CAM marked only
the reference to the source material, and omitted the sum-
mary included with the citation. For the authors of this paper,
Grad-CAM has proven superior here, yet many respondents
judged SHAP highly. 9 respondents gave it the lowest score
out of all three (with two giving 0), 8 graded it the highest
(with one person giving 10).
• While the heatmaps for the citation sentence from Grad-
CAM and SHAP seem to be very different in nature (also
pointed out by their F- metric), the respondents seem to
provide scores that vary only by a small margin.
• Based on the F and I metric scores for all the sentence types,
LIME in general seems to have higher scores compared to
Grad-CAM and then SHAP.
• Both SHAP and LIME seem to be more sensitive to the change
in threshold value in terms of their F metric when compared
to Grad-CAM.
• Based on respondents’ average scores, Grad-CAM and SHAP
seem to perform consistently between sentence types when
compared against LIME. But, based on F metric, both LIME
and SHAP seem to have consistent values between sentence
types as compared to Grad-CAM.
Herein we have presented a software system, based on CNN,
capable of classifying legally relevant texts and coupled the sys-
tem with an explainability module. Furthermore, this explainability
module was then subject to user study. The use of ANOVA (be-
cause D’Agostino’s and Pearson’s test did not allow to reject the
hypothesis that scores for each of the XAI methods have normal
distributions and the same conclusions were arrived at when using
Levene test for the equality of variances) confirms that the discrep-
ancy between various scores is not of practical importance. The
65
Ł. Górski, S. Ramakrishna
same results were achieved using the Friedman test (𝛼 = 0.05 in
all the cases). Therefore other factors should be taken into account
when choosing a particular method. From a software engineering
point of view, it should be noted that implementation of certain
methods is dependant not only on the user’s voice but also on
technical feasibility. In this respect, it is of importance that Grad-
CAM is a method that is used together only with CNNs and we are
unaware of any works that managed to use it with other neural
network architectures. LIME is dependent on a number of hyper-
parameters. In this work, we have chosen the largest values that
allowed us to carry calculations with our hardware resources and
software implementation. Yet, many works suggest that one should
fine-tune those values until explanations received are close to one’s
expectations. However, this introduces a risk of overfitting explana-
tory procedures and having such fine-tuned explanations would be
counterproductive with regard to our study. As far as SHAP goes,
the library solution we used is still under active development, with
documentation not always up-to-date and with limited support of
external libraries. Regarding some of the observations presented
hereinbefore, there are a few for which different evaluation tech-
niques (technical, empirical, or mixed) need to be performed for
arriving at conclusions. Such conclusions need to be backed by
domain experts. Those will form a basis for future work.
5 CONCLUSION
In this paper, we have presented a comparison of different explain-
ability methods when applied to legal classification neural network
(CNN) and provided an assessment of explainable artificial intel-
ligence as understood by lawyers. Different XAI methods were
graded similarly by the users, though certain variances can be spot-
ted. The metrics presented herein offer software engineers an option
to quantify explanations given by the system. When a more general
point of view is taken, one which comes from prospective users, it
should be noted that the lawyers are generally looking forward to
the implementation of (explainable) artificial intelligence systems
and solutions that allow them to be more efficient in their use of
time. Users are thus waiting for results of AI - and in particular XAI
- research.
ACKNOWLEDGMENTS
This research was carried out with the support of the Interdisci-
plinary Centre for Mathematical and Computational Modelling
(ICM), University of Warsaw, under grant no GR81-14.
DISCLAIMER
The views reflected in this article are the views of the author (SR)
and do not necessarily reflect the views of his employer organisation
or its member firms.
REFERENCES
[1] Amina Adadi and Mohammed Berrada. 2018. Peeking inside the black-box: a
survey on explainable artificial intelligence (XAI). IEEE access 6 (2018), 52138–
52160.
[2] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt,
and Been Kim. 2018. Sanity Checks for Saliency Maps. In Advances in
Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle,
Explainable Artificial Intelligence, Lawyer’s Perspective ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 4: Comparison of XAI methods on sample PTSD evidence-based-reasoning sentences

I metric I metric
Method Sentence User F metric vs. vs. F metric vs. vs.
score SHAP LIME SHAP LIME
(avg.)
𝑡 = 𝑡 1 = 𝑡 2 = 0.15 𝑡 = 𝑡 1 = 𝑡 2 = 0.5

Grad-CAM 4.43 0.51 0.46 0.48 0.14 0.18 0.3

SHAP 3.48 0.82 0.74 0.13 0.25

LIME 6.66 0.85 0.45

Grad-CAM 6.15 0.92 0.89 0.81 0.78 0.22 0.19

SHAP 5.15 0.96 0.85 0.26 0.31

LIME 5.8 0.85 0.4

K. Grauman, N. Cesa-Bianchi, and R. Garnett (Eds.), Vol. 31. Curran As-
sociates, Inc., 9505–9515. https://proceedings.neurips.cc/paper/2018/file/
294a8ed24b1ad22ec2e7efea049b8737-Paper.pdf
[3] Katie Atkinson, Trevor Bench-Capon, and Danushka Bollegala. 2020. Explanation
in AI and law: Past, present and future. Artificial Intelligence (2020), 103387.
[4] L Karl Branting, Craig Pfeifer, Bradford Brown, Lisa Ferro, John Aberdeen, Brandy
Weiss, Mark Pfaff, and Bill Liao. 2020. Scalable and explainable legal prediction.
Artificial Intelligence and Law (2020), 1–26.
[5] Núria Casellas. 2011. Legal ontology engineering: Methodologies, modelling trends,
and the ontology of professional judicial knowledge. Vol. 3. Springer Science &
Business Media.
[6] Jaekeol Choi, Jungin Choi, and Wonjong Rhee. 2020. Interpreting Neural Ranking
Models using Grad-CAM. arXiv preprint arXiv:2005.05768 (2020).
[7] Ashley Deeks. 2019. The judicial demand for explainable artificial intelligence.
Columbia Law Review 119, 7 (2019), 1829–1850.
[8] Lukasz Gorski, Shashishekar Ramakrishna, and Jedrzej M Nowosielski. 2020.
Towards Grad-CAM Based Explainability in a Legal Text Processing Pipeline.
arXiv preprint arXiv:2012.09603 (2020).
[9] Philipp Hacker, Ralf Krestel, Stefan Grundmann, and Felix Naumann. 2020. Ex-
plainable AI under contract and tort law: legal incentives and technical challenges.
Artificial Intelligence and Law (2020), 1–25.
[10] Linwei Hu, Jie Chen, Vijayan N. Nair, and Agus Sudjianto. 2020. Surrogate
Locally-Interpretable Models with Supervised Machine Learning Algorithms.
arXiv:2007.14528 [stat.ML]
[11] Mark T Keane and Eoin M Kenny. 2019. How case-based reasoning explains
neural networks: A theoretical analysis of XAI using post-hoc explanation-by-
example from a survey of ANN-CBR twin-systems. In International Conference
on Case-Based Reasoning. Springer, 155–171.
[12] Eoin M. Kenny, Courtney Ford, Molly Quinn, and Mark T. Keane. 2021. Explain-
ing black-box classifiers using post-hoc explanations-by-example: The effect of
explanations and error-rates in XAI user studies. Artificial Intelligence 294 (2021),
103459. https://doi.org/10.1016/j.artint.2021.103459
[13] Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification.
In Proceedings of the 2014 Conference on Empirical Methods in Natural Language
Processing (EMNLP). Association for Computational Linguistics, Doha, Qatar,
1746–1751. https://doi.org/10.3115/v1/D14-1181
[14] David Krakov and Dror G Feitelson. 2013. Comparing performance heatmaps. In
Workshop on Job Scheduling Strategies for Parallel Processing. Springer, 42–61.
[15] Scott M. Lundberg, Gabriel G. Erion, and Su-In Lee. 2019. Consistent Individual-
ized Feature Attribution for Tree Ensembles. arXiv:1802.03888 [cs.LG]
[16] Scott M Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting Model
Predictions. In Advances in Neural Information Processing Systems 30, I. Guyon,
U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett
(Eds.). Curran Associates, Inc., 4765–4774. http://papers.nips.cc/paper/7062-a-
unified-approach-to-interpreting-model-predictions.pdf
[17] Scott M. Lundberg and Su-In Lee. 2017. A Unified Approach to Interpreting
Model Predictions. In Proceedings of the 31st International Conference on Neural
Information Processing Systems (Long Beach, California, USA) (NIPS’17). Curran
Associates Inc., Red Hook, NY, USA, 4768–4777.

66
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Ł. Górski, S. Ramakrishna

Table 5: Comparison of XAI methods on sample PTSD legal rule sentences

I metric I metric
Method Sentence User F metric vs. vs. F metric vs. vs.
score SHAP LIME SHAP LIME
(avg.)
𝑡 = 𝑡 1 = 𝑡 2 = 0.15 𝑡 = 𝑡 1 = 𝑡 2 = 0.5

Grad-CAM 5.2 0.8 0.89 0.77 0.42 0.41 0.34

SHAP 6.05 0.9 0.87 0.52 0.41

LIME 5.9 0.93 0.48

Grad-CAM 5.45 0.73 0.47 0.8 0.12 0.14 0.18

SHAP 5.75 0.67 0.56 0.36 0.08

LIME 4.3 0.73 0.11

Grad-CAM 5.9 0.71 0.41 0.71 0.4 0.13 0.06

SHAP 4.2 0.63 0.63 0.29 0.25

LIME 6.15 0.93 0.54

[18] Dina Mardaoui and Damien Garreau. 2020. An Analysis of LIME for Text Data.
arXiv preprint arXiv:2010.12487 (2020).
[19] Victoria Hadfield Moshiashwili. 2015. The Downfall of Auer Deference: Veterans
Law at the Federal Circuit in 2014.
[20] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. [n.d.]. COLIEE 2020: Methods for Legal Document Retrieval
and Entailment. ([n. d.]).
[21] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2019. A Summary of the COLIEE 2019 Competition. In
JSAI International Symposium on Artificial Intelligence. Springer, 34–49.
[22] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why Should I
Trust You?": Explaining the Predictions of Any Classifier. In Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, San Francisco, CA, USA, August 13-17, 2016. 1135–1144.

67
Explainable Artificial Intelligence, Lawyer’s Perspective ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 6: Comparison of XAI methods on sample PTSD citation sentence

I metric I metric
Method Sentence User F metric vs. vs. F metric vs. vs.
score SHAP LIME SHAP LIME
(avg.)
𝑡 = 𝑡 1 = 𝑡 2 = 0.15 𝑡 = 𝑡 1 = 𝑡 2 = 0.5

Grad-CAM 5.2 0.26 0.26 0.24 0.13 0.13 0.11

SHAP 4.95 0.98 0.83 0.98 0.16

LIME 4.65 0.73 0.17

[23] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. 2016. "Why
Should I Trust You?": Explaining the Predictions of Any Classifier.
arXiv:1602.04938 [cs.LG]
[24] Edwina L Rissland, Kevin D Ashley, and Ronald Prescott Loui. 2003. AI and Law:
A fruitful synergy. Artificial Intelligence 150, 1-2 (2003), 1–15.
[25] Scott Robbins. 2019. A misdirected principle with a catch: explicability for AI.
Minds and Machines 29, 4 (2019), 495–514.
[26] Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. 2019.
DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter.
arXiv:1910.01108 [cs.CL]
[27] Jaromír Savelka, Vern R. Walker, Matthias Grabmair, and Kevin D. Ashley. 2017.
Sentence Boundary Detection in Adjudicatory Decisions in the United States.
[28] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedan-
tam, Devi Parikh, and Dhruv Batra. 2017. Grad-cam: Visual explanations from
deep networks via gradient-based localization. In Proceedings of the IEEE interna-
tional conference on computer vision. 618–626.
[29] Haebin Shin. [n.d.]. Grad-CAM for Text. https://github.com/HaebinShin/grad-
cam-text. Accessed: 2020-08-05.
[30] Dylan Slack, Sophie Hilgard, Emily Jia, Sameer Singh, and Himabindu Lakkaraju.
2020. Fooling lime and shap: Adversarial attacks on post hoc explanation methods.
In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society. 180–186.
[31] Jasper van der Waa, Elisabeth Nieuwburg, Anita Cremers, and Mark Neerincx.
2021. Evaluating XAI: A comparison of rule-based and example-based explana-
tions. Artificial Intelligence 291 (2021), 103404.
[32] Tom M van Engers and Dennis M de Vries. 2019. Governmental Transparency in
the Era of Artificial Intelligence.. In JURIX. 33–42.
[33] Martijn van Otterlo and Martin Atzmueller. 2018. On Requirements and Design
Criteria for Explainability in Legal AI. In XAILA@ JURIX.
[34] Vern R. Walker, Ji Hae Han, Xiang Ni, and Kaneyasu Yoseda. 2017. Semantic Types
for Computational Legal Reasoning: Propositional Connectives and Sentence
Roles in the Veterans’ Claims Dataset (ICAIL ’17). Association for Computing Ma-
chinery, New York, NY, USA, 217–226. https://doi.org/10.1145/3086512.3086535
[35] Vern R. Walker, Krishnan Pillaipakkamnatt, Alexandra M. Davidson, Marysa
Linares, and Domenick J. Pesce. 2019. Automatic Classification of Rhetorical
Roles for Sentences: Comparing Rule-Based Scripts with Machine Learning. In
Proceedings of the Third Workshop on Automated Semantic Analysis of Information
in Legal Texts, Montreal, QC, Canada, June 21, 2019 (CEUR Workshop Proceedings,
Vol. 2385). http://ceur-ws.org/Vol-2385/paper1.pdf
[36] Wei Zhao, Tarun Joshi, Vijayan N Nair, and Agus Sudjianto. 2020. SHAP values for
Explaining CNN-based Text Classification Models. arXiv preprint arXiv:2008.11825
(2020).
[37] Wei Zhao, Tarun Joshi, Vijayan N. Nair, and Agus Sudjianto. 2020. SHAP values
for Explaining CNN-based Text Classification Models. arXiv:2008.11825 [cs.CL]

68
 Unravel Legal References in Defeasible Deontic Logic
Guido Governatori
Data61
CSIRO
Dutton Park, Queensland, Australia
guido.governatori@data61.csiro.au
ABSTRACT
Legal documents often contain references to either other documents,
or other parts (of the same document). The use of references is
meant to reduce the complexity of the documents; however, they
pose serious concerns for the formal (logical) representation of the
norms stipulated in the document itself. We propose an approach
to directly model the references in a logic language and to resolve
them during the computation of the legal effects in force in a case.
The approach is proved to be computationally feasible and to have
an efficient algorithmic implementation.
CCS CONCEPTS
• Computing methodologies → Artificial intelligence; • Ap-
plied computing → Law; • Theory of computation → Proof
theory; Automated reasoning.
KEYWORDS
Defeasible Deontic Logic, legal references
ACM Reference Format:
Guido Governatori and Francesco Olivieri. 2021. Unravel Legal References in
Defeasible Deontic Logic. In Eighteenth International Conference for Artificial
Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM,
New York, NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466080
1 INTRODUCTION
A typical characteristic of legal documents aiming at the reduction
of the complexity of the documents themselves is the use of ref-
erences, either to other sections of the same document (internal
references), or to sections of other documents (external references).
The key idea behind references is that they are used to import “con-
tent” from relevant provisions in the provision where the reference
appears, without the need to repeat the content/text of the imported
provision. Also, frequently, the content is not simply imported, as
a reference may require a legal lens to unravel the actual (legal)
intent of the content, in the context where the reference appears.
The focus of this work is to examine the references in a legal
document under the legal lens that assesses the legal status of the
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466080
69
Francesco Olivieri
Institute for Integrated and Intelligent Systems
Griffith University
Nathan, Queensland, Australia
f.olivieri@griffith.edu.au
said references. Here, by the legal status we mean where the provi-
sion corresponding to the reference either is applicable, has been
complied with, or has been violated. The examples below provide a
wide set of instances of the types of references for which we are
going to develop techniques to represent them in a logic designed
to formalise legal reasoning.
Example 1.1 (Applicable). Section 9.8.2, Telecommunications Con-
sumer Protections Code.
Termination by a Customer: A Supplier must ensure that, if so
notified by the Customer who is exercising the applicable termi-
nation right in clause 9.9.1 h), if any, as a result of the move, the
Supplier terminates the relevant Customer Contract relating to the
Telecommunications Service within 5 Working Days of receiving
the Customer’s notice.
Example 1.2 (Complied with). Section 9.8.2, Telecommunications
Consumer Protections Code.
Provided that a Supplier complies with the terms of this clause
9.9 in circumstances of a move to an alternate wholesale network
provider, the Supplier is not required to comply with the other
provisions of this Chapter in relation to such a move except for
clauses 9.5, 9.6, and 9.7.
Example 1.3 (Violated). Section 26-105, Income Tax Assessment
Act 1997.
(2) You cannot deduct under this Act a non-cash benefit if:
(a) section 14–5 in Schedule 1 to the Taxation Administration
Act 1953 requires you to pay an amount to the Commis-
sioner before providing the benefit, because of any of the
following provisions in that Schedule:
(i) section 12–35 (about payments to employees);
(ii) section 12–40 (about payments to directors);
(iii) section 12–47 (about payments to religious practition-
ers);
(iv) section 12–60 (about payments under labour hire and
certain other arrangements);
(v) in relation to a supply, other than a supply referred to
in subsection (3) of this section–section 12-190 (about
quoting of ABN); and
(b) you fail to comply, or purportedly comply, with section
16–150 in that Schedule in relation to the amount.
Given a formal representation of a set of legal provisions as
logical expressions, one of the aims of a logic for legal reasoning is
to infer what conclusions are entailed by a set of given premises
(e.g., corresponding to the facts of a case). Said conclusions are
meant to represent: (a) the legal requirements (or effects) that hold,
or are in force, based on the facts of the case, and (b) the set of
norms (the given legal provisions).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
In general, following [14], provisions can be represented by “IF
. . . THEN . . . ” rules, where the IF part contains the conditions under
which a rule is able to produce its conclusion (or effect), which is
encoded by the THEN part. Provisions are, in general, defeasible;
this means that there are some provisions that give baseline condi-
tions under which the effect of the norm holds (or enters in force).
However, such baseline rules are subject to exceptions (or exclu-
sions, or derogations). We also adopt the usual distinction between
constitutive and prescriptive rules, where a constitutive rule defines
an institutional fact in the underlying normative system, while
a prescriptive rule stipulates the conditions under which a legal
effect is in force and what is the legal effect (either an obligation, a
prohibition, or a permission).
Based on the discussion so far, it should be clear that, to formalise
provisions containing citations (as the instances given in the exam-
ples), one has to understand what are the conditions corresponding
to the cited provisions before such conditions are represented by
logical expressions. One option is to formalise the cited provisions
and then copy the formalised conditions in the formalisation of the
citing provision. However, this is not enough because we have to
include the logical representation of the semantic of the citation.
Let us start by examining the case of “applicable”. It can have
multiple meanings. The first is that the IF part of the provision
formalisation holds (it is deemed to be true in the particular situ-
ation or case). The second meaning is that, in addition to the IF
part, also the THEN part holds. For the first case, the referring
provision should import the IF part, and for the second case the
import consists of the conjunction of both the IF part and the THEN
part. Suppose that section X is modelled by the rule
IF 𝐴1, . . . , 𝐴𝑛 THEN 𝐵 (1)
and that section Y, citing section X, requires conditions 𝐶 1, . . . , 𝐶𝑚
to produce the effect 𝐷 when section X applies. According to the
first reading the rule for section Y is
IF 𝐴1, . . . , 𝐴𝑛 , 𝐶 1, . . . , 𝐶𝑚 , THEN 𝐷. (2)
Based on the second interpretation the rendering of section Y is
IF 𝐴1, . . . , 𝐴𝑛 , 𝐵, 𝐶 1, . . . , 𝐶𝑚 , THEN 𝐷. (3)
However, in case there are multiple provisions to support the same
effect or conclusion, such rules might not offer an accurate model
of the normative system. To handle this situation, it is not enough
to use the conjunction of the IF and THEN parts, but we have to
include, in the formalisation, elements from the provisions that
might provide exceptions to the cited provision. Moreover, it is
possible to have exclusions of the exclusions, and this has a ripple
effect on the conditions to be included in the citing provision. When
we look at the examples above, specifically Example 1.3, and we
consider real-life Acts, where exclusion sections can run for several
pages with many rules encoding the exceptions (and the exceptions
themselves have further exclusions), it is evident that unravelling
citations based on the technique we just alluded to is untenable.
Consequently, the solution we advocate is to extend the logical
language with a new class of unary predicates whose argument is
the cited proposition; the new predicates correspond to the possible
legal statuses for their argument (the cited provision). Hence, for
70
Guido Governatori and Francesco Olivieri
section X in the modelling of section Y, we could use
IF 𝐶 1, . . . , 𝐶𝑚 , 𝐴𝑝𝑝𝑙𝑦 (𝑆𝑒𝑐𝑡𝑖𝑜𝑛 X) THEN 𝐷 (4)
The issue now is to determine when 𝐴𝑝𝑝𝑙𝑦 (𝑠𝑒𝑐𝑡𝑖𝑜𝑛 𝑋 ) holds and
can then be use to trigger rule above. A solution would be to add
the rule
IF 𝐴1, . . . , 𝐴𝑛 THEN 𝐴𝑝𝑝𝑙𝑦 (𝑆𝑒𝑐𝑡𝑖𝑜𝑛 X). (5)
This approach works well for the first meaning of applicable, but
it does not for the second interpretation, when it faces the same
issues discussed for the previous approach when exceptions exist.
Complied with and violated are types of references that apply
to prescriptive rules only. In the simplest case, the semantics for
complied with requires that if the provision is applicable, then the
effect (i.e., an obligation or a prohibition) has been fulfilled. This
amounts to that if the effect is the obligation O𝐴, then 𝐴 holds as
well; if the effect is a prohibition (i.e., F𝐴, equivalent to O¬𝐴), then
𝐴 does not hold. The case of violated is similar: the rule is violated if
the content of the obligation does not hold, or it holds if the effect
is a prohibition. Simple adaptations of the techniques exemplified
above are available. All we have to do is to include O𝐴 and 𝐴 (for
complied with), or O𝐴 and ¬𝐴 (for violated) in the IF part of the
appropriate rules.
Things get more complicated when we take into account a key
feature of normative systems: situations when it is possible to fail
to comply with some provisions, but compensatory measures are
contemplated. Hence, it is possible to fail in fulfilling the primary
obligation and still be compliant by fulfilling the compensatory mea-
sures. Nevertheless, we might encounter scenarios where, typically,
failing to comply with the primary obligation requires us to comply
with the compensatory obligation, but in specific circumstances
what constitutes the compensatory measure to recover from the vi-
olation of the primary obligation is forbidden, and then there is not
an effective way to compensate and restore compliance. According
to this discussion, we comply with an applicable provision when
we fulfil the primary obligation, or at least one of the compensatory
obligations that is actually in force and we fulfil it. Conversely, a
provision is not complied with when one of the obligations in force
(either the primary obligation or one of the compensatory obliga-
tions) is not fulfilled and it does not admit a compensation. While it
is possible to write rules importing the content of the related rules
(in this case rules preventing a compensation to be in force), the
process would require to use an oracle that simulate the underlying
reasoning to determine what are the rules that contribute to the
conditions of the imported citation.
The method we propose is to incorporate in the logic itself the
conditions for the resolution of the citation predicates. This means
that, while the logic determines when the conditions of a rule hold
and it possible to infer that the rule’s conclusion holds as well, then
we can assert: (i) that the rule applies (according to the second
reading of applicable), and (ii) that the citations are resolved at the
same time the inferences that can be drawn from a set of rules and
a set of facts are computed. The combination of (i) and (ii) avoids
the need to write additional rules to capture the intended semantics
(eventually using an oracle to solve the semantic conditions) of the
cited provisions.
Unravel Legal References in Defeasible Deontic Logic
2 LOGIC
In this section we introduce the logical apparatus, which is an ex-
tension of standard Defeasible Logic (DL) [1], and more specifically
is based on the modal, deontic frameworks proposed in [5], as we
shall need means to formalise prescriptive behaviours as well as to
determine which situations are compliant, and which ones are not.
2.1 Language of Defeasible Deontic Logic
Let PROP be the set of propositional atoms, then the set of literals
is Lit = PROP ∪ {¬𝑝 | 𝑝 ∈ PROP}. The complementary of a literal
𝑝 is denoted with ∼𝑝: if 𝑝 is a positive literal 𝑞 then ∼𝑝 is ¬𝑞, if 𝑝
is a negative literal ¬𝑞 then ∼𝑝 is 𝑞. The set of deontic literals is
ModLit = {𝑀𝑙, ¬𝑀𝑙 | 𝑙 ∈ Lit ∧ 𝑀 ∈ {O, P}}. Note that we will not
have specific rules nor modality for prohibitions, as we will treat
them according to the standard duality that something is forbidden
iff the opposite is mandatory (F𝑝 ≡ O∼𝑝).
Lab is the set of labels, which are names for rules and will be
denoted by small-capitals Greek letters. As we will be interested in
understanding (deriving) when a rule is violated, complied with, and
active, we extend the set of literals with reference literals RefLit =
{𝑌 (𝛼), ¬𝑌 (𝛼)| 𝛼 ∈ Lab ∧ 𝑌 ∈ {𝑎𝑐𝑡𝑖𝑣𝑒, 𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑, 𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑_𝑤𝑖𝑡ℎ}.
A defeasible theory 𝐷 is a tuple (𝐹, 𝑅, >). 𝐹 ⊆ Lit is the set of
facts (indisputable, constitutive statements). (In this version of the
logic, we do not admit obligations and permissions to be facts of
the theory.) We use two kinds of rules. Non-deontic (thus standard,
or constitutive) rules 𝑅 C model constitutive statements (count-as
rules). Deontic rules represent prescriptive behaviours; deontic
rules are either obligation rules 𝑅 O which determine when and
which obligations are in force, or permission rules 𝑅 P which repre-
sent strong (or explicit) permissions. Finally, > is a binary relation
over 𝑅 to solve conflicts in case of potentially conflicting informa-
tion.
Following the ideas of [8], obligation rules gain more expressive-
ness with the compensation operator ⊗ for obligation rules, which
is to model reparative chains of obligations. Intuitively, 𝑎 ⊗ 𝑏 means
that 𝑎 is the primary obligation, but if for some reason we fail to
obtain, to comply with, 𝑎 (by either not being able to prove 𝑎, or
by proving ∼𝑎), then 𝑏 becomes the new obligation in force. This
operator is used to build chains of preferences, called ⊗-expressions.
The formation rules for ⊗-expressions are: (i) every literal is
an ⊗-expression, (ii) if 𝐴 is an ⊗-expression and 𝑏 is a literal then
𝐴 ⊗ 𝑏 is an ⊗-expression. In addition, we stipulate that ⊗ obeys the
following properties: (a) Associativity 𝑎 ⊗ (𝑏 ⊗ 𝑐) = (𝑎 ⊗ 𝑏) ⊗ 𝑐, (b)
Ë Ë𝑘−1
Duplication
Ë𝑚 and contraction on the right 𝑚 𝑖=1 𝑐𝑖 = ( 𝑖=1 𝑐𝑖 ) ⊗
( 𝑖=𝑘+1 𝑐𝑖 ) for 𝑗 < 𝑘 and 𝑐 𝑗 = 𝑐𝑘 .
We adopt the standard DL’s definitions of strict rules, defeasible
rules and defeaters [1], where a rule is an expression 𝛼 : 𝐴(𝛼) ↩→𝑋
𝐶 (𝛼). 𝛼 ∈ Lab is a unique label for the name of the rule. ↩→∈
{→, ⇒, ;} is the type of rule: we use → for strict rules, ⇒ for
defeasible rules, and ; for defeaters. 𝑋 = {C, O, P} is the kind of
rule: if 𝑋 = C then ↩→ has no subscript and the rule is used to derive
non-deontic literals (constitutive statements), whilst if X is O or
P then the rule is used to derive deontic conclusions (prescriptive
statements). 𝐴(𝛼) = 𝑎 1, . . . , 𝑎𝑛 is the set of antecedents/premises
where each 𝑎𝑖 is either a literal, a deontic literal, or a reference
literal. Lastly, 𝐶 (𝛼) is the conclusion of the rule (referred to as the
71
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
head of the rule) which is a single literal in case 𝑋 = {C, P}, or an
⊗-expression in case 𝑋 = O.
We use several abbreviations on sets of rules. 𝑅s , 𝑅d , 𝑅sd denote,
respectively, the set of only strict, only defeasible, and both strict
and defeasible rules; 𝑅𝑋 [𝑙] is the set of constitutive rules (𝑋 = C) or
deontic rules (𝑋 ∈ {O, P}) whose head is 𝑙. Lastly, 𝑅 O [𝑙, 𝑖] denotes
the set of obligation rules, where 𝑙 is the 𝑖-th element in the ⊗-
expression.
The meaning of an ⊗-expression 𝐶 (𝛼) = 𝑐 1 ⊗ 𝑐 2 ⊗ · · · ⊗ 𝑐𝑚 as
consequent of a rule 𝐴(𝛼) ↩→O 𝐶 (𝛼) is that: if the rule is allowed
to draw its conclusion, then 𝑐 1 is the obligation in force, and only
when 𝑐 1 is violated then 𝑐 2 becomes the new in force obligation,
and so on for the rest of the elements in the chain. In this setting, 𝑐𝑚
represents the last chance to comply with the prescriptive behaviour
enforced by 𝛼 and, in case 𝑐𝑚 is violated as well, then we will result
in a non-compliant situation.
A conclusion of 𝐷 is either a tagged (deontic) literal, or a tagged
label; as it will be clear in the reminder of the section, reference
literals are not conclusions of rules and they can be derived based
on various conditions on the provability of particular literals in
the rule. For (deontic) literals, a conclusion can have one of the
following forms: (i) ±Δ𝑋 𝑙 which means that 𝑙 is definitely prov-
able/refutable in 𝐷, and (ii) ±𝜕𝑋 𝑙 which means that 𝑙 is defeasibly
provable/refutable in 𝐷. For labels, we use the ideas of ⊥ and ⊤
from [6] (that will provide a compact notation for compliance and
violation), a conclusion can have one of the following forms: (iii)
±𝜕A 𝛼 which means that 𝛼 is active/not active (various nuances
are possible, and some are given in Definitions 2.5–2.10), (iv) +⊤𝛼
which means that 𝛼 is complied with, (v) −⊤𝛼 which means that
𝛼 is not complied with - violated and not compensated -, (vi) +⊥𝛼
which means that 𝛼 is violated, and lastly (vii) −⊥𝛼 which means
that 𝛼 has never been violated.
A proof of length 𝑛 in 𝐷 is a finite sequence 𝑃 (1), 𝑃 (2), . . . , 𝑃 (𝑛)
of the tagged literals and tagged reference literals just described
above and formally defined hereafter; 𝑃 (1..𝑛) denotes the first 𝑛
steps of 𝑃. If, for instance, 𝐷 proves +𝜕𝑙 at proof step 𝑛, we write
𝑃 (𝑛) = +𝜕𝑙, and we also assume the conventional notation 𝐷 ⊢ +𝜕𝑙.
Before defining when a rule is applicable/discarded, we provide
the two definitions of being a trigger, and being a blank. These two
concepts are related to the elements in the set of antecedents for a
given rule. Conceptually, an antecedent is a trigger for a rule when
its provability allows such a rule to (potentially1 ) fire, whilst it is a
blank when its refutability prevents such a rule to fire. ¬O𝑙 (resp.
¬P𝑙) means that the obligation (resp. permission) of 𝑙 is not in force.
The meaning of being active depends on the interpretation,
which may vary from context to context. Notice that henceforth
we use active to refer to a rule that we describe as applicable in
Section 1 to distinguish it from the Defeasible Logic notion of a rule
being applicable. Various different nuances for the notion of active
are proposed afterwards, but we first start by formalising provabil-
ity/refutability for constitutive statements, obligations, permissions,
and finally violated and complied with rules.
Definition 2.1 (Trigger). Given a deontic defeasible theory 𝐷, a
proof 𝑃, a rule 𝛼, we say that 𝑎 ∈ 𝐴(𝛼) is a trigger for 𝛼 at 𝑃 (𝑛 + 1)
iff
1 Potentially because the rule itself may be defeated.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Guido Governatori and Francesco Olivieri

(1) if 𝑎 ∈ Lit, then
(a) if 𝛼 ∈ 𝑅𝑠 then +Δ𝑎 ∈ 𝑃 (1..𝑛);
(b) if 𝛼 ∈ 𝑅𝑑 then +𝜕𝑎 ∈ 𝑃 (1..𝑛);
(2) if 𝑎 = 𝑋𝑙 and 𝑋 ∈ {O, P}, then
(a) if 𝛼 ∈ 𝑅𝑠 then +Δ𝑋 𝑙 ∈ 𝑃 (1..𝑛);
(b) if 𝛼 ∈ 𝑅𝑑 then +𝜕𝑋 𝑙 ∈ 𝑃 (1..𝑛);
(3) if 𝑎 = ¬𝑋𝑙, then −𝜕𝑋 𝑙 ∈ 𝑃 (1..𝑛) with 𝑋 ∈ {O, P};
(4) if 𝑎 is 𝑎𝑐𝑡𝑖𝑣𝑒 (𝛼), then +𝜕EA 𝛼 ∈ 𝑃 (1..𝑛);
(5) if 𝑎 is ¬𝑎𝑐𝑡𝑖𝑣𝑒 (𝛼), then −𝜕EA 𝛼 ∈ 𝑃 (1..𝑛);
(6) if 𝑎 is 𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑 (𝛼), then +⊤𝛼 ∈ 𝑃 (1..𝑛);
(7) if 𝑎 is ¬𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑 (𝛼), then −⊤𝛼 ∈ 𝑃 (1..𝑛);
(8) if 𝑎 is 𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑 (𝛼), then +⊥𝛼 ∈ 𝑃 (1..𝑛);
(9) if 𝑎 is ¬𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑 (𝛼), then −⊥𝛼 ∈ 𝑃 (1..𝑛);
Definition 2.2 (Blank). Given a deontic defeasible theory 𝐷, a
proof 𝑃, and a rule 𝛼, we say that a literal 𝑎 ∈ 𝐴(𝛼) is a blank for 𝛼
at 𝑃 (𝑛 + 1) iff
(1) if 𝑎 ∈ Lit, then
(a) if 𝛼 ∈ 𝑅𝑠 then −Δ𝑎 ∈ 𝑃 (1..𝑛);
(b) if 𝛼 ∈ 𝑅𝑑 then −𝜕𝑎 ∈ 𝑃 (1..𝑛);
(2) if 𝑎 = 𝑋𝑙 and 𝑋 ∈ {O, P}, then
(a) if 𝛼 ∈ 𝑅𝑠 then −Δ𝑋 𝑙 ∈ 𝑃 (1..𝑛);
(b) if 𝛼 ∈ 𝑅𝑑 then −𝜕𝑋 𝑙 ∈ 𝑃 (1..𝑛);
(3) if 𝑎 = ¬𝑋𝑙, then +𝜕𝑋 𝑙 ∈ 𝑃 (1..𝑛) with 𝑋 ∈ {O, P};
(4) if 𝑎 is 𝑎𝑐𝑡𝑖𝑣𝑒 (𝛼), then −𝜕EA 𝛼 ∈ 𝑃 (1..𝑛);
(5) if 𝑎 is ¬𝑎𝑐𝑡𝑖𝑣𝑒 (𝛼), then +𝜕EA 𝛼 ∈ 𝑃 (1..𝑛);
(6) if 𝑎 is 𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑 (𝛼), then +⊤𝛼 ∈ 𝑃 (1..𝑛);
(7) if 𝑎 is ¬𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑 (𝛼), then −⊤𝛼 ∈ 𝑃 (1..𝑛);
(8) if 𝑎 is 𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑 (𝛼), then +⊥𝛼 ∈ 𝑃 (1..𝑛);
(9) if 𝑎 is ¬𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑 (𝛼), then −⊥𝛼 ∈ 𝑃 (1..𝑛);
Based on these two definitions, we are now ready to define when
a rule is applicable/discarded. Note that the definition of being
discarded is obtained by applying the strong negation principle2 to
its positive counterpart.
Definition 2.3 (Applicable & Discarded). Assume a deontic defea-
sible theory 𝐷.
We say that a rule 𝛼 ∈ 𝑅 C ∪ 𝑅 P is applicable at 𝑃 (𝑛 + 1) iff for all
𝑎 ∈ 𝐴(𝛼) then 𝑎 is a trigger for 𝛼 at 𝑃 (𝑛 + 1).
We say that 𝛼 is discarded at 𝑃 (𝑛 + 1) iff there exists 𝑎 ∈ 𝐴(𝛼)
such that 𝑎 is a blank for 𝛼 at 𝑃 (𝑛 + 1).
For obligation rules, we say that 𝛼 ∈ 𝑅 O is applicable at index 𝑖
and 𝑃 (𝑛+1) iff (1) for all 𝑎 ∈ 𝐴(𝛼) then 𝑎 is a trigger for 𝛼 at 𝑃 (𝑛+1),
and (2) for all 𝑐 𝑗 ∈ 𝐶 (𝛼), 𝑗 < 𝑖, (2.1) if 𝛼 ∈ 𝑅s then +ΔO𝑐 𝑗 ∈ 𝑃 (1..𝑛),
if 𝛼 ∈ 𝑅d then +𝜕O𝑐 𝑗 ∈ 𝑃 (1..𝑛), and (2.2) +𝜕∼𝑐 𝑗 ∈ 𝑃 (1..𝑛).
We say that 𝛼 ∈ 𝑅 O is discarded at index 𝑖 and 𝑃 (𝑛 + 1) iff
either (1) there exists 𝑎 ∈ 𝐴(𝛼) such that 𝑎 is a blank for 𝛼 at
𝑃 (𝑛 + 1), or (2) there exists 𝑐 𝑗 ∈ 𝐶 (𝛼), 𝑗 < 𝑖, such that (2.1) if 𝛼 ∈ 𝑅s
then −ΔO𝑐 𝑗 ∈ 𝑃 (1..𝑛), if 𝛼 ∈ 𝑅d then −𝜕O𝑐 𝑗 ∈ 𝑃 (1..𝑛), or (2.2)
+𝜕𝑐 𝑗 ∈ 𝑃 (1..𝑛).
Intuitively, a rule is applicable when all its premises are proven.
Moreover, for an obligation rule, we also need to take into consider-
ation its ⊗-chain. A rule being applicable at a certain index greater
than 1 reflects the idea of compensation: all the previous elements
2 The strong negation principle is closely related to the function that simplifies a
formula by moving all negations to the inner most position in the resulting formula,
and replaces its positive tags with negative ones, and the other way around [2].
were in force obligations that have been violated. We are thus to
establish whether the element at the current index is an in force
obligation or not. If it is, we have another chance to be compliant
with the rule; if it is not, the whole norm cannot be complied with
as the previous step was indeed our last possibility to be compli-
ant. These concepts will be formalised and further explained in the
following with Definition 2.4 and proof conditions ±⊤ and ±⊥.
We are now ready to provide the strict and defeasible proof
conditions for constitutive statements [1]. In the following, we
shall omit the explanation of some negative proof conditions, as
they can be obtained via the strong negation principle.
+Δ𝑙: If 𝑃 (𝑛 + 1) = +Δ𝑙 then
(1) 𝑙 ∈ 𝐹 , or
(2) ∃𝛼 ∈ 𝑅sC [𝑙] is applicable.
A constitutive statement is strictly proven if is either a fact, or
there exists an applicable, strict rule for it. Note that inconsistencies
within a deontic defeasible theories can arise only if derived from
the strict part of the theory.
−Δ𝑙: If 𝑃 (𝑛 + 1) = −Δ𝑙 then
(1) 𝑙 ∉ 𝐹 and
(2) ∀𝛼 ∈ 𝑅sC [𝑙] is discarded.
A constitutive statement is strictly refuted (disproven, or rejected)
if it is not a fact and all the strict rules for it are discarded.
+𝜕𝑙: If 𝑃 (𝑛 + 1) = +𝜕𝑙 then
(1) +Δ𝑙 ∈ 𝑃 (1..𝑛), or
(2) −Δ∼𝑙 ∈ 𝑃 (1..𝑛) and
C [𝑙] is applicable and
(3) ∃𝛼 ∈ 𝑅sd
C
(4) ∀𝛽 ∈ 𝑅 [∼𝑙] either
(1) 𝛽 is discarded, or
(2) ∃𝜁 ∈ 𝑅 C [𝑙] s.t. 𝜁 is applicable and 𝜁 > 𝛽.
A constitutive statement is defeasibly proven if either was already
strictly proven, or the opposite is not and there exists an applicable
rule for the conclusion itself such that every opposite rule is either
discarded or defeated by an applicable supporting rule. Note that
whilst 𝛽 and 𝜁 can be defeaters, 𝛼 may not.
−𝜕𝑙: If 𝑃 (𝑛 + 1) = −𝜕𝑙 then
(1) −Δ𝑙 ∈ 𝑃 (1..𝑛) and either
(2) +Δ∼𝑙 ∈ 𝑃 (1..𝑛) or
(3) ∀𝛼 ∈ 𝑅sdC [𝑙] is discarded, or
(4) ∃𝛽 ∈ 𝑅 C [∼𝑙] s.t.
(1) 𝛽 is applicable and
(2) ∀𝜁 ∈ 𝑅 C [𝑙] either 𝜁 is discarded, or 𝜁 ≯ 𝛽.
A constitutive statement is defeasibly refuted if it was not strictly
proven, and either the opposite statement was strictly proven, or
all the applicable rules are either discarded or there exists an ‘unde-
feated’ opposite, applicable rule.
Proof conditions for obligations and permissions are updated
versions of the ones proposed in [5, 7].
+ΔO𝑙: If 𝑃 (𝑛 + 1) = +ΔO𝑙 then
(1) ∃𝛼 ∈ 𝑅sO [𝑙, 𝑖] is applicable at index 𝑖.
−ΔO𝑙: If 𝑃 (𝑛 + 1) = −ΔO𝑙 then
(1) ∀𝛼 ∈ 𝑅sO [𝑙, 𝑖] is discarded at index 𝑖.
For the defeasible derivation, we need to consider attacks even from
permissive rules.

72
Unravel Legal References in Defeasible Deontic Logic ICAIL’21, June 21–25, 2021, São Paulo, Brazil

+𝜕O𝑙: If 𝑃 (𝑛 + 1) = +𝜕O𝑙 then of the content of the prescriptive behaviour is derivable. Therefore,
(1) +ΔO𝑙 ∈ 𝑃 (1..𝑛), or given 𝐷 ⊢ +𝜕O𝑙, a violation is 𝐷 ⊢ +𝜕∼𝑙. Naturally, when consider-
(2) −Δ𝑋 ∼𝑙 ∈ 𝑃 (1..𝑛) with 𝑋 ∈ {O, P} and ing chains of compensatory behaviours, we would like not just to
(3) ∃𝛼 ∈ 𝑅sd O [𝑙, 𝑖] is applicable at index 𝑖 and discern norms that have been complied with against norms that
(4) ∀𝛽 ∈ 𝑅 [∼𝑙, 𝑗] ∪ 𝑅 P [∼𝑙] either
O have been violated, but we also want to discriminate norms that
(1) 𝛽 is discarded (at index 𝑗), or have never been violated against norms that have been complied
(2) ∃𝜁 ∈ 𝑅 O [𝑙, 𝑘] s.t. with but where at least one compensation has occurred.
𝜁 is applicable at index 𝑘 and 𝜁 > 𝛽. +⊤𝛼: If 𝑃 (𝑛 + 1) = +⊤𝛼 then
Note that (i) in Condition (4) 𝛽 can be a permission rule as explicit, (1) ∃𝑖.𝛼 ∈ 𝑅 O [𝑙, 𝑖] s.t.
opposite permissions represent exceptions to obligations, whereas (1) If (1) 𝛼 is applicable at index 𝑖 and 𝑃 (1..𝑛), and
𝜁 must be an obligation rule as a permission rule cannot reinstate an (2) +𝜕O𝑙 ∈ 𝑃 (1..𝑛),
obligation, and that (ii) 𝑙 may appear at different positions (indices (2) Then +𝜕𝑙 ∈ 𝑃 (1..𝑛).
𝑖, 𝑗, and 𝑘) within the three ⊗-chains. A norm (obligation rule) is complied with if either it is not appli-
−𝜕O𝑙: If 𝑃 (𝑛 + 1) = −𝜕O𝑙 then cable, or there exists and element of its ⊗-chain that is an in force
(1) −ΔO𝑙 ∈ 𝑃 (1..𝑛), and either obligation and the content of the obligation holds.
(2) +Δ𝑋 ∼𝑙 ∈ 𝑃 (1..𝑛) with 𝑋 ∈ {O, P}, or −⊤𝛼: If 𝑃 (𝑛 + 1) = −⊤𝛼 then
(3) ∀𝛼 ∈ 𝑅sd O [𝑙, 𝑖] is discarded at index 𝑖, or
(1) +𝛼 is applicable at index 1 and ∈ 𝑃 (1..𝑛)
(4) ∃𝛽 ∈ 𝑅 O [∼𝑙, 𝑗] ∪ 𝑅 P [∼𝑙] (2) +𝜕O𝑐 1 ∈ 𝑃 (1..𝑛),
(1) 𝛽 is applicable (at index 𝑗), and (3) +𝜕∼𝑐 1 ∈ 𝑃 (1..𝑛), and
(2) ∀𝜁 ∈ 𝑅 O [𝑙, 𝑘] either (4) ∀𝑐𝑖 ∈ 𝐶 (𝛼), with 𝑖 ≥ 2
𝜁 is discarded at index 𝑘 or 𝜁 ≯ 𝛽. (1) If 𝛼 is applicable at index 𝑖 and 𝑃 (1..𝑛), and
Follow the proof conditions for permissions. +𝜕O𝑐𝑖 ∈ 𝑃 (1..𝑛),
+ΔP𝑙: If 𝑃 (𝑛 + 1) = +ΔP𝑙 then (2) Then +𝜕∼𝑐𝑖 ∈ 𝑃 (1..𝑛).
(1) +ΔO𝑙 ∈ 𝑃 (1..𝑛), or A norm is not complied with when is applicable and all the in force
(2) ∃𝛼 ∈ 𝑅sP [𝑙, 𝑖] is applicable. elements of its ⊗-chain have been violated.
We can derive a permission if it was already proven as an obligation. +⊥𝛼: If 𝑃 (𝑛 + 1) = +⊥𝛼 then
−ΔP𝑙: If 𝑃 (𝑛 + 1) = −ΔP𝑙 then (1) 𝛼 is applicable at index 1 and 𝑃 (1..𝑛), and
(1) −ΔO𝑙 ∈ 𝑃 (1..𝑛) and (2) ∃𝑐𝑖 ∈ 𝐶 (𝛼), 𝑖 ≥ 2, 𝛼 is applicable at index 𝑖 and 𝑃 (1..𝑛).
(2) ∀𝛼 ∈ 𝑅sP [𝑙, 𝑖] is discarded. A violation occurs if at least one in force element of the norm’s ⊗-
Permissive defeasible derivations: chain has been violated. By Definition 2.3 and the fact that the norm
+𝜕P𝑙: If 𝑃 (𝑛 + 1) = +𝜕P𝑙 then is applicable both at index 1 by Condition (1) and at an index greater
(1) +ΔP𝑙 ∈ 𝑃 (1..𝑛), or than 1 by Condition (2), it follows that a(t least one) violation has
(2) −ΔO ∼𝑙 ∈ 𝑃 (1..𝑛), and either occurred. Note that, trivially, −⊤𝛼 implies +⊥𝛼, but not the other
(1) +𝜕O𝑙 ∈ 𝑃 (1..𝑛), or way around.
(2) ∃𝛼 ∈ 𝑅sd P [𝑙] is applicable and −⊥𝛼: If 𝑃 (𝑛 + 1) = −⊥𝛼 then either
O
(3) ∀𝛽 ∈ 𝑅 [∼𝑙, 𝑗] either (1) 𝛼 is discarded at index 1 and 𝑃 (1..𝑛), or
(1) 𝛽 is discarded at index 𝑗, or (2) 𝛼 ∈ 𝑅 O [𝑐, 2] is discarded at index 2 and 𝑃 (1..𝑛).
(2) ∃𝜁 ∈ 𝑅 P [𝑙]∪ ∈ 𝑅 O [𝑙, 𝑘] s.t. A norm has never been violated if either it was never applicable by
𝜁 is applicable (at index 𝑘) and 𝜁 > 𝛽. Condition (1), or the first element is an in-force, and complied with
Condition (1) states that if something is obligatory then is permit- obligation (guaranteed by Condition (2) as the norm was not ap-
ted. Condition (3) considers as possible counter-arguments only plicable at index 2). Symmetrically to what commented previously,
obligation rules as situations where both P𝑙 and P∼𝑙 hold are legal. −⊥𝛼 implies +⊤𝛼, but not the other way around.
−𝜕P𝑙: If 𝑃 (𝑛 + 1) = −𝜕P𝑙 then
Definition 2.4. Given a deontic defeasible theory 𝐷 and an appli-
(1) −ΔP𝑙 ∈ 𝑃 (1..𝑛), and either
cable at index 1, obligation rule 𝛼 : 𝐴(𝛼) ⇒O 𝑐 1 ⊗ · · · ⊗ 𝑐𝑚 , we say
(2) +ΔO ∼𝑙 ∈ 𝑃 (1..𝑛), or
that 𝛼 is
(1) −𝜕O𝑙 ∈ 𝑃 (1..𝑛), and
(2) ∀𝛼 ∈ 𝑅sd P [𝑙] is discarded or Strongly complied with: (i) −⊥𝛼, and (ii) +𝜕O𝑐 1 .
(3) ∃𝛽 ∈ 𝑅 O [∼𝑙, 𝑗] s.t. Weakly complied with: (i) +⊤𝛼 and (ii) +⊥𝛼.
(1) 𝛽 is applicable at index 𝑗 and Violated: +⊥𝛼.
Not complied with: −⊤𝛼.
(2) ∀𝜁 ∈ 𝑅 P [𝑙]∪ ∈ 𝑅 O [𝑙, 𝑘] either
𝜁 is discarded (at index 𝑘), or 𝜁 ≯ 𝛽. A norm to be complied with, violated, or not complied with has
We now have the formal tools to define compliance and violations. to be applicable: note that we could not consider such a requirement,
We adhere to the principle that a violation occurs when the logics but a discarded norm is always vacuously applicable.
proves both (i) that the prescription holds and (ii) that the opposite A norm is violated when at least one violation has occurred.

73
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
A norm is strongly complied with if it has never been violated:
the theory thus achieves the very first literal of the ⊗-expression.
A norm is weakly complied with when, on the contrary, is com-
plied with but also violated; this means that all the first 𝑗 −1 elements
of the ⊗-expression were violated obligations, whereas the 𝑗-th ele-
ment is an in force obligation and the theory proves the content of
the obligation.
This is exactly what marks the difference between being weak
complied with and being not complied with . In fact, a norm is not
complied with in two cases. The former, and more straightforward,
case is when all elements (of the ⊗-expression) were in force, vio-
lated obligations. The latter case is when the first 𝑗 − 1 elements
were in force, violated obligations, whilst the 𝑗-th element is not
proven as obligation (i.e. 𝐷 ⊢ −𝜕O𝑐 𝑗 ). Accordingly, it does not mat-
ter whether the theory achieves it or not, as the 𝑗 − 1-th element
was the last chance to comply with the norm.
We now enter in the final part of this section, and we shift
our attention from compliance to active rules. Next Definition 2.5
identifies those applicable rules whose conclusion was actually
proved. We name such rules provisionally active.
Definition 2.5 (Provisionally Active).
+𝜕PA 𝛼: If 𝑃 (𝑛 + 1) = +𝜕PA 𝛼 then
(1) 𝛼 is applicable at 𝑃 (1..𝑛), and
(2) +𝜕𝑋 𝐶 (𝛼) ∈ 𝑃 (1..𝑛), 𝑋 ∈ {𝑛𝑢𝑙𝑙, O, P}.
Definition 2.6 (Provisionally Inactive).
−𝜕PA 𝛼: If 𝑃 (𝑛 + 1) = −𝜕PA 𝛼 then either
(1) 𝛼 is discarded at 𝑃 (1..𝑛), or
(2) −𝜕𝐶 (𝛼) ∈ 𝑃 (1..𝑛), 𝑋 ∈ {𝑛𝑢𝑙𝑙, O, P}.
A fairly straightforward limitation of Definition 2.5 is that it fails
to identify which rules ‘effectively contributes’ in proving a claim:
a rule may very well be applicable, but not really necessary as the
following example shows.
Example 2.7. Assume that, in a deontic defeasible theory 𝐷, the
following rules are all applicable: 𝛼 1 , 𝛼 2 , and 𝛼 3 are for 𝑙, while 𝛽 1
and 𝛽 2 are for ∼𝑙. It also holds that 𝛽 1 > 𝛼 1 , 𝛼 2 > 𝛽 1 , 𝛽 2 > 𝛼 2 and
𝛼 3 > 𝛽 2 . In this context, rule 𝛼 1 does not contribute in defeating any
𝛽-like rule (it is actually defeated by 𝛽 1 ), and thus its contribution
is possibly limited/superfluous. On the other hand, 𝛼 2 is indeed
defeated by 𝛽 2 , but its role is fundamental because it is the only
𝛼-like rule that defeats 𝛽 1 .
We hence try to identify those rules that are undefeated (by any
applicable, opposite rule).
Definition 2.8 (Provisionally Active v2).
+𝜕PA 𝛼: If 𝑃 (𝑛 + 1) = +𝜕PA 𝛼 then
(1) 𝛼 is applicable at 𝑃 (1..𝑛),
(2) +𝜕𝐶 (𝛼) ∈ 𝑃 (1..𝑛), and
(3) ∀𝛽 ∈ 𝑅 [∼𝐶 (𝛼)].
If 𝛽 is applicable at (1..𝑛), Then 𝛽 ≯ 𝛼.
Definition 2.9 (Provisionally Inactive v2).
−𝜕PA 𝛼: If 𝑃 (𝑛 + 1) = −𝜕PA 𝛼 then either
(1) 𝛼 is discarded at 𝑃 (1..𝑛), or
(2) −𝜕𝐶 (𝛼) ∈ 𝑃 (1..𝑛), or
(3) ∃𝛽 ∈ 𝑅 [∼𝐶 (𝛼)].
74
Guido Governatori and Francesco Olivieri
If 𝛽 is applicable at (1..𝑛), Then 𝛽 > 𝛼.
In many real-life contexts is indeed important to identify such
undefeated rules, but in some other contexts such conditions may
be too restrictive as they fail to include rules that are fundamental
in proving a given claim. Consider indeed theory 𝐷 of Example 2.7:
𝐷 ⊬ +𝜕PA 𝛼 2 according to Definition 2.8 as it is defeated by 𝛽 2 even
if without 𝛼 2 it is not possible to prove +𝜕𝑙, as no other rule defeats
𝛽 1 . Consequently, our next definition will identify which rules are
essential in proving a given claim, and we will call them effectively
active rules.
Definition 2.10 (Effectively Active).
+𝜕EA 𝛼: If 𝑃 (𝑛 + 1) = +𝜕EA 𝛼 then
(1) 𝛼 is applicable at 𝑃 (1..𝑛),
(2) +𝜕𝑋 𝐶 (𝛼) ∈ 𝑃 (1..𝑛), 𝑋 ∈ {𝑛𝑢𝑙𝑙, O, P}, and
(3) 𝐷 \ {𝛼 } ⊢ −𝜕𝑋 𝐶 (𝛼).
This definition includes the concept of sine qua non: as per Con-
dition (3), 𝛼 is indeed a requisite in proving its conclusion 𝑙 as the
theory without 𝛼 cannot prove 𝑙. Note that the contrapositive ver-
sion of Condition (3), (3 ′ ) 𝐷 \ {𝜁 ∈ 𝑅 [𝐶 (𝛼)] | 𝜁 ≠ 𝛼 } ⊢ +𝜕𝑙 would
give us slightly different results3 .
Definition 2.11 (Effectively Inactive).
+𝜕EA 𝛼: If 𝑃 (𝑛 + 1) = +𝜕EA 𝛼 then either
(1) 𝛼 is discarded at 𝑃 (1..𝑛), or
(2) −𝜕𝑋 𝐶 (𝛼) ∈ 𝑃 (1..𝑛), 𝑋 ∈ {𝑛𝑢𝑙𝑙, O, P}, or
(3) 𝐷 \ {𝛼 } ⊢ +𝜕𝑋 𝐶 (𝛼).
We end this section by providing an example that illustrates a
proof in our logical apparatus. The explanation hereafter is not com-
plete as the point is to explain every key notion without slavishly
describe the whole procedure.
Example 2.12. Let 𝐷 = (𝐹 = {𝑎, 𝑏, 𝑐, 𝑑 }, 𝑅, >= {(𝛼, 𝛾), (𝜒, 𝜁 ), (𝜂, 𝜇)})
be the deontic defeasible theory such that
𝑅 = {𝛼 : 𝑎 ⇒ 𝑒 𝛽 :𝑏 ⇒𝑒 𝛾 : 𝑐 ⇒ ∼𝑒 𝜑 :𝑑 ⇒𝑞
𝜈 : 𝑐 ⇒ ∼𝑧 𝜎 : 𝑎, 𝑏 ⇒ 𝑤 𝜌 : ∼𝑎 ⇒ 𝑓
𝜒 : 𝑒 ⇒O ∼𝑞 𝜇 : 𝑑, 𝑞 ⇒O ∼𝑤 ⊗ 𝑠 𝜁 : 𝑎, 𝑑 ⇒O 𝑞
𝜂 : 𝑎, active(𝛼), violated( 𝜒) ⇒O 𝑧 ⊗ 𝑤 ⊗ 𝑣 }.
As 𝑎 is a fact, we prove +Δ𝑎 and, in cascade, +𝜕𝑎 as well as −𝜕∼𝑎.
(The same applies for the other three literals in the set of facts.)
This makes 𝛼 applicable as all its antecedents are proved, and 𝜌
discarded as its sole antecedent is refuted. For the same reasoning,
𝛽, 𝛾, 𝜑, 𝜈, 𝜎 are all applicable, while 𝜒, 𝜇, 𝜁 , and 𝜂 are applicable
at index 1. Since 𝛼 > 𝛾 and since no other applicable rule for 𝑒 is
stronger than 𝛾, we have that 𝛼 is effectively active (𝐷 ⊢ +𝜕EA 𝛼): in
fact, the theory without 𝛼 would not prove +𝜕𝑒. This is not the case
for 𝛽, but we can state that 𝛽 is provisionally active as no applicable
rule for ∼𝑒 is stronger than 𝛽. Symmetrically, 𝛾 is provisionally
inactive. Both 𝜒 and 𝜁 are applicable at index 1, and, since 𝜒 > 𝜁 ,
we conclude 𝐷 ⊢ +𝜕O ∼𝑞. This makes 𝜁 vacuously complied with.
Given that the theory also proves +𝜕𝑞 and 𝜒’s ⊗-chain consists of
only one element, we conclude that 𝜒 is not just violated but also
3 Inboth Conditions (3) and (3′ ), we used the notational simplification 𝐷 \ Γ to denote
the revision process of removing Γ from 𝑅 , and adjusting the superiority relation
accordingly.
Unravel Legal References in Defeasible Deontic Logic ICAIL’21, June 21–25, 2021, São Paulo, Brazil

not complied with (𝐷 ⊢ +⊥𝜒 and 𝐷 ⊢ −⊤𝜒). This, in turn, makes Algorithm 1: Compliance
𝜂 applicable at index 1, and since there are no obligation rules for Input: A deontic defeasible theory 𝐷
∼𝑧, we conclude 𝐷 ⊢ +𝜕O𝑧. Again, as the theory proves +𝜕∼𝑧, then Output: The defeasible meta-extension 𝐸 (𝐷)
𝜂 is violated, 𝐷 ⊢ +⊥𝜂. This time the ⊗-chain has other elements; 1 ±𝜕2 ← ∅ with 2 ∈ {C, O, P};
consequently, 𝜂 is now applicable at index 2. The second element 2 ±⊤ ← ∅; ±⊥ ← ∅; ±𝜕A ← ∅; ±𝜕EA ← ∅;
of 𝜂’s ⊗-chain is 𝑤: 𝜇, the obligation rule for ∼𝑤, is applicable but 3 InitialiseHerbrandBase(𝐻 𝐵);
weaker than 𝜂. Lastly, as 𝐷 ⊢ +𝜕𝑤 through 𝜎, we can conclude that 4 for 2𝑙 ∈ Lit ∪ ModLit do 𝑅 2 [𝑙 ]𝑖𝑛𝑓 𝑑 ← ∅; // with 2 ∈ {C, O, P}
𝜂 is weakly complied with (𝐷 ⊢ +⊤𝜂). In this theory, no norm was 5 for 𝛼 ∈ 𝑅 O do Initialise 𝛼 [𝑘 ] [2] to null, 𝑘 = length(𝐶 (𝛼));
strongly complied with. 6 for 𝑙 ∈ 𝐹 do Prove(𝑙, C); Refute(∼𝑙, C);
7 repeat
3 ALGORITHMS 8
± ← ∅;
𝜕2
for 2𝑙 ∈ 𝐻 𝐵, 2 ∈ {C, O, P} do
The algorithms presented in this section, given a deontic defeasible 9
10 if 𝑅 2 [𝑙 ] = ∅ then Refute(𝑙, 2);
theory as input, compute: (1) the defeasible extension of the theory,
11 if ∃𝛼 ∈ +𝜕A ∩ 𝑅 C [𝑙 ] then // 𝑙 is a non-deontic
(2) which rules are applicable/discarded, (3) which rules are effec-
literal
tively active/inactive, (4) which norms are (not) complied with, and
12 𝑅 C [∼𝑙 ]𝑖𝑛𝑓 𝑑 ← 𝑅 C [∼𝑙 ]𝑖𝑛𝑓 𝑑 ∪ {𝛽 ∈ 𝑅 C [∼𝑙 ] | 𝛼 > 𝛽 };
lastly (5) which norms are (never) violated.
13 if {𝛽 ∈ 𝑅 C [∼𝑙 ] | 𝛽 > 𝛼 } = ∅ then
The extension of a defeasible theory is, in a sketch, all that 14 Refute(∼𝑙, C);
the theory can prove and refute (disprove). Typically, a defeasible 15 if 𝑅 [∼𝑙 ] \ 𝑅 [∼𝑙 ]𝑖𝑛𝑓 𝑑 = ∅ then
extension is limited to the (deontic) literals of the theory itself, 16 Prove(𝑙, C);
but as in this case we are interested in understanding which rules 17 Active;
are applicable and which norms are (effectively) active, complied 18 end
with, violated, we shall extend the standard definition to include 19 end
such nuances. In the algorithms, we will associate proof tag +𝜕A to 20 end
applicable rules, and −𝜕A to discarded rules (thus 𝐷 ⊢ ±𝜕A 𝛼 for a 21 if ∃𝛼 + 𝜕A ∩ ∈ 𝑅 O [𝑙, 𝑖 ] ∧ ∀𝑗 < 𝑖.(𝛼 [ 𝑗 ] [1] =
theory 𝐷 and a rule 𝛼). + ∧ 𝛼 [ 𝑗 ] [2] = −) then
Note that the algorithms do no compute proof conditions for 22 𝑅 O [∼𝑙 ]𝑖𝑛𝑓 𝑑 ← 𝑅 O [∼𝑙 ]𝑖𝑛𝑓 𝑑 ∪ {𝛽 ∈ 𝑅𝑋 [∼𝑙 ] | 𝛼 > 𝛽 };
Definitions 2.5-2.9 for space reasons, as such an addition is straight- // with 𝑋 = {O, P}
forward and does not result in any increase of the complexity of the 23 if {𝛽 ∈ 𝑅 [∼𝑙 ] | 𝛽 > 𝛼 } = ∅ then
algorithms themselves. Also, for space reasons, the algorithms pre- 24 Refute(∼𝑙, 2);
sented in this work do not compute the strict part of the extension: 25 if (𝑅 O [∼𝑙 ] ∪ 𝑅 P [∼𝑙 ]) \ 𝑅 [∼𝑙 ]𝑖𝑛𝑓 𝑑 = ∅ then
to include that is a mundane task as all the defeasibility checks are 26 Prove(𝑙, O);
not taken into consideration. 27 Prove(𝑙, P);
Given a deontic defeasible theory 𝐷, 𝐻 𝐵𝐷 is the set of literals 28 Refute(∼𝑙, P);
such that the literal or its complement appears in 𝐷, where ‘appears’ 29 Active;
means that is a sub-formula of a literal occurring in the theory. The 30 end
deontic Herbrand Base of 𝐷 is 𝐻 𝐵 = {𝑋𝑙 | 𝑙 ∈ 𝐻 𝐵𝐷 ∧ 𝑋 ∈ {O, P}}. 31 end
Note that we do not consider reference literals in the Herbrand 32 end
Base. Accordingly, the extension of a deontic defeasible theory is 33 if ∃𝛼 ∈ +𝜕A ∩ 𝑅 P [𝑙 ] then
defined as follows. 34 𝑅 P [∼𝑙 ]𝑖𝑛𝑓 𝑑 ← 𝑅 P [∼𝑙 ]𝑖𝑛𝑓 𝑑 ∪ {𝛽 ∈ 𝑅 O [∼𝑙 ] | 𝛼 > 𝛽 };
// with 𝑋 = {O, P}
Definition 3.1. Given a deontic defeasible theory 𝐷 = (𝐹, 𝐷, >), 35 if {𝛽 ∈ 𝑅 [∼𝑙 ] | 𝛽 > 𝛼 } = ∅ then
we say that the extension is 𝐸 (𝐷) = (±𝜕C, ±𝜕O, ±𝜕P, ±𝜕A, ±𝜕EA, ±⊤, 36 Refute(∼𝑙, P);
±⊥), where ±𝜕2 = {𝑙 ∈ 𝐻𝐵𝐷 : 𝐷 ⊢ ±𝜕2𝑙 } with 2 ∈ {C, O, P}, 37 Refute(∼𝑙, O);
±𝜕3 = {𝛼 ∈ 𝑅| 𝐷 ⊢ ±𝜕3 𝛼 } with 3 = {A, EA}, ±⊤ = {𝛼 ∈ 𝑅| 𝐷 ⊢ 38 if (𝑅 P [∼𝑙 ] ∪ 𝑅 O [∼𝑙 ]) \ 𝑅 [∼𝑙 ]𝑖𝑛𝑓 𝑑 = ∅ then
±⊤𝛼 }, and ±⊥ = {𝛼 ∈ 𝑅| 𝐷 ⊢ ±⊥𝛼 }. 39 Prove(𝑙, P);
We say that two theories 𝐷 and 𝐷 ′ are equivalent iff 𝐸 (𝐷)+𝐸 (𝐷 ′ ) 40 Active;
(i.e., they have the same extension). 41 end
42 end
The next definition extends the concept of complement presented 43 end
in Section 2, and its sole purpose is to ease the notation of the 44 end
algorithms by establishing the logical connection among proved 45 ±𝜕2 ← ±𝜕2 ∪ 𝜕2 ±;
and refuted literals. 46 until 𝜕2 = ∅ and 𝜕2 = ∅;
+ −

f of literal 𝑋𝑙 as
Definition 3.2. We define the complement 𝑋𝑙 47 return 𝐸 (𝐷) = (±𝜕C , ±𝜕O , ±𝜕P , ±⊤, ±⊥, ±𝜕A , ±𝜕EA )

f = {∼𝑙 }.
• Trivially if 𝑋 = C, then 𝑋𝑙
f = {O∼𝑙, ¬O𝑙, P∼𝑙 }.
• O𝑙 A few important comments for the reader before presenting the
e = {¬P𝑙, O∼𝑙 }.
• P𝑙 algorithms.

75
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Guido Governatori and Francesco Olivieri

The algorithms presented here determine the extension of a whether there exists an applicable, opposite rule that is not defeated
deontic defeasible theory by computing, at each iteration step, a by any other applicable rule (for 2𝑙) than 𝛼.
simpler theory than the one at the previous step. By simpler, we
mean that, by proving and refuting literals and standard rules,
we can progressively simplify the rules of the theory itself: (i) by Procedure Prove
progressively eliminating elements from the antecedents of rules, Input: 𝑙 ∈ Lit, 2 ∈ {C, O, P}
and (ii) by eliminating rules that we know are either discarded, or 1
+ ← 𝜕 + ∪ {𝑙 };
𝜕2 2
defeated. 2 f
𝐻 𝐵 ← 𝐻 𝐵 \ ( {2𝑙 } ∪ 2𝑙);
Important for this goal is to note that, trivially, a rule with empty 3 −𝜕3 ← −𝜕3 ∪ {𝜁 ∈ 𝑅 | 2𝑙 f ∈ 𝐴(𝜁 ) }; // with 3 = {A, EA}
antecedent is vacuously applicable. We thus want to achieve, for 4 f ⊆ 𝐴(𝜁 ) };
>←> \{ (𝜁 ,𝜓 ), (𝜓, 𝜁 ) ∈> | 2𝑙
any rule, the status where the antecedent’s rule is empty (as this 5 switch 2 do
will simplify solving the superiorities), and we will do so by pro- 6 case 2 = C do
gressively eliminating elements from the antecedent as soon as they 7 +𝜕A ← +𝜕A ∪ {𝜁 ∈ 𝑅 | 𝐴(𝜁 ) \ {𝑙 } = ∅ };
satisfy the proper condition in Definitions 2.1 and 2.3. 8 𝑅 ← {𝐴(𝜁 ) \ {𝑙 } ↩→ 𝐶 (𝜁 ) | 𝜁 ∈ 𝑅 };
Symmetrically, when a rule is discarded, it can no longer supports 9 for 𝜁 ∈ 𝑅 O [𝑙, 𝑛] do // for 𝑛 ≤ 𝑘 = length(𝐶 (𝜁 ))
its conclusion, nor rejects the opposite. Therefore, as soon as an 10 𝜁 [𝑛] [2] ← +;
element satisfies the proper condition of Definitions 2.2 and 2.3, we 11 if 𝑛 = 1 ∧ 𝜁 [1] [1] = + then
can eliminate the corresponding rule from the set of rules. 12 −⊥ ← −⊥ ∪ {𝜁 };
We begin by populating the Herbrand Base, for every (deontic) 13 𝑅 ← {𝐴(𝜙) \ {¬violated(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
literal we create the support set 𝑅𝑖𝑛𝑓 𝑑 , and for every obligation rule 𝑅 } \ {𝜙 ∈ 𝑅 | violated(𝜁 )∈ 𝐴(𝜙) };
a 2-dimensional array which will simplify in checking conditions 14 else if 𝜁 ∈ +𝜕A ∧ 𝜁 [𝑛] [1] = + ∧ (∀𝑗 < 𝑛.𝜁 [ 𝑗 ] [1] =
of applicability and compliance. + ∧ 𝜁 [ 𝑗 ] [2] = −) then
Let us consider the theory proposed in Example 2.12. As 𝑎 is 15 +⊤ ← +⊤ ∪ {𝜁 }; +⊥ ← +⊥ ∪ {𝜁 };
in the set of facts, at the first iteration loop for at Line 6 invokes 16 𝑅 ← {𝐴(𝜙) \ {complied(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
Procedure Prove to prove it as constitutive. There, it is added to the 𝑅 } \ {𝜙 ∈ 𝑅 | ¬complied(𝜁 )∈ 𝐴(𝜙) };
support set +𝜕C (Line 1). We then eliminate 𝑎 from the antecedent 17 𝑅 ← {𝐴(𝜙) \ {violated(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
of 𝛼, which is now empty and so 𝛼 is applicable (Lines 7 and 8). As 𝑅 } \ {𝜙 ∈ 𝑅 | ¬violated(𝜁 )∈ 𝐴(𝜙) };
𝑎 does not appear in any ⊗-chain, Procedure Prove terminates, and 18 end
the main algorithm invokes Refute on ∼𝑎. 19 end
The set of defeasibly refuted constitutive literals is updated with 20 end
∼𝑎, and 𝜌 is discarded as its antecedent contains ∼𝑎; 𝜌 is also ef- 21 case 2 = O do
fectively inactive. The idea of these simplifications is taken from 22 +𝜕A ← +𝜕A ∪ {𝜁 ∈ 𝑅 | 𝐴(𝜁 ) \ {O𝑙, ¬O∼𝑙, ¬P∼𝑙 } = ∅ };
[5, 7]. 23 𝑅 ← {𝐴(𝜁 ) \ {O𝑙, ¬O∼𝑙, ¬P∼𝑙 } ↩→2 𝐶 (𝜁 ) | 𝜁 ∈
𝑅 } \ {𝜁 ∈ 𝑅 | O𝑙f ⊆ 𝐴(𝜁 ) };
The algorithm now enters the main cycle Repeat-Until at Lines
7–46. For every literal 𝑙 in HB, depending on which type of con- 24 for 𝜁 ∈ 𝑅 O [𝑙, 𝑛] do // for 𝑛 ≤ 𝑘 = length(𝐶 (𝜁 ))
𝜁 [𝑛] [1] ← +;
clusion is (2), we first verify whether there is any rule supporting 25
if 𝑛 = 1𝜁 [1] [2] = + then
it and, if not, we refute it (Line 10). Otherwise, if there exists an 26
27 −⊥ ← −⊥ ∪ {𝜁 };
applicable rule 𝛼 supporting it (ifs at Lines 11 for C, 21 for O, and 33
28 𝑅 ← {𝐴(𝜙) \ {¬violated(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
for P), we update the set of defeated rules supporting the opposite
𝑅 } \ {𝜙 ∈ 𝑅 | violated(𝜁 )∈ 𝐴(𝜙) };
conclusion 𝑅 2 [∼𝑙]𝑖𝑛𝑓 𝑑 (Lines 12, 22, 34): (i) in case of obligations,
29 else if 𝜁 ∈ +𝜕A ∧ 𝜁 [𝑛] [2] = + ∧ (∀𝑗 < 𝑛.𝜁 [ 𝑗 ] [1] =
both obligation and permission opposite rules (Condition (4) of
+ ∧ 𝜁 [ 𝑗 ] [2] = −) then
+𝜕O ), (ii) in case of permissions, only obligation rules (Condition
30 +⊤ ← +⊤ ∪ {𝜁 }; +⊥ ← +⊥ ∪ {𝜁 };
(3.2) of +𝜕P ). Given that 𝑅 [∼𝑙] contains all the opposite rules, and
31 𝑅 ← {𝐴(𝜙) \ {complied(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
given that we have just verified that 𝛼 for 𝑙 is applicable, we store
𝑅 } \ {𝜙 ∈ 𝑅 | ¬complied(𝜁 )∈ 𝐴(𝜙) };
in 𝑅 [∼𝑙]𝑖𝑛𝑓 𝑑 all those rules defeated by 𝛼. The next step is to verify
32 𝑅 ← {𝐴(𝜙) \ {violated(𝜁 )} ↩→2 𝐶 (𝜙) | 𝜙 ∈
whether there actually exists any rule supporting ∼𝑙 stronger than
𝑅 } \ {𝜙 ∈ 𝑅 | ¬violated(𝜁 )∈ 𝐴(𝜙) };
𝛼: if not, ∼𝑙 can be refuted (Lines 13, 23, 35).
33 end
The idea behind the ifs at Lines 15–18, 25–30, and 38–41, is: if
34 end
𝐷 ⊢ +𝜕_2𝑙, eventually the repeat-until cycle will have added to
35 end
𝑅 2 [∼𝑙]𝑖𝑛𝑓 𝑑 enough rules to defeat all opposite supports. When
36 case 2 = P do
that is the case, we invoke Prove on 𝑙, 2, and Refute on ∼𝑙, 2 (but 37 +𝜕A ← +𝜕A ∪ {𝜁 ∈ 𝑅 | 𝐴(𝜁 ) \ {P𝑙, ¬O∼𝑙 } = ∅ };
not in case of permission as it is legal to have P𝑙 and P∼𝑙 at the e ⊆ 𝐴(𝜁 ) };
38 𝑅 ← {𝐴(𝜁 ) \ {P𝑙, ¬O∼𝑙 } | 𝜁 ∈ 𝑅 } \ {𝜁 ∈ 𝑅 | P𝑙
same time).
39 end
When something is proved, the algorithm verifies whether 𝛼 is
40 end
effectively active via procedure Active. Such a procedure controls

76
Unravel Legal References in Defeasible Deontic Logic ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Procedure Refute Prove will then save ‘+’ in 𝜂 [2] [2] at Line 10, but nothing can be
Input: 𝑙 ∈ Lit, 2 ∈ {C, O, P} determined yet on compliance or violation. Later on, it proves −𝜕C𝑧
− −
1 𝜕2 ← 𝜕2 ∪ {𝑙 }; (and 𝜂 [1] [2] = −, Line 8 of Refute), and finally +𝜕O𝑧. This implies
2 𝐻𝐵 ← 𝐻𝐵 \ {2𝑙 }; that 𝜂 [1] [1] = + at Line 25 of Prove. Once the algorithm proves
3 −𝜕3 ← −𝜕3 ∪ {𝜁 ∈ 𝑅| 2𝑙 ∈ 𝐴(𝜁 )}; // with 3 = {A, EA}
+𝜕O𝑤, the else if test at Line 29 will succeed, and the algorithm
4 𝑅 ← 𝑅 \ {𝜁 ∈ 𝑅| 2𝑙 ∈ 𝐴(𝜁 )};
correctly establishes that 𝜂 is complied with (since 𝜂 [2] [1, 2] = +),
but also is violated (since 𝜂 [1] [1] = + but 𝜂 [1] [2] = −).
5 >← > \{(𝜁 ,𝜓 ), (𝜓, 𝜁 ) ∈> | 2𝑙 ∈ 𝐴(𝜁 )};
6 if 2 = C then
3.1 Computational properties
7 for 𝜓 ∈ 𝑅 O [𝑙, 𝑛] do // for 𝑛 ≤ 𝑘 =length(𝐶 (𝜓 ))
We discuss the computational properties of Algorithm 1 Compli-
8 𝜓 [𝑛] [2] ← −;
ance. Due to space reason, we only sketch the proofs by providing
9 if 𝜓 ∈ +𝜕A ∧ (∀𝑗 < 𝑛 =length(𝐶 (𝜓 ))𝜓 [ 𝑗] [1] =
to the reader the motivations of why our algorithms are sound,
+ ∧ 𝜓 [ 𝑗] [2] = −) then complete, and terminate (but we leave out all the technical details).
10 −⊤ ← −⊤ ∪ {𝜓 }; In order to discuss termination and computational complexity,
11 𝑅 ← {𝐴(𝜇) \ {¬complied(𝜓 )} ↩→2 𝐶 (𝜇)| 𝜇 ∈ we start by defining the size of a meta-theory 𝐷 as Σ(𝐷), as number
𝑅} \ {𝜇 ∈ 𝑅 | complied(𝜓 )∈ 𝐴(𝜇)}; of the occurrences of literals plus the number of occurrences of
12 end rules plus 1 for every tuple in the superiority relation.
13 end Note that, by implementing hash tables with pointers to rules
14 else if 2 = O then where a given literal occurs, each rule can be accessed in constant
15 for 𝜓 ∈ 𝑅 O [𝑙, 𝑛] do // for 𝑛 ≤ 𝑘 =length(𝐶 (𝜓 )) time. We also implement hash tables for the tuples of the superiority
16 if 𝜓 ∈ +𝜕A ∧ 𝑛 = 1 then relation where a given rule appears as either of the two element,
17 𝑅 ← {𝐴(𝜇) \ {complied(𝜓 ), and thus even those can be accessed in constant time.
¬violated(𝜓 )} ↩→2 𝐶 (𝜇)| 𝜇 ∈ 𝑅} \ ({𝜙 ∈
Theorem 3.3. Algorithm 1 Compliance terminates and its com-
𝑅 | {¬complied(𝜓 ),
plexity is 𝑂 (Σ4 ).
violated(𝜓 )} ∈ 𝐴(𝜙)} ∪ {𝜓 });
18 +⊤ ← +⊤ ∪ {𝜓 }; −⊥ ← −⊥ ∪ {𝜓 }; Proof. Termination of Procedures Prove, Refute, and Active
19 else is straightforward, as the size of the input theory is finite, and, at
20 𝜓 [𝑛] [1] ← −; every step, we modify finite sets. The complexity of Prove is 𝑂 (Σ2 ),
21 end the complexity of Refute is 𝑂 (Σ3 ) (two inner for loops of 𝑂 (Σ)
22 end each), and, lastly, the complexity of Active is 𝑂 (Σ).
Termination of Algorithm 1 Compliance is bound to termination
23 end
of the repeat-until cycle at Lines 7–46, as all other cycles loop
24 if 2 ∈ {O, P} then
over finite sets of elements of the order of 𝑂 (Σ). Given that 𝐻 𝐵
25 𝑅 ← {𝐴(𝜁 ) \ {¬2𝑙 }| 𝜁 ∈ 𝑅};
and 𝑅 are finite, and since every time a literal is proved/refuted, it
26 end
is removed from the corresponding set, the algorithm eventually
empties such a set, and, at the next iteration, no modification to the
extension can be made. This proves the termination of Algorithm 1
Procedure Active Compliance.
Input: Regarding its complexity, note that: (1) all set modifications are
1 if {𝛽 ∈ 𝑅 [∼𝑙] 𝑖𝑛𝑓 𝑑 | 𝛼 > 𝛽} \ {𝜓 ∈ 𝑅 [∼𝑙] 𝑖𝑛𝑓 𝑑 | 𝜁 > 𝜓 ∧ 𝜁 ≠
made in linear time, and (ii) the aforementioned repeat-until cycle
𝛼 ∧ 𝜁 ∈ +𝜕A ∩ 𝑅 [𝑙]} ≠ ∅ then is iterated at most 𝑂 (Σ) times, and so are the two for loop at lines
2 +𝜕EA ← +𝜕EA ∪ {𝛼 }; 9–44. This would suggest that the repeat-until cycle runs in 𝑂 (Σ2 ).
3 𝑅 ← {𝐴(𝜁 ) \ {active(𝛼)} ↩→2 𝐶 (𝜁 )| 𝜁 ∈ 𝑅} \ {𝜁 ∈ A more discerning analysis shows that the complexity is actually
𝑅 | ¬active(𝛼)∈ 𝐴(𝜁 )}; 𝑂 (Σ): the complexity of the for cycle cannot be considered sepa-
4 else rately from the complexity of the external repeat-until loop, while
5 −𝜕EA ← −𝜕EA ∪ {𝛼 }; instead they are strictly dependent. Indeed, the overall number of
6 𝑅 ← {𝐴(𝜁 ) \ {¬active(𝛼)} ↩→2 𝐶 (𝜁 )| 𝜁 ∈ 𝑅} \ {𝜁 ∈ operations made by the sum of all loop iterations cannot outrun
𝑅 | active(𝛼)∈ 𝐴(𝜁 )}; the number of occurrences of the literals or rules (𝑂 (Σ) + 𝑂 (Σ)),
7 end
because the operations in the inner cycles directly decrease, iter-
ation after iteration, the number of the remaining repetitions of
the outmost loop, and the other way around. This sets the overall
complexity of Algorithm 1 Compliance to 𝑂 (Σ4 ). □
We conclude the analysis of the algorithms by seeing in more
details how compliance and violations are verified, and we do so Theorem 3.4. Algorithm 1 Compliance is sound and complete:
by continuing the analysis of Example 2.12. (1) 𝐷 ⊢ +𝜕𝑋 𝑝 iff 𝑝 ∈ +𝜕𝑋 of 𝐸 (𝐷), 𝑋 ∈ {C, O, P}, 𝑝 ∈ Lit
At a certain iteration, the algorithm will prove +𝜕C𝑤, but assume (2) 𝐷 ⊢ +𝜕𝑌 𝛼 iff 𝛼 ∈ +𝜕𝑌 of 𝐸 (𝐷), 𝑌 ∈ {A, EA}, 𝑝 ∈ Lit
that so far we have proven neither +𝜕O𝑧, nor +𝜕O𝑤. Procedure (3) 𝐷 ⊢ +⊤𝛼 iff 𝛼 ∈ +⊤ of 𝐸 (𝐷), 𝛼 ∈ Lab

77
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
(4) 𝐷 ⊢ +⊥𝛼 iff 𝛼 ∈ +⊥ of 𝐸 (𝐷), 𝛼 ∈ Lab
(5) 𝐷 ⊢ −𝜕𝑋 𝑝 iff 𝑝 ∈ −𝜕𝑋 of 𝐸 (𝐷), 𝑋 ∈ {C, O, P}, 𝑝 ∈ Lit
(6) 𝐷 ⊢ −𝜕𝑌 𝛼 iff 𝛼 ∈ −𝜕𝑌 of 𝐸 (𝐷), 𝑌 ∈ {A, EA}, 𝑝 ∈ Lit
(7) 𝐷 ⊢ −⊤𝛼 iff 𝛼 ∈ −⊤𝛼 of 𝐸 (𝐷), 𝛼 ∈ Lab
(8) 𝐷 ⊢ −⊥𝛼 iff 𝛼 ∈ −⊥𝛼 of 𝐸 (𝐷), 𝛼 ∈ Lab.
Proof. The aim of Algorithm 1 Compliance is to compute a
defeasible extension of the input theory through successive trans-
formations on the set of facts, rules and the superiority relation.
These transformations act in a way to obtain a simpler theory while
retaining the same extension. By simpler theory we mean a theory
with less symbol in it. For instance, given a theory 𝐷 such that
𝐷 ⊢ +𝜕O 𝑝, we can remove O𝑝 from the antecedent of the rules
since a such deontic literal no longer plays any role in the rule, and
we can delete all the rules where P¬𝑝 is in the antecedent, as such
rules can no longer conclude literals (and we also know that the
rule is no longer applicable); analogously, when 𝐷 ⊢ +⊤𝛼, we can
remove the instances of 𝑐𝑜𝑚𝑝𝑙𝑖𝑒𝑑 (𝛼) from the rule (and remove the
rules where 𝑣𝑖𝑜𝑙𝑎𝑡𝑒𝑑 (𝛼) is in the antecedent). The theory obtained
from 𝐷 from such operations is equivalent to 𝐷 ′ , with respect to
the elements of the Herbrand base and labels still in the theory.
The proof that the above transformation produces theories equiv-
alent to the original one is by induction on the length of derivations
and contrapositive. □ [4] Matteo Cristani, Francesco Olivieri, and Antonino Rotolo. 2017. Changes to
4 CONCLUSIONS AND RELATED WORK
References are widespread in legal documents, and they have been
a topic of intensive research in the field of AI and Law (e.g., citations
networks, automated detection of citations, citation and reference
navigations, . . . ). Despite their pervasive presence, the study of how
to logically represent them with the aim of exploit them for the
digitalisation of legislation has been largely neglected. The OASIS
LegalRuleML standard [11] introduces the terms “comply” and
“violated” (accepting an argument pointing to a legal rule), but the
development of a method to solve such references is well beyond the
scope of the standard. The task of solving such references has been
tackled by [10] adopting the technique of importing the content of
the citation (with the appropriate semantic layer), but the approach
is restricted to a shallow import (ignoring attacking rules).
The idea of employing terms denoting the legal status of pro-
visions goes back, at least, to the seminal work by Sartor [13].
However, when the approach is used, the information about such
terms is either given as part of the input of a case (and it is not de-
termined by the other “facts" of the case and the rules), or addressed
using the techniques exemplified by (5) and (4), and not dealt with
at the logic level, and focusing, at best, on the first reading of “ap-
plicable”. Often, the use of “applicable” is to facilitate some form
of non-monotonic reasoning, avoiding the adoption of a priority
relation over rules.
In this paper, we started from the same idea of [13], but with
the direct focus on handling the references and not to facilitate
some form of defeasible reasoning. Accordingly, references are
78
Guido Governatori and Francesco Olivieri
first class citizens in our logic with an equal treatment. This has
allowed us to provide an efficient computational treatment of the
references. The other major benefit of the approach we adopted is
that it enables encodings of pieces of legislation to strictly adhere to
the legal isomorphism principle [3] that facilitates the translation
from provisions in natural language to their formal representation
and their maintenance when the provisions are amended.
References are a prominent feature in norms amending other
norms. Norm change can be modelled by nested rules [4, 9] and
[12] develops algorithms to compute the extension of a defeasible
theory with nested norms. We plan to investigate how to integrate
the techniques proposed in this work with the algorithms in [12]
to implement the logics of [4, 9].
REFERENCES
[1] Grigoris Antoniou, David Billington, Guido Governatori, and Michael J. Maher.
2001. Representation results for defeasible logic. ACM Trans. Comput. Log. 2, 2
(2001), 255–287. https://doi.org/10.1145/371316.371517
[2] Grigoris Antoniou, David Billington, Guido Governatori, Michael J. Maher, and
Andrew Rock. 2000. A Family of Defeasible Reasoning Logics and its Implemen-
tation. In ECAI 2000, Proceedings of the 14th European Conference on Artificial
Intelligence, Berlin, Germany, August 20-25, 2000, Werner Horn (Ed.). IOS Press,
459–463.
[3] Trevor J. M. Bench-Capon and Frans Coenen. 1992. Isomorphism and legal
knowledge based systems. Artif. Intell. Law 1, 1 (1992), 65–86. https://doi.org/10
.1007/BF00118479
temporary norms. In Proceedings of the 16th edition of the International Conference
on Artificial Intelligence and Law, ICAIL 2017, London, United Kingdom, June 12-
16, 2017, Jeroen Keppens and Guido Governatori (Eds.). ACM, 39–48. https:
//doi.org/10.1145/3086512.3086517
[5] Guido Governatori, Francesco Olivieri, Antonino Rotolo, and Simone Scannapieco.
2013. Computing Strong and Weak Permissions in Defeasible Logic. J. Philos.
Log. 42, 6 (2013), 799–829. https://doi.org/10.1007/s10992-013-9295-1
[6] Guido Governatori, Francesco Olivieri, Simone Scannapieco, and Matteo Cristani.
2011. Designing for Compliance: Norms and Goals. In RuleML 2011-America
(LNCS, Vol. 7018). Springer, 282–297. https://doi.org/10.1007/978-3-642-24908-
2_29
[7] Guido Governatori, Francesco Olivieri, Simone Scannapieco, Antonino Rotolo,
and Matteo Cristani. 2016. The rationale behind the concept of goal. Theory Pract.
Log. Program. 16, 3 (2016), 296–324. https://doi.org/10.1017/S1471068416000053
[8] Guido Governatori and Antonino Rotolo. 2006. Logic of Violations: A Gentzen
System for Reasoning with Contrary-To-Duty Obligations. Australasian Journal
of Logic 4 (2006), 193–215. arXiv:9307/main.pdf http://ojs.victoria.ac.nz/ajl/artic
le/view/1780
[9] Guido Governatori and Antonino Rotolo. 2010. Changing legal systems: legal
abrogations and annulments in Defeasible Logic. Logic Journal of IGPL 18, 1
(2010), 157–194.
[10] Ho-Pun Lam and Mustafa Hashmi. 2019. Enabling reasoning with LegalRuleML.
Theory Pract. Log. Program. 19, 1 (2019), 1–26. https://doi.org/10.1017/S1471068
418000339
[11] OASIS. 2017. LegalRuleML core specification version 1.0. Standard Specification.
OASIS. http://docs.oasis-open.org/legalruleml/legalruleml-core-spec/v1.0/cspr
d02/legalruleml-core-spec-v1.0-csprd02.html
[12] Francesco Olivieri, Guido Governatori, Matteo Cristani, and Abdul Sattar. [n.d.].
Computing Defeasible Meta-logic. In Logics in Artificial Intelligence - 17th Euro-
pean Conference, JELIA (LNCS, Vol. 12678), Wolfgang Faber, Gerhard Friedrich,
Martin Gebser, and Michael Morak (Eds.). Springer, 69–84. https://doi.org/10.1
007/978-3-030-75775-5_6
[13] Giovanni Sartor. 1991. The Structure of Norm Conditions and Nonmonotonic
Reasoning in Law. In Proceedings of the Third International Conference on Artificial
Intelligence and Law, ICAIL ’91, Richard E. Susskind (Ed.). ACM, 155–164. https:
//doi.org/10.1145/112646.112665
[14] Giovanni. Sartor. 2005. Legal Reasoning: A Cognitive Approach to the Law.
Springer.
 Context-Aware Legal Citation Recommendation using Deep
Learning
Zihan Huang∗ Daniel E. Ho
Charles Low∗ Mark S. Krass
Mengqiu Teng∗ Stanford University
Hongyi Zhang∗
Language Technologies Institute
Carnegie Mellon University
ABSTRACT
Lawyers and judges spend a large amount of time researching the
proper legal authority to cite while drafting decisions. In this paper,
we develop a citation recommendation tool that can help improve
efficiency in the process of opinion drafting. We train four types
of machine learning models, including a citation-list based method
(collaborative filtering) and three context-based methods (text sim-
ilarity, BiLSTM and RoBERTa classifiers). Our experiments show
that leveraging local textual context improves recommendation,
and that deep neural models achieve decent performance. We show
that non-deep text-based methods benefit from access to struc-
tured case metadata, but deep models only benefit from such access
when predicting from context of insufficient length. We also find
that, even after extensive training, RoBERTa does not outperform a
recurrent neural model, despite its benefits of pretraining. Our be-
havior analysis of the RoBERTa model further shows that predictive
performance is stable across time and citation classes.
CCS CONCEPTS
• Applied computing → Law; Document analysis; • Information
systems → Data mining; Recommender systems; • Computing
methodologies → Natural language processing.
KEYWORDS
citation recommendation, citation normalization, legal text, legal
opinion drafting, neural natural language processing
ACM Reference Format:
Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho,
Mark S. Krass, and Matthias Grabmair. 2021. Context-Aware Legal Citation
Recommendation using Deep Learning. In Eighteenth International Con-
ference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São
∗ Authors contributed equally to the paper.
† Corresponding author (matthias.grabmair@tum.de). Current affiliation at TUM; work
largely conducted while employed at SINC as part of adjunct affiliation with Carnegie
Mellon University, Language Technologies Institute.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466066
79
Matthias Grabmair†
Department of Informatics
Technical University of Munich
SINC GmbH
Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/
3462757.3466066
1 INTRODUCTION
Government agencies adjudicate large volumes of cases, posing
well-known challenges for the accuracy, consistency, and fairness of
decisions [2, 27]. One of the prototypical mass adjudicatory agencies
in the U.S. context is the Board of Veterans’ Appeals (BVA), which
makes decisions on over fifty thousand appeals for disabled vet-
eran benefits annually. Due to these case volumes and constrained
resources, the BVA suffers from both a large backlog of cases and
large error rates in decisions. Roughly 15% of (single-issue) cases
are appealed and around 72% of appealed cases are reversed or
remanded by a higher court [14]. These challenges are typical for
agencies like the Social Security Administration, the Office of Medi-
care Hearings and Appeals, and the immigration courts, which
adjudicate far more cases than all federal courts combined. Lawyers
and judges are hence in great need of tools that can help them
reduce the cost of legal research as they draft decisions to improve
the quality and efficiency of the adjudication process.
Advancing the application of machine learning to suggesting
legal citations is essential to the broader effort to use AI to assist
lawyers. Citations are a critical component of legal text in common-
law countries. To show that a proposition is supported by law,
writers cite to statutes passed by a legislature; to regulations writ-
ten by agencies implementing statutes; and to cases applying legal
authorities in a particular context. Such is the importance of cita-
tions to legal writing that the traditional method of selecting law
students to edit law journals has been a gruelling test on the cor-
rect format of legal citations [30]. Achieving performance on more
difficult tasks, like text generation and summarization, depends on
a sophisticated treatment of citations.
This paper reports on experiments evaluating a series of machine
learning tools for recommending legal citations in judicial opinions.
We show that deep learning models beat ordinary machine learning
tools at recommending legal citations on a variety of metrics, which
suggests that the neural models have a stronger capability to exploit
semantics to understand which citation is the most appropriate.
We also demonstrate the importance of context in predicting
legal citations. For ordinary text-based machine learning models
with limited capacity for detecting semantic meaning, structured
contextual metadata improves virtually all predictions. For deep
learning models, the utility of structured metadata emerges only
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark S. Krass, and Matthias Grabmair
in sufficiently difficult settings where there may be a weaker se-
mantic link between the input and the target, and only for certain
models. Still, this result shows the potential importance of context
to citation predictions. Deep learning models that are able to bet-
ter incorporate contextual cues from semantic inputs are likely to
outperform methods without such capabilities.
Because the BVA corpus has never been made available to the
research community, we are releasing the text for single-issue de-
cisions, with legal citation tokenization, case metadata, and our
source code upon publication at: https://github.com/TUMLegalTech/
bva-citation-prediction. We believe many other advances can be
built on this as a benchmark for natural language processing in
law.
2 RELATED WORK
2.1 Citation Recommendation
Citation recommendation is a well-studied problem in the domain
of academic research paper recommendation, as researchers seek
help to navigate vast literatures in their fields. Many of the ap-
proaches are transferable to the legal context. They can be broadly
categorized into citation-list based methods, which characterize
a query document by the incomplete set of citations it contains
and provide a global recommendation of citations relevant to the
entire document, and context-based methods, which take a particu-
lar segment of text from the query document and provide a local
recommendation that is relevant to that specific context [26].
2.1.1 Citation-List Based Methods. In this setting, the researcher
is drafting a paper and has an incomplete set of citations on hand,
and seeks to find additional relevant papers.
An early approach in [28] applies collaborative filtering to this
task. There, citing papers are “users” and citations are “items.” Given
a new user, the algorithm locates existing users with similar prefer-
ences to the new user, and recommends items popular among the
existing users. Matrix factorization methods project the sparse, high-
dimensional user-item adjacency matrix onto a low-dimensional
latent space and compare similarity in this latent space. For exam-
ple, [4] uses Singular Value Decomposition to find the latent space,
and finds performance gains over ordinary collaborative filtering.
Graph-based approaches treat research papers as nodes and
citations as edges (directed or undirected), and use graph-based
measures of relevance to find relevant nodes to an input set corre-
sponding to the researcher’s incomplete set of citations. Examples
include the Katz measure [23], PageRank [31] and SimRank [19].
[12] applies a topic-sensitive version of the PageRank algorithm
by up-weighting papers in the incomplete set. [37] finds the Katz
measure of node proximity to be a significant feature.
The citation-list based approach has its drawbacks. It puts the
burden of creating a partial list of citations on the user. Attorneys
who are new to veterans’ law would face the well-known “cold-start”
problem, where they have difficulty generating enough citations as
input to receive quality recommendations. Second, attorneys draft-
ing an opinion may be more interested in local recommendations
relevant to their current section of work rather than global recom-
mendations that are generally relevant to the entire case. Third,
80
citation-list based approaches do not exploit the rich information
contained in the textual context of each citation.
2.1.2 Context-Based Methods. In this setting, the researcher inputs
a span of text (the query context), which can be a particular sentence
or paragraph, instead of a list of citations. The system recommends
local citations relevant to this query context.
Traditional information retrieval approaches directly compare
the words in the query context to the words in the title, abstract, or
full text of each cited document, and apply scoring models such as
Okapi BM25 [1] or Indri [38] to arrive at a similarity score that is
used to rank documents. However, as [33] observes, the full text of
cited documents is often noisy and may not contain words similar
to those used to describe the document as a whole. This problem is
especially pertinent in law. Legal decisions and statutes sometimes
lack informative titles, and the key legal implications are often
buried in a mountain of other factual or procedural details.
Intuitively, we expect the span of text preceding or surrounding
a citation (its citation context) to contain useful information pertain-
ing to the content of the cited document and the reason for citation.
This information can then be used for retrieval. [3, 33] demonstrate
that indexing academic papers using words found in their citation
contexts improves retrieval. He et al. [13] develop this idea further
by representing each paper as a collection of citation contexts, and
then using a non-parametric similarity measure between a query
context and each paper for recommendation. Huang et al. [16] use
a neural network to learn word and document representations to
perform similarity comparison in that space. More recently, in a
work most similar to our approach, Ebesu and Fang [10] directly
train an encoder-decoder network with attention for context-aware
citation prediction and find that adding embeddings representing
the citing and cited authors improves predictions.
2.2 Legal Citation Prediction
Because of the importance of citations to legal writing [8], prior
work has explored machine-generated recommendations for legal
authorities relevant to a given legal question.
A number of commercial tools claim to assist users in legal re-
search using citations. Zhang and Koppaka [43] describe a feature in
LexisNexis that allows users to traverse a semantics-based citation
network in which relevance is determined by textual similarity be-
tween citation contexts. Other commercial offerings include ROSS
Intelligence [18], CaseText’s CARA A.I. [17] and Parallel Search [6],
as well as Quick Check by Thomson Reuters [40]. The methodology
of such offerings is largely proprietary.
Winkels et al. [41] develop a prototype legal recommender sys-
tem for Dutch immigration law, which allows legal professionals
to search a corpus by clicking on articles of interest; the system
returns cases with the highest between-ness centrality with the
article. In [8], Dadgostari et al. consider the task of generating a
bibliography for a citation-free legal text by modelling the search
process as a Markov Decision Process in which an agent iteratively
selects relevant documents. At each step, the agent can choose
whether to explore a new topic in the original paper or to select a
relevant paper from the current topic of focus. An optimal policy
is learned using Q-learning. They find this adaptive algorithm to
Context-Aware Legal Citation Recommendation using Deep Learning
# Values Most Frequent Least Frequent
Class (# cases) Class (# cases)
Year 19 2009 (22,801) 2017 (3,651)
Issue Area 17 Service Con- Increased
nection for Rating for
Bodily Injury Nerve Damage
Claims (38,956) (2,921)
VLJ 289 Anonymized Anonymized
(6,159) (6)
Table 1: Summary Statistics of Corpus Metadata Variables.
outperform a simpler method, based on proximity to the original
document, on the task of retrieving U.S. Supreme Court decisions.
Other works [11, 21] have tangentially analyzed properties of
legal citation networks, exploring measures of authority and rele-
vance of precedents, as well as macro characteristics of the network,
such as degree distribution and shortest path lengths. Sadeghian et
al. [35] develop a system to automatically identify citations in legal
text, extract their context and predict the reason for the citation
(e.g., legal basis, exception) based on a curated label set.
3 DATA
3.1 The BVA Corpus
The BVA corpus we use contains the full text of over 1 million appeal
decisions from 1999 to 2017. Accompanying each decision is a set of
metadata derived from the Veterans Appeals Control and Locator
System (VACOLS), which includes fields such as the decision date,
diagnostic codes indicating the veteran’s injuries, the case outcome,
and an indicator for whether the case was subsequently re-appealed.
Each case also contains one or more ‘issue codes,’ which are hand-
coded by BVA attorneys and categorize the key legal or factual
questions raised (e.g., “entitlement to a burial benefit”). This paper
focuses on a subset of 324,309 cases that raise a single issue and
have complete metadata, although our methods can be generalized
to the full corpus.
We hypothesized that three metadata features would contribute
to model performance. First, we included the year of the decision, to
reflect changes in citation patterns as new legal precedents emerge
over time. Second, we constructed an issue area feature to reflect the
substantive issues presented in each case, which we hypothesize to
provide strong priors for the type of citations contained within as
well. The BVA has a hierarchical coding system comprising program
codes, issue codes, and diagnostic codes to categorize each issue.
For simplicity and class balancing, we curated a composite issue
area variable with 17 classes (see Figure 1). Third, we included a
feature referring to the Veterans’ Law Judge (VLJ) who handled the
case. This corresponds to the hypothesis that citation patterns vary
with the idiosyncrasies of individual judges, inspired in part by [10].
Judge names were anonymized and judges with 5 cases or fewer
were collapsed into a single unknown judge category. Summary
statistics for these metadata are included in Table 1.
3.2 Decision Text Preprocessing
American legal citations follow a predictable format governed by
[7]. Case citations, for instance, identify the parties to the case; the
81
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
reporter containing the case; and finally the page in the reporter
where the case begins. Thus, a citation to Brown v. Board of Education
of Topeka would begin as follows: Brown v. Board of Education, 347
U.S. 483. This indicates that the first page of Brown is found on
the 483rd page of the 347th volume of the United States Reports.
The volume-reporter-page citation is usually a unique identifier for
each case.1 Citations to statutory law follow a similar three-part
pattern: “18 U.S.C. § 46,” means the 46th section of the 18th title of
the United States Code.
These three-part citation patterns form the basis for our text
preprocessing pipeline. We first use a series of regular expressions
to identify, clean, and classify citations from opinions. We then
build a vocabulary of legal authority using publicly-available lists
of valid cases and statutes. We use this vocabulary to extract all
citations from case texts and represent them using standardized
indices. We describe this process in greater detail below.
3.3 Citation Preprocessing
The large raw citation vocabulary obtained from running regular
expression extractors on every case is normalized into classes of
case, statute, regulation, and unknown citations.
For cases, this normalization involves matching the volume, re-
porter, and first/last page interval derived from the citation string
with an authoritative list of cases found in the CaseLawAccess
(CLA) metadata.2 If an extracted citation can be matched to a CLA
metadata entry, it is replaced with a reference to that entry in the
citation vocabulary during tokenization. For example, the extrac-
tion ‘Degmetich v. Brown, 8 Vet. App. 208 (1995)’ is resolved to the
normalized ‘Degmetich v. Brown, 8 Vet. App. 208, CLA#6456776’
(i.e. CLA metadata entry 6456776), which becomes an entry in the
citation vocabulary that is used for all identifiable references to the
same case. Citations to the U.S. Code and to the Code of Federal
Regulations are extracted using patterns based on the ‘<chapter>
U.S.C. <tail>’ and ‘<chapter> C.F.R. <tail>’ anchors. The tail typi-
cally consists of one or more section elements, which we break into
individual elements that each become their own normalized citation
with the same anchor and chapter (e.g., ‘18 U.S.C. §§ 46(a), 46(b)
becomes the two entries ‘18 U.S.C. § 46(a)’ and ‘18 U.S.C. § 46(b)’).
All citations that cannot be normalized into either case, code, or
regulation classes will form the ‘unknown’ class. Once normalized,
the vocabulary is further reduced by removing all citation entries
which occur less than 20 times in the training cases and resolving
them to an ‘unknown citation’ token. This threshold was manu-
ally chosen as a suitable tradeoff between extensive coverage of
citations and baseline frequency to enable the model to learn.
The training data contains about 5M extracted citation instances
comprising roughly 97k unique strings. Our normalization proce-
dure reduces this to a citation vocabulary of size 4287, of which
4050 ( ≈ 94.5%) are normalized (1286 cases, 870 statutes, 1894 reg-
ulations). The normalized entries cover about 98.5% of citations
1 Summary dispositions of a case are sometimes reported in a table, such that multiple
cases appear on a single physical page.
2 CLA is a public-access project that has digitized the contents of major case reporters
[5]. We include the Vet. App. and F.3d reporters, which contain veterans’ law cases
and cases from the Federal Courts of Appeal, as these account for the vast majority of
cases cited in the corpus.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark S. Krass, and Matthias Grabmair
5: Accrued/Dental/New
1: Other
6: Bodily Injury
0: No 2: Dependents
7: Eye/Ear/Respiratory
3: Effective Date
8: Organ Damage
4: Total Disability
9: Nerve
10: Psychological
Compensation Claim?
Service Connection Issue?
Yes
Increased Rating Issue?
11: Not Schedular 12: Body Injury 13: Eye/Ear/Respiratory
14: Organ Damage 15: Nerve 16: Psych
Figure 1: Issue area categories.
occurring in the tokenized decisions. This reduction effect is pri-
marily due to (a) complex statutory citations breaking apart into a
smaller set of atoms, (b) different page-specific citations to a case
getting collapsed into a single CLA entry, and (c) different forms
of variation reduction (e.g., removal of trailing parentheses with
years, etc.).
The final vocabulary is then used to normalize all citations en-
countered in case texts. Citation strings are extracted and replaced
with a placeholder. The case/code/regulation procedure outlined
above is applied to each citation string to obtain a list of one or
more corresponding normalized citations. Each of these is kept if it
is contained in the final vocabulary, or replaced with the ‘unknown
citation’ otherwise. The resulting sequence of vocabulary index
tokens is re-inserted at the location of the general citation place-
holder after the text has been tokenized. Note that only citations
containing reporter and page references are extracted and regu-
larized. Short form citations (e.g., ‘id.’) are treated as ordinary text
and are excluded from the pool of prediction targets.3 We also do
not treat quotations in the text in any special way and rely on the
tokenizer to capture them as part of the context window.
4 PROBLEM DEFINITION
We model the legal citation prediction problem as follows. Suppose
a BVA staff attorney is drafting an opinion regarding an appeals
proceeding. We refer to this document as 𝑑. The incomplete draft
may already contain several citations to authority. We call this
incomplete set c𝑑 ⊂ 𝐶, where 𝐶 represents the entire corpus of
legal authorities, comprising possibly relevant cases, statutes, and
regulations. The first task we consider is to predict the next ci-
tation 𝑐 ∗ ∈ 𝐶 \ c𝑑 that is globally relevant to the opinion, given
the incomplete set c𝑑 . This corresponds to the citation-list based
approach. In our experiments, for a document that contains 𝑀 cita-
tions, we model the incomplete list c𝑑 by taking the first 𝑚 citations
(1 ≤ 𝑚 < 𝑀) from the document. We then seek to predict the next
citation, i.e., the (𝑚 + 1)-th citation 𝑐 ∗ given c𝑑 .
Alternatively, the attorney may be more interested in legal au-
thority specific to the current segment of the opinion he/she is
3 While this choice may limit the pool of prediction targets, it does not threaten the
integrity of predictions themselves. By convention, short-form citations always follow
full-form citations, which we detect. Because the system only has access to left context,
it cannot ‘cheat’ by reference to short-hand citations.
82
working on. We represent this text segment of interest by a se-
quence of tokens as the query context b𝑑 = {𝑏 1, ..., 𝑏𝑙 }. The second
task is thus to predict the next upcoming citation 𝑐 ∗ ∈ 𝐶 that is lo-
cally relevant to context b𝑑 . This corresponds to the context-aware
approach. Specifically in our experiments, given a query context b𝑑
of length 𝑙 in the document 𝑑, we seek to predict the first citation
that occurs in the upcoming forecast window of length 𝑤. We vary
length 𝑙 and forecast window 𝑤 depending on the method (see
Section 5).
In addition, metadata describing characteristics of the draft deci-
sion may also aid in citation prediction. For instance, the relevance
and validity of case citations can change over time as new prece-
dents emerge and others are overruled. Since many of the relevant
legal standards are specific to particular classes of claims, the issue
code feature may help identify relevant citations. Finally, different
VLJs may have different propensities to cite certain authorities.
5 METHODS
Our main metrics are recall at 1, recall at 5 and recall at 20, that is,
the proportion of data instances where the correct next citation is
among the model’s top 1, 5, and 20 predictions. Precision would
not be an informative metric as we are only seeking to predict the
single correct next citation. Recall at 1 reflects a restrictive user
that expects the system to predict a single citation only. Recall at
5 simulates what we think is the typical user, who benefits from a
small number of recommendations that can quickly be examined
for the most appropriate one. A longer list of 20 simulates users
seeking to get a bigger picture of what could possibly be relevant.
We split the 324,309 single-issue BVA cases into 233,506 (72%), 58,370
(18%) and 32,433 (10%) cases for the training, validation and test
set, respectively. Each model is trained on the training set, tuned
on the validation set, and tested on a 6-fold split of the test set to
measure statistical uncertainty.
We implement four different methods on the task of legal cita-
tion prediction on the BVA corpus, and examine their compara-
tive performance: a citation-only collaborative filtering system, a
context-similarity based model, a BiLSTM recurrent neural classi-
fier, and a RoBERTa-based classifier that has been pretrained on a
language model objective. We note that our task is related to legal
language generation (e.g., [32]). However, evaluating the citation
prediction ability of a language generation model is significantly
more difficult. Citations would need to be captured dynamically
during a parameter-dependent generation process, validated, and
resolved against the vocabulary. By contrast, the neural models
in this project are implemented as conceptually straightforward
classifiers, allowing us to test their ability to read the context well
enough to forecast what will be cited next. We plan to tackle citation
prediction as language generation in future work.
5.1 Collaborative Filtering
Our first experimental model uses collaborative filtering, a common
recommender system technique based on the assumption that sim-
ilar users will like similar items. Transferred to our setting, each
BVA decision document is treated as a user, and each citation is
seen as an item. The prediction task then takes as input the citations
that are already cited in a BVA draft opinion (which can be seen
Context-Aware Legal Citation Recommendation using Deep Learning
as the items that a user has liked), and returns other citations that
similar documents have also cited.
Formally, assume that the corpus of BVA cases 𝐶 has 𝑉 authori-
ties that can be cited. Then every document 𝑑 ′ can be represented
by a sparse vector v𝑑 ′ ∈ R𝑉 , each of whose dimensions v𝑑 ′,𝑐 in-
dicates an importance score of a citation 𝑐 to the document. If
citation 𝑐 is cited in a document, possible scoring functions could
include a binary representation (v𝑑 ′,𝑐 = 1), a term frequency vector
(tf), and a tf-idf vector that incorporates the inverse document fre-
quency (idf). With such a representation, a set of document vectors
{v𝑑 ′ : 𝑑 ′ ∈ D} can be constructed from a document collection D.
Given a draft of a BVA opinion 𝑑, its incomplete citation set
c𝑑 can also be summarized into a document vector v𝑑 . We use
a collaborative filtering approach known as the user-based top-𝐾
recommendation algorithm. The algorithm first identifies the 𝐾 doc-
uments D𝐾 (𝑑) that are most similar to 𝑑 from the collection based
on their vector representations v, based on their cosine similarity:
v𝑑 · v𝑑 ′
sim(v𝑑 , v𝑑 ′ ) = . uments, and words that contained digits. The most frequent 25,000
∥v𝑑 ∥ 2 ∥v𝑑 ′ ∥ 2
The algorithm then finds candidate citations based on what these
documents cite. An average of these document vectors weighted
by their similarities gives the final recommendation. Specifically,
the recommendation score of citation 𝑐 for document 𝑑 is given by
Í
𝑑 ′ ∈D (𝑑) sim(v𝑑 , v𝑑 ′ )v𝑑 ′,𝑐
score(𝑑, 𝑐) = Í 𝐾 . per citation. Metadata features are incorporated into the model in
𝑑 ′ ∈D𝐾 (𝑑) sim(v𝑑 , v𝑑 ′ )
In our experiments, the document vectors are collected from the
training set. The number of top similar documents 𝐾 is a hyper-
parameter that can be tuned, and 𝐾 = 50 is chosen for the results
reported. From our trials with three different scoring functions
for the document vectors, binary scoring proved to be the most
effective choice and was used throughout the experiments.
To incorporate metadata features, a score is assigned to each
categorical feature 𝑓𝑖 , namely the probability of citing the citation
𝑐 after conditioning on that feature:
score(𝑓𝑖 , 𝑐) = 𝑃 (𝑐 | 𝑓𝑖 ).
We take a weighted average of these features and the output
of the collaborative filtering algorithm. We adopt the commonly
used svmRank algorithm of [20] to learn weights for each feature.
We extract all citation occurrences in a random sample of 1000
documents from the training set, perform a pairwise transformation
on the data, apply min-max normalization on the pairwise data, and
train a linear Support Vector Machine (SVM) on the normalized data.
The final score is a linear combination of individual feature scores
using the learned weights. Citations suggested by the recommender
system are reranked by their final scores and the top citations are
chosen as final predictions.
5.2 Text Similarity
The second model uses a context-aware bag-of-words approach to
predict citations. Previous studies, such as [13, 34], have demon-
strated that the local context of words surrounding each citation
occurrence can be used as a compact representation of the cited
document to improve retrieval effectiveness, much like how in-link
text is used to improve web retrieval. By contrast to collaborative
filtering, this approach does not require the user to input an existing
83
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
set of citations. Instead, the words in a section of interest within the
draft opinion are used as a query to find the most relevant citation
based on textual similarity of the present context to the previous
contexts associated with each citation. Such local citation recom-
mendations have the added advantage of relevance to a particular
section of the opinion.
Formally, we adopt the approach of [13], which represents each
context by its tf-idf vector (normalized to have an L2-norm of
1). Each citation 𝑐 is represented by a collection of tf-idf vectors
{b 𝑗 : 𝑗 = 1, 2, · · · , 𝑘𝑐 }, where each b 𝑗 represents the local context
of one citation occurrence and 𝑘𝑐 is the number of times 𝑐 was
cited in the training set. Given a query context b𝑑 at test time, the
relevance of each citation 𝑐 to the query is then calculated as:
1 Õ
𝑘𝑐
score(b𝑑 , 𝑐) = (b · b 𝑗 ) 2
𝑘𝑐 𝑗=1 𝑑
We removed stopwords, words that occurred in less than 10 doc-
words were then chosen as a vocabulary. We used the 50 words
preceding (instead of surrounding) each citation as its context, in
line with our task to recommend relevant upcoming citations.4 Cita-
tions that occurred within each context were also used as part of the
vocabulary. As some citations were very frequently cited, we col-
lected at most 100 randomly chosen context vectors (i.e. 𝑘𝑐 ≤ 100)
a way similar to the Collaborative Filtering model (see Section 5.1).
Each feature is assigned a score and an SVM model is trained to
learn feature weights to produce the final score.
5.3 Bi-directional Long Short Term Memory
LSTMs [15] are a popular form of recurrent neural networks and
serve as a well-known baseline for deep neural network models.
Variants using LSTM remain competitive in various NLP tasks
[22, 25, 29]. BiLSTM (Bi-directional LSTM) improves on the original
LSTM by reading inputs in both forward and backward directions.
We adopted a two-layer BiLSTM on the BVA corpus for citation
prediction. Just like the text similarity baseline, this approach per-
forms local citation recommendation. It takes a sequence of words
within the draft opinion as the query context, and predicts which
citation is most likely to be cited next given the context. Going
beyond the text similarity model, we predict the first citation that
appears within a forecasting window of fixed length.
Formally, a sequence of tokens b𝑑 = {𝑏 1, ..., 𝑏𝑙 } is extracted from
each document 𝑑 as the query context and we seek to predict the
immediate next citation in the upcoming forecasting window of
length 𝑤. The query context is encoded using pre-trained byte-
level Byte Pair Encoding (BPE) [36]. For comparability with the
RoBERTa model, we use the ‘roberta-base’ tokenizer provided by
Huggingface [42], which has a vocabulary of about 50k tokens.
The citation vocabulary indices are re-inserted after encoding, re-
placing the general citation token to generate the final encoded
tokens as described in Section 3.3. The encoded tokens are fed into
an embedding layer followed by two stacked bi-directional LSTM
4 Notethat this means citations are always the very next word after the context. This
contrasts with the neural models presented below, where citations may appear at some
distance from the context.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark S. Krass, and Matthias Grabmair
layers to produce a sequence of hidden states. The hidden state
corresponding to the last token is used as the aggregate represen-
tation of the query context and flows into the classification head,
which consists of two dropout/linear combination layers separated
by a tanh activation, followed by a softmax layer to produce output
probabilities for each citation, indicating how likely they will be
cited next. Figure 2 illustrates the detailed architecture.
Figure 2: The BiLSTM model architecture.
To incorporate the metadata information, processed metadata
features are concatenated to the last hidden state before the clas-
sification layers as illustrated in Figure 2. The year and issue area
features are one-hot encoded. The VLJ feature is projected into
a three-dimensional vector space by feeding the VLJ ID into an
embedding layer that can be inspected after training.
Our training setup follows common settings for language anal-
ysis experiments of this size. We use an embedding size of 768
and a hidden size of 3072. We compute CrossEntropy loss against
a one-hot target vector of the same length as the vocabulary. To
facilitate stable convergence, we use an effective batch size of 512
(implemented via gradient accumulations across 4 batches of 128
to fit onto Nvidia P100 GPUs). We use an Adam optimizer with a
fixed learning rate of 1𝑒 −4 .
5.4 Pretrained RoBERTa-based Classifier
Since the introduction of BERT [9], language model pre-training
has gained immense popularity, leading to models with superior
performance on many NLP tasks and reductions in the amount
of task-specific training data required. Its core mechanism is to
compute a layer-wise self-attention across all tokens in the text,
which allows it to effectively capture long-distance interactions
without the architectural restrictions imposed by sequential mod-
els. RoBERTa [24] further improved BERT by employing certain
techniques, such as longer training, and key hyperparameter ad-
justment. We apply this model to our task via transfer learning
to test how a Transformer model pretrained on a language model
objective performs against our BiLSTM model trained from scratch.
We fine-tuned a pre-trained RoBERTa model (HuggingFace’s
‘roberta-base’ [42]) on the BVA corpus using the citation predic-
tion task. The model uses 24 layers, a hidden size of 1024, 16 self-
attention heads, leading to 355M parameters overall. We apply a
common sequence classification architecture and, similar to our
BiLSTM model, feed the final hidden layer’s output through two
84
Figure 3: The RoBERTa model architecture.
dropout/linear layers separated by a tanh activation to produce the
final hidden vector 𝐶 in the citation vocabulary size.
To fine-tune our RoBERTA model, we use the same data prepro-
cessing and loading as in the BiLSTM experiment. We tokenize the
BVA decisions with the pre-trained RoBERTa tokenizer provided
by Huggingface [42] and apply our citation extraction and normal-
ization procedure. Sequences are padded to the same length and an
attention mask is generated to indicate whether the corresponding
token is a padding token. Formally, a pair of tensors (b𝑑 , a𝑑 ) is
extracted from each document 𝑑, where b𝑑 represents the token
ids and a𝑑 represents attention mask. Label l𝑑 is the index in the
citation vocabulary of the first citation following the given context
b𝑑 . We compute cross-entropy loss between predictions p𝑑 for (b𝑑 ,
a𝑑 ) and label l𝑑 .
Again, to allow training with relatively large batches on NVidia
P100 GPUs, we use a batch size of 192 and accumulate gradients
for three steps before performing back-propagation, resulting an
effective batch size of 576. We use the AdamW optimizer with a
learning rate of 1𝑒 −4 .
5.5 Sampling-based Data Loading
Each data instance for the BiLSTM and RoBERTa models consists of
a context window and forecasting window to the left and right side
of an offset. During every training epoch, and during evaluation,
we sample a random offset for each case from all offsets whose fore-
casting window contains a citation token. We designed data loading
this way to mitigate the prohibitively large space of traversing all
possible context/forecasting window combinations for all citations
in all cases. Note that, because the target is always the first citation
within the forecasting window, our data loading is biased against
citations that rarely appear first in strings of successive citations.
We plan to address this imbalance in future work.
6 RESULTS AND DISCUSSION
6.1 General Performance
Table 2 shows the full citation prediction results of the four models.
We add a naive majority vote baseline, which always recommends
the 20 most popular citations in descending order of their number
of occurrences in the training data.
We first turn to our ordinary machine learning models. A com-
parison to their ‘original’ setting — without access to structured
metadata — shows the importance of semantic context for citation
Context-Aware Legal Citation Recommendation using Deep Learning ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Model Setting Recall@1 Recall@5 Recall@20

Majority Vote Original 1.73% (.02%) 7.35% (.03%) 26.4% (.02%)
Original 10.2% (.06%) 25.5% (.07%) 45.4% (.08%)
Collaborative Original + Year 9.68% (.05%) 24.8% (.08%) 45.2% (.09%)
Filtering Original + Year + IssueArea 9.64% (.05%) 24.7% (.08%) 45.2% (.08%)
Original + Year + IssueArea + VLJ 9.60% (.05%) 24.7% (.09%) 45.2% (.09%)
Original 16.4% (.03%) 41.1% (.04%) 66.2% (.05%)
Text Original + Year 20.4% (.05%) 48.2% (.05%) 79.5% (.03%)
Similarity Original + Year + Class 16.2% (.06%) 51.6% (.07%) 82.6% (.05%)
Original + Year + Class + VLJ 16.6% (.06%) 51.7% (.06%) 82.7% (.05%)
no metadata (47 epochs) 65.2% (.33%) 81.8% (.14%) 91.1% (.11%)
BiLSTM
all metadata (50 epochs) 65.8% (.35%) 82.4% (.26%) 91.3% (.16%)
no metadata (106 epochs) 65.6% (.33%) 82.8% (.31%) 91.7% (.21%)
RoBERTa
all metadata (126 epochs) 66.2% (.30%) 83.2% (.17%) 92.1% (.20%)
Table 2: Prediction results. Each model is evaluated on six folds of the test set and the numbers reported are the mean and the
standard error (in parentheses) of recall at 1, 5, and 20. Neural models are trained using 256/128 context/forecast windows. All
metadata includes year, issue area, and VLJ identifiers.

prediction. The collaborative filtering model uses only the previous
citations in a document as input. It returns the correct citation as its
top-ranked recommendation 10.2% of the time; recall@5 is 25.5%.
By contrast, the text similarity baseline achieves a recall@1 of 16.4%
and a recall@5 of 41.1%, on average. This is strong evidence that
the textual context preceding a citation is a critical signal. By con-
trast, the document-level statistical information on citation patterns
leveraged by collaborative filtering is less informative.
For the text similarity model, adding metadata information gen-
erally gives a noticeable improvement over predictions based on
text alone. For example, adding structured information on the year
of a decision improves performance, which suggests that the model
does not otherwise detect temporal information. But not all meta-
data is equally useful. Adding information on the identity of the
judge produces little or no marginal gain. Further, we do not find
evidence that metadata enhances the collaborative filtering model.
Interestingly, the benefit in recall@1 of case year information is
negated when class is added, although recall@5 and recall@20 im-
prove at the same time. If one were to pursue the baseline further,
this effect should be examined.
For purposes of this comparison experiment, we train our BiL-
STM and RoBERTa models on a context window of 256 tokens
and a forecast window of 128 tokens. They are trained until, in
our assessment, validation metrics indicated convergence, at which
point they dramatically outperform both baselines. Both predict the
correct citation roughly 65-66% of the time and produce a recall@5
of around 81-83% using the textual context alone. The neural mod-
els’ improvement over the text similarity baseline suggests that the
ability to encode more complex semantic meanings—and track long-
term dependencies across context windows of significant length—
noticeably improves performance in citation recommendations.
We experimented with different metadata combinations for the
neural models with 8 epochs of training time and observed no clear
differences, and decided to only train all-meta and no-meta models
until convergence. Giving the BiLSTM and RoBERTa models access
to metadata improves predictive performance by around 0.2-0.6%.
That delta, however, is mostly within two standard errors of the
two models. Our two possible explanations are (a) that the neural
models are capable of implicitly inferring some background features
from the legal text itself, and thus they will not benefit much from
us providing this information explicitly, and (b) that metadata may
not carry much signal for this task.
The superior neural model performance is intuitive in legal text
also because the text preceding a citation will typically paraphrase
a legal principle or statement that is reflective of that source. We
can assume that some portion of our context-forecast instances
consist of relatively easy examples. To some degree, short-distance
citation prediction can in fact be considered a sentence similarity
task. Commercial search engines even use text encoding similarity
to suggest cases to cite for a particular sentence (e.g., [6]). Similarly,
literal quotations from the source preceding the citation can be
certain indicators. However, a pure memorization approach will fail
for longer forecast distances, as one can anticipate an upcoming
cited source from the narrative progression in the text before it
becomes lexically similar to the source closer to the citation. An
exception to this consists of large spans of boilerplate text that
contain citations and are reused across decisions. To investigate the
capacity of our models to anticipate citations from further away,
we experiment with different forecasting lengths (see Section 6.2
and 6.4 below).
A final observation is the stability of predictive performance
across the six test set folds as evidenced by the low standard errors.
The neural models have slightly more deviation than the baselines
and the BiLSTM and RoBERTa models metric are generally within
the reach of ±2 standard errors within a given recall metric.
6.2 Context & Forecasting Window Sizes
To further explore the behavior of the deep neural models, we con-
ducted an ablation study, in which we varied the size of the context
and forecasting windows and varied the availability of structured
metadata information. We tested 12 different settings for BiLSTM

85
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark S. Krass, and Matthias Grabmair

Text Alone Text & Metadata Text Alone Text & Metadata Text Alone Text & Metadata
0.63 0.81
0.90
0.62 0.80

64 Ahead

64 Ahead

64 Ahead
0.61 0.79 0.89
0.60 0.78 0.88

Recall @ 20
Recall @ 1

Recall @ 5
0.59 0.77 0.87
0.58 0.76
0.76 0.88
0.56
0.87

128 Ahead

128 Ahead

128 Ahead
0.54 0.74 0.86
0.72 0.85
0.52
0.84
0.50 0.70 0.83
64 128 256 64 128 256 64 128 256 64 128 256 64 128 256 64 128 256
Context Window Context Window Context Window
Model BiLSTM RoBERTa

Figure 4: Results of the ablation study for Recall at 1, 5, and 20. Within each panel, the most difficult tasks are in the bottom
left corner and the easiest tasks are in the top right. The x-axis shows the context window. “64 ahead” and “128 ahead” refer to
the maximum number of tokens between the context window and the target citation. Error bars are 95% confidence intervals.

and RoBERTa, respectively. In this grid-search experiment, all mod- 1.0

case
els were trained for 8 epochs before test metrics were computed. regulation
The detailed settings and the results are illustrated in Figure 4. 0.8 statute
As expected, increasing the forecasting window hurts perfor-
mance by weakening the semantic link between input and target. citation recall @ 1 0.6
Also unsurprising is the upward slope in each panel, which simply
shows that providing the models with more context generally im- 0.4
proves predictions. But the utility of added context changes with
the difficulty of the forecasting task. When the target citation is 0.2
nearer to the context (‘64 ahead’), we observe diminishing returns
to context: A 128-token context window is only slightly better than 0.0
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
a 64-token context window. When the target is further away (‘128
decision year
ahead’), more context helps. We hypothesize that adding additional Figure 5: Per-class recall@1 for RoBERTa all-meta model
context helps to compensate for the difficulty of the task: the net- over time.
work models are able to infer more clues for the citation given the
extra information. As the size of the forecasting window increases,
the potential for a weaker semantic relationship between the im-
mediate context and the eventual citation makes it more helpful for
the model to have access to additional context.
decision that the loss decrease had slowed down enough to stop
We observe a similar story with respect to structured metadata:
training was a matter of judgment. Given the carbon footprint of
the harder the task, the more helpful it is to add metadata. In the
neural model training [39], we believe such ablation research on
BiLSTM framework, metadata is most helpful when the model is
large neural models should be conducted with care.
given little context and when the targets are far away. But when
the target is nearby (‘64 ahead’), performance is statistically indis-
tinguishable between models that do have access to metadata and 6.3 Pre-Training vs. Training From Scratch
those that do not. The findings align with our hypothesis that, when Despite its general English language-model pretraining, RoBERTa
enough context is given, the neural network models are able to models do not show noticeable superiority over BiLSTM in the ab-
derive the clues for citations from text snippets, and thus obviating lation experiments, even when a more challenging task and a more
the need for metadata information. complicated context is given. For the models trained until conver-
Although the ablation experiment was conducted with only 8 gence in Table 2, RoBERTa performs better than the BiLSTM model,
epochs of training, our experimental results on models trained un- but only by at most 1% and often with overlapping ±2 standard
til convergence are in line with the observation that the effect of errors. One possible explanation is that the pretraining of RoBERTa
metadata is only marginal. At the very least, however, our results is performed on non-legal text, negating the pretraining benefit
indicate that a conclusive exploration of the effectiveness of indi- for this domain-specific task. Alternatively, the task may not re-
vidual metadata features in training neural citation recommenders quire sophisticated language understanding and/or our supervised
may require considerable computational resources or the use of setup provides sufficient training to learn citation prediction from
advanced techniques to reduce training time. Even after 126 and 50 scratch. We leave an exploration of the effects of domain-specific
epochs, respectively, our models showed no signs of overfit and the pretraining (e.g. using [44]) in this task for future work.

86
Context-Aware Legal Citation Recommendation using Deep Learning ICAIL’21, June 21–25, 2021, São Paulo, Brazil

200000 1.0
175000 Distance 𝑁 Recall@1 Recall@5 Recall@20
citation count in training data

150000 0.8 1-16 13609 78.7 91.9 97.1

17-32 11237 75.6 89.8 95.9

125000
100000
75000
50000
25000
0 0.0
0 250 500 750 1000 1250 1500 1750 2000
citations sorted by recall
Figure 6: Per-citation recall@1 vs. number of instances in
training data for RoBERTa all-meta model.
6.4 Error Analysis
Figure 5 shows a relatively consistent recall at 𝑘 = 1 performance
across classes over time. We see a slight downward slope for the
case and regulation metric towards the end of our analysis period.
This may be due to opinions later in the time period potentially con-
taining new citations and patterns occurring less frequently in the
training data. The plot exhibit a single strong upwards oscillation
in 2002-2003. We believe this is likely due to litigation surrounding
the Veterans Claims Assistance Act of 2000, which sparked mass
remands by the BVA back to regional offices. This relative shape of
the per-class recall graphs stays roughly the same for larger values
of 𝑘, albeit shifted to higher absolute recall levels.
To assess the influence of the sampling distribution, the com-
bined scatterplot in Figure 6 plots the recall at 𝑘 = 1 achieved for
each citation against its frequency as a prediction target in the
training data. Of the 2037 different citations that were loaded in a
single pass over the test data (of the total of 4287; see Section 5.5),
only about 1200 citations are predicted with non-zero recall. At
𝑘 = 20 this number increases to about 1700 and the red curve shifts
right (not shown). The distribution of blue data points indicates
that almost all zero-recall citations occur with very low, or zero,
frequency. However, citations with high recall do not follow a rec-
ognizable frequency pattern. This is informative for the cold-start
problem of new sources becoming available that have not been cited
enough yet to be learned by models such as the ones presented
here. We are aware of this limitation and leave it for future work.
Finally, we examined whether the number of decisions in the test
data authored by a judge correlated with the model’s performance
in predicting citations from those decisions, but did not find clear
patterns. The three-dimensional judge embeddings also did not
reveal any clear separation with regard to the per-judge recall.
We intend to investigate the relationship between attributes of
individual VLJs and the behavior of trained models in future work.
To help characterize the underlying behavior of the models, we
drew a sample of 200 erroneous predictions generated by a long-
trained RoBERTa model similar to the one in Table 2.5 Two sets of
observations indicate that the model has developed some conceptual
mapping of citations. First, 16% of the erroneous predictions did
appear in the forecast window, somewhere after the first citation.
Idiosyncrasies in citation order might explain these errors, but there
5 After
qualitative error analysis was completed, a pre-processing bug was corrected,
leading to changes in recall values of less than 0.5%. Quantitative results and analyses
of converged models reported here are from this slightly improved version.
citation recall @ 1
0.6 33-48 9125 68.8 85.2 93.2
49-64 7082 63.3 81.1 91.0
0.4 65-80 5452 55.6 75.4 87.6
81-96 4534 52.7 73.1 87.1
0.2
97-112 3918 47.9 69.6 84.0
113-128 3403 42.1 66.2 82.7
Table 3: Roberta all-meta performance binned by token dis-
tance from beginning of forecasting window to target cita-
tion, based on single pass over validation set.
is no conceptual mismatch. Second, somewhere around 5% of the
errors involve regulations that implement a particular statute. For
example, one case cites 38 C.F.R. § 3.156(a), a regulation defining
when veterans may present “new and material evidence” to reopen
a claim. The model predicted a citation to 38 U.S.C. § 5108(a), which
is precisely the statute commanding the BVA to reopen claims when
veterans identify “new and material evidence.” Again, the erroneous
prediction is in exactly the right conceptual neighborhood.
Consistent with our ablation analysis, our review of the errors
suggests the critical role that topical changes in long texts play in
generating errors. Table 3 shows recall metrics for targets binned
by the position of the target citation within the forecast window
between minimum and maximum distances. Since legal analysis
is often addressed in a single section of an opinion, close citations
are more frequent than distant ones. Unsurprisingly, performance
decreases with distance from the context window. From closest to
farthest bin, recall@1 shrinks by a relative 47%, recall@5 by 28%,
and recall@20 by 15%. This behavior is intuitive and indicates that
the system may indeed memorize contexts immediately surround-
ing citations. Still, the gradual decline in performance, especially
for recall@5, suggests that the model is learning some amount of
longer-distance patterns. This forms evidence that effective citation
recommendation benefits from both a sophisticated representation
of context and supervised training on existing citation patterns.
7 CONCLUSION
In this paper, we have implemented and evaluated four models that
can recommend citations to lawyers drafting legal opinions. BiL-
STM and pretrained RoBERTa perform comparably and outperform
the collaborative filtering and bag-of-words baselines. Our ablation
experiments show that (a) adding metadata about case year, issue,
and judge only leads to insignificant performance improvements for
the neural models, and (b) predicting citations further away from
the context is more difficult, which can be compensated to some
degree by providing more context. Training for extended periods
continuously improves up to a recall@5 of 83.2%. As such, we have
shown that context-based citation recommendation systems can
be implemented as classifiers for a largely normalized citation vo-
cabulary with acceptable performance. Further, our error analysis
shows that even incorrect predictions may still be useful.
Our work also points to the next steps for legal citation predic-
tion. First, citation prediction can be conceived of more broadly
as language generation. Research should hence explore whether

87
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zihan Huang, Charles Low, Mengqiu Teng, Hongyi Zhang, Daniel E. Ho, Mark S. Krass, and Matthias Grabmair
neural models can go beyond pointing to an entry in the citation
vocabulary and write valid citation strings appropriate for a given
context, possibly as part of a continuation of the text. Second, as a
practical matter, it will be important to evaluate the usefulness of
the models trained here with expert users. Finally, we note that legal
sources and institutions form dynamic systems. Constant adapta-
tion, such as detecting and accounting for changes in precedent,
will be key to the future utility of citation systems.
These future directions could rapidly improve legal citation, and
our results here show that context-aware citation prediction can
play a significant role in improving the accuracy, consistency, and
speed of mass adjudication.
8 STATEMENT OF CONTRIBUTIONS
The project was conceived and planned by all authors. ZH, CL, MT,
and HZ conducted all model development and experimental work
under the mentorship of DEH, MK, and MG. MK and MG developed
the citation preprocessing functionality, as well as produced the
error analysis. All authors contributed to writing the paper.
9 ACKNOWLEDGMENTS
The authors thank CMU MCDS students Dahua Gan, Jiayuan Xu,
and Lucen Zhao for creating the issue typology, Anne McDonough
for supporting contributions around citation normalization, and
Dave Ames, Eric Nyberg, Mansheej Paul, and RegLab meeting par-
ticipants for helpful feedback.
REFERENCES
[1] Giambattista Amati. 2009. BM25. Springer US, Boston, MA, 257–260.
[2] David Ames, Cassandra Handan-Nader, Daniel E. Ho, and David Marcus. 2020.
Due Process and Mass Adjudication: Crisis and Reform. Stanford Law Review 72
(2020), 1–78.
[3] Shannon Bradshaw. 2004. Reference Directed Indexing: Redeeming Relevance
for Subject Search in Citation Indexes. In Research and Advanced Technology for
Digital Libraries, Vol. 2769. 499–510.
[4] Cornelia Caragea, Adrian Silvescu, Prasenjit Mitra, and C. Lee Giles. 2013. Can’t
See the Forest for the Trees? A Citation Recommendation System. In Proceedings
of the 13th ACM/IEEE-CS Joint Conference on Digital Libraries (JCDL ’13). 111–114.
[5] Caselaw Access Project. 2020. Caselaw Access Project. https://case.law.
[6] CaseText. 2020. The Machine Learning Technology Behind Parallel Search.
https://casetext.com/blog/machine-learning-behind-parallel-search/. Accessed:
2020-12-18.
[7] Columbia Law Review Ass’n, Harvard Law Review Ass’n, and Yale Law Journal.
2015. The Bluebook: A Uniform System of Citation (21st ed.).
[8] Faraz Dadgostari, Mauricio Guim, P. Beling, Michael A. Livermore, and D. Rock-
more. 2020. Modeling law search as prediction. Artif. Intell. Law 29 (2020),
3–34.
[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proceedings NAACL-HLT ’19. 4171–4186.
[10] Travis Ebesu and Yi Fang. 2017. Neural Citation Network for Context-Aware
Citation Recommendation. In Proceedings SIGIR ’17. 1093–1096.
[11] James Fowler, Timothy Johnson, James Spriggs, Sangick Jeon, and Paul Wahlbeck.
2007. Network Analysis and the Law: Measuring the Legal Importance of Prece-
dents at the U.S. Supreme Court. Political Analysis 15 (06 2007).
[12] Marco Gori and Augusto Pucci. 2006. Research Paper Recommender Systems: A
Random-Walk Based Approach. 2006 IEEE/WIC/ACM International Conference on
Web Intelligence (WI 2006 Main Conference Proceedings) (WI’06) (2006), 778–781.
[13] Qi He, Jian Pei, Daniel Kifer, Prasenjit Mitra, and Lee Giles. 2010. Context-Aware
Citation Recommendation. Proceedings of the 19th International Conference on
World Wide Web (2010), 421–430. https://doi.org/10.1145/1772690.1772734
[14] Daniel E. Ho, Cassandra Handan-Nader, David Ames, and David Marcus. 2019.
Quality Review of Mass Adjudication: A Randomized Natural Experiment at
the Board of Veterans Appeals, 2003–16. The Journal of Law, Economics, and
Organization 35, 2 (03 2019), 239–288. https://doi.org/10.1093/jleo/ewz001
[15] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
88
[16] Wenyi Huang, Zhaohui Wu, Chen Liang, Prasenjit Mitra, and C. Lee Giles. 2015.
A Neural Probabilistic Model for Context Based Citation Recommendation. In
Proceedings AAAI ’15. 2404–2410.
[17] Casetext Inc. 2020. CARA A.I. | Casetext. Retrieved December 17, 2020 from
https://casetext.com/cara-ai
[18] ROSS Intelligence Inc. 2020. ROSS Intelligence. Retrieved December 17, 2020
from https://blog.rossintelligence.com
[19] Glen Jeh and Jennifer Widom. 2002. SimRank: A Measure of Structural-Context
Similarity. In Proceedings KDD ’02. 538–543.
[20] Thorsten Joachims. 2002. Optimizing Search Engines Using Clickthrough Data.
Proceedings KDD ’02 (2002), 133–142.
[21] Marios Koniaris, Ioannis Anagnostopoulos, and Yannis Vassiliou. 2017. Network
analysis in the legal domain: a complex model for European Union legal sources.
Journal of Complex Networks 6, 2 (08 2017), 243–268.
[22] Peng-Hsuan Li, Tsu-Jui Fu, and Wei-Yun Ma. 2020. Why Attention? Analyze
BiLSTM Deficiency and Its Remedies in the Case of NER. In AAAI ’20. 8236–8244.
[23] David Liben-Nowell and Jon Kleinberg. 2007. The Link-Prediction Problem for
Social Networks. J. Am. Soc. Inf. Sci. Technol. 58, 7 (May 2007), 1019–1031.
[24] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen,
Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta:
A robustly optimized bert pretraining approach. CoRR abs/1907.11692 (2019).
http://arxiv.org/abs/1907.11692
[25] Ji Ma, Kuzman Ganchev, and David Weiss. 2018. State-of-the-art Chinese Word
Segmentation with Bi-LSTMs. In Proceedings EMNLP ’18. 4902–4908.
[26] Shutian Ma, Chengzhi Zhang, and Xiaozhong Liu. 2020. A review of citation
recommendation: from textual content to enriched context. Scientometrics 122, 3
(2020), 1445–1472.
[27] Jerry L Mashaw. 1985. Bureaucratic justice: Managing social security disability
claims. Yale University Press.
[28] Sean M. McNee, Istvan Albert, Dan Cosley, Prateep Gopalkrishnan, Shyong K.
Lam, Al Mamunur Rashid, Joseph A. Konstan, and John Riedl. 2002. On the
Recommending of Citations for Research Papers. In Proceedings of the 2002 ACM
Conference on Computer Supported Cooperative Work (CSCW ’02). 116–125.
[29] Gábor Melis, Chris Dyer, and Phil Blunsom. 2017. On the state of the art of
evaluation in neural language models. arXiv preprint arXiv:1707.05589 (2017).
[30] J.C. Oleson. 2003. You Make Me Sic: Confessions of a Sadistic Law Review Editor.
U.C. Davis Law Review 37 (2003).
[31] Larry Page, Sergey Brin, Rajeev Motwani, and Terry Winograd. 1998. The
PageRank Citation Ranking: Bringing Order to the Web. Technical Report, Stanford
University (1998).
[32] Lazar Peric, Stefan Mijic, Dominik Stammbach, and Elliott Ash. 2020. Legal
Language Modeling with Transformers. In Proceedings ASAIL 2020, Vol. 2764.
CEUR-WS.
[33] Anna Ritchie. 2009. Citation context analysis for information retrieval. PhD
thesis, University of Cambridge.
[34] Anna Ritchie, Stephen Robertson, and Simone Teufel. 2008. Comparing Citation
Contexts for Information Retrieval. Proceedings CIKM ’08 (2008), 213–222.
[35] Ali Sadeghian, Laksshman Sundaram, Daisy Zhe Wang, William F. Hamilton,
Karl Branting, and Craig Pfeifer. 2018. Automatic Semantic Edge Labeling over
Legal Citation Graphs. Artif. Intell. Law 26, 2 (2018), 127–144.
[36] Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural Machine
Translation of Rare Words with Subword Units. In Proceedings ACL ’16 (Volume
1: Long Papers). 1715–1725.
[37] Trevor Strohman, W. Bruce Croft, and David Jensen. 2007. Recommending
citations for academic papers. In Proceedings SIGIR ’07. 705–706.
[38] Trevor Strohman, Donald Metzler, Howard Turtle, and W. Bruce Croft. 2005.
Indri: a language-model based search engine for complex queries. Technical Report.
in Proceedings of the International Conference on Intelligent Analysis.
[39] Emma Strubell, Ananya Ganesh, and Andrew McCallum. 2019. Energy and Policy
Considerations for Deep Learning in NLP. In Proceedings ACL ’19. 3645–3650.
[40] Merine Thomas, Thomas Vacek, Xin Shuai, Wenhui Liao, George Sanchez, Paras
Sethia, Don Teo, Kanika Madan, and Tonya Custis. 2020. Quick Check: A Legal
Research Recommendation System. In Proceedings NLLP ’20, Vol. 2645. CEUR-WS.
[41] Radboud Winkels, Alexander Boer, Bart Vredebregt, and Alexander von Someren.
2014. Towards a Legal Recommender System. In Proceedings JURIX ’14.
[42] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,
Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe
Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu,
Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest,
and Alexander Rush. 2020. Transformers: State-of-the-Art Natural Language
Processing. In Proceedings EMNLP ’20: System Demonstrations. 38–45. https:
//doi.org/10.18653/v1/2020.emnlp-demos.6
[43] Paul Zhang and Lavanya Koppaka. 2007. Semantics-Based Legal Citation Network.
In Proceedings ICAIL ’07. 123–130.
[44] Lucia Zheng, Neel Guha, Brandon R. Anderson, Peter Henderson, and Daniel E.
Ho. 2021. When Does Pretraining Help? Assessing Self-Supervised Learning for
Law and the CaseHOLD Dataset. In Proceedings ICAIL ’21. arXiv:2104.08671 (in
press).
 A dynamic model for balancing values
Juliano Maranhão∗ Edelcio G. de Souza
julianomaranhao@usp.br edelcio.souza@usp.br
University of São Paulo Law School University of São Paulo - FFLCH
São Paulo, SP, Brazil São Paulo, SP, Brazil
ABSTRACT
We propose an additive model for balancing the impacts of actions
on values, where factors intensify or attenuate impacts on values,
and values are assigned degrees of relative importance (weights).
The balancing model induces axiological rules, consisting in prohi-
bitions or permissions that are justified according to the impacts of
the prohibited or permitted action on the values at stake. We also
propose eight different revision operators, which shift the balance
– and thus induce different norms – by expanding or contracting
either the set of factors or the set of values. We provide the con-
struction and prove some success properties of those operators.
KEYWORDS
balancing values, change functions, teleological interpretation and
argumentation
ACM Reference Format:
Juliano Maranhão, Edelcio G. de Souza, and Giovanni Sartor. 2021. A dy-
namic model for balancing values. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466143
1 INTRODUCTION
There is an apparent consensus in jurisprudence that legal decision-
making cannot be fully driven by rules alone; it calls for value-based
reasoning. Value-based reasoning is indeed relevant to the law in
multiple regards.
First of all, ethical and political values (the so-called political
morality) provide a critical framework for assessing the merit of
positive laws. Existing laws, as resulting from legislative enactment,
from the practice of legal officers or from custom can be critically
examined with regard to the extent that they meet or violate ideals
of justice and fairness, or that they promote or demote particular
human rights or social values. The divergence of laws from ideas of
justice may justify citizens in disobeying certain laws (as in cases of
civil disobedience) or even officers in refuse the application of such
laws. Legal theorists have provided different interpretations of this
phenomena, depending on views on the relation between law and
morality. Those endorsing a complete separation of law morality
argue that the laws departing from justice still count as perfectly
valid laws [13]; those affirming the intertwining of law and morality
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466143
89
Giovanni Sartor
giovanni.sartor@eui.eu
European University Institute
Firenze, FI, Italy
argue, on the contrary that such laws are legally defective and, in
extreme cases, legally invalid [3].
Second, values play a key role in the context of legal interpre-
tation. In determining the meaning of legal sources, a dominant
role is played by teleological approaches, where the ascription of
a meaning to a legal provision over other possible meanings is
justified on the ground that the selected meaning better promotes
desirable interests or goals, or better prevents undesired outcomes
[5].
Values also play a key role in in constitutional review where
assessments are often performed according to proportionality, i.e.,
by determining whether an infringement of constitutional rights
is justified by non-inferior advantages with regard to other consti-
tutional rights or values, provided that no less infringing choice
delivers a better trade-off [2]. More generally, values play a key
role in all instances of legal decision-making where there is a space
for discretion. In such cases the decision maker has to consider
the merit of alternative choices. This has to be done by taking into
account all legally relevant values (which the decision maker is
allowed to consider), and the extent to which each choice promotes
or demote such values, in the context provided by relevant features
(factors) of the case.
Value-based reasoning has been given an growing attention by
research on AI & Law.
Following the seminal contribution by Berman and Hafner [8],
AI & Law research has provided multiple models of the relation
between cases (and the factors that such cases include or express)
and the values at stake. Bench-Capon and Sartor [6] assign values
to factors and consequently to rules embedding such factors. They
explain precedents according to the applicable rules and the impor-
tance of the (sets of) values promoted by such rules. They compare
alternative sets of rules in terms of their coherence with precedents
and values. Bench Capon et al [19] formalise teleological reason-
ing using logics for defeasible argumentation, extended with the
possibility to express arguments about values, supported by cases.
Grabmair [12] defines functions representing the extent to which a
factor contributes to make it so that a certain outcome promotes a
certain value, and compares alternative outcomes accordingly. Sar-
tor [21] explores the proportional balance of constitutional rights
(as theorized by Alexy [2]), where a legal outcome is compared to
alternative outcomes based on its impact on the promotion and
demotion of values and examines consistency between value-based
decisions of cases, given the factors present in such cases [22].
AI & Law research on statutory interpretation has also explored
the relations between legal rules and the values underlying the
deliberation and moral/political justification of their enactment
based on dynamic approaches where values guide changes of the
content of rules. For instance, Boella et al. [9] introduce values
as coherence parameters guiding the change of conceptual rules.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Such models of statutory interpretation were also explored in the
AGM style [15] and in the framework of i/o logics [10]. In this line
Maranhão [16] proposed an architecture of i/o logics where values
are represented as rules, and constitutive and regulative rules are
the object of different revision functions. Walton, Macagno and
Sartor [23] analyse multiple argument schemes for interpretive
reasoning.
In this paper we propose an additive model of balancing values.
In this model, each relevant factor contributes to intensify or at-
tenuate the impact of the permissibility of the action at stake on a
set of relevant values which are promoted or demoted by the per-
missibility of that action. These influences are then proportionally
considered with respect to the relative importance of the values
being impacted. The resulting assessments of the action’s impacts
on single values are then aggregated to determine the action’s total
impact on the set of values it promotes and on the set of values it
demotes. The comparison of the action’s impacts on the promoted
and on the demoted values enables us to determine whether the
action is axiologically permitted or rather prohibited.
After presenting this model, we propose change functions that
shifts the axiological evaluation and consequently the axiological
rule that applies to it. Such shifts are operated by additions or
subtractions of factors or by additions or subtractions of values in
the model. These operations have some resemblance to argument
moves, where new features of the case or moral considerations are
brought about to oppose previously justified conclusions.
To illustrate our model, we shall use the some variations of the
leading case Riley v. California judged by the U.S. Supreme Court in
2014. In that precedent, the court concluded that a specific warrant
was needed to access the digital content of a mobile phone of an
arrestee, considering the significant amount of personal data usually
stored in such device. In its own words, such access “implicates
substantially greater privacy interests than a brief physical search”
of the items accessible to the arrestee, which is on the contrary
allowed.
The paper is structured as follows. First, in Section 2, we in-
troduce the additive model for balancing the impact of actions on
values. In Section 3, we describe how to build systems of value-rules
or axiological systems. Then, in Section 4, we introduce and dis-
cuss eight revision operators upon axiological systems and specify
their success conditions. We conclude the paper discussing some
limitations of our model and indicating possible paths for future
research.
2 AN ADDITIVE MODEL FOR BALANCING
VALUES
In this Section we shall first introduce the general idea and then
provide a formal account of an action’s impact on relevant values, in
given contexts. Then we show how such impact assessment induces
value-rules and finally discuss some properties of systems of such
rules and their relation with the deontological rules contained in
the positive legal system.
2.1 Introducing the additive model
An axiological model of balancing presupposes a determination
of the comparative moral merits of the choice of performing an
90
Maranhão, de Souza and Sartor
action rather than abstaining from it. The action may consist in
any behaviour, e.g., having an abortion rather than continuing the
pregnancy or accessing a text message stored in a mobile phone,
rather than respecting its confidentiality. The evaluation model
basically compares, for each given action, its impact on the set of
values it promotes against its impact on the set of values it demotes,
given the constellation of factors, i.e., the context in which the
action is performed. Two clarifications are of central importance to
understand the model here proposed.
First, we only consider the assessment of impact of a single action
on values and therefore we only compare the values promoted
against the values demoted by that specific action, so that a decision
takes place whether that action should or should not be performed
on moral grounds. There is no room in this model to compare
and decide among different and logically independent actions in
terms of their impacts on values. Typically, a claim before a court
questions the legality of a particular action and the court must
decide whether that action under evaluation should be performed
or not (should be forbidden or permitted, should be punished or
not be punished). So, we keep the same structure regarding its
axiological evaluation. We acknowledge that there may be contexts
where a judicial decision compares and chooses among alternative
courses of action, for instance, between the consumer’s right to
receive a new product or to have his money back. However we shall
leave this kind of value assessment to future work.
Second, we assume that the direction of impact of an action on a
value – i.e., whether the action promotes or demotes the value – is
invariant, although the extent of the promotion or demotion may
be intensified or attenuated by the presence of factors in the context
of performance. By saying that the direction of impact is invariant,
we mean that irrespective of how many attenuating factors are
taken into account, the impact of an action on the promotion of a
particular value never shifts to its demotion. And vice-versa the
impact of the action at stake on the demotion of a value never shifts
to its promotion.
Let us illustrate the model’s underlying rationality with an ex-
ample. Suppose the rules of a condominium forbid people to take
the elevator during the COVID-19 pandemics. Suppose now that
one inhabitant has a medical emergency. Then one could evaluate
whether following the rule would lead to immoral results. The fac-
tor “medical emergency” is an intensifier w.r.t the promotion of the
value of the patient’s health, which would lead to a permission to
use the elevator. But now consider that the emergency does not
hinder the patient’s ability to walk (for instance, it is a toothache)
and that she lives on the second floor. So the proportional influence
of the set of factors on the promotion of the patient’s health may
become null or negative, but one would not say that the action of
taking the elevator would now demote her health in that particular
context. Actually, the action still promotes health even in presence
of those attenuating factors. But in such cases the proportional
impact of the action is so low that it becomes morally irrelevant to
legal considerations, that is, it will not play a role in a consideration
whether to follow the rule or not. Hence, in the model here pro-
posed attenuating factors only affects the degree of moral impact
of the action on a value.
A dynamic model for balancing values
2.2 The model for balancing
The model of balancing may be described by the structure
V = ⟨𝐴𝑐𝑡, 𝐹𝑎𝑐𝑡, 𝑉 𝑎𝑙, 𝑃𝑟𝑜𝑚, 𝐷𝑒𝑚, {𝐼𝑥 }𝑥 ∈𝐴𝑐𝑡 , 𝑤⟩
whose elements are going to be detailed and discussed below.
We are going to work with two sorts of literals, actions and
factors. The set of action 𝐴𝑐𝑡 is the union 𝑃𝐴𝑐𝑡 ∪ 𝑁 𝐴𝑐𝑡, a set of
atomic actions {𝑥 1, 𝑥 2, . . .} and of their negations {¬𝑥 1, ¬𝑥 2, . . .}.
Similarly, the set of factors 𝐹𝑎𝑐𝑡 is the union 𝑃𝐹𝑎𝑐𝑡 ∪ 𝑁 𝐹𝑎𝑐𝑡 of a set
of atomic factors {𝑓1, 𝑓2, . . .} and of their negations {¬𝑓1, ¬𝑓2, . . .}.
We write 𝑥 to denote the complement (negation) of action 𝑥. The
set 𝑉 𝑎𝑙 = {𝑉1, 𝑉2, ..., 𝑉𝑛 } is a finite set whose elements are values.
Each action may also be subject to the usual deontic qualification.
By combining 𝑥 and the conjunction Φ∧ of all factors in a set Φ
we obtain diadic deontic formulae 𝑂 (𝑥 |Φ∧ ) and 𝑃 (𝑥 |Φ∧ ) stating,
respectively, that action 𝑥 is obligatory or permitted under condition
Φ∧ . We say that the formulas 𝑂 (𝑥 |Φ∧ ) and 𝑃 (𝑥 |Φ∧ ) are deontic
opposites.
In the following, we distinguish two possible deontic evaluations
of an action, under given factors. The first, which is indicated by
the operators 𝑂𝑑 and 𝑃𝑑 is the deontological evaluation, given by a
set of positive norms, which explicitly state obligations and permis-
sions (e.g., the norms stated by a legislator). The second, indicated
by the operators 𝑂 𝑣 and 𝑃 𝑣 , corresponds to the axiological eval-
uation. This evaluation refers to the connection between actions
and values: it considers to what extent the an action promotes or
demotes a value. Under a different reading, that may be more ap-
propriate when engaging in the axiological evaluation of positive
norms, the axiological evaluation considers whether making an
action permissible (in a positive legal code) promotes or rather de-
motes the values at stake. Consider for instance the case of abortion.
When considering the ethical merits of the legal permissibility of
abortion, we are not engaging in the moral merit of a woman’s
choice to have or not to have an abortion, but rather on the moral
merit of making abortion permissible rather than forbidden. This
assessment does not pertain to the morality of individuals, but
rather to political morality, i.e., to the morality of making public
choices, on what is to be imposed or not on citizens.
Thus if 𝑥 is an action, we define 𝑃𝑟𝑜𝑚(𝑥) ⊆ 𝑉 𝑎𝑙 as the set
of values promoted by (the legal permissibility of) action 𝑥 and
𝐷𝑒𝑚(𝑥) ⊆ 𝑉 𝑎𝑙 the set of values demoted by (the permissibility of)
the action 𝑥, where 𝑃𝑟𝑜𝑚(𝑥) ∩ 𝐷𝑒𝑚(𝑥) = ∅.
The comparison depends on the evaluations expressed by the
quantitative assignments of weights to values and of influence of
factors on the impact of actions on values. For generality’s sake
we assume that such indexes can take arbitrary numerical assign-
ments within given ranges. These numbers can be restricted to any
scales that may be convenient for the chosen domain of applica-
tion. Here we shall use the positions (0,.2,.4,.6,.8,1) in the examples.
What matters is that the numerical assignments reflect some rel-
ative importance of the elements at stake, as part of a reasoning
with dimensions and magnitudes, and how such assessment of rel-
ative importance affects the outputs of the systems and its overall
coherence.
Definition 2.1. (weight function) For the finite set of values 𝑉 𝑎𝑙
we define the weight function 𝑤 : 𝑉 𝑎𝑙 −→ [0, 1], where 𝑤 (𝑉𝑖 ) is
the weight of the value 𝑉𝑖 .
91
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Definition 2.2. (impact function) For each action 𝑥 we define
𝐼𝑥 : 𝐹𝑎𝑐𝑡 ∪ {∅} × 𝑉 𝑎𝑙 −→ [−1, 1], where 𝐼𝑥 (𝑓 , 𝑉𝑖 ) is the influence
of factor 𝑓 on the impact of action 𝑥 on the value 𝑉𝑖 .
Note that the function 𝐼𝑥 is fixed for a particular action 𝑥 and
it regards the influence that each factor has on the impact of the
action at stake on each value. The influence may be null, i.e. takes
value 0, when the factor has no influence on the action’s impact
on a particular value, that is, when it is a morally neutral factor
regarding that action and value. When the influence takes a positive
real number in the interval, we say that the factor is an intensifier
of the impact of the action on the value. And if the function assigns
the factor a negative real number in the interval, we say that the
factor is an attenuator of such impact. The action at stake also has
a baseline impact on each value, when the impact function takes as
argument the empty set.
Given these assignments of weight and impact we may define
the proportional impact of a factor on a value for a given action.
Definition 2.3. (Proportional influence of a factor on a value) Let
𝑥 be an action, 𝑓 a factor and 𝑉𝑖 a value in V ⊆ 𝑉 𝑎𝑙, we define:
𝑓
Δ𝑥 (𝑉𝑖 ) = 𝐼𝑥 (𝑓 , 𝑉𝑖 ) × 𝑤 (𝑉𝑖 )
The value assessment will compare the influence of all relevant
factors on all relevant values. On the one hand, those values pro-
moted by the action in a given constellation of factors, and on the
other hand those values promoted by that action in that context.
So we take the sum of the proportional impacts on each set of
promoted or demoted values.
Definition 2.4. (Proportional influence of a factor on a set of
values) Let 𝑥 be an action, 𝑓 a factor and a set V ⊆ 𝑉 𝑎𝑙, then we
define:
Õ
𝑓 ,𝑥
𝐵 𝑃𝑟𝑜𝑚 (V) =
𝑓
Δ𝑥 𝑉𝑖
𝑉𝑖 ∈𝑃𝑟𝑜𝑚 (𝑥)∩V
The same definition holds mutatis mutandis for the proportional
𝑓 ,𝑥
demotion of values, denoted by 𝐵𝐷𝑒𝑚 (V) . When V = 𝑉 𝑎𝑙 we write
𝑓 ,𝑥
simply 𝐵𝑃𝑟𝑜𝑚/𝐷𝑒𝑚 .
The proportional influence of a factor on the promotion (demo-
tion) of values may now be straightforwardly extended to a set of
factors:
Definition 2.5. (proportional influence of a set of factors a set of
values) Let Φ = {𝑓1, 𝑓2, ..., 𝑓𝑛 } be a set of factors, then:
Õ 𝑥,𝑓
𝐵𝑥,Φ
𝑃𝑟𝑜𝑚 = 𝐵𝑃𝑟𝑜𝑚
𝑓 ∈Φ
Again, the same definition holds mutatis mutandis for the propor-
tional influence of a set of factors on the demotion of value, denoted
by 𝐵𝑥,Φ
𝐷𝑒𝑚 .
2.3 Value-rules induced by balancing
We define now an entailment-like relation to extract from a given
value assessment whether the action under analysis would be axi-
ologically permitted, forbidden or indifferent. The action at stake
is forbidden if the proportional aggregate demotion of values is
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
positive and higher than the proportional aggregate promotion of
values. Otherwise, if the number corresponding to the aggregate
promotion of values is higher than the aggregate demotion, the
action is permitted. If the proportional aggregate promotion and
demotion are equivalent or if the aggregate promotion or demo-
tion is a non positive real number, then the action is not morally
relevant, and therefore also permitted.
Definition 2.6. Let 𝑥 ∈ 𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡, and Φ∧ be the conjunction
of all factors in Φ. Then the value-rule induced by Φ and 𝑥 is
(i) 𝑂 𝑣 (𝑥 |Φ∧ ), if 𝐵𝑥,Φ
𝐷𝑒𝑚 > 𝐵𝑃𝑟𝑜𝑚 and 𝐵 𝐷𝑒𝑚 > 0,
𝑥,Φ 𝑥,Φ
(ii) 𝑃 𝑣 (𝑥 |Φ∧ ), otherwise
We say that the balance shifts whenever the addition of a new
set of factors to the original set of factors changes the value rule
induced by the balancing. That is, if we have 𝑂 𝑣 (𝑥 |Φ∧ ) then the
addition of the set of factors Θ shifts the balance if it holds that
𝑃𝑟𝑜𝑚 ≥ 𝐵 𝐷𝑒𝑚 or 𝐵 𝐷𝑒𝑚 < 0. Likewise, if it holds that 𝑃 𝑣 (𝑥 |Φ ),
𝐵𝑥,Φ∪Θ 𝑥,Φ∪Θ 𝑥,Φ∪Θ ∧
then the balance shifts if we obtain 0 < 𝐵𝑥,Φ∪Θ
𝐷𝑒𝑚 > 𝐵𝑃𝑟𝑜𝑚 .
𝑥,Φ∪Θ
The dyadic deontic statements obtained by the balancing may
be taken as elements of a theory in a dyadic deontic logic, where
further deontic statements may be derived. If Φ = ∅, we express
the baseline evaluation of an action 𝑥 as 𝑂/𝑃 𝑣 (𝑥 |⊤), which we may
also be written as a monadic obligation/permission 𝑂/𝑃 𝑣 𝑥.
Definition 2.7. Let 𝑥 be an action, Φ a set o factors and V ⊆ 𝑉 𝑎𝑙.
Then VR𝑥Θ (Φ)is the set of value-rules induced by the balancing
model for Θ ⊆ Φ and VR𝑥 (Φ) is the class of sets of axiological rules
VR𝑥Θ (Φ) such that Θ𝑖 ⊆ Φ.
𝑖
Let us illustrate the model by an example. We adopt here a
convention to omit the reference to the argument of an impact
function when its value is zero.
According to the US case law before the judgment of Riley vs
California (2014) a police officer was allowed to access the content
of any items, including property, in the premises or surroundings
when arresting an individual due to any criminal offense. However,
in the Riley vs California case, the arrest included the seizure of a
mobile phone and, from the content accessed in that device, the
officers found evidence of another crime, which led to a conviction.
The U.S. Supreme Court concluded that a specific warrant was
needed to access the digital content of the mobile phone of the
arrestee. It considered that the significant amount of personal data,
usually stored in a mobile phone, would involve an inadmissible
impact on privacy. This rule could be explained by the following
considerations on the underlying value impacts: accessing items
(𝑎𝑐𝑐) in an arrest has a baseline impact on the promotion of public
safety (𝑆𝑎𝑓 ) and a baseline demotion of property rights (𝑃𝑟𝑖𝑔ℎ𝑡) and
privacy (𝑃𝑟𝑖𝑣); the factor “arrest” (𝑎𝑟𝑟 ) intensifies the promotion
of public safety so as to outweigh the extent to which the factor
“property” (𝑝𝑟𝑜𝑝) intensifies the demotion (through the same action)
of property rights and privacy respectively. However, as considered
by the court, if the item collected is a mobile phone (𝑚𝑜𝑏), then
the negative impact on privacy is intensified to the extent that the
promotion of public safety is outweighed. This led the court to
introduce an exception, forbidding access to the digital content
of mobile phones collected during an arrest without an specific
warrant.
92
Maranhão, de Souza and Sartor
Example 2.8 (Riley vs California). Consider the action 𝑎𝑐𝑐, the set
of factors Φ = {𝑎𝑟𝑟, 𝑝𝑟𝑜𝑝, 𝑚𝑜𝑏} and values V = {𝑆𝑎𝑓 , 𝑃𝑟𝑖𝑔ℎ𝑡, 𝑃𝑟𝑖𝑣 },
where 𝑃𝑟𝑜𝑚(V) = {𝑆𝑎𝑓 } and 𝐷𝑒𝑚(V) = {𝑃𝑟𝑖𝑔ℎ𝑡, 𝑃𝑟𝑖𝑣 }, with the
following weights 𝑤 (𝑃𝑟𝑖𝑣) = .6, 𝑤 (𝑃𝑟𝑖𝑔ℎ𝑡) = .4 and 𝑤 (𝑆𝑎𝑓 ) = .6.
Consider also that accessing items found with an individual has
only a baseline impact on privacy, as it would not impact property
if there are no property items neither promote safety if there is
no evidence that the individual is related to any criminal offence
(zero baseline impact). So, we have 𝐼𝑎𝑐𝑐 (∅, 𝑃𝑟𝑖𝑣) = .4. Now consider
the influence of the factors specified above on the impact of the
action on the relevant values. A considerable impact on property
right by accessing property 𝐼𝑎𝑐𝑐 (𝑝𝑟𝑜𝑝, 𝑃𝑟𝑖𝑔ℎ𝑡) = .4, a significant
promotion of safety if there is sufficient evidence of criminal offence
for an arrest 𝐼𝑎𝑐𝑐 (𝑎𝑟𝑟, 𝑆𝑎𝑓 ) = .8 and, as considered, by the court, an
extreme impact on privacy by accessing personal data stored in a
mobile phone 𝐼𝑎𝑐𝑐 (𝑚𝑜𝑏, 𝑃𝑟𝑖𝑣) = 1.
In the example above we have the following balance in the base-
line context, 𝐵𝑎𝑐𝑐,∅
𝑃𝑟𝑜𝑚 = 0 against 𝐵 𝐷𝑒𝑚 = .24, thus morally prohibit-
𝑎𝑐𝑐,∅
ing police officers to access items in the possession of any indi-
vidual (VR𝑎𝑐𝑐 = {𝑂¬𝑎𝑐𝑐}). In the context of an arrest, we would
∅
𝑎𝑐𝑐,{𝑎𝑟𝑟 } 𝑎𝑐𝑐,{𝑎𝑟𝑟 }
have 𝐵𝑃𝑟𝑜𝑚 = .48 against 𝐵𝐷𝑒𝑚 = .24, thus rendering the
access morally permitted (VR𝑎𝑐𝑐{𝑎𝑟𝑟 }
= {𝑃 (𝑎𝑐𝑐 |𝑎𝑟𝑟 )}. If the seizure
during the arrest includes property items the access would still be
𝑎𝑐𝑐,{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝 }
morally justified, given that we would have 𝐵𝑃𝑟𝑜𝑚 = .48
𝑎𝑐𝑐,{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝 }
against 𝐵𝐷𝑒𝑚 = .40, thus reflecting the previous US case
law (that is, VR𝑎𝑐𝑐
{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝 }
= {𝑃 𝑣 (𝑎𝑐𝑐 |𝑎𝑟𝑟 ∧ 𝑝𝑟𝑜𝑝)}). Finally, con-
sidering the additional factor that the property item is a mobile
phone, the finding of the court in Riley vs California would be
𝑎𝑐𝑐,{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝,𝑚𝑜𝑏 }
morally justified by the model, given that 𝐵𝑃𝑟𝑜𝑚 = .48
𝑎𝑐𝑐,{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝,𝑚𝑜𝑏 }
against 𝐵𝐷𝑒𝑚 = 1 (thus leading to VR𝑎𝑐𝑐
{𝑎𝑟𝑟,𝑝𝑟𝑜𝑝,𝑚𝑜𝑏 }
=
𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑎𝑟𝑟 ∧ 𝑝𝑟𝑜𝑝 ∧ 𝑚𝑜𝑏).
2.4 Consistency, coherence and the Radbruch’s
formula
Based on the aggregate impacts on the demotion and promotion of
the values, which are triggered by an action, we define the propor-
tional impact of an induced value-rule as the difference between
the values promoted and demoted in the assessment.
Definition 2.9. (Proportional impact of a rule) Consider an action
𝑥, a set of factors Φ and a set of values V.
Then, 𝜎 (𝑥, Φ, V) = 𝐵𝑥,Φ
𝑃𝑟𝑜𝑚 (V)
− 𝐵𝑥,Φ
𝐷𝑒𝑚 (V)
is the proportional impact
on values V of the axiological rules 𝑃 𝑣 (𝑥 |Φ∧ )/𝑂 𝑣 (𝑥 |Φ∧ ).
Given the definition above, the polarity (positive or negative) of
the proportional impact indicates the modality of the value-rule
induced by the balance model. If 𝜎 (𝑥, Φ, V) > 0, then 𝑃 𝑣 (𝑥 |Φ). If
𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥𝐷 𝑒𝑚 > 0 then 𝑂 𝑣 (𝑥 |Φ)
Considering that we may obtain value-rules from the balancing
model, we may now evaluate the content of positive rules expressed
in a dyadic deontic language with the operators 𝑂𝑑 (𝑥 |𝑎) for pos-
itive obligation to do 𝑥 in context 𝑎 and the operator 𝑃𝑑 (𝑥 |𝑎) for
permission of action 𝑥 in context 𝑎.
A dynamic model for balancing values
We may assume a dyadic deontic logic consequence relation ⊢
satisfying inclusion and factual detachment and to derive deontic
sentences from two different sets of rules.
Definition 2.10. (consistency) Let 𝑥 ∈ 𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡, and Φ∧ be
the conjunction of all factors in Φ. Then a set of rules R is consistent
iff
• R ⊬ 𝑂𝑑 (𝑥 |Φ∧ ) ∧ 𝑂𝑑 (𝑥 |Φ∧ ) and
• R ⊬ 𝑂𝑑 (𝑥 |Φ∧ ) ∧ 𝑃𝑑 (𝑥 |Φ∧ ).
Thus, the postulate of consistency is satisfied iff it is not the case
that an action 𝑥 is both forbidden and obligatory or both forbidden
and permitted by the normative system. The set of rules may be a
set of value-rules induced by the balancing model VR or a set of
positively enacted rules represented as dyadic deontic sentences
LR.
Definition 2.11. (coherence) Let 𝑥 ∈ 𝐴𝑐𝑡 and Φ ⊆ 𝐹𝑎𝑐𝑡. Then
(LR, VR) is coherent iff,
• LR and VR are both consistent and
• it is not the case that
– LR ⊢ 𝑂𝑑 (𝑥 |Φ∧ ) and VR ⊢ 𝑂 𝑣 (𝑥 |Φ∧ ) or
– LR ⊢ 𝑃𝑑 (𝑥 |Φ∧ ) and VR ⊢ 𝑂 𝑣 (𝑥 |Φ∧ ) or
– LR ⊢ 𝑂𝑑 (𝑥 |Φ∧ ) and VR ⊢ 𝑃 𝑣 (𝑥 |Φ∧ )
An interesting application of this model to legal theory consists
in the interpretation of a formula about the relation between law
and morality originally proposed by Gustav Radbruch [20] to deter-
mine the (in)validity of Nazi’s laws: laws enacted by proper authority
and power are legally valid unless they reach an unbearable degree
of immorality or injustice.
In our model, we may define an inference relation upon the basic
dyadic deontic logic, considering the proportional impact of the
value-rule corresponding to the legal rule. If the positively enacted
legal rule 𝑂𝑑 (𝑥 |Φ∧ ) /𝑃𝑑 (𝑥 |Φ∧ ) is inconsistent with the value-rule
induced by the value assessment 𝑃 𝑣 (𝑥 |Φ∧ ) /𝑂 𝑣 (𝑥 |Φ∧ ) and the value-
rule has a proportional impact that exceeds a given threshold, then
we say that the positively enacted rule is unbearably unjustified or
immoral and therefore is invalid, according to the Radbruch’s legal
theory. The inference relation consists in a restriction of the set of
positively enacted rules available to derive a normative solution
for a given action. There may be different interpretations of the
Radbruch’s formula leading to different associated legal theories,
for instance, a generative formula that not only censors positively
enacted rules but also generate valid legal content based on moral
considerations (see [17]).
Definition 2.12. (Radbruch’s Formula) Let 𝑥 be an action, Φ a set
of factors and V a valuation and 𝑟 a threshold index. Then:
• ⊢𝑟𝑎𝑑 𝑃𝑑 (𝑥 |Φ∧ ) iff
– LR ⊢ 𝑃𝑑 (𝑥 |Φ∧ ) and
– it is not the case that VR ⊢ 𝑂 𝑣 (𝑥 |Φ∧ ) and
– |𝜎 (𝑥, Φ, V)| ⩾ 𝑟 1 ;
• ⊢𝑟𝑎𝑑 𝑂𝑑 (𝑥 |Φ∧ ) iff
– LR ⊢ 𝑂𝑑 (𝑥 |Φ∧ ) and
– it is not the case that VR ⊢ 𝑃 𝑣 (𝑥 |Φ∧ ) and
– |𝜎 (𝑥, Φ, V)| ⩾ 𝑟
1 We consider |n| the absolute value of the (positive or negative) integer n.
93
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
In the Example 2.8, suppose we settle the Radbruch’s threshold at
.2. Then only the rule permitting police officers to access property
items would be legally enforceable based on the evaluation, while
all other rules indicated in the example – the permission to access
items in the possession of any individual, the prohibition to access
items with the arrestee and the permission to access the digital
content of the arrestee’s mobile phone in an arrest – would be
unbearably unjust and therefore not legally enforceable, according
to Radbruch’s theory.
To illustrate that in the model, consider a set of positively en-
acted rules LR = {𝑃𝑑 (𝑎𝑐𝑐 |⊤), 𝑃𝑑 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝), 𝑂𝑑 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ),
𝑃𝑑 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ∧ 𝑚𝑜𝑏)}. Now let us compare it with the rules
extracted in the evaluation model, indicating the proportional im-
pact of each in parenthesis:
AS𝑎𝑐𝑐 = {𝑂 𝑣 (¬𝑎𝑐𝑐) (−.24), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝) (−.4), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧
𝑎𝑟𝑟 ) (.08), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ∧ 𝑚𝑜𝑏) (−.52)}.
Hence, according to the value assessment, all positively enacted
rules are morally unjustified. Nevertheless, the only the rule which
would still be valid according to a threshold of .2 would be the pro-
hibition to search property items during an arrest. Following Alexy,
we would say that such prohibition would be morally defective
but still legally valid, while all others are both morally and legally
defective ([3], [4]).
3 SYSTEMS OF AXIOLOGICAL RULES
In Definition 2.7 we have specified singular sets of deontic sentences
VR𝑥Θ for each subset Θ of the set of relevant factors Φ. Clearly,
the union of these singular sets would lead to an inconsistency
if the underlying dyadic deontic logic is monotonic and the prin-
ciple of deontic non-contradiction holds. For instance, VR𝑎𝑐𝑐 ∪
∅
VR𝑎𝑐𝑐 = {𝑂 𝑣 (¬𝑎𝑐𝑐 |⊤), 𝑃 𝑣 𝑥 (𝑎𝑐𝑐 |𝑎𝑟𝑟 )}, from which one may infer
{𝑎𝑟𝑟 }
both 𝑂 (¬𝑎𝑐𝑐 |𝑎𝑟𝑟 ) and 𝑃 (𝑎𝑐𝑐 |𝑎𝑟𝑟 ).
Therefore we are going to define a construction of a system of
value-rules, which we shall call axiological system AS𝑥 , in which
all steps of factor additions are consistently compiled.
Note that each step of adding a factor to Θ𝑖 may or may not
shift the balance of values to the effect of changing the normative
solution to the action at stake in that specific context.
Thus there are two different situations to be covered. In the first
one, the addition of a new factor shifts the normative solution to the
opposite one. For instance, police officers are generally forbidden
to access items held by an individual. But in the context of an arrest,
the normative solution shifts to a permission to perform that action.
Were it not the case, that is, if the factor “arrest” were absent, then
the former solution would prevail. In its turn, if now we consider
that the item accessed is a property item, it is still permitted to
perform it if there is an arrest.
So, gathering the set of deontic sentences may reveal this “step
by step” feature of an argumentation process where new factors are
called into question or the presence of a factor or its relevance or
intensity regarding the values at stake is contested. We are going to
represent such argument moves by change functions of expansion
and contraction that impact the resulting normative solution.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
In order to do that, first we are going to define, by induction, an
axiological system for an action 𝑥 and a set of factors Φ. Notice that
we have to consider two cases for the inductive step. The first case
covers the hypothesis where there is no shifts in the balance in the
course of addition of new factors and accordingly new value-rules
to the axiological system. The second case covers the hypothesis
where there was at least one shift of balance (and the corresponding
induced value-rule) at some previous step of the construction of the
axiological system. For simplicity we write the definition only for
those cases where the antecedent step starts with an obligation, but
it is easy to adapt it, mutatis mutandis, for those cases where the
previous step in the construction of the axiological system delivers
a permission for the corresponding constellation of factors.
Definition 3.1. Let 𝑥 be an action, 𝐹𝑎𝑐𝑡 = {𝑓1, 𝑓2, ..., 𝑓𝑚 } and
Φ ⊆ 𝐹𝑎𝑐𝑡. We are going to define a set of axiological rules based
on an induction of an increasing sequence of subsets Φ0 = ∅, Φ1 =
{𝑓1 }, Φ2 = {𝑓1, 𝑓2 }, ..., Φ𝑛 = Φ. Then the Axiological Normative
System for 𝑥, AS𝑥 is inductively defined as follows:
• Basic step (Φ = ∅)
– AS𝑥1 = {𝑃 𝑣 (𝑥 |𝑓1 ), 𝑂 𝑣 (𝑥 |¬𝑓1 )}, if AS𝑥0 = 𝑂 𝑣 (𝑥 |⊤) and the
balance shifts
– AS𝑥1 = {𝑂 𝑣 (𝑥 |𝑓1 )}, otherwise.
• Inductive step with one element in AS𝑛𝑥 ; Φ = {𝑓1, 𝑓2, ..., 𝑓𝑛 }
– AS𝑛+1𝑥 = {𝑃 𝑣 (𝑥 |Φ∪{𝑓𝑛+1 }∧ ), 𝑂 𝑣 (𝑥 |Φ∪{¬𝑓𝑛+1 }∧ ), if AS𝑛𝑥 =
𝑂 𝑣 (𝑥 |Φ∧ ) and the balance shifts
– AS𝑛+1𝑥 = {𝑂 𝑣 (𝑥 |Φ ∪ {𝑓𝑛+1 })}∧ ), otherwise.
• Inductive step with two elements in AS𝑛𝑥 ; Φ = {𝑓1, 𝑓2, ..., 𝑓𝑛 }
– AS𝑛+1𝑥 = {𝑃 𝑣 (𝑥 |𝑓1 ∧ ... ∧ ¬𝑓𝑛 ∧ 𝑓𝑛+1 ), 𝑂 𝑣 (𝑥 |𝑓1 ∧ ... ∧ 𝑓𝑛 ∧
¬𝑓𝑛+1 )}, if AS𝑛𝑥 = {𝑂 𝑣 (𝑥 |𝑓1 ∧ ... ∧ 𝑓𝑛 ), 𝑃 (𝑥 |𝑓1 ∧ ... ∧ ¬𝑓𝑛 )}
and the balance shifts
– AS𝑛+1𝑥 = {𝑂 𝑣 (𝑥 |𝑓1 ∧ ... ∧ 𝑓𝑛 ∧ 𝑓𝑛+1 ), 𝑃 𝑣 (𝑥 |𝑓1 ∧ ... ∧ ¬𝑓𝑛 ∧
𝑓𝑛+1 )}, otherwise.
Ð
Then, AS𝑥 = AS𝑖𝑥 , for 0 ≤ 𝑖 ≤ 𝑛
Notice that the choice of a particular set of factors, i.e., the selec-
tion of a set Φ1 ⊆ 𝐹𝑎𝑐𝑡 rather than a different set Φ2 ⊆ 𝐹𝑎𝑐𝑡 may
deliver a different axiological system, since the assessment of the
action’s impact on the values is determined by the factors being
considered in each set. Moreover, notice that the particular order
in which the new factors are introduced in a given set Φ ⊆ 𝐹𝑎𝑐𝑡
determines a specific path leading to the axiological system based
on all factors in Φ ⊆ 𝐹𝑎𝑐𝑡.
Indeed, the sequence AS𝑥1 , . . . 𝐴𝑆 𝑛𝑥 reflects a strategy of argu-
mentation adopted by the parties, as far as this strategy consists
in the introduction of new factors (the removal of factors will be
considered in Section 3). For instance, that the items were collected
in an arrest is an argument for the justification of a permission to
the police to access the content of those items. In its turn arguing
that an item collected is a mobile phone favours the axiological
prohibition to access its digital content.
In the precedent discussed in example 2.8, the sequence ⟨∅,
{𝑝𝑟𝑜𝑝}, {𝑝𝑟𝑜𝑝, 𝑎𝑟𝑟 }, {𝑝𝑟𝑜𝑝, 𝑎𝑟𝑟, 𝑚𝑜𝑏}⟩ would result in AS𝑎𝑐𝑐 =
{𝑂 𝑣 (¬𝑎𝑐𝑐), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝∧𝑎𝑟𝑟 ), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝∧¬𝑎𝑟𝑟 ),
𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧𝑎𝑟𝑟 ∧𝑚𝑜𝑏), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧𝑎𝑟𝑟 ∧¬𝑚𝑜𝑏)}. In its turn,
the sequence ⟨∅, {𝑚𝑜𝑏}, {𝑚𝑜𝑏, 𝑝𝑟𝑜𝑝}, {𝑚𝑜𝑏, 𝑝𝑟𝑜𝑝, 𝑎𝑟𝑟 }⟩ would re-
sult in A𝑆 𝑎𝑐𝑐 = {𝑂 𝑣 (¬𝑎𝑐𝑐), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑚𝑜𝑏), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑚𝑜𝑏 ∧ 𝑝𝑟𝑜𝑝),
94
Maranhão, de Souza and Sartor
𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑚𝑜𝑏 ∧𝑝𝑟𝑜𝑝 ∧𝑎𝑟𝑟 )}. Nevertheless, each axiological system
resulting from the balancing is consistent.
Theorem 3.2. For every action 𝑥 and Φ ⊆ 𝐹𝑎𝑐𝑡, AS𝑥 is consistent.
Proof. (sketch) By construction, we notice that the basic step
𝐴𝑆 0𝑥 contains a single axiological rule and also that each step 𝐴𝑆𝑛+1
𝑥
preserves consistency of the extracted rules, provided that the con-
ditions of opposed modalities 𝑂¬𝑥 and 𝑃𝑥 are, by construction,
mutually exclusive. □
Although different consistent axiological systems AS𝑥 may be
built upon the same model V, according to the order of sets of
factors used in the construction, all systems will deliver the same
result in the presence of all relevant factors at hand. This result
may happen if all the steps in the construction have one element,
or if there is any step where two elements are included in the
axiological system. The end result may be thought of as the result
of an argumentation process, were in each step a new factor is
brought about that may or may not change the solution within the
balance of values. This property of the axiological systems may be
shown as a corollary of the following general theorem:
Theorem 3.3. Let Φ0 = ∅, Φ1 = {𝑓1 }, Φ2 = {𝑓1, 𝑓2 }, ..., Φ𝑛 = Φ be
an increasing sequence of subsets of Φ ⊆ 𝐹𝑎𝑐𝑡. Then for every natural
number 0 ⩽ 𝑖 ⩽ 𝑛, it holds that:
• if 0 < 𝐵𝑥,Φ
𝐷𝑒𝑚 (V)
> 𝐵𝑥,Φ
𝑃𝑟𝑜𝑚 (V)
then AS𝑥 ⊢ 𝑂 (𝑥 |Φ𝑖∧ )
• otherwise, AS𝑥 ⊢ 𝑃 (𝑥 |Φ𝑖∧ )
Proof. For 𝑛 = 0, the result is immediate. For the inductive
step, suppose that we have AS𝑥 ⊢ 𝑂 (𝑥 |Φ𝑘∧ ) for Φ𝑘 = {𝑓1, 𝑓2, ..., 𝑓𝑘 }
and therefore AS𝑥 ⊢ 𝑂 (𝑥 |Φ𝑘∧ ). Then for Φ𝑘 ∪ {𝑓𝑘+1 }, it holds,
by construction that if 0 < 𝐵𝑥,Φ > 𝐵𝑥,Φ , then AS𝑥 ⊢
𝐷𝑒𝑚 (V) 𝑃𝑟𝑜𝑚 (V)
𝑂 (𝑥 |Φ𝑖∧ ∧ 𝑓𝑘+1 ). Otherwise, it holds that AS𝑥 ⊢ 𝑃 (𝑥 |Φ𝑖∧ ∧ 𝑓𝑘+1 )
showing that new additions of factors in the axiological rules never
change the previous constellation of literals representing factors or
the previous order of increasing subsets of factors. The case where
AS𝑥 ⊢ 𝑃 (𝑥 |Φ𝑘∧ ) follows the same steps. □
As a corollary from theorem 3.3 it holds that the final evaluation
and corresponding induced value-rule does not depend on the order
of subsets of relevant factors used in the construction of the axio-
logical system, since it is obtained by summing all the differences
that each factor makes autonomously.
4 SHIFTING THE BALANCE
The ascription of weights to values and influence on the action’s
impact on values are the key aspects of the balancing model V. The
set of relevant factors Φ ⊆ 𝐹𝑎𝑐𝑡 and the set of relevant values V ⊆
𝑉 𝑎𝑙 are the building blocks that determine whether the outcome of
an evaluation is either an axiological prohibition or an axiological
permission.
Hence, provided an impact function and a weight assignment,
adding or excluding factors and adding or excluding values from
the respective sets considered in the evaluation may shift the bal-
ance and therefore change the induced value-rule. As mentioned
above, such moves may be though as the advancement of reasons or
A dynamic model for balancing values
arguments for the moral justification or disapproval of a permission
or a prohibition to the action at stake.
4.1 Possible shifts in the value assessment of
an action
To represent such reason-giving or argument dynamics we shall
introduce different change functions that either modify the set
of relevant factors or modify the set of values considered in the
assessment. In this paper we are not going to study combinations
of changes of values with changes in the set of factors.
We shall focus on those modifications that shift the balance
between the value-impacts of the action being considered, so that, in
virtue of the modification a prohibited action becomes permissible,
or a permissible action becomes prohibited:
• Before the change the action is axiologically prohibited since
its impacts on the values it demotes prevail over its impacts
on the values it promotes: 0 < 𝐵𝑥,Φ > 𝐵𝑥,Φ ); after
𝐷𝑒𝑚 (V) 𝑃𝑟𝑜𝑚 (V)
the change the action is axiologically permitted since its
impacts on the values it promotes prevail over its impacts
on the values it demotes (0 < 𝐵𝑥,Φ ≥ 𝐵𝑥,Φ ) of the balance between values demoted and promoted by the action.
𝑃𝑟𝑜𝑚 (V) 𝐷𝑒𝑚 (V)
• Before the change the action is axiologically permitted since
its impacts on the values it demotes prevail over its impacts
on the values it promotes (0 < 𝐵𝑥,Φ ≥ 𝐵𝑥,Φ ); after
𝑃𝑟𝑜𝑚 (V) 𝐷𝑒𝑚 (V)
the change the action is axiologically prohibited since its
impacts on the values it demotes prevail over its impacts on
the values it promotes 0 < 𝐵𝑥,Φ
𝐷𝑒𝑚 (V)
> 𝐵𝑥,Φ
𝑃𝑟𝑜𝑚 (V)
).
For a given evaluation 0 < 𝐵𝑥,Φ 𝐷𝑒𝑚 (V)
> 𝐵𝑥,Φ
𝑃𝑟𝑜𝑚 (V)
from which
we induce 𝑂 𝑣 (𝑥/Φ), there are different ways to shift the balance.
First, it is possible to add new factors to Φ that together are going
to attenuate the impact of the action on the demotion of the values
considered. If this attenuation effect shifts the balance either by
making the degree of the demotion negative, or by making it equal
or lower to the degree of the values promoted, we call this move a
factor-undercutting expansion of the relevant factors. Another move
to the same effect consists in adding a set of factors that together
intensify the promotion of the relevant values. We call this move a
factor-rebutting expansion.
Besides adding factors, one may also contest that a given factor
considered is present in the context, what could be reduced to the
exclusion of factors from the set of relevant factors Φ. This might
produce a shift of balance in two different ways. First it could be an
exclusion of a set of factors that together intensify the demotion
of the values by the action at stake. We call this move a factor-
undercutting contraction. On the other hand the same effect could
be obtained if one deletes a set of factors that together attenuates
the action’s promotion of the relevant values. We call this move a
factor-rebutting contraction.
Similar moves may be described that modify the set of val-
ues V which are considered relevant to the evaluation. A value-
undercutting expansion is an addition of values to V, which are
demoted by the action through attenuating factors, given the set of
factors Φ and both the impact and the weight function. In its turn,
a value-rebutting expansion is an addition of values to V, which
95
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
are promoted by the action through intensifying factors. The ex-
clusion of elements of V may also produce a shift in the previous
balance, changing the normative solution. For instance, a value-
undercutting contraction corresponds to the exclusion of values
which are demoted by the action through intensifying factors. In
its turn, a value-rebutting contraction corresponds to the exclusion
of values which are promoted by the action through attenuating
factors.
4.2 Factor expansion and contraction operators
We propose here eight change operators to capture the shifts of
balance described above. Four of such operators act upon factors,
and four upon values. In this subsection we consider the operators
on factor, two of which enlarge the set of available factors (factor ex-
pansion operators) and two restrict such a set (factors contractions
operators).
Definition 4.1. (factor expansion function) We define an expansion
function 𝑒 : P (𝐹𝑎𝑐𝑡) −→ P (𝐹𝑎𝑐𝑡) − {∅} such that for all Φ ⊂ 𝐹𝑎𝑐𝑡,
Φ ∩ 𝑒 (Φ) = ∅.
The first two operators add new attenuators the “winning” side
This expansion reduces the extent to which the action demotes or
promotes the values at stake, to an extent that is sufficient to invert
the original balance.
This can happen in two cases. In the first case, before the revision,
the action’s impact on the demoted values was greater than its
impact over the promoted values. The additional factors reduce the
demotion of the demoted values, to such an extent that the action’s
impact on the demoted value becomes smaller than its impact on
the promoted values. This means that in the context of the extended
set of factors the former prohibited action becomes permissible.
In the second case, before the revision, the action’s impact on
the promoted values was greater than its impact over the demoted
values. The additional factors reduce the promotion of the promoted
values to such an extent that the action’s impact on the promoted
values becomes smaller than its impact on the demoted values. This
means that in the context of the extended set of factors the former
permissible action becomes prohibited.
Definition 4.2. (factor undercutting expansion function) Let 𝑥 ∈
𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. An expansion function 𝑒 is an under-
cutting expansion function, denoted by 𝑢𝑒 (Φ), iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ >0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑒 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
𝑥,Φ∪𝑒 (Φ)
(ii) 𝜎 (𝑥, Φ ∪ 𝑒 (Φ), V) ⩾ 0 or 𝐵𝐷𝑒𝑚 (V) ⩽ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0
(i) for all 𝑓𝑖 ∈ 𝑒 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ ∪ 𝑒 (Φ), V) < 0
On this basis we define two factor contraction operators. The
first
Definition 4.3. (factor undercutting expansion operator) Let 𝑢𝑒 be
an undercutting expansion operation. Then we define an operator
+ on sets of factors such that Φ+ = Φ ∪ 𝑢𝑒 (Φ).
Now we turn to the rebutting expansion, which consists in
adding new intensifiers to the “losing” side in the balance between
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the action’s impacts on promoted vs demoted values, resulting in a
shift of such a balance.
Definition 4.4. (factor rebutting expansion function) Let 𝑥 ∈ 𝐴𝑐𝑡,
Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. An expansion function 𝑒 is an rebutting
expansion function, in this case denoted by, 𝑟𝑒 (Φ), iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ >0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑒 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ ∪ 𝑒 (Φ), V) ⩾ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ <0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑒 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
𝑥,Φ∪𝑒 (Φ)
(ii) 𝜎 (𝑥, Φ ∪ 𝑒 (Φ), V) < 0 and 𝐵𝐷𝑒𝑚 (V) > 0
Definition 4.5. (factor rebutting expansion operator) Let 𝑢𝑒 be an
undercutting expansion operation. Then we define an operator ⊕
on sets of factors such that Φ ⊕ = Φ ∪ 𝑟𝑒 (Φ).
Let us now introduce the two contraction operators. These op-
erators shift the balance either by weakening the “winning” side,
i.e., by deleting intensifying factors for that side (undercutting ex-
pansion), or by strengthening the “losing”, side, i.e., by deleting
attenuating factors for that side (rebutting expansion).Let us begin
by introducing the contraction function.
Definition 4.6. (factor contraction function) We define a factor
contraction function 𝑐 : P (𝐹𝑎𝑐𝑡) −→ P (𝐹𝑎𝑐𝑡) − {∅} such that for
every Φ ⊂ 𝐹𝑎𝑐𝑡, 𝑐 (Φ) ⊆ Φ.
Now we introduce the undercutting contraction operator.
Definition 4.7. (factor undercutting contraction function) Let 𝑥 ∈
𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. A contraction function 𝑐 is an under-
cutting contraction function, in this case denoted by 𝑢𝑐 (Φ) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ >0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
𝑥,Φ−𝑐 (Φ)
(ii) 𝜎 (𝑥, Φ − 𝑐 (Φ), V) ⩾ 0 ou 𝐵𝐷𝑒𝑚 (V) ⩽ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ − 𝑐 (Φ), V) < 0
Definition 4.8. (factor undercutting contraction operator) Let 𝑢𝑐
be a factor undercutting contraction function. Then we define and
operator − on sets of factors such that Φ− = Φ − 𝑢𝑐 (Φ).
We turn now to the construction of the rebutting contraction
operator.
Definition 4.9. (factor rebutting contraction function) Let 𝑥 ∈ 𝐴𝑐𝑡,
Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. A contraction function 𝑐 is a factor rebut-
ting contraction function, in this case denoted by 𝑟𝑐 (Φ) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ >0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ − 𝑐 (Φ), V) ⩾ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ <0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
𝑥,Φ−𝑐 (Φ)
(ii) 𝜎 (𝑥, Φ − 𝑐 (Φ), V) < 0 and 𝐵𝐷𝑒𝑚 (V) > 0
96
Maranhão, de Souza and Sartor
Definition 4.10. (factor rebutting contraction operator) Let 𝑟𝑒 be a
factor rebutting contraction function. Then we define and operator
⊖ on sets of factors such that Φ ⊖ = Φ − 𝑟𝑐 (Φ).
Introduce some examples for each contraction operator.
4.3 Value expansion and contraction operators
In the previous section we have seen how changes in set of the
relevant factors may modify the evaluation of an action. This may
happen either when some additional factors are considered relevant
or when some previously considered factors are rejected as being
irrelevant. Similarly, changes in the set of values, which have been
taken into account, may modify the valuation of an action. This
may happen when additional values are considered contextually
relevant, and thus added into the deliberation, or when some values
previously assumed to be relevant are discarded in the given context.
Hence, in the following, we define the value change operators
capturing these possible argument moves, which may provoke a
shift in the original balance of the proportional impact of actions
on values. Let us begin with a value expansion function.
Definition 4.11. (value expansion function) We define a value
expansion function 𝑣𝑒 : P (𝑉 𝑎𝑙) −→ P (𝑉 𝑎𝑙) − {∅} such that for
every V ⊆ 𝑉 𝑎𝑙 V ∩ 𝑣𝑒 (V) = ∅.
The first operator weakens the “winner” in the balance, which
compares promoted versus demoted values, thus provoking a shift
in the original balance.
Definition 4.12. (value undercutting expansion function) Let 𝑥 ∈
𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. An expansion function 𝑣𝑒 is a value
undercutting expansion function 𝑣𝑢𝑒 (V) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ 𝐷𝑒𝑚 (V)
>0
(i) for all 𝑣𝑖 ∈ 𝑣𝑒 (V), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
(ii) 𝜎 (𝑥, Φ, V ∪ 𝑣𝑒 (V)) ⩾ 0 or 𝐵𝑥,Φ 𝐷𝑒𝑚 (V∪𝑣𝑒 (V))
⩽0
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ <0
𝐷𝑒𝑚 (V)
(i) for all 𝑣𝑖 ∈ 𝑣𝑒 (V), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ, V ∪ 𝑣𝑒 (V)) < 0
Definition 4.13. (value undercutting expansion operator) Let 𝑣𝑢𝑒
be a value undercutting expansion function. Then we define and
operator ⊳ on sets of values such that V⊳ = V ∪ 𝑣𝑢𝑒 (V).
Now we define the operator which strengthen the “opponent”
in order to shift the balance.
Definition 4.14. (value rebutting expansion function) Let 𝑥 ∈ 𝐴𝑐𝑡,
Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. A value expansion function 𝑣𝑒 is an rebut-
ting expansion function 𝑣𝑟𝑒 (V) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ 𝐷𝑒𝑚 (V)
>0
(i) for all 𝑣𝑖 ∈ 𝑣𝑒 ((V)), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ, V ∪ 𝑣𝑒 (V)) ⩾ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ𝐷𝑒𝑚 (V)
<0
(i) for all 𝑣𝑖 ∈ 𝑣𝑒 ((V)), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
(ii) 𝜎 (𝑥, Φ, V ∪ 𝑣𝑒 (V)) < 0 and 𝐵𝑥,Φ 𝐷𝑒𝑚 (V∪𝑣𝑒 (V))
>0
Definition 4.15. (value rebutting expansion operator) Let 𝑣𝑟𝑒 be a
value rebutting expansion function. Then we define an operator ⊲
on sets of values such that V⊲ = V ∪ 𝑣𝑟𝑒 (V).
A dynamic model for balancing values
Now we are going to introduce the contraction operators that
subtract values from the original set of relevant values used in the
balance. The balance shifts either by deleting values influenced by
intensifying factors in the “winner” side of the balance between
demoted and promoted values (undercutting contraction) or by
strengthening the “opponent”, that is, by deleting values which
are affected by attenuating factors in the side that “lost” the bal-
ance (rebutting contraction). Let us begin by introducing the value
contraction function.
Definition 4.16. (value contraction function) Let V be a set of val-
ues, then a value contraction function is a function 𝑣𝑐 : P (𝑉 𝑎𝑙) −→
P (𝑉 𝑎𝑙) − ∅ such that 𝑣𝑐 (V) ⊆ V.
First we introduce the value undercutting contraction operator.
Definition 4.17. (value undercutting contraction function) Let 𝑥 ∈
𝐴𝑐𝑡, Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. A value contraction function 𝑣𝑐 is a
value undercutting contraction function 𝑣𝑢𝑐 (V) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ 𝐷𝑒𝑚 (V)
>0
(i) for all 𝑣𝑖 ∈ 𝑣𝑐 (V), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
(ii) 𝜎 (𝑥, Φ, V − 𝑣𝑐 (V)) ⩾ 0 or 𝐵𝑥,Φ ⩽0
𝐷𝑒𝑚 (V−𝑣𝑐 (V))
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ <0
𝐷𝑒𝑚 (V)
(i) for all 𝑣𝑖 ∈ 𝑣𝑐 (V), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) > 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ, V − 𝑣𝑐 (V)) < 0
Definition 4.18. (value undercutting contraction operator) Let 𝑣𝑢𝑐
be a value undercutting contraction function. Then we define an
operator ÷ on sets of factors such that V÷ = V − 𝑣𝑢𝑐 (V).
Finally, we turn to the construction of the value rebutting con-
traction operator.
Definition 4.19. (value rebutting contraction function) Let 𝑥 ∈ 𝐴𝑐𝑡,
Φ ⊆ 𝐹𝑎𝑐𝑡 and V ⊆ 𝑉 𝑎𝑙. A contraction function 𝑐 is a value rebutting
contraction function 𝑣𝑟𝑐 (Φ) iff:
• for 𝜎 (𝑥, Φ, V) < 0 and 𝐵𝑥,Φ 𝐷𝑒𝑚 (V)
>0
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝑃𝑟𝑜𝑚(V)
(ii) 𝜎 (𝑥, Φ, V − 𝑣𝑐 (V)) ⩾ 0
• for 𝜎 (𝑥, Φ, V) ⩾ 0 or 𝐵𝑥,Φ <0
𝐷𝑒𝑚 (V)
(i) for all 𝑓𝑖 ∈ 𝑐 (Φ), 𝐼𝑥 (𝑓𝑖 , 𝑣𝑖 ) < 0 and 𝑣𝑖 ∈ 𝐷𝑒𝑚(V)
(ii) 𝜎 (𝑥, Φ, V − 𝑣𝑐 (V)) < 0
Definition 4.20. (factor rebutting contraction operator) Let 𝑟𝑒 be a
factor rebutting contraction function. Then we define an operator
⊗ on sets of factors such that V ⊗ = V − 𝑣𝑟𝑐 (V).
Remark. The functions of undercutting expansion, rebutting ex-
pansion, undercutting contraction, rebutting contraction, value-
undercutting expansion, value-rebutting expansion, value-undercut-
ting contraction, value-rebutting contraction defined above may
not exist depending on the model V assumed. If the conditions to
apply these functions hold in the assumed model V we say that
the function is applicable in the model V.
4.4 Change operations on axiological systems
The change operators we have developed so far modify the set of
relevant factors or the set of relevant values, provoking, if the con-
ditions for application of each function are present in a particular
97
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
model, a shift in the solution provided by the axiological system
AS𝑥 in the modified constellation of factors or values considered
in the evaluation.
If those conditions are present and the functions apply, provided
that the sequence of factors of the original set of factors is preserved,
then a success and an inclusion result may be shown.
Theorem 4.21. Let AS𝑥Φ,V be an axiological system 2 based on
model V, 𝑒 an expansion function and 𝑐 a contraction function, then:
i. AS𝑥Φ,V ⊂ AS𝑥Φ+ ,V and, if 𝑒 is an 𝑢𝑒 applicable in V, then
Ó Ó
either AS𝑥Φ+ ,V ⊢ 𝑂 𝑣 (𝑥 | Φ) ∧ 𝑃 𝑣 (𝑥 | Φ+ ) or AS𝑥Φ+ ,V ⊢
Ó Ó +
𝑃 𝑣 (𝑥 | Φ) ∧ 𝑂 𝑣 (𝑥 | Φ )
ii. AS𝑥Φ,V ⊂ AS𝑥Φ⊕ ,V and if 𝑒 is an 𝑟𝑒 applicable in V, then either
Ó Ó Ó
AS𝑥Φ+ ,V ⊢ 𝑂 𝑣 (𝑥 | Φ)∧𝑃 (𝑥 | Φ ⊕ ) or AS𝑥Φ⊕ ,V ⊢ 𝑃 𝑣 (𝑥 | Φ)∧
Ó ⊕
𝑂 𝑣 (𝑥 | Φ )
iii. AS𝑥Φ− ,V ⊂ AS𝑥Φ,V and if 𝑐 is an 𝑢𝑐 applicable in V, then
Ó Ó
if AS𝑥Φ,V ⊢ 𝑂 𝑣 (𝑥 | Φ), then AS𝑥Φ− ,V ⊢ 𝑃 𝑣 (𝑥 | Φ− ) or if
Ó Ó
AS𝑥Φ,V ⊢ 𝑃 𝑣 (𝑥 | Φ), then AS𝑥Φ− ,V ⊢ 𝑂 𝑣 (𝑥 | Φ− )
iv. AS𝑥Φ⊖ ,V ⊂ AS𝑥Φ,V and if 𝑐 is an 𝑢𝑐 applicable in V, then
Ó Ó
if AS𝑥Φ,V ⊢ 𝑂 𝑣 (𝑥 | Φ), then AS𝑥Φ⊖ ,V ⊢ 𝑃 𝑣 (𝑥 | Φ ⊖ ) or if
Ó Ó
AS𝑥Φ,V ⊢ 𝑃 𝑣 (𝑥 | Φ), then AS𝑥Φ⊖ ,V ⊢ 𝑂 𝑣 (𝑥 | Φ ⊖ )
Proof. Straightforward from theorem 3.3. □
Changes in the set of values may determine changes in the rules
of the axiological system. For instance, in the precedent discussed at
example 2.8, the sequence ⟨∅, {𝑝𝑟𝑜𝑝}, {𝑝𝑟𝑜𝑝, 𝑎𝑟𝑟 }, {𝑝𝑟𝑜𝑝, 𝑎𝑟𝑟, 𝑚𝑜𝑏}⟩
would deliver the following axiological system AS𝑎𝑐𝑐 :
{𝑂 𝑣 (¬𝑎𝑐𝑐), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧
¬𝑎𝑟𝑟 ), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ∧ 𝑚𝑜𝑏), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ∧ ¬𝑚𝑜𝑏)}
Suppose one argues that the context of an arrest also strongly
impacts the promotion of national security 𝑆𝑒𝑐 with a proportional
impact 𝐵𝑎𝑐𝑐
{𝑝𝑟𝑜𝑝,𝑎𝑟𝑟,𝑚𝑜𝑏 },{𝑆𝑒𝑐 }
= .6, then we would have the follow-
ing axiological system AS𝑎𝑐𝑐 :
{𝑂 𝑣 (¬𝑎𝑐𝑐), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ), 𝑂 𝑣 (¬𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧
¬𝑎𝑟𝑟 ), 𝑃 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ 𝑎𝑟𝑟 ∧ 𝑚𝑜𝑏), 𝑂 𝑣 (𝑎𝑐𝑐 |𝑝𝑟𝑜𝑝 ∧ ¬𝑎𝑟𝑟 ∧ 𝑚𝑜𝑏)}
Hence inclusion simpliciter does not hold. But considering that
the modification is going to occur in a particular step of the evalua-
tion, we have a qualified form of inclusion where 𝐴𝑆𝑛−1,V
𝑥 𝑥
⊂ AS𝑛,V ∗,
for some 𝑛 in the sequence of the subsets of factors considered, and
where ∗ may be any value-change operator here defined. For con-
venience, we have denoted AS𝑛𝑥 as AS𝑛,V 𝑥 in order to express the
change of the set of relevant values.
Theorem 4.22. Let AS𝑥Φ,V be an axiological system 3 based on
model V, 𝑣𝑒 a value expansion function and 𝑣𝑐 a value contraction
function, then:
i. if 𝑣𝑒 is an 𝑣𝑢𝑒 or a 𝑣𝑟𝑒 function applicable in V, then if
Ó Ó
AS𝑥Φ+ ,V ⊢ 𝑂 𝑣 (𝑥 | Φ) then AS𝑥Φ+ ,V∗ ⊢ 𝑃 𝑣 (𝑥 | Φ) and if
Ó Ó
AS𝑥Φ+ ,V ⊢ 𝑃 (𝑣 𝑥 | Φ), then AS𝑥Φ+ ,V∗ ⊢ 𝑂 𝑣 (𝑥 | Φ), for ∗ ∈
{⊳, ⊲}
2 We are including V in the notation of AS𝑥 for convenience. Also to avoid clutter
Φ Ó
with superscripts in the notation, we are going to substitute Φ for Φ∧ .
3 We are including V in the notation of AS𝑥 for convenience.
Φ
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
ii. if 𝑣𝑒 is an 𝑣𝑢𝑐 or a 𝑣𝑟𝑐 applicable in V, then if AS𝑥Φ+ ,V ⊢
Ó Ó
𝑂 𝑣 (𝑥 | Φ) then AS𝑥Φ+ ,V∗ ⊢ 𝑃 𝑣 (𝑥 | Φ) and if AS𝑥Φ+ ,V ⊢
Ó Ó
𝑃 𝑣 (𝑥 | Φ), then AS𝑥Φ+ ,V∗ ⊢ 𝑂 𝑣 (𝑥 | Φ), for ∗ ∈ {÷, ⊗}
Proof. Straighforward from theorem 3.3. □ ACKNOWLEDGMENTS
5 FINAL REMARKS
The additive model we propose makes strong assumptions about
the behaviour of factors as reasons for the evaluation of the action
at stake. It assumes an atomist conception to the effect that a factor
that is a reason for the action in one case always remains a reason
when new factors (reasons) are considered. Also, a factor always
keeps the same polarity and its contribution in terms of impact
in each value considered. That is, not only the action is invariant
with respect to each value 𝑉𝑖 in the assessment, but also the factors
retain the same contribution with the same polarity and intensity
(it intensifies or attenuates in a given degree the action’s promotion
or demotion of the value 𝑉𝑖 ) in the assessment of new evaluations
considering different constellations of factors. So, the model as-
sumes a strong atomist conception in the assessment of the moral
evaluation of actions, and consequently, the moral evaluation of
positively enacted rules.
Different conceptions are possible, relaxing in different degrees
the atomist assumptions. For instance, a multiplicative model may
make the contribution of a factor void if a new factor with propor-
tional impact 0 is considered ([14]). One may also assume a model
where the intensity of the contributions of factors may vary or
invert polarity, changing from an intensifier to an attenuator of the
impact of the action on a value, or vice-versa. Or even the polarity
of the impact of an action may change, for instance an action that,
given a constellation of factors, promotes a value, may demote the
same value given a new constellation of factors.
Such steps would move the model towards an holist conception
where a feature which is a reason for or against the action in one
case, may not be a reason or be an opposite reason in another
case, and where each contribution for the evaluation is variant and
contextual. Holist conceptions are usually, although not necessar-
ily, connected to moral particularism, according to which moral
evaluations of actions do not depend on the subsumption of moral
principles, as opposed to moral universalism, according to which
moral judgment is intrinsically connected to the instantiation of
moral principles ([11]).
We shall explore such variations in future developments of the
model.
In this paper we have limited ourselves to use the model to
induce value-rules and build axiological systems, which may be
assumed as premises in any inference system of dyadic deontic
logic, satisfying some conditions. Hence another relevant path to
explore is to embed the model into an inferential system [7]. We see
two interesting alternatives. One of them is to follow the statutory
interpretation line of research in AI & Law and embed the model
in a deontic logic with revision operators applied on the logical
consequences of normative systems in the AGM-style ([1]). Steps in
this direction have been made by [18]. The other is to explore the
resemblance of the revision operators and the axiological system
with an argumentation structure and embed the model or explore
its connections with a logic of defeasible argumentation. Finally,
98
Maranhão, de Souza and Sartor
the model could be extended in order to compare different and in-
dependent actions, or evaluate if positively enacted rules regarding
different and independent actions are justified.
Juliano Maranhão acknowledges the support by the Fundação de
Apoio à Pesquisa do Estado de São Paulo (FAPESP 2019/07665-4)
and the IBM Corporation to the Center for Artificial Intelligence
(C4AI/USP. Giovanni Sartor has been supported by the H2020 Euro-
pean Research Council (ERC) Project “CompuLaw” (G.A. 833647).
REFERENCES
[1] Carlos E Alchourrón, Paul Gärdenfors, and David Makinson. 1985. On the Logic
of Theory Change: Partial Meet Contraction and Revision Functions. Journal of
Symbolic Logic 50 (1985), 510–530.
[2] R. Alexy. 2002. A Theory of Constitutional Rights. Oxford University Press.
[3] Robert Alexy. 2003. The Argument from Injustice. {A} Reply to legal positivism.
Oxford University Press.
[4] Robert Alexy. 2010. The Dual Nature of Law. Ratio Juris 23 (2010), 167–182.
[5] Aharon Barak. 2005. Purposive Interpretation in Law. Princeton University Press.
[6] T. J. M. Bench-Capon and G. Sartor. 2003. A Model of Legal Reasoning with Cases
Incorporating Theories and Values. Artificial Intelligence 150 (2003), 97–142.
[7] Christoph Benzmüller, David Fuenmayor, and Bertram Lomfeld. 2021. Encod-
ing Legal Balancing: Automating an Abstract Ethico-Legal Value Ontology in
Preference Logic. arXiv:2006.12789 [cs.AI]
[8] D. H. Berman and C. D. Hafner. 1993. Representing Teleological Structure in Case-
based Reasoning: The Missing Link. In Proceedings of the Fourth International
Conference on Artificial Intelligence and Law (ICAIL). ACM, 50–9.
[9] G. Boella, G. Governatori, A. Rotolo, and L. van der Torre. 2010. Lex minus dixit
quam voluit, lex magis dixit quam voluit: A formal study on legal compliance
and interpretation. In AICOL-I/IVR-XXIV’09 Proceedings of the 2009 international
conference on AI approaches to the complexity of legal systems: complex systems,
the semantic web, ontologies, argumentation, and dialogue. Springer, 162–183.
[10] G. Boella, L. van der Torre, and G. Pigozzi. 2016. AGM Contraction and Revision
of Rules. J Log Lang Inf 25 (2016), 273–297.
[11] J. Dancy. 2004. Ethics Without Principles. Oxford University Press.
[12] M. Grabmair. 2017. Predicting Trade Secret Case Outcomes using Argument
Schemes and Learned Quantitative Value Effect Tradeoffs. In Proceedings of
ICAIL-2017. ACM, 89–98.
[13] Herbert L. A. Hart. 1994. The Concept of Law (2nd ed.). Oxford University Press.
[14] Shelly Kagan. 1988. The additive Falacy. Ethics 99 (1988), 5–31.
[15] J. Maranhão. 2001. Refinement: a tool to deal with inconsistencies. In Proceedings
of the Eight International Conference on AI and Law. ICAIL-2001. ACM Press,
52–59.
[16] J. Maranhão. 2017. A logical Architecture for dynamic legal interpretation. In
Proceedings of the Eight International Conference on AI and Law ICAIL ’17,. ACM
Press, 129–38.
[17] JSA Maranhão and G Sartor. [n.d.]. Interpretive Normative Systems. In 15th
International Conference on Deontic Logic and Normative Systems- DEON 2020-
2021, F Liu, A Marra, P Portner, and F Van de Putte (Eds.). College Publications.
[18] Juliano Maranhão and Giovanni Sartor. 2019. Value assessment and revision in
legal interpretation. In Proceedings of the 17th International Conference on Artificial
Intelligence and Law, ICAIL 2019. Association for Computing Machinery, Inc, New
York, New York, USA, 219–223. https://doi.org/10.1145/3322640.3326709
[19] H. Prakken, A. Wyner, T.R Bench-Capon, and K. Atkinson. 2015. A formalisation
of argumentation schemes for legal case-based reasoning in ASPIC+. Journal of
Logic and Computation 25 (2015), 1141–1166.
[20] G. Radbruch. 2006. Statutory Lawlessness and Supra-Statutory Law. Oxford
Journal of Legal Studies 6 (2006), 1–11. (1st ed. 1946.).
[21] G. Sartor. 2013. The Logic of Proportionality: Reasoning with Non-Numerical
Magnitudes. German Law Journal (2013), 1419–57.
[22] G. Sartor. 2018. Consistency in balancing: from value assessments to factor-
based rules. In Proportionality in Law: An Analytical Perspective, D. Duarte and
S. Sampaio (Eds.). Springer, 121–36.
[23] Douglas Walton, Giovanni Sartor, and Fabrizio Macagno. 2018. Statutory Inter-
pretation as Argumentation. In Handbook of Legal Reasoning and Argumentation,
G. Bongiovanni, G. Postema, A. Rotolo, G. Sartor, C. Valentini, and D. Walton
(Eds.). Springer, 519–60.
 Case-level Prediction of Motion Outcomes in Civil Litigation
Devin J. McConnell
devin.mcconnell@uconn.edu
Department of Computer Science
University of Connecticut
Storrs, Connecticut, USA
Sachin Pandya
sachin.pandya@uconn.edu
School of Law
University of Connecticut
Hartford, Connecticut, USA
ABSTRACT
Lawyers regularly predict court outcomes to make strategic deci-
sions, including when, if at all, to sue or settle, what to argue, and
how to reduce their clients’ liability risk. Yet, lawyer predictions
tend to be poorly calibrated and biased, which exacerbate unjustifi-
able disparities in civil case outcomes. Current machine learning
(ML) approaches for predicting court outcomes are typically con-
strained to final dispositions or are based on features unavailable in
real-time during litigation, like judicial opinions. Here, we present
the first ML-based methods to support lawyer and client decision
making in real-time for motion filings in civil proceedings. Using
the State of Connecticut Judicial Branch administrative data and
court case documents, we trained six classifiers to predict motion
to strike outcomes in tort and vehicular cases between July 1, 2004
and February 18, 2019. Integrating dense word embeddings from
complaint documents, which contain information specific to the
claims alleged, with the Judicial Branch data improved classifica-
tion accuracy across all models. Subsequent models defined using a
novel attorney case-entropy feature, dense word embeddings using
corpus specific TF-IDF weightings, and algorithmic classification
rules yielded the best predictor, Adaboost, with a classification
accuracy of 64.4%. An analysis of feature importance weights con-
firmed the usefulness of incorporating attorney case-entropy and
natural language features from complaint documents. Since all fea-
tures used in model training are available during litigation, these
methods will help lawyers make better predictions than they oth-
erwise could given disparities in lawyer and client resources. All
ML models, training code, and evaluation scripts are available at
https://github.com/aguiarlab/motionpredict.
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Machine learning approaches; Natural language processing.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466101
99
James Zhu
james.zhu@uconn.edu
Department of Computer Science
University of Connecticut
Storrs, Connecticut, USA
Derek Aguiar
derek.aguiar@uconn.edu
Department of Computer Science
University of Connecticut
Storrs, Connecticut, USA
ACM Reference Format:
Devin J. McConnell, James Zhu, Sachin Pandya, and Derek Aguiar. 2021.
Case-level Prediction of Motion Outcomes in Civil Litigation. In Eighteenth
International Conference for Artificial Intelligence and Law (ICAIL’21), June
21–25, 2021, Sao Paulo, Brazil. ACM, New York, NY, USA, 16 pages. https:
//doi.org/10.1145/3462757.3466101
1 INTRODUCTION
In the late nineteenth century, American jurist Oliver Wendell
Holmes declared that law is “nothing more pretentious” than pre-
dictions of what courts in fact do [38]. Lawyers regularly predict
what courts will do and, on that basis, make strategic decisions,
including what to argue, when, if at all, to sue or prosecute, set-
tle cases, and how to reduce their clients’ liability risk. Indeed,
because most civil proceedings settle before any trial occurs, litiga-
tion outcomes (civil settlements) vary on how attorneys estimate
what would happen if the case went to trial and how a trial judge
or jury would rule. Yet, despite a literature in litigation decision
analysis [13], lawyers’ actual predictions tend to be overconfident,
not well calibrated [34, 39], and surprisingly vulnerable to bias
associated with advocating for a particular view [55].
While lawyers have always made strategic decisions based on
what they believed about outcomes in similar cases, the collection
of large and diverse data from civil proceedings facilitates a new
computational view of legal decision making. This computational
perspective of court outcome prediction has been bolstered by
the rise of machine learning (ML), which is a branch of artificial
intelligence (AI) focused on the development of computational
systems that learn from historical data to make predictions or infer
patterns in big, noisy, or high-dimensional data sets [28, 74].
Combining ML with collections of dockets, judicial opinions,
and court documents presents two significant opportunities. First,
ML techniques can increase how accurately lawyers can predict
legal outcomes to help support real-time decision making. Compu-
tational modeling in law has relied primarily on the text of formal
judicial opinions, with or without annotation from domain experts,
to classify the outcome reached or the type of legal reasoning used
to reach that outcome [5, 35, 54, 62]. In some scenarios, ML methods
can predict judicial decisions with an accuracy that exceeds legal
scholars [41, 65]. However, court opinions are not available during
litigation, and thus this retrospective process stands in contrast to
predicting judge or jury decision-making itself [25, 57].
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Second, since court outcomes vary with attorney quality and
client resources, predictions using ML may reduce the litigation
disadvantages faced by the poor, racial minorities, and other vulner-
able groups [20]. These problems are exacerbated by overwhelming
workloads faced by civil courts and lawyers. Increased workloads
contribute to workplace strain, which has detrimental effects on
the ability to function effectively [40] and, for lawyers, a decreased
perception of ability to uphold the law [12]. However, to ensure
widespread acceptability and trustworthiness of algorithmic deci-
sions [60], models must be accurate and explainable to all parties.
Methods for characterizing judicial decisions in previous work
have focused on court opinions, ignoring the many important for-
mal procedures that lead to a final judgment. In this work, we
present the first client and lawyer support methods that predict
court outcomes at the level of individual motions. Motions are for-
mal requests to judges for an official ruling on a contested issue.
They can be submitted before, during, or after the trial and can
have a significant impact on the final disposition of the lawsuit. For
example, motions to strike petition for the removal of all or a subset
of the opposing party’s pleading. While many court documents
are filed over the course of legal proceedings that are relevant to
motion outcomes, we focus on complaint documents because (a)
they contain the facts alleged and legal claims asserted and (b) are
available to all parties when a lawsuit begins and can therefore be
used to support decision making.
We present a general overview of computational prediction of
litigation outcomes and our contributions in Section 2. Section
3 provides details of our methods, where we describe the court
administrative data and legal documents as well as our approach to
feature engineering and predictive modelling. We present results
on approximately 15 years of Connecticut civil case data in Section
4, followed by a discussion and conclusions in Sections 5 and 6.
2 BACKGROUND
Predicting legal outcomes has traditionally been the purview, not
only of practicing lawyers, but also researchers of judicial behavior
in law, political science, and recently, computer science [37]. Early
efforts to computationally model legal decision making focused on
representations of rules obtained from case law and legislation [68,
32, 64]. When legal decisions can be modelled as a deterministic
process, rule-based AI has achieved considerable success [16, 23].
More recently, ML has made a large impact on the research and
practice of law in general, and in predictive litigation analysis
specifically [2, 3, 28, 63].
Several efforts are focused on extracting standardized data sets to
support ML in law. The Supreme Court Database contains over two
hundred years of U.S. Supreme Court cases each containing hun-
dreds of variables [70]. The CASELAW4 data set contains 350,000
common law judicial decisions extracted from US State appellate
courts [62]. The University of Oxford is constructing a database
of 100,000 US court case decisions with features that include the
facts of the case, judgements, location, timing, and judicial opin-
ions [24]. These and other similar works [77] provide benchmarks
that will accelerate the use of ML in litigation in a similar manner
as CIFAR [44] and MNIST [49] for image classification.
100
McConnell et al.
The most prevalent features used by ML methods to model court
decisions are extracted from legal documents. Legal documents, typ-
ically court records, judicial opinions, and legislation, are difficult
to model due to their high dimensionality. For instance, modelling
documents as bags-of-words, which treats a document as an un-
ordered multiset of words, is commonly assumed in applications
like text classification [43] and topic modelling [9]. With this simpli-
fying assumption, the dimension of a document is proportional to
the vocabulary size, |𝑉 |, which is prohibitively large for many ML
methods. When the ordering of words in a document is considered,
the dimensionality grows exponentially. Therefore, methods typi-
cally seek lower dimensional representations of legal documents
that preserve relevant structure of the underlying text.
Early legal document representations focused on summary sta-
tistics, like word length [8] or other metadata including document
complexity [22], publication date, and amendment counts [47].
Knowledge representations of the dyadic citation relationships
between documents are typically modelled as citation networks,
where vertices correspond to legal documents and a directed edge
exists from document 𝐴 to document 𝐵 if 𝐴 cites 𝐵 [27]. Important
features can be extracted from the connectivity structure of legal
citation networks, e.g., directed paths can be interpreted as chains
of legal precedent or network centrality and in-degree can indicate
case importance.
More recently, representations from natural language processing
(NLP) have been used to compute richer representations of legal
documents. Term frequency–inverse document frequency (TD-IDF)
is a statistic computed for a word 𝓌 and document 𝒹:
 
|D|
𝑓𝑇 (𝓌,𝒹) = 𝒻𝒹 (𝓌) · log Í
𝒹 ∈D 1 (𝓌 ∈ 𝒹)
where 𝒹 ∈ D is a document in corpus D, 𝒻𝑑 (𝑤) is the frequency
of word 𝓌 in document 𝒹, and 1 (𝓌 ∈ 𝒹) is an indicator function
that is 1 if 𝓌 appears in 𝒹 and 0 otherwise. TD-IDF is often used
as a corpus-specific importance weighting for words [80].
State-of-the-art language embedding models have seen recent
success by providing lower 𝑑 dimensional embeddings of words
and documents, where 𝑑 ≈ 102 << |𝑉 |. These architectures are pre-
trained on large general corpora and then either applied directly
to legal documents or fine-tuned using transfer learning to legal
specific applications. The law2vec model is a neural embedding
architecture based on word2vec [56] that was pre-trained on a large
legal corpus consisting of mostly legislative documents [14]. Taking
advantage of recent developments in NLP, some researchers have
used transfer learning techniques from pre-trained transformer
models [21] to classify U.S. Fourth Amendment cases [35].
Machine learning methods for predicting court outcomes have,
thus far, been mostly trained using judicial opinions. One study
developed a random forest classifier to predict over 240,000 justice
votes and about 28,000 case outcomes for the U.S. Supreme Court
from 1816 through 2015 [41]. The method predicted court decisions
with 70.2% accuracy and justice votes with 71.9% accuracy. By
comparison, legal experts at best accurately predicted about 66%
of the outcomes in sixty-eight cases argued in the U.S. Supreme
Court’s 2002 Term [65]. It is a common practice to use a court’s
past decisions to predict its future decisions, as was done with data
from the European Court of Human Rights [54]. French Supreme
Case-level Prediction of Motion Outcomes in Civil Litigation ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Court Admin Hyperparameter ML Model Training

Feature Engineering
Database Optimization and Evaluation

SQL Parameter 1 AdaBoost Decision

Minimal Feature Set Trees
Subset Feature Set Gradient Random
Boosting Forests

Parameter 2
SVM XGBoost
word2vec
OCR
TF-IDF
rule-based algorithm Grid Search
Complaint
Documents

Figure 1: Motion prediction pipeline overview. Two sets of features were computed from the State of Connecticut Judicial
Branch court administrative database (minimal and subset, Table 1) and combined with natural language features extracted
from complaint documents using word2vec, TF-IDF, and a rule-based algorithm. We optimized the hyperparameters of six ML
models using grid search. Here, grid search is described in two dimensions where the circles denote parameter configurations
and the curves on each axis denote the marginal classification accuracy. In this example, classification accuracy has higher
variability across values of parameter 1, which primarily determines the choice of the best parameter setting (red point). A
toy decision tree on the attorney specialization and major code features is shown under our six ML models.

Court decisions have been modelled from historical rulings data
using a linear support vector machine (SVM) classifier, assuming a
bag-of-words representation for the rulings documents [73].
Importantly, these methods require data from decisions and opin-
ions, which distinguish them from other uses of ML to support
real-time litigation support in court cases [25, 57]. Other tools, like
MyOpenCourt, provide an AI platform directly to clients for an-
swering legal questions [45]. While these tools do provide real-time
support for decision making, the focus is on data mining and learn-
ing legal recommendations to support self-represented litigants,
not predictive analytics.
2.1 Our Contributions
Existing approaches to predicting court case outcomes focus on
final judgements of a trial and rely on retrospective court data from
appellate or national court decisions, and thus are not amenable
for informing motion-level decision making [4, 41, 54, 73]. In this
work, we contribute to the field of legal analytics in several ways:
(1) we define new lower dimensional features to support predic-
tive modelling at the case-level;
(2) we develop and benchmark the first computational pipeline
to assist lawyer and client decision making through the pre-
diction of motion outcomes in district court data (Figure 1);
(3) we analyze the predictability of motions to strike using both
court administrative data and natural language features ex-
tracted from complaint documents;
(4) we provide this pipeline (code, trained models, evaluation
scripts) freely available and open-source.
In this work, we consider 𝑘 = 2 where 1 and 2 correspond to
motion denied and granted respectively. Let the observed data be
(𝑥𝑥 1, . . . ,𝑥𝑥 𝑖 , . . . ,𝑥𝑥 𝑛 ) = 𝑋 ∈ R𝑛×𝑝 , an 𝑛 × 𝑝 matrix of 𝑛 civil court
cases each containing 𝑝 covariates. Note that 𝑋 can, in general,
contain real, nominal, ordinal, or integer valued variables. Given a
training set (𝑋 𝑋 ,𝑌 𝑌 ), the goal is to build a model that predicts class
labels 𝑌¯ 𝑡𝑒𝑠𝑡 from held-out test data 𝑋 𝑡𝑒𝑠𝑡 to maximize classifica-
tion accuracy 𝑇|𝑌𝑌𝑃 +𝑇 𝑁| where TP and TN are true positives and true
𝑡𝑒𝑠𝑡
negatives respectively.
3.1 Connecticut Civil Court Data
The values for 𝑋 and 𝑌 were collected from the State of Connecticut
Judicial Branch, which provides access to materials such as public
records and court case documents, as well as researcher access to
their civil court administrative data [71]. The court administrative
data is populated by courthouse staff and stored in a centralized
relational database, which was rebuilt locally in mySQL. Court case
documents are scanned and made available through the Judicial
Branch Law Library API.
We focus on predicting the outcomes for motions to strike. A
motion to strike has significant influence on case outcomes and
are therefore an important factor in legal decision making. In Con-
necticut, a motion to strike is a written petition typically from a
defendant to a judge to remove part or all of a plaintiff’s complaint
allegations based on legal insufficiency. While we restrict our at-
tention to civil cases filed in Connecticut, the methods apply more
generally to other states as long as minimal docket information and
complaint documents are available.

3 METHODS 3.2 Feature Engineering: Court Administrative

Consider 𝑛 motions filed in civil court whose outcomes are rep- Data
resented as a collection of random variables (𝑦1, . . . ,𝑦𝑖 , . . . ,𝑦𝑛 )𝑇 = Most features in the court administrative data were either not
𝑌 ∈ {1, . . . ,𝑘 }𝑛 where 𝑘 is the number of distinct judicial rulings. relevant to motion outcomes, had high missingness rates, or were

101
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
categorical variables with a number of levels proportional to the
size of the data. Therefore, we developed custom SQL scripts to
extract 3 informative features based on domain expertise and low
missing data rates (< 0.6% missingness). In total, we considered
four court administrative features: juris number, major code, case
location, and attorney specialization (Table 1). The juris number is a
unique identifier for the attorney or firm representing the defendant.
The major code represents the case type encoded as a Bernoulli
variable for tort or vehicular cases. The case location encodes a 15
dimensional categorical variable denoting the Connecticut superior
court location for the case.
The attorney specialization is derived based on the entropy of
the case type (i.e., major case code) distribution for each attorney.
Formally, let the number of different major case codes (e.g. tort
or vehicular) associated with an attorney be 𝑚 and the counts of
cases litigated by an attorney in each major case code be 𝑤 =
(𝑤 1, . . . ,𝑤𝑚 ). Then, we model the case counts 𝑤 for attorney 𝑎 as
a multinomial distribution with a Dirichlet prior,
𝑤𝑎 ∼𝑀𝑢𝑙𝑡𝑖𝑛𝑜𝑚𝑖𝑎𝑙 (𝜃 𝑎 )
𝜃 𝑎 ∼𝐷𝑖𝑟𝑖𝑐ℎ𝑙𝑒𝑡 (𝛼 1, . . . ,𝛼𝑚 )
In this work, we set (𝛼 1, . . . ,𝛼𝑚 ) = 1 . After observing 𝑤𝑎 , the
posterior is a Dirichlet-multinomial distribution
𝑃 (𝜃 𝑎 |𝑤𝑎 ) = 𝐷𝑖𝑟 (𝛼 1 + 𝑤𝑎1, . . . ,𝛼𝑚 + 𝑤𝑎𝑚 )
We compute the specialization for attorney 𝑎 and major case code
𝑗 as the entropy of the posterior expectation:
! set. We consider two functions (or criteria) to optimize: a simple
  Õ𝑚
𝑤𝑎 𝑗 + 𝛼 𝑗 𝑤𝑎 𝑗 + 𝛼 𝑗
𝐻 𝐸 [𝜃 𝑎 𝑗 |𝑤𝑎 ] = − Í𝑚 log Í𝑚
(𝑤 + 𝛼𝑘 )
𝑗=1 𝑘=1 𝑎𝑘
(𝑤 + 𝛼𝑘 )
𝑘=1 𝑎𝑘
where 𝐻 is Shannon’s entropy whose probability vector is the rel-
ative frequency of the 𝑗 𝑡ℎ major case code smoothed by taking
the expectation of a Dirichlet-multinomial distribution with prior
𝐷𝑖𝑟𝑖𝑐ℎ𝑙𝑒𝑡 (𝛼 1, . . . ,𝛼𝑚 ); this smoothing reduces the variance of en-
tropy calculations for attorneys with small caseloads. Note, that
this generates 𝑚 − 1 features.
We consider two feature sets derived from these four court ad-
ministrative features, minimal and subset (see Table 1).
Feature Description S the RIPPER algorithm [17, 18]. The FOIL criteria is less greedy than
Juris Number Defendant attorney or firm identifier
Attorney Case type specialization smoothed by a
Specialization Dirichlet-multinomial prior
Major Code Classifies the case type (e.g. tort) ✓ 𝑛2 + 𝑛2
Case Location Superior court location for the case ✓ where
Table 1: Features of the CT Civil Case Administrative Data-
base. The column identified as S denotes whether or not the
feature was only included in the subset feature set (and not
in the minimal feature set).
3.3 Feature Engineering: Complaint
Documents
We downloaded 7904 complaint documents associated with each
case in our data as PDF files from the State of Connecticut Judicial
102
McConnell et al.
Branch website using custom crawling scripts [71]. If a PDF con-
tained text, we used pdftotext (version 0.26.5) to convert the PDF to
a text file (see Supplemental Methods for additional details). If a PDF
contained an image, we first converted the PDF to a TIFF file using
ImageMagick (version 6.9.10-68 Q16) [72]. Then, we converted the
TIFF to text using tesseract (version 4.1.1-rc2-20-g01fb) [69]. The
tesseract optical character recognition (OCR) engine is based on
LSTM neural networks and maintained by Google. It has shown to
have high accuracy on machine-written characters and black and
white images, both of which categorize complaint documents [75].
3.3.1 Rule-based features for complaint documents. We also con-
sider algorithmically generated natural language features based
on a sequential covering rule generating algorithm [1]. A rule 𝑅
maps a condition (antecedent) to a class (consequent). Here, an-
tecedents are a conjunction of Boolean conditions indicating the
presence of a word in a complaint document and the consequent is
motion granted or denied. For example, the rule (𝑐𝑎𝑟 ∈ 𝐷) AND
(𝑎𝑐𝑐𝑖𝑑𝑒𝑛𝑡 ∈ 𝐷) AND (𝑛𝑒𝑔𝑙𝑖𝑔𝑒𝑛𝑐𝑒 ∈ 𝐷) ⇒ 1 would map a motion
to strike associated with complaint 𝐷 to denied if the words “car”,
“accident”, and “negligence” are contained within 𝐷.
The sequential covering algorithm proceeds by learning a rule
that maximizes some function. Then, the rule is added to a list and
all of the complaint documents covered by this rule are removed
from the data. This process is repeated until all documents are
removed or sufficient coverage of the data is reached. In this work,
we learned the number of rules by cross-validation on the training
criteria and the First Order Inductive Learner (FOIL) criteria.
The function with which the simple sequential covering algo-
rithm optimizes is
𝑛+ + 1
(1)
𝑛∗ + 𝑘
where 𝑛 + is the frequency of a word that appears in the granted
(GR) documents, 𝑛 ∗ is the total number of occurrences of the word
across all documents, and k is the number of classes (here, 𝑘 = 2).
In other words, the simple algorithm greedily adds a term to the
antecedent that increases the rule’s accuracy the most.
A separate function is the FOIL criteria, which is optimized by
Equation 1, attempting to balance information gain with document
coverage
 
𝑛+ 𝑛+
𝑛 +2 log2 + 2 − − log2 + 1 −
𝑛1 + 𝑛1
• 𝑛 +1 (𝑛 −
1 ) is the number of complaint documents associated
with a motion to strike that were granted (denied) that the
rule covers;
• 𝑛 +2 (𝑛 −
2 ) is the number of complaint documents associated
with a motion to strike that are changed to positive (negative)
with the addition of a prospective word to the antecedent.
3.3.2 Word embeddings for complaint documents. We considered
three architectures to construct complaint document features from
neural embeddings: word2vec [56], doc2vec [48], and law2vec [14].
The word2vec model maximizes log 𝑃 (𝑤𝑂 |𝑤 𝐼 ), or the log probabil-
ity of a given word 𝑤𝑂 given an input word 𝑤 𝐼 . The doc2vec model
Case-level Prediction of Motion Outcomes in Civil Litigation
is similar but instead of conditioning on an input word 𝑤 𝐼 , it condi-
tions on a vector representing the document. We also consider the
law2vec model which was trained on 123,066 documents, including
53,000 UK legislative documents, 62,000 European legislative doc-
ument, and thousands of other English legislative, U.S. code, and
opinion documents. We computed 300 dimensional embeddings
for each complaint document using doc2vec and word2vec models
that were pre-trained on AP News, Google News, and Wiki arti-
cles. The law2vec repository provides pretrained models for 100
and 200 dimensional vectors, thus we selected 200 dimensional
vectors. To compute a document representation using word2vec or
law2vec, we computed an average word2vec vector weighted by
term frequency–inverse document frequency (TF-IDF) [53].
We also produce models trained on a combination of word embed-
ding and rule classifiers. A simple word2vec or simple law2vec TF-
IDF model computes an average word2vec vector for each document
that is weighted by TF-IDF and including only those words identi-
fied by the simple rule-based classifier. Likewise, a FOIL word2vec
or FOIL law2vec TF-IDF model computes an average word2vec vec-
tor for each document that is weighted by TF-IDF and including
only those words identified by the FOIL rule-based classifier.
3.4 Modelling Motion Outcomes
We considered six classification models to predict motion to strike
outcomes: adaBoost [29], decision trees [66], gradient boosting [31],
random forests [11], SVM [10, 19], and XGBoost [15]. These meth-
ods were selected based on their relatively high explainability [81],
applicability in small-data settings, and ability to model mixed data
types (e.g. real and categorical-valued variables).
3.4.1 Tree Classifiers. Tree classifiers aim to construct a hierarchy
of decision rules that recursively split the data until a leaf is reached
that denotes the inferred class label. Formally, the hierarchy is rep-
resented as a tree where internal nodes denote a bifurcation of
a subset of samples based on maximum information gain splits
(although splitting based on other criteria like Gini index is also
common). Information gain is the difference between the infor-
mation entropy of the samples at an internal node in the decision
tree and the conditional entropy of the samples split on a feature.
Intuitively, after an internal node split, the bifurcation should yield
subsets that have higher purity. Since finding an optimal decision
tree is NP-hard [36, 46], trees are typically constructed in a greedy
fashion with successive maximum entropy gain splits and leaves
that define a sample’s classification. In this work we build decision
tree classifiers [66], which produce a single decision tree, and ran-
dom forest classifiers [11], which are ensembles of decision trees
built from a bootstrap subset of the training data.
3.4.2 Boosting Methods. Boosting is a general ML technique based
on building an accurate learner from an ensemble of weak learn-
ers [30]. Boosting classifiers learn an ensemble of weak learners
(e.g. shallow decision trees) sequentially. In this work, we train Ad-
aboost, gradient boosting trees, and XGBoost classifiers. Adaboost,
the first practical boosting algorithm [67], builds new decision trees
with successive boosting iterations re-weighting training instances
such that newly built decision trees focus more on the samples that
were previously misclassified. Gradient boosting trees generalize
103
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Adaboost, optimizing an approximate negative gradient of the bi-
nomial deviance loss function [31]. XGBoost, or Extreme Gradient
Boosting, is a highly efficient implementation of gradient boosting
trees with an adjusted loss function to control the complexity of
the decision trees [7]
Õ
𝑛 Õ
𝑀
𝐿𝑥𝑔𝑏 = 𝐿 (𝑦𝑖 , 𝐹 (𝑥𝑖 )) + Ω(ℎ𝑚 )
𝑖=1 𝑚=1
1
Ω(ℎ𝑚 ) = 𝛾𝑇𝑚 + 𝜆||𝑠 || 2
2
where 𝐿 (𝑦𝑖 , 𝐹 (𝑥𝑖 )) is a loss function computed from observed mo-
tion outcome 𝑦𝑖 and computed outcome 𝐹 (𝑥𝑖 ), ℎ𝑚 is the 𝑚𝑡ℎ weak
learning model with 𝑇𝑚 leaves and 𝑠 leaf output scores.
3.4.3 Support Vector Machines. Support vector machines classify
data by solving a convex optimization problem yielding a hyper-
plane that optimally separates two classes [10, 19]. Optimality is
measured with respect to the size of the margin between granted
and denied motions. They can efficiently compute non-linear deci-
sion boundaries with kernel functions, representing similarities in
an inner product space.
4 RESULTS
The court administrative data from the State of Connecticut Judicial
Branch contained 916,805 observations from 184,125 unique cases.
First, we extracted 7904 motions to strike filed by a defendant
against a plaintiff between July 1, 2004 to February 18, 2019. In
the results that follow, we focus on predicting the outcome of a
motion to strike in civil tort and vehicular cases based on court
administrative data and complaint documents.
Classifier performance can greatly depend on model selection [26].
For example, when using default parameters, the performance of
XGBoost is significantly worse than models with optimized hyper-
parameters [7]. For model selection, we performed a grid search for
each method, with 10 fold cross validation on 70% of the data used
minimal
0.60 subset
0.59
Motion Classification Accuracy
0.58
0.57
0.56
0.55
0.54
0.53
0.52
ab dt gb rf svm xgb
Figure 2: Performance comparison by feature set using court
administrative database features. We compared the motion
to strike classification accuracy on minimal and subset fea-
ture sets for six ML models (adaBoost (ab), decision trees (dt),
gradient boosting (gb), random forests (rf), support vector
machines (svm), and XGBoost (xgb)). Box plots show the dis-
tribution of 100 bootstrapped samples with Tukey whiskers
(median ± 1.5 times interquartile range).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Database
Doc2vec
Motion Classification Accuracy
Word2vec
2
0.6
0
0.6
8
0.5
6
0.5
4
0.5
ab
Figure 3: Predicting motion to strike outcomes across court administrative data and complaint documents. Distinct classifiers
were trained on court administrative data and dense features computed from complaint documents. Document embedding
features based on doc2vec [48] (black) and word2vec [56, 53] (blue) largely improved the classification accuracy of motion
outcomes versus court administrative database features alone (red) for six classifiers [61]: AdaBoost (ab), decision trees (dt),
gradient boosting (gb), random forests (rf), support vector machines (svm), Xgboost (xgb). Box plots are drawn with Tukey
whiskers (median ± 1.5 times interquartile range).
for training and validation. To estimate variability in classification
accuracy, we computed 100 bootstrapped samples for each model
selected from our grid search.
First, we evaluated the minimal and subset feature sets (Table 1)
associated with legal cases to determine feature relevancy for mo-
tion outcome prediction. We varied the feature composition for the
court administrative data and compared classifier accuracy (Fig. 2).
Adding in the case location and major code features improved
median accuracy for all methods besides random forests. Overall,
decisions trees exhibited the highest motion to strike classification
accuracy (mean: 0.583, median: 0.583) with slightly higher perfor-
mance than boosting methods. Given these results, we primarily
focus our analysis on the subset featurization.
To assist interpretation of model performance, we compared
our ML models with a naive baseline. The naive classifier predicts
motion outcomes using the empirical frequency of the training
set; with 52% of the motions being granted, we observed a naive
baseline accuracy of 0.501. During model selection, we observed
a maximum classification accuracy of 0.644 using Adaboost with
dense word embeddings, corpus specific TF-IDF weightings, and
FOIL algorithmic rules. This same model had a mean accuracy of
0.605 over 100 bootstraps. However, the highest mean accuracy
score from the same group of features was found in decision tree
classifiers with 0.606.
Next, we evaluated whether court administrative data alone
was sufficient for learning accurate motion to strike classification
models. Using only database features, all methods produced classifi-
cation accuracies less than 0.60 (Fig. 3). Subsequently, we evaluated
these same methods, but including dense natural language fea-
tures extracted from complaint documents. While concatenating
doc2vec features to the database feature vectors improved classifi-
cation accuracy for most methods, a more careful model defined
over word2vec features using corpus specific weighting (TF-IDF)
McConnell et al.
dt gb rf svm xgb
improved classification accuracy for all models. These results high-
light the utility of incorporating natural language features in the
prediction of motion outcomes.
Interestingly, the difference between minimal and subset fea-
turization was diminished when including complaint document
features into the model. We observed this behavior when consider-
ing both doc2vec (Supp. Fig. 1) and word2vec (Supp. Fig. 2) features,
although word2vec features continued to yield better performing
models. These findings suggest that complaint document embed-
dings effectively capture major code and case location features.
We next investigated feature significance for motion to strike
outcome prediction for decision trees built on word2vec features.
We chose word2vec features because models built on word2vec
embeddings produced more accurate results than database-only or
database and doc2vec features; we selected decisions trees since
these are the most explainable of the six models tested with similar
performance to the boosting models. We found that features derived
from the complaint documents were universally important across
all decision tree classifiers (Fig. 4). The case type (Major Code) was
also an important feature for a subset of models, as was the attorney
specialization (entropy).
Next, we quantified whether tort or vehicular cases had subtle
differences that made prediction easier by stratifying classification
accuracy by case type in our best performing feature configura-
tion (database features, FOIL and TF-IDF weighted word2vec). All
methods, excluding SVM, predicted motion to strike outcomes in
vehicular cases with a significantly higher accuracy than tort cases
(one-way paired t-test, 𝑝 ≤ 2.2 × 10−16 ) (Fig. 5). This is likely due
to inherent properties of vehicular cases and not class imbalance
since vehicular cases encompassed approximately 47% of the total
cases.
Lastly, we investigated if legal domain-specific word2vec mod-
els would improve classifier performance compared to word2vec
104
Case-level Prediction of Motion Outcomes in Civil Litigation ICAIL’21, June 21–25, 2021, São Paulo, Brazil

1.0
dt (AP = 0.6274)
gb (AP = 0.5734)
0.8 ab (AP = 0.6186)
0.9 svm (AP = 0.6042)
ub

xgb (AP = 0.6273)

p,s p,m IL,m IL,s p,s p,m IL,m IL,s

rf (AP = 0.5661)
,si F,FO F,FO
in

0.8
-ID -ID
TF TF

Precision
0.6
in
m

0.7
ub
F
-ID
TF

i m
F,s
ub

0.4
-ID

0.6
TF
sim sim FO FO
in

0.5
in

0.0 0.2 0.4 0.6 0.8 1.0

0.2 Recall
ub
er

py

V
de

Lo T)

Lo HB)

Lo D)

Figure 6: Precision-Recall curves for law2vec classifier on

NL

W2
mb

(FS

tro
H
Co

(K
(H

(H

En
Nu

subset and TF-IDF weighted simple features. We trained and

jor

c.

c.
c.

c.
Lo
Ma
ris

0.0
Ju

tested with 100 bootstrapped samples the law2vec model us-

ing TF-IDF weighted simple and subset features. Precision
and recall were computed for several thresholds for six ML
models (adaBoost (ab), decision trees (dt), gradient boosting
Figure 4: Decision tree feature importance for word2vec
(gb), random forests (rf), support vector machines (svm), and
models. The vertical axis denotes the model configuration
XGBoost (xgb)). Average precision (AP) for each method is
from TF-IDF weighted, simple (simp) or FOIL algorithmic
given in the legend.
rules, and minimum (min) or subset (sub) feature sets. The
x-axis gives features with at least one model yielding a non-
zero importance score. The case major code, attorney spe-
cialization (entropy), and word2vec (W2V) features exhib- law2vec improved classifier performance. Using court adminis-
ited the highest average feature importance across models. trative data, FOIL, and TF-IDF weighted features, word2vec classi-
The word2vec feature importance weights were summed fier median performance was greater than or approximately equal
across all word2vec features. across all methods besides XGBoost (Supp. Fig. 3); the behavior
was similar when removing FOIL features (Supp. Fig. 4). While
tree and boosting classifiers outperformed our SVM model overall,
V when we compared the performance of each classifier by vary-
T ing the classification threshold, all methods had similar average
0.650
precision when including TF-IDF weighted features from the sim-
Motion Classification Accuracy

0.625
ple (Fig. 6) and FOIL (Supp. Fig. 5) rule based classifiers or when
including (Supp. Fig. 6) and excluding (Supp. Fig. 7) FOIL features
0.600

without TF-IDF weighting. However, SVMs have noticeably higher

0.575
0.550
average precision than their classification accuracy would suggest,
0.525
providing strong justification for SVMs when multiple classification
0.500 thresholds are considered.
0.475
ab dt gb rf svm xgb
5 DISCUSSION
In this work we focused on motions to strike, but the prediction
Figure 5: Classifier accuracy by case type. We stratified clas-
of other high impact motions may be of interest to lawyers and
sifier performance on motion to strike prediction by vehic-
their clients. For example, outcomes of interest may include the
ular or tort case types (major code). Here, we include classi-
file dates of a disposition, the cost or activity associated with the
fiers trained on court administrative and word2vec features.
discovery period (which is typically a function of complexity of
Box plots show the distribution of 100 bootstrapped sam-
motions during discovery), or motions for default judgement –
ples with Tukey whiskers (median ± 1.5 times interquartile
which can happen when a party does not respond to a court filing.
range).
A motion for summary judgement is another high-impact motion
which occurs with high frequency; this motion is a written petition
which argues that, based on the evidence gathered thus far, no
trained on general corpora [42]. Overall, we did not find that reasonable jury could find that the plaintiff has proved one or

105
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
more genuinely-disputed facts that matter to one or more of the
plaintiff’s legal claim. However, careful consideration must be taken
when deciding which features or court documents are relevant. For
example, a motion for summary judgement depends on the quality
of the evidence collected by the parties thus far, which may not be
well captured by the text of court documents.
Our comparisons between doc2vec and law2vec suggested there
are potential benefits to incorporating neural embeddings from
models trained specifically on legal corpora. In modern transformer
architectures, domain specific pre-trained models, e.g. BioBERT [51],
have been shown to outperform transfer learning fine-tuning ap-
proaches from general corpora like Wikipedia. While some of these
models are in the legal domain, e.g. patentBERT [50], no such model
exists for court documents. However, one significant challenge
to applying such models in the legal domain is that the memory
requirements for transformers scale quadratically with sequence
length, prohibiting them from being applied directly to the longer
texts that are common in complaints and judicial opinions [6]. Re-
cent work on extending the range of transformers provides some
evidence that this issue will be addressed [79, 76].
A limitation of the data in our analysis is that courts may occa-
sionally grant a motion to strike in part. This can occur, for example,
if two or more legal claims levied against a defendant were formally
challenged by a motion to strike. The court may grant the motion
to strike with respect to a single legal claim and deny it for others.
In these cases, the motion to strike order code in the Law Library
data is noisy since there is only a single value that is provided.
Furthermore, the procedure with which courts interpret this order
code is heterogeneous. Some courts interpret a granted motion as
any motion granted in part or in full; other courts only use the
granted order code when the motion is granted in full. Addressing
this issue either requires reforming and unifying the data process-
ing procedures across Connecticut courts or developing methods
to parse out distinct legal claims from complaint documents and
then matching them to judicial order documents.
The relevant features for the motion prediction problem can
also likely be improved. For example, an attorney is defined based
on their juris number and a derived attorney specialization fea-
ture. There are other features that are likely relevant for predicting
motion outcomes, e.g., attorney experience, record, or case load.
Similar feature engineering can be implemented for judges. Case
location and other high dimensional categorical features can be
one-hot encoded, but may also benefit from a descriptive, lower
dimensional set of features based on, e.g., court culture.
With the rise of ML in the legal domain, governments across the
globe are placing new emphases on ensuring AI-assisted decision
making is done in an ethical, transparent, and nondiscriminatory
manner. The European Commission for the Efficiency of Justice
adopted 5 principles in the European Ethical Charter on the use of
AI in judicial systems [82]. These principles guarantee that judi-
cial AI is compatible with fundamental rights, nondiscriminatory,
transparent, impartial, fair, and explainable. In the U.S., the Na-
tional Center for State Courts has identified data transparency
and investigating how AI transforms judicial processes as national
priorities [59].
Fundamental to upholding these ideals is developing methods
that are interpretable. Interpretability provides a mechanism to
106
McConnell et al.
ensure ethical and legal compliance in situations where the ethics
or legal implications are ill-defined or cannot be mathematically
modelled. Model-based interpretability focuses on restricting the set
of models such that a trained model directly informs relationships
among model variables [58]. The methods developed in this work
were selected, in part, due to high model-based interpretability (e.g.
decision trees), but, incorporating causal reasoning and explicit
Bayesian modelling of relevant variables in the judicial decision
making process would only increase explainability, interpretability,
and ultimately, trustworthiness of the model. One possibility is to
consider the recent work in causal frameworks for decision trees [33,
52, 78].
6 CONCLUSIONS
By developing ML workflows with feature engineering rooted in
legal domain expertise, we developed methods to help researchers
better understand the predictability of trial motions and practition-
ers the ability to make more informed decisions. We developed and
benchmarked the first ML methods to predict motion outcomes
using only data that is available to all parties at trial. Our work
demonstrated that motion to strike outcomes are predictable with
high accuracy when new features like attorney specialization are
combined with complaint document embeddings.
We expect these methods will be a valuable resource for lawyers
and their clients by enabling the estimation of case strength. For
example, our methods can be used to predict if the case will sur-
vive a motion to strike after a complaint document is filed. Based
on predicted motion outcomes, both parties can make more in-
formed settlement decisions and lawyers representing the plaintiff
can revise the language in their complaint documents. Fitted ML
models, training code, and benchmarking code can be accessed at
https://github.com/aguiarlab/motionpredict.
REFERENCES
[1] Charu C Aggarwal. 2018. Machine learning for text. Springer.
[2] Sharan Agrawal et al. 2017. Affirm or reverse? using machine
learning to help judges write opinions. NBER Working Paper,
29.
[3] Benjamin Alarie et al. 2016. Using machine learning to pre-
dict outcomes in tax law. Can. Bus. LJ, 58, 231.
[4] Nikolaos Aletras et al. 2016. Predicting judicial decisions
of the european court of human rights: a natural language
processing perspective. PeerJ Computer Science, 2, e93.
[5] Katie Atkinson et al. 2020. Explanation in AI and law: past,
present and future. Artificial Intelligence, 103387.
[6] Iz Beltagy et al. 2020. Longformer: the long-document trans-
former. arXiv preprint arXiv:2004.05150.
[7] Candice Bentejac et al. 2020. A comparative analysis of gra-
dient boosting algorithms. Artificial Intelligence Review, 1–
31.
[8] Ryan C Black and James F Spriggs. 2008. An empirical anal-
ysis of the length of US Supreme Court opinions. Hous. L.
Rev., 45, 621.
[9] David M Blei et al. 2003. Latent dirichlet allocation. the Jour-
nal of Machine Learning Research, 3, 993–1022.
Case-level Prediction of Motion Outcomes in Civil Litigation
[10] Bernhard E Boser et al. 1992. A training algorithm for optimal
margin classifiers. In Proceedings of the fifth annual workshop
on Computational learning theory, 144–152.
[11] Leo Breiman. 2001. Random forests. Machine learning, 45, 1,
5–32.
[12] Shelagh MR Campbell. 2017. Exercising discretion in the
context of dependent employment: assessing the impact of
workload on the rule of law. Legal Studies, 37, 2, 305–323.
[13] John Celona. 2016. Winning at Litigation through Decision
Analysis: Creating and Executing Winning Strategies in any
Litigation or Dispute. Springer Series in Operations Research
and Financial Engineering. Springer.
[14] Ilias Chalkidis. 2018. Law2Vec: Legal Word Embeddings.
(2018). https://archive.org/details/Law2Vec.
[15] Tianqi Chen and Carlos Guestrin. 2016. Xgboost: a scalable
tree boosting system. In Proceedings of the 22nd acm sigkdd
international conference on knowledge discovery and data
mining, 785–794.
[16] Cary Coglianese. 2004. E-rulemaking: information technol-
ogy and the regulatory process. Admin. L. Rev., 56, 353.
[17] William W Cohen et al. 1996. Learning rules that classify
e-mail. In AAAI spring symposium on machine learning in
information access. Volume 18. Stanford, CA, 25.
[18] William W Cohen and Yoram Singer. 1999. Context-sensitive
learning methods for text categorization. ACM Transactions
on Information Systems (TOIS), 17, 2, 141–173.
[19] Corinna Cortes and Vladimir Vapnik. 1995. Support-vector
networks. Machine learning, 20, 3, 273–297.
[20] Lindsey Devers. 2011. Plea and charge bargaining. Research
summary for Bureau of Justice Assistance, U.S. Department of
Justice, 1.
[21] Jacob Devlin et al. 2018. Bert: pre-training of deep bidi-
rectional transformers for language understanding. arXiv
preprint arXiv:1810.04805.
[22] Michael Evans et al. 2007. Recounting the courts? apply-
ing automated content analysis to enhance empirical legal
research. Journal of Empirical Legal Studies, 4, 4, 1007–1039.
[23] Frank Fagan and Saul Levmore. 2019. The impact of artificial
intelligence on rules, standards, and judicial discretion. S.
Cal. L. Rev., 93, 1.
[24] Felix Steffek. 2021. Law and Autonomous Systems Series:
Paving the Way for Legal Artificial Intelligence – A Common
Dataset for Case Outcome Predictions. University of Oxford.
(2021). https://www.law.ox.ac.uk/business-law-blog/blog/
2018 / 05 / law - and - autonomous - systems - series - paving -
way-legal-artificial.
[25] Norman Fenton et al. 2016. Bayes and the law. Annual Review
of Statistics and Its Application, 3, 1, 51–77. https://doi.org/
10.1146/annurev-statistics-041715-033428.
[26] Matthias Feurer and Frank Hutter. 2019. Hyperparameter op-
timization. In Automated Machine Learning. Springer, Cham,
3–33.
[27] James H Fowler et al. 2007. Network analysis and the law:
measuring the legal importance of precedents at the us
supreme court. Political Analysis, 324–346.
[28] Jens Frankenreiter and Michael A. Livermore. 2020. Compu-
tational methods in legal analysis. Annual Review of Law and
107
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Social Science, 16, 1, 39–57. https://doi.org/10.1146/annurev-
lawsocsci-052720-121843.
[29] Yoav Freund and Robert E Schapire. 1997. A decision-theoretic
generalization of on-line learning and an application to
boosting. Journal of computer and system sciences, 55, 1, 119–
139.
[30] Yoav Freund et al. 1999. A short introduction to boosting.
Journal-Japanese Society For Artificial Intelligence, 14, 771-
780, 1612.
[31] Jerome H Friedman. 2002. Stochastic gradient boosting. Com-
putational statistics & data analysis, 38, 4, 367–378.
[32] Anne von der Lieth Gardner. 1984. Artificial intelligence
approach to legal reasoning. Technical report. Stanford Univ.
[33] Tim Genewein et al. 2020. Algorithms for Causal Reasoning
in Probability Trees. en. arXiv:2010.12237 [cs], (November
2020). arXiv: 2010.12237. Retrieved 12/10/2020 from http :
//arxiv.org/abs/2010.12237.
[34] Jane Goodman-Delahunty et al. 2010. Insightful or wishful:
lawyers’ ability to predict case outcomes. Psychology, Public
Policy, and Law, 16, 2, 133–157.
[35] Evan Gretok et al. 2020. Transformers for classifying fourth
amendment elements and factors tests. In Legal Knowledge
and Information Systems: JURIX 2020: The Thirty-third An-
nual Conference, Brno, Czech Republic, December 9-11, 2020.
Volume 334. IOS Press, 63–72.
[36] Thomas Hancock et al. 1996. Lower bounds on learning
decision lists and trees. Information and Computation, 126, 2,
114–122.
[37] Allison P. Harris and Maya Sen. 2019. Bias and judging. An-
nual Review of Political Science, 22, 1, 241–259. https://doi.
org/10.1146/annurev-polisci-051617-090650.
[38] Oliver Wendell Holmes. 1897. The path of the law. Harvard
Law Review, 10, 8, 457–478.
[39] Jonas Jacobson et al. 2011. Predicting civil jury verdicts: how
attorneys use (and misuse) a second opinion. Journal of Em-
pirical Legal Studies, 8, S1, 99–119. http://dx.doi.org/10.1111/
j.1740-1461.2011.01229.x.
[40] Robert A Karasek Jr. 1979. Job demands, job decision latitude,
and mental strain: implications for job redesign. Administra-
tive science quarterly, 285–308.
[41] Daniel Martin Katz et al. 2017. A general approach for pre-
dicting the behavior of the Supreme Court of the United
States. PLOS ONE, 12, 4, (April 2017), 1–18. https://doi.org/
10.1371/journal.pone.0174698.
[42] Nari Kim and Hyoung Joong Kim. 2017. A study on the
law2vec model for searching related law. Journal of Digital
Contents Society, 18, 7, 1419–1425.
[43] Sang-Bum Kim et al. 2006. Some effective techniques for
naive Bayes text classification. IEEE transactions on knowl-
edge and data engineering, 18, 11, 1457–1466.
[44] Alex Krizhevsky, Geoffrey Hinton, et al. 2009. Learning mul-
tiple layers of features from tiny images.
[45] Jason T Lam et al. 2020. The gap between deep learning and
law: predicting employment notice. NLLP KDD, 7, 10.
[46] Hyafil Laurent and Ronald L Rivest. 1976. Constructing opti-
mal binary decision trees is NP-complete. Information pro-
cessing letters, 5, 1, 15–17.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
[47] David S Law and David Zaring. 2009. Law Versus Ideology:
The Supreme Court and the Use of Legislative History. Wm.
& Mary L. Rev., 51, 1653.
[48] Quoc Le and Tomas Mikolov. 2014. Distributed representa-
tions of sentences and documents. In International conference
on machine learning, 1188–1196.
[49] Yann LeCun and Corinna Cortes. 2010. MNIST handwritten
digit database. http://yann.lecun.com/exdb/mnist/.
[50] Jieh-Sheng Lee and Jieh Hsiang. 2019. Patentbert: patent clas-
sification with fine-tuning a pre-trained bert model. arXiv
preprint arXiv:1906.02124.
[51] Jinhyuk Lee et al. 2020. Biobert: a pre-trained biomedical
language representation model for biomedical text mining.
Bioinformatics, 36, 4, 1234–1240.
[52] Jiuyong Li et al. 2016. Causal decision trees. IEEE Transactions
on Knowledge and Data Engineering, 29, 2, 257–271.
[53] Joseph Lilleberg et al. 2015. Support vector machines and
word2vec for text classification with semantic features. In
2015 IEEE 14th International Conference on Cognitive Infor-
matics & Cognitive Computing (ICCI* CC). IEEE, 136–140.
[54] Masha Medvedeva et al. 2020. Using machine learning to
predict decisions of the European Court of Human Rights.
Artificial Intelligence and Law, 28, 2, 237–266.
[55] David E. Melnikoff and Nina Strohminger. 2020. The auto-
matic influence of advocacy on lawyers and novices. Nature
Human Behaviour, (September 7, 2020), 1–7.
[56] Tomas Mikolov et al. 2013. Efficient estimation of word rep-
resentations in vector space. (2013). http://arxiv.org/abs/
1301.3781.
[57] Jane Mitchell et al. 2020. Machine learning for determining
accurate outcomes in criminal trials. Law, Probability and
Risk, 19, 1, (March 2020), 43–65.
[58] W James Murdoch et al. 2019. Definitions, methods, and
applications in interpretable machine learning. Proceedings
of the National Academy of Sciences, 116, 44, 22071–22080.
[59] National Center for State Courts. 2021. Joint technology com-
mittee priority topics. Accessed on 2021-03-01. (2021). https:
//www.ncsc.org/about-us/committees/joint-technology-
committee/priority-topics-old-page.
[60] Patrick W Nutter. 2018. Machine learning evidence: admissi-
bility and weight. U. Pa. J. Const. L., 21, 919.
[61] F. Pedregosa et al. 2011. Scikit-learn: machine learning in
Python. Journal of Machine Learning Research, 12, 2825–2830.
[62] Alina Petrova et al. 2020. Extracting Outcomes from Ap-
pellate Decisions in US State Courts. In Legal Knowledge
and Information Systems: JURIX 2020: The Thirty-third An-
nual Conference, Brno, Czech Republic, December 9-11, 2020.
Volume 334. IOS Press, 133–142.
[63] Arti K Rai. 2018. Machine learning at the patent office: lessons
for patents and administrative law. Iowa L. Rev., 104, 2617.
[64] Edwina L Rissland. 1990. Artificial intelligence and law: step-
ping stones to a model of legal reasoning. The Yale Law
Journal, 99, 8, 1957–1981.
[65] Theodore W Ruger et al. 2004. The supreme court forecasting
project: legal and political science approaches to predicting
supreme court decision making. Columbia Law Review, 1150–
1210.
108
McConnell et al.
[66] S. R. Safavian and D. Landgrebe. 1991. A survey of decision
tree classifier methodology. IEEE Transactions on Systems,
Man, and Cybernetics, 21, 3, 660–674.
[67] Robert E Schapire. 2013. Explaining adaboost. In Empirical
inference. Springer, 37–52.
[68] Marek J. Sergot et al. 1986. The British Nationality Act as a
logic program. Communications of the ACM, 29, 5, 370–386.
[69] Ray Smith. 2007. An overview of the tesseract ocr engine.
In Ninth international conference on document analysis and
recognition (ICDAR 2007). Volume 2. IEEE, 629–633.
[70] Harold Spaeth et al. 2014. Supreme court database code book.
(2014).
[71] State of Connecticut Judicial Branch. 2021. Public Records
Online. Accessed on 2021-01-01. (2021). https://jud.ct.gov/
lawlib/publicrecords.htm.
[72] Michael Still. 2006. The definitive guide to ImageMagick. Apress.
[73] Octavia-Maria Şulea et al. 2017. Predicting the law area and
decisions of French Supreme Court cases. In Proceedings
of the International Conference Recent Advances in Natural
Language Processing, RANLP 2017. INCOMA Ltd., Varna, Bul-
garia, (September 2017), 716–722. https://doi.org/10.26615/
978-954-452-049-6_092.
[74] Harry Surden. 2014. Machine learning and law. Wash. L. Rev.,
89, 87.
[75] Ahmad P Tafti et al. 2016. OCR as a service: an experi-
mental evaluation of Google Docs OCR, Tesseract, ABBYY
FineReader, and Transym. In International Symposium on
Visual Computing. Springer, 735–746.
[76] Yi Tay et al. 2020. Long Range Arena: A Benchmark for
Efficient Transformers. en. arXiv:2011.04006 [cs], (November
2020). arXiv: 2011.04006. Retrieved 12/12/2020 from http :
//arxiv.org/abs/2011.04006.
[77] Thomas Vacek et al. 2019. Litigation Analytics: Case out-
comes extracted from US federal court dockets. In Proceed-
ings of the Natural Legal Language Processing Workshop 2019,
45–54.
[78] Stefan Wager and Susan Athey. 2018. Estimation and in-
ference of heterogeneous treatment effects using random
forests. Journal of the American Statistical Association, 113,
523, 1228–1242.
[79] Sinong Wang et al. 2020. Linformer: Self-Attention with
Linear Complexity. en. arXiv:2006.04768 [cs, stat], (June 2020).
arXiv: 2006.04768. Retrieved 12/12/2020 from http://arxiv.
org/abs/2006.04768.
[80] Ho Chung Wu et al. 2008. Interpreting tf-idf term weights
as making relevance decisions. ACM Transactions on Infor-
mation Systems (TOIS), 26, 3, 1–37.
[81] Feiyu Xu et al. 2019. Explainable ai: a brief survey on history,
research areas, approaches and challenges. In CCF interna-
tional conference on natural language processing and Chinese
computing. Springer, 563–574.
[82] Irina Moroianu Zlatescu and Petru Emanuel Zlatescu. 2019.
Implementation of the European ethical charter on the use
of artificial intelligence in judicial systems and their envi-
ronment. Current Issues of the EU Political-Legal Space, 237.
 Evaluating Document Representations for Content-based Legal
Literature Recommendations
Malte Ostendorff Elliott Ash
Open Legal Data ETH Zurich
Germany Switzerland
mo@openlegaldata.io ashe@ethz.ch
Bela Gipp Julian Moreno-Schneider
University of Wuppertal DFKI GmbH
Germany Germany
gipp@uni-wuppertal.de julian.moreno_schneider@dfki.de
ABSTRACT
Recommender systems assist legal professionals in finding relevant
literature for supporting their case. Despite its importance for the
profession, legal applications do not reflect the latest advances in
recommender systems and representation learning research. Si-
multaneously, legal recommender systems are typically evaluated
in small-scale user study without any public available benchmark
datasets. Thus, these studies have limited reproducibility. To address
the gap between research and practice, we explore a set of state-of-
the-art document representation methods for the task of retrieving
semantically related US case law. We evaluate text-based (e.g., fast-
Text, Transformers), citation-based (e.g., DeepWalk, Poincaré), and
hybrid methods. We compare in total 27 methods using two silver
standards with annotations for 2,964 documents. The silver stan-
dards are newly created from Open Case Book and Wikisource and
can be reused under an open license facilitating reproducibility.
Our experiments show that document representations from aver-
aged fastText word vectors (trained on legal corpora) yield the best
results, closely followed by Poincaré citation embeddings. Com-
bining fastText and Poincaré in a hybrid manner further improves
the overall result. Besides the overall performance, we analyze the
methods depending on document length, citation count, and the
coverage of their recommendations.
CCS CONCEPTS
• Information systems → Recommender systems; Similarity mea-
sures; Clustering and classification; • Applied computing → Law.
KEYWORDS
Legal literature, document embeddings, document similarity, rec-
ommender systems, Transformers, WikiSource, Open Case Book
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466073
109
Terry Ruas
University of Wuppertal
Germany
ruas@uni-wuppertal.de
Georg Rehm
DFKI GmbH
Germany
georg.rehm@dfki.de
ACM Reference Format:
Malte Ostendorff, Elliott Ash, Terry Ruas, Bela Gipp, Julian Moreno-Schneider,
and Georg Rehm. 2021. Evaluating Document Representations for Content-
based Legal Literature Recommendations. In Eighteenth International Con-
ference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São
Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/
3462757.3466073
1 INTRODUCTION
Legal professionals, e.g., lawyers and judges, frequently invest con-
siderable time to find relevant literature [23]. More so than most
other domains, in law there are high stakes for finding the most
relevant information (documents) as that can drastically affect the
outcome of a dispute. A case can be won or lost depending on
whether or not a supporting decision can be found. Recommender
systems assist in the search for relevant information. However,
research and development of recommender systems for legal cor-
pora poses several challenges. Recommender system research is
known to be domain-specific, i.e., minor changes may lead to unpre-
dictable variations in the recommendation effectiveness [4]. Like-
wise, legal English is a peculiarly obscure and convoluted variety of
English with a widespread use of common words with uncommon
meanings [30]. Recent language models like BERT [15] may not
be equipped to handle legal English since they are pretrained on
generic corpora like Wikipedia or cannot process lengthy legal doc-
uments due to their limited input length. This raises the question
of whether the recent advances in recommender system research
and underlying techniques are also applicable to law.
In this paper, we empirically evaluate 27 document representa-
tion methods and analyze the results with respect to the aforemen-
tioned possible issues. In particular, we evaluate for each method
the quality of the document representations in a literature recom-
mender use case. The methods are distinguished in three categories:
(1) word vector-based, (2) Transformer-based, and (3) citation-based
methods. Moreover, we test additional hybrid variations of the afore-
mentioned methods. Our primary evaluation metric comes from
two silver standards on US case law that we extract from Open
Case Book and Wikisource. The relevance annotations from the
silver standards are provided for 2,964 documents.
In summary, our contributions are: (1) We propose and make
available two silver standards as benchmarks for legal recommender
systems that currently do not exist. (2) We evaluate 27 methods of
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
which the majority have never been investigated in the legal context
with a quantitative study and validate our results qualitatively. (3)
We show that the hybrid combination of text-based and citation-
based methods can further improve the experimental results.
2 RELATED WORK
Recommender systems are a well-established research field [3] but
relatively few publications focus on law as the application domain.
Winkels et al. [52] are among the first to present a content-based
approach to recommend legislation and case law. Their system
uses the citation graph of Dutch Immigration Law and is evaluated
with a user study conducted with three participants. Boer and
Winkels [9] propose and evaluate Latent Dirichlet Allocation (LDA)
[7] as a solution to the cold start problem in collaborative filtering
approaches. In an experiment with 28 users, they find the user-
based approach outperforms LDA. Wiggers and Verberne [49] study
citations for legal information retrieval and suggest citations should
be combined with other techniques to improve the performance.
Kumar et al. [22] compare four methods to measure the similar-
ity of Indian Supreme Court decision: TF-IDF [43] on all document
terms, TF-IDF on only specific terms from a legal dictionary, Co-
Citation, and Bibliographic Coupling. They evaluate the similarity
measure on 50 document pairs with five legal domain experts. In
their experiment, Bibliographic Coupling and TF-IDF on legal terms
yield the best results. Mandal et al. [28] extend this work by evaluat-
ing LDA and document embeddings (Paragraph Vectors [25]) on the
same dataset, whereby Paragraph Vectors was found to correlate
the most with the expert annotations. Indian Supreme Court deci-
sions are also used as evaluation by Wagh and Anand [47], where
they use document similarity based on concepts instead of full-text.
They extract concepts (groups of words) from the decisions and
compute the similarity between documents based on these concepts.
Their vector representation, an average of word embeddings and
TF-IDF, shows IDF for weighting word2vec embeddings improve
results. Also, Bhattacharya et al. [6] compare citation similarity
methods, i.e., Bibliographic Coupling, Co-citation, Dispersion [32]
and Node2Vec [17]), and text similarity methods like Paragraph Vec-
tors. They evaluate the algorithms and their combinations using a
gold standard of 47 document pairs. A combination of Bibliographic
Coupling and Paragraph Vectors achieves the best results.
With Eunomos, Boella et al. [8] present a legal document and
knowledge management system for searching legal documents.
The document similarity problem is handled using TF-IDF and co-
sine similarity. Other experiments using embeddings for document
similarity include Nanda et al. [33] or Ash and Chen [2].
Even though different methods have been evaluated in the legal
domain, most results are not coherent and rely on small-scale user
studies. This finding emphasizes the need for a standard bench-
mark to enable reproducibility and comparability [4]. Moreover,
the recent Transformer models [46] or novel citation embeddings
have not been evaluated in legal recommendation research.
3 METHODOLOGY
In this section, we describe our quantitative evaluation of 27 docu-
ment recommendations methods. We define the recommendation
scenario as follows: The user, a legal professional, needs to research
110
Ostendorff et al.
a particular decision, e.g., to prepare a litigation strategy. Based on
the decision at hand, the system recommends other decisions to its
users such that the research task is easy to accomplish. The recom-
mendation is relevant when it covers the same topic or provides
essential information, e.g., it overruled the seed decision [45].
3.1 Case Corpus and Silver Standard
Most of the previous works (Section 2) evaluate recommendation
relevance by asking domain experts to provide subjective annota-
tions [9, 22, 28, 52]. Especially in the legal domain, these expert
annotations are costly to collect and, therefore, their quantity is
limited. For the same reason, expert annotations are rarely pub-
lished. Consequently, the research is difficult to reproduce [4]. In
the case of the US court decisions, such expert annotations between
documents are also not publicly available. We construct two ground
truth datasets from publicly available resources allowing the evalua-
tion of more recommendations to mitigate the mentioned problems
of cost, quantity, and reproducibility.
3.1.1 Open Case Book. With Open Case Book, the Harvard Law
School Library offers a platform for making and sharing open-
licensed casebooks 1 . The corpus consists of 222 casebooks contain-
ing 3,023 cases from 87 authors. Each casebook contains a manually
curated set of topically related court decisions, which we use as
relevance annotations. The casebooks cover a range from broad
topics (e.g., Constitutional law) to specific ones (e.g., Intermediary
Liability and Platforms’ Regulation). The decisions are mapped to
full-texts and citations retrieved from the Caselaw Access Project
(CAP)2 . After duplicate removal and the mapping procedure, rele-
vance annotations for 1,601 decisions remain.
3.1.2 Wikisource. We use a collection of 2,939 US Supreme Court
decisions from Wikisource as ground truth [50]. The collection is
categorized in 67 topics like antitrust, civil rights, and amendments.
We map the decisions listed in Wikisource to the corpus from
CourtListener3 . The discrepancy between the two corpora decreases
the number of relevance annotations to 1,363 court decisions.
Table 1: Distribution of relevant annotations for Open Case
Book and Wikisource.
Relevant annotations per document
Mean Std. Min. 25% 50% 75% Max.
Open Case Book 86.42 65.18 2.0 48.0 83.0 111.0 1590.0
Wikisource 130.01 82.46 1.0 88.0 113.0 194.0 616.0
We derive a binary relevance classification from Open Case Book
and Wikisource. When decisions A and B are in the same casebook
or category, A is relevant for B and vice versa. Table 1 presents the
distribution of relevance annotations. This relevance classification
is limited since a recommendation might still be relevant despite
not being assigned to the same topic as the seed decision. Thus,
we consider the Open Case Book and Wikisource annotations as a
silver standard rather than a gold one.
1 https://opencasebook.org
2 https://case.law
3 https://courtlistner.com
Evaluating Document Representations for Content-based Legal Literature Recommendations
3.2 Evaluated Methods
We evaluate 27 methods, each representing legal document 𝑑 as a
numerical vector 𝑑® ∈ R𝑠 , with 𝑠 denoting the vector size. To retrieve
the recommendations, we first obtain the vector representations (or
document embeddings). Next, we compute the cosine similarities
of the vectors. Finally, we select the top 𝑘 = 5 documents with the
highest similarity through nearest neighbor search4 . Mean Average
Precision (MAP) is the primary and Mean Reciprocal Rank (MRR) is
the second evaluation metric [29]. We compute MAP and MRR over
a set of queries 𝑄, whereby 𝑄 is equivalent to the seed decisions
with |𝑄 WS | = 1363 available in Wikisource and |𝑄 OCB | = 1601 for
Open Case Book. In addition to the accuracy-oriented metrics, we
evaluate the coverage and Jaccard index of the recommendations.
The coverage for the method 𝑎 is defined as in Equation 1 where
𝐷 denotes the set of all available documents in the corpus and 𝐷𝑎
denotes the recommended documents by 𝑎 [16].
|𝐷𝑎 |
𝐶𝑜𝑣 (𝑎) = (1)
|𝐷 |
We define the Jaccard index [19] for the similarity and diversity
of two recommendation sets 𝑅𝑎 and 𝑅𝑏 from methods 𝑎 and 𝑏 for
the seed 𝑑𝑠 in Equation 2:
|𝑅𝑎 ∩ 𝑅𝑏 |
𝐽 (𝑎, 𝑏) = (2)
|𝑅𝑎 ∪ 𝑅𝑏 |
We divide the evaluated methods into three categories: Word
vector-, Transformer-, and citation-based methods.
3.2.1 TF-IDF Baseline. As a baseline method, we use the sparse
document vectors from TF-IDF [43], which are commonly used in
related works [22, 33]5 .
3.2.2 Word vector-based Methods. The following methods are de-
rived from word vectors, i.e., context-free word representations.
Paragraph Vectors [25] extend the idea of word2vec [31] to learn-
ing embeddings for word sequences of arbitrary length. Paragraph
Vectors using distributed bag-of-words (dbow) performed well in
text similarity tasks applied on legal documents [2, 28] and other
domains [24]. We train Paragraph Vectors’ dbow model to gen-
erate document vectors for each court decision. Like word2vec,
GloVe [38] and fastText [10, 20] produce dense word vectors but
they do not provide document vectors. To embed a court decision
as a vector, we compute the weighted average over its word vectors,
𝑤®𝑖 , whereby the number of occurrences of the word 𝑖 in 𝑑 defines
the weight 𝑐𝑖 . Averaging of word vectors is computationally effec-
tive and yields good results for representing even longer documents
[1]. For our experiments, we use word vectors made available by
the corresponding authors and custom word vectors. While GloVe
vectors are pretrained on Wikipedia and Gigaword [38], fastText
is pretrained on Wikipedia, UMBC webbase corpus and statmt.org
news dataset [10]. Additionally, we use custom word vectors6 for
both methods (namely fastTextLegal and GloVeLegal ) pretrained
on the joint court decision corpus extracted from Open Case Book
and Wikisource (see Section 3.1). Using word vectors pretrained
4 We
set 𝑘 = 5 due to the UI of our legal recommender system [35].
5 We use the TF-IDF implementation from the scikit-learn framework [37].
6 The legal word vectors can be downloaded from our GitHub repository.
111
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
on different corpora, allows the evaluation of the method’s cross-
domain applicability.
3.2.3 Transformer-based Methods. As the second method cate-
gory, we employ language models for deep contextual text rep-
resentations based on the Transformer architecture [46], namely
BERT [15], RoBERTa [27], Sentence Transformers (Sentence-
BERT and Sentence-RoBERTa) [41], LongFormer [5] and vari-
ations of them. In contrast to Paragraph Vectors and average word
vectors, which neglect the word order, the Transformers incorporate
word positions making the text representations context-dependent.
BERT significantly improved the state-of-the-art for many NLP
tasks. In general, BERT models are pretrained on large text corpora
in an unsupervised fashion to then be fine-tuned for specific tasks
like document classification [36]. We use four variations of BERT.
The original BERT [15] as base and large version (pretrained on
Wikipedia and BookCorpus) and two BERT-base models pretrained
on legal corpora. Legal-JHU-BERT-base from Holzenberger et al.
[18] which is a BERT base model but fine-tuned on the CAP corpus.
Similarly, Legal-AUEB-BERT-base from Chalkidis et al. [14] is
as well fine-tuned on the CAP corpus but also on other corpora
(court cases and legislation from the US and EU, and US contracts).
RoBERTa improves BERT with longer training, larger batches, and
removal of the next sentence prediction task for pretraining. Sen-
tence Transformers are fine-tuned BERT and RoBERTa models in
a Siamese setting [12] to derive semantically meaningful sentence
embeddings that can be compared using cosine similarity (Sentence-
BERT and Sentence-RoBERTa). The provided Sentence Transform-
ers variations are nli- or stsb-version that are either fine-tuned on
the SNLI and MNLI dataset [11, 51] or fine-tuned on the STS bench-
mark [13]. As the self-attention mechanism scales quadratically
with the sequence length, the Transfomer-based methods (BERT,
RoBERTa and Sentence Transformers) bound their representation
to 512 tokens. Longformer includes an attention mechanism that
scales linearly with sequence length, which allows to process longer
documents. We use pretrained Longformer models as provided by
Beltagy et al. [5] and limited to 4096 tokens. All Transformer models
apply mean-pooling to derive document vectors. We experimented
with other pooling strategies but they yield significantly lower re-
sults. These findings agree with Reimers and Gurevych [41]. We
investigate each Transformer in two variations depending on their
availability and w.r.t. model size and document vector size (base
with 𝑠 = 768 and large with 𝑠 = 1024).
3.2.4 Citation-based Methods. We explore citation-based graph
methods in which documents are nodes and edges correspond
to citations to generate document vectors. Like text-based repre-
sentations, citation graph embeddings have the vector size 𝑑® ∈
R300 . With DeepWalk, Perozzi et al. [39] were the first to borrow
word2vec’s idea and applied it to graph network embeddings. Deep-
Walk performs truncated random walks on a graph and the node em-
beddings are learned through the node context information encoded
in these short random walks similar to the context sliding window
in word2vec. Walklets [40] explicitly encodes multi-scale node
relationships to capture community structures with the graph em-
bedding. Walklets generates these multi-scale relationships by sub-
sampling short random walks on the graph nodes. BoostNE [26] is
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
a matrix factorization-based embedding technique combined with
gradient boosting. In [26], BoostNE is applied on a citation graph
from scientific papers and outperforms other graph embeddings
such as DeepWalk. Hence, we expect comparable results for the
legal citation graph. Nickel and Kiela [34] introduced Poincaré
embeddings for learning embeddings in the hyperbolic space of
the Poincaré ball model rather than the Euclidean space used in
the aforementioned methods. Embeddings produced in hyperbolic
space are naturally equipped to model hierarchical structures [21].
Such structures can also be found in the legal citation graph in the
form of different topics or jurisdictions. For DeepWalk, Walklets,
BoostNe, we use the Karate Club implementation [42].
3.2.5 Variations & Hybrid Methods. Given the conceptional differ-
ences in the evaluated methods, each method has its strength and
weakness. For further insights on these differences, we evaluate
all methods with limited text, vector concatenation, and score
summation: Unlike the Transformers, the word vector-based meth-
ods have no maximum of input tokens. Whether an artificial limi-
tation of the document length improves or decreases the results is
unclear. Longer documents might add additional noise to the repre-
sentation and could lead to worse results [44]. To make these two
method categories comparable, we include additional variations of
the word vector-based methods that are limited to the first 512 or
4096 tokens of the document. For instance, the method fastTextLegal
(512) has only access to the first 512 tokens.
Additionally, we explore hybrid methods that utilize texts and
citations. Each of the single methods above yields a vector rep-
resentation 𝑑® for a given document 𝑑. We combine methods by
concatenating their vectors. For example, the vectors from fastText
𝑑®fastText and Poincaré 𝑑®Poincaré can be concatenated as in Equation 3:
𝑑® = 𝑑®fastText ||𝑑®Poincaré
The resulting vector size is the sum of the concatenated vector
sizes, e.g., 𝑠 = 300 + 300 = 600. Recommendations based on the
concatenated methods are retrieved in the same fashion as the other
methods, with cosine similarity. Moreover, we combine methods by
adding up their cosine similarities [48]. The combined score of two
methods is the sum of the individual scores, e.g., for method X and
method Y the similarity of two documents 𝑑𝑎 and 𝑑𝑏 is computed
as in Equation 4. Methods with score summation are denoted with
𝑋 + 𝑌 , e.g., Poincaré + fastTextLegal .
𝑠𝑖𝑚(𝑑®𝑎 , 𝑑®𝑏 ) = 𝑠𝑖𝑚(𝑑®X𝑎 , 𝑑®X𝑏 ) + 𝑠𝑖𝑚(𝑑®Y𝑎 , 𝑑®Y𝑏 )
Lastly, we integrate citation information into Sentence Trans-
formers analog to the fine-tuning procedure proposed by Reimers
and Gurevych [41]. Based on the citation graph, we construct a
dataset of positive and negative document pairs. Two documents
𝑑𝑎 , 𝑑𝑏 are considered as positive samples when they are connected
through a citation. Negative pairs are randomly sampled and do
not share any citation. Sentence-Legal-AUEB-BERT-base is the
Sentence Tranformer model with Legal-AUEB-BERT-base as base
model and trained with these citation information.
Ostendorff et al.
4 RESULTS
For our evaluation, we obtain a list of recommendations for each
input document and method and then compute the performance
measures accordingly. We compute the average number of relevant
recommendations, precision, recall, MRR, MAP, and coverage.
4.1 Quantitative Evaluation
4.1.1 Overall Results. Table 2 presents the overall evaluation met-
rics for 27 methods and the two datasets. From the non-hybrid
methods, fastTextLegal yields with 0.05 the highest MAP score on
Open Case Book, whereas on Wikisource, fastTextLegal , Poincaré,
and Walklets all achieve the highest MAP score of 0.031. The hybrid
method of Poincaré ∥ fastTextLegal outperforms the non-hybrids
for Wikisource with 0.035 MAP. For Open Case Book, the MAP of
Poincaré + fastTextLegal and fastTextLegal are equally high.
Due to space constraints, we remove 14 methods from Table 2
(excluded methods are on GitHub9 ). From the word vector-based
methods, we discard the 512 and 4096 tokens variations of Para-
graph Vectors, GloVe and GloVeLegal , as they show a similar per-
formance deterioration as fastTextLegal . The base versions of some
Transformers are also excluded in favour of the better performing
large versions. Similarly, the nli version always outperform the stsb
version of Sentence Transformers (sBERT and sRoBERTa). For the
hybrid variations, we show only the best methods. We also tested
Node2Vec [17] and but exclude it given its low MAP scores.
Regarding the word vector-based methods, we see that the meth-
ods which are trained on the legal corpus (Paragraph Vectors,
fastTextLegal , GloVeLegal ) perform similarly well with a minor ad-
vantage by fastTextLegal . Moreover, there is a margin between the
generic and legal word vectors even though the legal word vectors
are trained on a small corpus compared to ones from the generic
vectors. The advantage of Paragraph Vectors over TF-IDF is consis-
(3) tent with the results from Mandal et al. [28]. Limiting the document
length to 512 or 4096 decreases the effectiveness of fastTextLegal .
A limit of 512 tokens decreases the MAP score to 59% compared
to all tokens on Open Case Book. With 4096 tokens, the perfor-
mance decline is only minor (90% compared to all tokens). The
token limitation effect is also larger on Open Case Book than Wiki-
source. The 4096 tokens version of fastTextLegal even outperforms
all Transformer methods.
Longformer-large is the best Transformer for Open Case Book
with 0.031 MAP. For Wikisource, Legal-AUEB-BERT achieves the
highest MAP of 0.022, closely followed by Legal-JHU-BERT. The
Longformer’s theoretical advantage of processing 4096 instead of
512 tokens does not lead to better results for Wikisource, for which
(4) even BERT scores the same MAP of 0.018. We generally observe
that large models outperform their base counterparts7 . Likewise,
RoBERTa has higher scores than BERT as Liu et al. [27] suggested.
From the Transformers category, Sentence Transformers yield the
worst results. We assume that fine-tuning on the similarity datasets
like NLI or STSB does not increase the performance since the mod-
els do not generalize well to other domains. However, the language
model fine-tuning from Legal-JHU-BERT and Legal-AUEB-BERT
7 Legal-JHU-BERT and Legal-AUEB-BERT are only available as base version.
112
Evaluating Document Representations for Content-based Legal Literature Recommendations ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 2: Overall scores for top 𝑘 = 5 recommendations from Open Case Book and Wikisource as the number of relevant
documents, precision, recall, MRR, MAP and coverage for the 27 methods and the vector sizes. The methods are divided into:
baseline, word vector-based, Transformer-based, citation-based, and hybrid. High scores according to the exact numbers are
underlined (or bold for category-wise). ∗ values were rounded up.

Datasets → Open Case Book Wikisource

Methods ↓ Size Rel. Prec. Recall MRR MAP Cov. Rel. Prec. Recall MRR MAP Cov.
TF-IDF 500000 1.60 0.320 0.032 0.363 0.020 0.487 1.59 0.318 0.026 0.389 0.015 0.446
Paragraph Vectors 300 2.78 0.555 0.056 0.729 0.049 0.892 2.39 0.477 0.036 0.629 0.030 0.841
fastText 300 2.66 0.532 0.053 0.713 0.045 0.811 2.11 0.422 0.031 0.581 0.025 0.772
fastTextLegal 300 2.87 0.574 0.059 0.739 0.050 0.851 2.39 0.478 0.037 0.631 0.031 0.815
fastTextLegal (512) 300 1.97 0.394 0.037 0.591 0.028 0.835 2.16 0.433 0.034 0.587 0.027 0.809
fastTextLegal (4096) 300 2.76 0.552 0.054 0.727 0.045 0.867 2.33 0.466 0.035 0.620 0.029 0.817
GloVe 300 2.68 0.536 0.054 0.702 0.046 0.814 2.06 0.412 0.033 0.577 0.026 0.789
GloVeLegal 300 2.82 0.564 0.057 0.724 0.048 0.834 2.31 0.461 0.037 0.621 0.030 0.804
BERT-base 768 1.26 0.253 0.021 0.428 0.015 0.815 1.62 0.323 0.021 0.485 0.015 0.784
BERT-large 1024 1.35 0.270 0.022 0.443 0.016 0.841 1.82 0.364 0.023 0.530 0.018 0.794
Legal-JHU-BERT-base 768 1.47 0.295 0.025 0.482 0.018 0.848 1.85 0.371 0.027 0.537 0.020 0.796
Legal-AUEB-BERT-base 768 1.66 0.331 0.028 0.506 0.021 0.884 2.01 0.401 0.027 0.573 0.022 0.813
Longformer-base 768 1.91 0.382 0.033 0.572 0.026 0.892 1.65 0.329 0.020 0.514 0.016 0.841
Longformer-large 1024 2.09 0.419 0.039 0.614 0.031 0.885 1.80 0.360 0.023 0.535 0.018 0.826
RoBERTa-large 1024 1.52 0.305 0.026 0.481 0.019 0.843 1.93 0.387 0.026 0.553 0.020 0.782
Sentence-BERT-large-nli 1024 1.03 0.206 0.018 0.352 0.013 0.872 1.37 0.273 0.017 0.443 0.012 0.782
Sentence-BERT-large-nli-stsb 1024 0.98 0.196 0.018 0.338 0.013 0.848 1.36 0.272 0.015 0.434 0.011 0.777
Sentence-RoBERTa-large-nli 1024 0.92 0.183 0.016 0.321 0.011 0.884 1.18 0.236 0.013 0.409 0.009 0.795
BoostNE 300 1.29 0.258 0.022 0.442 0.016 0.800 1.24 0.248 0.016 0.398 0.013 0.832
DeepWalk 300 1.34 0.267 0.028 0.473 0.021 0.818 1.82 0.364 0.030 0.533 0.025 0.856
Poincaré 300 2.24 0.447 0.044 0.629 0.036 0.930 2.33 0.465 0.038 0.598 0.031 0.837
Walklets 300 2.24 0.448 0.043 0.636 0.035 0.816 2.35 0.470 0.038 0.611 0.031 0.826
Poincaré ∥ fastTextLegal 600 2.36 0.473 0.048 0.656 0.041 0.737 2.52 0.505 0.041 0.638 0.035 0.818
Longformer-large ∥ fastTextLegal 1324 2.26 0.451 0.043 0.642 0.035 0.876 1.91 0.383 0.025 0.547 0.020 0.829
300
Poincaré + fastTextLegal 2.85 0.571 0.058 0.746 0.050 0.860 2.48 0.497 0.040 0.646 0.034 0.835
300
300
Poincaré + Longformer-large 2.09 0.419 0.039 0.630 0.033 0.885 1.80 0.360 0.023 0.548 0.019 0.826
1024
Sentence-Legal-AUEB-BERT-base 768 2.19 0.438 0.039 0.603 0.031 0.917 2.36 0.471 0.038 0.602 0.032 0.849

does improve the performance, whereby Legal-AUEB-BERT gen-
erally outperforms Legal-JHU-BERT. For Open Case Book, Legal-
AUEB-BERT is the best model in the Transformer category in terms
of MAP even though it is only used as base version.
Poincaré and Walklets are by far the best methods in the citation
category. For Wikisource, the two citation-based methods, score the
same MAP of 0.031 as fastTextLegal . Compared to the word vector-
based methods, the citation methods do better on Wikisource than
on Open Case Book.
In the category of hybrid methods, the combination of text and
citations improves the performance. For Open Case Book, the score
summation Poincaré + fastTextLegal has the same MAP of 0.05
as fastTextLegal but a higher MRR of 0.746. The MRR of Poincaré
+ fastTextLegal is even higher than the MRR of its sub-methods
Poincaré (0.629) and fastTextLegal (0.739) individually. The concate-
nation of Poincaré ∥ fastTextLegal is with 0.035 MAP the best method
on Wikisource. Using citation as training signal as in Sentence-
Legal-AUEB-BERT also improves the performance but not as much
as concatenation or summation. When comparing the three hybrid
variations, score summation achieves overall the best results. In the
case of Wikisource, the concatenation’s scores are below its sub-
methods, while summation has at least the best sub-methods score.
Moreover, combining two text-based methods such as Longformer-
large and fastTextLegal never improves its sub-methods.
4.1.2 Document Length. The effect of the document length on the
performance in terms of MAP is displayed in Figure 1. We group

113
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Paragraph Vectors
fastText
MAP (OpenCaseBook)
0.06
0.04
0.02
Paragraph Vectors
0.00
(187, 2327] fastText
(2327, 3499] (3499, 4772] fastText Legal-4096
0.06
MAP (WikiSource)
0.05
0.04
0.03
0.02
0.01
0.00
(31, 1777] (1777, 2666] (2666, 3520]
Figure 1: MAP wrt. words in the seed document of Open Case Book (top) and Wikisource (bottom). The more words, the better
the results, no peak at medium length. fastTextLegal outperforms Legal-BERT and Longformer for short documents.
the seed documents into eight equal-sized buckets (each bucket
represents the equal number of documents) depending on the word
count in the document text to make the two datasets comparable.
Both datasets, Open Case Book and Wikisource, present a similar
outcome. The MAP increases as the word count increases. Table 2
presents the average overall documents and, therefore, the overall
best method is not equal to the best method in some subsets. For
instance, Paragraph Vectors achieve the best results for several buck-
ets, e.g., 4772-6172 words in Open Case Book or 6083-8659 words
in Wikisource. The text limitation of fastTextLegal (4096 tokens)
in comparison to fastText is also clearly visible. The performance
difference between the two methods increases as the document
length increases. For the first buckets with less than 4096 words,
e.g., 187-2327 words in Open Case Book, one could expect no differ-
ence since the limitation does not affect the seed documents in these
buckets. However, we observe a difference since target documents
are not grouped into the same buckets. Remarkable is that the per-
formance difference for very long documents is less substantial.
When comparing Longformer-large and Legal-AUEB-BERT, we
also see an opposing performance shift with changing word count.
While Legal-AUEB-BERT’s scores are relatively stable throughout
all buckets, Longformer depends more on the document length. On
the one hand, Longformer performs worse than Legal-AUEB-BERT
for short documents, i.e., 187-2327 words in Open Case Book, and
31-1777 words in Wikisource. On the other hand, for documents
with more words, Longformer mostly outperforms Legal-AUEB-
BERT by a large margin. The citation-based method Poincaré is as
well affected by the document length. However, this effect is due to
a positive correlation between word count and citation count.
4.1.3 Citation Count. Figure 2 shows the effect of the number of
in- and out-citations (i.e., edges in the citation graph) on the MAP
score. The citation analysis for Wikisource confirms the word count
analysis. More data leads to better results. Instead, for Open Case
Book, the performance of the citation-based methods peak for 31-51
Ostendorff et al.
fastText Legal Legal-AUEB-BERT-base Poincaré
fastText Legal-4096 Longformer-large Poincaré + fastText Legal
fastText Legal Legal-AUEB-BERT-base Poincaré
(4772, 6172] Longformer-large
(6172, 7859] (7859, Poincaré
11070] + fastText Legal
(11070, 16785] (16785, 88269]
Text length as word count (8 equal-sized buckets)
(3520, 4532] (4532, 6083] (6083, 8659] (8659, 12930] (12930, 136017]
Text length as word count (8 equal-sized buckets)
citations and even decrease at 67-89 citations. When comparing
Poincaré and Walklets there is no superior method and no depen-
dency pattern is visible. The performance effect on DeepWalk is
more substantial. The number of citations must be above a cer-
tain threshold to allow DeepWalk to achieve competitive results.
For Open Case Book, the threshold is at 51-67 citations, and for
Wikisource, it is at 30-50 citations. Figure 2 also shows the on aver-
age higher MAP of Poincaré + fastTextLegal in comparison to the
other approaches. Citation-based methods require citations to work,
whereas text methods do not have this limitation (see 0-14 citations
for Open Case Book). When no citations are available, citation-
based methods cannot recommend any documents, whereas the
text methods still work (see 0-14 citations for Open Case Book).
Our citation-based methods use only a fraction of original cita-
tion data, 70,865 citations in Open Case Book, and 331,498 citations
in Wikisource, because of limitation to the documents available in
the silver standards. For comparison, the most-cited decision from
CourtListener (the underlying corpus of Wikisource) has 88,940 ci-
tations, whereas in experimental data of Wikisource the maximum
number of in- and out-citations is 386. As a result, we expect the
citation-based methods, especially DeepWalk, to work even better
when applied on the full corpus.
4.1.4 Coverage and Similarity of Recommendations. In addition
to the accuracy-oriented metrics, Table 2 reports also the cover-
age of the recommendation methods. A recommender systems for
an expert audience should not focus on small set of most-popular
items but rather provide a high coverage of the whole item collec-
tion. However, coverage alone does not account for relevancy and,
therefore, it must be contextualized with other metrics, e.g., MAP.
Overall, two citation-based methods yield the highest coverage
for both datasets, i.e., Poincaré for Open Case Book and DeepWalk
for Wikisource. In particular, Poincaré has not only a high coverage
but also high MAP scores. Yet, the numbers do not indicate that
citation-based methods have generally a higher coverage since the
114
Evaluating Document Representations for Content-based Legal Literature Recommendations ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Paragraph Vectors Longformer-large DeepWalk Walklets

fastText Legal BoostNE Poincaré Poincaré + fastText Legal

0.06
MAP (OpenCaseBook)

0.05

0.04

0.03

0.02

0.01
Paragraph Vectors Longformer-large DeepWalk Walklets
0.00
(0, 14] (14, 23] fastText(23,
Legal
31] BoostNE
(31, 40] Poincaré
(40, 51] Poincaré + fastText Legal
(51, 67] (67, 89] (89, 425]
In- and out- citations(8 equal-sized buckets)
0.06
MAP (WikiSource)

0.05

0.04

0.03

0.02

0.01

0.00
(0, 3] (3, 6] (6, 12] (12, 21] (21, 30] (30, 50] (50, 82] (82, 386]
In- and out- citations(8 equal-sized buckets)

Figure 2: MAP scores wrt. citation count for Open Case Book (top) and Wikisource (bottom). Among citation-based methods,
Poincaré and Walklets perform on average the best, while DeepWalk outperforms them only for Wikisource and when more
than 82 citations are available (rightmost bucket).

Le
g
Po Po
inc inc Besides the coverage, we also analyze the similarity or diversity
Pa al-A ar a
fas ragr UEB
é | ré +
|f
as fas
of the recommendations between two methods. Figure 3 shows the
Gl a
similarity measured as Jaccard index for selected methods. Method
o t Te ph BE- D tT tT
e e
TF Ve L fast xt L Ve RT- eep Wal Poin xt L xt L
-ID eg Te eg cto ba Wa kle ca e e
F al xt al rs se lk ts ré gal gal
pairs with 𝐽 (𝑎, 𝑏) = 1 have identical recommendations, whereas
TF-IDF 1.00 0.17 0.15 0.16 0.10 0.04 0.06 0.06 0.06 0.11 0.13
𝐽 (𝑎, 𝑏) = 0 means no common recommendations. Generally speak-
ing, the similarity of all method pairs is considerably low (𝐽 < 0.8).
The highest similarity can be found between a hybrid method and
GloVe Legal 0.17 1.00 0.40 0.67 0.27 0.08 0.09 0.12 0.11 0.23 0.52

fastText 0.15 0.40 1.00 0.41 0.21 0.07 0.07 0.10 0.09 0.18 0.33
fastText Legal 0.16 0.67 0.41 1.00 0.28 0.09 0.09 0.13 0.11 0.24 0.76
Paragraph Vectors 0.10 0.27 0.21 0.28 1.00 0.09 0.09 0.13 0.12 0.19 0.24
Legal-AUEB-BERT-base 0.04 0.08 0.07 0.09 0.09 1.00 0.04 0.06 0.05 0.07 0.08
DeepWalk 0.06 0.09 0.07 0.09 0.09 0.04 1.00 0.20 0.14 0.14 0.12
Walklets 0.06 0.12 0.10 0.13 0.13 0.06 0.20 1.00 0.32 0.27 0.18
Poincaré 0.06 0.11 0.09 0.11 0.12 0.05 0.14 0.32 1.00 0.39 0.32
Poincaré || fastText Legal 0.11 0.23 0.18 0.24 0.19 0.07 0.14 0.27 0.39 1.00 0.23
Poincaré + fastText Legal 0.13 0.52 0.33 0.76 0.24 0.08 0.12 0.18 0.32 0.23 1.00
Figure 3: Jaccard index for similarity or diversity of two
recommendation sets (average over all seeds from the two
datasets).
one of its sub-methods, e.g., Poincaré + fastTextLegal and fastText-
Legal with 𝐽 = 0.76. Apart from that, substantial similarity can be
only found between pairs from the same category. For example, the
pair of the two text-based methods of GloVeLegal and fastTextLegal
yields 𝐽 = 0.67. Citation-based methods tend to have a lower sim-
ilarity compared to the text-based methods, whereby the highest
Jaccard index between two citation-based methods is achieved for
Walklets and Poincaré with 𝐽 = 0.32. Like the coverage metric,
the Jaccard index should be considered in relation to the accu-
racy results. GloVeLegal and fastTextLegal yield equally high MAP
scores, while having also a high recommendation’s similarity. In
contrast, the MAP for Wikisource from fastTextLegal and Poincaré
is equally high, too. However, their recommendation’s similarity
is low 𝐽 = 0.11. Consequently, fastTextLegal and Poincaré provide
relevant recommendations that are diverse from each other. This
explains the good performance of their hybrid combination.

4.2 Qualitative Evaluation

Due to lack of openly available gold standards, we conduct our
text-based Paragraph Vectors or Longformer-base also achieve a
quantitative analysis using silver standards. Thus, we additionally
considerably high coverage. The lowest coverage has by far the
conduct a qualitative evaluation with domain experts to estimate
TF-IDF baseline. Notable, the hybrid methods with concatenation
the quality of our silver standards.
and summation have a different effect on the coverage as on the ac-
Table 3 lists one of the randomly chosen seed decisions (Mu-
curacy metrics. While the hybrid methods generally yield a higher
gler vs. Kansas8 ), and five recommended similar decisions, each
MAP, their coverage is lower compared to their sub-methods. Only,
from fastTextLegal and Poincaré. In Mugler vs. Kansas (1887), the
the Sentence-Legal-AUEB-BERT-base yields a higher coverage com-
pared to Legal-AUEB-BERT-base. 8 https://www.courtlistener.com/opinion/92076/mugler-v-kansas/

115
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Ostendorff et al.

Table 3: Examples from fastTextLegal and Poincaré (other methods are in the supplementary material) for Mugler v. Kansas
with relevance annotations by the silver standards (S) and domain expert (D).

Open Case Book Wikisource

# Recommendations Year S D Recommendations Year S D
1 Yick Wo v. Hopkins 1886 N N Kidd v. Pearson 1888 N Y
fastTextLegal

2 Munn v. Illinois 1876 Y Y Lawton v. Steele 1894 N Y

3 LS. Dealers’ & Butchers’ v. Crescent City LS. 1870 N Y Yick Wo v. Hopkins 1886 N N
4 Butchers’ Benevolent v. Crescent City LS. 1872 Y Y Geer v. Connecticut 1896 N Y
5 Lochner v. New York 1905 Y Y Groves v. Slaughter 1841 Y N
1 Yick Wo v. Hopkins 1886 N N Rast v. Van Deman & Lewis Co. 1916 Y N
2 Allgeyer v. Louisiana 1897 Y Y County of Mobile v. Kimball 1881 N N
Poincaré

3 Calder v. Wife 1798 N N Brass v. North Dakota Ex Rel. Stoeser 1894 Y Y

4 Davidson v. New Orleans 1877 Y Y Erie R. Co. v. Williams 1914 Y Y
5 Muller v. Oregon 1908 Y Y Hall v. Geiger-Jones Co. 1917 Y Y

court held that Kansas could constitutionally outlaw liquor sales
with constitutional issues raised on substantive due process (Four-
teenth Amendment) and takings (Fifth Amendment). We provide a
description of the cases and their relevance on GitHub9 .
The sample verification indicates the usefulness of both text-
based and citation-based methods and does not contradict our quan-
titative findings. Each of the recommendations have a legal impor-
tant connection to the seed case (either the Fourteenth Amendment
or Fifth Amendment), although it is difficult to say whether the
higher-ranked cases are more similar along an important topical
dimension. The rankings do not appear to be driven by facts pre-
sented in the case as most of them have not to do with alcohol
bans. Only Kidd vs. Pearson (1888) is about liquor sales as the seed
decision. The samples also do not reveal considerable differences
between text- and citation-based similarity. With regards to the
silver standards, the domain expert agrees in 14 of 20 cases (70%).
In only two cases the domain expert classify a recommendation as
irrelevant despite being classified as relevant in the silver standard.
5 DISCUSSION
Our experiments explore the applicability of the latest advances in
research to the use case of legal literature recommendations. Exist-
ing studies on legal recommendations typically rely on small-scale
user studies and are therefore limited in the number of approaches
that they can evaluate (Section 2). For this study, we utilize rele-
vance annotations from two publicly available sources, i.e., Open
Case Book and Wikisource. These annotations does not only enable
us to evaluate the recommendations of 2,964 documents but also
the comparison of in total 41 methods and their variations of which
27 methods are presented in this paper.
Our extensive evaluation shows a large variance in the recom-
mendation performance. Such a variance is known from other stud-
ies [4]. There is no single method that yields the highest scores
across all metrics and all datasets. Despite that, fastTextLegal is on
average the best of all 41 methods. fastTextLegal yields the highest
MAP for Open Case Book, while for Wikisource only hybrid meth-
ods outperform fastTextLegal . Also, the coverage of fastTextLegal
is considerably high for both datasets. Simultaneously, fastText-
Legal is robust to corner cases since neither very short nor very
long documents reduce fastTextLegal ’s performance substantially.
These results confirm the findings from Arora et al. [1] that average
word vectors are “simple but tough-to-beat baseline”. Regarding
baselines, our TF-IDF baseline yields one of the worst results. In
terms of accuracy metrics, only some Transformers are worse than
TF-IDF, but especially TF-IDF’s coverage is the lowest by a large
margin. With a coverage below 50%, TF-IDF fails to provide diverse
recommendations that are desirable for legal literature research.
The transfer of research advances to the legal domain is one
aspect of our experiments. Thus, the performance of Transformers
and citation embeddings is of particular interest. Despite the success
of Transformers for many NLP tasks, Transformers yield on average
the worst results for representing lengthy documents written in
legal English. The other two method categories, word vector-based,
and citation-based methods, surpass Transformers.
The word vector-based methods achieve overall the best results
among the non-hybrid methods. All word vectors with in-domain
training, i.e., Paragraph Vectors, fastTextLegal , and GloVeLegal , per-
form similarly good with a minor advantage by fastTextLegal . Their
similar performance aligns with the large overlap among their re-
commendations. Despite a small corpus of 65,635 documents, the
in-domain training generally improves the performance as the gap
between the out-of-domain fastText and fastTextLegal shows. Given
that the training of custom word vectors is feasible on commodity
hardware, in-domain training is advised. More significant than the
gap between in- and out-of-domain word vectors is the effect of
limited document lengths. For Open Case Book, the fastTextLegal
variation limited to the first 512 tokens has only 52% of the MAP
of the full-text method. For Wikisource, the performance decline
exists as well but is less significant. This effect highlights the advan-
tage of the word vector-based methods that they derive meaningful
representations of documents with arbitrary length.

116
Evaluating Document Representations for Content-based Legal Literature Recommendations
The evaluated Transformers cannot process documents of arbi-
trary length but are either limited to 512 or 4096 tokens. This limi-
tation contributes to Transformers’ low performance. For instance,
Longformer-large’s MAP is almost twice as high as BERT-large’s
MAP on Open Case Book. However, for Wikisource both models
yield the same MAP scores. For Wikisource, the in-domain pretrain-
ing as a larger effect than the token limit since Legal-AUEB-BERT
achieves the best results among the Transformers. Regarding the
Transformer pretraining, the difference between Legal-JHU-BERT
and Legal-AUEB-BERT shows the effect between two pretraining
approaches. The corpora and the hyperparameter settings used
during pretraining are crucial. Even though Legal-JHU-BERT was
exclusively pretrained on the CAP corpus, which has a high over-
lap with Open Case Book, Legal-AUEB-BERT still outperforms
Legal-JHU-BERT on Open Case Book. Given these findings, we
expect the performance of Transformers could be improved by in-
creasing the token limit beyond the 4096 tokens and by additional
in-domain pretraining. Such improvements are technically possible
but add significant computational effort. In contrast to word vectors,
Transformers are not trained on commodity hardware but on GPUs.
Especially long-sequence Transformers such as the Longformer
require GPUs with large memory. Such hardware may not be avail-
able in production deployments. Moreover, the computational effort
must be seen in relation to the other methods. Put differently, even
fastTextLegal limited to 512 tokens outperforms all Transformers.
Concerning the citation embeddings, we consider Poincaré, clo-
sely followed by Walklets, as the best method. In particular, the two
methods outperform the other citation methods even when only a
few citations are available, which makes them attractive for legal
research. Poincaré also provides the highest coverage for Open Case
Book, emphasizing its quality for literature recommendations. For
Wikisource, DeepWalk has the highest coverage despite yielding
generally low accuracy scores. As Figure 2 shows, DeepWalk’s MAP
score improves substantially as the number of citations increases.
Therefore, we expect that DeepWalk but also the other citation
methods would perform even better when applied on larger cita-
tion graph. The analysis of recommendation similarity also shows
little overlap between the citation-based methods and the text-
based methods (Figure 3). This indicates that the two approaches
complement each other and motivates the use of hybrid methods.
Related work has already shown the benefit of hybrid methods
for literature recommendations [6, 49]. Our experiments confirm
these findings. The simple approaches of score summation or vector
concatenation can improve the results. In particular, Poincaré +
fastTextLegal never leads to a decline in performance. Instead, it
increases the performance for corner cases in which one of the
sub-methods performs poorly. Vector concatenation has mixed
effects on the performance, e.g., positive effect for Wikisource and
negative effect for Open Case Book. Using citations as training
data in Sentence Transformers can also be considered as a hybrid
method that improves the performance. However, this requires
additional effort for training a new Sentence Transformer model.
As we discuss in Section 3.1, we consider Open Case Book and
Wikisource more of silver than gold standards. With the qualitative
evaluation, we mitigate the risk of misinterpreting the quantitative
results, whereby we acknowledge our small sample size. The overall
117
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
agreement with the domain expert is high. The expert tends to clas-
sify more recommendations as relevant than the silver standards,
i.e., relevant recommendations are missed. This explains the rela-
tively low recall from the quantitative evaluation. In a user study,
we would expect only minor changes in the ranking of methods
with similar scores, e.g., fastTextLegal and GloVeLegal . The category
ranking would remain the same. The benefit of our silver standards
is the number of available relevance annotations. The number of
annotations in related user studies is with up to 50 annotations
rather low. Instead, our silver standards provide a magnitude more
relevance annotations. Almost 3,000 relevance annotations enable
evaluations regarding text length, citation count, or other proper-
ties that would be otherwise magnitudes more difficult. Similarly,
the user studies are difficult to reproduce as their data is mostly
unavailable. This leads to reproducibility being an issue in recom-
mender system research [4]. The open license of the silver standards
allows the sharing of all evaluation data and, therefore, contributes
to more reproducibility. In summary, the proposed datasets bring
great value to the field, overcoming eventual shortcomings.
6 CONCLUSION
We present an extensive empirical evaluation of 27 document rep-
resentation methods in the context of legal literature recommen-
dations. In contrast to previous small-scale studies, we evaluate
the methods over two document corpora containing 2,964 docu-
ments (1,601 from Open Case Book and 1,363 from Wikisource). We
underpin our findings with a sample-based qualitative evaluation.
Our analysis of the results reveals fastTextLegal (averaged fastText
word vectors trained on our corpora) as the overall best performing
method. Moreover, we find that all methods have a low overlap be-
tween their recommendations and are vulnerable to certain dataset
characteristics like text length and number of citations available. To
mitigate the weakness of single methods and to increase recommen-
dation diversity, we propose hybrid methods like score summation
of fastTextLegal and Poincaré that outperforms all other methods on
both datasets. Although there are limitations in the experimental
evaluation due to the lack of openly available ground truth data, we
are able to draw meaningful conclusions for the behavior of text-
based and citation-based document embeddings in the context of
legal document recommendation. Our source code, trained models,
and datasets are openly available to encourage further research9 .
ACKNOWLEDGMENTS
We would like to thank Christoph Alt, Till Blume, and the anony-
mous reviewers for their comments. The research presented in this
article is funded by the German Federal Ministry of Education and
Research (BMBF) through the project QURATOR (Unternehmen
Region, Wachstumskern, no. 03WKDA1A) and by the project LYNX,
which has received funding from the EU’s Horizon 2020 research
and innovation program under grant agreement no. 780602.
REFERENCES
[1] Sanjeev Arora, Yingyu Liang, and Tengyu Ma. 2017. A simple but though Baseline
for Sentence Embeddings. In 5th International Conference on Learning Representa-
tions (ICLR 2017), Vol. 15. 416–424.
9 GitHub repository: https://github.com/malteos/legal-document-similarity
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
[2] Elliott Ash and Daniel L. Chen. 2018. Case Vectors: Spatial Representations of
the Law Using Document Embeddings. SSRN Electronic Journal 11, 2017 (may
2018), 313–337. https://doi.org/10.2139/ssrn.3204926
[3] Xiaomei Bai, Mengyang Wang, Ivan Lee, Zhuo Yang, Xiangjie Kong, and Feng
Xia. 2019. Scientific paper recommendation: A survey. IEEE Access 7 (2019),
9324–9339.
[4] Joeran Beel, Corinna Breitinger, Stefan Langer, Andreas Lommatzsch, and Bela
Gipp. 2016. Towards reproducibility in recommender-systems research. User
Modeling and User-Adapted Interaction (UMAI) 26 (2016).
[5] Iz Beltagy, Matthew E. Peters, and Arman Cohan. 2020. Longformer: The Long-
Document Transformer. (2020). arXiv:2004.05150
[6] Paheli Bhattacharya, Kripabandhu Ghosh, Arindam Pal, and Saptarshi Ghosh.
2020. Methods for Computing Legal Document Similarity: A Comparative Study.
(2020). arXiv:2004.12307
[7] David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation.
Journal of machine Learning research 3, Jan (2003), 993–1022.
[8] Guido Boella, Luigi Di Caro, Llio Humphreys, Livio Robaldo, Piercarlo Rossi,
and Leendert van der Torre. 2016. Eunomos, a legal document and knowledge
management system for the Web to provide relevant, reliable and up-to-date
information on the law. Artificial Intelligence and Law 24, 3 (2016), 245–283.
[9] Alexander Boer and Radboud Winkels. 2016. Making a cold start in legal recom-
mendation: An experiment. Frontiers in Artificial Intelligence and Applications
294 (2016), 131–136. https://doi.org/10.3233/978-1-61499-726-9-131
[10] Piotr Bojanowski, Edouard Grave, Armand Joulin, and Tomas Mikolov. 2017. En-
riching Word Vectors with Subword Information. Transactions of the Association
for Computational Linguistics 5 (2017), 135–146.
[11] Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Man-
ning. 2015. A large annotated corpus for learning natural language inference.
Proc. of EMNLP (2015), 632–642.
[12] Jane Bromley, J.W. Bentz, Leon Bottou, I. Guyon, Yann Lecun, C. Moore, Eduard
Sackinger, and R. Shah. 1993. Signature verification using a Siamese time de-
lay neural network. International Journal of Pattern Recognition and Artificial
Intelligence 7, 4 (1993).
[13] Daniel Cer, Mona Diab, Eneko Agirre, Iñigo Lopez-Gazpio, and Lucia Specia. 2017.
SemEval-2017 Task 1: Semantic Textual Similarity Multilingual and Crosslingual
Focused Evaluation. In Proc. of the 11th International Workshop on Semantic
Evaluation (SemEval-2017). ACL, Vancouver, Canada, 1–14.
[14] Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras,
and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of Law
School. In Findings of the Association for Computational Linguistics: EMNLP 2020.
ACL, Stroudsburg, PA, USA, 2898–2904.
[15] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proc. of the 2019 Conf. of the NAACL. ACL, Minneapolis, Minnesota, 4171–4186.
[16] Mouzhi Ge, Carla Delgado-Battenfeld, and Dietmar Jannach. 2010. Beyond
accuracy: evaluating recommender systems by coverage and serendipity. In Proc.
of RecSys ’10. ACM Press, New York, New York, USA, 257.
[17] Aditya Grover and Jure Leskovec. 2016. node2vec: Scalable Feature Learning for
Networks. In Proc. of the 22nd ACM SIGKDD Int. Conf. on Knowledge Discovery
and Data Mining - KDD ’16. ACM Press, New York, New York, USA, 855–864.
[18] Nils Holzenberger, Andrew Blair-Stanek, and Benjamin Van Durme. 2020. A
dataset for statutory reasoning in tax law entailment and question answering. In
Proc. of the 2020 Natural Legal Language Processing Workshop. 31–38.
[19] Paul Jaccard. 1912. The Distribution of the Flora in the Alpine Zone. New
Phytologist 11, 2 (feb 1912), 37–50.
[20] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. 2017. Bag
of Tricks for Efficient Text Classification. In Proc. of EACL 2017. ACL, Stroudsburg,
PA, USA, 427–431.
[21] Dmitri Krioukov, Fragkiskos Papadopoulos, Maksim Kitsak, Amin Vahdat, and
Marián Boguñá. 2010. Hyperbolic geometry of complex networks. Physical
Review E - Statistical, Nonlinear, and Soft Matter Physics 82, 3 (2010), 1–18.
[22] Sushanta Kumar, P. Krishna Reddy, V. Balakista Reddy, and Aditya Singh. 2011.
Similarity analysis of legal judgments. Compute 2011 - 4th Annual ACM Bangalore
Conference (2011). https://doi.org/10.1145/1980422.1980439
[23] Steven A. Lastres. 2013. Rebooting Legal Research in a Digital Age. https:
//www.lexisnexis.com/documents/pdf/20130806061418_large.pdf
[24] J. H. Lau and T. Baldwin. 2016. An Empirical Evaluation of doc2vec with Practical
Insights into Document Embedding Generation. In Proc. Workshop on Representa-
tion Learning for NLP. https://doi.org/10.18653/v1/w16-1609
[25] Quoc V. Le and Tomas Mikolov. 2014. Distributed Representations of Sentences
and Documents. Int. Conf. on Machine Learning 32 (2014), 1188–1196.
[26] Jundong Li, Liang Wu, Ruocheng Guo, Chenghao Liu, and Huan Liu. 2019. Multi-
level network embedding with boosted low-rank matrix approximation. In Proc.
of the 2019 IEEE/ACM International Conference on Advances in Social Networks
Analysis and Mining. ACM, New York, NY, USA, 49–56.
[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A
Robustly Optimized BERT Pretraining Approach. (2019). arXiv:1907.11692
118
Ostendorff et al.
[28] Arpan Mandal, Raktim Chaki, Sarbajit Saha, Kripabandhu Ghosh, Arindam Pal,
and Saptarshi Ghosh. 2017. Measuring Similarity among Legal Court Case
Documents. In Proc. of Compute ’17. 1–9.
[29] Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schutze. 2008. Intro-
duction to Information Retrieval. Vol. 16. Cambridge University Press, Cambridge.
100–103 pages. https://doi.org/10.1017/CBO9780511809071
[30] David Mellinkoff. 1963. The language of the law. Boston: Little Brown and
Company (1963).
[31] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Esti-
mation of Word Representations in Vector Space. (2013), 1–12. arXiv:1301.3781
[32] Akshay Minocha, Navjyoti Singh, and Arjit Srivastava. 2015. Finding Relevant
Indian Judgments using Dispersion of Citation Network. In Proc. of WWW ’15.
ACM Press, New York, New York, USA, 1085–1088.
[33] Rohan Nanda, Giovanni Siragusa, Luigi Di Caro, Guido Boella, Lorenzo Grossio,
Marco Gerbaudo, and Francesco Costamagna. 2019. Unsupervised and supervised
text similarity systems for automated identification of national implementing
measures of European directives. Artificial Intelligence and Law 27, 2 (2019),
199–225. https://doi.org/10.1007/s10506-018-9236-y
[34] Maximilian Nickel and Douwe Kiela. 2017. Poincaré embeddings for learning
hierarchical representations. Advances in Neural Information Processing Systems
2017-Decem, Nips (2017), 6339–6348. arXiv:1705.08039
[35] Malte Ostendorff, Till Blume, and Saskia Ostendorff. 2020. Towards an Open
Platform for Legal Information. In Proc. of the ACM/IEEE Joint Conference on
Digital Libraries in 2020. ACM, New York, NY, USA, 385–388.
[36] Malte Ostendorff, Peter Bourgonje, Maria Berger, Julian Moreno-Schneider, Georg
Rehm, and Bela Gipp. 2019. Enriching BERT with Knowledge Graph Embeddings
for Document Classification. In Proc. of the 15th Conference on Natural Language
Processing (KONVENS 2019). GSCL, Erlangen, Germany, 305–312.
[37] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M.
Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cour-
napeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine
Learning in Python. Journal of Machine Learning Research 12 (2011), 2825–2830.
[38] Jeffrey Pennington, Richard Socher, and Christopher Manning. 2014. Glove:
Global Vectors for Word Representation. In Proc. of the 2014 Conference on Em-
pirical Methods in Natural Language Processing (EMNLP). ACL, Stroudsburg, PA,
USA, 1532–1543. https://doi.org/10.3115/v1/D14-1162
[39] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. 2014. DeepWalk: online learning
of social representations. In Proc. of KDD ’14. ACM Press, New York, New York,
USA, 701–710.
[40] Bryan Perozzi, Vivek Kulkarni, Haochen Chen, and Steven Skiena. 2017. Don’t
Walk, Skip!: Online Learning of Multi-scale Network Embeddings. In Proc. of the
2017 IEEE/ACM International Conference on Advances in Social Networks Analysis
and Mining 2017. ACM, New York, NY, USA, 258–265.
[41] Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings
using Siamese BERT-Networks. In The 2019 Conference on Empirical Methods in
Natural Language Processing (EMNLP 2019). arXiv:1908.10084
[42] Benedek Rozemberczki, Oliver Kiss, and Rik Sarkar. 2020. An API Oriented
Open-source Python Framework for Unsupervised Learning on Graphs. (2020).
arXiv:2003.04819
[43] G. Salton, A. Wong, and C. S. Yang. 1975. Vector Space Model for Automatic
Indexing. Information Retrieval and Language Processing. Commun. ACM 18, 11
(1975), 613–620.
[44] Malte Schwarzer, Moritz Schubotz, Norman Meuschke, and Corinna Breitinger.
2016. Evaluating Link-based Recommendations for Wikipedia. Proc. of the 16th
ACM/IEEE Joint Conference on Digital Libraries (JCDL‘16) (2016), 191–200.
[45] Marc van Opijnen and Cristiana Santos. 2017. On the concept of relevance in
legal information retrieval. Artificial Intelligence and Law 25, 1 (2017), 65–87.
[46] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser,
and I. Polosukhin. 2017. Attention Is All You Need. Advances in Neural Information
Processing Systems 30 (Jun 2017), 5998–6008.
[47] Rupali S. Wagh and Deepa Anand. 2020. Legal document similarity: A multicri-
teria decision-making perspective. PeerJ Computer Science 2020, 3 (2020), 1–20.
https://doi.org/10.7717/peerj-cs.262
[48] Lidan Wang, Ming Tan, and Jiawei Han. 2016. FastHybrid: A hybrid model for effi-
cient answer selection. Proc. of the 26th International Conference on Computational
Linguistics (2016), 2378–2388.
[49] Gineke Wiggers and Suzan Verberne. 2019. Citation Metrics for Legal Information
Retrieval Systems. In BIR@ECIR. 39–50.
[50] Wikisource. 2020. United States Supreme Court decisions by topic.
https://en.wikisource.org/wiki/Category:United_States_Supreme_Court_
decisions_by_topic
[51] Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A Broad-Coverage
Challenge Corpus for Sentence Understanding through Inference. (2018), 1112–
1122. https://doi.org/10.18653/v1/n18-1101
[52] Radboud Winkels, Alexander Boer, Bart Vredebregt, and Alexander Van Someren.
2014. Towards a Legal Recommender System. In Frontiers in Artificial Intelligence
and Applications, Vol. 271. 169–178.
 From Data to Information: Automating Data Science to Explore
the U.S. Court System
Andrew Paley Andong L. Li Zhao
andrewpaley@u.northwestern.edu andong@u.northwestern.edu
Northwestern University Northwestern University
Sergio Servantez Rachel F. Adler
servantez@u.northwestern.edu r-adler@neiu.edu
Northwestern University Northeastern Illinois University
Northwestern University
Adam Pah David Schwartz
a-pah@kellogg.northwestern.edu david.schwartz@law.northwestern.edu
Northwestern University Northwestern University
Alexander Einarsson
aeinarsson@u.northwestern.edu Kristian.Hammond@northwestern.edu
Northwestern University
ABSTRACT
The U.S. court system is the nation’s arbiter of justice, tasked with
the responsibility of ensuring equal protection under the law. But
hurdles to information access obscure the inner workings of the
system, preventing stakeholders – from legal scholars to journalists
and members of the public – from understanding the state of justice
in America at scale. There is an ongoing data access argument here:
U.S. court records are public data and should be freely available.
But open data arguments represent a half-measure; what we really
need is open information. This distinction marks the difference
between downloading a zip file containing a quarter-million case
dockets and getting the real-time answer to a question like “Are pro
se parties more or less likely to receive fee waivers?” To help bridge
that gap, we introduce a novel platform and user experience that
provides users with the tools necessary to explore data and drive
analysis via natural language statements. Our approach leverages an
ontology configuration that adds domain-relevant data semantics
to database schemas to provide support for user guidance and for
search and analysis without user-entered code or SQL. The system
is embodied in a “natural-language notebook” user experience,
and we apply this approach to the space of case docket data from
the U.S. federal court system. Additionally, we provide detail on
the collection, ingestion and processing of the dockets themselves,
including early experiments in the use of language modeling for
docket entry classification with an initial focus on motions.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466100
119
Harper Pack
harper.pack@northwestern.edu
Northwestern University
Marko Sterbentz
marko.sterbentz@u.northwestern.edu
Northwestern University
Cameron Barrie
cameron.barrie@u.northwestern.edu
Northwestern University
Kristian Hammond
Northwestern University
CCS CONCEPTS
• Information systems → Decision support systems; • Ap-
plied computing → Law; • Computing methodologies → Nat-
ural language processing; • Human-centered computing →
Natural language interfaces.
KEYWORDS
notebook interface, information extraction, data analytics, natural
language processing, visualization
ACM Reference Format:
Andrew Paley, Andong L. Li Zhao, Harper Pack, Sergio Servantez, Rachel
F. Adler, Marko Sterbentz, Adam Pah, David Schwartz, Cameron Barrie,
Alexander Einarsson, and Kristian Hammond. 2021. From Data to Informa-
tion: Automating Data Science to Explore the U.S. Court System. In Eigh-
teenth International Conference for Artificial Intelligence and Law (ICAIL’21),
June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 10 pages.
https://doi.org/10.1145/3462757.3466100
1 INTRODUCTION
In the United States, the federal judicial system serves as a vital
umpire, ideally ensuring equal protection and justice under the law.
Its mechanics are a tapestry of countless unique decisions made
by individuals across 94 district courts, 13 circuit courts and the
Supreme Court. While this system is at the core of the management
of justice in the United States, its operation is essentially inaccessi-
ble. Distributed decision making, inaccessible data, and the general
public’s lack of technical skills sufficient to analyze data mean that
the actual mechanics of U.S. justice are largely obscured. As citi-
zens, we trust our laws are being enforced equally but – absent the
occasional headline-grabbing case – it’s all opaque to the majority
of us.
One issue is that data access is prohibitively expensive. Public
Court records are available only through a paywall called the Public
Access to Court Electronic Records (PACER) system. Other research
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
has explored such limitations on access [1, 35], as well as the ques-
tionable completeness of the data available [30, 32]. Legislation to
eliminate the PACER fees is progressing through Congress – one
step towards opening the courts to public scrutiny and understand-
ing.
But making court records free won’t eliminate all barriers to
access. Many open-government initiatives in the U.S. and abroad
have yielded a growing array of public datasets [2, 24], and work to-
wards data transparency is an ongoing effort [14]. However, while
access to data is necessary, it’s insufficient: the applied value of
that data to the end goal of increased public understanding – of
access to information – remains stymied by the limited analytical
skills and resources of the majority of those afforded that data. A
survey detailed in [47] found in part that while citizens acknowl-
edge and appreciate moves towards open data, most don’t know
people in their social circles who take advantage of it. Further, the
authors note “most open data released by the government is avail-
able in the raw format, which restricts its understandability by all
people” and that “this data is mostly usable by experts with some
technical knowledge to interpret and develop applications” [47].
Separately, in a case study of Data.gov, [23] argue that open data
“generates its value when it is not only available and accessible but
also made sense by its users to solve problems” and conclude that
“public agencies should invest in new technologies and craft new
data management techniques to make data readily accessible to
users...providing real-time analysis and updates.”
To date, the bridge between raw data and meaningful infor-
mation has generally been built ad-hoc and on-demand by data
scientists, but that resource-intensive approach doesn’t scale when
considering the information needs of a broader subset of the public.
And, in the space of the legal system, even questions as simple as,
“Are there differences in how judges handle fee waiver requests?” or
“Is there any correlation between a judge’s tenure and the length of
cases they oversee?” are impossible to answer without significant
data expertise or the resources to pay for it. Clearly, open data ac-
cess isn’t enough; we need a mechanism to access the information
contained within.
To build that mechanism, in essence, is to automate work that
would be done by a data scientist to extract information. Thus, we
endeavor to outline what the data scientist’s role entails and identify
those functions as requirement sets for building the platform.
1.1 The Domain Expert/Data Scientist
Interaction
One set of requirements mirrors the domain expert/data scientist
interaction: the ability to understand the user’s intent and trans-
late that into queries and analysis and to provide guidance and
guardrails around what’s possible given a dataset – and then to
translate the results of analysis back to users in a way that is intel-
ligible to them.
To help us better understand and frame the set of potential
users in the space of the U.S. court system, we conducted 28 sets of
interviews with a total of 38 people (25 male and 13 female). Some of
those interviewed included faculty in law, sociology, and economics;
lawyers; and journalists. Participants generally reported wanting
to answer advanced questions beyond their analytical capabilities.
120
Paley et al.
They felt they were limited by the tools they were currently using
and wanted to ask questions of the data that they weren’t able to.
To help bridge that gap, we introduce a novel platform and
user experience that provides users with the tools necessary to ex-
plore data and drive analysis via natural language statements. Our
approach leverages an ontology configuration that adds domain-
relevant data semantics to a database schema for the sake of sup-
porting search and analysis without user-entered code or SQL. This
configuration allows us to abstract away the underlying schema
complexities from user concern, understand what filters and analy-
sis are possible and domain-relevant, infer relevant analytics from
the data semantics, and provide guided outcomes during both search
and analysis.
The associated notebook-style experience is an early embodi-
ment of a new form of human-data, or human-information, interface
– a user experience imbued with a set of assistive capabilities where
interactions happen in natural language rather than code. The sys-
tem also generates responses in modalities intuitively appropriate
to the nature of the analysis results – from text to various types of
visualizations.
1.2 The Data Scientist/Data Interaction
The second set of requirements mirrors the data scientist/data inter-
action: the wrangling of data into coherent and controlled schemas
through various modes of ETL (extract, transform, load), text extrac-
tion, data cleaning, and the more complicated arenas of machine
learning and language modeling.
To support explorations of the U.S. court system, this includes
the structuring and harmonization of court records, with the initial
focus here on a snapshot of roughly 270,000 case dockets. This
involves consultation with domain experts; the definition of a com-
plex schema (across 30 tables ranging in size from two to thirty-one
columns); a pipeline to extract, transform and harmonize the un-
structured and semi-structured components of dockets; the integra-
tion of additional datasets to expand the information space (starting
with background information on federal judges); the creation of a
novel dataset for training language models for classification tasks
(initially for classifying various types of motions within the scope of
a case); and the model training/fine-tuning and validation process
in pursuit of proving the utility of framing motion type detection
as a classification task.
1.3 Automating Data Science
In sum, those two tracks build to our end goal: to democratize access
to information associated with the U.S. court system, eliminating
barriers to access and understanding, and providing journalists, le-
gal scholars, lawyers, government officials, social justice advocates
and others with relevant information derived from data about the
mechanics of the federal courts.
We first detail our primary novel contribution – the platform
for information exploration and analysis, and the natural language
notebook frontend – and then provide detail about the ETL and data
enrichment processes that serve as a backdrop for the search and
analyses this instantiation supports. We engage in user testing and
report preliminary results as well as explore how our platform’s
capabilities map to existing data science approaches through a case
From Data to Information: Automating Data Science to Explore the U.S. Court System
study. Our discussion elaborates on the goals of our work, including
challenges to be addressed.
Our approach to court docket search and analysis is one early
step in the development of an open-source platform aimed at de-
mocratizing access to information. In discussion of future work,
we outline dual and distinct tracks: the first aimed at continuing to
build and augment our U.S. court records database, and the second
focused on the ongoing development of the core platform. On the
platform side, we point to a future in which additional data can
be brought in by technical users who manage data wrangling and
define data semantics – the steps we now think of as getting to
“open data,” but with a newly imagined purpose – and our system
scales to new domains, communities, and geographies.
2 RELATED WORKS
Reducing the costs associated with PACER has been pursued as a
way to achieve judicial transparency. However, studies have shown
the limits of open data in providing greater transparency [44].
Notably, problems persist across many user personas, from citi-
zens to data scientists, government agents, and even academics
[7, 18, 19, 23]. We aim to address a subset of these challenges –
pertaining to data utility and barriers to information access – by
applying automated analytical and visualization capabilities on top
of the data.
Much research has focused on automating legal processes [28],
predicting outcomes [5], or assessing the value of AI for the two
former areas [13, 45]. There have been some recent developments
in legal question-answering (QA) systems [11, 21]. However, these
have had limited data analytics capabilities [22] and often rely
on simple data retrieval for generating answers [33]. While some
commercial tools support exploration of court documents they are
prohibitively expensive, and limited in terms of scope and consis-
tency of results [1].
More broadly, general QA systems have been the subject of re-
search for decades [16, 40, 46] and are some of the most prominent
examples of AI systems [12]. There has been significant progress
in neural QA systems [10], with transformer-based models [27]
achieving state-of-the-art results on benchmark tasks [39]. How-
ever, these QA systems are best suited for unstructured text data
where the answer is plainly stated in the corpus itself, unlike our
system which can infer or derive the answer through follow-on
analysis. Other approaches aim to understand and decompose the
structure of complex questions into discrete parts as a plan for
deriving an answer [49]; however, the representation is high-level
and distinct from our approach which constructs runnable queries
against a given datasource.
Extensive research has parsed natural language queries into SQL
queries [20], using techniques from deep learning [17], rules-based
methods [42], or a mixture of both [43]. Instead, our approach
automatically generates the space of possible analysis from an
ontology configuration, and then translates the underlying analysis
plans to natural language, drawing inspiration from prior work
[37, 41].
Beyond current work in information retrieval via conversational
systems, our approach utilizes a notebook-style interface, with
inspiration coming from Jupyter notebooks [38] as well as their
121
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
forebear, IPython [36]. Automated visualization is a related area of
research [31, 50] focusing primarily on presentation layers for a
given dataset rather than intent-driven question-answering.
3 THE NATURAL-LANGUAGE NOTEBOOK
Notebook-style interfaces are a standard part of the modern data
science toolkit, and for good reason: they support a logical process
flow and marry exploratory and presentation layers in one cohesive
experience. However, they are the tools of experts – users who
know how to code, run analysis, interpret stack traces and explore
complex results. They bring order in the form of scaffolding, but
remain largely agnostic about content or the specifics of a particular
dataset or domain.
We borrow from that scaffolding, but our system leverages sim-
plified data filtering mechanisms and natural language statements.
And where other notebooks display a variety of outputs (defined by
the near-infinite space of possibility supported by arbitrary code),
our system outputs natural language and annotated visualization
as a means of conveying information.
This approach maintains the intuitive flow of the notebook user
experience but brings its power to people unfamiliar with program-
ming. Our notebooks are domain- and dataset-aware, and the user
experience speaks the language not of the data scientist, but of a
user reasonably fluent in the domain. Further, they provide assis-
tive mechanisms to surface what the system knows it is capable of,
guiding even novice users to understand the range of capabilities
available to them.
An exploration of the current iteration of interface mechanisms
and output capabilities can be found in Figure 1 and Figure 2, and a
deeper discussion of the paradigm follows.
3.1 The UX Paradigm: “Search First, Then
Converse”
Our approach separates concerns between search (winnowing the
available dataset to a space of interest) and converse (the user in-
putting statements that drive analysis upon the filtered dataset and
the generation of responses). This approach embodies the strengths
of the notebook format in focusing on one task at a time and pre-
senting interstitial output as feedback. Further, this has the indi-
rect effect of separating concerns on the backend, supporting a
generalizable approach to the specification of filter and analysis
configuration.
As depicted in Figure 1, each exploration in our notebook inter-
face starts from a “search” (or “filter”) panel: a paginated view of the
dataset that matches the current set of filters. The primary entities
presented in this view are court cases in the Northern District of
Illinois. On initialization, the user is presented with the full space
of available data absent any applied filters and can opt to apply
filters or skip right to adding analyses of the full dataset.
Below that is the partitioned “converse” step, where the user can
enter natural-language statements that drive analysis (Figure 2).
Of note, users can enter multiple analysis statements against one
data view, stepping through a set of questions while maintaining a
thread of prior exploration.
This paradigm means our system does not have to manage state-
ments like “Average case duration grouped by judge tenure for cases
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paley et al.

Figure 1: The primary interface for search/filtering within

our notebook format. Annotations: 1) Notebook title, 2) Ap- Figure 2: The primary interface for running analysis within
plied filter example, 3) Mechanism for adding filters, 4) Re- our notebook. Annotations: 1) Applied filter example, 2) Col-
sults interface with sortable columns, 5) Pagination and full lapsed results panel (seen in detail in Figure 1), 3) Analy-
results space, 6) Download button for raw data access, 7) sis statement, 4) Mechanism for removing previous analysis
Mechanism for adding analysis output, 5) Result of analysis (basic NLG type to deliver sin-
gle value), 6) Second analysis statement, 7) Result of analysis
(interactive line chart with rollover states to display change
over time, including legend with terms of art and associated
that occurred in the Northern District of Illinois since February definition), 8) Mechanism for adding additional analysis
2015 and involved property rights,” (or require the user to repeat
parts of this cumbersome statement for additional analysis), but
instead simple, widely applicable statements like the “average case
duration grouped by judge tenure” in the context of a previously
filtered set of data. Thus, given an ontology and available analyt-
ics, the space of analysis statements is finite, but is made virtually
infinite by the possibility of applying them to any slice of the data.

3.2 Mechanisms and Underlying Configuration

Necessary elements are abstracted out of the platform’s core search
and analysis engines as well as the user experience framework, API
mechanics, and proactive caching and pre-fetching mechanisms.
The system requires only a pointer to an SQL database, an object-
relational mapping (ORM) defined in the open-source SQLAlchemy
library [34], and an ontology configuration that references that Figure 3: A simplified architecture diagram to illustrate the
ORM in order to provide all functionality, from generating the relationship between the different components of the plat-
available filters and analysis statements to building queries and form and the dataset-specific Ontology Configuration and
running analytics based on user input. These capabilities mean ORM
that additions to the underlying data schema or the onboarding of
complementary datasets can be made available through the plat-
form with only a small addendum to the already-required work 3.2.1 Search/Filter. As depicted in Figure 1, filters are applied by 1)
of data management: the creation of or updates to the ontology adding them to the filter bar above a given data view, 2) selecting a
configuration. See Figure 3 for a high-level architecture; specifics filter type via the dropdown, and 3) entering values in the associated
about the configuration follow for both the search and analysis input. The set of additive filters that can be applied to winnow down
components. the list of case dockets includes: district, circuit, case name, cause

122
 From Data to Information: Automating Data Science to Explore the U.S. Court System
of action, case status, filing date, nature of suit, party name, judge
name, attorney name, as well as free text search in the docket entries
associated with the case. Ultimately, users can make a few targeted
selections and fill in a few inputs to get to searches equivalent to
“all cases in the Northern District of Illinois between 2015 and 2017
where Kennelly served as judge” or “all cases with nature of suit
property rights where one of the parties is Apple” – the SQL query
versions of which only a fraction of those users could generate
themselves.
The focus on case dockets being the “primary” searchable unit (as
opposed to judges, parties, attorneys, etc.) and all associated filters
are entirely configuration-driven and distinct from complexities
of the underlying schema. The ontology config maps the machine
representation to a user-friendly set of names and attends to the
scaffolding of ids, foreign keys and joins. Thus, the filterable fields
are a subset of those that exist at the schema level on various
tables that join against the case table, and in some cases (such as
Judge Name, depicted in Figure 4) actually span multiple fields in
the schema (first_name, middle_name, last_name). The key point
here is that the user does not need to consider the schema but
simply makes decisions about domain-relevant ways to search with
guidance from the system about the relevant search space (and the
system then generates runnable queries of various types, including
string matching and range finding, such as with dates). For domain
expert users, our approach is a significant convenience over having
to learn or write SQL, and, for less knowledgeable users, it also
serves as guidance about relevance in the domain.
1 {... ,
2 " judgeName " : {
3 " nicename " : " Judge Name "
4 " type ": " text " ,
5 " allowMultiple " : True ,
6 " autocomplete " : acs . getJudges ,
7 " model " : [ db . JudgeOnCase , db . Judge ] ,
8 " fromTargetModel " : [ " judges " , " judge " ] ,
9 " fields " : [" first_name " , " middle_name " , "
last_name " ] ,
10 },
11 ... }
Figure 4: The config entry for the “Judge Name” entity for
search/filter capabilities and the results view. 1) “nicename”
is the user-facing name of this entity type, 2) “type” and "al-
lowMultiple" inform the input style and query generation
mechanisms, 3) “autocomplete” maps to a method on the
autocomplete class (can be default or a plugin) and powers
the autocomplete API endpoint, 4) “model” and “fromTar-
getModel” map the model join and relationship feature path
from the db.Case table at the ORM level, 5) “fields” defines
the field(s) this entity’s name/id maps to (affording support
for multi-field queries)
3.2.2 Analysis Statements and Query Generation. Once the user
has arrived at filtered data they are interested in, they can add
multiple analysis statements below the dataview panel. As seen
123
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 5: Two examples of the primary interface for adding
analysis statements. Annotations: 1) The user-entered state-
ment, having been auto-completed progressively via gener-
ated statement candidates, 2) Candidate matches given the
previously auto-completed statement (“Average Fee Waiver
Grant Rate Year-Over-Year”) and the subsequently user-
appended “grouped by,” 3) A user-entered string that hasn’t
yet been auto-completed, demonstrating fuzzy string match-
ing, 4) A set of fuzzily matched results
in Figure 5, this is realized on the UX as a fuzzy (i.e., approximate
string matching) search across a set of natural language statements,
each of which is generated dynamically by the system through
inferring relevant analysis possibilities based on the underlying
ontology configuration and a core model of analysis types. Because
the system is inferring and defining the analysis space based on the
ontology components, each generated statement corresponds to an
underlying plan representation that is interpretable by the analysis
engine for the sake of generating queries and running analytics on
any set of filters.
As users select and add additional analysis statements to the
notebook, the system responds with answers in the form of text
and visualizations, as depicted in Figure 2. As per standard notebook
mechanics, each analysis statement is tied to the active filtered set
in the panel above such that changes to the filters (and thus the slice
of the data presented) will flow through and update each linked
result.
The core platform’s model of analysis includes a growing set
of available operations (e.g., average), as well as specifications on
how the operation ought to be performed (e.g., can only be done
on numeric fields, how many fields are needed). As seen in Figure
6, the ontology configuration then defines the fields relevant for
analysis and their user-friendly names, as well as their attributes
(e.g., semantic type, possible transformations into other data types,
relevant units, and – in the case of discrete entities delimited by id
– how to generate their user-friendly names) and their relationship
to the primary model.
To illustrate the generation of the analysis space, in the instanti-
ation referenced in our figures, an analysis configuration that lists
ten relevant features for analysis (e.g., Judge Tenure, Nature of Suit,
Case Duration, Fee Waiver Grant Status) alongside the metadata
 ICAIL’21, June 21–25, 2021, São Paulo, Brazil
1 {... ,
2 " caseDuration " : {
3 " model " : db . Case ,
4 " field " : " case_duration " ,
5 " type ": " float " ,
6 " name ": [ " Case Duration " , " Case Durations "
],
7 " unit ": [ " day " , " days "] ,
8 },
9 ... }
Figure 6: The config entry for the “Case Duration” feature
as an analysis target. 1) “model” and “field” define where in
the schema the relevant field(s) exist, 2) “type” informs the
available analyses for that field, 3) “name” and “unit” define
the singular and plural forms for user interaction/presenta-
tion.
associated with each is sufficient to generate 120 different possible
analyses, each of which can be applied to any filtered view of the
data. Augmenting this list of generated analysis statements requires
simply adding new elements to the configuration or augmenting
the core system’s analytics library and tying new analytics to data
or semantic types.
To compute an analytics statement at runtime, our system steps
through the process of building an analysis chain and SQL queries
based on the filters and the statement’s underlying plan. The steps
to do so are: 1) Do any necessary filtering (as specified by the
search context the given analysis statement is run in); 2) Query
the necessary fields from the analytics statement; 3) If needed,
transform data to ensure data type compatibility; 4) Perform the
necessary operations and grouping; and 5) Format the results based
on the nature of the information to be conveyed.
For instance, if we wanted to know the "Average Fee Waiver
Grant Rate Grouped by District" for cases where the Nature of Suit
is “Property Rights,” the steps would be: 1) We filter to get only
cases where "Property Rights" is listed in the nature of suit field
(leveraging the ontology mapping to the schema); 2) We query the
fields associated with Fee Waiver Grant and Court District (again
leveraging the ontology mapping); 3) Since we are taking an average
rate and Fee Waiver Grant is stored as booleans, we convert it into
integers; 4) We compute the average rate of Fee Waiver Grants
for each Court District; 5) We convert the internal database Court
District id into human-readable labels leveraging the name and
units information from the ontology config.
Finally, the system output – the result of running the analysis
statements above – is delivered in a form best-suited to conveying
the nature of the information for each result (and includes a de-
scription of the filters applied for added clarity). This is keyed off of
features of the results themselves. As depicted in Figure 2, running
“Average Case Duration” – a bit of analysis that will yield a single
value – results in the system rendering the results via basic natural
language generation mechanics. However, when looking at some-
thing “Year-Over-Year,” the system pivots into change-over-time
behavior, leading to the generation of an interactive line chart.
124
Paley et al.
4 THE DATA PROCESSING PIPELINE
PACER, the official source for federal judicial records, houses a
variety of document types and charges a per-page fee for access.
Following the recommendations set forth in [35], we focused on
the docket reports, “essentially a lawsuit’s table of contents.” [35]
identified efforts to improve accessibility of these docket reports as
the “most impactful” work to be done in building a more open justice
system. Thus, using docket reports as our primary data source, we
designed a 30-table database schema (plus relevant join tables) to
represent them and all relevant entities (e.g., judges, attorneys,
defendants, and districts) as well as key fields (e.g., nature of suit,
date of filing, and docket entries).
Our initial dataset captures samples of both depth (ten years
of docket reports from Northern Illinois district courts from 2007
to 2016) and breadth (docket reports from every district court in
2016). In total, our sample draws from more than a quarter-million
case dockets in HTML format acquired through purchase and batch
downloading from PACER. Taking advantage of their semi-regular
structure, we parsed the files into meaningful sections (e.g., the
docket header) and extracted information. While most informa-
tion we extracted was listed explicitly on the docket report (e.g.,
case title), we also captured implicit (e.g., case duration) and in-
terpreted (e.g., party acting as their own attorney) information.
In this latter case, we relied on guidance from domain experts to
analyze both docket text (e.g., recognizing “Pro Se” designations
for attorneys) and docket contents (e.g., verifying instances of self-
representation where a party has only one attorney, who is also the
party). Such interpretation often required triangulation between
multiple approaches (e.g., a “Pro Se” designation alone is insufficient
for classifying self-representation; one must also count and check
the attorneys).
5 FURTHER DATA ENRICHMENT
To garner a more complete data-level representation of the mechan-
ics of the judicial system, we sought to enrich the initial core docket
dataset in two distinct ways: blending additional sources and the
use of language modeling for classification tasks.
5.1 Additional Sources
We blended additional data into the core database, serving both
as supplemental fodder for search and analysis and in support of
initial forms of entity disambiguation. We leveraged the Federal
Judicial Center’s database of appointed federal judges [6], which
includes birthdate, gender, race/ethnicity, history of appointments,
appointing parties, education, and professional career. We normal-
ized the data into a multi-table schema, linked it to the extracted
representations of judges, and then leveraged that join to expand
the space of available analysis. For example, based on the judge’s
appointment date and the start of a given case, we derived how
long a judge was on the bench prior to the start of that case, and
then used that “judge tenure” as a metric in subsequent analysis
(e.g., to derive “fee waiver grant rate grouped by judge tenure”).
5.2 Classification Tasks
We identified a variety of information targets that we believe can
be culled from the unstructured text components of dockets by
From Data to Information: Automating Data Science to Explore the U.S. Court System
reframing them as classification tasks. For context, the main body
of a docket is a series of time-stamped text entries, each marking
events in the arc of a given case. These text-snippet representations
contain various sorts of useful information, including motions (ef-
fectively discrete requests for a judicial decision), the outcome of
a given motion, changes of representation or venue or presiding
judge, references to evidence or testimony, eventual outcomes, and
so on. Being able to identify and classify such information would
prove highly valuable for both search and analysis.
To explore approaches, we started with the classification of mo-
tion types as our initial target. At first glance, it could be tempting
to envision a solution to this classification task based on regular
expression where motions are explicitly identified by name. How-
ever, as depicted in Table 1, a pure regex approach is far too rigid
to capture the many complexities found in the docket entry sample
space, including multiple motions being named in a single entry,
non-motion events referencing motions by name (e.g., notices, or-
ders), and obfuscation of the motion type through varying levels of
docket entry metadata. These complexities are further compounded
by naming convention variations across districts and the trappings
of error-prone human data entry [1].
Thus we pivoted to language modeling. As no training dataset
exists for such a task, we created a web application to view and tag
the motions pulled from our docket dataset. For both the definition
of the space of possible motion types and for the sake of actually
tagging the motion entries, we solicited help from legal scholars and
their law students. We implemented a voting mechanism in the app
such that each motion will be tagged three times by three distinct
users as a means of ensuring accuracy. Our dataset continues to
grow through use of the application, though the experiments that
follow leverage a subset of this data.
In order to effectively utilize this data for our classification ex-
periments, we performed some preprocessing on the raw dataset.
First, the raw dataset contained several motion classes with few
data points. To address these rare motion classes in these initial
tests, we set a threshold of 25 data points and merged all classes
below this threshold into the “Other Motion” class. Second, we re-
moved all duplicate docket entries that arose as a byproduct of the
voting mechanism from the dataset to ensure that the models were
not training on some docket entries more than others. After this
preprocessing, the smallest motion class contained 25 samples, the
largest contained 951, and the median and average of the motion
classes were 50 and 152, respectively. For each of the models we
used a train/validation/test split of 80/10/10 per class. In total after
preprocessing, there were 2,064 training samples with 524 testing
samples across 17 distinct motion classes.
Making use of two pretrained transformers, the 110M parameter
BERT-base [9] and 125M parameter RoBERTa [29] models, we fine-
tuned each on this processed dataset. We made use of the AllenNLP
framework [15] as a wrapper around the Huggingface Transformers
library [48] to fine-tune the models for 10 epochs, using a batch
size of 8, and the AdamW optimizer. The RoBERTa model achieved
training accuracy of 95.69%, validation accuracy of 91.22%, and
test accuracy of 90.08%. The BERT-base model achieved training
accuracy of 96.95%, validation accuracy of 89.69%, and test accuracy
of 89.31%. These results exceeded the baseline bag of embeddings
classification model, which achieved a test accuracy of 80.50%. For
125
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the transformer models, the training accuracy is slightly higher
than the validation/test accuracy for both models, but we believe
this margin is reasonable given the small size of the dataset. To
reduce the likelihood of overfitting in the future, we continue to
grow the tagged motion dataset.
6 EVALUATION
We evaluated our system’s effectiveness in handling both search
and analysis of data across two separate tracks: 1) usability testing
in which target users completed tasks with the system and provided
survey feedback, and 2) a case-study comparative analysis to assess
the system’s efficacy when benchmarked against a data scientist’s
ad hoc analysis.
6.1 Usability Testing
We gave 15 subjects (14 legal professionals, one journalist) a set
of prompts (e.g., “For all cases in the ’N.D. IL’ district, which year
had the highest average case duration?”) and assessed their experi-
ences in: (1) using the search filter, (2) conducting analysis on all
the records, and (3) conducting analyses based on specific search
criteria of varying complexity. In addition, we gave them time to
test their own scenarios while “thinking aloud” so we could capture
their intentions and strategies. Participants were then presented
with a survey to complete at the end of the session consisting of the
modified System Usability Scale (SUS), an evaluation framework
shown to be effective at quantifying the complexity and ease of use
of interfaces [4]. The average SUS score for our participants’ overall
experience across (1), (2), and (3) was 72.83, which is considered
good usability [3, 26]. When answering the statement regarding
whether they would use our system frequently, all but two partic-
ipants (87%) agreed or strongly agreed with that statement (one
was neutral and one disagreed and wrote that docket sheets are not
used in their research). These results represent a preliminary round
of user testing, and we intend to further analyze the associated
feedback and conduct additional user tests targeting users with a
wider variety of backgrounds.
6.2 System Evaluation: A Case Study
To further weigh the benefits of our approach, we compare it with
prior work done by data scientists examining how fee waiver grant
rates vary among judges using ad hoc data processing and analysis
[35]. We answer the same question using our system (see Figure
7 for an example of the output) and through observation compare
both approaches across three dimensions: speed to insight, flexibil-
ity of exploration, and barrier to entry.
The initial data processing pipeline looks similar for both. A
systematic analysis of this issue requires paying to download case
documents, creating an ETL process to structure the data, and
identifying the fee waiver status of each case [35]. However, where
the ad hoc method attends to ETL, aggregation, and visualization
for a single target task, our approach looks to leverage that upfront
data work to support a wide array of possible downstream analyses.
Thus, when considering a one-off query or single data point, we
cannot definitively say that the ETL, schema and ontology work in
support of our system will require less time than a data scientist
taking the ad hoc approach. But one-offs aren’t the goal of our
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Paley et al.

Motion Entry Motion Class

MOTION by Defendant [Name Omitted] for extension of time to file respon- Motion for Extension
se/reply as to motion for summary judgment 20 (Unopposed) ([Name Omitted])
(Entered: 05/18/2017)
Proposed Order re 13 MOTION to Stay Proceedings Pending Transfer by [Name Not a Motion
Omitted]. ([Name Omitted]) (Filed on 4/16/2012) [Transferred from California
Northern on 4/18/2012.] (Entered: 04/16/2012)
Motion by [Name Omitted], Cook County Board of Review, [Name Omitted] Motion for Leave
for Leave to Cite Additional Authority ([Name Omitted]) (Entered: 01/13/2011)
Plaintiff’s motion for leave to file a first amended complaint 54 is granted. Not a Motion
WRITTEN Opinion entered by the Honorable George W. Lindberg on 4/29/2011:
Signed by the Honorable George W. Lindberg on 4/29/2011:Mailed notice(pm, )
(Entered: 04/29/2011)
Table 1: Sample motion entries illustrating some complexities of classifying motion types.

platform and what matters to us is the speed to information for our

end users – and once the setup of our system is complete, speed
from question to information for end users will be much faster and
easier on a per-query basis. Thus, our system demonstrates a clear
advantage given that the cost of defining a configuration can be
amortized over every question answered.
On flexibility of exploration, we consider the state of things once
the fee waiver question is answered. For the ad hoc approach, we
have access to the information needed to answer the question at
hand, but any subsequent question or filter amendment requires
fresh code and additional work on the part of the data scientist. In
contrast, once a configuration has been defined, our system sup-
ports running new types of analysis or perhaps the same analysis on
different slices of the data (e.g., “How do fee waiver grant rates vary
among judges in the Northern District of Illinois vs the Southern
Figure 7: The system’s response to the analysis statement
District of Illinois?”) – without a data scientist in the loop.
“Average Fee Waiver Grant Rate Grouped by Judge” in a fil-
Last, and we believe most important, is the barrier to entry.
tered data context. Each bar represents a given judge’s fee
Regardless of time or effort, there will always be those who lack
waiver grant rate – judge names are available in rollover
the technical skills or resources necessary to convert data into
states (not depicted).
information. Since most are not data scientists and cannot afford to
hire one, the ad hoc approach doesn’t scale. In contrast, our system
provides a path to minimize the barrier to entry by abstracting
away these technical skills through a one-off, upfront setup. This To reach those goals, we must automate. We need systems that
decouples the data scientist from the exploration of data and by bridge the gap between questions and answers – routinely employ-
doing so democratizes access to the underlying information. ing individual data scientists to serve that function simply doesn’t
scale. The ultimate objective is repositioning as much of the bur-
den of that complexity on the machine as possible, and our work
7 DISCUSSION here is a step in that direction. And while our approach certainly
Though in its early stages, our work already demonstrates signif- won’t eliminate the role of data scientists, we believe a significant
icant promise. User testing among legal scholars, attorneys and amount of analysis can be standardized, systematized and auto-
journalists in the U.S. confirms both the value and usability of our mated, bringing access where previously there was none for lack
system. Beyond that, our approach has the potential to generalize of expertise or resources. And of note, where access does already
well to a wide array of data sources, providing a new platform for exist, our approach has the potential to free data scientists from
the democratization of information across communities, sectors some of the repetitive “query generation” aspects of their roles,
and geographies. We see broad opportunities for such an approach affording them more time to drive novel exploration. In some sense
in the space of open government data, a domain rife with available the approach detailed here could scale their expertise, allowing
datasets but with chronic challenges in terms of accessibility and them to teach our platform about their datasets and then offload
use [18, 23]. This is a push towards the realization of the true goal stakeholder questions to the system whenever possible.
of such initiatives: from code to content and from open data to open That said, we see challenges in this approach to the U.S. court
information. system, and anticipate them in scaling to new domains as well:

126
From Data to Information: Automating Data Science to Explore the U.S. Court System
• Issues of ethics and responsibility: One such example is
privacy. Court documents are rife with personally identifi-
able information, and reliably de-identifying documents at
scale is a non-trivial problem. Further, the tension between
de-identification and information completeness (say, for the
sake of mapping to geographies) adds another complication.
The use of highly regulated medical records data in research
and machine learning provides a promising precedent [8, 25]
for reference as we move forward.
• Issues of information misuse: Protecting against misuse
of analysis, especially when the barrier of expertise to arriv-
ing at such analysis has been lowered, is a significant issue
in our increasingly fraught information landscape. In the
realm of law, the politicization of judicial decision making or
the use of judicial analytics as a means of influencing future
outcomes are both potential issues.
• Issues of explainability and data quality: Our scalable
approach to data analysis adds a new layer of importance to
the explainability of results and also runs the risk of obscur-
ing incomplete or deficient data. To fully realize the promise
of data science automation, additional research will be fo-
cused on ensuring our system can explain itself and handle
issues of data quality gracefully and transparently.
• Issues associated with novel analysis: Inarguably, data
scientists can flexibly address novel questions or analysis
requirements on the fly, and while our platform’s library
of analytics will grow, there will continue to be question
types it can’t answer. In future work, we will expand our
nascent plugin framework to support custom analysis and
continuously grow the built in libraries.
8 FUTURE WORK
Going forward, various members of our team are pursuing in tan-
dem the dual roadmap we laid out in the introduction.
One thread is aimed at making the raw data emitted from the
U.S. court system increasingly machine-readable. This entails ev-
erything from the continued evolution of the ingestion pipeline
(sourcing data from a wider variety of districts and tackling corner
cases in the data) to improvements to the data already obtained
through various forms of enrichment. In the near term, we intend
to pursue entity disambiguation on parties and attorneys, as well
as the creation of additional datasets to train language models for
classification outside the scope of the motions described above
(such that we can attempt to capture additional data points such
as judicial rulings, charge severity, changes in representation, and
various forms of case outcome).
The other thread is the work with the core platform itself. This
will take a number of forms, including: 1) Expansions to the ana-
lytics capabilities and plugins (including the introduction of new
response types and visualizations); 2) An evolution of the ontology
configuration and support for ontology management through the
user experience, allowing for user-driven updates as well as the
introduction of new data sources; 3) Ontology-driven derived fields,
providing support for adding new data points dynamically and intro-
ducing new possibilities for downstream explanations; 4) Support
for localization such that the platform could be used by non-English
127
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
speakers (of note, our ontology-driven approach means very little
actual language is coded into the UI, making this an easier pursuit),
opening up the possibility of legal documents and open data from
other countries being made available through the platform; 5) UX
improvements, including changes to analysis statement selection
(with fuzzy semantic matching on colloquial terms against terms of
art), support for more interactivity in visualizations, and additional
explanations associated with analysis results; 6) Support for inter-
active machine learning by bringing the capabilities of our separate
motion tagging application directly to the platform and augmenting
them to cover both the extraction/creation of novel tagged datasets
and in-platform model training/fine-tuning, validation and testing.
This presents a significant opportunity for research in the space of
making machine learning accessible to non-technical users.
While we believe deeply in the importance of bringing trans-
parency to the U.S. court system and will continue the data work
necessary to do so, we also see this data-information rift throughout
the government and public sector in the United States and globally.
Thus, we are excited by the prospect of platform improvements to
support bringing a variety of new datasets to our application.
9 CONCLUSION
In this work we’ve detailed a novel platform and user experience to
allow non-data scientists to drive exploration and analysis of data
associated with the U.S. Court system. In support of that experi-
ence, we defined the process by which we ingested, extracted and
structured the data from 270,000 case dockets. Given the results of
usability testing presented in our evaluation, we believe we have
early confirmation that this new natural language notebook ap-
proach marks a step in the direction of democratizing access to data
analysis and could have significant impact not only in the space
of the U.S. court system, but also more broadly across a variety of
publicly available data. Subsequent work is already underway to
further develop the capabilities, refine the UX mechanics, and stand
up new components of the ecosystem. In tandem, the ingestion,
structuring and enrichment of U.S. court records continues as we
work towards a comprehensive database mirroring the federal court
system.
ACKNOWLEDGMENTS
This material is based upon work supported by the National Sci-
ence Foundation Convergence Accelerator Program under grant
no. 1937123 and grant no. 2033604.
REFERENCES
[1] Charlotte Alexander and Mohammed Javad Feizollahi. 2019. On Dragons, Caves,
Teeth, and Claws: Legal Analytics and the Problem of Court Data Access. Com-
putational Legal Studies: The Promise and Challenge of Data-Driven Legal Research
(Ryan Whalen, ed., Edward Elgar, 2019, Forthcoming) (2019).
[2] Judie Attard, Fabrizio Orlandi, Simon Scerri, and Sören Auer. 2015. A systematic
review of open government data initiatives. Government Information Quarterly
32, 4 (2015), 399–418.
[3] Aaron Bangor, Philip Kortum, and James Miller. 2009. Determining what indi-
vidual SUS scores mean: Adding an adjective rating scale. Journal of usability
studies 4, 3 (2009), 114–123.
[4] Aaron Bangor, Philip T Kortum, and James T Miller. 2008. An empirical evaluation
of the system usability scale. Intl. Journal of Human–Computer Interaction 24, 6
(2008), 574–594.
[5] Karl Branting, Brandy Weiss, Bradford Brown, Craig Pfeifer, A Chakraborty, Lisa
Ferro, M Pfaff, and A Yeh. 2019. Semi-supervised methods for explainable legal
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
prediction. In Proceedings of the Seventeenth International Conference on Artificial
Intelligence and Law. 22–31.
[6] Federal Judicial Center. 2011. Biographical directory of federal judges.
[7] Jonathan Crusoe, Anthony Simonofski, Antoine Clarinval, and Elisabeth Gebka.
2019. The impact of impediments on open government data use: insights from
users. In 2019 13th International Conference on Research Challenges in Information
Science (RCIS). IEEE, 1–12.
[8] Franck Dernoncourt, Ji Young Lee, Ozlem Uzuner, and Peter Szolovits. 2017.
De-identification of patient notes with recurrent neural networks. Journal of the
American Medical Informatics Association 24, 3 (2017), 596–606.
[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
Pre-training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805 (2018).
[10] Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and
Michael Auli. 2019. ELI5: Long Form Question Answering. In Proceedings of the
57th Annual Meeting of the Association for Computational Linguistics. 3558–3567.
[11] Biralatei Fawei, Jeff Z Pan, Martin Kollingbaum, and Adam Z Wyner. 2018. A
methodology for a criminal law and procedure ontology for legal question answer-
ing. In Joint International Semantic Technology Conference. Springer, 198–214.
[12] David Ferrucci, Eric Brown, Jennifer Chu-Carroll, James Fan, David Gondek,
Aditya A Kalyanpur, Adam Lally, J William Murdock, Eric Nyberg, John Prager,
et al. 2010. Building Watson: An overview of the DeepQA project. AI magazine
31, 3 (2010), 59–79.
[13] Anthony W Flores, Kristin Bechtel, and Christopher T Lowenkamp. 2016. False
positives, false negatives, and false analyses: A rejoinder to machine bias: There’s
software used across the country to predict future criminals. and it’s biased
against blacks. Fed. Probation 80 (2016), 38.
[14] World Wide Web Foundation. 2018. Open Data Barometer - Leaders Edition. World
Wide Web Foundation.
[15] Matt Gardner, Joel Grus, Mark Neumann, Oyvind Tafjord, Pradeep Dasigi, Nelson
H S Liu, Matthew E. Peters, Michael Schmitz, and Luke Zettlemoyer. 2017. A
Deep Semantic Natural Language Processing Platform.
[16] Bert F Green Jr, Alice K Wolf, Carol Chomsky, and Kenneth Laughery. 1961.
Baseball: an automatic question-answerer. In Papers presented at the May 9-11,
1961, western joint IRE-AIEE-ACM computer conference. 219–224.
[17] Srinivasan Iyer, Ioannis Konstas, Alvin Cheung, Jayant Krishnamurthy, and Luke
Zettlemoyer. 2017. Learning a neural semantic parser from user feedback. arXiv
preprint arXiv:1704.08760 (2017).
[18] Maxat Kassen. 2018. Adopting and managing open data: Stakeholder perspec-
tives, challenges and policy recommendations. Aslib Journal of Information
Management (2018).
[19] Muhammad Mahboob Khurshid, Nor Hidayati Zakaria, Ammar Rashid, and
Muhammad Nouman Shafique. 2018. Examining the Factors of Open Govern-
ment Data Usability From Academician’s Perspective. International Journal of
Information Technology Project Management (IJITPM) 9, 3 (2018), 72–85.
[20] Hyeonji Kim, Byeong-Hoon So, Wook-Shin Han, and Hongrae Lee. 2020. Natural
language to SQL: Where are we today? Proceedings of the VLDB Endowment 13,
10 (2020), 1737–1750.
[21] Mi-Young Kim, Randy Goebel, and S Ken. 2015. COLIEE-2015: evaluation of
legal question answering. In Ninth International Workshop on Juris-informatics
(JURISIN 2015).
[22] Mi-Young Kim, Ying Xu, and Randy Goebel. 2014. Legal question answering using
ranking svm and syntactic/semantic similarity. In JSAI International Symposium
on Artificial Intelligence. Springer, 244–258.
[23] Rashmi Krishnamurthy and Yukika Awazu. 2016. Liberating data for public value:
The case of Data. gov. International Journal of Information Management 36, 4
(2016), 668–672.
[24] Karim R Lakhani, Robert D Austin, and Yumi Yi. 2002. Data. gov. Harvard
Business School.
[25] Joffrey L Leevy, Taghi M Khoshgoftaar, and Flavio Villanustre. 2020. Survey on
RNN and CRF models for de-identification of medical free text. Journal of Big
Data 7, 1 (2020), 1–22.
[26] James R Lewis. 2018. The system usability scale: past, present, and future. Inter-
national Journal of Human–Computer Interaction 34, 7 (2018), 577–590.
[27] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman
Mohamed, Omer Levy, Ves Stoyanov, and Luke Zettlemoyer. 2019. Bart: Denoising
sequence-to-sequence pre-training for natural language generation, translation,
and comprehension. arXiv preprint arXiv:1910.13461 (2019).
[28] Tomer Libal and Matteo Pascucci. 2019. Automated reasoning in normative
detachment structures with ideal conditions. In Proceedings of the Seventeenth
International Conference on Artificial Intelligence and Law. 63–72.
[29] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A
robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692
(2019).
[30] Lynn M LoPucki. 2001. Politics of Research Access to Federal Court Data. Tex. L.
Rev. 80 (2001), 2161.
128
Paley et al.
[31] Jock Mackinlay, Pat Hanrahan, and Chris Stolte. 2007. Show me: Automatic
presentation for visual analysis. IEEE transactions on visualization and computer
graphics 13, 6 (2007), 1137–1144.
[32] Peter W Martin. 2018. District Court Opinions That Remain Hidden Despite a
Long-Standing Congressional Mandate of Transparency-the Result of Judicial
Autonomy and Systemic Indiffernece. Law Libr. J. 110 (2018), 305.
[33] Gayle McElvain, George Sanchez, Sean Matthews, Don Teo, Filippo Pompili,
and Tonya Custis. 2019. WestSearch Plus: A Non-factoid Question-Answering
System for the Legal Domain. In Proceedings of the 42nd International ACM SIGIR
Conference on Research and Development in Information Retrieval. 1361–1364.
[34] Michael Bayer. [n.d.]. SQLAlchemy. https://www.sqlalchemy.org/
[35] Adam R Pah, David L Schwartz, Sarath Sanga, Zachary D Clopton, Peter DiCola,
Rachel Davis Mersey, Charlotte S Alexander, Kristian J Hammond, and Luís
A Nunes Amaral. 2020. How to build a more open justice system. Science 369,
6500 (2020), 134–136.
[36] Fernando Pérez and Brian E Granger. 2007. IPython: a system for interactive
scientific computing. Computing in science & engineering 9, 3 (2007), 21–29.
[37] Antonella Poggi, Domenico Lembo, Diego Calvanese, Giuseppe De Giacomo,
Maurizio Lenzerini, and Riccardo Rosati. 2008. Linking data to ontologies. In
Journal on data semantics X. Springer, 133–173.
[38] Min Ragan-Kelley, F Perez, B Granger, T Kluyver, P Ivanov, J Frederic, and M
Bussonnier. 2014. The Jupyter/IPython architecture: a unified view of computa-
tional research, from interactive exploration to communication and publication.
AGUFM 2014 (2014), H44D–07.
[39] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016.
SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceed-
ings of the 2016 Conference on Empirical Methods in Natural Language Processing.
2383–2392.
[40] Deepak Ravichandran and Eduard Hovy. 2002. Learning surface text patterns for
a question answering system. In Proceedings of the 40th Annual meeting of the
association for Computational Linguistics. 41–47.
[41] Mariano Rodriguez-Muro, Roman Kontchakov, and Michael Zakharyaschev. 2013.
Ontology-based data access: Ontop of databases. In International Semantic Web
Conference. Springer, 558–573.
[42] Diptikalyan Saha, Avrilia Floratou, Karthik Sankaranarayanan, Umar Farooq
Minhas, Ashish R Mittal, and Fatma Özcan. 2016. ATHENA: an ontology-driven
system for natural language querying over relational data stores. Proceedings of
the VLDB Endowment 9, 12 (2016), 1209–1220.
[43] Jaydeep Sen, Chuan Lei, Abdul Quamar, Fatma Özcan, Vasilis Efthymiou, Ayushi
Dalmia, Greg Stager, Ashish Mittal, Diptikalyan Saha, and Karthik Sankara-
narayanan. 2020. ATHENA++ natural language querying for complex nested
SQL queries. Proceedings of the VLDB Endowment 13, 12 (2020), 2747–2759.
[44] Md Shamim Talukder, Liang Shen, Md Farid Hossain Talukder, and Yukun Bao.
2019. Determinants of user acceptance and use of open government data (OGD):
An empirical investigation in Bangladesh. Technology in Society 56 (2019), 147–
156.
[45] Songül Tolan, Marius Miron, Emilia Gómez, and Carlos Castillo. 2019. Why ma-
chine learning may lead to unfairness: Evidence from risk assessment for juvenile
justice in catalonia. In Proceedings of the Seventeenth International Conference on
Artificial Intelligence and Law. 83–92.
[46] Ellen M Voorhees et al. 1999. The TREC-8 question answering track report. In
Trec, Vol. 99. 77–82.
[47] Vishanth Weerakkody, Zahir Irani, Kawal Kapoor, Uthayasankar Sivarajah, and
Yogesh K Dwivedi. 2017. Open data and its usability: an empirical view from the
Citizen’s perspective. Information Systems Frontiers 19, 2 (2017), 285–300.
[48] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,
Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, Joe
Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu,
Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest,
and Alexander M. Rush. 2019. HuggingFace’s Transformers: State-of-the-art
Natural Language Processing. ArXiv abs/1910.03771 (2019).
[49] Tomer Wolfson, Mor Geva, Ankit Gupta, Matt Gardner, Yoav Goldberg, Daniel
Deutch, and Jonathan Berant. 2020. Break it down: A question understanding
benchmark. Transactions of the Association for Computational Linguistics 8 (2020),
183–198.
[50] Kanit Wongsuphasawat, Zening Qu, Dominik Moritz, Riley Chang, Felix Ouk,
Anushka Anand, Jock Mackinlay, Bill Howe, and Jeffrey Heer. 2017. Voyager 2:
Augmenting visual analysis with partial view specifications. In Proceedings of the
2017 CHI Conference on Human Factors in Computing Systems. 2648–2659.
 Lex Rosetta: Transfer of Predictive Models Across Languages,
Jurisdictions, and Legal Domains
Jaromir Savelka Hannes Westermann Charlotte S. Alexander
jsavelka@cs.cmu.edu Karim Benyekhlef Jayla C. Grant
Carnegie Mellon University Université de Montréal Georgia State University
USA Canada USA

David Restrepo Amariles Sébastien Meeùs Michał Araszkiewicz

Rajaa El Hamdani Aurore Troussel Uniwersytet Jagielloński
HEC Paris HEC Paris Poland
France France

Kevin D. Ashley Karl Branting Mattia Falduti

Alexandra Ashley MITRE Corporation Libera Università di Bolzano
University of Pittsburgh USA Italy
USA

Matthias Grabmair Jakub Harašta Elizabeth Tippett

Technische Universität München Tereza Novotná Shiwanni Johnson
Germany Masarykova univerzita University of Oregon
Czech Republic USA

ABSTRACT CCS CONCEPTS

In this paper, we examine the use of multi-lingual sentence embed-
dings to transfer predictive models for functional segmentation of
adjudicatory decisions across jurisdictions, legal systems (common
and civil law), languages, and domains (i.e. contexts). Mechanisms
for utilizing linguistic resources outside of their original context
have significant potential benefits in AI & Law because differences
between legal systems, languages, or traditions often block wider
adoption of research outcomes. We analyze the use of Language-
Agnostic Sentence Representations in sequence labeling models
using Gated Recurrent Units (GRUs) that are transferable across
languages. To investigate transfer between different contexts we
developed an annotation scheme for functional segmentation of
adjudicatory decisions. We found that models generalize beyond
the contexts on which they were trained (e.g., a model trained on
administrative decisions from the US can be applied to criminal law
decisions from Italy). Further, we found that training the models on
multiple contexts increases robustness and improves overall perfor-
mance when evaluating on previously unseen contexts. Finally, we
found that pooling the training data from all the contexts enhances
the models’ in-context performance.
• Applied computing → Law; Annotation; • Information sys-
tems → Document structure; Structure and multilingual text
search; Data mining.
KEYWORDS
multi-lingual sentence embeddings, transfer learning, domain adap-
tation, adjudicatory decisions, document segmentation, annotation
ACM Reference Format:
Jaromir Savelka, Hannes Westermann, Karim Benyekhlef, Charlotte S. Alexan-
der, Jayla C. Grant, David Restrepo Amariles, Rajaa El Hamdani, Sébastien
Meeùs, Aurore Troussel, Michał Araszkiewicz, Kevin D. Ashley, Alexandra
Ashley, Karl Branting, Mattia Falduti, Matthias Grabmair, Jakub Harašta,
Tereza Novotná, Elizabeth Tippett, and Shiwanni Johnson. 2021. Lex Rosetta:
Transfer of Predictive Models Across Languages, Jurisdictions, and Legal
Domains. In Eighteenth International Conference for Artificial Intelligence
and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York,
NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466149

1 INTRODUCTION
This paper explores the ability of multi-lingual sentence embed-
Permission to make digital or hard copies of part or all of this work for personal or dings to enable training of predictive models that generalize beyond
classroom use is granted without fee provided that copies are not made or distributed individual languages, legal systems, jurisdictions, and domains (i.e.,
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. contexts). We propose a new type schema for functional segmenta-
For all other uses, contact the owner/author(s). tion of adjudicatory decisions (i.e., decisions of trial and appellate
ICAIL’21, June 21–25, 2021, São Paulo, Brazil court judges, arbitrators, administrative judges and boards) and use
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06. it to annotate legal cases across eight different contexts (7 countries,
https://doi.org/10.1145/3462757.3466149 6 languages). We release the newly created dataset (807 documents

129
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
with 89,661 annotated sentences) including the annotation schema
to the public.1
In the area of AI & Law, research typically focuses on a single
context, such as decisions of a specific court on a specific issue
within a specific time range. This is justified by the complexity of
legal work and the need for nuanced solutions to particular prob-
lems. At the same time, this narrow focus can limit the applicability
of the research outcomes, since a proposed solution might not be
readily transferable to a different context. In text classification, for
example, a model might simply memorize a particular vocabulary
characteristic of a given context, rather than acquiring the seman-
tics of a predicted type. Adaptation of such a model to a new context
would then require the assembly of a completely new dataset. This
may be both time-consuming and expensive, since the annotation
of legal documents relies on legal expertise.
Certain tasks appear to be of interest to researchers from multiple
countries with different legal traditions (e.g., deontic classification
of legal norms embodied in statutory law, argument extraction from
case law, summarization/simplification of legal documents, etc.).
This suggests that there may be several core tasks in AI & Law that
are of general interest in almost any context. One such task is a
functional segmentation of adjudicatory decisions, which has been
the subject of numerous studies in the past (see Section 2). In this
paper, we show that for this particular task it is possible to leverage
linguistic resources created in multiple contexts.
This has wide-reaching implications for AI & Law research. Since
annotation of training data is expensive, models that are able to use
existing data from other contexts might be instrumental in enabling
real-world applications that can be applied across contexts. Such
approaches may further enable international collaboration of re-
searchers, each annotating their own part of a dataset to contribute
to a common pool (as we do in this work) that could be used to
train strong models able to generalize across contexts.
1.1 Functional Segmentation
We investigate the task of segmenting adjudicatory decisions based
on the functional role played by their parts. While there are signifi-
cant differences in how decisions are written in different contexts,
we hypothesize that selected elements might be universal, such as,
for example, sections:
(1) describing facts that give rise to a dispute;
(2) applying general legal rules to such facts;
(3) stating an outcome of the case (i.e., how was it decided).
This conjecture is supported by the results of the comparative
project titled Interpreting Precedents [23], which aimed to analyze
(among other things) the structure in 11 different jurisdictions. The
findings of this project suggest that the structure indicated above
may be considered a general model followed in the investigated
jurisdictions, although variations exist that are characteristic of
particular legal systems and types of courts and their decisions.
The ability to segment cases automatically could be beneficial
for many tasks. It could support reading and understanding of
legal decisions by students, legal practitioners, researchers, and
the public. It could facilitate empirical analyses of the discourse
structure of decisions. It could enhance the performance, as well
1 https://github.com/lexrosetta/caselaw_functional_segmentation_multilingual
130
Savelka and Westermann, et al.
as the user experience, of legal search tools. For example, if a user
searches for an application of a legal rule, they might restrict the
search to the section where a judge applies the rule to a factual
situation. Judges themselves might find the technique useful, within
their own jurisdictions but also in transnational disputes involving
the application of different legal standards. The same benefits apply
to non-court settings, e.g., international arbitration, where many
jurisdictions’ laws and their interpretation matter. Further, the
segmentation of decisions into meaningful sections could serve as
an important step in many legal document processing pipelines.
1.2 Hypotheses
To investigate how well predictive models, based on multi-lingual
sentence embeddings, learn to segment cases into functional parts
across different contexts, we evaluated the following hypotheses:
(H1) A model trained on a single context can generalize when
transferred to other, previously unseen, contexts.
(H2) A model trained on data pooled from multiple contexts is
more robust and generalizes better to unseen contexts than
a model trained on a single context.
(H3) A context-specific model benefits from pooling the in-domain
data with data from other contexts.
1.3 Contributions
By carrying out this work, we provide the following contributions
to the AI & Law research community:
• Detailed definition and analysis of a functional segmentation
task that is widely applicable across different contexts.
• A new labeled dataset consisting of 807 documents (89,661
sentences) from seven countries in six different languages.
• Evidence of the effectiveness of multi-lingual embeddings
on processing legal documents.
• Release of the code used for data preparation, analysis, and
the experiments in this work.
2 RELATED WORK
Segmenting court decisions into smaller elements according to their
function or role is an important task in legal text processing. Prior
research utilizing supervised machine learning (ML) approaches
or expert crafted rules can roughly be distinguished into two cat-
egories. First, the task could be to segment the text into a small
number of contiguous parts typically comprising multiple para-
graphs (this work). Different variations of this task were applied
to several legal domains from countries, such as Canada [15], the
Czech Republic [17], France [8], or the U.S. [27]. Second, the task
could instead be labeling smaller textual units, often sentences,
according to some predefined type system (e.g., rhetorical roles,
such as evidence, reasoning, conclusion). Examples from several
domains and countries include administrative decisions from the
U.S. [33, 41], multi-domain court decisions from India [6], inter-
national arbitration decisions [9], or even multi-{domain,country}
adjudicatory decisions in English [28]. Identifying a section that
states an outcome of the case has also received considerable atten-
tion separately [25, 38]. To the best of our knowledge, existing work
on functional segmentation of court decisions is limited to a single
Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdictions, and Legal Domains
language—ours being the first paper exploring the task jointly on
legal documents in multiple languages.
In NLP, the success of word embeddings was followed by an
increasing interest in learning continuous vector representations of
longer linguistic units, such as sentences (a trend that has been re-
flected in AI & Law research as well [34, 41]). Multi-lingual represen-
tations recently attracted ample attention. While most of the earlier
work was limited to a few close languages or pairwise joint em-
beddings for English and one foreign language, several approaches
to obtain general-purpose massively multi-lingual sentence repre-
sentations were proposed [5, 11, 13]. Such representations were
utilized in many downstream applications, such as document clas-
sification [21], machine translation [2], question answering [22],
hate speech detection [4], or information retrieval (IR) in the legal
domain [40]. Our work is one of the first such applications in the
legal domain and to the best of our knowledge the first dealing with
more than two languages.
Approaches other than language-agnostic sentence embeddings
(this work) were used in AI & Law research focused on texts in
multiple languages. A recent line of work mapped recitals to arti-
cles in EU directives and normative provisions in Member states’
legislation [24]. There, mono-lingual models were used (i.e., one
model per language). Other published applications in multi-lingual
legal IR were based on thesauri [14, 29]. A common technique to
bridge the language gap was the use of ontologies and knowledge
graphs [1, 3, 7, 16]. The multi-lingual environments, such as EU
or systems established by international treaties, attracted work on
machine translation [20], meaning equivalence verification [32],
and building of parallel corpora [30, 31].
3 DATASET
In creating the dataset, the first goal was to identify a task that
would be useful across different contexts. After extensive litera-
ture review, we identified the task of functional segmentation of
adjudicatory decisions as a viable candidate. To make the task gen-
eralizable, we decided to include only a small number of core types.
(1) Out of Scope – Parts outside of the main document body (e.g.,
metadata, editorial content, dissents, end notes, appendices).
(2) Heading – Typically an incomplete sentence or marker start-
ing a section (e.g., “Discussion,” “Analysis,” “II.”).
(3) Background – The part where the court describes procedural
history, relevant facts, or the parties’ claims.
(4) Analysis – The section containing reasoning of the court,
issues, and application of law to the facts of the case.
(5) Introductory Summary – A brief summary of the case at the
beginning of the decision.
(6) Outcome – A few sentences stating how the case was decided
(i.e, the overall outcome of the case).
We created detailed annotation guidelines defining the individ-
ual types as well as describing the annotation workflow (tooling,
steps taken during annotation). Eight teams of researchers from six
different countries (14 persons) were trained in the annotation pro-
cess through online meetings. After this, each annotator conducted
a dry-run annotation on 10 cases and received detailed feedback.
Then, each team was tasked with assembling approximately 100
131
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
adjudicatory decisions. Each team developed specifications for the
decisions to be included in their part of the dataset.
Four of the contexts were double-annotated by two annotators
(Canada, Czech R., France, U.S.A. I); the remaining four by just
one. Each team had at least one member with a completed law
degree. When a team had more than one member, law students
were allowed to be included.
A high-level description of the resulting dataset is provided in
Table 1. It consists of eight contexts from seven different countries
(two parts are from the U.S.) with 807 documents in six languages
(three parts are in English). Most of the contexts include judicial
decisions, while U.S.A. II was the only context that consisted solely
of administrative decisions.There are considerable variations in the
length of the documents. While an average document in the U.S.A. I
context comprises of 530.6 sentences, an average document in the
France context is about ten times shorter (59.0 sentences).
The four double-annotated parts enabled us to examine the inter-
annotator agreement. Table 2 shows the raw agreement on a charac-
ter level. While it appears that recognizing the Outcome was rather
straightforward in the France and U.S.A. I contexts, it was more
complicated in case of Canada and the Czech R. This might be due
to a presence/absence of some structural clue. We also observe that
in the Czech R. context it was presumably much easier to distin-
guish between the Background and Analysis than in case of the
other three contexts.
In this paper, we focus on prediction of the Background, Analysis,
and Outcome types. We decided to exclude the Introductory Sum-
mary type, since it is mainly present in the data from the United
States. For the double-annotated datasets, we picked the annota-
tions that appeared to be of higher quality (either by consensus
between the annotators themselves or by a decision of a third unbi-
ased expert).
We first removed all the spans of text annotated with either of
the Out of Scope or Heading types. The removal of Out of Scope
leaves the main body of a decision stripped of potential metadata
or editorial content at the beginning of the document as well as
dissents, or end notes at the end. The removal of the text spans
annotated with the Heading type might appear counter-intuitive
since headings often provide a clear clue as to the content of the
following section (e.g., “Outcome”, “Analysis” etc.). We remove
these (potentially valuable) headings because we want to focus
on the more interesting task of recognizing the sections purely
by the semantics of their constitutive sentences. This task is more
challenging, and more closely emulates generalization to domains
where headings are not used or not present in all cases, or are not
reliable indicators.
The transformed documents are separated into several segments
based on the annotations of the three remaining types. Each seg-
ment is then split into sentences.2 A resulting document is a se-
quence of sentences labeled with one of the Background, Analysis,
or Outcome types. The highlighted (green) part of Table 1 provides
2 We used the processing pipeline from https://spacy.io/ (large models). For the Czech
language we used https://github.com/TakeLab/spacy-udpipe with the Czech model
(PDT) from https://universaldependencies.org/. The output was further processed
with several regular expressions. A different method was used for the French dataset,
which consists of a few very long sections, internally separated by a semicolon. After
consultation with an expert we decided to split the cases by the semicolon as well.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Savelka and Westermann, et al.

Table 1: Descriptive statistics of the created dataset. Each entry provides information about the country, the language of the
decisions (Lang), and the number of documents (Docs) in a specific context. The Sentence-Level Statistics subsection reports
basic descriptive statistics focused on sentences as well as the number of sentences labeled with each type (OoS - Out of Scope,
Head - Heading, Int.S. - Introductory Summary, Back - Background, Anl - Analysis, Out - Outcome). The part highlighted in
green contains the counts of sentences labeled with the types we focus on in this work.

Sentence-Level Statistics
Country Lang Docs Count Avg Min Max OoS Head Int.S. Back Anl Out Description
Canada EN 100 12168 121.7 8 888 873 438 20 3319 7190 328 Random selection of cases retrieved from www.canlii.org from multiple provinces.
7.2% 3.6% 0.2% 27.3% 59.1% 2.7% The selection is not limited to any specific topic or court.
Czech R. CS 100 11283 112.8 10 701 945 1257 2 3379 5422 278 A random selection of cases from Constitutional Court (30), Supreme Court (40), and
8.4% 11.1% 0.0% 29.9% 48.1% 2.5% Supreme Administrative Court (30). Temporal distribution was taken into account.
France FR 100 5507 55.1 8 583 3811 220 0 485 631 360 A selection of cases decided by Cour de cassation between 2011 and 2019. A stratified
69.2% 4.0% 0.0% 8.8% 11.4% 6.5% sampling based on the year of publication of the decision was used to select the cases.
Germany DE 104 10724 103.1 12 806 406 333 38 2960 6697 290 A stratified sample from the federal jurisprudence database spanning all federal courts
3.8% 3.1% 0.4% 27.6% 62.4% 2.7% (civil, criminal, labor, finance, patent, social, constitutional, and administrative).
Italy IT 100 4534 45.3 10 207 417 1098 0 986 1903 130 The top 100 cases of the criminal courts stored between 2015 and 2020 mentioning
9.2% 24.2% 0.0% 21.7% 42.0% 2.9% “stalking” and keyed to the Article 612 bis of the Criminal Code.
Poland PL 101 9791 96.9 4 1232 796 303 0 2736 5820 136 A stratified sample from trial-level, appellate, administrative courts, the Supreme Court,
8.1% 3.1% 0% 27.9% 59.4% 1.4% and the Constitutional tribunal. The cases mention “democratic country ruled by law.”
U.S.A. I EN 102 24898 244.1 34 1121 574 1235 475 6042 16098 474 Federal district court decisions in employment law mentioning “motion for summary
2.3% 5.0% 1.9% 24.3% 64.7% 1.9% judgment," “employee,” and “independent contractor.”
U.S.A. II EN 100 10756 107.6 24 397 1766 650 639 3075 4402 224 Administrative decisions from the U.S. Department of Labor. Top 100 ordered in
1.6% 6.0% 5.9% 28.6% 40.9% 2.1% reverse chronological rulings order, starting in October 2020, were selected.
Overall 6 807 89661 105.6 4 1232 9588 5534 1174 22982 48163 2220

Table 2: Raw agreement on a character level for the four encoder itself has no information on the language or writing script
datasets with two human annotators. The agreement is com- of the tokenized text, while the tokenizer is language specific. It
puted as a percentage of characters where both the annota- is even possible to mix multiple languages in one sentence. The
tors agree on a specific type over all the characters annotated focus of the LASER model is to produce vector representations
by that type by any of the annotators. (NM=Not Marked) of sentences that are general with respect to two dimensions: the
input language and the NLP task [5]. An interesting property of
OoS Head Int.S. Back Anl Out NM such universal multi-lingual sentence embeddings is the increased
Canada 97.2 68.2 44.0 83.3 92.2 79.9 43.4 focus on the sentence semantics, as the syntax or other surface
Czech R. 80.3 54.6 0.0 92.6 94.5 46.9 10.0 properties are unlikely to be shared among languages.
France 93.5 92.5 N/A 43.0 72.2 99.1 1.0
U.S.A. I 90.8 71.0 74.2 78.4 93.7 91.1 18.4 The GRU neural network [10] is an architecture based on a
recurrent neural network (RNN) that is able to learn the mapping
Overall 91.8 70.4 72.1 82.1 92.6 77.3 3.8
from a sequence of an arbitrary length to another sequence. GRUs
are able to either score a pair of sequences or to generate a target
basic descriptive statistics of the resulting dataset per the individual sequence given a source sequence (this work). In a bidirectional
contexts. Our final dataset for analysis consists of 807 cases split GRU, two separate sequences are considered (one from right to
into 74,539 annotated sentences. left and the other from left to right). Traditional RNNs work well
for shorter sequences but cannot be successfully applied to long
4 MODELS sequences due to the well-known problem of vanishing gradients.
Long Short-Term Memory (LSTM) networks [18] have been used
In our experiments we use the Language-Agnostic Sentence Repre- as an effective solution to this problem (the forget gate, along with
sentations (LASER) model [5] to encode sentences from different the additive property of the cell state gradients). GRUs have been
languages into a shared semantic space. Each document becomes proposed as an alternative to LSTMs with a reduced number of
a series of vectors which represent the semantic content of a sin- parameters. In GRUs there is no explicit memory unit, and the
gle sentence. We use these vectors to train a bidirectional Gated forget gate and the update gate are combined. The performance of
Recurrent Unit (GRU) model [10] for predicting sentence labels. GRUs was shown to be superior to that of LSTMs in the scenario
The LASER model is a language-agnostic bidirectional LSTM of long texts and small datasets [39], which is the situation in this
encoder coupled with an auxiliary decoder and trained on parallel work. For these reasons, we chose to use GRUs over LSTMs.
corpora. The sentence embeddings are obtained by applying a max- The overall structure of the employed model is shown in Fig-
pooling operation over the output of the encoder. The resulting ure 1.4 Each case is transformed into a 1080 × 1024 matrix. The
sentence representations (after concatenating both directions) are number 1080 represents the maximum length (in sentences) of any
1024-dimensional. The released trained model,3 which we use in case in our dataset. Shorter cases are padded to be of uniform length.
this work, supports 93 languages (including the six in our dataset) The vectors are passed to the model in batches of size 32. They first
belonging to 30 different families and written in 28 different scripts. go through a masking layer, which masks the sentences used for
The model was trained on 223 million parallel sentences. The joint
3 https://github.com/facebookresearch/LASER 4 The model was implemented using the Keras framework (https://keras.io/).

132
Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdictions, and Legal Domains ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Out-Context Experiment (H1) Out-Context Experiment (H1)

(i = 1) (i = 2)
Context 1 Context n Target Ctx Context 1 Context n Target Ctx
1 1 Test 1 Train 1 1 1

Val 2 2 2 2 2 Test 2

3 3 3 Val 3 3 3

... ... ... ... ... ...

Train
9 9 9 Train 9 9 9

10 10 10 10 10 10

Pooled Out-Context (H2) Pooled With In-Context (H3)

(i = 1) (i = 1)
Context 1 Context n Target Ctx Context 1 Context n Target Ctx
1 1 Test 1 1 1 Test 1

Val 2 2 2 Val 2 2 2

3 3 3 3 3 3
Figure 1: The structure of the sequential model used for pre-
diction. Each case 𝑚 is split into 𝑛 sentences, which are con- Train
... ... ...
Train
... ... ...

verted to language-independent LASER vector embeddings. 9 9 9 9 9 9

These are fed through a bidirectional recurrent GRU-model, 10 10 10 10 10 10

which predicts one of the three labels per input sentence.

Figure 2: A visualization of the three experimental setups.
The figure shows how training contexts are selected for H1,
padding. The data is then passed to a bidirectional GRU model
H2 and H3, and examples of how the folds are assigned to
with 256 units and a dropout of 0.2. Finally, the model contains a
Train, Val and Test.
time-distributed dense output layer with softmax activation, which
outputs the predicted label (i.e., Background, Analysis, Outcome) for
each sentence. As a loss function, we use categorical cross-entropy. Table 3: Description of different training data selection and
As optimizer we use the Adam algorithm [19] with initial learning baselines used for H1, H2 and H3. The random (H1), the best
rate set to 4𝑒 −3 (reduced to 4𝑒 −4 once validation loss has stopped single Out-Context models (H2), and the In-Context models
decreasing, with a patience of 50 epochs). We train the model for up (H3) baselines are highlighted.
to 1000 epochs. We halt the training once the validation accuracy
has not increased in 80 epochs. For prediction, we use the best Name Trained on Hyp. Baseline
model as determined by validation accuracy. Random Target Context - -
In-Context Target Context - -
Out-Context Non-Target Context H1 Random
5 EXPERIMENTAL DESIGN Pooled Out-Context Pooled Non-Target H2 Best performing Out-
Contexts Context model per context
We performed three experiments to test the hypotheses (Section 1.2). Pooled with In-Context Non-Target and Target H3 In-Context
The first experiment (H1) focused on model generalization across Pooled Contexts
different contexts (Out-Context). The second experiment (H2) as-
sessed model robustness when trained on multiple pooled contexts
different from the one where the model was applied (Pooled Out- (6) The 𝑖 fold from the test context is designated as test data.
Context). Finally, the third experiment (H3) analyzed the effects (7) The models are trained and evaluated.
of pooling the target context data with data from other contexts (8) Index 𝑖 = 𝑖 + 1 is set.
(Pooled with In-Context). The different pools of training data, and (9) If 𝑖 ≤ 10 go to (4) else finish.
the baselines we compare them against, are summarized in Table 3 Note that from here on we highlight the baselines using the colors
and described in Sections 5.1-5.3. as shown in Table 3 to improve clarity.
Since our dataset is limited, we performed a 10-fold cross-vali-
dation. The folds were kept consistent across all the experiments. 5.1 Out-Context Experiment (H1)
The experiments were conducted in the following manner: In this experiment, we investigated the ability of the models to
(1) Index 𝑖 = 1 is set. generalize beyond the training context. The pool of training data
(2) Pool of training context(s) is selected (see Sections 5.1-5.3). consisted of a single context at a time (see Figure 2). As a baseline,
(3) A single test context is selected (see Sections 5.1-5.3). we used a random classifier that learns the distribution of the target
(4) Eight of the folds from the training context(s) are used as dataset (train and validation folds) and generates predictions based
training data (index different from 𝑖 and (𝑖 + 1) mod 10). on that distribution. The baseline thus had a slight advantage in
(5) The folds with index (𝑖 + 1) mod 10 from the training con- having access to the signal from the target dataset. The Out-Context
text(s) are designated as validation data. models performing better than the random baseline would show

133
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
that the models are able to transfer (at least some) knowledge from
one context to another.
5.2 Pooled Out-Context Experiment (H2)
The second experiment focused on the ability of the models to gain
robustness as they learn from more than one context, when applied
to an unseen context. Therefore, the training pool for the exper-
iments consisted of data from all the contexts except the current
target context (see Figure 2). As a baseline for each context, we used
the performance measurement taken from the best Out-Context
model for that context (see Section 5.1). This was a very competitive
baseline since the best Out-Context model likely stems from the
context that is the most similar to the target context. If the Pooled
Out-Context model performed better than the best Out-Context
model, this would indicate that pooling datasets increases the ro-
bustness of the model, allowing it to perform better on previously
unseen contexts.
5.3 Pooled With In-Context Experiment (H3)
The third experiment focused on pooling the target context’s data
with data from the other contexts. The training pool therefore
incorporates all of the contexts, including the target one (see Fig-
ure 2). As a baseline, we used the In-Context models, trained on
the target context. Again, this is a very strong baseline since the
In-Context model should be able to learn the most accurate signal
from the target context (including context specific peculiarities). If
the model trained on the pooled contexts (including target) is able
to outperform the In-Context model, this would indicate that pool-
ing contexts is beneficial in terms of both robustness and absolute
performance on the individual contexts.
6 EVALUATION METHOD
To evaluate the performance of the systems trained on data from
different contexts, we used Precision (𝑃), Recall (𝑅), and 𝐹 1 -measure.
We compute the evaluation metrics for each class (i.e., Background,
Analysis, Outcome) per fold. We then report per class 𝐹 1 -scores as
well as an overall 𝐹 1 -score (micro) including its standard deviation.
For statistical evaluation we used the Wilcoxon signed-rank
test [37] to perform pair comparisons between methods and the
baselines as suggested by [12]. We used the overall (micro) 𝐹 1 -score
as the evaluation metric. The null-hypothesis states that the mean
performance of the two compared methods (i.e., an assessed model
and a baseline) is equivalent.
Since the number of samples is rather small (7 contexts in testing
H1, and 8 contexts in testing H2 and H3), we are not able to reject the
null hypotheses at 𝛼 = 0.05 for any of H1–H3. This is as expected
given the size of the dataset, and it is not unacceptable given the
exploratory nature of our work at this stage. We will be in a position
to formally evaluate the hypotheses once we extend the dataset
(see Section 10). Instead, we report raw p-values produced by the
testing as an evidence of the assessed methods’ effectiveness. The
p-value is the probability of seeing the results at least as extreme
as those we observed given the null hypothesis is correct (i.e., their
mean performance is the same).
134
Savelka and Westermann, et al.
Figure 3: A description of performance metrics reported in
Table 4.
7 RESULTS
The results of all three experiments are shown in Table 4. Each row
reports a performance of a specific model across different contexts
(columns). Each cell shows a bold number on the first row which
correspond to the (micro) average F1 -score over the three predicted
classes, with standard deviation across the 10 folds. The F1 -scores
of the three classes are reported in the second row of each cell
ordered as Background, Analysis, Outcome. A visual explanation of
cell contents can be found in Figure 3.
7.1 Out-Context Experiment (H1)
The performance of the models trained during the Out-Context
experiment is reported by the eight rows of Table 4 starting with
Canada and ending with U.S.A. II. Here, the models are trained on
a single context (row), and then evaluated on all of the contexts
(columns). It appears the models perform reasonably well under
the Out-Context condition. The application of the trained models
outperforms the random baseline across the board (in 54 out of
56 instances). This is further corroborated by the low p-values
obtained for these methods when compared to the random baseline
(reported in Table 4).
Several interesting patterns emerge from the results. First, it
appears that models trained on contexts with the same language
or a language from the same family perform better. For example,
the models trained on Canada and the two U.S.A. contexts perform
well among each other. A similar observation applies to the models
trained on the Poland and the Czech R. contexts. Quite surprisingly
the models succeeded in identifying the Outcome sentences to some
extent, despite a heavy under-representation of these sentences
in our dataset. For example, a model trained on Canada predicts
the Outcome sentences of the both U.S.A. contexts with an average
F1 -score close to 0.7. At the same time the Outcome sentences
are by far the most challenging ones. On multiple occasions the
models completely fail to predict these (e.g., Canada→Germany, or
Poland→France).
Overall, the results demonstrate the ability of the models to
effectively transfer the knowledge learned on one context to another.
At the same time, it is clear that the models trained on one context
and applied to a different one do not perform as well as the In-Con-
text models.
Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdictions, and Legal Domains ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 4: Results of the Out-Context, Pooled Out-Context, and Pooled with In-Context experiments. Each row reports a perfor-
mance of a model across contexts. A bold number in each cell reports the (micro) average F1 -score over the predicted classes
and the standard deviation across the 10 folds. The F1 -scores of the three classes are reported below ordered as Background,
Analysis, Outcome. A visual explanation of cell contents can be found in Figure 3. The random (H1), the best single Out-Con-
text models (H2), and the In-Context (H3) baselines are highlighted. The p-values are color coded to match the baselines.

Canada Czech R. France Germany Italy Poland U.S.A. I U.S.A. II Avg (-test) Avg (+test)
Random .54 ± .04 .49 ± .02 .33 ± .05 .56 ± .04 .50 ± .04 .55 ± .04 .58 ± .03 .51 ± .02 .51 ± .07
(dist.) .31 .66 .06 .36 .59 .04 .32 .37 .25 .29 .68 .03 .31 .62 .04 .32 .67 .00 .26 .72 .02 .37 .61 .02 .32 .62 .10
Canada .82 ± .09 .68 ± .08 .64 ± .09 .81 ± .06 .73 ± .08 .73 ± .07 .87 ± .05 .88 ± .03 .76 ± .09 .77 ± .08
𝑝 = .016 .75 .87 .70 .53 .80 .39 .64 .66 .57 .71 .89 .00 .55 .82 .66 .50 .85 .00 .75 .92 .69 .87 .90 .70 .65 .83 .43 .66 .84 .46
Czech R. .76 ± .09 .91 ± .04 .47 ± .08 .82 ± .08 .82 ± .05 .83 ± .07 .84 ± .06 .86 ± .05 .77 ± .13 .79 ± .13
𝑝 = .016 .71 .80 .31 .90 .92 .64 .52 .36 .48 .75 .89 .01 .81 .84 .40 .73 .90 .00 .73 .90 .28 .86 .88 .41 .73 .80 .27 .75 .81 .32
France .71 ± .07 .61 ± .06 .86 ± .08 .66 ± .10 .73 ± .08 .65 ± .07 .72 ± .07 .68 ± .08 .68 ± .04 .70 ± .07
𝑝 = .016 .45 .83 .69 .37 .78 .45 .81 .83 .98 .37 .82 .00 .59 .81 .69 .33 .82 .00 .32 .86 .69 .48 .81 .73 .42 .82 .46 .47 .82 .53
Germany .72 ± .10 .64 ± .09 .29 ± .12 .88 ± .11 .69 ± .09 .77 ± .09 .73 ± .10 .83 ± .07 .67 ± .16 .69 ± .17
𝑝 = .031 .68 .76 .01 .42 .81 .01 .42 .32 .00 .82 .93 .66 .50 .84 .00 .54 .88 .54 .47 .85 .00 .82 .88 .00 .55 .76 .08 .58 .78 .15
Italy .55 ± .12 .76 ± .09 .63 ± .08 .78 ± .09 .95 ± .02 .73 ± .10 .53 ± .13 .74 ± .08 .67 ± .10 .71 ± .13
𝑝 = .047 .57 .55 .49 .74 .81 .12 .69 .63 .52 .73 .83 .00 .92 .96 .94 .66 .78 .00 .50 .54 .42 .74 .74 .63 .66 .70 .31 .69 .73 .39
Poland .76 ± .08 .83 ± .05 .38 ± .11 .85 ± .08 .83 ± .05 .93 ± .05 .73 ± .08 .83 ± .07 .74 ± .15 .77 ± .16
𝑝 = .016 .66 .84 .00 .82 .89 .01 .48 .52 .00 .73 .91 .44 .80 .91 .00 .89 .95 .88 .44 .86 .00 .82 .87 .00 .68 .83 .06 .71 .84 .17
U.S.A. I .83 ± .06 .65 ± .08 .47 ± .14 .81 ± .07 .65 ± .15 .67 ± .09 .91 ± .03 .89 ± .03 .71 ± .13 .74 ± .14
𝑝 = .016 .76 .87 .59 .45 .79 .49 .35 .61 .35 .71 .89 .00 .40 .80 .58 .38 .83 .00 .84 .94 .73 .87 .91 .68 .56 .81 .38 .60 .83 .43
U.S.A. II .81 ± .08 .67 ± .06 .53 ± .15 .84 ± .10 .75 ± .10 .70 ± .08 .86 ± .05 .94 ± .02 .74 ± .11 .76 ± .12
𝑝 = .016 .74 .86 .65 .49 .80 .31 .50 .63 .41 .76 .91 .00 .57 .84 .75 .43 .84 .00 .72 .92 .59 .93 .96 .82 .60 .83 .39 .64 .85 .44
Pooled .83 ± .06 .87 ± .03 .66 ± .08 .90 ± .04 .85 ± .04 .88 ± .05 .81 ± .10 .92 ± .03 .84 ± .08
𝑝 = .148 .77 .86 .66 .87 .91 .03 .59 .67 .71 .87 .95 .01 .81 .89 .65 .83 .94 .01 .64 .88 .73 .91 .93 .65 .79 .88 .43
Pooled+ .88 ± .05 .94 ± .03 .82 ± .09 .96 ± .02 .94 ± .04 .94 ± .04 .92 ± .03 .96 ± .02 .92 ± .04
𝑝 = .195 .83 .91 .77 .94 .95 .70 .76 .78 .96 .94 .98 .64 .91 .95 .90 .92 .97 .65 .86 .95 .84 .95 .96 .80 .89 .93 .78

7.2 Pooled Out-Context Experiment (H2)
The performance of Pooled Out-Context models is reported in
the Pooled row of Table 4. The experiment concerns the resulting
models’ robustness, i.e., if a model trained on multiple contexts
adapts well to unseen contexts. We are especially interested if such
a model adapts better than the models trained on single contexts.
The results suggest that training on multiple contexts leads to
models that are robust and perform better than the models trained
on a single context. The multi-context models outperform the best
single extra-context model baseline in 7 out of 8 cases. The 𝑝 = 0.148
needs to be understood in terms of the small number of samples
(contexts) and competitiveness of the baseline.
Interestingly, the Pooled Out-Context models even appear to
be competitive with several In-Context models (Canada, Czech R.,
Germany, U.S.A. II). The overall average F1 -scores are often quite
high (over 0.80 or 0.90). This is a surprising outcome considering
the fact that no data from the context on which a model is evaluated
is used during training.
7.3 Pooled With In-Context Experiment (H3)
The performance of the Pooled with In-Context model is reported
in the Pooled+ row of Table 4. This experiment models a scenario
where a sample of labeled data from the target context is available.
The question is whether combining In-Context data with data from
other contexts leads to improved performance.
The results appear to suggest that pooling the target context
with data from other contexts does lead to improved performance.
In case of 3 out of the 8 contexts (Canada, Czech R., and U.S.A. I)
the improvement is clear and substantial across all the three classes
(Background, Analysis, Outcome). For three additional contexts (Ger-
many, Poland, and U.S.A. II), the performance also improved in
terms of overall (micro) F1 -score, but took a slight hit for the chal-
lenging Outcome class. With respect to the two remaining contexts
(France, Italy) the overall performance of the pooled models is lower
than that of the In-Context models. As in the previous experiment,
the 𝑝 = 0.195 needs to be understood in terms of the small number
of samples (contexts) and high competitiveness of the In-Context
baseline.

135
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 4: Average LASER embeddings for each case doc-
ument, projected to 2-dimensional space using Principal
Component Analysis.
8 DISCUSSION
It appears that the multilingual sentence embeddings generated
by the LASER model excel at capturing the semantics of a sen-
tence. A model trained on a single context could in theory capture
the specific vocabulary used in that context. This would almost
certainly lead to poor generalization across different contexts. How-
ever, the performance statistics we observed when transferring the
Out-Context models to other contexts suggest that the sentence
embeddings provide a representation of the sentences that enable
the model to learn aspects of the meaning of a sentence, rather than
mere surface linguistic features.
The results clearly point to certain relationships where contexts
within the same or related languages appear to work well together,
e.g., {Canada, U.S.A. I, and U.S.A. II} or {Czech R. and Poland}. This
could indicate that the multi-lingual embeddings work better when
the language is the same or similar. It could also point to similarities
in legal traditions, e.g. the use of somewhat similar sentences to
indicate a transition from one type of section to the next. Note that
we removed headings, which means that explicit cues could not be
relied on by the models we trained on such transformed documents.
Finally, the cause could also be topical (domain) similarity of the
contexts (e.g., both the U.S.A. contexts deal with employment law).
Also note that the above are just possible explanations. We did not
perform feature importance analysis on the models.
To gain insight into this phenomenon, we visualized the relation-
ships among the contexts on a document level. We first calculated
the average sentence embedding for each document. This yielded
1024-dimensional vectors for 807 documents representing their se-
mantics. We arranged the resulting vectors in a matrix (1024 × 807)
and performed a Principal Component Analysis (PCA) reducing
the dimensionality of the document vectors to 2. This operation
enabled a convenient visualization shown in Figure 4.
136
Savelka and Westermann, et al.
Figure 5: Distribution of labels across datasets. The X-axis
corresponds to unique cases, whereas the Y-axis corresponds
to the normalized length of the cases. The colors correspond
to the different labels of the sentences (Background, Analy-
sis, Outcome).
Overall, the cases from the same contexts appear to cluster to-
gether. This is expected as the documents written in the same
language, having similar topics, or sharing some similarities due
to legal traditions, are likely to map to vectors that are closer. The
documents from both the U.S.A. contexts occupy the same region
in the embedding space. This is not surprising as they come from
the same jurisdiction, are written in the same language, and deal
with similar topics (employment law). The Canadian cases, which
are also in English, occupy a nearby space. This could be linked to
the language as well as likely similarities in legal traditions. French,
German, and Italian documents occupy the middle space; they are
closer to the English documents than those from the Czech R. and
Poland. Interestingly, the Czech and Polish documents occupy al-
most the same space. The Polish context focuses on the rule of
law while the Czech one is supposed to be more general. As the
latter deals with the decisions of the top-tier courts (one of them
Constitutional), it is possible that the topics substantially overlap.
Moreover, Poland and the Czech R. share similar legal traditions
and languages from the same family (Slavic), so the close proximity
of the documents in the embedding space might not be unexpected.
Finally, German, French, and Canadian cases occupy wider areas
than documents from other contexts. This could be due to their
lack of focus on specific legal domains.
We observed a peculiar phenomenon where the Out-Context
models trained on the German and Polish contexts failed to detect
Outcome sentences on the six remaining contexts and vice-versa.
The cause is readily apparent from the visualization shown in Fig-
ure 5. Each segment of the figure depicts a spatial distribution of
sentences color coded with their labels across documents for a par-
ticular context. As can be seen, the cases typically follow a pattern
of a contiguous Background segment, followed by a long Analysis
section. The several Outcome sentences are placed at the very end
of the documents. In the Polish and German decisions, however,
the Outcome sentences come first. The GRU models we use rely on
the structure as well as semantics in making their predictions. As
we can see, a model trained exclusively on cases that begin with a
Background might therefore have difficulties correctly identifying
outcome sections at the beginning, and vice-versa. However, as we
Lex Rosetta: Transfer of Predictive Models Across Languages, Jurisdictions, and Legal Domains
will see below, a model trained with data featuring both structures
can learn to correctly identify the correct structure based on the
semantics of the sentences.
The model trained on the French context appears to perform
better on detecting Outcome sentences than models trained on other
contexts. This is somewhat surprising as the French model’s overall
performance is among the weakest (e.g. compare the Czech model’s
average F1 = 0.77 to the F1 = 0.68 of the French model). Again,
Figure 5 provides an insight into why this happens. The French
context is the only one where the count of Outcome sentences is
comparable to those of the other two categories. For all the other
contexts, the Outcome sentences are heavily underrepresented. This
reveals an interesting direction for future work where the use of
re-sampling may yield models with better sensitivity for identifying
Outcome sentences.
In two instances models trained on a single context under-per-
formed the random baseline. The model trained on the Germany
context achieved the average F1 = 0.29 when applied to the French
context (Random 𝐹 1 = 0.33). As the model trained on Polish data
also performed poorly on the French context (𝐹 1 = 0.38) the cause
appears to be the inability of the two models to detect the Outcome
sentences at the end (discussed above). As the Outcome sentences
are heavily present in the French context, this problem manifests
in a score lower than the random baseline. The second instance
is the model trained on the Italy context applied to the U.S.A. I
data (F1 = 0.53 versus F1 = 0.58 on Random). Here, the cause
appears to be different. Note that the Italian context likely has a very
specific notion of Outcome sentences (F1 = 0.94 on Italy→Italy).
It appears that many Analysis sentences from the U.S.A. I context
were labeled as Outcome by the model. Summary judgments often
address multiple legal issues with their own conclusions which
could have triggered the model to label such sentences as Outcome.
An important finding is the performance of the Pooled Out-
Context model (H2) shown in the Pooled row of Table 4. The ex-
periment simulates training of a model on several contexts, and
then applying it to an unseen context. The Pooled Out-Context
models, having no access to the data from a target context, reliably
outperform the best single Out-Context models. They appear to
be competitive with several In-Context models. These results are
achieved with a fairly small dataset of 807 cases. We expect that
expanding the dataset would lead to further improved performance.
The Pooled with In-Context experiment (H3) models the situ-
ation where data from a target context is available in addition to
labeled data from other contexts. Our experiments indicate that
the use of data from other contexts (if available) in addition to data
from the target context is preferable to the use of the data from the
target context only. This is evidenced by the improved performance
of the models trained on the pooled contexts over the single In-Con-
text models. The models have an interesting property of being able
to identify the Outcome sentences with effectiveness comparable
to (or higher than) the models trained on the same context only.
This holds for all the contexts, except Poland where the Outcome
performance is a bit lower (0.65 vs. 0.88). This indicates that the
model is able to learn the two possible modes of the Outcome section
placement. It successfully distinguishes cases where the section is
at the beginning from the cases where the Outcome sentences are
found toward the end.
137
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The inclusion of the In-Context data in the pooled data leads
to a remarkable improvement over only using the pooled Out-
Context data. The magnitude of the improvement highlights the
importance of including such data in the training. We envision that
the models trained on different contexts used in combination with
high-speed similarity annotation frameworks [35, 36] could enable
highly cost efficient annotation in situations where resources are
scarce. Perhaps, adapting a model to an entirely new context could
be as simple as starting with a model trained on other contexts, and
spending a few hours correcting the misconceptions of the model
to teach it the particularities of the new context.
9 CONCLUSIONS
We analyzed the use of multi-lingual sentence embeddings in se-
quence labeling models to enable transfer across languages, juris-
dictions, legal systems (common and civil law), and domains. We
created a new type schema for functional segmentation of adjudi-
catory decisions and used it to annotate legal cases across eight
different contexts. We found that models generalize beyond the
contexts they were trained on and that training the models on
multiple contexts increases their robustness and improves the over-
all performance when evaluating on previously unseen contexts.
We also found that pooling the training data of a model with data
from additional contexts enhances its performance on the target
context. The results are promising in enabling re-use of annotated
data across contexts and creating generalizable and robust models.
We release the newly created dataset (807 documents with 89,661
annotated sentences), including the annotation schema and the
code used in our experiments, to the public.
This work suggests a promising path for the future of inter-
national collaboration in the field of AI & Law. While previous
annotation efforts have typically been limited to a single context,
the experiments presented here suggest that researchers can work
together by annotating cases from many different contexts at the
same time. Such a combined effort could aid researchers in creating
models that perform well on the data from the context they care
about, while at the same time helping other groups train even better
models for other contexts. We encourage these research directions
and hope to form such collaborations under the Lex Rosetta project.
10 FUTURE WORK
The application of multi-lingual sentence embeddings to functional
segmentation of case law across different contexts yielded promis-
ing results. At the same time, the work is subject to limitations
and leaves much room for improvement. Hence, we suggest several
directions for future work:
• Extension of the datasets from different contexts used in this
work beyond ~100 documents per context.
• Annotation of data from contexts beyond the eight used here
(multi-lingual models support close to 100 languages).
• Analysis of automatic detection of Introductory Summary,
Headings, and Out of Scope.
• Identification and investigation of other tasks applicable
across different contexts.
• Evaluation of the application of other multilingual models
(e.g., those mentioned in Section 2).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
• Exploring other transfer learning strategies beyond simple
data pooling, such as the framework proposed in [26].
• Using multi-lingual models for annotation tasks with high-
speed annotation framework, such as [35, 36].
• Performing the transfer across contexts with related (but
different) tasks, such as in [28].
• Further exploring the differences in the distribution of the
multilingual embeddings for purposes of comparing and
analyzing domains, languages, and legal traditions.
ACKNOWLEDGMENTS
Hannes Westermann, Karim Benyeklef, and Kevin D. Ashley would
like to thank the Cyberjustice Laboratory at Université de Montréal,
the LexUM Chair on Legal Information and the Autonomy through
Cyberjustice Technologies (ACT) project for their support of this
research. Kevin D. Ashley also thanks the Canadian Legal Infor-
mation Institute for providing the corpus of legal cases. Matthias
Grabmair thanks the SINC GmbH for supporting this research.
Jakub Harašta and Tereza Novotná acknowledge the support of
the ERDF project “Internal grant agency of Masaryk University”
(No. CZ.02.2.69/0.0/0.0/19_073/0016943).
REFERENCES
[1] Tommaso Agnoloni, Lorenzo Bacci, Enrico Francesconi, P Spinosa, Daniela Tis-
cornia, Simonetta Montemagni, and Giulia Venturi. 2007. Building an ontological
support for multilingual legislative drafting. Frontiers in Artificial Intelligence
and Applications 165 (2007), 9.
[2] Roee Aharoni, Melvin Johnson, and Orhan Firat. 2019. Massively Multilingual
Neural Machine Translation. In NAACL-HLT, Vol. 1 (Long and Short Papers).
3874–3884.
[3] Gianmaria Ajani, Guido Boella, Luigi Di Caro, Livio Robaldo, Llio Humphreys,
Sabrina Praduroux, Piercarlo Rossi, and Andrea Violato. 2016. The European
Legal Taxonomy Syllabus: A multi-lingual, multi-level ontology framework to
untangle the web of European legal terminology. Applied Ontology 11, 4 (2016).
[4] Sai Saket Aluru, Binny Mathew, Punyajoy Saha, and Animesh Mukherjee. 2020.
Deep learning models for multilingual hate speech detection. arXiv preprint
arXiv:2004.06465 (2020).
[5] Mikel Artetxe and Holger Schwenk. 2019. Massively multilingual sentence
embeddings for zero-shot cross-lingual transfer and beyond. Transactions of the
Association for Computational Linguistics 7 (2019), 597–610.
[6] Paheli Bhattacharya, Shounak Paul, Kripabandhu Ghosh, Saptarshi Ghosh, and
Adam Wyner. 2019. Identification of rhetorical roles of sentences in Indian legal
judgments. In JURIX 2019, Vol. 322. IOS Press, 3.
[7] Guido Boella, Luigi Di Caro, Michele Graziadei, Loredana Cupi, Carlo Emilio
Salaroglio, Llio Humphreys, Hristo Konstantinov, Kornel Marko, Livio Robaldo,
Claudio Ruffini, et al. 2015. Linking legal open data: breaking the accessibility and
language barrier in european legislation and case law. In ICAIL 2015. 171–175.
[8] Paul Boniol, George Panagopoulos, Christos Xypolopoulos, Rajaa El Hamdani,
David Restrepo Amariles, and Michalis Vazirgiannis. 2020. Performance in the
Courtroom: Automated Processing and Visualization of Appeal Court Decisions
in France. In Proceedings of the Natural Legal Language Processing Workshop 2020.
[9] Karl Branting, Brandy Weiss, Bradford Brown, Craig Pfeifer, A Chakraborty, Lisa
Ferro, M Pfaff, and A Yeh. 2019. Semi-supervised methods for explainable legal
prediction. In ICAIL 2019. 22–31.
[10] Kyunghyun Cho, Bart van Merrienboer, Çaglar Gülçehre, Dzmitry Bahdanau,
Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2014. Learning Phrase
Representations using RNN Encoder-Decoder for Statistical Machine Translation.
In EMNLP 2014.
[11] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guil-
laume Wenzek, Francisco Guzmán, Édouard Grave, Myle Ott, Luke Zettlemoyer,
and Veselin Stoyanov. 2020. Unsupervised Cross-lingual Representation Learning
at Scale. In ACL 2020. 8440–8451.
[12] Janez Demšar. 2006. Statistical comparisons of classifiers over multiple data sets.
Journal of Machine learning research 7, Jan (2006), 1–30.
[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
NAACL 2019, Volume 1 (Long and Short Papers). 4171–4186.
[14] Luca Dini, Wim Peters, Doris Liebwald, Erich Schweighofer, Laurens Mommers,
and Wim Voermans. 2005. Cross-lingual legal information retrieval using a
138
Savelka and Westermann, et al.
WordNet architecture. In ICAIL 2005. 163–167.
[15] Atefeh Farzindar and Guy Lapalme. 2004. LetSum, an Automatic Text Summa-
rization system in Law field. JURIX 2004.
[16] Jorge González-Conejero, Pompeu Casanovas, and Emma Teodoro. 2018. Business
Requirements for Legal Knowledge Graph: the LYNX Platform.. In TERECOM@
JURIX 2018. 31–38.
[17] Jakub Harašta, Jaromír Šavelka, František Kasl, and Jakub Míšek. 2019. Automatic
Segmentation of Czech Court Decisions into Multi-Paragraph Parts. Jusletter IT
4, M (2019).
[18] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
[19] Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti-
mization. arXiv preprint arXiv:1412.6980 (2014).
[20] Philipp Koehn, Alexandra Birch, and Ralf Steinberger. 2009. 462 Machine Trans-
lation Systems for Europe. In Proceedings of the Twelfth Machine Translation
Summit. Association for Machine Translation in the Americas, 65–72.
[21] Guokun Lai, Barlas Oguz, Yiming Yang, and Veselin Stoyanov. 2019. Bridg-
ing the domain gap in cross-lingual document classification. arXiv preprint
arXiv:1909.07009 (2019).
[22] Patrick Lewis, Barlas Oguz, Ruty Rinott, Sebastian Riedel, and Holger Schwenk.
2020. MLQA: Evaluating Cross-lingual Extractive Question Answering. In ACL
2020. 7315–7330.
[23] D.N. MacCormick, R.S. Summers, and A.L. Goodhart. 2016. Interpreting Precedents:
A Comparative Study. Taylor & Francis.
[24] Rohan Nanda, Llio Humphreys, Lorenzo Grossio, and Adebayo Kolawole John.
2020. Multilingual Legal Information Retrieval System for Mapping Recitals and
Normative Provisions. In Proceedings of Jurix 2020. IOS Press, 123–132.
[25] Alina Petrova, John Armour, and Thomas Lukasiewicz. 2020. Extracting Out-
comes from Appellate Decisions in US State Courts. In JURIX 2020. 133.
[26] Jaromír Šavelka and Kevin D Ashley. 2015. Transfer of predictive models for
classification of statutory texts in multi-jurisdictional settings. In ICAIL 2015.
216–220.
[27] Jaromir Savelka and Kevin D Ashley. 2018. Segmenting US Court Decisions into
Functional and Issue Specific Parts.. In JURIX 2018. 111–120.
[28] Jaromır Šavelka, Hannes Westermann, and Karim Benyekhlef. 2020. Cross-
Domain Generalization and Knowledge Transfer in Transformers Trained on
Legal Data. In ASAIL@ JURIX 2020.
[29] Párai Sheridan, Martin Braschlert, and Peter Schauble. 1997. Cross-language
information retrieval in a Multilingual Legal Domain. In International Conference
on Theory and Practice of Digital Libraries. Springer, 253–268.
[30] Ralf Steinberger, Mohamed Ebrahim, Alexandros Poulis, Manual Carrasco-
Benitez, Patrick Schluter, Marek Przybyszewski, and Signe Gilbro. 2014. An
overview of the European Union’s highly multilingual parallel corpora. Lan-
guage Resources and Evaluation 48, 4 (2014), 679–707.
[31] Kyoko Sugisaki, Martin Volk, Rodrigo Polanco, Wolfgang Alschner, and Dmitriy
Skougarevskiy. 2016. Building a Corpus of Multi-lingual and Multi-format Inter-
national Investment Agreements. In JURIX 2016.
[32] Linyuan Tang and Kyo Kageura. 2019. Verifying Meaning Equivalence in Bilingual
International Treaties. In JURIX 2019. 103–112.
[33] Vern R Walker, Krishnan Pillaipakkamnatt, Alexandra M Davidson, Marysa
Linares, and Domenick J Pesce. 2019. Automatic Classification of Rhetorical
Roles for Sentences: Comparing Rule-Based Scripts with Machine Learning.. In
ASAIL@ ICAIL 2019.
[34] Hannes Westermann, Jaromír Šavelka, and Karim Benyekhlef. 2021. Paragraph
Similarity Scoring and Fine-Tuned BERT for Legal Information Retrieval and
Entailment. In New Frontiers in Artificial Intelligence (Lecture Notes in Computer
Science). Springer International Publishing.
[35] Hannes Westermann, Jaromír Šavelka, Vern R Walker, Kevin D Ashley, and Karim
Benyekhlef. 2019. Computer-Assisted Creation of Boolean Search Rules for Text
Classification in the Legal Domain. In JURIX 2019, Vol. 322. IOS Press, 123.
[36] Hannes Westermann, Jaromír Šavelka, Vern R Walker, Kevin D Ashley, and Karim
Benyekhlef. 2020. Sentence Embeddings and High-Speed Similarity Search for
Fast Computer Assisted Annotation of Legal Documents. In JURIX 2020, Vol. 334.
IOS Press, 164.
[37] Frank Wilcoxon. 1992. Individual comparisons by ranking methods. In Break-
throughs in statistics. Springer, 196–202.
[38] Huihui Xu, Jaromír Šavelka, and Kevin D Ashley. 2020. Using Argument Mining
for Legal Text Summarization. In JURIX 2020, Vol. 334. IOS Press.
[39] Shudong Yang, Xueying Yu, and Ying Zhou. 2020. LSTM and GRU Neural Network
Performance Comparison Study: Taking Yelp Review Dataset as an Example. In
IWECAI 2020. IEEE, 98–101.
[40] Vladimir Zhebel, Denis Zubarev, and Ilya Sochenkov. 2020. Different Approaches
in Cross-Language Similar Documents Retrieval in the Legal Domain. In Interna-
tional Conference on Speech and Computer. Springer, 679–686.
[41] Linwu Zhong, Ziyi Zhong, Zinian Zhao, Siyuan Wang, Kevin D Ashley, and
Matthias Grabmair. 2019. Automatic summarization of legal decisions using
iterative masking of predictive sentences. In ICAIL 2019. 163–172.
 Converting Copyright Legislation into Machine-Executable
Code: Interpretation, Coding Validation and Legal Alignment
Alice Witt
Law School
Queensland University of Technology
Brisbane, Queensland, Australia
ae.witt@qut.edu.au
Guido Governatori
Data61
CSIRO
Dutton Park, Queensland, Australia
ABSTRACT
A critical challenge in “Rules as Code” (“RaC”) initiatives is enhanc-
ing legal accuracy. In this paper, we present the preliminary results
of a two-week, first of its kind experiment that aims to shed light
on how different legally trained people interpret and convert Aus-
tralian Commonwealth legislation into machine-executable code.
We find that coders collaboratively agreeing on key legal terms, or
atoms, before commencing independent coding work can signifi-
cantly increase the similarity of their encoded rules. Participants
nonetheless made a range of divergent interpretive choices, which
we argue are most likely due to: (1) the complexity of statutory
interpretation, (2) encoded provisions having varying levels of gran-
ularity, and (3) the functionality of our coding language. Based on
these findings, we draw an important distinction between processes
for technical validation of encoded rules, which focus on ensuring
rules adhere to select coding languages and conventions, and pro-
cesses of legal alignment, which we conceptualise as enhancing
congruence between the encoded provisions and the true meaning
of the statutory text in line with the modern approach to statutory
interpretation. We argue that these processes are distinct but both
critically important in enhancing the accuracy of encoded rules. We
conclude by underlining the need for multi-disciplinary expertise
across specific legal subject matters, statutory interpretation and
technical programming in RaC initiatives.
CCS CONCEPTS
• Computing methodologies → Artificial intelligence; • Ap-
plied computing → Law.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466083
139
Anna Huggins
Law School
Queensland University of Technology
Brisbane, Queensland, Australia
Joshua Buckley
Data61
CSIRO
Dutton Park, Queensland, Australia
KEYWORDS
Rules as code, machine-executable code, statutory interpretation,
legal alignment, technical processes
ACM Reference Format:
Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley. 2021.
Converting Copyright Legislation into Machine-Executable Code: Inter-
pretation, Coding Validation and Legal Alignment. In Eighteenth Interna-
tional Conference for Artificial Intelligence and Law (ICAIL’21), June 21–
25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 10 pages. https:
//doi.org/10.1145/3462757.3466083
1 BACKGROUND
In recent years, there has been significant growth in the “Rules
as Code” (“RaC”) movement, a label for diverse initiatives that re-
evaluate how, when, and for whom government rules are made [24].
There are two main RaC approaches, the first of which involves
converting existing regulation, including statutes, into machine-
executable code: “a coded representation of the actual rules in the
legislation, written in a computer language, so that computers can
read it and then use it to carry out programs”[29, p 27]. An applica-
tion of this approach is regulatory technology (“RegTech”) [31] that
can help firms comply and stay up to date with the rules governing
their commercial activities [24, 34]. The second approach involves
“co-drafting” regulation in both natural language and machine-
consumable format, one that can “enable computers to model the
effect of the law”[3, p. 76], at the same time. A distinguishing feature
of co-drafting is the potential for “digital first” rules [33], which
not only represent an output, but also a “strategic and deliberate
approach to rulemaking”[24, p. 81]. RaC is therefore related to, yet
separate from, “computational law”, which investigates whether
regulation can and should be represented in computer code, among
other lines of inquiry [30], and “automated decision making", which
refers to decisions made by automated means with varying levels
of human involvement [18].
While decades of research inform these overlapping areas, and
governments including New Zealand, Australia, France and Canada
have made significant inroads with the practical RaC movement
[27, 30], challenges persist in ensuring that encoded rules are trans-
parent, traceable, appealable and legally accurate. It can be particu-
larly difficult to enhance the legal accuracy of digitised legislation
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
in a way that aligns with the correct legal interpretation of the se-
lect statute(s). In the Australian legal system, this challenge is made
more complex by the strict separation of judicial, legislative and ex-
ecutive powers in the Australian Constitution, under which only the
judiciary can authoritatively interpret the meaning of statutes [18].
A result is that other interpreters, including individual coders and
RegTech companies, are expected to mirror the courts’ approach to
interpreting statutes [5].
Against this backdrop, we commenced a multi-disciplinary and
collaborative research project that seeks to identify the legal and
coding challenges of converting Australian Commonwealth legisla-
tion into a machine-executable format. Our research largely falls
within the first RaC approach given that we are focusing on legisla-
tion that already exists. In this paper, as part of the broader research
project, we present a first of its kind experiment that aims to shed
light on how different legally trained people (T = 3 participants
over a two-week period) interpret and convert select provisions of
the Copyright Act 1968 (Cth) (“the CA”, “the Act") into computer
code. By examining the choices that interpreters make when con-
verting existing legislation into machine-executable code, we aim
to provide critical insights into the coding, statutory interpretation
and other issues that arise in the encoding process. This can, in
turn, facilitate new understandings of how stakeholders can pro-
mote alignment between the languages and logics of statutes and
encoded rules for the RaC movement more broadly.
This paper proceeds in five sections. In Section 2, we provide a
brief overview of Australian copyright law and, in Section 3, we
explain the Defeasible Deontic Logic and Turnip encoding soft-
ware that participants used to convert select provisions of the CA
into computer code. Next, in Section 4, we explain our experiment
design and methods. Then, in Section 5, we outline our prelimi-
nary results that support our hypothesis that coders collaboratively
agreeing on key atoms, or legal terms, before commencing inde-
pendent coding work increases the similarity of their coded output.
Despite a significant increase in the average similarity of atoms
and, to a lesser extent, rules in Week 2, participants made a range
of divergent interpretive choices, most likely due to: (1) the com-
plexity of statutory interpretation, (2) encoded provisions having
varying levels of granularity, and (3) the functionality of our coding
language, Turnip. These differences underline the complexity of
attempting to reproduce the languages and logics of a statutory
text in machine-executable code.
Overall, we argue that processes for technical coding validation
and legal alignment are distinct but both critically important in
enhancing the accuracy of RaC. On the one hand, processes for
technical validation can be automated and/or manual, and aim to
ensure encoded rules adhere to select coding languages and conven-
tions. On the other hand, a process of legal alignment is concerned
with enhancing congruence between the encoded rules and the true
meaning of the select legislation in line with the modern approach
to statutory interpretation, an undertaking that heavily relies on
human judgment. This approach, as articulated and applied by the
High Court of Australia in Project Blue Sky v Australian Broadcast-
ing Corporation (1998) 194 CLR 355, “requires a combined exercise
involving analysis of the text, context and purpose (or policy) of the
statute in question”[20, p. 116]. The key task of statutory interpre-
tation is therefore to find the legal meaning of the statutory words
140
Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley
(the text) taking into account their context and purpose [26]. We
conclude this paper, in Section 6, by highlighting the need in RaC
initiatives for multi-disciplinary expertise across specific legal sub-
ject matters, statutory interpretation and technical programming.
We also identify important areas for future research, such as testing
our hypothesis with a larger number of participants, and further
exploring what the separate yet interrelated processes for technical
coding validation and legal alignment might entail in practice.
2 COPYRIGHT LAW
Copyright is a body of intellectual property law that “confers rights
in relation to the reproduction and dissemination of material that
expresses ideas or information” [8, p. 7]. For material to be protected
under Australian copyright law, it must: (1) fall under one of the
established categories of subject matter, including literary, dramatic,
musical or artistic "works" (known as “Part III (of the CA) works"),
which we focus on in this experiment; and (2) be sufficiently con-
nected to Australia (see ss 10 and 32 of the CA). Additionally, Part
III works must be: (3) recorded in “material form” and (4) “original"
(s 32 of the CA) [8, ch. 6]. When the relevant criteria for subsistence
of copyright are satisfied, the copyright owner has certain exclusive
rights (see, e.g., Divisions 1 and 2, Part III of the CA), which a party
can directly or indirectly infringe [8, pp. 268, 285].
We focus on copyright law for several reasons, including our
research team having some copyright law expertise, and because
this body of law is widely applicable to a range of stakeholders but
often difficult for lay audiences to understand [7]. In practical terms,
our research could be of benefit to large technology companies,
galleries, libraries, educational institutions and archives that often
deal with copyright issues in bulk [28]. It could also potentially
benefit a large number of small content creators, who routinely
rely on existing material [2]. In research terms, by focusing on law
that already exists, we are generating new knowledge particularly
relevant to the first RaC approach. This knowledge can in turn
inform the RaC movement more broadly.
3 FROM LEGISLATION TO ENCODED RULES
To convert select provisions of the CA into machine-executable
code, participants used a language and program (reasoner) called
“Turnip”, which is based on Defeasible Deontic Logic (“DDL”) [14].
The presentation of DDL and Turnip in this paper is a based on
[12, 14]. DDL is an extension of Defeasible Logic [1], which refers to
an interest that can be defeated, and Deontic Logic, which pertains
to “the study of those sentences in which only logical words and
normative expressions occur essentially. Normative expressions in-
clude the words ‘obligation’, ‘duty’, ‘permission’, ‘right’, and related
expressions”[10, p. 1]. Defeasible Deontic Logic therefore extends
defeasible logic “by adding deontic and other modal operators” [16,
p. 47]. More specifically, DDL enables coders to integrate reasoning
with exceptions; to model deontic concepts, such as obligations [O],
permissions [P], prohibitions [F], and exemptions [E]; and to rep-
resent both definitional norms (also known as “constitutive rules”)
and prescriptive norms [12, p. 178], all of which are present in the
CA. Coders can also use a non-classical compensation operator to
model obligations in force after a (potential) violation [14, 15]. De-
feasible Deontic Logic has been applied in several studies that aim
Converting Copyright Legislation into Machine-Executable Code
to convert different types of Australian regulation into computer
code [12, 13, 19].
A rule in DDL takes the form of an IF . . . THEN . . . statement
in which “IF” represents the condition(s) of the rule and "THEN"
models the effect of the norm [11, p. 284]. Coders can divide rules
into constitutive rules that, for instance, define important terms
in a normative document (e.g., different types of regulation) or
outline condition(s) (i.e., the IF part) that might give rise to legal
requirements (i.e., the THEN part, such as obligations, permissions
and prohibitions). Rules can be further classified according to their
strength: specifically, as strict rules, defeasible rules and defeaters. A
strict rule is a rule in the classical sense. Defeasible rules are rules
subject to exceptions: the conclusion of the rule holds unless there
are other (applicable) rules (for the same conclusion) that defeat
the rule. Defeaters are a special kind of rule, they do not support
conclusions, but prevent the conclusion to the opposite [1, p. 257].
For more information about these classifications, see [14].
As previously noted, DDL enables coders to represent both defi-
nitional norms, also known as “constitutive rules”, and prescriptive
norms. Constitutive rules are those in standard defeasible logic [12,
p. 179]. Normative rules can be prescriptive, such as rules estab-
lishing that something is obligatory or forbidden, and permissive,
including rules establishing that certain activities are explicitly
permitted, derogating rules for prohibitions or obligations to the
contrary. The standard form of normative rules follows:
𝑟 : 𝐴1, . . . , 𝐴𝑛 ↩→□ 𝐶 1 ⊙ · · · ⊙ 𝐶𝑚
In this rule, 𝐴1, . . . , 𝐴𝑛 are the condition(s) of the rule expressed
as literals or deontic literals (e.g., an obligation [O] or permission
[P]), □ is a deontic modality, and the 𝐶𝑖 are literals (𝐶 1 ⊙ · · · ⊙ 𝐶𝑚
is a “reparation chain"). ↩→ is a placeholder for the type of rule,
and → stands for a strict rule, ⇒ for a defeasible rule, and { for
a defeater. The mode of the rule □ determines the scope of the
conclusion. When the mode is [O], the meaning of the right-hand
side of the rule is that when the rule applies [O]𝐶 1 is in force (i.e.,
𝐶 1 is obligatory). If the rule is violated (i.e., ¬𝐶 1 holds), then [O]𝐶 2
is in force (𝐶 2 is obligatory, and 𝐶 2 compensates for the violation
of [O]𝐶 1 ). We can repeat this reasoning when [O]𝐶 2 is potentially
violated [14, 15].
DDL is a type of skeptical non-monotonic formalism, which
means that when there are applicable rules with conflicting con-
clusions (i.e., 𝐴 and ¬𝐴), the logic does not provide a conclusion.
To solve conflicts, DDL employs a so-called superiority relation: a
binary relation over rules establishing the relative strength of rules.
For example, if we have an applicable rule 𝑟 for 𝐴 and a second
applicable rule 𝑠 for ¬𝐴, we can use 𝑟 > 𝑠 to indicate that 𝑟 is
stronger than 𝑠. Accordingly, 𝑟 defeats 𝑠 when both apply, solves
the conflict and allows a coder to conclude for 𝐴.
The superiority relation also provides a simple and effective
mechanism to encode exceptions. Consider the following defeasible
rule:
𝑟 : 𝐴1, . . . , 𝐴𝑛 ⇒ 𝐶
We can model an exception to this rule using a second rule (let us
say 𝑠), where the conclusion is the opposite of the conclusion of
𝑟 , and the IF part of s contains the conditions when the exception
holds. We can formalise this second rule, with the instance 𝑠 > 𝑟 of
141
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the superiority relation, as follows:
𝑠 : 𝐵 1, . . . , 𝐵𝑚 ⇒ ¬𝐶
The reasoning mechanism of DDL, which is based on an argumen-
tation structure, extends the proof theory of Defeasible Logic [1].
To prove a conclusion, there must be an applicable rule for the
said conclusion., and a rule is applicable if all the elements of the
antecedent of the rule hold (i.e., have been proved). All counter-
arguments must also be rebutted or defeated. A counter-argument
is a rule for a conflicting conclusion; that is, the negation of the
conclusion, or in case of deontic conclusions, conflicting deontic
modalities. A counter-argument is rebutted if its premise(s) do not
hold, or a coder proves that the premise(s) do not hold, and the
counter-argument is defeated when the rule is weaker than an
applicable rule for the conclusion. Having outlined the basics of
DDL, as it applies to this experiment, we now turn to our encoding
software.
3.1 Turnip
Turnip1 is a modern (typed) functional programming implementa-
tion of Defeasible Deontic Logic that is written in the programming
language Haskell2 . As previously noted, this software facilitates
the conversion of norms (e.g., different types of regulation) into
computer code.
Turnip requires coders to define all terms before using them in a
set of rules. The basic structure for defining an individual term is:
Type Name description_string
Type is defined in the following table:
Type Keyword Sample Values
Boolean Atom True, False
String String "anything in double quotation marks"
Numeric Numeric 123.456,-5,0
Date Date 1995-02-01
DateTime DateTime 1995-02-01T13:35
Duration Duration 10w, 1d, 5h, 30m
In the experiment that is the subject of this paper, participants
largely used “atoms”, which correspond to literals in DDL and rep-
resent (atomic Boolean) statements that can be either true of false
(e.g., Atom person "is a person"). The description string, which is
the optional text in double quotation marks (" "), defines the atom in
natural language. Coders can also use arithmetic operators (i.e., +,
-, *, /,) for numeric terms and values; comparison operators (i.e.,
==, !=, <, <=, >, >=) to create Boolean types from numeric and
duration terms; and conversion functions (e.g., interval, toDays,
after) that can operate on dates, times and duration terms. Con-
sider, for example, the interval function that takes two dates as
input and returns a duration:
publication . date := 1919 -09 -01
usage . date := 2010 -12 -03
interval ( usage . date , publication . date ) >= 70 y
Here the assignment operator := gives values to two terms of type
date. Then, we use the interval operator to compute the duration
1 An online environment to run Turnip rulesets, with samples of the features it offers
is available at http://turnipbox.netlify.com/.
2 https://www.haskell.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
(i.e., time elapsed between the two dates), and we compare it with
a given duration (70 years).
Rules also have a basic structure that generally includes a label,
a condition list, and a conclusion list. For example:
label : condition_list = > conclusion_list
The arrow (=>) determines the type of rule (e.g., strict, defeasible
or a defeater). It is important to note that rules are designed to
represent norms: a norm prescribes multiple, simultaneous effects,
and different norms can prescribe the same effect [12, p. 180]. To
make the work of coding in DDL more efficient, a condition list
can include a conjunction (&) or disjunction (|) of Boolean, and
a conclusion list can be either an assignment, a single Boolean,
a conjunction of assignments, or a conjunction of Boolean. The
following example illustrates the equivalence between rules with
conjunctions (&) and a disjunction (|) in a condition list:
A & B => C & D A | B => C
A & B => C A => C
A & B => D B => C
Turnip syntax, including negation ~ and numeric and temporal
expression, ultimately allows for deontic expression. A deontic
expression is based on the combination of one of four deontic
modalities: namely, [O], [P], [F], [E] (i.e., Obliged, Permitted,
Forbidden, Exempt) and an atom. For the modalities, "notice that
[F]A is equivalent to [O]~A (and ~[P]A) and [E]A is equivalent to
[P]~A". Given two rule labels, label1 and label2, label1>>label2
denotes the superiority relation between the rules identified by the
labels.
Turnip is of course one of several languages and logics that
coders can use to convert different types of regulation into machine-
executable code [4, 25]. We used this language because it is particu-
larly useful for coders attempting to express complex rule structures.
Take, for example, a statutory provision that establishes multiple
obligations, permissions and at least one prohibition. We argue that
coders can more accurately represent the effects of this provision by
using Turnip’s deontic expressions and disjunctive and/or conjunc-
tive conditions lists. Thus, although Turnip is not the only encoding
software available, it has clear advantages for the purposes of this
experiment.
4 EXPERIMENT DESIGN AND METHODS
We conducted this experiment over a two-week period in late 2020.
The population for this experiment was the pool of legally trained
research assistants attached to the broader research project. While
participation was voluntary, in line with our university ethics ap-
proval3 , participants were paid for their time. We had three partici-
pants in total.
In Week/Phase 1, participants independently coded sections (“ss”)
40, 41, 41A and 42 of the CA, which are among several fair dealing
exceptions to copyright infringement. These exceptions are those
for the purpose of research or study, criticism or review, parody
or satire, and reporting the news, respectively. By “independent”
coding, we mean that participants worked in remote files, rather
3 Wereceived ethics approval for this research at the Queensland University of Tech-
nology (QUT Approval 2000000763).
142
Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley
than the project’s shared GitHub4 repository, and could not com-
municate with each other, or members of the broader research team,
in any way over the course of Week 1. Participants could, however,
raise questions or concerns with the first author (the contact per-
son for the experiment) at any time. Additionally, we instructed
participants to assume that the elements for determining whether
copyright subsists in a work are satisfied (i.e., to assume copyright
subsistence), and to encode the select legislative provisions only.
This means that participants did not encode relevant case law.
In Week/Phase 2, participants coded ss 31, 32 and 36 of the CA.
These provisions outline the nature of copyright in original works,
original works in which copyright subsists, and infringement by
doing acts comprised in the copyright, respectively. Like in Week
1, participants encoded the select provisions only (i.e., they did not
encode case law), and could raise questions or concerns with the
first author at any time. Aside from participants encoding different
provisions in each phase, a choice that we made largely to avoid
participants becoming overly familiar with the statutory text, the
main difference between the two phases of this experiment is that
Week 2 had an intervention. The intervention was a two-hour com-
pulsory meeting during which participants collaboratively drafted
key atoms for manual coding in a single remote file (“Agreed Atom
File”). The Agreed Atom File, which set out and defined key atoms
for ss 31, 32 and 36 of the CA, was separate from documents outlin-
ing our established coding conventions. After this meeting, which
the first author facilitated, participants independently coded the
select provisions like in Week 1. Participants could refer back to, but
not edit, the Agreed Atom File. The purpose of the intervention was
to test our hypothesis that coders collaboratively agreeing on key
atoms before commencing independent coding work increases the
similarity of their coded output (H1). Our null hypothesis was that
coders collaboratively agreeing on key atoms before commencing
independent coding work does not increase the similarity of their
coded output (H0).
In Weeks 1 and 2, participants were allocated up to 7.5 hours
to encode the select provisions. For each phase, participants had
approximately one week to complete their 7.5 hours of work, which
they submitted via email in the form of a single coding file (i.e.,
each participant had one file for Week 1 and another for Week
2). We assigned each participant a pseudonym and deidentified
the participants’ files for both the technical validation and legal
alignment processes, to which we now turn.
4.1 Coding Validation Processes
An important part of converting legislation into machine-executable
code is measuring the apparent success of encoded rules in terms
of coding validation. From a technical perspective, we argue that
an encoded provision is “validated” when it adheres, in a formal
sense, to our select Turnip language and other relevant coding
conventions. In practical terms, this means that the code runs, or
works, and produces a definitive outcome. Coding validation is also
concerned with the degree of internal consistency between coders
to ensure that encoded rules for the same piece of legislation work
together.
4 https://github.com/about.
Converting Copyright Legislation into Machine-Executable Code
Our coding validation process involves automated and manual
analyses of encoded rules. For automated analysis, we created a
program that uses string manipulation to parse each participant’s
atoms into string arrays and, then, compare the participants’ en-
coded rules to find similarities between their approaches. The mea-
sure of apparent “similarity” between any two coders is the number
of shared atoms and rules, respectively, between the coders divided
by the average number of total atoms between the two datasets.
This automated analysis focused on syntax and not semantic mean-
ing. We also manually “cleaned” some aspects of the participants’
coding files to attempt to enhance the comparability of encoded
rules. For example, we adjusted atom names, which used a mixture
of periods ( . ) and camel case (e.g., camelCase), in line with our
project-specific naming conventions. We also filtered out common
stop words, such as “the” and “a”, normalised the tense of atoms
and, where possible, corrected issues with the structure of certain
rules.
4.2 Legal Alignment Processes
We draw an important distinction between coding validation, which
is principally a technical process, and legal alignment, which we
define as the extent to which encoded provisions align with the
languages and logics of the select statutory text [18]. At the heart
of legal alignment is the accurate representation of laws and other
regulation in computer code. Our legal alignment processes are
based on the modern approach to statutory interpretation and
incorporate the Turnip coding language. As previously mentioned,
this approach requires interpreters to examine the text, context and
purpose of select legislation [20, 26], which we break down into a
7-step process:
(1) Locate and read a specific statutory provision, and identify key
words/elements/conditions (known in coding terms as “atoms”).
(2) Read the provision in the context of the Act as a whole. Inter-
preters refine the legal meaning of key words by interpreting
their meaning within the legislation’s full scope. For example,
is a key word defined in the legislation? Under what part of
the legislation is the provision located? This step includes in-
terpreters reading intertextual legislation: for example, other
statutes that are referenced in an Act.
(3) Consider Parliament’s purpose in enacting the legislation. In-
terpreters further refine the legal meaning of key words by
interpreting the provision in line with the legislation’s pur-
pose. For example, the legislation’s object clause, identified
at the start of every statute, can indicate a stated purpose
or purposes. If a statute has multiple purposes, then inter-
preters read the legislation’s establishing documents, such
as speeches in Parliament or the legislation’s explanatory
memorandum, to determine a hierarchy of purposes.
(4) Evaluate interpretive choices. Interpreters identify all possible
interpretations of key words and apply them to findings from
Steps 1 to 3 to evaluate which interpretation best aligns with
the legislation’s context and purpose.
(5) Consider the “canons of construction”. If Step 4 findings are
still ambiguous, then interpreters consider whether and how
syntactical presumptions (e.g., noscitur a sociis and ejusdem
generis) [26, pp. 212, 215], which relate to the meaning of the
143
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
statutory text, and/or statutory presumptions (e.g., legisla-
tion not operating retrospectively), which relate to the scope
and effect of legislation rather than the statutory language,
should apply [26, p. 225].
(6) Return to the provision. Interpreters assign a meaning to the
provision, and its key words, based on findings from Steps 1
to 5. After assigning meaning, interpreters define atoms and
identify modalities (e.g., a permission [P], prohibition [F],
obligation [O] or exemption [E]) and key words/elements/-
conditions for conversion into code.
(7) Acceptance testing of coded provisions. Testing in this con-
text involves coders developing a series of unit tests for
select provisions based on case law or examples in explana-
tory memoranda. Encoded rules "passing" relevant tests is a
strong indicator of legal alignment.
At Stage 6, interpreters start to manually convert natural language
legislation into the machine-executable Turnip language. In prac-
tice, the Turnip language requires interpreters to convert a statutory
provision into a set of if-then statements or rules (rulesets). The
online runtime environment (the Turnip reasoner) can take a set of
rules and facts, respectively, and produce a set of results that are
what the logic can infer from applying the facts to the rules.
While acceptance testing is Step 7, we recommend that inter-
preters start testing encoded rules as early as possible, ideally in
parallel with Stage 6 to optimise legal alignment processes. Devel-
oping and applying rigorous acceptance testing can, for example,
help interpreters to identify edge cases and subtle errors that might
be easily overlooked. While copyright has a doctrinally deep and
rich body of case law[8], participants did not undertake acceptance
testing, principally due to time constraints. Rigorous acceptance
testing for select copyright provisions therefore remains an impor-
tant topic for future research. We provide select examples of our
subsequent acceptance testing in Section 5.2.
4.3 Limitations
The findings in this paper should be interpreted with some limita-
tions in mind. First, there is a small number of participants (N =
3) who encoded different provisions in Week 1 and Week 2. This
means that we cannot directly compare atoms and rules across
weeks, and the results cannot be used to make generalised find-
ings about all coders that might apply our methodology. Second,
the measure of apparent “similarity” between atoms and rules is
based on our methodology, including Turnip syntax, which is a
non-standardised benchmark. Finally, we are not able to reach a
definitive conclusion about the extent to which the encoded rules
align with the statutory text due to the authoritative interpretive
role of the courts under Australia’s constitutional framework [18],
a limitation that will apply to any attempt to encode legislation.
Despite these limitations, which we expand upon in the Results
and Discussion section below, this study makes significant inroads
with developing and applying a methodology that fuses the modern
approach to statutory interpretation with DDL. This paper also
provides valuable insights into sources of apparent coding differ-
ences, legal alignment issues and potential solutions, as part of the
broader RaC movement.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley

5 RESULTS AND DISCUSSION
Overall, the results of our automated syntactical analysis show
a significant increase in the similarity of atoms and shared rules
drafted by participants after the Week 2 intervention, as illustrated
in Table 1. More specifically, Table 2 shows that the similarity of
atoms increased from an average of 4.27% in Phase 1 to 57.64% in
Phase 2. A corollary of this is that the number of unique atoms,
or atoms that are used by one coder only, decreased from Phase
1 to Phase 2. The similarity of rules drafted by participants also
increased, albeit marginally, from 0% in Phase 1 to 1.01% in Phase 2.
These results support our hypothesis that participants agreeing on
key atoms before commencing individual coding work increases
the similarity of their encoding choices (H1). Importantly, as the
individual coders encoded different statutory provisions in Weeks
1 and 2, this cannot simply be attributed to the coders’ increasing
familiarity with the statutory text.
meaning of a statute”[20, p. 118], unless the courts have reached
definitive conclusions about, for instance, what different provisions
mean and how they apply in certain cases. More specific challenges
include converting evidentiary burdens, the exercise of discretion,
and ambiguous and open-textured rules in computer code, intra-
and intertextual provisions and drafting errors, all of which under-
line the potential margin of error in interpreting statutory language
[17]. Due to the nature of statutory interpretation, it is unsurprising
that there were differences in coders’ interpretive choices.
We also observed that participants coded the select provisions
with varying levels of granularity. By “granularity”, we mean the
extent to which coders split, or broke down, the statutory text into
discrete parts. Take, for example, s 40 of the CA that establishes the
fair dealing exception to copyright infringement for the purpose
of research or study. A selection of key atoms originating from ss
40(1), (1A) and (1B) follow:

Week 1 Week 2 Participant One (P1):

Atoms Rules Atoms Rules purposeOf . researchOrStudy
Participant One (P1) 60 68 62 43
Participant Two (P2) 47 59 26 24
Participant Three (P3):
Participant Three (P3) 106 70 51 43
work . producedForPurposeOf . theCourseOfStudy
Table 1: Total Number of Rules and Atoms by Participant work . producedForThePurposeOf . theCourseOfResearch
and Week/Phase work . producedBy . personLecturing . inCourseOfStudy
work . producedBy . personLecturing . inCourseOfResearch
work . producedBy . personTeaching . inCourseOfStudy
work . producedBy . personTeaching . inCourseOfResearch
work . producedBy . personInConnectionWith . courseOfStudy
Week 1 Week 2 work . producedBy . personInConnectionWith . courseOfResearch
P1-P2 atom similarity 11.54% 58.06% Here P1 adopts a high-level approach that combines research and
P1-P3 atom similarity 1.26% 51.85% study into one atom, while P3 uses multiple fine-grained atoms to
P2-P3 atom similarity 0% 62.99% distinguish between, inter alia, research and study. In general, to
Average 4.27% 57.64% be as comprehensive as possible in the first instance, we advocate
Week 1 Week 2 for a fine-grained approach that coders can abstract up if neces-
P1-P2 rule similarity 0% 3.03% sary. This does not mean that this approach is always optimal: for
P1-P3 rule similarity 0% 0% example, some coders might excessively split atoms, or potential
P2-P3 rule similarity 0% 0% end-users might not require or want fine-grained encoded provi-
Average 0% 1.01% sions. Nonetheless, it is clear that differing levels of granularity
Table 2: Percentage Similarity of Atoms and Rules by Partic- contributed to the low average similarity of atoms in Phase 1, with
ipant and Week/Phase P1, P2 and P3 creating 60, 47 and 106 atoms, respectively. Part of this
could also be due to coders’ individual backgrounds, as some partic-
ipants had varying levels of experience in statutory interpretation
and in computer science. It is worth noting that we can potentially
Despite this notable increase in the similarity of coding ap- reconcile the coding differences explored in this paragraph with a
proaches in Week 2, the results show that participants continued to constituent rule like this:
make a range of divergent interpretive choices, for which there are
several potential explanations. The first is the relative difficulty of work . producedForPurposeOf . theCourseOfStudy |
work . producedForThePurposeOf . theCourseOfResearch |
applying the modern approach to statutory interpretation, which,
work . producedBy . personLecturing . inCourseOfStudy |
as previously explained, involves interpreters deriving the mean- work . producedBy . personLecturing . inCourseOfResearch |
ing of a statute from close examination of the text, context and work . producedBy . personTeaching . inCourseOfStudy |
purpose (policy) of select legislation [6, 26]. This approach, as the work . producedBy . personTeaching . inCourseOfResearch |
Honourable Justice Michael Kirby contends, is “an art and not a work . producedBy . personInConnectionWith . courseOfResearch |
work . producedBy . personInConnectionWith . courseOfStudy =>
science”[20, p. 113] and stands in stark contrast to outdated “literal purposeOf . researchOrStudy
positivist”, “plain English” or seemingly “objective” understandings
of statutory interpretation [20, p. 116]. A result is that “[t]he task of Such a rule is arguably concise, in line with P1’s encoded rules for
construing statutory language is notorious for generating opposing ss 40(1), (1A) and (1B), and fine-grained, much like P3’s rendering
answers, no one of which is indisputably correct construing the of conditions.

144
Converting Copyright Legislation into Machine-Executable Code ICAIL’21, June 21–25, 2021, São Paulo, Brazil

P1:
s31_1_c : literaryDramaticMusicalWork . copyrightSubsists & ~ computerProgram & copyright . inRelationTo . work = >
exclusiveRightTo . enterInto . commercialRentalArrangement . workReproducedInSoundRecording

Participant 2 (P2):
s_31_1c : prescribedWorkSection31_1c & copyright . inRelationTo . work = >
exclusiveRightTo . enterInto . commercialRentalArrangement . inSoundRecording

s_31_1c_exception : computerProgram = >

~ exclusiveRightTo . enterInto . commercialRentalArrangement . inSoundRecording

s_31_1c_exception >> s_31_1c

Table 3: Encodings for Section 31(1)(c) CA

Another possible explanation for the differences we observe is Week 1 Week 2

that coders can use the Turnip language in different ways to achieve
the same outcome(s). Take, for instance, the basic statement if A or
B, then C. There are two main coding options:
Option One (Using Disjunction ( | )):
A | B => C
Option Two:
A => C
B => C
A more complex example comes from Week 2 of the experiment
for which participants encoded, inter alia, s 31(1)(c) of the CA. This
provision establishes that copyright in a literary work “(other than
a computer program)”, musical or dramatic work is the exclusive
right “to enter into a commercial rental arrangement in respect of
the work reproduced in a sound recording”. Select encodings for
this provision are presented in Table 3.
Of particular note is how P1 encoded the provision in one rule
and dealt with the statutory text "other than a computer program”
using the ~ (not) operator. P2 adopted a more complex approach by
creating three rules, two of which establish that the exclusive right
does not apply to a computer program (i.e., s_31_1c_exception, and
s_31_1c_exception >>s_31_1c). While these approaches are both
technically sound, or valid when using Turnip syntax as the relevant
benchmark, the former is arguably more straightforward. This
example highlights that differences can arise even when coders
undertake similar training, use the same language and work in the
same team. We suggest that it is likely that there will be comparable
findings if participants used another coding language, but this
would need to be investigated through further in-depth, empirical
research.
5.1 Coding Validation
After undertaking automated syntactical analysis, we manually
“cleaned” the dataset to attempt to improve internal consistency, as
part of the coding validation process explained above. We found
that these measures improved the baseline average similarity of
P1-P2 atom similarity 56.07% 85.71%
P1-P3 atom similarity 30.12% 81.16%
P2-P3 atom similarity 32.68% 87.61%
Average 39.62% 84.86%
Week 1 Week 2
P1-P2 rule similarity 11.36% 35.82%
P1-P3 rule similarity 10.62% 53.49%
P2-P3 rule similarity 28.57% 26.87%
Average 16.85% 38.72%
Table 4: Results of Manual Syntactical Analysis. Percentage
Similarity of Atoms and Rules
atoms from 4.27% to 39.62% in Week 1 and, in Week 2, from 57.64%
to 84.86%. Most notably, coding validation measures significantly
improved the average similarity of rules from 0% to 16.85% in Week
1 and, in Week 2, from 1.01% to 38.72% (See Table 4 for the detailed
results of manual analysis). These results not only underline the im-
portance of shared vocabulary and conventions, but also teams hav-
ing processes in place for managing alternative technically sound
rule structures, with a view to reducing interpretative differences
down the line. While it appears that coding validation improved the
internal consistency of encoded provisions, it is important to note
that “playing with data”[22] can give rise to potential risks, includ-
ing biases and erroneous outcomes. These risks lead us to explore
the importance of interpreters achieving legal alignment between
encoded provisions and the statutory text, which we discuss in the
following subsection.
When coding legislation, we argue that following best practice
and standardised approaches to coding is critical in the long-term.
Transparency, a well-established value of the Anglo-American legal
ideal of the rule of law [32], “is concerned with the quality of being
clear, obvious and understandable without doubt or ambiguity”5 .
At the core of this value is the principle that individuals should be
able to access and understand the rules that impact them [18, 35].
5 Belgium v Commission (C-110/03) [2005] ECR I-2801, [44] (Advocate General Ruiz-
Jarabo Colomer).

145
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Coders using common conventions and developing clean, well-
documented practices is likely to significantly improve the ability of
future users to understand how the legislation has been interpreted
in the encoding process. It would also greatly simplify the task
of future developers who may be called upon to review, audit, or
amend the code.
5.2 Enhancing Legal Alignment
Legal alignment, as distinct from and in addition to coding valida-
tion, is vital for ensuring the accuracy and legitimacy of encoded
legislation. As a further step in checking the accuracy of individual
coding choices, we adopted a process of legal alignment, based on
the methodology above, to attempt to reproduce the legal logic
of select copyright provisions in machine-executable code. Legal
alignment focuses only on our interpretation of the provisions us-
ing a judicially confirmed process as the basis for our interpretive
approach and, critically, is distinct from a claim of “legal validity”
that encoded rules correctly reflect the law. It is impossible to guar-
antee the validity of any encoded legal rules [18]. As noted above,
under Australia’s constitutional framework, only the judiciary can
conclusively interpret the legal meaning of a statute. Even if a body
of case law exists, as it does for copyright law, the nature of the
common law system means that while similar cases will generally
be treated alike [5], future cases with slightly different facts may
trigger a reinterpretation of the law. An added complication is that
the construction of statutes is a question of law and therefore open
to appeal on the basis of errors in statutory interpretation [20].
This underlines that interpreters cannot authoritatively determine
the extent to which their encoding choices are “legally valid”[17].
Instead, we contend that RaC stakeholders should aim for “legal
alignment” between encoded rules and the true construction of the
statutory text.
As noted above, coders were not asked to undertake acceptance
testing in Weeks 1 and 2 due to time constraints. This process was
undertaken by the research team after the technical coding valida-
tion process. The subsequent legal alignment process highlighted
the complexity of attempting to evaluate the extent to which en-
coded rules align with the languages and logics of a statue. To
illustrate this point, it is useful to return to the encoding choices
for section 40 of the CA, for which P1 adopted a high-level ap-
proach and P3 drafted several fine-grained atoms. This provision,
which is one of several exceptions to copyright infringement in
Part III, Division 3 of the Act, establishes that copyright in a work
or an adaptation of a literary, dramatic or musical work is not in-
fringed by a fair dealing for the purpose of research or study. The
wording of this provision is significant; in particular, the legisla-
ture’s use of “research or study” (emphasis added). This raises the
question of whether an interpreter should code these terms con-
junctively (i.e., research and study) or disjunctively (i.e., research or
study). In De Garis v Neville Jeffress Pidler Pty Ltd (1990) 37 FCR 99
ALR 625; 18 IPR 292, Beaumont J stated that the terms “research”
and “study” take their dictionary meanings and, most critically,
should be considered disjunctively[9]. Indeed, the court considered
whether the activities at issue could be characterised as “research”
or “study” for the purposes of s 40 of the CA, separately. This sug-
gests that P1’s decision to encode the terms in one atom; namely,
146
Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley
purposeOf.researchOrStudy, risks deviating from the true meaning
of the statutory text as articulated by the courts. P3’s fine-grained
approach to construing and encoding the statutory language ap-
pears to better align with the court’s interpretation of the statutory
text.
The divergent coding choices for Section 40(2) of the CA provide
another useful illustration of nuanced legal issues that interpreters
could overlook in formal coding validation processes. Section 40(2)
outlines five “. . . matters to which regard shall be had, in determin-
ing whether a dealing constitutes a fair dealing with the work or
adaptation for the purpose of research or study” (emphasis added).
We direct our attention to the word “shall” because it confers a
mandatory or directory obligation [26, p. 333]. We noticed, how-
ever, that only two of the three participants encoded the five matters
as obligations ([O]). Consider the rules in Table 5. Significantly, this
suggests that P1’s choice to encode the matters in s 42(2) as general
atoms rather than obligations deviates from the meaning the legisla-
ture seeks to convey. Such a decision matters because the statutory
text “has specific legal authority”[21, p. 159] and its component
parts — words, symbols and, at times, images — are there for a
reason [23]. It can also have significant flow on effects for poten-
tial end users who might, for example, rely on RegTech to comply
and stay up to date with regulation governing their commercial
activities.
In Australia, like other societies that strive to uphold the legal
ideals of the rule of law and separation of powers [32, 35], public
law provides a range of checks and balances to assess whether a
decision-maker is exercising power in accordance with established
rules and principles [6, 18]. We argue that governments, technology
companies and other key stakeholders must take steps to ensure
that adequate measures are in place to promote alignment between
encoded statutory provisions and the true meaning of the statute as
interpreted by the courts. This underscores the importance of RaC
initiatives having multi-disciplinary expertise across the fields of
law, computer science and public policy [17]. From our perspective
as legally trained coders, we argue that it is vital to have both
subject matter and statutory interpretation expertise as part of a
multi-disciplinary RaC team.
6 CONCLUSION AND FUTURE RESEARCH
The results from our experiment illustrate the complexities of at-
tempting to reproduce the languages and logics of a statutory text
in machine-executable format. While our analyses are preliminary,
we have provided a first of its kind experiment that examines how
different legally trained people interpret and convert legislation
into computer code in practice. After our intervention in Week 2, a
meeting during which participants collaboratively agreed on key
atoms for manual coding, we identified a significant increase in the
average similarity of atoms — from 4.27% in Week 1 to 57.64% in
Week 2 — and, to a lesser extent, rules. This finding, among others,
supports our hypothesis that coders collaboratively agreeing on
key atoms before commencing independent coding work increases
the similarity of their coded output. Importantly, as the individ-
ual participants encoded different statutory provisions in Weeks 1
and 2 of the experiment, the greater similarity we observed cannot
Converting Copyright Legislation into Machine-Executable Code ICAIL’21, June 21–25, 2021, São Paulo, Brazil

P1:
s40_2_aTOd : determining . fairDealing = > regardTo . purposeAndCharacter & regardTo . natureOfWorkOrAdaptation &
regardTo . possibilityOfObtainingWorkOrAdaptation . withinReasonableTime . atOrdinaryCommercialPrice &
regardTo . effectOfDealingOn . potentialMarketOrValue

P2:
s_40_2 : determiningPotentialFairDealing . researchStudy = > [O] regardWorkPurposeCharacter &
[O] regardWorkNature & [O] regardPossibilityOfPurchasing &
[O] regardMarketValueEffect & [O] regardSubstantialityOfCopiedPart

P3:
s_40_2_work_a_to_e : entity . determining . whetherDealingIsFair . forPurposesOf . copyrightAct &
work . isLiteraryOrDramaticOrMusicalOrArtistic & dealing . isReproduction = >
[O] entity . toConsider . dealing . purpose & [O] entity . toConsider . dealing . character &
[O] entity . toConsider . dealing . natureOfWorkOrAdaptation &
[O] entity . toConsider . effectOfDealing . onValueOfWorkOrAdaptation &
[O] entity . toConsider . possibilityOfObtainingWorkOrAdaptationWithinReasonableTimeAtOrdinaryCommercialPrice

Table 5: Encodings for Section 40(2) CA

simply be attributed to the coders’ increasing familiarity with the
statutory text.
Notwithstanding these increases, participants made a range of
divergent interpretive choices, which we argue are most likely due
to: (1) the complexity of statutory interpretation, (2) encoded provi-
sions having varying levels of granularity, and (3) the functionality
of our coding language, Turnip. This underlines that interpretive
differences can arise even when coders undertake similar training,
use the same language and work in the same team. We explained
that interpreters can, to some extent and not without risks, improve
the internal consistency of encoded rules by automatically and/or
manually cleaning the dataset.
Overall, we contend that RaC initiatives should have processes
for technical coding validation and legal alignment, both of which
are critically important in enhancing the accuracy of digitising
legislation. The former helps to ensure that encoded rules adhere
to select coding languages and conventions. It is particularly im-
portant that interpreters not only follow select coding languages
and conventions, but also develop clean, well-documented code to
enable future users to understand how the legislation has been inter-
preted in the encoding process. A second critical step is to engage in
a process of legal alignment, which we conceptualise as enhancing
congruence between the encoded rules and the true meaning of the
select legislation in line with the modern approach to statutory in-
terpretation. The results of this experiment suggest that a rigorous
assessment of legal alignment requires multi-disciplinary expertise
across specific legal subject matters, statutory interpretation and
technical programming.
There are a range of important opportunities for future research
that arise from this study. First, there is scope to expand this ex-
periment to test our hypothesis (H1) across different bodies of law
and types of expertise, and with a larger number of participants.
Secondly, further exploration of what the separate yet interrelated
processes for technical coding validation and legal alignment might
entail in practice is warranted. For example, in terms of coding
validation, our experiment raises practical questions about how
exactly coding teams should choose between different yet equally
technically valid coding choices, and the most appropriate and effi-
cient syntactic and semantic conventions and methodologies for
identifying and representing terms, predicates and propositions in
legal texts. Finally, in terms of coding validation, further research
is needed to shed light on acceptance testing options, including en-
coding case law, for copyright law and beyond. An important part
of this work will be clarifying best practices from both a technical
and legal perspective.
REFERENCES
[1] Grigoris Antoniou, David Billington, Guido Governatori, and Michael J. Ma-
her. 2001. Representation Results for Defeasible Logic. ACM Transactions on
Computational Logic 2, 2 (2001), 255–287.
[2] Patricia Aufderheide, Kylie Pappalardo, Nicolas Suzor, and Jessica Stevens. 2018.
Calculating the consequences of narrow Australian copyright exceptions: mea-
surable, hidden and incalculable costs to creators. Poetics 69 (2018), 15–26.
[3] Tom Barraclough, Hamish Fraser, and Curtis Barnes. 2021. Legislation as Code
for New Zealand: Opportunities, Risks and Recommendations. Report. Brainbox
and The New Zealand Law Foundation, New Zealand.
[4] Sotiris Batsakis, George Baryannis, Guido Governatori, Tachmazidis Ilias, and
Grigoris Antoniou. 2018. Legal Representation and Reasoning in Practice: A Crit-
ical Comparison. In Legal Knowledge and Information Systems, Monica Palmirani
(Ed.). IOS Press, 31–40.
[5] Lisa Burton Crawford and Dan Meagher. 2020. Statutory Precedents under the
“Modern Approach” to Statutory Interpretation. Sydney Law Review 42, 2 (2020),
209–239.
[6] Lisa Burton Crawford, Maria O’Sullivan, Janina Boughey, and Melissa Castan.
2017. Public Law and Statutory Interpretation: Principles and Practice. Federation
Press, NSW, Australia.
[7] Commonwealth of Australia. 2013. Copyright and the Digital Economy: Discussion
Paper. ALRC Discussion Paper 79. https://www.alrc.gov.au/wp-content/upload
s/2019/08/dp79_whole_pdf_.pdf
[8] Mark Davidson, Ann Monotti, and Leanne Wiseman. 2012. Australian Intellectual
Property Law. Cambridge University Press, Cambridge, U.K.
[9] Bronwen Claire Ewen. 2017. 240 – Intellectual Property, III COPYRIGHT, (8)
DEFENCES TO INFRINGEMENT (B) Fair Dealing – Defences to Infringement of
Copyright. In Halsbury’s Laws of Australia. LexisNexis Australia, [240–2357].

147
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
[10] Dagfinn Føllesdal and Risto Hilpinen. 1971. Deontic Logic: An Introduction. In
Deontic Logic: Introductory and Systematic Readings, Risto Hilpinen (Ed.). North
Holland, 1–35.
[11] Thomas F. Gordon, Guido Governatori, and Antonino Rotolo. 2009. Rules and
Norms: Requirements for Rule Interchange Languages in the Legal Domain
(LNCS, 5858), Guido Governatori, John Hall, and Adrian Paschke (Eds.). Springer,
Heidelberg, 282–296.
[12] Guido Governatori, Pompeu Casanovas, and Louis de Koker. 2020. On the Formal
Representation of the Australian Spent Conviction Scheme. In Rules and Reasoning
(LNCS, Vol. 12173), Víctor Gutiérrez Basulto, Tomáš Kliegr, Ahmet Soylu, Martin
Giese, and Dumitru Roman (Eds.). Springer International, 177–185.
[13] Guido Governatori, Mustafa Hashmi, Ho-Pun Lam, Serena Villata, and Mon-
ica Palmirani. 2016. Semantic Business Process Compliance Checking Using
LegalRuleML. In Knowledge Engineering and Knowledge Management (LNAI,
10024), Eva Blomqbvist, Paolo Ciancarini, Francesco Poggi, and Fabio Vitali (Eds.).
Springer International, 746–761.
[14] Guido Governatori, Francesco Olivieri, Antonino Rotolo, and Simone Scannapieco.
2013. Computing Strong and Weak Permissions in Defeasible Logic. Journal of
Philosophical Logic 42, 6 (2013), 799–829.
[15] Guido Governatori and Antonino Rotolo. 2006. Logic of Violations: A Gentzen
System for Reasoning with Contrary-To-Duty Obligations. Australasian Journal
of Logic 4 (2006), 193–215.
[16] Guido Governatori, Antonino Rotolo, and Erica Calardo. 2012. Possible World
Semantics for Defeasible Deontic Logic. In Deontic Logic in Computer Science
(DEON 2012) (Lecture Notes in Computer Science, Vol. 7393), Thomas Ågotnes, Jan
Broersen, and Dag Elgesem (Eds.). Springer, Heidelberg, 46–60.
[17] Anna Huggings, Alice Witt, Nicholas Suzor, Mark Burdon, and Guido Governatori.
2020. Financial Technology and Regulatory Technology, Issues Paper Submission.
Submission 196. Select Senate Committee on Financial Technology and Regula-
tory Technology. https://www.aph.gov.au/DocumentStore.ashx?id=30153f40-
c456-4398-99f2-0dd627f86401&subId=699554
[18] Anna Huggins. 2020. Executive Power in the Digital Age: Automation, Statutory
Interpretation and Administrative Law. In Interpreting Executive Power, Janina
Boughey and Lisa Burton Crawford (Ed.). Federation Press, 111–128.
[19] Mohammad Badiul Islam and Guido Governatori. 2018. RuleRS: A rule-based ar-
chitecture for decision support systems. Artificial Intelligence and Law 26, 4 (2018),
315–344. https://doi.org/10.1007/s10506-018-9218-0 arXiv:http://rdcu.be/HIvL
[20] Michael Kirby. 2011. Statutory Interpretation: The Meaning of Meaning. Mel-
bourne University Law Review 35, 1 (2011), 113–133.
148
Alice Witt, Anna Huggins, Guido Governatori, and Joshua Buckley
[21] Michael Kirby. 2012. The Never-Ending Challenge of Drafting and Interpreting
Statutes – A Meditation on the career of John Finemore QC. Melbourne University
Law Review 36, 1 (2012), 140.
[22] David Lehr and Paul Ohm. 2017. Playing with the Data: What Legal Scholars
Should Learn About Machine Learning. University of California Davis Law Review
51, 2 (2017), 653–717.
[23] John Middleton. 2016. Statutory Interpretation: Mostly Common Sense? Mel-
bourne University Law Review 40, 2 (2016), 626–656.
[24] James Mohun and Alex Roberts. 2020. Cracking the code: Rulemaking for humans
and machines. OECD Working Papers on Public Governance. OECD, Paris, France.
https://doi.org/10.1787/3afe6ba5-en
[25] Jason Morris. 2020. Spreadsheets for Legal Reasoning: The Continued Promise of
Declarative Logic Programming in Law. Masters Thesis, University of Alberta.
[26] Michelle Sanson. 2016. Statutory Interpretation (2nd ed.). Oxford University Press„
Oxford, U.K.
[27] Service Innovation Lab (LabPlus). 2018. Better Rules for Government Discovery
Report. Technical Report.
[28] Nicolas Suzor. 2014. The only way to fix copyright is to make it fair. The
Conversation (21 February 2014). https://theconversation.com/the-only-way-to-
fix-copyright-is-to-make-it-fair-23402
[29] Matthew Waddington. 2019. Machine-Consumable Legislation: A Legislative
Drafter’s Perspective – Human v Artificial Intelligence. The Loophole 2 (2019),
21–52.
[30] Matthew Waddington. 2020. Rules as Code. Law in Context 37, 1 (2020), 179–186.
[31] Vicki Waye. 2019. Regtech: A New Frontier in Legal Scholarship. Adelaide Law
Review 40, 1 (2019), 363–386.
[32] Alice Witt, Nicolas Suzor, and Anna Huggins. 2019. The Rule of Law on Instagram:
An Evaluation of the Moderation of Images Depicting Women’s Bodies. UNSW
Law Journal, 42, 2 (2019), 557–596.
[33] Meng Weng (HUANG Mingrong) Wong. 2020. Rules as Code – Seven Levels of
Digitisation. Report. Singapore Management University, School of Law, Singa-
pore.
[34] World Government Summit. 2018. RegTech for Regulators: Re-Architect the System
for Better Regulation. Technical Report. 2018. https://www.worldgovernmentsum
mit.org/api/publications/document?id=5ccf8ac4-e97c-6578-b2f8-ff0000a7ddb6
[35] Monika Zalnieriute, Lyria Bennett Moses, and George Williams. 2019. The Rule
of Law and Automation of government Decision-Making. Modern Law Review
82, 3 (2019), 425–427.
 Hardness of Case-Based Decisions: a Formal Theory
Heng Zheng Davide Grossi
Artificial Intelligence, Bernoulli Artificial Intelligence, Bernoulli
Institute, University of Groningen Institute, University of Groningen
The Netherlands ILLC/ACLE, University of
h.zheng@rug.nl Amsterdam
The Netherlands
d.grossi@rug.nl
ABSTRACT
Stare decisis is a fundamental principle of case-based reasoning. Yet
its application varies in complexity and depends, in particular, on
whether relevant past decisions agree, or exist at all. The contribution
of this paper is a formal treatment of types of the hardness of case-
based decisions. The typology of hardness is defined in terms of
the arguments for and against the issue to be decided, and their
kind of validity (conclusive, presumptive, coherent, incoherent). We
apply the typology of hardness to Berman and Hafner’s research on
the dynamics of case-based reasoning and show formally how the
hardness of decisions varies with time.
CCS CONCEPTS
• Computing methodologies → Knowledge representation and
reasoning; • Applied computing → Law.
KEYWORDS
Case-based reasoning, computational argumentation, hard cases
ACM Reference Format:
Heng Zheng, Davide Grossi, and Bart Verheij. 2021. Hardness of Case-
Based Decisions: a Formal Theory . In Eighteenth International Confer-
ence for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São
Paulo, Brazil. ACM, New York, NY, USA, 11 pages. https://doi.org/10.1145/
3462757.3466071
1 INTRODUCTION
Legal decision-making can be hard, very hard. Its complexities,
which have been well-recognized in AI and Law since its early
days, are numerous. An early contribution to the discussion of the
hardness of legal decision-making in AI and Law is ‘An Artificial
Intelligence Approach to Legal Reasoning’, a book by Gardner [9].
In that book, Gardner addresses the distinction between hard and
easy cases using ideas from jurisprudence. Following Rissland’s
review [18] of the landmark work by Gardner, legal decisions are
guided rather than governed by existing law; legal terms are open
textured; legal questions can have more than one answer, but a
reasonable and timely answer must be given; and the answers to legal
questions can change over time. These, and other, such complexities
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466071
149
Bart Verheij
Artificial Intelligence, Bernoulli
Institute, University of Groningen
The Netherlands
bart.verheij@rug.nl
of legal decision-making have been object of much work in AI and
Law (cf. also [23]).
But some cases are easier to decide than others. For instance,
when all past cases agree on a given legally relevant fact situa-
tion, decision-making using the principle of stare decisis can be
straightforward. When relevant past cases disagree, things get harder.
Sometimes such a conflict of precedents can be resolved, for instance
when one precedent is considered a landmark case overturning ex-
isting doctrine, or when a precedent comes from a higher level
court. But not all conflicts can be resolved, making decision-making
harder. Also it can happen that a legally relevant fact situation has no
matching precedent, so the stare decisis principle gives no answer.
Paper contribution. As the above examples show, the hardness
of case-based decisions comes in different types. It is the topic
of this paper to provide a formal theory of the hardness of cases
in case-based decision-making. Significant work has been devoted
to the nature and dynamics of case-based reasoning (e.g., [1, 2, 5–
8, 11, 13–17, 21]), and to the topic of hard cases (e.g., [4, 9, 12]). Yet,
to the best of our knowledge, no formal theory has been proposed
so far of what makes a current case harder, or less hard, than other
cases. We provide such a theory here, by focusing on the following
question: is there a typology of how hard it is to make a decision
about an issue in case-based reasoning? To answer this question,
we propose a formal approach based on the case model formalism
[20–22, 25, 26]. We describe a decision-making issue as an argument
and its counterargument, and formalize its hardness with the validity
of these arguments. We also illustrate the approach with a case study
in the dynamics of case-based reasoning (following an example by
Berman and Hafner [6–8, 11] as formalized in [21]), and show how
hardness varies over time.
Paper outline. Section 2 introduces earlier work in the case model
formalism. Section 3 develops a formal theory of hardness. Section 4
shows an application of our approach to a series of concrete legal
cases highlighting the development of hardness over time. Section 5
positions our theory within existing literature on case-based decision-
making in law. Section 6 concludes. Detailed proof sketches of
relevant formal properties are provided throughout the paper.
2 PRELIMINARIES: CASE MODELS
Our approach uses case models [20], a formalism based on a propo-
sitional logic language L generated from a finite set of constants. We
write ¬ for negation, ∧ for conjunction, ∨ for disjunction, ↔ for
equivalence, ⊤ for a tautology, and ⊥ for a contradiction. The asso-
ciated classical logical consequence relation is denoted |=. Cases can
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
π2
π0
π1
P∧Q ¬P
P ∧ ¬Q
Figure 1: Example of a case model. Larger boxes denote cases
that are preferred over the cases denoted by smaller boxes.
be compared through the preference relation between cases in case
models. A case model is a set of logically consistent, incompatible
cases forming a total preorder (i.e., a transitive, total binary relation)
representing a preference relation among the cases.
Definition 1 (Case models [20]) A case model is a pair C = (C, ≥)
with finite C ⊆ L, such that, for all π, π ′ and π ′′ ∈ C:
(1) ̸|= ¬π;
(2) If ̸|= π ↔ π ′ , then |= ¬(π ∧ π ′ );
(3) If |= π ↔ π ′ , then π = π ′ ;
(4) π ≥ π ′ or π ′ ≥ π;
(5) If π ≥ π ′ and π ′ ≥ π ′′ , then π ≥ π ′′ .
As customary, the asymmetric part of ≥ is denoted >. The symmetric
part of ≥ is denoted ∼. Intuitively, ≥ means ‘at least as preferred
as’. For instance, P ∧ Q ≥ P ∧ ¬Q means that the case expressed by
P ∧ Q is higher than or equally high in the preference ordering than
the case expressed by P ∧ ¬Q.
Example 1 ([21]) Cases π0 = P ∧ Q, π1 = P ∧ ¬Q, and π2 = ¬P, and
preference relation π2 > π0 > π1 form a case model (Figure 1).
We move now to the definition of arguments and their validities.
Definition 2 (Arguments [20]) An argument from χ to ρ is a pair
(χ, ρ) with χ and ρ ∈ L. For λ ∈ L, if χ |= λ , λ is a premise of the
argument; if ρ |= λ , λ is a conclusion; if χ ∧ ρ |= λ , λ is a position
made by the argument. We say that χ expresses the full premise of
the argument, ρ the full conclusion, and χ ∧ ρ its full position made
by the argument.
Example 2 For instance, (P, Q) is an argument with P as its full
premise, and Q as its full conclusion. Sentence P ∧ Q is the full
position by the argument (P, Q).
Arguments have three kinds of validities. If the full position made
by an argument is logically implied by one of the precedents in a
case model, then the argument is coherently valid in the case model.
If an argument’s conclusion is logically implied by a precedent
which is weakly preferred over all precedents that logically imply
the argument’s full premise, then the argument is presumptively
valid in the case model. If all precedents that logically imply the full
premise of a coherently valid argument, also logically imply its full
conclusion, then the argument is conclusively valid in the model.
Definition 3 (Validity of arguments [20]) Let C = (C, ≥) be a case
model and (χ, ρ) an argument:
(1) (χ, ρ) is coherent w.r.t. C if and only if ∃π ∈ C : π |= χ ∧ρ. We
then write C |= χ⊤ρ. If C ̸|= χ⊤ρ, then (χ, ρ) is incoherent
w.r.t. C . We then write C |= χ⊥ρ.
(2) (χ, ρ) is presumptive w.r.t. C if and only if ∃π ∈ C s.t.:
(a) π |= χ ∧ ρ; and
(b) for all π ′ ∈ C: if π ′ |= χ, then π ≥ π ′ .
We then write C |= χ { ρ.
150
H. Zheng, D. Grossi and B. Verheij
All arguments
Coherent arguments
Presumptive arguments
Conclusive arguments
Figure 2: Validity of arguments in a Venn diagram.
(3) (χ, ρ) is conclusive w.r.t. C if and only if ∃π ∈ C : π |= χ ∧ ρ
and for all π ∈ C: if π |= χ, then π |= χ ∧ ρ. We then write
C |= χ ⇒ ρ.
Example 3 If we consider arguments (P, Q), (¬P, Q), and (P ∨ R, P)
with the case model C shown in Figure 1, we have:
(1) C |= P⊤Q, since π0 |= P ∧ Q;
(2) C |= ¬P⊥Q, since there is no case in C can imply ¬P ∧ Q;
(3) C |= P { Q, since π0 |= P ∧ Q, and π0 is the most preferred
case in C that implies P;
(4) C |= P ∨ R ⇒ P, since for all cases in C , if they imply P ∨ R,
they also imply its conclusion P.
The following proposition shows relations between validities.
Proposition 1 Let C be a case model, (χ, ρ) be an argument:
(1) If C |= χ ⇒ ρ, then C |= χ { ρ, but not in general vice versa;
(2) If C |= χ { ρ, then C |= χ⊤ρ, but not in general vice versa.
(3) For all χ, ρ ∈ L, either C |= χ⊤ρ or C |= χ⊥ρ.
P ROOF. We prove the first claim. The other cases are similar.
Observe that since (χ, ρ) is conclusively valid, by Definition 3, for
all π ∈ C, if π |= χ, then π |= χ ∧ ρ. Furthermore, a case π |= χ ∧ ρ
exists again by the definition of conclusive validity. There are two
cases. Either such a π is maximally preferred among the cases
logically implying χ, and we have π |= χ ∧ ρ as desired. Or it is not.
There exists then a maximally preferred π ′ ∈ C such that π ′ |= χ and
π ′ ≥ π. But by the above observation we also have that π ′ |= ρ as
desired. We conclude that (χ, ρ) is presumptively valid. □
Figure 2 is an illustration of Proposition 1. Conclusive arguments
are presumptively valid, and presumptive arguments are coherently
valid. All arguments are either incoherent with respect to the case
model (outside the set of coherent arguments), or coherent with
respect to the model (inside the set of coherent arguments).
3 A FORMAL THEORY OF HARDNESS
We turn to the main theoretical contribution of the paper: a formal
approach to compare issues in case models by their hardness.
We start by introducing the key definitions underpinning our
theory, in two steps. First we introduce a natural way of ordering
arguments by their strength, in essence based on Proposition 1. We
show then how this notion can be used also to develop ways in which
issues can be compared. We think of issues essentially as pairs of
arguments (e.g., the plaintiff’s and the defendant’s arguments): the
argument that points to the truth of the issue, an the one that points
Hardness of Case-Based Decisions: a Formal Theory
to its falsity. Here arguments are premise-conclusion pairs, without
considering a possible internal stepwise structure. Based on two
natural ways to compare issues, we will define and study an ordering
of issues representing their relative hardness.
3.1 Comparing arguments
First we introduce labels for arguments based on their validity.
Definition 4 (Validity labels for arguments) Let (χ, ρ) be an argu-
ment and C a case model. Then the validity label of (χ, ρ) in C is
denoted AC (χ, ρ) and is defined as follows:
(1) AC (χ, ρ) = conc if C |= χ ⇒ ρ;
(2) AC (χ, ρ) = pres if C |= χ { ρ and C ̸|= χ ⇒ ρ;
(3) AC (χ, ρ) = coh if C |= χ⊤ρ and C ̸|= χ { ρ;
(4) AC (χ, ρ) = incoh if C ̸|= χ⊤ρ.
For instance, pres is the label for arguments that are presumptively
valid, but not conclusive. Using Figure 2 as an illustration, these
arguments are in the presumptive arguments set, but not in the con-
clusive arguments set. The label incoh is used for arguments that
are not coherent. In Figure 2, these arguments are in the set of all
arguments, but not in the set of coherent arguments.1
Validity labels come with a natural ordering:
Definition 5 (Validity ordering) The validity ordering is the total or-
der ≥ on set {conc, pres, coh, incoh} characterized by the following
property:2 conc ≥ pres ≥ coh ≥ incoh.
Intuitively, based on their validity, arguments with label conc are
stronger than arguments with label pres. Similarly, arguments with
label pres are stronger than arguments with label coh, and arguments
with label coh are stronger than arguments with label incoh.
The following proposition relates argument validities (Defini-
tion 3) to validity labels and validity ordering (Definitions 4 and 5).
Proposition 2 Let (χ, ρ) be an argument and C a case model. Then
the following hold:
(1) (χ, ρ) is conclusive if and only if AC (χ, ρ) = conc;
(2) (χ, ρ) is presumptive if and only if AC (χ, ρ) ≥ pres;
(3) (χ, ρ) is coherent if and only if AC (χ, ρ) ≥ coh;
(4) (χ, ρ) is incoherent if and only if AC (χ, ρ) = incoh.
P ROOF. Immediate using the definitions and Proposition 1.
Using the ordering of arguments their validity label, we can quantify
how far apart they are in terms of strength, as follows:
Definition 6 (Validity distance) Let v0 and v1 ∈ {conc, pres, coh,
incoh}. Then we define the validity distance vd(v0 , v1 ) as the length
of the shortest path from v0 to v1 in the validity ordering.
Intuitively, the validity distance between two arguments is the num-
ber of steps in the validity ordering ≥ that are needed to go from
the label of the weaker argument to the label of the stronger one
(of course avoiding loops). Clearly, as ≥ has length 3, the maximal
validity distance is 3 and the minimal is 0.
1
These labels and their ordering in the following Definition 5 are related to the quantita-
tive representation of [20, Section 3.3]. Arguments with label conc correspond to those
arguments with strength equal to 1 [20]. Arguments with label pres corresponds to those
arguments with strength above a given threshold but less than 1. Arguments with label
coh corresponds to arguments with strength less than the given threshold but still above
0. And arguments with label incoh corresponds to arguments with strength equal to 0.
2
Recall that a total order is a binary relation which is transitive, total and antisymmetric.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Example 4 In the case model of Figure 1, argument (P, Q) has type
pres, and argument (P, ¬Q) has type coh. Hence the validity distance
between the validity labels of these two arguments is 1.
3.2 Issues
We introduce now the notion of issue, that is, the specific proposition
a case model is supposed to decide about. Our theory concerns
precisely the hardness of deciding about the truth or falsity of an
issue given a situation in a case model.
Issues and issue types. So an issue is a sentence whose truth or falsity
we would like to establish, given a situation. Formally:
Definition 7 (Issues) A situation is a sentence σ ∈ L. A sentence
ι ∈ L is an issue given situation σ (denoted σ ± ι) if and only if
σ ̸|= ι and σ ̸|= ¬ι.
In other words, an issue for a given situation is a sentence whose
truth or falsity is not logically settled by the situation. It is worth
observing that σ ± ι is an issue if and only if σ ± ¬ι is an issue.
Example 5 For instance, P ± Q represents an issue Q with respect
to a situation P.
Importantly, for every issue σ ± ι there are two naturally associated
arguments: (σ , ι) and (σ , ¬ι). We will study hardness as a relation
on the types of those pairs of arguments that correspond to issues.
Given the pair of arguments induced by an issue, we call a type
the multi-set (i.e., a set admitting multiple copies of an element)
consisting of the validity labels of the two arguments. Formally:
Definition 8 (Types of issues) Let C be a case model and σ ± ι an
issue. Then the type of the issue, denoted TC (σ ± ι), is the multiset:
{AC (σ , ι), AC (σ , ¬ι)}.
Example 6 The type of issue P ± Q with respect to the case model
in Figure 1 is {pres,coh}, as AC (P, Q) = pres and AC (P, ¬Q) = coh.
Not all types are logically possible. The following proposition shows
there are in total 7 possible issue types.
Proposition 3 Let C be a case model and σ ± ι an issue. Then
TC (σ ± ι) equals one of the following:
(1) {conc, incoh};
□ (2) {pres, pres};
(3) {pres, coh};
(4) {pres, incoh};
(5) {coh, coh};
(6) {coh, incoh};
(7) {incoh, incoh}.
P ROOF. By Definition 4, an argument can have 4 possible labels
(42 −4)
conc, pres, coh, incoh. There are therefore 2 +4 = 10 possible
multisets of size 2. Let a case model C = (C, ≥) and an issue σ ± ι
be given. We split the proof in two parts. First, we reason by cases
and for a multiset {v0 , v1 } (for which we can assume v0 ≥ v1 ) we
show what values among conc, pres, coh, incoh v1 can take once
we fix v0 . Then we display examples showing the types listed in the
statement are possible.
v0 = conc Then v1 ∈ {conc, pres, coh, incoh}. If AC (σ , ι) =
conc, by Definition 3, for each π ∈ C with π |= σ , we have π |= σ ∧ ι.
Then since cases are consistent, π ̸|= σ ∧ ¬ι, so AC (σ , ¬ι) = incoh
151
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
and the types {conc, conc}, {conc, pres}, and {conc, coh} are not
possible. Similarly, when AC (σ , ¬ι) = v0 . Therefore, TC (σ ± ι) is
equal to {conc, incoh}.
v0 = pres Then v1 ∈ {pres, coh, incoh} and AC (σ , ¬ι) ∈ {pres,
coh, incoh}. Similarly, when AC (σ , ¬ι) = v0 . Therefore, TC (σ ± ι)
is equal to {pres, pres}, {pres, coh}, or {pres, incoh}.
v0 = coh Then v1 ∈ {coh, incoh}, and TC (σ ± ι) is equal to
{coh,coh} or {coh,incoh}.
v0 = incoh Then v1 = incoh and TC (σ ± ι) is {incoh,incoh}.
Now we give an example for each possible type. For {conc, incoh},
let C be a case model with only one case π0 = P ∧ Q, P ± Q an issue.
TC (P ± Q) = {conc, incoh}.
Let C be a case model with cases π0 = P ∧ Q and π1 = P ∧ ¬Q,
P ± Q an issue. For {pres, pres}, if C has a preference relation
π0 ∼ π1 , then TC (P ± Q) = {pres, pres}. For {pres, coh}, if C has
a preference relation π0 > π1 , then TC (P ± Q) = {pres, coh}.
For {pres, incoh}, let C be a case model with cases π0 = P ∧ Q
and π1 = P∧R, P±Q an issue. If C has a preference relation π0 > π1 ,
then TC (P ± Q) = {pres, incoh}.
For {coh, coh}, let C be a case model with cases π0 = P, π1 =
P∧Q, and π2 = P∧¬Q, P±Q an issue. If C has a preference relation
π0 > π1 ∼ π2 , then TC (P ± Q) = {coh, coh}.
For {coh, incoh}, let C be a case model with cases π0 = P and
π1 = P ∧ Q, P ± Q an issue. If C has a preference relation π0 > π1 ,
then TC (P ± Q) = {coh, incoh}.
For {incoh, incoh}, let C be a case model only with case π0 = P,
P ± Q an issue. Then TC (P ± Q) = {incoh, incoh}. □ of ≥ (Definition 5) and is therefore reflexive, antisymmetric, and
Proposition 3 depends on the preference relation between cases
in C to be general. If restrictions are imposed on that preference
relation, fewer types may be possible. In particular, if such a pref-
erence relation is trivial (in the sense that all cases are at least as
preferred as all other cases), like in the case models representing
HYPO examples [25], then only 4 types are possible: {conc,incoh},
{pres,incoh}, {pres,pres} and {incoh,incoh} since then a coherent
argument is always presumptive.
Comparing issues. There are two natural ways in which to compare
types. They can be compared by the relative strength of their validity
labels, or by the distance of their labels. Formally:
Definition 9 (Type orderings) Let {v0 , v1 } and {v′0 , v′1 } be types. We
define two binary relations ⪰v and ⪰d ∈ {conc, pres, coh, incoh}2 as
follows:
(1) {v0 , v1 } ⪰v {v′0 , v′1 } if and only if: v0 ≥ v′0 and v1 ≥ v′1 , or
v0 ≥ v′1 and v1 ≥ v′0 ;
(2) {v0 , v1 } ⪰d {v′0 , v′1 } if and only if vd(v0 , v1 ) ≥ vd(v′0 , v′1 ).
The asymmetric parts of ⪰v , ⪰d are respectively denoted ≻v and
≻d . Their symmetric parts are respectively denoted ∼v and ∼d .
Intuitively, relation ⪰v orders types by the strength of their va-
lidity labels defined in Definition 5. So higher types in ⪰v pertain
issues involving stronger arguments, while lower types in ⪰v pertain
issues involving weaker arguments. Instead, relation ⪰d orders types
by how far apart the labels within the type are from each other. So
higher types in ⪰d pertain issues involving arguments containing a
strong and a weak argument, while lower types in ⪰d pertain issues
involving arguments of similar strength.
152
H. Zheng, D. Grossi and B. Verheij
3 2 1 0
{conc, incoh} {pres, pres}
{pres, coh}
{pres, incoh} {coh, coh}
{coh, incoh}
{incoh, incoh}
Figure 3: Hasse diagram of ⪰v . The numbered columns depict
the equivalence classes of ⪰d (one per distance from 0 to 3).
Example 7 Continuing on Example 6, we have:
{pres,coh} ⪰v {incoh,incoh};
{pres,coh} ⪰d {incoh,incoh}.
We will see these relations at work in the definition of hardness
ordering provided in the next subsection. For now it is important to
observe that these relations are well-behaved:
Proposition 4 We have that:
(1) ⪰v is a partial order, which is not total;
(2) ⪰d is a total preorder, which is not antisymmetric.
P ROOF. Claim 1. By Definition 9, ⪰v inherits the properties
transitive; hence a partial order. However, it is not total since some
types are incomparable, for instance {conc,incoh} and {pres,coh}
can not be compared since conc≥pres while incoh≤coh .
Claim 2. By Definitions 9 and 6, an integer is associated to each
type. Then ⪰d inherits the properties of the integer ≥ relation: reflex-
ivity, transitivity and totality. However, it is not antisymmetric, for in-
stance, {pres,coh} ⪰d {coh,incoh} and {coh,incoh} ⪰d {pres,coh}
(both have validity distance 1), but {pres,coh} ≠ {coh,incoh}. □
Relations ⪰v and ⪰d are depicted in Figure 3.
3.3 Hardness
We view hardness as a relation between issues: harder vs. easier
issues. The intuition that guides our definition is based on the re-
lations ⪰d and ⪰v introduced in the previous section. An issue is
‘easy’ if the validity labels of the two arguments involved in the
issue are far apart by their validity distance: the stronger argument
prevails. By means of illustration, an issue where the two arguments
are one of type conc and one of type incoh will be easy to decide.
It is ‘hard’ if, vice versa, the validity labels of the two arguments
involved in the issue are close by validity distance. The prototypical
case consists of arguments with the same validity label. It is in such
cases that one can then use relation ⪰v to distinguish among issues
whose arguments have same validity distance. Intuitively, if two
issues both involve arguments with the same validity distance—like
for instance {pres, coh} and {coh, incoh}—it is the issue involving
stronger arguments that is arguably ‘easier’.
These intuitions back the definition of hardness as an ordering
over issue types, which is based on the lexicographic combination
Hardness of Case-Based Decisions: a Formal Theory
of ⪰d and ⪰v . In our definition of the hardness ordering, an issue
type that is higher in the ordering is ‘easier’.
Definition 10 (Hardness ordering) Let C , C ′ be two case models
and σ ± ι and σ ′ ± ι ′ two issues. The hardness ordering ⪰h is a
binary relation over types defined as follows.
′ ′
TC (σ ± ι) ⪰h TC ′ (σ ± ι )
if and only if:
(1) TC (σ ± ι) ≻d TC ′ (σ ′ ± ι ′ ), or
(2) TC (σ ± ι) ∼d TC ′ (σ ′ ± ι ′ ) and TC (σ ± ι) ⪰v TC ′ (σ ′ ± ι ′ ).
The asymmetric (symmetric) part of ⪰h is denoted ≻h (∼h ).
Example 8 In the case model shown in Figure 1, TC (P ± Q) =
{pres, coh}, TC (P ± R) = {incoh, incoh}. By the hardness order-
ing, TC (P ± Q) ≻h TC (P ± R).
We now show that our definition of hardness is well-behaved. In
particular, it is transitive and can compare any pair of issue types.
Theorem 1 The relation ⪰h is a total order.
P ROOF. We need to prove that ⪰h is transitive, antisymmetric
and total. Let x, y, z be issue types.
Transitivity Assume x ⪰h y and y ⪰h z. There are 4 circum-
stances:
(1) If x ≻d y and y ≻d z, then by Proposition 4, x ≻d z, hence by
Definition 10, x ⪰h z;
(2) If x ≻d y and y ∼d z, then by Definition 9 x ≻d z, hence by
Definition 10, x ⪰h z;
(3) If x ∼d y and y ≻d z, then by Definition 9, x ≻d z, hence by
Definition 10 x ⪰h z;
(4) If x ∼d y and y ∼d z, then by Definition 10, x ⪰v y and y ⪰v z.
Then by Proposition 4, x ⪰v z, hence x ⪰h z;
Totality For types x and y, if the validity distance between
the arguments in x is not equal to the validity distance in y, then
by Definition 10 and Proposition 4, x and y are comparable via
⪰h because of the relation ⪰d . If the validity distance between
the arguments in x is equal to the validity distance in y, then by
Definition 10, x and y are comparable via ⪰h using ⪰v .
Antisymmetry If x ⪰h y and y ⪰h x, then by Definition 10, x ⪰d
y, y ⪰d x, hence x ∼d y and therefore x ⪰v y, y ⪰v x. Then using
Proposition 4, we find x = y. Hence ⪰h is antisymmetric. □ where there are decisions based on both an expert and a non-expert.
The hardness ordering is depicted in Figure 4. The formal notion of
hardness captured by Definition 10 provides us with a systematic
way to categorize easy and hard decisions in case-based reasoning,
once represented within the case model formalism. The remaining
of the paper is dedicated to putting this formal notion of hardness
to the test by further discussing the intuitions underpinning it, and
putting it at work in concrete examples of case-based decisions.
3.4 An illustration of the theory
We now give examples that are meant to illustrate all possible types
of an issue (summarized in Figure 5). In the next section, we will
give a realistic example for discussing the hardness of issues.
We assume there are two possible outcomes for the current situa-
tion to be considered. We represent the current situation as SITUATION,
the two possible outcomes as OUTCOME_1 and OUTCOME_2.
153
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
{conc, incoh}
{pres, incoh}
easier
{pres, coh}
{coh, incoh}
{pres, pres}
harder
{coh, coh}
{incoh, incoh}
Figure 4: Hasse diagram of ⪰h .
The issues to consider are as follows:
(1) Whether the current situation should have OUTCOME_1 or not,
represented as SITUATION ± OUTCOME_1;
(2) Whether the current situation should have OUTCOME_2 or not,
represented as SITUATION ± OUTCOME_2.
The situation is decided based on testimony, where the witness can
have expert knowledge, or not. Therefore, depending on the knowl-
edge of the witness, the value of the witness varies. If the witness is
an expert (represented as EXPERT), the outcome has stronger support
than if the person is not an expert (represented as ¬EXPERT).
Now we focus on issue SITUATION ± OUTCOME_1. Figure 5 shows
examples of case models for each type of the issue listed in Proposi-
tion 3, from which we can see that the cases with an expert witness
(EXPERT) are more preferred than the cases with ¬EXPERT, as sug-
gested by the size of boxes.
In the {conc,incoh} model, there is only one case, namely, an ex-
pert predicts that the situation should have OUTCOME_1. In this model,
argument (SITUATION, OUTCOME_1) in SITUATION ± OUTCOME_1
is conclusive, namely all decisions in this model are for having
OUTCOME_1, which makes OUTCOME_1 seem like a natural conse-
quence. And there is no support for other decisions.
Comparing the {conc,incoh} model with the {pres,incoh} model,
For issue SITUATION ± OUTCOME_1 in the {pres,incoh} model, ar-
gument (SITUATION, OUTCOME_1) has label pres, which makes
the decision of OUTCOME_1 stronger. However, comparing with the
same issue in the {conc,incoh} model, it becomes less strong, since
there is also a case that implies another outcome (OUTCOME_2) in
the {pres,incoh} model, even though the argument for OUTCOME_2
is only with coh. As the cases do not all point to OUTCOME_1, issue
SITUATION ± OUTCOME_1 is harder in the {pres,incoh} model.
In the {pres,coh} model, the issue SITUATION ± OUTCOME_1 is
again harder than in the {pres,incoh} model. Argument (SITUATION,
OUTCOME_1) is still stronger (pres), however, the other argument
about SITUATION ± OUTCOME_1, i.e., (SITUATION, ¬OUTCOME_1), is
not incoherent anymore. There is a case in the model with ¬OUTCOME_1,
which indicates that it is possible that the situation should not have
OUTCOME_1, even though it is from a non-expert source. The coherent
but opposite decisions make the issue become harder.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil H. Zheng, D. Grossi and B. Verheij

{conc, incoh} {incoh, incoh}

EXPERT ∧ SITUATION ∧ OUTCOME_1 EXPERT ∧ SITUATION ∧ OUTCOME_2

{pres, incoh} {coh, incoh}

EXPERT ∧ SITUATION ∧ OUTCOME_1 ¬EXPERT ∧ SITUATION ∧ OUTCOME_2 EXPERT ∧ SITUATION ∧ OUTCOME_2 ¬EXPERT ∧ SITUATION ∧ OUTCOME_1

{pres, pres} {pres, coh}

EXPERT ∧ SITUATION ∧ OUTCOME_1 EXPERT ∧ SITUATION ∧ ¬OUTCOME_1 EXPERT ∧ SITUATION ∧ OUTCOME_1 ¬EXPERT ∧ SITUATION ∧ ¬OUTCOME_1

{coh, coh}
EXPERT ∧ SITUATION ∧ OUTCOME_2 ¬EXPERT ∧ SITUATION ∧ OUTCOME_1 ¬EXPERT ∧ SITUATION ∧ ¬OUTCOME_1

Figure 5: The hardness of SITUATION ± OUTCOME_1 in 7 case models characterized by the type of issues

In the {coh,incoh} model, SITUATION±OUTCOME_1 is harder than Kerfoot v. Kelley 294 N.Y. 288, 62 N.E.2d 74 (1945): The claim
in the {pres,coh} model, since for argument (SITUATION, OUTCOME_1), was in tort law (driver negligence). The territorial rule applies.
it is less preferable as the testimony for OUTCOME_1 is made by a non- Auten v. Auten 308 N.Y. 155, 124 N.E.2d 99 (1954): The claim
expert. For (SITUATION, ¬OUTCOME_1), there is no support. Compar- was in contract law (enforce a child support agreement). The
ing with the {pres,coh} model, there is no expert testimony about center-of-gravity rule applies.
the issue, which makes the consideration of this issue harder. Kaufman v. American Youth Hostels 5 N.Y.2d 1016 (1959):
In the {pres,pres} model, issue SITUATION±OUTCOME_1 is harder The claim was in tort law (travel guide negligence). The territorial
than in the {coh,incoh} model, even though the terstimony for rule applies.
OUTCOME_1 is from an expert, namely a more preferable source. Haag v. Barnes 9 N.Y.2d 554, 175 N.E.2d 441, 216 N.Y.S. 2d
This is because its counterargument (SITUATION, ¬OUTCOME_1) is 65 (1961): The claim was in contract law (reopen a child support
also based on an expert, which makes both having OUTCOME_1 and agreement). The center-of-gravity rule applies.
not having OUTCOME_1 have strong support, hence harder to solve Kilberg v. Northeast Airlines 9 N.Y.2d 34, 172 N.E.2d 526, 211
than in the {coh,incoh} model, where there is no one who testifies N.Y.S.2d 133 (1961): The claim was in tort law (common carrier
that the situation should not have OUTCOME_1. negligence). The territorial rule is partly applied, and there is an
In the {coh,coh} model, issue SITUATION ± OUTCOME_1 is harder exception for the damages part of the case.
than in the {pres,pres} model. Even though in both of the models, the Babcock v. Jackson 12 N.Y.2d 473, 191 N.E.2d 279, 473 N.Y.S.2d
arguments for having OUTCOME_1 and for not having it are as strong 279 (1963): The claim was in tort law (driver negligence).
as each other, type {coh,coh} still indicates that the testimonies are
Considering the case model constructed in [21], also shown in Fig-
from less preferable sources (non-expert), and because of this, the
ure 6, where it consists of 7 cases. They are represented by factors for
consideration of the issue becomes harder.
the plaintiff’s name (SMITH, KERFOOT, etc.), the year of the decision
Type {incoh,incoh} is the hardest one. As shown in Figure 5,
(1938, 1945, etc.), the kind of case (TORT for a tort case, CONTRACT
there is completely no decision about issue SITUATION±OUTCOME_1,
for a contract case), and the jurisdiction choice rule (TERRITORY for
hence the consideration of the issue is the hardest as there is nothing
entirely applying the territorial rule, EXCEPTION for partly applying
that can be referred to.
the territorial rule while making an exception for the damages part
of the case, and GRAVITY for applying the center-of-gravity rule).
The preference relation among these cases is denoted by the size
4 HARDNESS OVER TIME
of the boxes directly, namely, the Babcock case is more preferred
In this section, we apply our approach to model case-based decision-
making in a real legal domain from the United States, and discuss
the development of precedential value in a series of relevant cases
by following the research developed by Berman and Hafner [8, 11] SMITH ∧ 1938 ∧ TORT ∧ TERRITORY
and Verheij [21]. KERFOOT ∧ 1945 ∧ TORT ∧ TERRITORY
The cases we show here are tort cases from New York, which are AUTEN ∧ 1954 ∧ CONTRACT ∧ GRAVITY
about car accidents, and which rule should be applied when different KAUFMAN ∧ 1959 ∧ TORT ∧ TERRITORY
jurisdictions are relevant. For instance, when people drive from New HAAG ∧ 1961 ∧ CONTRACT ∧ GRAVITY
York and have an accident in Ontario, which rule should be followed, KILBERG ∧ 1961 ∧ TORT ∧ EXCEPTION
Ontario’s or New York’s? BABCOCK ∧ 1963 ∧ TORT ∧ GRAVITY

Smith v. Clute 277 N.Y. 407, 14 N.E.2d 455 (1938): The claim
was in tort law (driver negligence). The territorial rule applies. Figure 6: The development of precedential values in cases [21]

154
Hardness of Case-Based Decisions: a Formal Theory
than other cases, which are preferentially equivalent. Since the Bab-
cock case is a landmark case that overriding previous cases, by which
the center-of-gravity approach is established for tort law [11, 21].
We also apply the background theory of all cases in the case model
set in [21], namely, the plaintiff names exclude each other pairwise
(¬(SMITH ∧ KERFOOT), etc.), and similarly for the decision years
(¬(1938 ∧ 1945), etc.), the kinds of cases (¬(TORT ∧ CONTRACT))
and the choice rules (¬(TERRITORY ∧ EXCEPTION), etc.).
We analyze the development of the jurisdiction choice rule by
restricting the case model to the cases up and until a particular year.
For instance, we write C (1954) for the set consisting of the three
cases Smith, Kerfoot and Auten dating from 1954 or before [21].
The issues that we want to analyze in this series of cases are about
the development of the jurisdiction choice rule in tort law cases.
They are shown as follows:
(1) TORT ± TERRITORY, which is associated with arguments:
(TORT, TERRITORY) and (TORT, ¬TERRITORY);
(2) TORT ± GRAVITY, which is associated with arguments:
(TORT, GRAVITY) and (TORT, ¬ GRAVITY);
(3) TORT ± EXCEPTION, which is associated with arguments:
(TORT, EXCEPTION) and (TORT, ¬ EXCEPTION);
(4) ⊤ ± GRAVITY, which is associated with arguments:
(⊤, GRAVITY) and (⊤, ¬ GRAVITY).
Issue TORT ± TERRITORY is about whether a tort law case should
entirely apply the territorial rule or not. TORT ± GRAVITY is about
whether a tort law case should apply the center-of-gravity rule or not.
TORT ± EXCEPTION is about whether a tort law case should partly
follow the territorial rule and make an exception for the damages
part of the case. And ⊤ ± GRAVITY is about the applied status of the
center-of-gravity rule in a general sense.
The validity of the arguments listed above has been discussed in
[21]. As we show in Section 3, the hardness of an issue is determined
by the validity of the arguments that it associates with. For instance,
the hardness of issue TORT ± TERRITORY in 1938 with respect to case
model C (1938) is:
TC (1938) (TORT ± TERRITORY) = {conc, incoh}
which is determined by the validity of the following arguments:
AC (1938) (TORT, TERRITORY) = conc
AC (1938) (TORT, ¬TERRITORY) = incoh
TORT ± TERRITORY becomes harder in 1961, since the hardness of
the issue in C (1961) is {pres, pres}. Notice that according to Def-
inition 4, we consider the labels of arguments (TORT, TERRITORY)
and (TORT, ¬TERRITORY) in C (1961) as pres rather than coh.
Based on the validity of the relevant arguments, we summarize
the hardness of issues with respect to case models by years in Table 1.
The trends of the hardness of issues is shown in Figure 7, from which
we have the following observations about the hardness of issues:
(1) When the center-of-gravity rule is introduced in general (1954)
and into the tort law domain (1963), the issues about the
GRAVITY rule become harder.
(2) The issues related to the territorial rule become harder corre-
spondingly when the court shows doubt on the rule by making
an exception for the damages part of a tort case.
(3) When the center-of-gravity rule is introduced by a landmark
case (with higher preference), which makes the rule more
155
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
preferred, not only the issue about this rule becomes harder,
but also other issues are affected (become easier).
(4) In general, after finally making the GRAVITY rule as the pri-
mary one in 1963, the 4 tort law-relevant issues remain the
same hardness, as in 1945 when the primary one is the terri-
torial rule. However, more options makes the consideration
of applying rules becomes harder.
The first observation can be illustrated by the introduction of the
center-of-gravity rule. In 1954, the center-of-gravity rule starts to
be considered by the New York courts in a general sense, even
it has no effect on the hardness of issues TORT ± TERRITORY and
TORT ± GRAVITY, it does make issue ⊤ ± GRAVITY become harder
than in 1945:
TC (1954) (⊤ ± GRAVITY) ≻h TC (1945) (⊤ ± GRAVITY).
This is because, before 1954, it is clear that the GRAVITY rule is not
considered in the court as argument (⊤, GRAVITY) is with incoh and
(⊤, ¬GRAVITY) is with conc. However, the Auten case introduces this
rule to the series of case models and makes both of the arguments
become presumptive. The introduction not only makes (⊤, GRAVITY)
stronger and (⊤, ¬GRAVITY) weaker but also shortens the validity
distance between the validity labels of the two opposite arguments in
the issue. Because of the shorter distance, considering whether the
center-of-gravity rule should be generally considered or not becomes
harder than before. Similarly, after introducing the center-of-gravity
rule into the tort law domain (by the Babcock case in 1963), we can
see the same trend as in 1954.
The second observation is about the exception in a tort law case
that applied the territorial rule, which is introduced by the Kilberg
case in 1961. After this case is added into the model, it has no effect
on the hardness of issues that related to the GRAVITY rule:
TC (1961) (TORT ± GRAVITY) ∼h TC (1959) (TORT ± GRAVITY);
TC (1961) (⊤ ± GRAVITY) ∼h TC (1959) (⊤ ± GRAVITY).
Both TORT ± TERRITORY and TORT ± EXCEPTION become harder:
TC (1961) (TORT ± TERRITORY) ≻h TC (1959) (TORT ± TERRITORY);
TC (1961) (TORT ± EXCEPTION) ≻h TC (1959) (TORT ± EXCEPTION).
This is because the exception makes the consideration of the territo-
rial rule in the tort law domain become more complex, as now the
courts need to think of whether there will be an exception or not.
The third observation is for the introduction of the landmark
case (Babcock), introducing the center-of-gravity rule in the tort law
domain. This makes issue TORT ± GRAVITY harder:
TC (1963) (TORT ± GRAVITY) ≻h TC (1961) (TORT ± GRAVITY).
and other relevant issues easier:
TC (1963) (TORT ± TERRITORY) ≻h TC (1961) (TORT ± TERRITORY);
TC (1963) (TORT ± EXCEPTION) ≻h TC (1961) (TORT ± EXCEPTION);
TC (1963) (⊤ ± GRAVITY) ≻h TC (1961) (⊤ ± GRAVITY).
These trends can be explained from an intuitive perspective. Since
from 1963, the GRAVITY rule becomes primary, for other options, the
more preferable way is not applying them, hence make the issues that
they associated with easier. But for the TORT ± GRAVITY, it becomes
harder as we explained in the first observation above.
In the last observation, we find that after the GRAVITY becomes
primary in 1963, all the tort law-relevant issues remain the same
hardness as they were before 1954. The only difference is that the
ICAIL’21, June 21–25, 2021, São Paulo, Brazil H. Zheng, D. Grossi and B. Verheij

Table 1: Hardness of issues in different years

TORT ± TERRITORY TORT ± GRAVITY

(TORT, TERRITORY): conc (TORT, GRAVITY): incoh
1938 - 1959 {conc,incoh} 1938 - 1959 {conc,incoh}
(TORT, ¬TERRITORY): incoh (TORT, ¬GRAVITY): conc
(TORT, TERRITORY): pres (TORT, GRAVITY): incoh
1961 {pres,pres} 1961 {conc,incoh}
(TORT, ¬TERRITORY): pres (TORT, ¬GRAVITY): conc
(TORT, TERRITORY): coh (TORT, GRAVITY): pres
1963 {pres,coh} 1963 {pres,coh}
(TORT, ¬TERRITORY): pres (TORT, ¬GRAVITY): coh
TORT ± EXCEPTION ⊤ ± GRAVITY
(TORT, EXCEPTION): incoh (⊤, GRAVITY): incoh
1938 - 1959 {conc,incoh} 1938 - 1945 {conc,incoh}
(TORT, ¬EXCEPTION): conc (⊤, ¬GRAVITY): conc
(TORT, EXCEPTION): pres (⊤, GRAVITY): pres
1961 {pres,pres} 1954 - 1961 {pres,pres}
(TORT, ¬EXCEPTION): pres (⊤, ¬GRAVITY): pres
(TORT, EXCEPTION): coh (⊤, GRAVITY): pres
1963 {pres,coh} 1963 {pres,coh}
(TORT, ¬EXCEPTION): pres (⊤, ¬GRAVITY): coh

{conc, incoh}

{pres, incoh}
easier
{pres, coh}
: TORT±TERRITORY
: TORT±GRAVITY
{coh, incoh} : TORT±EXCEPTION
: ⊤±GRAVITY
{pres, pres}
harder
{coh, coh}

{incoh, incoh}
1938 1945 1954 1959 1961 1963

Figure 7: Development of hardness over time for different issues in the series of tort cases

more preferred rule shifts from the TERRITORY rule to the GRAVITY
rule. Moreover, we can see that making the choice of which rule
should be applied in 1963 is harder than before 1954. This can be
connected to our first observation. Before 1954, there is only one
jurisdiction choice rule to be considered, as the argument for apply-
ing the TERRITORY rule is conclusive (in C (1938) and C (1945)),
and the argument for applying the GRAVITY rule is incoherent (in
C (1938) and C (1945)). Therefore, if a current case is given in the
year before 1954, by following the precedents, applying the territo-
rial rule in the current case will be a more preferable choice. In 1963,
even though the GRAVITY rule has already become more preferred
than other choices, the consideration of which rule should be applied
still becomes more complex. This is because there are new cases that
introduce new choices about the rule application during this period.
More choices make the consideration of the application harder.
5 DISCUSSION
In this section, we position our formal theory of the hardness of
case-based decisions with respect to related research.
The development of hardness over time. The dynamics of case-based
reasoning has for instance been addressed in [11, 13, 14] in terms of
rules, values, and reasons, and there changes in these elements are
associated with the handling of hard cases. As we show in Section 4,
our approach is also relevant for the dynamics in case-based reason-
ing. We extend the analysis of a series of New York tort cases in
[11, 21] with types of issues, from where we find that even though
new cases can strengthen an argument’s validity, the hardness of
issues may also increase.
Our approach gives insight into the five temporal patterns listed
in [11, Section 4.2] (also discussed in [21]), providing a different
angle in terms of the hardness typology.
(1) A general shift in the relative priority of competing purposes.
The Auten and Haag cases introduce the center-of-gravity rule
into the contract cases, and let GRAVITY become a presump-
tive conclusion in general, whereas it was incoherent. How-
ever, argument (TORT, GRAVITY) is not yet coherent. From
the issue perspective, the Auten case makes the general con-
sideration of the GRAVITY rule (⊤ ± GRAVITY) harder (from

156
Hardness of Case-Based Decisions: a Formal Theory
{conc,incoh} to {pres,pres}). After 1954, the consideration
of the GRAVITY rule becomes more complicate. However, the
hardness of handling GRAVITY rule in tort law cases has not
changed yet, as TORT ± GRAVITY is still as hard as before.
(2) A shift in the relative priority of competing purposes by find-
ing exceptions. The Kilberg case makes that TERRITORY, rep-
resenting the entire application of the territorial rule, is no
longer a conclusive consequence of tort cases, but only a
presumptive consequence. From the issue perspective, after
the Kilberg case is added to the model, the type of issue
TORT ± TERRITORY shifts from {conc,incoh} to {pres,pres},
in the sense that EXCEPTION (partly applying the territorial
rule) makes the territorial rule harder to handle.
(3) The ratio decidendi of an older case is overruled, although it
is significantly different. The example of this pattern discussed
in [11, 21] is that the Babcock case overrules the Kaufman
case. The formal case model we use doesn’t distinguish tort
cases from passenger cases, thus the pattern is not visible
here. But as shown in Figure 7, we can still figure out that
the landmark Babcock case makes the consideration of some
issues in a sense harder than right after the Kaufman case.
(4) A case is implicitly overruled. The rule that applied in tort
cases has changed since 1961 because of the Kilberg case and
the Babcock case. The territorial rule is no longer a presump-
tive conclusion, and the center-of-gravity rule becomes more
preferred. If the Kerfoot case is decided after Kilberg or Bab-
cock, it may have come with a different outcome. As shown
in Figure 7, both issues are with type {pres,coh} in 1963,
but notice that the conclusions of presumptive arguments in
the issues are GRAVITY and ¬TERRITORY. Furthermore, both
issues in 1963 are harder than in the period 1938 ∼ 1959, if
the Kerfoot case is decided after 1963, though it will more
likely apply the GRAVITY rule, the decision-making process
still becomes more complicated than before.
(5) A case is explicitly overruled. As discussed by [8, 11], this
pattern occurs rarely, and is not shown in our case model.
Comparing with other research developed recently, in particular
Horty’s reason model for precedential constraint [14], Henderson
and Bench-Capon’s model [13] following Levi’s idea to consider
case-based reasoning as a “moving classification system” [15], the
approach we applied here is based on a different mechanism in terms
of an elementary propositional logical language. Moreover, their
research focuses more on the development of cases and the involved
rules, and the hardness of issues as they occur in the decision-making
process—the main contribution of this paper—have not been for-
mally analyzed. It seems interesting to follow up on Levi’s three
stage life cycle of rules (creation, refinement, replacement) as dis-
cussed in [13] using the hardness typology in this paper.
The focus of other research on case-based reasoning is often
rather different from our paper, and the hardness approach presented
here may supplement that research. For instance, [13, 14] emphasize
the role of reasons and rules in legal cases, and how they favor
the different parties in the court, which may be connected with our
approach, in which issues and their hardness are associated with
arguments with equal premises, but opposite conclusions.
157
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The approach we present here continues the discussion in [25],
where we find that ‘using an incoherent argument can make sense
and break new ground. A decision based on such an argument can
be considered as going beyond the current legal status modeled in
the precedent model.’ We can further interpret this idea with the
results we get from the case study in Section 4. For instance, after
introducing the center-of-gravity rule into the tort law domain in
1963 by the Babcock case, the validity of the argument for applying
the rule in a tort law case, namely (TORT, GRAVITY), shifts from
incoh to pres. Even though the validity of the argument becomes
stronger, the associated issue TORT±GRAVITY becomes harder, and
the validity distance between the validity labels of the two arguments
in the issue is shortened, namely, making the other argument weaker.
It could be interesting to enrich the series of cases discussed in
Section 4 to include what happened after the Babcock case. Other
series of cases that are well-known in AI and Law are also interesting
to look at using the hardness theory we developed, for instance, the
cases about product liability and privity [3, 13, 15]. Also, since the
case model we show in the case study does not have all the possi-
ble types in a real legal domain, it can be interesting to investigate
whether such a complete case study can be made, in order to better
understand the hardness of issues in an actual decision-making pro-
cess. Natural developments are also to connect our hardness typology
in terms of kinds of validity to proof standards [10] and to consider
the development of hardness over time in terms of argumentation
schemes for case-based reasoning [24]. It would also be interesting
to explore the hardness of issues under different preference orderings
than significance, for instance, in terms of court levels.
From hardness of issues to easy and hard cases. As discussed by
Gardner [9], hard cases is a main topic in law. Rissland summarizes
that hard cases in law can arise in three ways [18]:
(1) there exist competing legal rules;
(2) there exist unresolved predicates; and
(3) there exist competing cases.
Our formalism has the potential for modeling the hard cases dis-
cussed by Gardner. Competing cases and legal rules can be associ-
ated with issues with types {pres,coh}, {pres,pres}, and {coh,coh}.
As in these types, the conclusions of arguments involved are op-
posite to each other, hence form a competing relation. Unresolved
predicates can be associated with issues that do not contain con-
clusive arguments, namely, the same premise can lead to different
conclusions. If we treat these predicates as the premises, their in-
dicated meaning as the conclusions, we can analyze the meaning
of unresolved predicates as leaving room for debate, hence leading
to cases that are harder in the typology. Hence it seems interesting
for future research to connect the hardness of issues to insights on
easy and hard cases. For instance, there is a connection to Dworkin’s
famous idea (see e.g. [19, p. 488f.]) that for the perfect, Herculean
judge, there is one right solution for all cases, including the hard
ones. In our hardness typology, there is a variety of options. Some-
times there is exactly one solution, namely in the types {conc,incoh},
{pres,incoh} and {coh,incoh}. In {incoh,incoh} there is no stare
decisis solution. In {pres,pres} and {coh,coh}, there are two equally
preferred solutions, and in {pres,coh}, there are two of which one
is strictly preferred over the other. Also, consider the characteriza-
tion of hard cases in [12] that they require an a-rational decision
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
making process. In our typology, this characterization applies to
{incoh,incoh}, where there is no solution, and also to {pres,pres}
and {coh,coh}, where none of the two choices is preferred. These
situations require the construction of a new, persuasive theory of the
case and its solution, as suggested by [16].
Also, our hardness theory uses a fixed preference ordering, and it
seems relevant to consider a dynamic perspective as a topic of future
research in order to address changes legal and societal changes.
Hardness of issues in case-based reasoning with factors. The discus-
sion about current situations with precedents in case-based reasoning
with factors, such as HYPO/CATO, is also relevant to our approach
about hardness of issues. As discussed in [25], HYPO examples can
be modeled in case models in which all cases are equally preferred.
In Section 3.2, we show that this kind of case model constrains the
possible types of issues. Therefore, issues that occurred in HYPO-
style reasoning will have a special hardness typology, a subset of
the general typology. Further investigation seems to be in place. For
instance, connecting the hardness of issues to argument moves, such
as analogizing and distinguishing a current situation with precedents,
could be an interesting line of further research.
6 CONCLUSION
In this paper, we model the hardness of case-based decisions in terms
of arguments, and their kind of validity. In the approach, we describe
a decision-making problem as an issue in a situation in terms of
an argument and a counterargument. The hardness of an issue is
represented by the validity of the two associated arguments (conclu-
sive, presumptive, coherent, incoherent). We also define an ordering
that shows which issues are harder, and which easier. Building on
work by Berman and Hafner, we apply our approach to discuss the
hardness of the issues that arose in a series of legal cases. It turns
out that we can formally show the varying hardness of issues in the
temporal development of case-based reasoning.
The hardness approach is relevant in the understanding of case-
based decision-making using stare decisis as it formally describes
the complexity of decision making in different circumstances. In the
discussion, we further suggested that it seems interesting to connect
our hardness approach to other research. Although the approach
here has been applied to issues in hard cases in law, it could be
interesting to consider whether our hardness typology, based on
propositional logic, is relevant in other domains where example-
based reasoning is relevant (such as in medical diagnosis). As the
analysis of argument validity we use, is consistent with probabilistic
methods [20], the connection between the hardness of issues and the
validity of arguments may also lead to insights on the hardness of
decision-making in hybrid AI systems involving both knowledge
and data. In this way, the approach can be developed to support the
relevance of AI & Law research for AI generally (cf. [23]).
ACKNOWLEDGMENTS
The authors would like to thank the reviewers and Wijnand van
Woerkom for their valuable feedback on earlier versions of this pa-
per. This research was partially funded by the Hybrid Intelligence
Center, a 10-year programme funded by the Dutch Ministry of Edu-
cation, Culture and Science through the Netherlands Organization
for Scientific Research, https://hybrid-intelligence-centre.nl.
158
H. Zheng, D. Grossi and B. Verheij
REFERENCES
[1] V. Aleven. 1997. Teaching Case-Based Argumentation Through a Model and
Examples. Ph.D. Dissertation. University of Pittsburgh.
[2] K. D. Ashley. 1990. Modeling Legal Arguments: Reasoning with Cases and
Hypotheticals. MIT Press, Cambridge.
[3] K. D. Ashley. 2018. Precedent and Legal Analogy. In Handbook of Legal
Reasoning and Argumentation, G. Bongiovanni, G. Postema, A. Rotolo, G. Sartor,
C. Valentini, and D. Walton (Eds.). Springer Netherlands, Dordrecht, 673–710.
[4] K. Atkinson and T. Bench-Capon. 2019. Reasoning with Legal Cases: Analogy or
Rule Application? In Proceedings of the Seventeenth International Conference on
Artificial Intelligence and Law. ACM Press, New York, 12–21.
[5] T. Bench-Capon and G. Sartor. 2003. A Model of Legal Reasoning with Cases
Incorporating Theories and Values. Artificial Intelligence 150, 1 (2003), 97–143.
[6] D. Berman and C. Hafner. 1991. Incorporating Procedural Context into a Model
of Case-based Legal Reasoning. In Proceedings of the Third International Confer-
ence on Artificial Intelligence and Law. ACM Press, New York, 12–20.
[7] D. Berman and C. Hafner. 1993. Representing Teleological Structure in Case-
Based Legal Reasoning: The Missing Link. In Proceedings of the 4th International
Conference on Artificial Intelligence and Law. ACM Press, New York, 50–59.
[8] D. Berman and C. Hafner. 1995. Understanding Precedents in a Temporal Context
of Evolving Legal Doctrine. In Proceedings of the Fifth International Conference
on Artificial Intelligence and Law. ACM, New York, 42–51.
[9] A. Gardner. 1987. An Artificial Intelligence Approach to Legal Reasoning. MIT
Press, Cambridge.
[10] T. F. Gordon and D. N. Walton. 2009. Proof Burdens and Standards. In Argu-
mentation in Artificial Intelligence, I. Rahwan and G. R. Simari (Eds.). Springer,
Berlin, 239–258.
[11] C. Hafner and D. Berman. 2002. The Role of Context in Case-Based Legal
Reasoning: Teleological, Temporal, and Procedural. Artificial Intelligence and
Law 10, 1 (2002), 19–64.
[12] J. Hage, R. Leenes, and A. R. Lodder. 1993. Hard Cases: A Procedural Approach.
Artificial Intelligence and Law 2, 2 (1993), 113–167.
[13] J. Henderson and T. Bench-Capon. 2019. Describing the Development of Case
Law. In Proceedings of the Seventeenth International Conference on Artificial
Intelligence and Law. ACM Press, New York, 32–41.
[14] J. Horty and T. Bench-Capon. 2012. A Factor-Based Definition of Precedential
Constraint. Artificial Intelligence and Law 20, 2 (2012), 181–214.
[15] E. Levi. 1948. An Introduction to Legal Reasoning. University of Chicago Law
Review 15, 3 (1948), 501–574.
[16] L. T. McCarty. 1995. An Implementation of Eisner v. Macomber. In Proceedings
of the Fifth International Conference on Artificial Intelligence and Law. ACM
Press, New York, 276–286.
[17] H. Prakken. 2019. Comparing Alternative Factor- and Precedent-based Accounts
of Precedential Constraint. In Legal Knowledge and Information Systems. JURIX
2019: The Thirty-Second Annual Conference, M. Araszkiewicz and V. Rodriguez-
Doncel (Eds.). IOS Press, Amsterdam, 73–82.
[18] E. Rissland. 1988. Artificial Intelligence and Legal Reasoning: A Discussion of
the Field and Gardner’s Book. AI Magazine 9, 3 (Sep. 1988), 45–55.
[19] G. Sartor. 2005. Legal reasoning: a cognitive approach to the law. Vol 5 of Treatise
on legal philosophy and general jurisprudence. Springer, Berlin.
[20] B. Verheij. 2016. Correct Grounded Reasoning with Presumptive Arguments.
In 15th European Conference on Logics in Artificial Intelligence, JELIA 2016.
Larnaca, Cyprus, November 9–11, 2016. Proceedings (LNAI 10021), L. Michael
and A. Kakas (Eds.). Springer, Berlin, 481–496.
[21] B. Verheij. 2016. Formalizing Value-Guided Argumentation for Ethical Systems
Design. Artificial Intelligence and Law 24, 4 (2016), 387–407.
[22] B. Verheij. 2017. Formalizing Arguments, Rules and Cases. In Proceedings of the
Sixteenth International Conference on Artificial Intelligence and Law. ACM, New
York, 199–208.
[23] B. Verheij. 2020. Artificial Intelligence as Law. Presidential Address to the
Seventeenth International Conference on Artificial Intelligence and Law. Artificial
Intelligence and Law 28, 2 (2020), 181–206.
[24] A. Wyner and T. J. M. Bench-Capon. 2019. Argument Schemes for Legal Case-
based Reasoning. In Legal Knowledge and Information Systems: JURIX 2007:
The Twentieth Annual Conference, A. R. Lodder and L. Mommers (Eds.). IOS
Press, Amsterdam, 139–149.
[25] H. Zheng, D. Grossi, and B. Verheij. 2020. Case-Based Reasoning with Precedent
Models: Preliminary Report. In Computational Models of Argument. Proceedings
of COMMA 2020, H. Prakken, S. Bistarelli, F. Santini, and C. Taticchi (Eds.).
Vol. 326. IOS Press, Amsterdam, 443–450.
[26] H. Zheng, D. Grossi, and B. Verheij. 2020. Precedent Comparison in the Precedent
Model Formalism: A Technical Note. In Legal Knowledge and Information
Systems. JURIX 2020: The Thirty-third Annual Conference, S. Villata, J. Harašta,
and P. Kšemen (Eds.). Vol. 334. IOS Press, Amsterdam, 259–262.
 When Does Pretraining Help? Assessing Self-Supervised
Learning for Law and the CaseHOLD Dataset of 53,000+ Legal
Holdings
Lucia Zheng∗ Neel Guha∗
zlucia@stanford.edu nguha@stanford.edu
Stanford University Stanford University
Stanford, California, USA Stanford, California, USA
Peter Henderson
phend@stanford.edu
Stanford University
Stanford, California, USA
ABSTRACT
While self-supervised learning has made rapid advances in natural
language processing, it remains unclear when researchers should
engage in resource-intensive domain-specific pretraining (domain
pretraining). The law, puzzlingly, has yielded few documented in-
stances of substantial gains to domain pretraining in spite of the
fact that legal language is widely seen to be unique. We hypothesize
that these existing results stem from the fact that existing legal NLP
tasks are too easy and fail to meet conditions for when domain
pretraining can help. To address this, we first present CaseHOLD
(Case Holdings On Legal Decisions), a new dataset comprised of
over 53,000+ multiple choice questions to identify the relevant
holding of a cited case. This dataset presents a fundamental task
to lawyers and is both legally meaningful and difficult from an
NLP perspective (F1 of 0.4 with a BiLSTM baseline). Second, we
assess performance gains on CaseHOLD and existing legal NLP
datasets. While a Transformer architecture (BERT) pretrained on
a general corpus (Google Books and Wikipedia) improves perfor-
mance, domain pretraining (on a corpus of ≈3.5M decisions across
all courts in the U.S. that is larger than BERT’s) with a custom legal
vocabulary exhibits the most substantial performance gains with
CaseHOLD (gain of 7.2% on F1, representing a 12% improvement
on BERT) and consistent performance gains across two other legal
tasks. Third, we show that domain pretraining may be warranted
when the task exhibits sufficient similarity to the pretraining corpus:
the level of performance increase in three legal tasks was directly
tied to the domain specificity of the task. Our findings inform when
researchers should engage in resource-intensive pretraining and
show that Transformer-based architectures, too, learn embeddings
suggestive of distinct legal language.
∗ These authors contributed equally to this work.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466088
159
Brandon R. Anderson
banderson@law.stanford.edu
Stanford University
Stanford, California, USA
Daniel E. Ho
dho@law.stanford.edu
Stanford University
Stanford, California, USA
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Natural language processing; Neural networks.
KEYWORDS
law, natural language processing, pretraining, benchmark dataset
ACM Reference Format:
Lucia Zheng, Neel Guha, Brandon R. Anderson, Peter Henderson, and Daniel
E. Ho. 2021. When Does Pretraining Help? Assessing Self-Supervised Learn-
ing for Law and the CaseHOLD Dataset of 53,000+ Legal Holdings. In Eigh-
teenth International Conference for Artificial Intelligence and Law (ICAIL’21),
June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 10 pages.
https://doi.org/10.1145/3462757.3466088
1 INTRODUCTION
How can rapid advances in Transformer-based architectures be
leveraged to address problems in law? One of the most significant
advances in natural language processing (NLP) has been the advent
of “pretrained” (or self-supervised) language models, starting with
Google’s BERT model [12]. Such models are pretrained on a large
corpus of general texts — Google Books and Wikipedia articles —
resulting in significant gains on a wide range of fine-tuning tasks
with much smaller datasets and have inspired a wide range of
applications and extensions [27, 38].
One of the emerging puzzles for law has been that while general
pretraining (on the Google Books and Wikipedia corpus) boosts
performance on a range of legal tasks, there do not appear to be any
meaningful gains from domain-specific pretraining (domain pre-
training) using a corpus of law. Numerous studies have attempted to
apply comparable Transformer architectures to pretrain language
models on law, but have found marginal or insignificant gains on
a range of legal tasks [7, 14, 49, 50]. These results would seem to
challenge a fundamental tenet of the legal profession: that legal lan-
guage is distinct in vocabulary, semantics, and reasoning [28, 29, 44].
Indeed, a common refrain for the first year of U.S. legal education
is that students should learn the “language of law”: “Thinking like
a lawyer turns out to depend in important ways on speaking (and
reading, and writing) like a lawyer.” [29].
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
We hypothesize that the puzzling failure to find substantial gains
from domain pretraining in law stem from the fact that existing fine-
tuning tasks may be too easy and/or fail to correspond to the domain
of the pretraining corpus task. We show that existing legal NLP
tasks, Overruling (whether a sentence overrules a prior case, see
Section 4.1) and Terms of Service (classification of contractual terms
of service, see Section 4.2), are simple enough for naive baselines
(BiLSTM) or BERT (without domain-specific pretraining) to achieve
high performance. Observed gains from domain pretraining are
hence relatively small. Because U.S. law lacks any benchmark task
that is comparable to the large, rich, and challenging datasets that
have fueled the general field of NLP (e.g., SQuAD [36], GLUE [46],
CoQA [37]), we present a new dataset that simulates a fundamental
task for lawyers: identifying the legal holding of a case. Holdings are
central to the common law system. They represent the governing
legal rule when the law is applied to a particular set of facts. The
holding is precedential and what litigants can rely on in subsequent
cases. So central is the identification of holdings that it forms a
canonical task for first-year law students to identify, state, and
reformulate the holding.
This CaseHOLD dataset (Case Holdings on Legal Decisions)
provides 53,000+ multiple choice questions with prompts from
a judicial decision and multiple potential holdings, one of which
is correct, that could be cited. We construct this dataset using the
rules of case citation [9], which allow us to match a proposition to a
source through a comprehensive corpus of U.S. case law from 1965
to the present. Intuitively, we extract all legal citations and use the
“holding statement,” often provided in parenthetical propositions
accompanying U.S. legal citations, to match context to holding [2].
CaseHOLD extracts the context, legal citation, and holding state-
ment and matches semantically similar, but inappropriate, holding
propositions. This turns the identification of holding statements
into a multiple choice task.
In Table 1, we show a citation example from the CaseHOLD
dataset. The Citing Text (prompt) consists of the context and legal
citation text, Holding Statement 0 is the correct corresponding
holding statement, Holding Statements 1-4 are the four similar,
but incorrect holding statements matched with the given prompt,
and the Label is the 0-index label of the correct holding statement
answer. For simplicity, we use a fixed context window that may
start mid-sentence.
We show that this task is difficult for conventional NLP ap-
proaches (BiLSTM F1 = 0.4 and BERT F1 = 0.6), even though law
students and lawyers are able to solve the task at high accuracy.
We then show that there are substantial and statistically significant
performance gains from domain pretraining with a custom vocabu-
lary (which we call Legal-BERT), using all available case law from
1965 to the present (a 7.2% gain in F1, representing a 12% relative
boost from BERT). We then experimentally assess conditions for
gains from domain pretraining with CaseHOLD and find that the
size of the fine-tuning task is the principal other determinant of
gains to domain-specific pretraining.
The code, the legal benchmark task datasets, and the Legal-BERT
models presented here can be found at: https://github.com/reglab/
casehold.
Our paper informs how researchers should decide when to en-
gage in data and resource-intensive pretraining. Such decisions
160
Zheng and Guha, et al.
Table 1: CaseHOLD example
Citing Text (prompt)
They also rely on Oswego Laborers’ Local 214 Pension Fund v. Marine
Midland Bank, 85 N.Y.2d 20, 623 N.Y.S.2d 529, 647 N.E.2d 741 (1996), which
held that a plaintiff “must demonstrate that the acts or practices have a
broader impact on consumers at large." Defs.’ Mem. at 14 (quoting Oswego
Laborers’, 623 N.Y.S.2d 529, 647 N.E.2d at 744). As explained above, how-
ever, Plaintiffs have adequately alleged that Defendants’ unauthorized
use of the DEL MONICO’s name in connection with non-Ocinomled
restaurants and products caused consumer harm or injury to the public,
and that they had a broad impact on consumers at large inasmuch as
such use was likely to cause consumer confusion. See, e.g., CommScope,
Inc. of N.C. v. CommScope (U.S.A) Int’l Grp. Co., 809 F. Supp.2d 33, 38
(N.D.N.Y 2011) (<HOLDING>); New York City Triathlon, LLC v. NYC
Triathlon
Holding Statement 0 (correct answer)
holding that plaintiff stated a 349 claim where plaintiff alleged facts
plausibly suggesting that defendant intentionally registered its corporate
name to be confusingly similar to plaintiffs CommScope trademark
Holding Statement 1 (incorrect answer)
holding that plaintiff stated a claim for breach of contract when it alleged
the government failed to purchase insurance for plaintiff as agreed by
contract
Holding Statement 2 (incorrect answer)
holding that the plaintiff stated a claim for tortious interference
Holding Statement 3 (incorrect answer)
holding that the plaintiff had not stated a claim for inducement to breach
a contract where she had not alleged facts sufficient to show the existence
of an enforceable underlying contract
Holding Statement 4 (incorrect answer)
holding plaintiff stated claim in his individual capacity
pose an important tradeoff, as cost estimates for fully pretraining
BERT can be upward of $1M [41], with potential for social harm
[4], but advances in legal NLP may also alleviate huge disparities in
access to justice in the U.S. legal system [16, 34, 47]. Our findings
suggest that there is indeed something unique to legal language
when faced with sufficiently challenging forms of legal reasoning.
2 RELATED WORK
The Transformer-based language model, BERT [12], which lever-
ages a two step pretraining and fine-tuning framework, has achieved
state-of-the-art performance on a diverse array of downstream NLP
tasks. BERT, however, was trained on a general corpus of Google
Books and Wikipedia, and much of the scientific literature has
since focused on the question of whether the Transformer-based
approach could be improved by domain-specific pretraining.
Outside of the law, for instance, Lee et al. [25] show that BioBERT,
a BERT model pretrained on biomedicine domain-specific corpora
(PubMed abstracts and full text articles), can significantly outper-
form BERT on domain-specific biomedical NLP tasks. For instance,
it achieves gains of 6-9% in strict accuracy compared to BERT [25]
for biomedical question answering tasks (BioASQ Task 5b and Task
5c) [45]. Similarly, Beltagy et al. show improvements from domain
When Does Pretraining Help?
pretraining with SciBERT, using a multi-domain corpus of scientific
publications [3]. On the ACL-ARC multiclass classification task [22],
which contains example citations labeled with one of six classes,
where each class is a citation function (e.g., background), SciBERT
achieves gains of 7.07% in macro F1 [3]. It is worth noting that this
task is constructed from citation text, making it comparable to the
CaseHOLD task we introduce in Section 3.
Yet work adapting this framework for the legal domain has not
yielded comparable returns. Elwany at el. [14] use a proprietary
corpus of legal agreements to pretrain BERT and report “marginal”
gains of 0.4 - 0.7% on F1. They note that in some settings, such gains
could still be practically important. Zhong et al. [49] uses BERT
pretrained on Chinese legal documents and finds no gains relative
to non-pretrained NLP baseline models (e.g., LSTM). Similarly, [50]
finds that the same pretrained model performs poorly on a legal
question and answer dataset.
Hendrycks et al. [19] found that in zero-shot and few-shot set-
tings, state-of-the-art models for question answering, GPT-3 and
UnifiedQA, have lopsided performance across subjects, performing
with near-random accuracy on subjects related to human values,
such as law and morality, while performing up to 70% accuracy
on other subjects. This result motivated their attempt to create a
better model for the multistate bar exam by further pretraining
RoBERTa [27], a variant of BERT, on 1.6M cases from the Harvard
Law Library case law corpus. They found that RoBERTa fine-tuned
on the bar exam task achieved 32.8% test accuracy without domain
pretraining and 36.1% test accuracy with further domain pretrain-
ing. They conclude that while “additional pretraining on relevant
high quality text can help, it may not be enough to substantially
increase . . . performance.” Hendrycks et al. [18] highlight that future
research should especially aim to increase language model perfor-
mance on tasks in subject areas such as law and moral reasoning
since aligning future systems with human values and understand-
ing of human approval/disapproval necessitates high performance
on such subject specific tasks.
Chalkidis et al. [7] explored the effects of law pretraining using
various strategies and evaluate on a broader range of legal NLP tasks.
These strategies include (a) using BERT out of the box, which is
trained on general domain corpora, (b) further pretraining BERT on
legal corpora (referred to as LEGAL-BERT-FP), which is the method
also used by Hendrycks et al. [19], and (c) pretraining BERT from
scratch on legal corpora (referred to as LEGAL-BERT-SC). Each
of these models is then fine-tuned on the downstream task. They
report that a LEGAL-BERT variant, in comparison to tuned BERT,
achieves a 0.8% improvement in F1 on a binary classification task
derived from the ECHR-CASES dataset [5], a 2.5% improvement in
F1 on the multi-label classification task derived from ECHR-CASES,
and between a 1.1-1.8% improvement in F1 on multi-label classifi-
cation tasks derived from subsets of the CONTRACTS-NER dataset
[6, 8]. These gains are small when considering the substantial data
and computational requirements of domain pretraining. Indeed,
Hendrycks et al. [19] concluded that the documented marginal
difference does not warrant domain pretraining.
This existing work raises important questions for law and arti-
ficial intelligence. First, these results might be seen to challenge
the widespread belief in the legal profession that legal language is
distinct [28, 29, 44]. Second, one of the core challenges in the field
161
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
is that unlike general NLP, which has thrived on large benchmark
datasets (e.g., SQuAD [36], GLUE [46], CoQA [37]), there are few
large and publicly available legal benchmark tasks for U.S. law.
This is explained in part due to the expense of labeling decisions
and challenges around compiling large sets of legal documents
[32], leading approaches above to rely on non-English datasets
[49, 50] or proprietary datasets [14]. Indeed, there may be a kind
of selection bias in available legal NLP datasets, as they tend to
reflect tasks that have been solved by methods often pre-dating the
rise of self-supervised learning. Third, assessment standards vary
substantially, providing little guidance to researchers on whether
domain pretraining is worth the cost. Studies vary, for instance, in
whether BERT is retrained with custom vocabulary, which is partic-
ularly important in fields where terms of art can defy embeddings
of general language models. Moreover, some comparisons are be-
tween (a) BERT pretrained at 1M iterations and (b) domain-specific
pretraining on top of BERT (e.g., 2M iterations) [25]. Impressive
gains might hence be confounded because the domain pretrained
model simply has had more time to train. Fourth, legal language
presents unique challenges in substantial part because of extensive
and complicated system of legal citation. Work has shown that con-
ventional tokenization that fails to account for the structure of legal
citations can improperly present the legal text [20]. For instance,
sentence boundary detection (critical for BERT’s next sentence pre-
diction pretraining task) may fail with legal citations containing
complicated punctuation [40]. Just as using an in-domain tokenizer
helps in multilingual settings [39], using a custom tokenizer should
improve performance consistently for the “language of law.” Last,
few have examined differences across the kinds of tasks where
pretraining may be helpful.
We address these gaps for legal NLP by (a) contributing a new,
large dataset with the task of identification of holding statements
that comes directly from U.S. legal decisions, (b) assessing the con-
ditions under which domain pretraining can help.
3 THE CASEHOLD DATASET
We present the CaseHOLD dataset as a new benchmark dataset for
U.S. law. Holdings are, of course, central to the common law system.
They represent the governing legal rule when the law is applied
to a particular set of facts. The holding is what is precedential and
what litigants can rely on in subsequent cases. So central is the
identification of holdings that it forms a canonical task for first-year
law students to identify, state, and reformulate the holding. Thus,
as for a law student, the goal of this task is two-fold: (1) understand
case names and their holdings; (2) understand how to re-frame the
relevant holding of a case to back up the proceeding argument.
CaseHOLD is a multiple choice question answering task derived
from legal citations in judicial rulings. The citing context from the
judicial decision serves as the prompt for the question. The answer
choices are holding statements derived from citations following
text in a legal decision. There are five answer choices for each citing
text. The correct answer is the holding statement that corresponds
to the citing text. The four incorrect answers are other holding
statements.
We construct this dataset from the Harvard Law Library case
law corpus (In our analyses below, the dataset is constructed from
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 2: Dataset overview
Dataset Source Task Type
Overruling Casetext Binary classification
Terms of Service Lippi et al. [26] Binary classification
CaseHOLD Authors Multiple choice QA
the holdout dataset, so that no decision was used for pretraining
Legal-BERT.). We extract the holding statement from citations (par-
enthetical text that begins with “holding”) as the correct answer and
take the text before it as the citing text prompt. We insert a <HOLD-
ING> token in the position of the citing text prompt where the
holding statement was extracted. To select four incorrect answers
for a citing text, we compute the TD-IDF similarity between the
correct answer and the pool of other holding statements extracted
from the corpus and select the most similar holding statements, to
make the task more difficult. We set an upper threshold for simi-
larity to rule out indistinguishable holding statements (here 0.75),
which would make the task impossible. One of the virtues of this
task setup is that we can easily tune the difficulty of the task by
varying the context window, the number of potential answers, and
the similarity thresholds. In future work, we aim to explore how
modifying the thresholds and task difficulty affects results. In a hu-
man evaluation, the benchmark by a law student was an accuracy
of 0.94.1
A full example of CaseHOLD consists of a citing text prompt, the
correct holding statement answer, four incorrect holding statement
answers, and a label 0-4 for the index of the correct answer. The
ordering of indices of the correct and incorrect answers are random
for each example and unlike a multi-class classification task, the
answer indices can be thought of as multiple choice letters (A, B,
C, D, E), which do not represent classes with underlying meaning,
but instead just enumerate the answer choices. We provide a full
example from the CaseHOLD dataset in Table 1.
4 OTHER DATASETS
To provide a comparison on difficulty and domain specificity, we
also rely on two other legal benchmark tasks. The three datasets
are summarized in Table 2.
In terms of size, publicly available legal tasks are small compared
to mainstream NLP datasets (e.g., SQuAD has 100,000+ questions).
The cost of obtaining high-fidelity labeled legal datasets is precisely
why pretraining is appealing for law [15]. The Overruling dataset,
for instance, required paying attorneys to label each individual
sentence. Once a company has collected that information, it may
not want to distribute it freely for the research community. In
the U.S. system, much of this meta-data is hence retained behind
proprietary walls (e.g., Lexis and Westlaw), and the lack of large-
scale U.S. legal NLP datasets has likely impeded scientific progress.
We now provide more detail of the two other benchmark datasets.
4.1 Overruling
The Overruling task is a binary classification task, where positive
examples are overruling sentences and negative examples are non-
overruling sentences from the law. An overruling sentence is a
1 This
human benchmark was done on a pilot iteration of the benchmark dataset and
may not correspond to the exact TF-IDF threshold presented here.
Zheng and Guha, et al.
statement that nullifies a previous case decision as a precedent, by
Size a constitutionally valid statute or a decision by the same or higher
ranking court which establishes a different rule on the point of law
2,400
9,414 involved.
53,137 The Overruling task dataset was provided by Casetext, a com-
pany focused on legal research software. Casetext selected positive
overruling samples through manual annotation by attorneys and
negative samples through random sampling sentences from the
Casetext law corpus. This procedure has a low false positive rate for
negative samples because the prevalence of overruling sentences
in the whole law is low. Less than 1% of cases overrule another
case and within those cases, usually only a single sentence contains
overruling language. Casetext validates this procedure by estimat-
ing the rate of false positives on a subset of sentences randomly
sampled from the corpus and extrapolating this rate for the whole
set of randomly sampled sentences to determine the proportion of
sampled sentences to be reviewed by human reviewers for quality
assurance.
Overruling has moderate to high domain specificity because
the positive and negative overruling examples are sampled from
the Casetext law corpus, so the language in the examples is quite
specific to the law. However, it is the easiest of the three legal
benchmark tasks, since many overruling sentences are distinguish-
able from non-overruling sentences due to the specific and explicit
language judges typically use when overruling. In his work on
overruling language and speech act theory, Dunn cites several ex-
amples of judges employing an explicit performative form when
overruling, using keywords such as “overrule”, “disapprove”, and
“explicitly reject” in many cases [13]. Language models, non-neural
machine models, and even heuristics generally detect such key-
word patterns effectively, so the structure of this task makes it less
difficult compared to other tasks. Previous work has shown that
SVM classifiers achieve high performance on similar tasks; Sulea et
al. [31] achieves a 96% F1 on predicting case rulings of cases judged
by the French Supreme Court and Aletras et al. [1] achieves 79%
accuracy on predicting judicial decisions of the European Court of
Human Rights.
The Overruling task is important for lawyers because the process
of verifying whether cases remain valid and have not been overruled
is critical to ensuring the validity of legal arguments. This need has
led to the broad adoption of proprietary systems, such as Shepard’s
(on Lexis Advance) and KeyCite (on Westlaw), which have become
important legal research tools for most lawyers [11]. High language
model performance on the Overruling tasks could enable further
automation of the shepardizing process.
In Table 3, we show a positive example of an overruling sentence
and a negative example of a non-overruling sentence from the
Overruling task dataset. Positive examples have label 1 and negative
examples have label 0.
Table 3: Overruling examples
Passage Label
for the reasons that follow, we approve the first district in the 1
instant case and disapprove the decisions of the fourth district.
a subsequent search of the vehicle revealed the presence of an 0
additional syringe that had been hidden inside a purse located
on the passenger side of the vehicle.
162
When Does Pretraining Help?
4.2 Terms of Service
The Terms of Service task is a binary classification task, where
positive examples are potentially unfair contractual terms (clauses)
from the terms of service in contract documents. The Unfair Terms
in Consumer Contracts Directive 93/13/EEC [17] defines an unfair
contractual term as follows. A contractual term is unfair if: (1)
it has not been individually negotiated; and (2) contrary to the
requirement of good faith, it causes a significant imbalance in the
parties rights and obligations, to the detriment of the consumer.
The Terms of Service dataset comes from Lippi et al. [26], which
studies machine learning and natural language approaches for au-
tomating the detection of potentially unfair clauses in online terms
of service and implements a system called CLAUDETTE based on
the results of the study. The dataset was constructed from a corpus
of 50 online consumer contracts. Clauses were manually annotated
as clearly fair, potentially unfair, and clearly unfair. Positive exam-
ples were taken to be potentially unfair or clearly unfair clauses
and negative examples were taken to be clearly fair clauses to di-
chotomize the task. Lippi et al. [26] also studies a multi-class setting
in which each clause is additionally labeled according to one of
eight categories of clause unfairness (e.g. limitation of liability). We
focus on the more general setting where clauses are only labeled
according to whether they encompass any type of unfairness.
Terms of Service has low domain specificity relative to the Over-
ruling and CaseHOLD tasks because examples are drawn from the
terms of service text in consumer contracts. Extensive contracting
language may be less prevalent in the Casetext and Harvard case
law corpora, although contracts cases of course are. The Terms of
Service task is moderately difficult. Excluding ensemble methods,
the classifier that achieves highest F1 performance in the general
setting of Lippi et al. [26] is a single SVM exploiting bag-of-words
features, which achieves a 76.9% F1.
The Terms of Service task is useful for consumers, since automa-
tion of the detection of potentially unfair contractual terms could
help consumers better understand the terms they agree to when
signing a contract and make legal advice about unfair contracts
more accessible and widely available for consumers seeking it. It
could also help consumer protection organizations and agencies
work more efficiently [26].
In Table 4, we show a positive example of a potentially unfair
clause and a negative example of a fair clause from the Terms
of Service dataset. Positive examples have label 1 and negative
examples have label 0.
Table 4: Terms of Service examples
Passage
occasionally we may, in our discretion, make changes to the
agreements.
this section contains service-specific terms that are in addition
to the general terms.
5 METHODS
Our basic approach to understand the conditions for when domain
pretraining may help is to use a series of pretrained BERT models,
but to carefully vary one key modeling decision at a time. This is
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
computationally expensive requiring approximately 16 TPU (64
GPU) core-days per 1M steps. First, we assess performance with
base BERT. Second, we train BERT with twice the number of itera-
tions to be able to compare the value of additional training. Third,
we ingest the entire Harvard Law case corpus from 1965 to the
present and pretrain Legal-BERT on the corpus. The size of this
dataset (37GB) is substantial, representing 3,446,187 legal decisions
across all federal and state courts, and is larger than the size of
the BookCorpus/Wikipedia corpus originally used to train BERT
(15GB). Fourth, we train a custom vocabulary variant of Legal-BERT.
We provide a comparison to a BiLSTM baseline. We now provide
details of these methods.
5.1 Baseline
Our baseline architecture is a one-layer BiLSTM, with 300D word2vec
vectors [30]. For single-sentence tasks, Overruling and Terms of
Service, we encode the sentence and pass the resulting vector to
a softmax classifier. For CaseHOLD, each citation prompt has five
answer choices associated with it. We concatenate the prompt with
each one of the five answers, separated by the <SEP> token, to get
five prompt-answer pairs. We independently encode each prompt-
answer pair and pass the resulting vector through a linear layer,
then apply softmax over the concatenated outputs for the five pairs.
We choose this architecture because it is comparable to the design
suggested for fine-tuning BERT on multiple choice tasks in Rad-
ford et al. [35], where prompt-answer pairs are fed independently
through BERT and a linear layer. In this architecture, we replace
BERT with the BiLSTM.
5.2 BERT
We use the base BERT model (uncased, 110M parameters) [12]
as our baseline BERT model. Because researchers in other disci-
plines have commonly performed domain pretraining starting with
BERT’s parameter values, we also train a model initialized with
base BERT and pretrained for an additional 1M steps, using the
same English Wikipedia corpus that BERT base was pretrained on.
This facilitates a direct comparison to rule out gains solely from
increased pretraining. We refer to this model, trained for 2M total
steps, as BERT (double), and compare it to our two Legal-BERT
variants, each pretrained for 2M total steps. Using 2M steps as our
comparison point for pretraining also allows us to address findings
from Liu et al. [27] that BERT was significantly undertrained and
exhibited improved performance with RoBERTa, a set of modifica-
tions to BERT training procedure which includes pretraining the
model longer.
Label
5.3 Legal-BERT
1
We pretrain two variants of BERT with the Harvard Law case corpus
(https://case.law/) from 1965 to the present.2 We randomly sample
0
10% of decisions from this corpus as a holdout set, which we use
to create the CaseHOLD dataset. The remaining 90% is used for
pretraining.
We preprocess the case law corpus with the sentence segmen-
tation procedure and use the pretraining procedure described in
2 Weuse this period because there is a significant change in the number of reporters
around this period and it corresponds to the modern post-Civil Rights Act era.
163
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Devlin et al. [12]. One variant is initialized with the BERT base
model and pretrained for an additional 1M steps using the case law
corpus and the same vocabulary as BERT (uncased). The other vari-
ant, which we refer to as Custom Legal-BERT, is pretrained from
scratch for 2M steps using the case law corpus and has a custom
legal domain-specific vocabulary. The vocabulary set is constructed
using SentencePiece [24] on a subsample (appx. 13M) of sentences
from our pretraining corpus, with the number of tokens fixed to
32,000. We pretrain both variants with sequence length 128 for 90%
and sequence length 512 for 10% over the 2M steps total.
Both Legal-BERT and Custom Legal-BERT are pretrained us-
ing the masked language model (MLM) pretraining objective, with
whole word masking. Whole word masking and other knowledge
masking strategies, like phrase-level and entity-level masking, have
been shown to yield substantial improvements on various down-
stream NLP tasks for English and Chinese text, by making the MLM
objective more challenging and enabling the model to learn more
about prior knowledge through syntactic and semantic informa-
tion extracted from these linguistically-informed language units
[10, 21, 43]. More recently, Kang et al. [23] posit that whole-word
masking may be most suitable for domain adaptation on emrQA
[33], a corpus for question answering on electronic medical records,
because most words in emrQA are tokenized to sub-word Word-
Piece tokens [48] in base BERT due to the high frequency of unique,
domain-specific medical terminologies that appear in emrQA, but
are not in the base BERT vocabulary. Because the case law corpus
shares this property of containing many domain-specific terms rel-
evant to the law, which are likely tokenized into sub-words in base
BERT, we chose to use whole word masking for pretraining the
Legal-BERT variants on the legal domain-specific case law corpus.
The second pretraining task is next sentence prediction. Here, we
use regular expressions to ensure that legal citations are included
as part of a segmented sentence according to the Bluebook system
of legal citation [9]. Otherwise, the model could be poorly trained
on improper sentence segmentation [40].3
6 RESULTS
6.1 Base Setup
After pretraining the models as described above in Section 5, we
fine-tune on the legal benchmark target tasks and evaluate the
performance of each model.
6.1.1 Hyperparameter Tuning. We provide details on our hyperpa-
rameter tuning process at https://github.com/reglab/casehold.
6.1.2 Fine-tuning and Evaluation. For the BERT-based models, we
use the input transformations described in Radford et al. [35] for
fine-tuning BERT on classification and multiple choice tasks, which
convert the inputs for the legal benchmark tasks into token se-
quences that can be processed by the pretrained model, followed
by a linear layer and a softmax. For the CaseHOLD task, we avoid
making extensive changes to the architecture used for the two clas-
sification tasks by converting inputs consisting of a prompt and
five answers into five prompt-answer pairs (where the prompt and
answer are separated by a delimiter token) that are each passed
3 Where the vagaries of legal citations create detectable errors in sentence segmentation
(e.g., sentences with fewer than 3 words), we omit the sentence from the corpus.
164
Zheng and Guha, et al.
independently through our pretrained models followed by a linear
layer, then take a softmax over the five concatenated outputs. For
Overruling and Terms of Service, we use a single NVIDIA V100
(16GB) GPU to fine-tune on each task. For CaseHOLD, we used
eight NVIDIA V100 (32GB) GPUs to fine-tune on the task.
We use 10-fold cross-validation to evaluate our models on each
task. We use F1 score as our performance metric for the Overruling
and Terms of Service tasks and macro F1 score as our performance
metric for CaseHOLD, reporting mean F1 scores over 10 folds. We
report our model performance results in Table 5 and report statisti-
cal significance from (paired) 𝑡-tests with 10 folds of the test data
to account for uncertainty.
From the results of the base setup, for the easiest Overruling
task, the difference in F1 between BERT (double) and Legal-BERT
is 0.5% and BERT (double) and Custom Legal-BERT is 1.6%. Both
of these differences are marginal. For the task with intermediate
difficulty, Terms of Service, we find that BERT (double) with fur-
ther pretraining BERT on the general domain corpus increases
performance over base BERT by 5.1%, but the Legal-BERT vari-
ants with domain-specific pretraining do not outperform BERT
(double) substantially. This is likely because Terms of Service has
low domain-specificity, so pretraining on legal domain-specific text
does not help the model learn information that is highly relevant to
the task. We note that BERT (double), with 77.3% F1, and Custom
Legal-BERT, with 78.7% F1, outperform the highest performing
model from Lippi et al. [26] for the general setting of Terms of
Service, by 0.4% and 1.8% respectively. For the most difficult and
domain-specific task, CaseHOLD, we find that Legal-BERT and Cus-
tom Legal-BERT both substantially outperform BERT (double) with
gains of 5.7% and 7.2% respectively. Custom Legal-BERT achieves
the highest F1 performance for CaseHOLD, with a macro F1 of
69.5%.
We run paired 𝑡-tests to validate the statistical significance of
model performance differences for a 95% confidence interval. The
mean differences between F1 for paired folds of BERT (double) and
base BERT are statistically significant for the Terms of Service task,
with 𝑝-value < 0.001. Additionally, the mean differences between
F1 for paired folds of Legal-BERT and BERT (double) with 𝑝-value
< 0.001 and the mean differences between F1 for paired folds of
Custom Legal-BERT and BERT (double) with 𝑝-value < 0.001 are
statistically significant for the CaseHOLD task. The substantial per-
formance gains from the Legal-BERT model variants were achieved
likely because the CaseHOLD task is adequately difficult and highly
domain-specific in terms of language.
6.1.3 Domain Specificity Score. Table 5 also provides a measure of
domain specificity of each task, which we refer to as the domain
specificity (DS) score. We define DS score as the average difference
in pretrain loss between Legal-BERT and BERT, evaluated on the
downstream task of interest. For a specific example, we run pre-
diction for the downstream task of interest on the example input
using Legal-BERT and BERT models after pretraining, but before
fine-tuning, calculate loss on the task (i.e., binary cross entropy
loss for Overruling and Terms of Service, categorical cross entropy
loss for CaseHOLD), and take the difference between the loss of
the two models. Intuitively, when the difference is large, the gen-
eral corpus does not predict legal language very well. DS scores
When Does Pretraining Help?
Table 5: Test performance, with ±1.96 × standard error, aggregated across 10 folds. Mean F1 scores are reported for Overruling
and Terms of Service. Mean macro F1 scores are reported for CaseHOLD. The best scores are in bold.
Model Baseline BERT
Overruling — DS=-0.028 0.910 ± 0.012 0.958 ± 0.005
Terms of Service — DS=-0.085 0.712 ± 0.020 0.722 ± 0.015
CaseHOLD — DS=0.084 0.399 ± 0.005 0.613 ± 0.005
Number of Pretraining Steps - 1M
Vocabulary Size (domain) - 30,522 (general)
serve as a heuristic for task domain specificity. A positive value
conveys that on average, Legal-BERT is able to reason more accu-
rately about the task compared to base BERT after the pretraining
phase, but before fine-tuning, which implies the task has higher
legal domain-specificity.
The rank order from least to most domain-specific is: Terms of
Service, Overruling, and CaseHOLD. This relative ordering makes
substantive sense. CaseHOLD has high domain specificity since a
holding articulates a court’s precise, legal statement of the holding
of a decision. As noted earlier, the language of contractual terms-of-
service may not be represented extensively in the case law corpus.
The results in Table 5 outline an increasing relationship between
the legal domain specificity of a task, as measured by the DS score
(compatible with our qualitative assessments of the tasks), and
the degree to which prior legal knowledge captured by the model
through unsupervised pretraining improves performance. Addition-
ally, the Overruling results suggest that there exists an interplay
between the legal domain specificity of a task and the difficulty of
the task, as measured by baseline performance on non-attention
based models. Gains from attention based models and domain pre-
training may be limited for lower difficulty tasks, even those with
intermediate DS scores, such as Overruling, likely because the task
is easy enough provided local context that increased model domain
awareness is only marginally beneficial.
6.2 Task Variants
To provide a further assessment on the conditions for pretraining,
we evaluate the performance and sensitivity of our models on three
task variants of CaseHOLD, the task for which we observe the most
substantial gains from domain pretraining. We vary the task on
three dimensions: the volume of training data available for fine-
tuning (train volume), the difficulty of the prompt as controlled
by the length of the prompt (prompt difficulty), and the level of
domain specificity of the prompt (domain match). We hypothesize
that these dimensions — data volume, prompt difficulty, and domain
specificity — capture the considerations practitioners must account
for in considering whether pretraining is beneficial for their use
case. For the task variants, we split the CaseHOLD task dataset into
three train and test set folds using an 80/20 split over three random
seeds and evaluate on each fold. We report results as the mean F1
over the three folds’ test sets.
6.2.1 Train Volume. For the train volume variant, keeping the test
set constant, we vary the train set size to be of size 1, 10, 100, 500,
165
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
BERT (double) Legal-BERT Custom Legal-BERT
0.958 ± 0.005 0.963 ± 0.007 0.974 ± 0.005
0.773 ± 0.019 0.750 ± 0.018 0.787 ± 0.013
0.623 ± 0.003 0.680 ± 0.003 0.695 ± 0.003
2M 2M 2M
30,522 (general) 30,522 (general) 32,000 (legal)
Figure 1: Mean macro F1 scores over 3 folds, with ±1.96 ×
standard error, for train volume variant.
1,000, 5,000, 10,000, and the full train set. We find that the Legal-
BERT gains compared to BERT (double) are strongest with low
train volume and wash out with high train volume. As we expect,
Legal-BERT gains are larger when the fine-tuning dataset is smaller.
In settings with limited training data, the models must rely more
on prior knowledge and Legal-BERT’s prior knowledge is more
relevant to the highly domain-specific task due to pretraining on
legal domain-specific text, so we see stronger gains from Legal-
BERT compared to BERT (double). For a training set size of 1, the
mean gain in Legal-BERT is 17.6% ± 3.73, the maximal gain across
train set sizes.
This particular variant is well-motivated because it has often
been challenging to adapt NLP for law precisely because there is
limited labeled training data available. Legal texts typically require
specialized legal knowledge to annotate, so it can often be prohibi-
tively expensive to construct large structured datasets for the legal
domain [16].
6.2.2 Prompt Difficulty. For the difficulty variant, we vary the cit-
ing text prompt difficulty, by shortening the length of the prompt to
the first 𝑥 words. The average length of a prompt in the CaseHOLD
task dataset is 136 words, so we take the first 𝑥 = 5, 10, 20, 40, 60,
80, 100 words of the prompt and the full prompt. We take the first
𝑥 words instead of the last 𝑥 words closest to the holding, as the
latter could remove less relevant context further from the holding
and thus make the task easier. We find that the prompt difficulty
variant does not result in a clear pattern of increasing gains from
Legal-BERT over BERT (double) above 20 words, though we would
expect to see the gains grow as the prompt is altered more. How-
ever, a 2% drop in gain is seen in the 5 word prompt (the average
F1 gap above 20 words is 0.062, while at 5 is 0.0391).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 2: Mean macro F1 scores over 3 folds, with ±1.96 ×
standard error, for prompt difficulty variant.
One possible reason we do not observe a clear pattern may be
that the baseline prompt length constrains the degree to which we
can manipulate the prompt and vary this dimension; the expected
relationship may be more clearly observed for a dataset with longer
prompts. Additionally, BERT models are known to disregard word
order [42]. It is possible that beyond 5 words, there is a high like-
lihood that a key word or phrase is encountered that Legal-BERT
has seen in the pretraining data and can attend to.
6.2.3 Domain Match. For the domain match variant, we weight
the predictions for the test set when calculating F1 by sorting the
examples in ascending order by their DS score and weighting each
example by its rank order. Intuitively, this means the weighted F1
score rewards correct predictions on examples with higher domain
specificity more. This method allows us to keep train volume con-
stant, to avoid changing the domain specificity distribution of train
set examples (which would occur if the test set was restricted to a
certain range of DS scores), and still observe the effects of domain
specificity on performance in the test set. We expect that the gains
in Legal-BERT compared to BERT are stronger for the weighted F1
than the unweighted F1. We find that the mean gain in Legal-BERT
over three folds is greater for the weighted F1 compared to the
unweighted F1, but only by a difference of 0.8% ± 0.154, as shown
in Table 6.
Table 6: Mean gain in Legal-BERT over 3 folds, for domain
match variant.
Mean macro F1 BERT (double) Legal-BERT Mean Gain
Unweighted 0.620 0.679 0.059
Weighted 0.717 0.784 0.067
One possible reason this occurs is that the range of DS scores
across examples in the CaseHOLD task is relatively small, so some
similarly domain-specific examples may have fairly different rank-
based weights. In Figure 3, we show histograms of the DS scores
of examples for Terms of Service, CaseHOLD, and 5,000 examples
sampled (without replacement) from each task. Notice that the
Terms of Service examples are skewed towards negative DS scores
and the CaseHOLD examples are skewed towards positive DS scores
so the range of DS scores within a task is limited, while the examples
sampled from both tasks span a larger range, explaining the small
gains from the domain match variant, but more substantial gains
for CaseHOLD from Legal-BERT compared to Terms of Service. In
166
Zheng and Guha, et al.
Figure 3: Density histograms of DS scores of examples for
Terms of Service, CaseHOLD, and both tasks.
other words, because the CaseHOLD task is already quite domain
specific, variation within the corpus may be too range-restricted to
provide a meaningful test of domain match.
Further work could instead examine domain match by pretrain-
ing on specific areas of law (e.g., civil law) and fine-tuning on other
areas (e.g., criminal law), but the Harvard case law corpus does not
currently have meaningful case / issue type features.
6.3 Error Analysis
We engage in a brief error analysis by comparing the the breakdown
of errors between Legal-BERT and BERT (double). In the test set,
the breakdown was: 55% both correct, 13% Legal-BERT correct /
BERT (double) incorrect, 7% BERT (double) correct / Legal-BERT
incorrect, and 25% both incorrect. We read samples of instances
where model predictions diverged, with a focus on the examples
Legal-BERT predicted correctly and BERT (double) predicted in-
correctly. While we noticed instances indicative of Legal-BERT
attending to legal language (e.g., identifying a different holding
because of the difference between a “may" and “must" and because
the “but see" citation signal indicated a negation), we did not find
that such simple phrases predicted differences in performance in
a bivariate probit analysis. We believe there is much fruitful work
to be done on further understanding what Legal-BERT uniquely
attends to.
6.4 Limitations
While none of the CaseHOLD cases exist in the pretraining dataset,
some of Legal-BERT gains on the CaseHOLD task may be attribut-
able to having seen key words tied to similar holding formulations
in the pretraining data. As mentioned, this is part of the goal of
the task: understanding the holdings of important cases in a min-
imally labeled way and determining how the preceding context
may affect the holding. This would explain the varying results in
the prompt difficulty variant of the CaseHOLD task: gains could
be mainly coming from attending to only a key word (e.g., case
name) in the context. This may also explain how Legal-BERT is able
to achieve zero-shot gains in the train volume variant of the task.
BERT, may have also seen some of the cases and holdings in English
Wikipedia,4 potentially explaining its zero-shot performance im-
provements over random in the train volume variant. Future work
on the CaseHOLD dataset may wish to disentangle memorization
of case names from the framing of the citing text, but we provide a
strong baseline here. One possible mechanism for this is via a future
variant of the CaseHOLD task where a case holding is paraphrased
to indicate bias toward a different viewpoint from the contextual
4 See,
e.g., https://en.wikipedia.org/wiki/List_of_landmark_court_decisions_in_the_
United_States which contains a list of cases and their holdings.
When Does Pretraining Help?
framing. This would reflect the first-year law student exercise of
re-framing a holding to persuasively match their argument and
isolate the two goals of the task.
7 DISCUSSION
Our results resolve an emerging puzzle in legal NLP: if legal lan-
guage is so unique, why have we seen only marginal gains to do-
main pretraining in law? Our evidence suggests that these results
can be explained by the fact that existing legal NLP benchmark
tasks are either too easy or not domain matched to the pretrain-
ing corpus. Our paper shows the largest gains documented for
any legal task from pretraining, comparable to the largest gains
reported by SciBERT and BioBERT [3, 25]. Our paper also shows
the highest performance documented for the general setting of
the Terms of Service task [26], suggesting substantial gains from
domain pretraining and tokenization.
Using a range of legal language tasks that vary in difficulty and
domain-specificity, we find BERT already achieves high perfor-
mance for easy tasks, so that further domain pretraining adds little
value. For the intermediate difficulty task that is not highly domain-
specific, domain pretraining can help, but gain is most substantial
for highly difficult and domain-specific tasks.
These results suggest important future research directions. First,
we hope that the new CaseHOLD dataset will spark interest in
solving the challenging environment of legal decisions. Not only
are many available benchmark datasets small or unavailable, but
they may also be biased toward solvable tasks. After all, a com-
pany would not invest in the Overruling task (baseline F1 with
BiLSTM of 0.91), without assurance that there are significant gains
to paying attorneys to label the data. Our results show that domain
pretraining may enable a much wider range of legal tasks to be
solved.
Second, while the creation of large legal NLP datasets is impeded
by the sheer cost of attorney labeling, CaseHOLD also illustrates an
advantage of leveraging domain knowledge for the construction of
legal NLP datasets. Conventional segmentation would fail to take
advantage of the complex system of legal citation, but investing in
such preprocessing enables better representation and extraction of
legal texts.
Third, our research provides guidance for researchers on when
pretraining may be appropriate. Such guidance is sorely needed,
given the significant costs of language models, with one estimate
suggesting that full pretraining of BERT with a 15GB corpus can
exceed $1M. Deciding whether to pretrain itself can hence have
significant ethical, social, and environmental implications [4]. Our
research suggests that many easy tasks in law may not require do-
main pretraining, but that gains are most likely when ground truth
labels are scarce and the task is sufficiently in-domain. Because
estimates of domain-specificity across tasks using DS score match
our qualitative understanding, this heuristic can also be deployed to
determine whether pretraining is worth it. Our results suggest that
for other high DS and adequately difficult legal tasks, experimen-
tation with custom, task relevant approaches, such as leveraging
corpora from task-specific domains and applying tokenization /
sentence segmentation tailored to the characteristics of in-domain
text, may yield substantial gains. Bender et al. [4] discuss the signif-
icant environmental costs associated in particular with transferring
167
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
an existing large language models to a new task or developing new
models, since these workflows require retraining to experiment
with different model architectures and hyperparameters. DS scores
provide a quick metric for future practitioners to evaluate when
resource intensive model adaptation and experimentation may be
warranted on other legal tasks. DS scores may also be readily ex-
tended to estimate the domain-specificity of tasks in other domains
with existing pretrained models like SciBERT and BioBERT [3, 25].
In sum, we have shown that a new benchmark task, the Case-
HOLD dataset, and a comprehensively pretrained Legal-BERT model
illustrate the conditions for domain pretraining and suggests that
language models, too, can embed what may be unique to legal
language.
ACKNOWLEDGMENTS
We thank Devshi Mehrotra and Amit Seru for research assistance,
Casetext for the Overruling dataset, Stanford’s Institute for Human-
Centered Artificial Intelligence (HAI) and Amazon Web Services
(AWS) for cloud computing research credits, and Pablo Arredondo,
Matthias Grabmair, Urvashi Khandelwal, Christopher Manning,
and Javed Qadrud-Din for helpful comments.
REFERENCES
[1] Nikolaos Aletras, Dimitrios Tsarapatsanis, Daniel Preoţiuc-Pietro, and Vasileios
Lampos. 2016. Predicting judicial decisions of the European Court of Human
Rights: A natural language processing perspective. PeerJ Computer Science 2
(2016), e93.
[2] Pablo D Arredondo. 2017. Harvesting and Utilizing Explanatory Parentheticals.
SCL Rev. 69 (2017), 659.
[3] Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciBERT: A Pretrained Language
Model for Scientific Text. In Proceedings of the 2019 Conference on Empirical
Methods in Natural Language Processing and the 9th International Joint Conference
on Natural Language Processing (EMNLP-IJCNLP). Association for Computational
Linguistics, Hong Kong, China, 3615–3620.
[4] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret
Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be
Too Big? . In Proceedings of the 2021 ACM Conference on Fairness, Accountability,
and Transparency. Association for Computing Machinery, New York, NY, USA,
610–623.
[5] Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. 2019. Neural Legal
Judgment Prediction in English. In Proceedings of the 57th Annual Meeting of
the Association for Computational Linguistics. Association for Computational
Linguistics, Florence, Italy, 4317–4323. https://www.aclweb.org/anthology/P19-
1424
[6] Ilias Chalkidis, Ion Androutsopoulos, and Achilleas Michos. 2017. Extracting
Contract Elements. In Proceedings of the 16th Edition of the International Con-
ference on Articial Intelligence and Law (London, United Kingdom) (ICAIL ’17).
Association for Computing Machinery, New York, NY, USA, 19–28.
[7] Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras,
and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of
Law School. In Findings of the Association for Computational Linguistics: EMNLP
2020. Association for Computational Linguistics, Online, 2898–2904. https:
//www.aclweb.org/anthology/2020.findings-emnlp.261
[8] Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, and Ion Androut-
sopoulos. 2019. Neural Contract Element Extraction Revisited. Workshop on
Document Intelligence at NeurIPS 2019. https://openreview.net/forum?id=
B1x6fa95UH
[9] Columbia Law Review Ass’n, Harvard Law Review Ass’n, and Yale Law Journal.
2015. The Bluebook: A Uniform System of Citation (21st ed.). The Columbia Law
Review, The Harvard Law Review, The University of Pennsylvania Law Review,
and The Yale Law Journal.
[10] Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Ziqing Yang, Shijin Wang, and
Guoping Hu. 2019. Pre-Training with Whole Word Masking for Chinese BERT.
arXiv:1906.08101 [cs.CL]
[11] Laura C. Dabney. 2008. Citators: Past, Present, and Future. Legal Reference
Services Quarterly 27, 2-3 (2008), 165–190.
[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Proceedings of the 2019 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Volume 1 (Long and
Short Papers). Association for Computational Linguistics, Minneapolis, Minnesota,
4171–4186. https://www.aclweb.org/anthology/N19-1423
[13] Pintip Hompluem Dunn. 2003. How judges overrule: Speech act theory and the
doctrine of stare decisis. Yale LJ 113 (2003), 493.
[14] Emad Elwany, Dave Moore, and Gaurav Oberoi. 2019. BERT Goes to Law School:
Quantifying the Competitive Advantage of Access to Large Legal Corpora in
Contract Understanding. arXiv:1911.00473 http://arxiv.org/abs/1911.00473
[15] David Freeman Engstrom and Daniel E Ho. 2020. Algorithmic accountability in
the administrative state. Yale J. on Reg. 37 (2020), 800.
[16] David Freeman Engstrom, Daniel E. Ho, Catherine Sharkey, and Mariano-
Florentino Cuéllar. 2020. Government by Algorithm: Artificial Intelligence in
Federal Administrative Agencies. Administrative Conference of the United States,
Washington DC, United States.
[17] European Union 1993. Council Directive 93/13/EEC of 5 April 1993 on unfair terms
in consumer contracts. European Union.
[18] Dan Hendrycks, Collin Burns, Steven Basart, Andrew Critch, Jerry Li, Dawn
Song, and Jacob Steinhardt. 2021. Aligning AI With Shared Human Values.
arXiv:2008.02275 [cs.CY]
[19] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn
Song, and Jacob Steinhardt. 2021. Measuring Massive Multitask Language Un-
derstanding. arXiv:2009.03300 [cs.CY]
[20] Michael J. Bommarito II, Daniel Martin Katz, and Eric M. Detterman. 2018.
LexNLP: Natural language processing and information extraction for legal and
regulatory texts. arXiv:1806.03688 http://arxiv.org/abs/1806.03688
[21] Mandar Joshi, Danqi Chen, Yinhan Liu, Daniel S. Weld, Luke Zettlemoyer, and
Omer Levy. 2020. SpanBERT: Improving Pre-training by Representing and
Predicting Spans. Transactions of the Association for Computational Linguistics 8
(2020), 64–77. https://www.aclweb.org/anthology/2020.tacl-1.5
[22] David Jurgens, Srijan Kumar, Raine Hoover, Dan McFarland, and Dan Jurafsky.
2018. Measuring the Evolution of a Scientific Field through Citation Frames.
Transactions of the Association for Computational Linguistics 6 (2018), 391–406.
https://www.aclweb.org/anthology/Q18-1028
[23] Minki Kang, Moonsu Han, and Sung Ju Hwang. 2020. Neural Mask Generator:
Learning to Generate Adaptive Word Maskings for Language Model Adaptation.
In Proceedings of the 2020 Conference on Empirical Methods in Natural Language
Processing (EMNLP). Association for Computational Linguistics, Online, 6102–
6120. https://www.aclweb.org/anthology/2020.emnlp-main.493
[24] Taku Kudo and John Richardson. 2018. SentencePiece: A simple and language
independent subword tokenizer and detokenizer for Neural Text Processing.
arXiv:1808.06226 [cs.CL]
[25] Jinhyuk Lee, Wonjin Yoon, Sungdong Kim, Donghyeon Kim, Sunkyu Kim,
Chan Ho So, and Jaewoo Kang. 2019. BioBERT: a pre-trained biomedical language
representation model for biomedical text mining. Bioinformatics 36, 4 (2019),
1234–1240.
[26] Marco Lippi, Przemysław Pałka, Giuseppe Contissa, Francesca Lagioia, Hans-
Wolfgang Micklitz, Giovanni Sartor, and Paolo Torroni. 2019. CLAUDETTE:
an automated detector of potentially unfair clauses in online terms of service.
Artificial Intelligence and Law 27, 2 (2019), 117–139.
[27] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A
Robustly Optimized BERT Pretraining Approach. arXiv:1907.11692 [cs.CL]
[28] David Mellinkoff. 2004. The language of the law. Wipf and Stock Publishers,
Eugene, Oregon.
[29] Elizabeth Mertz. 2007. The Language of Law School: Learning to “Think Like a
Lawyer”. Oxford University Press, USA.
[30] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient
Estimation of Word Representations in Vector Space. http://arxiv.org/abs/1301.
3781
[31] Octavia-Maria, Marcos Zampieri, Shervin Malmasi, Mihaela Vela, Liviu P. Dinu,
and Josef van Genabith. 2017. Exploring the Use of Text Classification in the
Legal Domain. Proceedings of 2nd Workshop on Automated Semantic Analysis
of Information in Legal Texts (ASAIL).
[32] Adam R. Pah, David L. Schwartz, Sarath Sanga, Zachary D. Clopton, Peter DiCola,
Rachel Davis Mersey, Charlotte S. Alexander, Kristian J. Hammond, and Luís
A. Nunes Amaral. 2020. How to build a more open justice system. Science 369,
6500 (2020), 134–136.
168
Zheng and Guha, et al.
[33] Anusri Pampari, Preethi Raghavan, Jennifer Liang, and Jian Peng. 2018. emrQA:
A Large Corpus for Question Answering on Electronic Medical Records. In
Proceedings of the 2018 Conference on Empirical Methods in Natural Language
Processing. Association for Computational Linguistics, Brussels, Belgium, 2357–
2368. https://www.aclweb.org/anthology/D18-1258
[34] Marc Queudot, Éric Charton, and Marie-Jean Meurs. 2020. Improving Access to
Justice with Legal Chatbots. Stats 3, 3 (2020), 356–375.
[35] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. 2018. Im-
proving language understanding by generative pre-training.
[36] Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016.
SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceed-
ings of the 2016 Conference on Empirical Methods in Natural Language Pro-
cessing. Association for Computational Linguistics, Austin, Texas, 2383–2392.
https://www.aclweb.org/anthology/D16-1264
[37] Siva Reddy, Danqi Chen, and Christopher D Manning. 2019. Coqa: A conversa-
tional question answering challenge. Transactions of the Association for Compu-
tational Linguistics 7 (2019), 249–266.
[38] Anna Rogers, Olga Kovaleva, and Anna Rumshisky. 2021. A primer in bertol-
ogy: What we know about how bert works. Transactions of the Association for
Computational Linguistics 8 (2021), 842–866.
[39] Phillip Rust, Jonas Pfeiffer, Ivan Vulić, Sebastian Ruder, and Iryna Gurevych. 2020.
How Good is Your Tokenizer? On the Monolingual Performance of Multilingual
Language Models. arXiv:2012.15613 [cs.CL]
[40] Jaromir Savelka, Vern R Walker, Matthias Grabmair, and Kevin D Ashley. 2017.
Sentence boundary detection in adjudicatory decisions in the United States.
Traitement automatique des langues 58 (2017), 21.
[41] Or Sharir, Barak Peleg, and Yoav Shoham. 2020. The Cost of Training NLP Models:
A Concise Overview. arXiv:2004.08900 [cs.CL]
[42] Koustuv Sinha, Prasanna Parthasarathi, Joelle Pineau, and Adina Williams. 2020.
Unnatural Language Inference. arXiv:2101.00010 [cs.CL]
[43] Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin
Tian, Danxiang Zhu, Hao Tian, and Hua Wu. 2019. ERNIE: Enhanced Represen-
tation through Knowledge Integration. arXiv:1904.09223 [cs.CL]
[44] P.M. Tiersma. 1999. Legal Language. University of Chicago Press, Chicago,
Illinois. https://books.google.com/books?id=Sq8XXTo3A48C
[45] George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioannis Partalas,
Matthias Zschunke, Michael R. Alvers, Dirk Weissenborn, Anastasia Krithara,
Sergios Petridis, Dimitris Polychronopoulos, Yannis Almirantis, John Pavlopou-
los, Nicolas Baskiotis, Patrick Gallinari, Thierry Artiéres, Axel-Cyrille Ngonga
Ngomo, Norman Heino, Eric Gaussier, Liliana Barrio-Alvers, Michael Schroeder,
Ion Androutsopoulos, and Georgios Paliouras. 2015. An overview of the BIOASQ
large-scale biomedical semantic indexing and question answering competition.
BMC Bioinformatics 16, 1 (April 2015), 138.
[46] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel
Bowman. 2018. GLUE: A Multi-Task Benchmark and Analysis Platform for
Natural Language Understanding. In Proceedings of the 2018 EMNLP Workshop
BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP. Association
for Computational Linguistics, Brussels, Belgium, 353–355. https://www.aclweb.
org/anthology/W18-5446
[47] Jonah Wu. 2019. AI Goes to Court: The Growing Landscape of AI for Access
to Justice. https://medium.com/legal-design-and-innovation/ai-goes-to-court-
the-growing-landscape-of-ai-for-access-to-justice-3f58aca4306f
[48] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V. Le, Mohammad Norouzi,
Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, Jeff
Klingner, Apurva Shah, Melvin Johnson, Xiaobing Liu, Łukasz Kaiser, Stephan
Gouws, Yoshikiyo Kato, Taku Kudo, Hideto Kazawa, Keith Stevens, George Kurian,
Nishant Patil, Wei Wang, Cliff Young, Jason Smith, Jason Riesa, Alex Rudnick,
Oriol Vinyals, Greg Corrado, Macduff Hughes, and Jeffrey Dean. 2016. Google’s
Neural Machine Translation System: Bridging the Gap between Human and
Machine Translation. arXiv:1609.08144 [cs.CL]
[49] Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and
Maosong Sun. 2020. How Does NLP Benefit Legal System: A Summary of Legal
Artificial Intelligence. In Proceedings of the 58th Annual Meeting of the Association
for Computational Linguistics. Association for Computational Linguistics, Online,
5218–5230. https://www.aclweb.org/anthology/2020.acl-main.466
[50] Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and
Maosong Sun. 2020. JEC-QA: A Legal-Domain Question Answering Dataset. ,
9701-9708 pages.
 Part II

Short Papers
 Practical Tools from Formal Models: The ECHR as a Case Study
Katie Atkinson
Joe Collenette∗
Trevor Bench-Capon∗
{katie,j.m.collenette,tbc}@liverpool.ac.uk
Department of Computer Science
University of Liverpool
Liverpool, UK
ABSTRACT
One approach to building legal support systems is to run an ex-
ecutable model of the relevant knowledge through an interface
designed to collect information from the user and provide explana-
tions. The usability of such systems depends on the terms used in
the law being represented: often only users familiar with the prac-
tice and application of the law will be able to provide the required
information. Earlier work applied this approach to the European
Convention on Human Rights (ECHR). Although the performance
of the tool built for that domain was good, the questions posed to
the user demanded a good deal of knowledge and experience of
the ECHR. Here we use the knowledge of an expert with extensive
experience of the ECHR to extend the model, through intermediate
levels, to identify questions that are appropriate to the target user.
We have undertaken a pilot evaluation in which a small number
of lawyers have used the prototype program and provided very
positive feedback, showing that they are receptive to AI solutions
that give effective, explainable decision support.
CCS CONCEPTS
• Applied computing → Law.
KEYWORDS
ADFs, case-based reasoning, explainability, ECHR
ACM Reference Format:
Katie Atkinson, Joe Collenette, Trevor Bench-Capon, and Kanstantsin Dze-
htsiarou. 2021. Practical Tools from Formal Models: The ECHR as a Case
Study. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY,
USA, 5 pages. https://doi.org/10.1145/3462757.3466095
1 INTRODUCTION
Since the pioneering formalisation of the British Nationality Act
(BNA) [16], one popular approach to building legal support systems
has been to formalise the law and then execute the formalisation
through a program able to gather facts about particular cases from
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466095
170
Kanstantsin Dzehtsiarou
dzeh@liverpool.ac.uk
Department of Law
University of Liverpool
Liverpool, UK
the user and provide explanations. In [16] the law in question was
statute law, but the approach can also be applied to case law, e.g. [1].
Similar structures are found in hybrid systems based on reasoning
with cases such as CABARET [17], IBP [7] and VJAP [12] which
also represent the law as high level rules but instead of directly
querying the user, the leaf predicates are resolved according to
factors found in precedent cases. In all these approaches, we be-
gin with a high level question, such as ‘is Peter a British Citizen?’
and then unfold it through a series of intermediate concepts until
a base level is reached. At this point, hybrid systems will launch
their case based reasoning mechanism, but if we are using a rule
based approach which expects the user to answer these base level
questions, the terms must be readily understood by the user. For a
successful system, two kinds of knowledge are required: knowledge
as to the questions that should be asked, and knowledge as to how
these questions should be answered. A formalisation of the legislation
provides the first but not the second, and so if the law does not
use terms familiar to the users, they will be unable to answer the
questions. This is also true of hybrid systems; the users will need
to ascribe the factors, which is itself a substantial and not entirely
straightforward task. Moreover the knowledge of how to answer
the questions or ascribe the factors may vary according to the back-
ground and skills of the user answering the questions. This suggests
that the formalisation of the law will need to be supplemented by
a further sets of intermediate concepts, one for each type of user
which will unfold into questions appropriate to the users.
With the BNA, the formalisation of the law resulted in questions
such as ‘where was Peter born?’ and ‘who is the father of Peter?’
which appeared to be readily answerable by applicants, or by ad-
judicators on the basis of an application form. Much of the appeal
of this system came from the fact that the questions that resulted
directly from executing the formalisation were immediately and
intuitively understandable in both these situations. However, those
who followed in their footsteps found that this was not always the
case. Problems became apparent in a follow up exercise to the BNA
[5]. Two kinds of problems arose.
First, many of the questions were difficult to answer for, or even,
unintelligible to, the lay user, such as ‘did Peter pay the qualifying
level of contributions in the relevant tax year?’. Such questions might,
however, be answerable by an adjudicator, especially if there was
access to the contributions record database. With other questions
the adjudicator would have the problem: lay users will know their
birth details, but this may require investigation and verification
by the adjudicator, who will therefore need to know what kind of
evidence is required. So the questions may need further refinement
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
to enable the target users to answer them, and may need different
refinements for different target users. For a lay person, the contri-
butions condition will need to be expressed in terms of working in
a particular year, while for the adjudicator, date of birth will need
to ask about acceptable forms of proof, such as birth certificates.
The second problem is that the question may have some partic-
ular legal interpretation. For example, UK Housing Benefit at one
time had an addition if the house was “hard to heat” [4]. While a
lay user may well have an opinion on this, in fact there was a very
technical definition, spelled out in secondary legislation and case
law, in terms of factors such as size and type of house, number of
rooms and age of occupants. Clearly this information is needed
both to enable lay users to answer the question, and to support the
decision making of adjudicators, who may be unfamiliar with the
relevant case law.
In [9] an executable model of Article 6 of the European Con-
vention on Human Rights (ECHR) was presented. However the
questions that resulted from that formal model are unlikely to be
answered with confidence by a potential applicant, and some might
even test an experienced lawyer. To move towards appropriate ques-
tions requires knowledge of the theory and practice of the ECHR.
In this paper we will discuss what is needed for that model to be-
come usable, in particular for those processing initial applications.
Section 2 will briefly describe the ECHR. Section 3 will describe the
formal model of [9]. Section 4 will describe the additional analysis
we have carried out to move to a usable system, and Section 5
the resulting prototype. Section 6 gives an initial evaluation from
members of the intended user group who have used the prototype.
Finally Section 7 will offer some concluding remarks.
2 DOMAIN OVERVIEW
The European Convention on Human Rights (ECHR) is a regional
human rights treaty that is now ratified by 47 European states
and covers almost the whole of Europe. Since 1960, when it was
established, the European Court of Human Rights (ECtHR) has
delivered judgments in thousands of cases and created a significant
body of legal precedents.
The ECHR has proved very popular for experimentation with
machine learning techniques for legal judgment predication tasks;
for example, see [2], [14], [8], [15] and [13]. These studies all report
success, with correct predictions being achieved in around 70-85%
of cases. JURI Says [15] reports a success rate of 69% over the last
year, although it fell to 60.9% for March 20211 .
3 ADFS FOR REASONING ABOUT ECHR
CASES ON ARTICLE 6
The aim of the work described in [9] was to encapsulate Article 6
in an Abstract Dialectical Framework (ADF) [6]. A program based
on the developed ADF predicted whether a particular case was
admissible, and if so whether there was a violation of Article 6. One
of the main strengths of the program was that it was able to justify
its reasoning in terms of the ADF nodes and acceptance conditions,
similar to the how? explanation of a rule based system. The ADF
was created using the ANGELIC methodology [1], designed to cap-
ture case law in a manner that supports argumentation techniques
1 JURI Says can be found at https://jurisays.com/ (accessed 2021/03/01).
171
Katie Atkinson, Joe Collenette, Trevor Bench-Capon, and Kanstantsin Dzehtsiarou
and is maintainable. The ANGELIC ADF corresponds to the factor
hierarchy of traditional case based systems such as CATO [3]. In
an ADF the nodes are connected into a graph structure, and there
are acceptance conditions that determine the acceptability of each
parent in terms of its children. The ADFs produced by ANGELIC
are always trees. In ANGELIC the acceptance conditions take the
form of a set of prioritised sufficient conditions, together with a de-
fault to make the conditions collectively necessary. Maintainability
comes from the modular nature of this structure.
Testing on a limited set of cases achieved a 100% success rate.
Although the test set was small, the benefits of the approach were
clear; the ADF was able to predict Article 6 cases with a high
success rate, and fully justify its reasoning for the predictions. The
developed ADF in [9] included a part relating to admissibility. The
questions developed for admissibility were at a very high level. The
model described in this paper, also developed using the ANGELIC
methodology, focuses entirely on admissibility. This focus allows a
much more detailed analysis to reduce the level of expertise needed
to answer the questions with confidence. The decomposition of
the high level questions obtained from the documentation used in
[9] does, however, require guidance from someone expert in the
relevant statute and case law. We have also replaced the command
line prototype of [9] with a visual interface that uses standard
interactions to make the program more accessible to non-computer
specialists.
4 LEGAL FOUNDATIONS OF THE MODEL
The model described here considers whether an application is ad-
missible or not, which is itself a substantial task. All applications
submitted to the ECtHR need to be admissible in order to be consid-
ered on merits. In other words, the Court needs to establish that the
application complies with a set of formal rules before it can examine
the substance of this application [10]. Admissibility includes two
types of rules: first, the ECtHR needs to establish that the applica-
tion falls within its jurisdiction to confirm that the Court can deal
with this application. Secondly, the ECHR has established a set of
formal rules that the application itself needs to comply with, such
as that the application was first submitted at the national level and
was rejected by the national judicial bodies and that it should be
submitted within 6 months after the highest judicial body rejected
the same application on the national level. This application should
not be abusive, anonymous or trivial. This application also should
not be clearly without merits or – in the ECHR terms – manifestly
ill-founded. Again, if these conditions are not satisfied, the Court
declares an application inadmissible. The Court’s decision as to
inadmissibility is final and cannot be appealed against.
The importance of admissibility is often underestimated. On
average about 90% of all applications submitted to the ECtHR are
declared inadmissible. For instance, in 2019, 44,500 applications
were submitted to the Court and in the same year 38,480 applica-
tions were declared inadmissible. At the same time, in 2019, the
Court delivered only 2,187 meritorious judgments [11]. A large
number of applications is declared inadmissible every year, so our
project has potential importance for both the applicants wanting
to avoid inadmissibility and for the Court for which considering of
inadmissible applications takes a significant proportion of its time
Practical Tools from Formal Models: The ECHR as a Case Study
and resources which could be re-allocated to the meritorious cases
and so reduce the backlog (currently 50,000). Next we describe the
implemented tool that we have produced to enable decision support
for the important issue of admissibility of cases submitted to the
ECtHR. The model used in the tool captures the factors discussed
above that need to be examined to determine admissibility and is a
result of close consultation with our expert on the ECtHR.
5 PROTOTYPE IMPLEMENTATION
The current prototype consists of two parts; the ADF model, exe-
cuted by a JAVA program, and the JAVA front-end. The ADF model
makes the predictions from a set of answers to the questions gath-
ered from the user through the front-end, which then presents an
explanation of the prediction to the user. The aim of the application
is to allow people with minimal legal training to quickly identify
whether a case that is being submitted to the ECtHR is likely to be
admissible. The current ADF is more extensive in the modelling of
admissibility when compared to the previous ADF developed in [9].
There are 61 questions which when answered allow the 26 factors
to resolve to a prediction. In the previous ADF of [9] there were
only 5 factors relating to admissibility and only 7 questions needed
to be answered.
When developing the ADF we consulted a legal expert in order
to decompose the high level nodes used in [9] and so capture the
legal knowledge required to ensure that the list of questions is both
as complete as possible, and appropriate to the target users. Table 1
shows the text and the acceptance conditions associated with the
nodes. The root of the ADF is “V1" which indicates whether sub-
mitting the application to the ECHR is recommended. The program
will recommend submission if both issues, nodes I1 and I2, are ac-
cepted. That is, the application is admissible (I2) and the application
complies with rule 47 of the court (I1). In turn, node I1 is accepted
if the base level factors, I1Q1 and I1Q2, are accepted and the ab-
stract factors that represents that all necessary signatures (I1F1)
and all documentation (I1F2) have been accepted. Abstract Factor
I1F1 has a number of different base level factors, offering different
possibilities for I1F1 to be accepted. These represent the different
signatures that are needed depending on different situations.
The questions used as part of the ADF are not only based on the
requirements of the ECHR but also take into account its case law.
The case law makes the questions more nuanced and integrates
those aspects of admissibility that do not obviously flow from the
text of the Convention. In practice it is almost impossible to submit
a successful application without familiarising oneself with the case
law, which can be quite broad and diverse, or consulting with a
professional lawyer specialising in the law of the ECtHR. This
model presents the rules enshrined in the case law of the Court in
a simplified form.
In Table 1, I2F2Q4 asks “Is the applicant a potential victim of a
violation?" which affects whether the applicant has victim status.
It is not obvious to a lay person that potential victims can apply
to the Court, in certain limited circumstances. The scope of these
circumstances was made clear in the case law: the question has
arisen in the context of extradition cases, where the question was
whether there was a risk of human rights violations in the receiving
state if extradition were permitted e.g. Soering v. the United Kingdom
172
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 1: The results screen of the JAVA prototype. The user
is presented with a high level result and the reasoning be-
hind the decision is shown in the text below.
(1989). Identifying such questions requires a good knowledge of
the legal practice.
When the JAVA program loads, it parses a text representation of
the ADF, similar to Table 1. This supports maintenance by enabling
any changes to the ADF to be automatically reflected in the program.
The program asks the user a series of yes/no questions and once all
the relevant questions have been answered the results are shown
to the user. The results screen also displays the reasoning as to
why that decision was made. Figure 1 shows the results for an
example where the program informs the user that a signature is
required for the form and explains that as the applicant does not
have legal representation and they have not signed the application
form, not all signatures have been provided, which in turn means
that the application does not comply with rule 47 of the rules of
the court and therefore the program recommends not to submit the
application.
To give a better user experience than the previous work [9],
only questions that are needed to generate a recommendation are
asked: if a node can be resolved, the program moves on to the
next. Different paths lead to different questions, e.g. the questions
needed if the application is submitted by a company rather, than as
an individual, are notably different.
Using the example in Figure 1 and referring to Table 1, we will
show how the explanation is generated in Figure 1. The last ques-
tion to be answered is I1F1Q2. We start by printing all the base
level factors that have the same parent as the last base level factor,
as long as the user has answered the corresponding question. The
parent of I1F1Q2 is I1F1, and the only other base level factor that
has an answer is I1F1Q1. This gives us the first two lines of explana-
tion. Next we print the abstract factors and issues in a hierarchical
manner, adding the word “So” before each explanation. Finally we
print the root and add the word “Therefore” at the start of the root’s
explanation.
This prototype has a number of benefits when compared to the
system in [9]. Despite the two systems tackling two different aspects
of legal reasoning within the ECHR domain, they are comparable
as both systems use an underlying ADF that is presented to a user.
The major benefit of the current approach is the ease of use when
compared to the previous work. We have now incorporated both
knowledge of the law, and how that law is applied in practice,
which allows expression in terms of questions which the users can
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Katie Atkinson, Joe Collenette, Trevor Bench-Capon, and Kanstantsin Dzehtsiarou

Table 1: Nodes of the ADF that are referenced, along with their acceptance conditions.

ID Factor Text Accepting Logic

V1 Submission Recommendation ACCEPT IF I1 AND I2
REJECT OTHERWISE
I1 Application compiles with Rule 47 ACCEPT IF I1Q1 AND I1Q2 AND I1F1 AND I1F2
REJECT OTHERWISE
I1Q1 Has the applicant identified the state against which the appli- ACCEPT IF TRUE
cation is brought to the Court (p2 of the application form)? REJECT OTHERWISE
I1Q2 Has the applicant ticked an appropriate box on p2 of the appli- ACCEPT IF TRUE
cation form? REJECT OTHERWISE
I1F1 All Signatures Provided ACCEPT IF (I1F1Q1 AND I1F1Q4 AND (I1F1Q2 OR I1F1Q3))
OR (!I1F1Q1 AND I1F1Q3) OR (I1F1Q8 AND ((!I1F1Q1 AND
I1F1Q7) OR (I1F1Q1 AND I1F1Q5 AND (I1F1Q6 OR I1F1Q7))))
REJECT OTHERWISE
I1F1Q1 Does the applicant have a legal representative? ACCEPT IF TRUE
REJECT OTHERWISE
I1F1Q2 Was the application form signed (p13) by the applicant’s legal ACCEPT IF TRUE
representative? REJECT OTHERWISE
I1F2 All Documentation ACCEPT IF I1F2Q1 AND I1F2Q2
REJECT OTHERWISE
I2 Application is Admissible ACCEPT IF I2F1 AND I2F2 AND I2F4 AND I2F5 AND (I2F3F1
AND I2F3F2 AND !I2F3F3 AND I2F3F4 AND I2F3F5)
REJECT OTHERWISE
I2F2 Victim Status ACCEPT IF !I2F2Q3 AND (I2F2Q1 OR I2F2Q2 OR I2F2Q4)
REJECT OTHERWISE
I2F2Q4 Is the applicant a potential victim of a violation? ACCEPT IF
TRUE REJECT OTHERWISE

be expected to answer . For example, in [9] there is a question “Was

No. of Responses

3
the person bringing the case the victim”. To answer the question
effectively the user needs to understand what constitutes a person, 2
a case, and a victim. The current ADF has decomposed the base 1
level factor into abstract factors such as I2F1 which describes what
constitutes a valid petitioner, and I2F2 which describes victim status 0
Q1
Q2
Q3
Q4
Q5
Q6
Q7
Q8
Q9

0
1
(see Table 1 for questions associated with these abstract factors). Q1
Q1
The questions associated with these abstract factors require less Positive Responses Question Number
specialist knowledge. Improvements have been made to the method Negative Responses
of interacting with the application, which is more in line with what
non-specialist computer users would expect from an application. Figure 2: Graph showing number of feedback responses to
The previous work used a text-only, command line interface [9]. each question, as positive or negative feedback
The new application uses a visual, mouse-driven, interface.

6 PILOT STUDY
Before embarking upon a full evaluation, we have conducted a pilot
study in which the prototype has been tested by a sample of our
target audience, which is a small group of lawyers who work within
the ECHR. Three independent lawyers (not, of course, including our
domain expert) tested the prototype and completed a questionnaire
that covered five different aspects of the prototype.
The following is the list of questions that the users of the pro-
totype were presented with, along with the question category in
brackets. Each question had four possible responses, for different
levels of agreement:
(1) (Functionality) Does the program have a reasonable response
time?
(2) (Functionality) Did the program run to completion without
any interruptions?
(3) (Usability) How easy was the program to use?
(4) (Usability) How intuitive was the program to start using?
(5) (Explainability) How effective was the explanation given for
describing the program’s decisions?
(6) (Explainability) How easy was the information to parse?
(7) (Usefulness) How useful would you find this program for
assisting you in your work?

173
Practical Tools from Formal Models: The ECHR as a Case Study
(8) (Usefulness) Generally how useful would additional technol-
ogy be for assisting with legal work?
(9) (Questions about Questions) How clear were the questions
that you answered within the program? Additionally the
users were able to say which questions were unclear.
(10) (Questions about Questions) How much time would you
save if you used a fully functional program for your work
on deciding on the admissibility of cases?
(11) (Questions about Questions) Does the program reflect how
you decide on the admissibility of cases that you process?
In Figure 2 the responses to the questionnaire have been con-
densed into positive or negative responses, where the top two
answers to a question are positive and the bottom two are negative.
Though the results of the questionnaire come from a very small
sample, with only three lawyers completing the survey, we can con-
clude that the program developed worked well and was functional,
as all the responses received on functionality (Q1, Q2) and usability
(Q3, Q4) were positive. Another positive outcome is that two of the
three ECHR lawyers responded that they trusted that the justifica-
tions for the decisions made were sensible and understandable (Q5)
and all three respondents agreed that the information was easy to
parse (Q6). All respondents saw the usefulness of our prototype
(Q7), with two respondents stating they would use the program as it
currently stands, and the other affirming the usefulness but saying
that some (as opposed to many) changes are needed. Again all the
respondents agreed that technology has a role to play in the legal
domain (Q8): one respondent said technology is needed rapidly,
while two cautioned that careful development will be needed. The
positive responses to Q9 and Q10 were particularly pleasing, since
these questions directly concerned the central aims of this exercise:
the users all agreed that the questions were suitable for them and
that the program would save them time when assessing admis-
sibility. While the majority of the feedback has been positive, it
has also highlighted the need for domain experts to be a part of
the development process (Q11): although two respondents felt the
program reflected all or part of their own process of dealing with
admissibility, one felt that only some aspects had been covered.
Overall the response to the program is very positive and indicates
that it is a sound basis for further developments for our legal de-
cision support tools, giving potential for collaborations between
computer scientists and lawyers. Encouraged by these results, we
will now extend the study to a larger group of potential users, to
ensure that this initial evaluation is reflected in the law community.
7 SUMMARY AND NEXT STEPS
Our aim in this paper has been to conduct a deep dive into the legal
analysis required to provide a robust, executable model of Article 6
of the ECHR. We focused on the theory and practice of a particular
issue within the the ECHR, namely admissibility of cases. In con-
sultation with our legal expert, we defined an ADF that captures
the domain knowledge relevant for the issue of admissibility and
transformed our model into an implemented tool that is able to ask
questions appropriate to the target user. This is a necessary step
to enable academic exercises on legal case-based reasoning to be
transformed into usable tools. Our prototype tool implements the
back-end reasoning of the underpinning ADF and provides us with
174
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
solid grounding on the road to development of a tool for use in
practice.
The prototype tool has been tested by a small group of the tar-
get audience of lawyers. Their response has been very positive,
highlighting that lawyers are receptive to tools that are carefully
designed and targeted at problems they encounter with their work,
all of which paves the way for a fuller evaluation.
The continuation of the work will focus on expanding on the use
of the prototype. We are currently organising, with the assistance of
our ECHR expert, a field evaluation of our tool by users who work
within the ECtHR. This will provide a greater range of feedback
in order to adjust our tool as needed to best fit the audience’s
needs. Once we have a satisfactory tool for use by those assessing
applications – which would provide a way of addressing the current
significant backlog of unprocessed cases in the ECtHR – we will
develop a set of questions for use by applicants themselves. Such
a facility should reduce the number of inadmissible applications
by enabling applicants to gain a better understanding of what is
required to make an admissible application. In this way, access to
to the ECtHR could be improved, both by assisting applicants, and
by speeding up the decision process for cases submitted.
REFERENCES
[1] Latifa Al-Abdulkarim, Katie Atkinson, and Trevor Bench-Capon. 2016. A method-
ology for designing systems to reason with legal cases using ADFs. AI and Law
24, 1 (2016), 1–49.
[2] Nikolaos Aletras, Dimitrios Tsarapatsanis, Daniel Preoţiuc-Pietro, and Vasileios
Lampos. 2016. Predicting judicial decisions of the European Court of Human
Rights: A natural language processing perspective. PeerJ Computer Science 2
(2016), e93.
[3] Vincent Aleven. 1997. Teaching case-based argumentation through a model and
examples. Ph.D. thesis. University of Pittsburgh.
[4] Trevor Bench-Capon. 1991. Practical legal expert systems: the relation between
a formalisation of legislation and expert knowledge. In Law, Computer Science
and Artificial Intelligence, M Bennun and A Narayanan (Eds.). Ablex, 191–201.
[5] Trevor Bench-Capon, Gwen Robinson, Tom Routen, and Marek Sergot. 1987.
Logic programming for large scale applications in law: A formalisation of sup-
plementary benefit legislation. In Proceedings of the 1st ICAIL. 190–198.
[6] Gerhard Brewka and Stefan Woltran. 2010. Abstract dialectical frameworks. In
Twelfth International Conference on the Principles of Knowledge Representation
and Reasoning. 102–111.
[7] Stefanie Brüninghaus and Kevin D Ashley. 2003. Predicting outcomes of case
based legal arguments. In Proceedings of the 9th ICAIL. ACM, 233–242.
[8] Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. 2019. Neural legal
judgment prediction in English. arXiv preprint arXiv:1906.02059 (2019).
[9] Joe Collenette, Katie Atkinson, and Trevor Bench-Capon. 2020. An explainable
approach to deducing outcomes in European Court of HumanRights cases using
ADFs. In Proceedings COMMA 2020. IOS Press, 21–32.
[10] Fiona De Londras and Kanstantsin Dzehtsiarou. 2018. Great Debates on the Euro-
pean Convention on Human Rights. Macmillan International Higher Education.
[11] European Court of Human Rights. 2019. Analysis of statistics. (2019). https:
//www.echr.coe.int/Documents/Stats_analysis_2019_ENG.pdf.
[12] Matthias Grabmair. 2017. Predicting trade secret case outcomes using argument
schemes and learned quantitative value effect tradeoffs. In Proceedings of the 16th
ICAIL. 89–98.
[13] Arshdeep Kaur and Bojan Bozic. 2019. Convolutional Neural Network-based
Automatic Prediction of Judgments of the European Court of Human Rights.. In
27th AIAI Irish Conference on AI and Cognitive Science. CEUR 2563, 458–469.
[14] Masha Medvedeva, Michel Vols, and Martijn Wieling. 2019. Using machine
learning to predict decisions of the European Court of Human Rights. AI and
Law (2019), 1–30.
[15] Masha Medvedeva, Michel Vols, and Martijn Wieling. 2020. URI SAYS: An
Automatic Judgement Prediction System for the European Court of Human
Rights.. In Proceedings of JURIX 2020. 277–280.
[16] Marek J. Sergot, Fariba Sadri, Robert A. Kowalski, Frank Kriwaczek, Peter Ham-
mond, and H Terese Cory. 1986. The British Nationality Act as a logic program.
Commun. ACM 29, 5 (1986), 370–386.
[17] David Skalak and Edwina L Rissland. 1992. Arguments and cases: An inevitable
intertwining. AI and Law 1, 1 (1992), 3–44.
 On the relevance of algorithmic decision predictors for judicial
decision making
Floris Bex
Utrecht University, The Netherlands
Tilburg University, The Netherlands
f.j.bex@uu.nl
ABSTRACT
In this article, we discuss case decision predictors, algorithms which,
given some features of a legal case predict the outcome of the
case (i.e. the decision of the judge). We discuss whether, and if so
how, such prediction algorithms can be used to support judges
in their decision making process. We conclude that case decision
predictors can only be useful in individual cases if they can give
legal justifications for their predictions, and that only these legal
justifications are what should matter for a judge.
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Natural language processing; Machine learning.
KEYWORDS
legal prediction, legal decision making, application of algorithms
ACM Reference Format:
Floris Bex and Henry Prakken. 2021. On the relevance of algorithmic de-
cision predictors for judicial decision making. In Eighteenth International
Conference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021,
São Paulo, Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.
1145/3462757.3466069
1 INTRODUCTION
The prediction of the decision of legal cases by means of machine-
learning algorithms has become a hot topic [1, 3, 4, 12, 16]. Such
algorithmic predictors can have various uses in the law. In this
paper we discuss their application to support judges in individual
cases, focusing on algorithmic decision predictors: algorithms that
predict the final decision of a legal case, a decision that would oth-
erwise be made by the judge(s) (such as guilty/not guilty, rule for
plaintiff/defendant). Algorithmic decision predictors are sometimes
claimed to improve the predictability and consistency of judicial
decision making, which is demanded by the principle of equality
(cf. [10]). According to these claims, judges can use decision predic-
tors in order to come to more consistent, more informed and less
biased judgments [4, 8, 17]. Others, however, fear that when judges’
decisions are informed by algorithmic case predictors, people will
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466069
175
Henry Prakken
Utrecht University, The Netherlands
University of Groningen, The Netherlands
h.prakken@uu.nl
not be judged any more on the legal merits of their individual case
but on the basis of general statistics [19]. This is related to O’Neill’s
[18] criticism of ‘bucketing’, the practice of basing a decision about
an individual (e.g., about granting the person a loan) on the fact
that the individual is member of a particular class of which a statis-
tical frequency is known instead of on the particular situation of
that individual. O’Neill [18, pp. 145–6] argues that, although this
strategy might optimise the decision maker’s profit in the long run,
it may lead to unjust decisions in individual cases.
To be able to evaluate this debate it is necessary to have a clear
picture on what information a prediction of a decision by an algo-
rithm in a particular case gives to the judge deciding the case. One
answer is given in [4]: “an AI system can be trained to accurately
forecast based on past behaviour what a user’s decision would be in
a situation absent lapses in rationality.” So if an algorithm performs
well on a test set and if it predicts a particular decision in a new
case, then an arbitrary rationally-thinking judge would if assigned
to the case, take the predicted decision. Of course, algorithms are
rarely 100% accurate, so we look at the probability that an arbitrary
competent judge assigned to the case would take a predicted de-
cision. We want to investigate to what extent an algorithmic case
prediction can yield such a decision probability: how, and under
which assumptions, does a prediction in a particular case combined
with information about an algorithm’s performance on a test set
yield a decision probability for a new case?
This last question immediately gives rise to a new question: why
would judges be interested in probabilities at all when deciding a
case? After all, we expect judges not to give probabilistic reasons
for their decisions (except perhaps on matters of fact) but legal
reasons. Still, judges have always looked at what their colleagues
decide in similar cases and there are good reasons for doing so,
such as improving the consistency of intra-judicial decision making
[10, par. 8]. Underlying this is the assumption that if the great
majority of their colleagues would take the same decision, then
it presumably is the right decision. Of course, this assumption is
at best defeasible and this leads to a second idea, namely, that if
an algorithmic decision predictor performs well in the test phase,
then its predictions yield the ‘normal’ decision of the case, so that
a judge could only deviate from a prediction if there are special
circumstances in the case. We also want to investigate to what
extent such thinking is justified.
To address these questions, we first in Section 2 give a brief
overview of the main types of algorithmic case predictors for legal
cases. We then discuss in Section 3 the various senses in which
probabilities can be derived given an algorithm and its evaluation
with a test set. The heart of our paper is Section 4, in which we
discuss to which extent the probabilities derived from an algorithm
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
and its evaluation can be applied to a new case that is to be decided
in court. Our main conclusion will be that in practice such an ap-
plication is almost never warranted. We then in Section 5 discuss
what this means for the hope that the use of algorithmic case pre-
dictors by judges in individual cases will improve the consistency
and predictability of judicial decision making.
2 ALGORITHMIC DECISION PREDICTORS
Algorithmic decision predictors come in, roughly, three types: pre-
dictors on the basis of legally relevant factors, predictors on the
basis of features unrelated to the merits of a case and predictors on
the basis of the textual description of a case (for a recent overview
see [3]). We focus on supervised classification algorithms that pre-
dict a categorical outcome – one of multiple possible decisions,
such as affirm/reverse, guilty/not guilty – and not on algorithms
that provide a continuous output – such as a regression algorithm
that predicts the length of a sentence or the amount of damages to
be paid. Furthermore, note that we only focus on predictors that,
given some features of a case, predict the final decision of the case,
and that we do not include, for example, algorithms for estimating
recidivism risk, as these do not provide a final case decision.
Predicting on the basis of legally relevant factors. One approach
predicts decisions on the basis of legally relevant factors in a case,
by using either machine-learning techniques or a symbolic model
of legal reasoning. 1 This approach describes the facts of a case at a
higher level of abstraction than the concrete facts. The factors are
assumed to be legally relevant for the case decision, so they can be
used for generating informative explanations of a prediction.
The first studies into prediction on the basis of factors applied
general machine-learning techniques to encodings of cases in terms
of legally relevant factors. An early AI & law example is Mackaay
& Robillard [14], who studied the prediction of a type of Cana-
dian tax case with the nearest-neighbor rule. In AI & Law, various
factor-based models for case-based reasoning have been used for
generating knowledge-based case decision predictions without the
use of machine learning techniques. Examples are the studies of
Ashley and his PhD students on the case law concerning misuse of
trade secrets in American law [2, 7]. Accuracy levels were obtained
of up to 88% [2] and 92% [7]. An advantage of this approach is that
the arguments generated about the predicted decision can be used
as explanations of the prediction based on legal knowledge and in
a form not unlike the arguments of human judges or lawyers.
Predicting on the basis of case metadata. Several authors have
used supervised machine learning based on case features that are
not related to the merits of the case. An example is the algorithm
that predicts decisions of the American Supreme Court on the basis
of structured metadata such as the kind of case, the date at which it
was decided and which lower court decided the original case [12].
This algorithm, which correctly predicted 70% of the decisions,
cannot explain the predicted decisions in a legally meaningful way,
since the features on the basis of which it makes its predictions are
‘extra-legal’, that is, they are not related to the merits of the case.
1 ‘Factors’
are here not just CATO-style boolean factors but any abstract fact pattern
that can have two or more values.
176
Bex and Prakken
Predicting on the basis of the textual description of a case. Other
algorithms predict decisions based on the text of case law, where
statistical correlations are identified between, for example, word
combinations in the text and the case decision. Examples are algo-
rithms that predict whether the European Court of Human Rights
(ECHR) will for a specific article from the Convention with the
same name decide whether that article was violated, on the basis of
part of the text of the decision by the Court [1, 16] or the facts of
the case as communicated to the parties [15]. The performance of
these different algorithms is largely comparable, with accuracy and
F-measures ranging between 75% and 80%. Although it would seem
that this kind of algorithm looks at the legal aspects of the case
(procedural history, facts), the identified statistical correlations do
not say anything about the legally relevant reasons for the decision
of a case. Therefore these algorithms can also not explain their
predicted decisions in a legally meaningful way.
3 FROM ALGORITHM PERFORMANCE TO
PROBABILITIES
Recall that we want to investigate to which extent the performance
of an algorithm on a test set justifies the idea that an arbitrary
competent judge assigned to a case will likely take the predicted
decision. We call the probability at stake here the decision probabil-
ity, the probability that an arbitrary competent and rational judge
assigned to a particular case will take decision 𝑋 in that case, given
that the algorithm predicts “𝑋 ”, that is, that the case will receive
decision 𝑋 . In formulas this is 𝑃𝑟 (𝑋 |“𝑋 ”), where “𝑋 ” stands for the
algorithm predicting decision 𝑋 . The precise way in which this
decision probability can be determined is for present purposes irrel-
evant, but the idea is that this probability can somehow be derived
from the algorithm’s performance on the test set. One candidate
method is using Precision, the percentage of predictions “𝑋 ” on the
test set that are correct (i.e. the true positives divided by the total
number of positively predicted cases). Interpreted as a frequency-
type probability, the precision is 𝑃𝑟 (𝑋 |“𝑋 ”), which looks like the
decision probability we are after. However, we do not commit to ex-
actly this way of determining the decision probability – for present
purposes, all that is relevant is that this decision probability will be
defined in terms of an algorithm’s application to a test set, and the
crucial thing to note is that this makes the step to a probability of
the same form for a new case that is not in the test set non-trivial.
4 APPLYING GENERAL FREQUENCY-TYPE
PROBABILITIES TO NEW CASES
For the answer to the question how the step from a probability
derived from performance on a test to a probability for a new
case can be made, we turn to the philosophy of probability theory.
Philosophers distinguish frequency-type and belief-type probability.
Probabilities are of the frequency type if they are based on relative
frequencies. Usually, the frequencies are relative to outcomes of
experiments that can be repeated indefinitely, such as tossing a coin
or rolling a dice, but we consider the special case where they are
derived from a given finite set of test cases. Dawid [9] calls such
probabilities ‘statistical’ probabilities. In contrast, probabilities are
of the belief type if they are about the degree to which a proposition
is believable. Such probabilities can also be attached to propositions
On the relevance of algorithmic decision predictors for judicial decision making
that a single event occurs. The probabilities that can be defined
in terms of an algorithm’s performance on the test set are all of
the frequency type, since they are based on the relative number of
true/false positives/negatives. However, what we want is a belief-
type probability, namely, the probability that a given new case will
be decided as predicted by the algorithm.
So what we are interested in is what information a prediction of
a decision gives to a judge in a particular case that the judge has to
decide. The italicised words are crucial, since when a probability is
interpreted as a frequency (or in Dawid’s [9] terms as a statistical
probability), it does not by itself say anything about a particular case.
As is well known (e.g. [11, p. 137]), there is a logical gap between
frequencies and an individual probability: turning a frequency-type
probability into a probability about a particular case is a decision,
which has to be justified. Now how can this decision be justified?
It turns out that this requires a number of assumptions.
4.1 From the test set to the set of future cases
Clearly, the move from the past to the future is only justified if the
set of future cases has the same proportions as the test set. However,
this is not guaranteed (see also e.g. [5, 6]). First, the decisions of
judges can change in that they start deciding on different grounds
or weighing reasons in different ways than they used to do. This can
happen, for instance, when moral or political opinions in society
change, or because different judges with different legal opinions
are assigned to the same type of case. Also, the distribution of types
of cases can change because of changes in the world. Moreover, the
algorithm could be overfitted on inessential features of the training
data (a well-known problem in statistics and machine learning). So
(as is well known in the literature on machine learning) in order to
accept a probability based on the test set as a probability for a future
set of cases, we have to make at least the following assumptions:
judges continue to decide cases on the same grounds; the frequency
of the various types of cases remains the same; and the algorithm
made its predictions on the test set for the right reasons.
4.2 Yielding a decision probability for an
individual case
This is not yet all. If the assumptions listed in the previous section
are justified, then all we know is that the frequency-type probability
derived from the test set can also be applied to a future set of cases
(which can be open-ended). However, we are not after a probability
of a kind of event (decisions predicted by this algorithm) but after
the probability of a single event (this decision predicted by this
algorithm). The former can be frequency-type but the latter must
be belief-type. We could apply the so-called frequency principle
[11, p. 137] and let the latter equate the former. However, if we
do so, that is, if we base our probabilities concerning individual
case decision predictions on frequencies, then we in fact make a
crucial assumption. This assumption is that the only ways in which
cases can relevantly differ is in the properties on which the relative
frequencies are defined, that is, on their real and predicted decision,
just as in familiar text book examples about urns with coloured balls
the only relevant way in which the balls can differ is in their colour.
While in the textbook examples this assumption is justified, for
legal cases it is not. Judges who have to decide a case know much
177
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
more about it than its predicted decision. And the point is that if a
judge has more information than just membership of the ‘reference
class’ of the relative frequency (for instance, ‘80% of the cases with
predicted decision 𝑋 have decision 𝑋 ’), then it is irrational to rely
on the frequency-based probability concerning that class. Instead,
one should look at the probability of the decision conditional on the
more specific reference class that corresponds to one’s knowledge
about the case. And this, of course, amounts to thinking about the
particulars of case as judges are used to do.
Our argument is an instance of what philosophers call the prob-
lem of finding the right reference class when performing ‘direct
inference’. It is this reference-class argument that gives a philosoph-
ical justification for O’Neill’s [18] criticism of ‘bucketing’ and more
generally for the fear of trial by statistics. In essence it means that
if nothing more is known of an algorithmic decision predictor than
its performance on the test set, then its predicted decisions cannot
be regarded as the decision that an arbitrary judge assigned to the
case would likely take. So a judge who wants to know what his
or her colleagues would likely decide in an individual case, should
not consult the algorithm since it does not provide the correct de-
cision probability for the case. This in turn means that there is no
meaningful sense in which an algorithmically predicted decision is
the ‘normal’ decision for the case, from which a judge could only
deviate if he or she can point at special circumstances that make
this case different than a normal case of this kind.
To explain this further, imagine that cases are distributed in such
a way that many cases are ‘clear’, for which a decision predictor
would always be correct, but many other cases are ‘hard’, for which
a decision predictor would often be incorrect, but the algorithm
cannot explain to which type a new case belongs. Then only in the
clear cases can the predicted decision be said to be the ’normal’ one.
But how can the judge know which case is easy and which case is
hard? To know this, the judge has to think about the particulars of
the case as judges always do. But then the judge can just as well
ignore the algorithmic prediction.
4.3 Objections to the reference class argument
In the previous subsection we concluded that in practice it will be
impossible to rationally derive a case-specific decision probability
from frequency-type probabilities based on experiments with a test
set, so that judges who want to know what their colleagues would
likely decide in the case cannot obtain an answer to their question
by consulting a case decision predictor. We now discuss possible
objections to our reference-class argument for this conclusion.
First, it might be argued that it is still rational to stick to a
statistical decision probability for a new case, since there often are
no statistics on which a more specific frequency-type probability
can be based. Yet this is a reasoning fallacy: if one wants to express
a decision probability in such cases, one should take the additional
information into account. If this cannot be done on the basis of
known frequencies, one should form a probability based about one’s
information about the specific case, on the penalty of making the
unfounded assumption that this additional information is irrelevant
for the decision of the case (cf. [11, p. 137]).
A variant of this argument is the argument that a belief-type
probability is always less well founded than a frequency-based
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
probability, so that a judge who wants to know what his or her
colleagues would likely decide can still look at what an algorithmic
decision predictor with a high precision predicts. However, this
argument fails, since if one knows more about the case, then sticking
to the frequency is even less well-founded. Consider the analogy of
an urn with 80% red balls and 20% blue balls. If this is all one knows
and one draws a ball from the top of the urn, then it is rational
to assume that there is an 80% probability that it will be red. But
suppose now that the person who filled the urn tells you that he
first put all the red balls in and then all the blue balls, and that he
did not shake or stir, and that you take the ball from the top of the
urn. It is now irrational to stick to the 80% probability that the ball
will be red. In fact, the inverse probability (just 20% chance that the
ball will be red) seems more rational.
One may also consider technical solutions to the reference class
problem. The first is to inspect the test set to check the algorithm’s
performance on subsets of test cases of particular types, as an at-
tempt to make it more likely that the class memberships considered
for the algorithm’s performance coincide with the knowledge the
judge has about a particular case. This is a good idea in theory,
but note that this approach in fact amounts to building a legal-
knowledge model of the reasons relevant for a decision. Moreover,
the created subclasses may be too small to yield reliable probabili-
ties, since in the law the collections of cases usually are not very
big [5]. Furthermore, a too fine-grained feature set may lead to an
overfitted model that does not easily generalise [6, p. 9].
A second technical solution is to obtain the probability for a
single case directly from the algorithm, that is, the probability of a
certain decision 𝑋 given the set of features 𝐹 that represents the
case, or 𝑃𝑟 (𝑋 |𝐹 ). Simpler predictive algorithms directly output such
a prediction probability for a single case, and for e.g. neural net-
works or support vector machines (SVMs) it is possible to estimate
this probability based on the output of the model (cf. [20]). It can be
argued that it is exactly this probability the judge needs: the algo-
rithm captures the behaviour of the judges in the training set cases
and then directly outputs the probability that these judges would
rule 𝑋 in a case like the current one with features 𝐹 . However, this
still does not yield the probability that an arbitrary judge would
rule in that way given the case, because there need not be a relation
between the correctness of predictions and the prediction probabil-
ity. For example, the algorithm can predict the wrong decision with
a high probability, or the algorithm may over- or underestimate
individual probabilities simply because this leads to better classifi-
cation performance. Furthermore, using such advanced techniques
brings along even more assumptions and makes it even harder to
determine what exactly the given probability means, particularly
for a judge with no background in statistics or machine learning.
So instead of relying on this probability, the judge would be better
off thinking about the particulars of case as normal.
A final objection is that an algorithm does not have to be perfect,
as long as it performs better than human decision makers. Here
sometimes the medical domain is mentioned, in which it is widely
accepted that, for instance, a human oncologist has to consult a
data-driven predictive algorithm for recognising skin cancer if this
algorithm has been proven to perform better than humans [21].
However, this analogy beaks down, since unlike in the medical
example, a legal predictive algorithm and a judge perform different
178
Bex and Prakken
tasks. In the medical example human and algorithm perform the
same task, namely, recognising cancer in images of, for instance,
birthmarks. Moreover, the estimates of human and algorithm are
compared to the same (objective) truth: by examining the cells un-
der a microscope it can be determined with certainty whether there
is cancer. Thus a human expert and an algorithmic expert are com-
pared in terms of the same standard. In such a case a comparison
between how humans and algorithms perform is meaningful and
the algorithm can be said to perform better than the human doctor,
namely, by recognising malign spots missed by the human doctor.
However, an algorithmic decision predictor performs a different
task than the judge. A decision predictor predicts which decision a
judge would take, which is a different task than the task the judge
performs, which is deciding the case. Then it is meaningless to
compare the performance of the algorithm and the human judge.
What is more, even a correct prediction of a legally incorrect deci-
sion would count as a success for the predictive algorithm. Such
situations may arise, for instance, since the test set contains legally
incorrect decisions [5]. Correctly predicting a decision is not the
same as predicting a correct decision.
5 CAN A DECISION PREDICTOR IMPROVE
PREDICTABILITY AND CONSISTENCY?
In Section 4 we concluded that a judge who has to decide a case
and who wants to know what an arbitrary rational judge assigned
to the case would probably decide, cannot rely on the statistics
provided by (the evaluation of) an algorithmic decision predictor.
However, this leaves the question what other benefits consulting
such an algorithm can have for a judge in an individual case. This
we discuss below, focusing on the alleged benefit of improving the
predictability and consistency of judicial decision making.
First, we have to determine what the terms predictability and
consistency mean in this context. Assuming that they mean the
same, there are two interpretations. One interpretation is that the
same case is decided the same by different judges. Another interpre-
tation is that similar cases are decided in the same way (or a similar
way) by the same or different judges. The second interpretation
implies the first but not vice versa.
We can now ask how an algorithmic decision predictor can be
used in order to improve predictability and consistency. If these
terms mean that the same case is decided the same by different
judges, then a sure way to guarantee predictability and consistency
is to give all judges the same algorithmic decision predictor and to
require that they all follow its predictions in all cases. Then different
judges would, when assigned to the same case, be guaranteed to
take the same decision. However, this does not make sense, since
as we argued in Section 4.3 we do not know whether all decisions
in the training and test set were correct. If all judges blindly follow
the algorithm’s prediction, then both its accuracy and precision
will increase to 100%, and this would further lead to a tendency
to make the predicted decision the legally correct one even if this
cannot be justified.
What if predictability and consistency mean that similar cases
should be decided the same? Is this improved if we require judges
to consult decision predictors as a source of information? Again, for
mere decision predictors we cannot know. Suppose an algorithm
On the relevance of algorithmic decision predictors for judicial decision making
with 90% precision predicts decision X for case C. Does the judge
then treat like cases alike if s/he follows the prediction? We cannot
know, since the prediction in itself would not give any information
about similarity with other cases. In fact, it might well be that
an algorithm treats cases that judges would regard as similar as
different or vice versa (likewise [6, p. 6]). For example, text-based
decision predictors like the ECHR predictor could fail to recognise
that linguistically small differences are legally very relevant.
However, is this different if the prediction is combined with an
explanation for it? The answer is negative if the explanation cannot
be given in terms of reasons related to the merit of the case. So a
SCOTUS-like predictor is ruled out. But this implies that an ECHR-
type predictor is also ruled out, since it cannot extract any legally
relevant information from the texts to which it is applied, so there is
no way to identify whether its prediction is based on legal grounds
or on extraneous factors. Only decision predictors that base their
predictions on legally relevant factors could possibly yield legally
relevant information about similar cases to a judge.
However, we believe that only these legal explanations are what
should matter for a judge, and that the judge should ignore the fact
that a decision was predicted by an algorithm with good statisti-
cal performance on a test set. This use of such algorithms is not
much different from how judges currently use other information
sources, such as books, journals and peer consultation. Numeri-
cal performance indicators like accuracy, precision and recall can
justify a degree of trust in algorithms in this general sense, but can-
not indicate the quality of individual predictions or explanations.
Moreover, evaluating the quality of algorithmic explanations for
individual predictions requires validation studies of a kind that goes
far beyond the current trend to focus on numerical performance
measures like accuracy, precision and recall and is more akin to an
older AI tradition of carrying out empirical validation studies with
potential or actual users of the algorithm [13].
6 CONCLUSION
In this paper we argued that a judge who has to decide a case
and who wants to know what an arbitrary rational judge assigned
to the case would probably decide, cannot rely on the statistics
provided by (the evaluation of) an algorithmic decision predictor.
The idea that an algorithmic prediction that performed well on a
test set yields the ‘normal’ decision of the case, from which a judge
could only deviate if there are special circumstances in the case, is
unfounded. Moreover, we argued that relying on the predictions of
such algorithms cannot improve the predictability and consistency
of judicial decision making in desirable ways. We believe that mere
decision predictors, that is, predictors that cannot explain their
predictions in legally meaningful terms, should not be used at
all by judges as decision-support tools for individual cases. Such
algorithms do not give any useful information to judges and may
in fact be misleading and cause intellectual laziness.
If an algorithmic decision predictor gives any useful information
to judges at all, it is not in its predictions but in its explanations
for these predictions. However, we noted that whether algorithmic
explanations can indeed improve the quality of judicial decision
making requires validation studies of a kind that goes far beyond
the current trend to focus on numerical performance measures
179
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
like accuracy, precision and recall, and instead involves potential
or actual users of the algorithms. More generally, we believe that
it is important to inform the legal world in transparent language
about not only the potential benefits but also the limitations of
algorithmic outcome predictors.
Finally, we like to emphasise that our conclusions are confined
to the use of algorithmic decision predictors for informing judges
on what they could decide in particular cases. Other uses of such
algorithms may well have benefits, but this requires another paper.
REFERENCES
[1] N. Aletras, D. Tsarapatsanis, D. Preoţiuc-Pietro, and V. Lampos. 2016. Predicting
judicial decisions of the European Court of Human Rights: A natural language
processing perspective. PeerJ Computer Science 2 (2016), e93.
[2] V. Aleven. 2003. Using background knowledge in case-based legal reasoning: a
computational model and an intelligent learning environment. Artificial Intelli-
gence 150 (2003), 183–237.
[3] K. Ashley. 2019. A brief history of the changing roles of case prediction in AI
and law. Law in Context. A Socio-legal Journal 36, 1 (2019), 93–112.
[4] B. Babic, D. Chen, T. Evgeniou, and A.-L. Fayard. 2021. The better way to onboard
AI. Harvard Business Review (2021). http://nber.org/~dlchen/papers/The_Better_
Way_to_Onboard_AI.pdf To appear.
[5] T. Bench-Capon. 2020. The need for good-old fashioned AI and law. In Interna-
tional Trends in Legal Informatics: Festschrift for Erich Schweighofer, W. Hötzen-
dorfer, C. Tschol, and F. Kummer (Eds.). Editions Weblaw, Bern, 23–36.
[6] R. Binns. 2020. Analogies and disanalogies between machine-driven and human-
driven legal judgement. Journal of Cross-disciplinary Research in Computational
Law 1, 1 (2020).
[7] S. Brueninghaus and K. Ashley. 2009. Automatically classifying case texts and
predicting outcomes. Artificial Intelligence and Law 17 (2009), 125–165.
[8] I. Chalkidis, I. Androutsopoulos, and N. Aletras. 2019. Neural legal judgment
prediction in English. In Proceedings of the 57th Annual Meeting of the Association
for Computational Linguistics. 4317–4323.
[9] P. Dawid. 2005. Probaility and Proof. (2005). http://tinyurl.com/tz85o Appendix
to Analysis of Evidence, by T. J. Anderson, D. A. Schum and W. L. Twining.
[10] European Commission for the Efficiency of Justice (CEPEJ). 2018. European
ethical Charter on the use of Artificial Intelligence in judicial systems and their
environment.
[11] I. Hacking. 2001. An Introduction to Probability and Inductive Logic. Cambridge
University Press, Cambridge.
[12] D. Katz, M. Bommarito, and J. Blackman. 2017. A general approach for predicting
the behavior of the Supreme Court of the United States. PloS one 12, 4 (2017),
e0174698.
[13] R. O. Keefe. 1993. Issues in the verification and validation of knowledge-based
systems. In Advances in Software Engineering and Knowledge Engineering, V. Am-
briola and G. Tortora (Eds.). Series on Software Engineering and Knowledge
Engineering, Vol. 2. World Scientific Publishing Co, 173–189.
[14] E. Mackaay and P. Robillard. 1974. predicting judicial decisions: The nearest
neighbor rule and visual representation of case patterns. Datenverarbeitung im
Recht 3 (1974), 302–331.
[15] M. Medvedeva, , X. Xu, M. Vols, and M. Wieling. 2020. JURI SAYS: an automatic
judgement prediction system for the European Court of Human Rights. In
Legal Knowledge and Information Systems. JURIX 2020: The Thirty-Third Annual
Conference, S. Villata, J. Harašta, and P. Křemen (Eds.). IOS Press, Amsterdam
etc., 277–280.
[16] M. Medvedeva, M. Vols, and M. Wieling. 2020. Using machine learning to predict
decisions of the European Court of Human Rights. Artificial Intelligence and Law
28, 2 (2020), 237–266.
[17] F. Muhlenbach and I. Sayn. 2019. Artificial Intelligence and law: What do people
really want?: Example of a French multidisciplinary working group. In Proceedings
of the 17th International Conference on Artificial Intelligence and Law. ACM Press,
New York, 224–228.
[18] C. O’Neil. 2016. Weapons of Math Destruction: How Big Data Increases Inequality
and Threatens Democracy. Crown.
[19] F. Pasquale and G. Cashwell. 2018. Prediction, persuasion, and the jurisprudence
of behaviourism. University of Toronto Law Journal 68, supplement 1 (2018),
63–81.
[20] J. Platt et al. 1999. Probabilistic outputs for support vector machines and com-
parisons to regularized likelihood methods. Advances in large margin classifiers
10, 3 (1999), 61–74.
[21] J. Susskind. 2018. Future Politics: Living Together in a World Transformed by Tech.
Oxford University Press, Oxford.
 The Burden of Persuasion in Structured Argumentation
Roberta Calegari Régis Riveret
Alma Mater Research Institute for Commonwealth Scientific and
Human-Centered Artificial Industrial Research Organisation
Intelligence Brisbane, Australia
Bologna, Italy regis.riveret@data61.csiro.au
roberta.calegari@unibo.it
ABSTRACT
In this paper we provide an account of the burden of persuasion in
the context of structured argumentation. A formal model for the
burden of persuasion is defined, discussed, and used to capture the
role of the burden of persuasion in adjudicating conflicts between
conflicting arguments and in determining the dialectical status of
arguments. We consider how our model can also capture adversarial
burdens of proof, namely, those cases in which failure to establish
an argument for a proposition burdened with persuasion entails
establishing the complementary proposition.
KEYWORDS
Burden of persuasion, argumentation, legal reasoning
ACM Reference Format:
Roberta Calegari, Régis Riveret, and Giovanni Sartor. 2021. The Burden
of Persuasion in Structured Argumentation. In Eighteenth International
Conference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021,
São Paulo, Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.
1145/3462757.3466078
1 INTRODUCTION
The burden of proof is a central feature of many dialectical con-
texts. It is particularly relevant in those domains, such as legal
disputations or political debates, in which controversial issues are
discussed in order to adopt a decision, see [20] on burdens of proof
in different dialogue types. Research in AI & law has devoted a
number of contributions to the formal analysis of burdens of proof:
models of defeasible legal reasoning have been criticised for not
taking burdens of proof into account [11], the distinction between
different standards of proof has been addressed [7], formal accounts
of burdens of proof have been developed within models for formal
argumentation [9, 17]. However, it seems to us that a comprehensive
model of burdens of proof in legal reasoning is still missing.
In the legal domain, two types of burdens can be distinguished:
the burden of production (also called burden of providing evidence,
or ‘evidential’ burden), and the burden of persuasion [17]. This ter-
minology is used in common law systems [21], but the distinction
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466078
180
Giovanni Sartor
Alma Mater Research Institute for
Human-Centered Artificial
Intelligence
Bologna, Italy
giovanni.sartor@unibo.it
is also recognised in civil law jurisdiction, possibly using a different
terminology [10]. The focus of this paper is on the burden of persua-
sion. We will show how an allocation of the burden of persuasion
may induce single outcomes in contexts in which the assessment
of conflicting arguments would, without such an allocation, remain
undecided. Our model combines Prakken and Sartor’s [17] model
with the insight from Carneades’ [8], and takes into account the
fact that the persuasiveness of an argument, in a dialectical context,
is determined not only by the internal strength of the argument, as
determined by the strength of the inference rules used for building
the argument (according, for instance, to the last link criterion),
but also by the applicable counterarguments. Our model originates
from legal considerations and is applied to legal examples [4, 5].
However, the issue of the burden of proof has a significance that
goes beyond the legal domain involving other domains – such as
public discourse, risk management, etc. [3] – in which evidence
and arguments are needed, and corresponding responsibilities are
allocated, according to types of dialogues and dialectical or organi-
sational roles [19, 20]. The novelty of this contribution consists of
a new definition of defeat relations involving arguments burdened
with persuasion, and a corresponding definition of the criteria for
labelling such arguments.
2 BURDENS OF PERSUASION
Let us illustrate how the burden of persuasion works through two
examples, one from criminal law and one from civil law.
Burden of persuasion in criminal law. In criminal law, the burden
of production is distributed between prosecution and defence, while
the burden of persuasion (in most legal systems) is always on pros-
ecution. More exactly, in criminal law, the burden of production
falls on the prosecution relative to the two constitutive elements of
crime, namely, the criminal act (actus reus) and the required mental
state (mens rea), be it intention/recklessness or negligence, while
it falls to the defendant relative to justifications or exculpatory
defences (e.g., self-defence, state of necessity, etc.). In other words,
if both actus reus and mens rea are established, but no exculpatory
evidence is provided, the decision should be criminal conviction. On
the other hand, the burden of persuasion falls on the prosecution
for all determinants of criminal responsibility, including not only
for the constitutive elements of a crime but also for the absence of
justifications of exculpatory defences.
Example 2.1. Let us consider a case in which a woman, Hellen,
has shot and killed an intruder in her home. The applicable law
consists of (a) the rule according to which intentional killing consti-
tutes murder, and (b) the exception according to which there is no
murder if the victim was killed in self-defence. Assume that it has
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
been established with certainty that Hellen shot the intruder and
that she did so intentionally. However, it remains uncertain whether
the intruder was threatening Hellen with a gun, as claimed by the
defence, or had turned back and was running away on having been
discovered, as claimed by prosecution. The burden of persuasion is
on prosecution, who needs to provide a convincing argument for
murder. Since in this case it remains uncertain whether there was
self-defence, prosecution has failed to provide such an argument.
Therefore, the legally correct solution is that there should be no
conviction: Hellen needs to be acquitted. □ and issues that can be found in the complete model [14]. A frame-
Burden of persuasion in civil law. In civil law, burdens of produc-
tion and burdens of persuasion may be allocated in different ways.
The general principle is that the plaintiff only has the burden of
proof (both of production and persuasion) relatively to the operative
facts that ground its claim, while the defendant has the burden of
proof relative to those exceptions which may prevent the operative
facts from delivering their usual outcomes, such as justifications
with regard to torts, or incapability and vices of consent in contracts.
However, derogations from this principle may be established by
the law, in order to take into account various factors, such as the
presumed ability of each party to provide evidence in favour of his
or her claim, the need to protect weaker parties against abuses, etc.
In matters of civil liability, for example, it is usually the case that
the plaintiff, who asks for compensation, has to prove both that the
defendant caused the harm, and that this was done intentionally
or negligently. However, in certain cases, the law establishes an
inversion of the burden of proof for negligence. This means that in
order to obtain compensation, the plaintiff only has to prove that
s/he was harmed by the defendant. This will be sufficient to win
the case unless the defendant provides a convincing argument that
s/he was diligent (not negligent).
Example 2.2. Let us consider a case in which a doctor caused
harm to a patient by misdiagnosing his case. Assume that there is
no doubt that the doctor harmed the patient: she failed to diagnose
cancer, which consequently spread and became incurable. However,
it is uncertain whether or not the doctor followed the guidelines
governing this case: it is unclear whether she prescribed all the
tests that were required by the guidelines, or whether she failed to
prescribe some tests that would have enabled cancer to be detected.
Assume that, under the applicable law, doctors are liable for any
harm suffered by their patients, but they can avoid liability if they
show that they were diligent (not negligent) in treating the patient,
i.e., that they exercised due care. Thus, rather than the patient
having the burden of proving that doctors have been negligent (as
it should be the case according to the general principles), doctors
have the burden of providing their diligence. Let us assume that
the law also says that doctors are considered to be diligent if they
followed the medical guidelines that govern the case. In this case,
given that the doctor has the burden of persuasion on her diligence,
and that she failed to provide a convincing argument for it, the
legally correct solution is that she should compensate the patient.□
These two examples share a common feature. In both, uncertainty
remains concerning a decisive issue. However, this uncertainty does
not preclude the law from prescribing a single legal outcome in
each case. This outcome can be achieved by discarding the argu-
ments that fail to meet the required burden of persuasion, i.e., the
181
Calegari, Riveret and Sartor
prosecution’s argument for murder and the doctor’s argument for
her diligence, respectively.
3 ARGUMENTATION FRAMEWORK
We introduce a structured argumentation framework relying on a
lightweight ASPIC+ -like argumentation system [14]. For the sake
of simplicity, we assume that arguments only consist of defeasible
rules, to the exclusion of strict rules and of some constituents of a
knowledge base—such as axioms, ordinary premises, assumptions,
work based on defeasible rules is sufficient for our purposes and
can be extended as needed with further structures.
3.1 Argumentation graphs
Let any literal be an atomic proposition or its negation. Literals are
brought into relation through defeasible rules.
Notation 3.1. For any literal 𝜙, its complement is denoted by 𝜙, ¯
i.e., if 𝜙 is a proposition 𝑝, then 𝜙¯ is ¬𝑝, while if 𝜙 is ¬𝑝, then 𝜙¯ is 𝑝.
Definition 3.1. A defeasible rule 𝑟 is a construct of the form:
𝜌 : 𝜙 1, ..., 𝜙𝑛 , ∼ 𝜙 1′ , ..., ∼ 𝜙𝑚
′ ⇒𝜓
with 0 ≤ 𝑛 and 0 ≤ 𝑚, and where
• 𝜌 is the unique identifier for 𝑟 , denoted by N(𝑟 );
• each 𝜙 1, . . . 𝜙𝑛 , 𝜙 1′ , ..., 𝜙𝑚 ′ ,𝜓 is a literal;
• 𝜙 1, . . . 𝜙𝑛 , ∼ 𝜙 1, ..., ∼ 𝜙𝑚 are denoted by 𝐴𝑛𝑡𝑒𝑐𝑒𝑑𝑒𝑛𝑡(𝑟 ) and 𝜓 by
′ ′
𝐶𝑜𝑛𝑠𝑒𝑞𝑢𝑒𝑛𝑡(𝑟 );
• ∼ 𝜙 denotes the weak negation (negation by failure) of 𝜙: 𝜙 is
an exception that would block the application of the rule whose
antecedent includes ∼ 𝜙.
The identifier of a rule can be understood as the name of the rule.
It can be used as a literal to specify that the named rule is appli-
cable, and its negation correspondingly to specify that the rule is
inapplicable [13].
A superiority relation ≻ is defined over rules: 𝑠 ≻ 𝑟 states that
rule 𝑠 prevails over rule 𝑟 .
Definition 3.2. A superiority relation ≻ over a set of rules
𝑅𝑢𝑙𝑒𝑠 is an antireflexive and antisymmetric binary relation over
𝑅𝑢𝑙𝑒𝑠, i.e., ≻⊆ 𝑅𝑢𝑙𝑒𝑠 × 𝑅𝑢𝑙𝑒𝑠.
A defeasible theory consists of a set of rules and a superiority
relation over the rules.
Definition 3.3. A defeasible theory is a tuple ⟨𝑅𝑢𝑙𝑒𝑠, ≻⟩ where
𝑅𝑢𝑙𝑒𝑠 is a set of rules, and ≻ is a superiority relation over 𝑅𝑢𝑙𝑒𝑠.
We can construct arguments by chaining rules from the defeasi-
ble theory, as specified in the following definition; cf. [6, 13, 18].
Definition 3.4. An argument 𝐴 constructed from a defeasible
theory ⟨𝑅𝑢𝑙𝑒𝑠, ≻⟩ is a finite construct of the form:
𝐴 : 𝐴1, . . . 𝐴𝑛 ⇒𝑟 𝜙
with 0 ≤ 𝑛, and where
• 𝐴 is the argument’s unique identifier;
• 𝐴1, . . . , 𝐴𝑛 are arguments constructed from the defeasible theory
⟨𝑅𝑢𝑙𝑒𝑠, ≻⟩;
• 𝜙 is the conclusion of the argument, denoted by Conc(𝐴);
The Burden of Persuasion in Structured Argumentation
• 𝑟 : Conc(𝐴1 ), . . . , Conc(𝐴𝑛 ) ⇒ 𝜙 is the top rule of 𝐴, denoted by
TopRule(𝐴).
Notation 3.2. Given an argument 𝐴 : 𝐴1, . . . 𝐴𝑛 ⇒𝑟 𝜙 as in
definition 3.4, Sub(𝐴) denotes the set of subarguments of 𝐴, i.e.,
Sub(𝐴) = Sub(𝐴1 ) ∪ . . . ∪ Sub(𝐴𝑛 ) ∪ {𝐴}.
Different types of inconsistencies can appear between arguments,
causing them to attack each other. In the ASPIC family of argumen-
tation frameworks, attack is differentiated from defeat, with the
latter taking preferences between arguments into account. Prefer-
ences over arguments are defined in the work reported here via a
last-link ordering: an argument 𝐴 is preferred over another argu-
ment 𝐵 if the top rule of 𝐴 is stronger than the top rule of 𝐵.
Definition 3.5. A preference relation ≻ is a binary relation
over a set of arguments A, such that an argument 𝐴 is preferred to
argument 𝐵, denoted by 𝐴 ≻ 𝐵, iff TopRule(𝐴) ≻ TopRule(𝐵).
Before specifying the notion of defeat between arguments, let
us first identify burdens of persuasion, i.e., those literals the proof
of which requires a convincing argument. We assume that such
literals are consistent: it cannot be the case that there is a burden
of persuasion both on 𝜙 and 𝜙.
Definition 3.6 (Burdens of persuasion). Let BurdPers, the set
of burdens of persuasion, be a set of literals such that if 𝜙 ∈
BurdPers then 𝜙 ∉ BurdPers. We say that an argument 𝐴 is bur-
dened with persuasion if Conc(𝐴) ∈ BurdPers.
We now consider possible collisions between arguments, i.e.,
those cases in which an argument 𝐴 challenges an argument 𝐵: (a)
by contradicting the conclusion of a 𝐵’ subargument (rebutting), or
(b) by denying (the application of) the top rule of a 𝐵’ subargument
or by contradicting a weak negation in the body of the top rule of
a 𝐵’ subargument (undercutting). Note that our notion of rebutting
corresponds to the notion of successful rebutting in [14].
Definition 3.7 (bp-rebut). Argument 𝐴 bp-rebuts argument 𝐵 iff
∃𝐵 ′ ∈ 𝑆𝑢𝑏(𝐵) such that Conc(𝐴) = Conc(𝐵 ′ ) and
(1) Conc(𝐴) ∉ BurdPers, and 𝐵 ′ ⊁ 𝐴, or
(2) Conc(𝐴) ∈ BurdPers and 𝐴 ≻ 𝐵 ′ .
According to Definition 3.7, for an unburdened argument 𝐴 to
rebut 𝐵 by contradicting the latter’s subargument 𝐵 ′ , it is sufficient
that 𝐵 ′ is non-superior to 𝐴. For a burdened argument 𝐴 to rebut 𝐵
by contradicting 𝐵 ′ , it is necessary that 𝐴 is superior to 𝐵 ′ . Thus,
burdens of persuasion supplement priorities in deciding conflicts
between arguments having opposed conclusions. They dictate the
outcome of such conflicts when priorities do not already determine
which argument is to prevail: when two arguments contradict one
another, the one burdened with persuasion fails to bp-rebut the
other, while the latter will succeed in bp-rebutting the first.
Undercutting is defined as usual, including both the case in
which an the attacker excludes the application of the top rule of
the attacked argument (by denying the rule’s the name) and the
case in which it contradicts a weakly negated literal in the body of
that rule.
Definition 3.8 (bp-undercut). Argument 𝐴 undercuts argument
𝐵 iff ∃𝐵 ′ ∈ 𝑆𝑢𝑏(𝐵) such that: 1) Conc(𝐴) = ¬N(𝑟 ) and TopRule(𝐵 ′ ) =
𝑟 ; or 2) Conc(𝐴) = 𝜙 and ∼ 𝜙 ∈ 𝐴𝑛𝑡𝑒𝑐𝑒𝑑𝑒𝑛𝑡(TopRule(𝐵 ′ )).
182
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The notions of bp-rebuttings and undercuttings can then be
used to define a defeat relation comprising bp-defeats and strict
bp-defeats between arguments.
Definition 3.9 (bp-defeat). A defeat relation { over a set of
arguments A is a binary relation over A, i.e. {⊆ A × A, such
that ∀𝐴, 𝐵 ∈ A, 𝐴 defeats 𝐵, i.e. 𝐴 { 𝐵, iff 𝐴 bp-defeats 𝐵 or 𝐴
strictly-bp-defeats 𝐵:
(1) 𝐴 bp-defeats 𝐵 iff 𝐴 bp-rebuts 𝐵 or 𝐴 undercuts 𝐵
(2) 𝐴 strictly-bp-defeats 𝐵 iff 𝐴 bp-defeats 𝐵 and 𝐵 does not
bp-defeats 𝐴 .
Example 3.10 (Civil law example: rules and arguments). To ex-
emplify the notions just introduced, let us formalise Example 2.2
through a set of rules. We assume that sufficient evidence is pro-
vided to support (in the absence of evidence to the contrary) the
factual claims at issue (𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠, ¬𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠, ℎ𝑎𝑟𝑚), i.e., that the
corresponding burdens of production are satisfied.
f1 : ⇒ ¬𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 r1 : ¬𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 ⇒ ¬𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒
f2 : ⇒ 𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 r2 : 𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 ⇒ 𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒
f3 : ⇒ ℎ𝑎𝑟𝑚 r3 : ℎ𝑎𝑟𝑚, ∼ 𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒 ⇒ 𝑙𝑖𝑎𝑏𝑙𝑒
We can then build the following arguments:
A1 : ⇒f1 ¬𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 A2 : A1 ⇒r1 ¬𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒
B1 : ⇒f2 𝑔𝑢𝑖𝑑𝑒𝑙𝑖𝑛𝑒𝑠 B2 : B1 ⇒r2 𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒
C1 : ⇒f3 ℎ𝑎𝑟𝑚 C2 : C1 ⇒r3 𝑙𝑖𝑎𝑏𝑙𝑒
If there were no burden of persuasion, the relations would be the
following: arguments A1 and B1 defeat one another, B1 defeats A2,
A1 defeats B2, A2 and B2 defeat one another, B2 strictly defeats C2.
If on the contrary, there is burden of is on the doctors’ diligence
(𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒 ∈ BurdPers), then B2 fails to defeats A2, so that A2
strictly defeats B2. □
Given a defeasible theory, arguments built from it and defeats be-
tween these arguments are gathered into an argumentation graph.
Definition 3.11. An argumentation graph constructed from a
defeasible theory 𝑇 is a tuple ⟨A, {⟩, where A is the set of all
arguments constructed from 𝑇 , and { is a defeat relation over A.
Notation 3.3. Given an argumentation graph 𝐺 = ⟨A, {⟩, we
write A𝐺 , and {𝐺 to denote A and { respectively.
3.2 Labelling semantics
Let us now introduce the notion of {IN, OUT, UND}-labellings of an
argumentation graph, so that each argument in the graph is labelled
IN, OUT or UND, depending on whether it is accepted, rejected, or
undecided, respectively.
Definition 3.12. A {IN, OUT, UND}-labelling 𝐿 of an argumenta-
tion graph 𝐺 is a total function 𝐿 : A𝐺 → {IN, OUT, UND}.
Notation 3.4. Given a labelling 𝐿, we write IN(𝐿) for {𝐴 | 𝐿(𝐴) =
IN }, OUT(𝐿)
for {𝐴 | 𝐿(𝐴) = OUT} and UND(𝐿) for {𝐴 | 𝐿(𝐴) = UND}.
There are various ways to specify {IN, OUT, UND}-labelling func-
tions [1]. For example, they can be complete or grounded.
Definition 3.13. A complete {IN, OUT, UND}-labelling of an argu-
mentation graph 𝐺 is a {IN, OUT, UND}-labelling such that ∀𝐴 ∈ A𝐺
(1) 𝐴 is labelled IN iff all defeaters of 𝐴 are labelled OUT, and
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Calegari, Riveret and Sartor

(2) 𝐴 is labelled OUT iff 𝐴 has a defeater labelled IN.
Definition 3.14. A grounded {IN, OUT, UND}-labelling of an ar-
gumentation graph 𝐺 is a complete {IN, OUT, UND}-labelling 𝐿 of 𝐺
such that IN(𝐿) is minimal.
Remark that any argument not labelled IN or OUT must be labelled
UND, since any { IN, OUT, UND }-labelling is a total function.
While common specifications of {IN, OUT, UND}-labellings define
reasonable positions [1], they do not cater for burdens of persuasion.
We now specify the notion of bp-labelling, namely, a labelling which
persuasion. Let us assume that (as under Italian law) we have
BurdPers = {𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒, 𝑙𝑖𝑎𝑏𝑙𝑒}, i.e., the doctor has to provide
a convincing argument that she was diligent, and the patient has
to provide a convincing argument for the doctor’s liability. As the
burdened doctor’s argument for 𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒 is labelled OUT, her li-
ability can be established even though it remains uncertain whether
the guidelines were followed. □
A2 B2 C2

takes into account a set of burdens of persuasion. UND UND UND

Definition 3.15. A bp-labelling of an argumentation graph

𝐺, relative to a set of burdens of persuasion BurdPers, is a A1 B1 C1

{IN, OUT, UND}-labelling 𝐿 such that ∀𝐴 ∈ A𝐺 UND UND IN

(1) 𝐴 ∈ IN(𝐿) iff ∀𝐵 ∈ A𝐺 such that 𝐵 bp-defeats 𝐴 : 𝐵 ∈ OUT(𝐿)

A2 B2 C2
(2) 𝐴 ∈ OUT(𝐿) iff
(a) Conc(𝐴) ∈ BurdPers and ∃ 𝐵 ∈ A𝐺 such that UND OUT IN

• 𝐵 bp-defeats 𝐴, and 𝐵 ∈ IN(𝐿) or 𝐵 ∈ UND(𝐿)

(b) Conc(𝐴) ∉ BurdPers and ∃ 𝐵 ∈ A𝐺 such that A1 B1 C1
• 𝐵 bp-defeats 𝐴 and 𝐵 ∈ IN(𝐿) UND UND IN

Burdens of persuasion affect conditions for rejection, as speci-

fied in condition 3.15 (2) (a): the rejection (the OUT labelling) of an
argument burdened of persuasion may be determined by any de-
feating counterargument 𝐵 that is accepted (IN) or also is uncertain
(UND). However, as specified in condition 3.15 (2) (b), the rejection
of an argument which is not burdened with persuasion requires a
defeating counterargument 𝐵 that is labelled IN.
The semantics just described does not always deliver a unique
labelling. Multiple labellings may exist when arguments rebut each
other, none of them being burdened with persuasion. If one of
these arguments is labelled IN the other is labelled OUT, and vice
versa. To address such a situation, we focus on IN-minimal labelling
semantics, where for example both such arguments are labelled
UND. Let us call such a labelling a grounded bp-labelling.
Definition 3.16. A bp-labelling 𝐿 of an argumentation graph 𝐺 is
a grounded bp-labelling iff UND(𝐿) is maximal.
Proposition 3.17. Let 𝐿1 be the grounded {IN, OUT, UND}-labelling
of an argumentation graph 𝐺, and 𝐿2 the grounded bp-labelling of 𝐺.
If BurdPers = ∅ then IN(𝐿1 ) = IN(𝐿2 ).
Proof. It is easy to see that if condition 3.15(1) concerning argu-
ments burdened with persuasion is removed from definition 3.15,
we obtain the definition of grounded {IN, OUT, UND}-labellings. □
Example 3.18 (Civil law example: graphs and bp-labelling). Let
us consider again Example 2.2 and the corresponding rules and
arguments built in Example 3.10. The argumentation graph and
its grounded {IN, OUT, UND}-labelling are depicted in Figure 1 (top),
in which all arguments are UND except arguments for undisputed
facts. The result is not satisfactory according to the law, since
it does not take into account the applicable burdens of persua-
sion. The doctor should have lost the case – i.e., be found liable
– since she failed to discharge her burden of proving that she
was diligent (non-negligent). The doctor’s failure results from the
fact that it remains uncertain whether she followed the guide-
lines. To capture this aspect, we need to specify the burdens of
Figure 1: Grounded {IN, OUT, UND}-labelling of Example 2.2
in the absence of burdens of persuasion (top), and its bp-
labelling with BurdPers = {𝑑𝑢𝑒𝐷𝑖𝑙𝑖𝑔𝑒𝑛𝑐𝑒, 𝑙𝑖𝑎𝑏𝑙𝑒} (bottom).
This example shows how the model presented here allows us
to deal with the inversion of the burden of proof, i.e., a situation in
which one argument 𝐴 is presented for a claim 𝜙 being burdened
with persuasion, and 𝐴 (or a subargument of it) is attacked by a
counterargument 𝐵, of which the conclusion 𝜓 is also burdened
with persuasion. If no convincing argument for 𝜓 can be found, then
the attack fails, and the uncertainty on 𝜓 does not affect the status
of 𝐴. In the example, the argument for the doctor’s due diligence
fails to meet its burden of persuasion. Consequently, it fails to defeat
the argument for the doctor’s liability, which succeeds, meeting its
burden of persuasion.
Example 3.19 (Criminal law example: rules, graphs and bp-la-
belling). Referring to Example 2.1, let us consider the following
rules (for simplicity’s sake, we do not specify pieces of evidence
here, but we assume that all factual claims are supported by evi-
dence):
f1: ⇒ 𝑘𝑖𝑙𝑙𝑒𝑑 f2: ⇒ 𝑖𝑛𝑡𝑒𝑛𝑡𝑖𝑜𝑛
f3: ⇒ 𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑 f4: ⇒ ¬𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑
r1: 𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑 ⇒ 𝑠𝑒𝑙 𝑓 𝐷𝑒 𝑓 𝑒𝑛𝑐𝑒 r2: 𝑘𝑖𝑙𝑙𝑒𝑑, 𝑖𝑛𝑡𝑒𝑛𝑡𝑖𝑜𝑛 ⇒ 𝑚𝑢𝑟𝑑𝑒𝑟
r3: 𝑠𝑒𝑙 𝑓 𝐷𝑒 𝑓 𝑒𝑛𝑐𝑒 ⇒ ¬𝑚𝑢𝑟𝑑𝑒𝑟 r4: ¬𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑 ⇒ ¬𝑠𝑒𝑙 𝑓 𝐷𝑒 𝑓 𝑒𝑛𝑐𝑒
with r3 ≻ r2. We can build the following arguments:
A1 : ⇒f1 𝑘𝑖𝑙𝑙𝑒𝑑 B1 : ⇒f3 𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑
A2 : ⇒f2 𝑖𝑛𝑡𝑒𝑛𝑡𝑖𝑜𝑛 B2 : B1 ⇒r1 𝑠𝑒𝑙 𝑓 𝐷𝑒 𝑓 𝑒𝑛𝑐𝑒
A3 : A1, A2 ⇒r2 𝑚𝑢𝑟𝑑𝑒𝑟 B3 : B2 ⇒r3 ¬𝑚𝑢𝑟𝑑𝑒𝑟
C1 : ⇒f4 ¬𝑡ℎ𝑟𝑒𝑎𝑡𝑒𝑑
C2 : C1 ⇒r4 ¬𝑠𝑒𝑙 𝑓 𝐷𝑒 𝑓 𝑒𝑛𝑐𝑒
In the grounded {IN, OUT, UND}-labelling of Figure 2 (top), all argu-
ments are UND except for the undisputed facts. Thus, in the absence

183
The Burden of Persuasion in Structured Argumentation ICAIL’21, June 21–25, 2021, São Paulo, Brazil

of burdens of persuasion, we do not obtain the legally correct an-
swer, namely, acquittal. To obtain acquittal we need to introduce
burdens of persuasion. Prosecution has the burden of persuasion on
murder: it therefore falls to the prosecution to persuade the judge
that there was killing, that it was intentional, and that the killer did
not act in self-defence.
we plan to study the properties of our semantics, and its connection
with other semantics for argumentation [1, 2].
ACKNOWLEDGMENTS
R. Calegari and G. Sartor have been supported by the H2020 ERC
Project “CompuLaw” (G.A. 833647).

A3 B3
REFERENCES
[1] Pietro Baroni, Martin Caminada, and Massimiliano Giacomin. 2011. An introduc-
UND UND
tion to argumentation semantics. The knowledge engineering review 26, 4 (2011),
365–410. https://doi.org/10.1017/S0269888911000166
[2] Pietro Baroni and Régis Riveret. 2019. Enhancing Statement Evaluation in Argu-
A2 B2 C2 mentation via Multi-labelling Systems. Journal Artificial Intelligence Research 66
IN UND UND (2019), 793–860. https://doi.org/10.1613/jair.1.11428
[3] Roberta Calegari, Andrea Omicini, and Giovanni Sartor. 2020. Argumentation
and Logic Programming for Explainable and Ethical AI. In XAI.it 2020 – Italian
A1 B1 C1 Workshop on Explainable Artificial Intelligence 2020 (CEUR Workshop Proceedings,
Vol. 2742). Sun SITE Central Europe, RWTH Aachen University, Italy, 55–68.
IN UND UND [4] Roberta Calegari and Giovanni Sartor. 2020. Burden of Persuasion in Ar-
gumentation. In Proceedings 36th International Conference on Logic Program-
ming (Technical Communications), ICLP 2020 (Electronic Proceedings in Theo-
A3 B3
retical Computer Science, Vol. 325). OPA, Rende (CS), Italy, 151–163. https:
OUT UND //doi.org/10.4204/EPTCS.325.21
[5] Roberta Calegari and Giovanni Sartor. 2020. A Model for the Burden of Persuasion
in Argumentation. In Legal Knowledge and Information Systems. JURIX 2020: The
A2 B2 C2 Thirty-third Annual Conference (Frontiers in Artificial Intelligence and Applications,
Vol. 334), Serena Villata, Jakub Harašta, and Petr Křemen (Eds.). IOS, Brno, Czech
IN UND UND Republic, 13–22. https://doi.org/10.3233/FAIA200845
[6] Martin Caminada and Leila Amgoud. 2007. On the Evaluation of Argumentation
Formalisms. Artificial Intelligence 171, 5—6 (2007), 286–310. https://doi.org/10.
A1 B1 C1 1016/j.artint.2007.02.003
IN UND UND [7] Arthur M. Farley and Kathleen Freeman. 1995. Burden of Proof in Legal Argumen-
tation. In Proceedings of the 5th International Conference on Artificial Intelligence
and Law. ACM, Maryland USA, 156—164. https://doi.org/10.1145/222092.222227
[8] Thomas F. Gordon, Henry Prakken, and Douglas Walton. 2007. The Carneades
Figure 2: Grounded {IN, OUT, UND}-labelling of Example model of argument and burden of proof. Artificial Intelligence 171, 10 (2007),
2.1 in the absence of burdens of persuasion (top), and 875–896. https://doi.org/10.1016/j.artint.2007.04.010
bp-labelling with the burden of persuasion BurdPers = [9] Thomas F. Gordon and Douglas N. Walton. 2009. Proof Burdens and Standards.
In Argumentation in Artificial Intelligence. Springer, Boston, MA, 239–258. https:
{𝑚𝑢𝑟𝑑𝑒𝑟 } (bottom). //doi.org/10.1007/978-0-387-98197-0_12
[10] Ulrike Hahn and Mike Oaksford. 2007. The Burden of Proof and Its Role in
Argumentation. Argumentation 21 (2007), 36–61. https://doi.org/10.1007/s10503-
The bp-labelling is depicted in Figure 2 (bottom). The prosecution 007-9022-6
failed to meet its burden of proving murder, i.e., its argument is [11] Ronald E. Leenes. 2001. Burden of Proof in Dialogue Games and Dutch Civil Pro-
cedure. In Proceedings of the 8th International Conference on Artificial Intelligence
not convincing, since it remains undetermined whether there was and Law. ACM, Missouri USA, 109–18. https://doi.org/10.1145/383535.383549
self-defence. Therefore, the argument supporting murder is labelled [12] Sanjay Modgil and Henry Prakken. 2010. Reasoning about Preferences in Struc-
tured Extended Argumentation Frameworks. In Proceedings of COMMA 2010,
OUT, and the presumed killer is to be acquitted. □ Computational Models of Argumentation. IOS, Italy, 347–58. https://doi.org/10.
3233/978-1-60750-619-5-347
4 CONCLUSION [13] Sanjay Modgil and Henry Prakken. 2014. The ASPIC + framework for structured
argumentation: a tutorial. Argument & Computation 5, 1 (2014), 31–62. https:
We have presented a formal model for the burden of persuasion. The //doi.org/10.1080/19462166.2013.869766
model is based on the idea that arguments burdened with persuasion [14] Henry Prakken. 2010. An Abstract Framework for Argumentation with Struc-
tured Arguments. Argument and Computation 1 (2010), 93–124. https://doi.org/
have to be rejected when there is uncertainty about them. We have 10.1080/19462160903564592
shown how an allocation of the burden of persuasion may lead to a [15] Henry Prakken, Chris Reed, and Douglas N. Walton. 2005. Dialogues about the
single outcome (IN arguments) in contexts in which the assessment Burden of Proof. In Proceedings of the 10th International Conference on Artificial
Intelligence and Law. ACM, Bologna, Italy, 115–124. https://doi.org/10.1145/
of conflicting arguments would otherwise remain undecided. We 1165485.1165503
have also shown how our model is able to address inversions of [16] Henry Prakken and Giovanni Sartor. 1996. Rules about Rules: Assessing Con-
flicting Arguments in Legal Reasoning. Artificial Intelligence and Law 4 (1996),
burdens of proof, namely, those cases in which the burden shifts 331–68. https://doi.org/10.1007/BF00118496
from one party to the other. In such cases, there is the burden of [17] Henry Prakken and Giovanni Sartor. 2010. A Logical Analysis of Burdens of
persuasion over the conclusion of a multistep argument, and at Proof. Legal Evidence and Proof: Statistics, Stories, Logic 1 (2010), 223–253.
[18] Gerard Vreeswijk. 1997. Abstract Argumentation Systems. Artificial Intelligence
the same time a burden of persuasion over the conclusion of an 90, 1–2 (1997), 225–279. https://doi.org/10.1016/S0004-3702(96)00041-0
attacker against a subargument of that multistep argument. The [19] Douglas Walton. 1996. Arguments from Ignorance. Pennsylvania State University
model can be expanded in various ways, to capture further aspects Press, Pennsylvania. https://doi.org/10.1007/978-3-319-15013-0_3
[20] Douglas Walton. 2014. Burden of proof, presumption and argumentation. Cam-
of legal reasoning. For instance, it can also be supplemented with bridge University Press, USA. https://doi.org/10.1017/CBO9781107110311
argumentation over burdens of persuasion [15], in a manner similar [21] C.R. Williams. 2003. Burdens and standards in civil litigation. Sydney Law Review
25 (2003), 165–188.
to the way in which argumentation systems can be expanded to
include argumentation about priorities, see [12, 16]. More generally

184
 Prediction of monetary penalties for data protection cases in
multiple languages
Aaron Ceross
University of Oxford
Oxford, United Kingdom
aaron.ceross@cs.ox.ac.uk
ABSTRACT
As the use of personal data becomes further entrenched in the func-
tion of societal interaction, the regulation of such data continues to
grow as an important area of law. Nevertheless, it is unfortunately
the case that data protection authorities have limited resources to
address an increasing number of investigations. The leveraging of
appropriate data-driven models, coupled with the automation of
decision making, has the potential to help in such circumstances.
In this paper, we evaluate machine learning models in the litera-
ture (such as Support Vector Machine (SVM), Random Forest, and
Multinomial Naive Bayes (MNB) classifiers) for natural language
processing in order to predict whether a monetary penalty was
levied based on a description of case facts. We tested these models
on a novel data set collected from the data protection authority of
Macao across the three languages (i.e., Chinese, English, and Por-
tuguese). Our experimental results show that the machine learning
models provide the necessary predictability in order to automate the
evaluation of data protection cases. In particular, SVM has consis-
tent performance across three languages and achieving an AUROC
of 0.725, 0.762, and 0.748 for Chinese, English, and Portuguese, re-
spectively. We further evaluated the interpretability of the results
independently for each of the languages and found that the salient
texts that were identified are shared across the three languages.
ACM Reference Format:
Aaron Ceross and Tingting Zhu. 2021. Prediction of monetary penalties
for data protection cases in multiple languages. In Eighteenth International
Conference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021,
São Paulo, Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.
1145/3462757.3466097
1 INTRODUCTION
The capture, storage, and processing of personal data within in-
formation systems has become a fundamental feature of societal
interaction, including social relationships, commerce, government,
and education. The increasing multiplicity and complexity of these
interactions across different entities necessitates the promulgation
of rules regarding the use and storage of personal data in order to
prevent misuse of the data, as well as redress power asymmetries
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466097
185
Tingting Zhu
University of Oxford
Oxford, United Kingdom
tingting.zhu@eng.ox.ac.uk
in the relationship between the data subject and the data controller.
Correspondingly, data protection law has gained increased visibility
in recent years, notably the enactment of the European Union’s
General Data Protection Regulation (GDPR) [20].
In general, there exists a disparity between regulators and the
objects of regulation. This includes access to information, resources
to contest regulatory action, and technical expertise — with the bal-
ance of power often favouring corporate entities over the regulating
authority. It has been argued that this leads to inefficient regulatory
action for such corporate entities [4]. Constrained resources are
particularly acute in data protection regulation, with the effect of
increasing pressure on the prioritisation of cases, therefore requir-
ing prudence when taking on an investigation in order to maximise
the effectiveness of regulatory action [5]. The widespread use of
personal data by innumerable entities has resulted in a need to be
selective in terms of determining which cases to take forward so
as to maximise effectiveness of regulation of personal data [14].
Further pressure on resources also comes from newer competences
and regulatory expectation, such as mandatory breach notification.
In this paper we evaluate text classification methods that have the
potential to facilitate this aspect of the regulatory process. We are
unaware of any work in the available literature examining the use
of machine learning methods for data protection regulatory actions.
2 BACKGROUND
2.1 Automation of data protection regulation
Within the available literature we do not find examples using a
data-driven approach for regulatory actions in data protection. The
wider literature regarding empirical analysis of data protection
judgements and regulatory actions has attracted only modest aca-
demic interest. There are few data sets which are readily available,
and this may contribute to the limited research in this area. For
example, Ceross and Simpson [6] provided summary statistics on
civil penalties provided by the United Kingdom’s data protection
authority, however they did not provide a model in this work. Nev-
ertheless, the authors found that regulatory actions were focused
primarily on health and government-held data, with the causes of
breach being non-technological (e.g. improper disposal of records
and unintended disclosure), which may suggest the priorities of the
data protection authority.
2.2 Automated prediction of legal outcomes
through textual analysis
There has been increasing interest in the research literature regard-
ing the prediction of the outcomes of legal cases through machine
learning and natural language processing. This may be due to the
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
greater availability of court data as well as widespread use of ma-
chine learning and natural language processing libraries. Aletras et
al. [1] constructed a dataset of 584 cases decided by the European
Court of Human Rights (ECtHR), focusing on those cases involving
Articles 3, 6, and 8 of the European Convention on Human Rights.
The authors extracted n-grams and semantic topics as features, and
classified the cases using an SVM with 10-fold cross validation. The
approach established by Aletras et al. [1] has acted as a template
for other studies. For example, using a data-set of 5,990 German
tax cases, Waltl et al. [23] employed a Naive Bayes classifier, using
11 linguistic features, 10 fold cross validation. As another example,
Virtucio et al. [22] examined 27,492 cases decided by the Philippine
Supreme Court, achieving a 59% accuracy using a Random Forest
classifier using topics as a feature. More recently, Medvedeva et
al. [16] adopted the approach in [1], expanding the scope to anal-
yse ECtHR decisions across 14 articles. The authors in this work
selected an equal number of decision outcomes in order to maintain
a balanced dataset.
There exist multiple jurisdictions wherein there are multiple
official languages of legal action (e.g, Canada and Belgium). In the
works examining the ECtHR data, the authors did not consider
whether the outcomes were similar for different languages. We add
to the literature by comparing the same facts of a case in three
languages to determine whether predictors of the cases are the
same across the language versions.
3 DATA
3.1 Description
Macao is a Special Administrative Region of the People’s Republic
of China. It is a former colony of Portugal which was transferred to
Chinese rule in 1999. Macao’s designation as a special administra-
tive region has meant that Macao retains a degree of autonomy in its
public administration and legislation. As such, the city has adopted
an independent approach to data protection regulation which is
different from China. The Personal Data Protection Act (Act 8/2005)
(PDPA) [17] regulates the processing and storage of personal data
in Macao and establishes a regulator, the Office for Personal Data
Protection (GPDP).1 The data protection regulation in Macao is
based on the approach in the European Union, which envisions
more prominent roles for data protection authorities [12]. Accord-
ingly, the GPDP has a wide regulatory remit with regard to data
protection, including education, advice, and consultation. As part
of its regulatory functions, Article 33 of the PDPA provides that the
GPDP may impose a monetary penalty ranging from MOP$4,000
to MOP$40,000 (approximately $500 – $5,000 USD) for acts which
(i) infringe on the rights of the data subject (Articles 5, 10, 11, 12,
13); (ii) contravene rules related to security and confidentiality of
processing (Articles 16, 17); or (iii) where the processing of personal
data has not been appropriately publicised (Article 25, paragraph
3). In 2019, the GPDP reported an unusual increase in received en-
quiries regarding data protection (2,940) — which was an increase
of 60% from the previous year and the highest for five years [10].
Such an increase inevitably gives rise to issues of prioritisation and
resource allocation, as discussed in Section 1.
1 Acronym used in English is taken from the authority’s Portuguese name, Gabinete
para a Protecção de Dados Pessoais.
186
Aaron Ceross and Tingting Zhu
3.2 Data collection
In this study, we extracted the text of completed cases-investigations
by the GPDP from 2007 to 2019. As Chinese and Portuguese are
the two co-equal official languages of Macao, the GPDP provides
case-files on its website provided in these languages. Additionally,
the GPDP also provides English translations of the cases, although
the this translation is unofficial. Overall, the case report itself is
divided into three parts: (i) the brief, which explains the facts of the
case; (ii) the analysis, which gives an explanation of the applicable
law and consideration of factors; and (iii) the case outcome, which
details the decision of the GPDP.
The GPDP provides reports for cases whether an infraction of
the PDPA is found or not. This allows for classification of penalties
and closed cases, which represents the full scope of possibilities
with regard to assessment, that data protection authorities make
when deciding on cases. From the retrieved cases, there was a dis-
crepancy between the number of translated cases available from the
website (316 in written Chinese; 309 Portuguese; and 292 English),
indicating that the website does not provide every case across the
three languages. Using the unique case identification numbers, we
kept only those cases common to all three languages, resulting in a
dataset of 281 cases, representing investigations between 2007 and
2019. Out of these 281 cases, 73 (26%) were given a penalty, with the
remaining 208 cases closed without one (75%). Most of the cases in
the data set arise from complaints (185, 66%), followed by reports
(58, 21%), referrals (21, 8%), and active investigations (17, 6%).
4 METHODS
4.1 Data processing and feature extraction
In addition to any of the regulatory actions the GPDP may take, the
case outcomes may result in the authority finding that no further
action is necessary. In this work, we only make use of the case briefs
in order to predict case outcomes. Other sections within the case
documentation address the merits of the claims (the analysis) or
provide the case outcome which may strongly influence the results.
Table 1 describes the summary features of the linguistic properties
of a brief. Chinese text has the shortest mean length of the three
languages, as well as having the least number of tokens. Despite
this, it has a larger vocabulary than the Portuguese and English
text (3,423 types). These smaller tokens and vocabulary may help in
reducing the feature set for the purposes of classification, thereby
allowing for more effective prediction of classes.
Table 1: Descriptive statistics of case briefs.
Language #tokens #types min max mean
length length length
Chinese 19,287 3,423 20 459 68.64
English 26,002 2,256 30 240 92.53
Portuguese 26,157 2,856 26 249 93.09
The cases briefs are written in short sentences describing the
reasons the case has been brought to the GPDP. The brief does not
Prediction of monetary penalties for data protection cases in multiple languages
make an evaluation as to the merits of the case nor do the briefs
provide an indication as to the outcome. The input feature to our
classification models can be defined as a document term frequency
matrix which describes the frequency of terms that occur in a
collection of documents. In order to generate this, we lowercase the
text and remove ‘stop words’, which are frequently occurring words
of no semantic significance (e.g. articles and prepositions). Both
unigram and bigram word features are extracted. While stemming
is a common approach to reduce the number of features in language
processing tasks, research has demonstrated that lemmatisation,
using the infinitive root of a word may produce better outcomes
for tasks [15]. Finally, term-frequency inverse document frequency
(tf-idf) of each word was extracted to reflect how important the
word is in a collection of case briefs.
4.2 Evaluation of classification models
We evaluated the case briefs using the classification models in the
identified literature (see Section 2.2). This includes (i) a Support
Vector Machine (SVM) classifier, which is a non-probabilistic classi-
fier that maps feature values into a higher dimension to maximise
the discrimination between two classes; (ii) a Random Forest (RF)
classifier, which is an ensemble of decision trees via bootstrap-
ping subsets of features; and (iii) Multinomial Naive Bayes (MNB),
a probabilistic classifier which combines Bayes’ theorem with a
multinomial event model, to allow for explicit modelling of the
frequency/count of each of our features.
In our experimental setup, each model was trained and tested
using a stratified 3-fold cross validation. Stratification was consid-
ered as the portion of penalty versus non-penalty case briefs was
highly imbalanced with a ratio of 73:281. The averaged performance
across 3 folds is then computed for each model as well as for each
language. In model performance evaluation, we considered com-
mon metrics including: specificity, recall, the F1 score, and the Area
Under the Receiver Operating Characteristic (AUROC). We also in-
cluded the Matthews correlation coefficient (MCC), which provides
a correlation coefficient between the observed and predicted binary
classifications. MCC is robust and therefore particularly useful in
scenarios where classes exhibit a strong imbalance [18].
4.3 Interpretability of classification
The interpretability of the resultant machine learning models is
a challenging but worthwhile endeavour [11]. As such we utilise
the LIME methodology proposed by Ribeiro et al [21]. LIME allows
for the identification of areas which provided the most influence
to a classification, which is an attractive feature for legal decision
making. With the identification of predictive features, we qualita-
tively assess whether the identified features for correctly predicted
positive cases are shared across the three languages. From the com-
parison of classification models, we select the most performant
model across all three languages by assessing the classifier with
the best value in by each metric for each language, and then check
for consensus among metrics. We re-run a fold from the model in
order to examine the quality of predictions across the entire data
set. We utilise the predictions of the entire dataset to (i) compare
the amount of true positives identified in each language and (ii)
to determine whether the semantic meaning of the determinative
187
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
features of any particular prediction (i.e. which words were most
influential) are shared between the languages. We draw an illustra-
tive case example and compare the top ten features, assessing the
extent to which these translate into each language.
5 RESULTS
5.1 Performance of classification models
Table 2 shows the mean and standard deviation of the performance
metrics on the test data sets derived from the stratified 3-fold cross
validation. Given structural differences between languages, it is
expected that different models will be more effective for one lan-
guage over another. For instance, SVM performs mostly well across
all metrics for Portuguese (AUROC of 0.748 and F1 of 0.605) and
Chinese (AUROC of 0.725 and F1 of 0.577). For English, MNB was
considered to be a better option (AUROC of 0.774 and F1 of 0.649).
However, it is difficult to conclude if the results are significant due
to the limited number of case briefs available. It was further ex-
pected that the specificity would be high across different models
(a range of 0.875 to 0.981) due to the number of non-penalty briefs
available in comparison to those with penalty. In the case of recall,
SVM performed the best mostly across three languages (0.576 for
Chinese, 0.630 for English, and 0.602 for Portuguese), where other
models have variable results. The MCC values were also similar
across three languages with MNB providing the highest values
ranging from 0.48 to 0.534. We determine that the SVM performs
the best scores across all languages when considering all metrics.
5.2 Comparison of predictions across languages
Out of the 57 cases tested, the performance of the languages in-
significantly varied: English, 49 (86%), Portuguese, 50 (88%), and
Chinese, 50 (88%). The classification of No Penalty is high but is due
to the imbalanced class. Complaints, which make up the majority
of reasons initiating a case investigation, receive the most penalties
when compared to other reasons. Across all languages, true predic-
tions of penalties for complaints were high, with the most being
identified by English (75%). The largest disparity in scores exists in
Active Investigation; only the Chinese language classifier was able
to positively identify one of the two penalties.
5.3 Assessment of interpretability across
languages
From the above results in Section 5.2, there is some evidence for the
feasibility of utilising machine learning for penalty prediction. One
of the questions posed by this paper was whether these predictions
converge on semantic meaning with regard to the most discrimina-
tive features used to make these predictions. We make a qualitative
check regarding the features by selecting a case which was correctly
identified by the classifier in all languages. For this purpose we
select Case No 0002/2014/IP [9], which concerned the uploading
of a photo onto a social networking site without the consent of
the complainant. The complainant had family photographs profes-
sionally taken by a company. Without the knowledge or consent of
the complainant, the company uploaded the photo to social media
as promotional material. The complainant asked the company to
remove the photo but did not receive any response. The complaint
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Aaron Ceross and Tingting Zhu

Table 2: Test performance of classification models across three languages.

Language Model F1 AUROC Recall Specificity MCC

SVM 0.577 (± 0.107) 0.725 (± 0.075) 0.576 (± 0.204) 0.875 (± 0.056) 0.463 (± 0.1)
Chinese RF 0.397 (± 0.136) 0.627 (± 0.053) 0.274 (± 0.119) 0.981 (± 0.013) 0.398 (± 0.081)
MNB 0.555 (± 0.136) 0.711 (± 0.081) 0.493 (± 0.204) 0.928 (± 0.042) 0.48 (± 0.1)

SVM 0.620 (± 0.126) 0.762 (± 0.095) 0.630 (± 0.272) 0.894 (± 0.083) 0.564 (± 0.081)
English RF 0.480 (± 0.122) 0.666 (± 0.067) 0.385 (± 0.17) 0.947 (± 0.036) 0.423 (± 0.084)
MNB 0.649 (± 0.13) 0.774 (± 0.098) 0.631 (± 0.256) 0.918 (± 0.065) 0.588 (± 0.116)

SVM 0.605 (± 0.142) 0.748 (± 0.095) 0.602 (± 0.239) 0.894 (± 0.048) 0.514 (± 0.126)
Portuguese RF 0.561 (± 0.147) 0.709 (± 0.084) 0.453 (± 0.19) 0.966 (± 0.024) 0.528 (± 0.108)
MNB 0.602 (± 0.163) 0.747 (± 0.102) 0.575 (± 0.255) 0.918 (± 0.56) 0.534 (± 0.135)
Note: the mean and standard deviation are computed on the test sets from the 3-fold stratified cross validation.

photo foto
several tirar
networking mensagem
publish rede
message companhia
account exigir
specify social
take publicar
social atender
afterwards acto
0.1 0.0 0.1 0.2 0.3 0.4 0.5 0.1 0.0 0.1 0.2 0.3
Coef Coef

(a) Chinese (b) English (c) Portuguese

Figure 1: Influence of word features on the correct classification of Case No 0002/2014/IP.

was submitted to the GPDP who found that the situation warranted
a penalty.
In Figure 1 the most predictive features are shown for the out-
come of Case No 0002/2014/IP [9]. The top feature for English and
Portuguese is ‘photo’ and ‘foto’ respectively. In this instance we
have semantic convergence as these mean the same thing in both
languages. However, Chinese it is ‘投訴’, which translates to ‘to
complain’. In this example, most of the terms are shared between
English and Portuguese, e.g., ‘publish’ / ‘publicar’, ‘message’ / ‘men-
sagem’. Portuguese and Chinese share the word for ‘company’:
companhia / ‘公司’. It is noted that many of the Chinese terms do
not provide much semantic meaning. For example, the ‘甲’ and ‘乙’
characters are related to the ordering of items and in legal practice
these characters are often used to anonymise names in a text (e.g.
Person A/甲 and Person B/乙). The character ‘均’ is used to qualify
other nouns and thus may have multiple meanings depending on
context such as ‘all’ or ‘even’; by itself, it does not provide any
meaning. In short, while there may be some semantic meaning
derived from the prediction (e.g., ‘photo’, ’foto’, ‘公司’), many of
the features have limited semanitic utility such as ‘take’, ‘均’, and
‘acto’. This may have implications for the provision of explanations
in support of a prediction when used in a legal setting.
6 DISCUSSION
This work has demonstrated the effectiveness of text classification
of data protection cases. The binary classification of penalties has
been shown to be effective across multiple language translations of
the same facts. Despite the small dataset, this paper’s experiment
has, at the very least, indicated the strong possibility that machine
learning may be included in some manner to facilitate case prioriti-
sation for regulatory action. Data protection authorities may utilise
this method in order to pre-screen complaints thus more effectively
prioritise the use of the authority’s resources.
While an argument may be made to use text classification for pri-
oritising cases, there are considerations whether such an undertak-
ing impinges on the nature of regulation and the law. Bayamlıoğlu
and Leenes [2] regard the use of the technology in legal decisions
and data-driven law as degrading the “moral enterprise” of the

188
Prediction of monetary penalties for data protection cases in multiple languages
law, which has an impact on the human trust and value placed not
only with the law itself but by extension those institutions charged
with its execution. This is echoed by Hildebrandt [13] who argues
that data-driven law creates a crisis for law in that data-driven
predictions may result in atrophy of the ability to make judgements
congruent with the lived experience of individuals; data-driven law
is beholden to a type of logic that may not lend easily itself to the
rule of law.
Explanability of machine learning models is often suggested as
a counter-balance to what may be perceived as the negative effects
of automated decisions. Edwards and Veale [8] maintain that an
explanation of the logic of a model and its outcomes may have
no meaning for an affected individual, giving little recourse by
which to challenge the outcome of the automated decision. There
is also the question as to whether such explanations are helpful or
informative to individuals as experiments on human perception
of algorithmic decisions and found that even where explanations
are provided to algorithmic decision-making, such explanations
may not contribute to a sense of ‘fairness’ in the decision made.
Binns et al. [3] investigated the human perception of algorithmic
decisions and found that even where explanations are provided to
algorithmic decision-making, such explanations may not contribute
to a sense of ‘fairness’ in the decision made.
From the results, we would strongly caution against adopting
the methods outlined in the experiment as the sole determination
of a penalty allocation. The results from our experiment shows
that the model are able to indicate which cases may be of more
interest to a regulatory authority than others. It is, however, ques-
tionable whether the description of factors amounts to a sufficient
explanation for legal purposes. This is due to the models utilising
text processed in a bag-of-words approach which accounts for the
frequency of words, not the semantic meaning of those words. As
such, many predictive features, as those detailed in Section 5.3, may
be nonsensical when shown in isolation.
7 CONCLUSIONS AND FUTURE WORK
In this work, we introduced a novel dataset, cases from the data
protection authority of Macao, and evaluated multiple machine
learning classifiers for binary text classification for text in three
language versions. Our results show that MNB and SVM performed
well across all metrics for all languages with SVM being considered
the most performant. Assessing interpretability was difficult given
the bag-of-words model used in the text preprocessing, although
there is some overlap between the semantic meaning of features
between the languages. The models evaluated are not without their
limitations and the size of training datasets remains a challenge. In
future work, we aim to employ language models such as BERT [7]
to experiment with test different language tasks in data protection
regulation. Another avenue of future work may include assessing
the utility and interpretability of the outputs of case prediction with
case investigators themselves.
ACKNOWLEDGEMENTS
The authors thank the anonymous reviewers for their feedback as
well as Tasos Papastylianou and Andrew Simpson for comments
on earlier drafts of this work. Special thanks to Ethan Ceross.
189
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
REFERENCES
[1] Aletras, N., Tsarapatsanis, D., Preoţiuc-Pietro, D., and Lampos, V. Pre-
dicting judicial decisions of the European Court of Human Rights: A natural
language processing perspective. PeerJ Computer Science 2 (2016), e93.
[2] Bayamlioğlu, E., and Leenes, R. The ‘rule of law’ implications of data-driven
decision-making: a techno-regulatory perspective. Law, Innovation and Technol-
ogy 10, 2 (2018), 295–313.
[3] Binns, R., Van Kleek, M., Veale, M., Lyngs, U., Zhao, J., and Shadbolt, N.
’It’s reducing a human being to a percentage’: Perceptions of justice in algorith-
mic decisions. In Proceedings of the 2018 CHI Conference on Human Factors in
Computing Systems, CHI ’18, p. 1–14.
[4] Braithwaite, J. Enforced self-regulation: A new strategy for corporate crime
control. Michigan Law Review 80, 7 (1982), 1466–1507.
[5] Ceross, A. Examining data protection enforcement actions through qualitative
interviews and data exploration. International Review of Law, Computers &
Technology 32, 1 (2018), 99–117.
[6] Ceross, A., and Simpson, A. C. The use of data protection regulatory actions as
a data source for privacy economics. In Computer Safety, Reliability, and Security
(SAFECOMP) (2017), S. Tonetta, E. Schoitsch, and F. Bitsch, Eds., vol. 10489 of
Lecture Notes in Computer Science (LNCS), Springer, pp. 350–360.
[7] Devlin, J., Chang, M.-W., Lee, K., and Toutanova, K. BERT: Pre-training of deep
bidirectional transformers for language understanding. In Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers)
(Minneapolis, Minnesota, June 2019), Association for Computational Linguistics,
pp. 4171–4186.
[8] Edwards, L., and Veale, M. Enslaving the algorithm: From a “Right to an
Explanation” to a “Right to Better Decisions”? IEEE Security & Privacy 16, 3
(2018), 46–54.
[9] Gabinete para a Protecção de Dados Pessoais. Case No: 0002/2014/IP:
Uploaded clients’ photos by mistake. https://www.gpdp.gov.mo/index.php?m=
content&c=index&a=show&catid=209&id=775, 2014. English version.
[10] Gabinete para a Protecção de Dados Pessoais. 個案調查. https://www.
gpdp.gov.mo/uploadfile/2020/1009/20201009040422173.pdf, Oct. 2020.
[11] Gilpin, L. H., Bau, D., Yuan, B. Z., Bajwa, A., Specter, M., and Kagal, L.
Explaining explanations: An overview of interpretability of machine learning.
In 2018 IEEE 5th International Conference on data science and advanced analytics
(DSAA) (2018), IEEE, pp. 80–89.
[12] Greenleaf, G. Macao’s EU-influenced Personal Data Protection Act. Privacy
Laws & Business International Newsletter 96 (2008), 21–22.
[13] Hildebrandt, M. Law as computation in the era of artificial legal intelligence:
Speaking law to the power of statistics. University of Toronto Law Journal 68,
Supplement 1 (2018), 12–35.
[14] Hustinx, P. The role of data protection authorities. In Reinventing Data Pro-
tection?, S. Gutwirth, Y. Poullet, P. De Hert, C. de Terwange, and S. Nouwt, Eds.
Springer, 2009, pp. 131–137.
[15] Jianqiang, Z., and Xiaolin, G. Comparison research on text pre-processing
methods on Twitter sentiment analysis. IEEE Access 5 (2017), 2870–2879.
[16] Medvedeva, M., Vols, M., and Wieling, M. Using machine learning to predict
decisions of the European Court of Human Rights. Artificial Intelligence and Law
28, 2 (2020), 237–266.
[17] Personal Data Protection Act. https://www.gpdp.gov.mo/uploadfile/2016/0302/
20160302033801814.pdf, 2005.
[18] Powers, D. M. Evaluation: From precision, recall and F-measure to ROC, in-
formedness, markedness and correlation. arXiv preprint arXiv:2010.16061 (2020).
[19] Qi, P., Zhang, Y., Zhang, Y., Bolton, J., and Manning, C. D. Stanza: A Python
natural language processing toolkit for many human languages. In Proceedings of
the 58th Annual Meeting of the Association for Computational Linguistics: System
Demonstrations (2020).
[20] Regulation on the protection of natural persons with regard to the processing of
personal data and on the free movement of such data, and repealing Directive
95/46/EC (General Data Protection Regulation). L119, 4/5/2016, p. 1–88, 2016.
[21] Ribeiro, M. T., Singh, S., and Guestrin, C. "Why should I trust you?": Explain-
ing the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD
International Conference on Knowledge Discovery and Data Mining, San Francisco,
CA, USA, August 13-17, 2016 (2016), pp. 1135–1144.
[22] Virtucio, M. B. L., Aborot, J. A., Abonita, J. K. C., Avinante, R. S., Copino,
R. J. B., Neverida, M. P., Osiana, V. O., Peramo, E. C., Syjuco, J. G., and Tan,
G. B. A. Predicting decisions of the Philippine Supreme Court using natural
language processing and machine learning. In 2018 IEEE 42nd Annual Computer
Software and Applications Conference (COMPSAC) (2018), vol. 2, IEEE, pp. 130–135.
[23] Waltl, B., Bonczek, G., Scepankova, E., Landthaler, J., and Matthes, F.
Predicting the outcome of appeal decisions in Germany’s tax law. In International
Conference on Electronic Participation (2017), Springer, pp. 89–99.
 Regulating Artificial Intelligence:
A Technology Regulator’s Perspective
Joshua Ellul Gordon Pace Stephen McCarthy
joshua.ellul@um.edu.mt gordon.pace@um.edu.mt stephen.mccarthy@mdia.gov.mt
Malta Digital Innovation Authority Department of Computer Science, Malta Digital Innovation Authority
& University of Malta University of Malta Malta
Malta Malta

Trevor Sammut Juanita Brockdorff Matthew Scerri

trevor.sammut@mdia.gov.mt JuanitaBrockdorff@kpmg.com.mt
Malta Digital Innovation Authority KPMG
Malta Malta
ABSTRACT
Artificial Intelligence (AI) and the regulation thereof is a topic that
is increasingly being discussed and various proposals have been
made in literature for defining regulatory bodies and/or related
regulation. In this paper, we present a pragmatic approach for
providing a technology assurance regulatory framework. To the
best of our knowledge, this work presents the first national AI
technology assurance legal and regulatory framework that has
been implemented by a national authority empowered through law
to do so. Aiming to both provide assurances where required and
not stifling innovation yet supporting it, it is proposed that such
regulation is not to be mandated for all AI-based systems but rather
should provide a voluntary framework and only be mandated in
sectors and activities as deemed necessary by other authorities or
laws for regulated and critical areas.
CCS CONCEPTS
• Social and professional topics → Governmental regulations.
KEYWORDS
artificial intelligence, regulation
ACM Reference Format:
Joshua Ellul, Gordon Pace, Stephen McCarthy, Trevor Sammut, Juanita
Brockdorff, and Matthew Scerri. 2021. Regulating Artificial Intelligence: A
Technology Regulator’s Perspective. In Eighteenth International Conference
for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo,
Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.
3466093
1 INTRODUCTION
Issues concerning the design of legal and regulatory frameworks
for Artificial Intelligence (AI) have been a topic of discussion and
debate for the past few decades. Much of the debate inherits from
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466093
MatthewScerri@kpmg.com.mt
KPMG
Malta
discussions on regarding how to regulate technology and the regu-
lation of computer systems, but reaches further due to the very na-
ture of the potential for AI. In fact, one can argue that a substantial
portion of the debate is due to this very potential, which brings to-
gether ethical issues, rights, perils and other aspects. Regardless of
which school of thought one subscribes to, in a spectrum ranging
from the requirement of generic principles [1], to specific laws1 , to
advocating that regulation of such technology should be avoided,
and focus should be on safety mechanisms [16], when the technol-
ogy is used for applications that can directly or indirectly impact so-
ciety then sufficient regulation (whether through law or otherwise)
should be investigated (whether applied directly or indirectly to the
technology). At the same time whilst some argue for mandatory
regulation, many warn that regulation could stifle innovation [11].
In this paper, we do not purport to present a contribution to
this philosophical debate, but rather our aim is a more pragmatic
one — that of outlining and explaining the rationale behind a legal
and regulatory framework addressing AI systems adopted by Malta.
Whilst other regulatory frameworks and bodies have been proposed
in literature (discussed in Section 5), it is to the best knowledge
of the authors that the framework being presented herein is the
first AI technology assurance legal and regulatory framework that
has been implemented by a national authority (the Malta Digital
Innovation Authority2 ) empowered through law [10] to do so.
Towards the end of the 2010’s Malta built a framework for ad-
dressing the regulation of Innovative Technology Arrangements
[9], in order to ensure better end user protection through the adop-
tion of appropriate due diligence on the underlying technologies.
Initially focusing on Blockchain, Smart Contracts and other Dis-
tributed Ledger Technologies (DLTs) [6], the legislation has since
been extended to cover critical systems (through a legal notice3 )
and regulatory guidelines have been issued by the Malta Digital
Innovation Authority (MDIA) for the regulation of arrangements
which use an element of AI.
The aim of the paper is to provide a review of the regulatory
framework proposed and to put it in the context of the ongoing AI
regulation debate. One of the primary observations is that the need
1 https://www.europarl.europa.eu/doceo/document/E-9-2019-002411_EN.html
2 https://mdia.gov.mt
3 https://legislation.mt/eli/ln/2020/389/eng/pdf

190
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Joshua Ellul, Gordon Pace, Stephen McCarthy, Trevor Sammut, Juanita Brockdorff, and Matthew Scerri
for precise definitions and objective measures in the legal frame-
work, meant that the Maltese regulatory approach is founded on
practical and auditable aspects, and is intended to address concerns
with existing technology (as opposed to attempting to address pos-
sible issues arising from future development of AI technology, for
example Artificial General Intelligence). To implement an AI reg-
ulatory framework intended for modern day technology and also
in aim of not stifling innovation, the framework is primarily vol-
untary however may be mandated based upon the sector and/or
risk associated with the activity within which the AI system is used
or as deemed necessary by another lead authority or governing
legislation. This sets the tone of much of the paper, but it is nat-
urally endemic to any discussion of practical implementations of
the regulation of technologies. The fast evolving nature of technol-
ogy requires law-makers to address existing technology in a sound
manner, but also in a way that is expected to be future-proof. A full
version of this paper can be found in [7].
2 THE CASE FOR AI ASSURANCES
We start by highlighting issues related to AI-based systems which
could result in systems operating incorrectly in relation to its in-
tended functionality and thereafter build the case for instilling as-
surances. We concentrate on Artificial Narrow Intelligence (ANI)
given that the state-of-the-art has not yet reached levels of Artifi-
cial General Intelligence (AGI) [3]. We will use the term AI through-
out the rest of the paper to refer to AI that exists today — ANI.
Since the inception of software development, the fact that such
systems occasionally fail has been accepted to be the norm. Al-
though much work has gone into developing techniques to reduce
the frequency and severity of such occurrences, we continue to ex-
perience software malfunction on a daily basis. The impact of such
failure is contained as long as the software functions in a closed
system i.e. it has no direct impact on the real world, but frequently
software affects the real world in a direct or indirect manner. One
finds reports of many catastrophic failures in literature and news-
paper reports, with effects ranging from huge financial losses to
critical infrastructure failure and even loss of human life. AI sys-
tems are no exception when it comes to incorrect behaviour and
even when the algorithms themselves are correctly implemented,
incorrect behaviour might emerge. For instance, a correct imple-
mentation of a machine learning based algorithm may still learn
wrong due behaviour due to insufficient training, biases and unbal-
ances that may exist within datasets, etc.
Undoubtedly AI systems should undergo standard quality as-
surances processes, not only for functional correctness of the algo-
rithms themselves but also with respect to the behaviour emergent
following training. However, testing of AI systems is only as good
as the coverage of training data, iterations and permutations and
use cases which are undertaken. Once an AI system is deployed
and it encounters an event that it was not trained to handle it may
well end up handling it incorrectly. More so, if it is continuously
learning in a live environment it may be exposed to certain situa-
tions which could affect its behaviour negatively.
Part of the challenge is that many AI-based techniques function
as black-boxesfor which reason one finds extensive research to-
wards explainable AI. The past decade has seen various infamous
191
cases where unexpected behaviour emerged from AI systems, some-
times of a controversial or even safety-critical nature. The increas-
ing concern is not only to do with cases that have emerged but also
based on the reality that more and more systems are becoming com-
puterised and automated. One often-referenced cliché is that of au-
tomated scoring systems [13] in which discrimination is unaccept-
able, highlighting the need to ensure bias in datasets is removed and
attempts made to remove discriminatory features during training.
The concerns highlighted above demonstrate the need to ensure
that sufficient assurances are put in place to ensure that AI algo-
rithms are implemented correctly, and that their behaviour is as ex-
pected and does not introduce any unwanted biases. Indeed, many
are advocating for such regulatory frameworks to be developed
and applied to AI systems, but it whether such frameworks should
be mandatory for all AI-based systems is debatable. We now follow
with a case for why such frameworks should not always be man-
dated, and that they should not be focused on the technology but
should be focused on the sector or activity that the AI is being used
within/for.
3 THE CASE FOR VOLUNTARY ASSURANCES
OF AI AND MANDATORY ASSURANCES OF
REGULATED AND CRITICAL ACTIVITIES
Setting aside AGI, when it comes to ANI should such frameworks
always be mandated? The same AI framework, for instance identi-
fying user preferences, can be the engine behind a wide range of
applications, from a personal movie recommendation system, to a
social network targeted advertising campaign to influence users in
an upcoming election. The underlying infrastructure is application
agnostic, but should such an underlying infrastructure be required
to be regulated? More so, what difference does it make if an algo-
rithm is AI-based or not and yet can be used for the same activity?
Then, should we be talking about AI regulation at all? Or should
we be focusing on software — or rather, the activity it is used for
irrespective of how it is implemented?
Regulating all forms of AI would result in shackling and stifling
innovation [11]. The definition of AI itself is controversial, and
even if a definition is chosen, is it going to be clear what software
is AI and what software is not? There are some algorithms which
we can ascertain are universally accepted as AI, and some systems
which are universally considered to not have aspects of AI within
them, however what should be done about the rest? Could this
approach not only stifle AI-innovation, but also other software
based innovation?
Looking back at the principles of regulation though, we need to
ask ourselves why is regulation of AI being proposed? Is it only be-
cause of end-of-the-world scenarios being painted which require
AGI, which the state-of-the-art is currently not capable of? If so,
then perhaps we should differentiate between any regulatory re-
quirements for AGI and ANI. We propose that this should be done,
at least in the interim until AGI is deemed to be upon us. We leave
considerations for AGI as future work, and here will continue dis-
cussing aspects pertaining to ANI.
If AI is regulated even when applied to unregulated and non-
critical activities, given the line between AI and software in general
is blurred, and given that non-AI based techniques processes may
Regulating Artificial Intelligence: A Technology Regulator’s Perspective
yield the same sort of undesirable outcomes, then why should not
all software be regulated? We propose that mandatory regulation
should be sector/activity-based and not on technology.
The question of what constitutes high-risk or defines whether
a sector or activity should mandate this framework arises. This
is to be left up to other lead authorities and laws of the land to
decide. For financial affairs, a financial services authority (a separate
body) may impose when a sector or activity should be mandated
to undertake a technology audit (as proposed herein), or even if
any levels of enhanced due diligence is required. Therefore, based
on the above we make the argument that mandatory regulatory
frameworks should not be technology-specific (or AI-specific), yet
should be activity or sector-specific as defined and required per
activity/sector.
AI technology-based assurances may not only be required for
regulated activities, however various AI-based products and ser-
vices may see benefit in providing assurances to various stakehold-
ers. Therefore, the regulatory approach enables for technology-
based assurances to also be offered on a voluntary basis (besides be-
ing mandated from lead authorities of respective sectors/activities).
Now, we present the AI technology assurance framework imple-
mented by the Malta Digital Innovation Authority4 which offers
certification of AI systems on a voluntary basis where sought, or on
a mandatory basis where other lead authorities or laws require it.
4 AN AI TECHNOLOGY ASSURANCE
REGULATORY FRAMEWORK
We now present the AI Innovative Technology Arrangement (AI-
ITA) regulatory technology assurance framework. Approaches for
providing software assurances will invariably have a degree of com-
monality irrespective of the technology domain and also applica-
tion domain within which the solution is categorised under. As
such, this framework builds on the Innovative Technology Arrange-
ment (ITA) [6] regulatory assurance framework overseen by the
Malta Digital Innovation Authority (MDIA). Rather than mandating
compliance and certification of all AI based systems, the regulatory
framework is a voluntary one — unless a lead authority deems that
such technology assurances are required. It is in this manner we be-
lieve innovation can still flourish, by only requiring mandatory over-
sight of sectors and activities that should require such oversight.
AI Innovative Technology Arrangement. The challenge with
Artificial Intelligence ITAs (AI-ITAs), primarily revolves around
identifying what constitutes AI. Rather than define what is an AI-
ITA as a hard and fast rule, the guidelines take the approach of
defining qualities and criteria that qualify software as an AI-ITA:
(a) the ability to use knowledge acquired in a flexible manner in
order to perform specific tasks and/or reach specific goals; (b) evo-
lution, adaptation and/or production of results based on interpret-
ing and processing data; (c) systems logic based on the process
of knowledge acquisition, learning, reasoning, problem solving,
and/or planning; (d) prediction forecast and/or approximation of
results for inputs that were not previously encountered.
The above ensures that techniques and algorithms commonly as-
sociated with the wider AI field are captured and include anything
4 https://mdia.gov.mt
192
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
from Deep Learning to Natural Language Processing and Optimisa-
tion Algorithms. The MDIA will also continue to monitor develop-
ments and update guidelines as required to include (and potentially
exclude) defining features of what is/not classifed as an AI-ITA.
System Audits and Subject Matter Experts. The framework pro-
vides a structure for the Authority and applicant to work with inde-
pendent (and approved) system auditors to be able to scrutinise to
a fairly high level of detail the software itself as well as the manner
with which it is being operated under the ISAE 3000 [12] standard
for assurance. The audit of the software system itself is primarily
conducted via a code review, whose aim is to ensure that the man-
ner with which the AI-ITA is implemented accurately reflects what
the organisation behind the AI-ITA are claiming in their technology
blueprints. The rationale behind this is to ensure that any claims be-
ing done are truly reflected in the code, which enables the general
public, who may not know what AI really is to gain trust in the sys-
tem given that it stood up to scrutiny prior to the certificate being is-
sued. Beyond the software, the certification mandates depending on
the type of audit being undertaken and associated controls, to also
give the general public assurances that the AI-ITA creator and oper-
ator are running the organisation in a manner that meets the stan-
dards set out by the MDIA. The certification therefore enables the
general public to trust the creator, in the manner they build, main-
tain and run the AI system. Two main types of audits are required
throughout an AI-ITA’s lifetime: (i) first a ‘Type 1 Systems Audit’
is required which focuses on providing assurances with respect to
functional correctness typically undertaken as an AI-ITA’s first au-
dit; and (ii) a ‘Type 2 Systems Audit’ which focuses on renewing pre-
vious assurances provided through a previous audit which factors
in live data and operations associated with the system to assure the
system assurances are still in place within the period under audit.
The audit process begins with the applicant submitting a request
(in the form of an application) to the Authority, upon which the
Authority will assess the applicant by reviewing the provided doc-
umentation around the AI-ITA and conduct its due diligence. Fol-
lowing this, the MDIA issues a Letter of Intent upon which the ap-
plicant will be able to appoint an MDIA approved Systems Auditor,
and notify the MDIA of the appointment, for the MDIA to verify
that the Systems Auditor has the required competencies (which the
Authority has tested the system auditor for). The Systems Auditor
will then conduct the audit as per the Authority’s guidelines5 and
compile a report with their findings, which is issued to the MDIA
for a review and a subsequent decision on whether the certificate
is to be issued. Once issued, a further follow up audit must be con-
ducted every time there is a material change in the AI-ITA (and on
renewal after every two years).
Systems Audits are an integral part of the certification process as
they provide the MDIA with an independent report on the particu-
larities of the AI-ITA, specifically the code (and data) and whether it
accurately reflects what is being disclosed in the blueprint, and the
ongoing operations of the AI-ITA. Systems Audits are conducted by
Systems Auditors, who must be independent from the AI-ITA and
its operator, that are subject to approval by the Authority, and who
need to meet a set of requirements (defined in the Systems Auditor
guidelines) through their combined complement of Subject Matter
5 https://mdia.gov.mt/wp-content/uploads/2019/10/AI-ITA-Guidelines-03OCT19.pdf
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Joshua Ellul, Gordon Pace, Stephen McCarthy, Trevor Sammut, Juanita Brockdorff, and Matthew Scerri
Experts (SMEs) in the fields of IT audit, cybersecurity and technol-
ogy with specialisation, in this case, in AI. The SMEs will be the
primary individuals responsible for conducting systems audits, and
must adhere to a set of requirements, such as ensuring that they
meet a level of continuous professional education in the AI field6 .
This section describes the requirements that an AI-ITA must
meet in order to qualify for certification.
ITA Blueprint. The Blueprint document is an essential document
in the certification process as it is meant to provide a detailed
description to the Authority on what the system does, how it is
designed, and operated. Other than allowing the MDIA to evaluate
whether AI-ITA certification is applicable, it is further intended to
be used by the Systems Auditors as the document against which
aspects such as the code is reviewed against. The blueprint also
defines a minimum set of disclosures that must be disclosed to
direct users (in English) in a non-technical manner, to be able to
communicate the features and functionalities of the system and
how it respects the ethical AI framework7 , limitations to prevention
of bias, and the expected accuracy of the AI-ITA.
In a general (AI agnostic) sense, the detailed description must
cover the functional capabilities of the AI-ITA, how the system is to
be verified and tested to ensure the results meet expectations and
what the operational limitations of the systems are. More specif-
ically, for an AI solution the blueprint must include a disclosure
of the AI techniques used and to justify why certification is being
sought, and how specific risks are being managed and mitigated e.g.
what is being done to ensure that the underlying dataset is unbiased.
In a broader sense, the Blueprint must highlight the safety mecha-
nisms in place and alignment with Malta’s Ethical AI Framework.
ITA Harness. A crucial element that the AI-ITA framework pro-
poses, and which needs to be highlighted clearly in the Blueprint is
the ITA Harness. The ITA Harness provides a safety net for the pro-
cess by monitoring activity inputs and outputs to ensure that the
boundaries (which must also be disclosed in the Blueprint) are re-
spected. Furthermore, the ITA harness must also be able to handle
any anomalies it detects (such as outputs outside expected bound-
aries) in a manner which is also disclosed. The AI-ITA harness must
also communicate with the Forensic Node (discussed next) to ensure
that any anomalies are appropriately logged and can be investigated
and rectified. While the harness may not apply to all AI-ITAs, the
Authority requires that when it does not apply it must be justified ad-
equately in the blueprint and accompanied with alternative plans of
how the behaviour of the AI-ITA will be monitored and contained8 .
Forensic Node. The Forensic Node is another requirement man-
dated by the MDIA, and whose implementation and operation is
also subject to the audit. The purpose of the Forensic Node is to
“store all relevant information on the runtime behaviour of the AI-ITA
in real-time such as recording of inputs and outputs, and supporting
data related to potential explainability of how an output was derived
from a given input wherever applicable”. This means that any inputs,
outputs as well as data that supports how the system achieved the
results it did must be stored in a secure data store in real-time. This
6 https://mdia.gov.mt/sa-guidelines/
7 https://malta.ai/wp-content/uploads/2019/08/Malta_Towards_Ethical_and_
Trustworthy_AI.pdf
8 https://mdia.gov.mt/wp-content/uploads/2019/10/AI-ITA-Blueprint-Guidelines-
03OCT19.pdf
193
highlights that the Forensic Node is not only used to support the
assessment of (some of) the operating effectiveness of the controls
during an audit, but may also be used to support legal compliance
by the MDIA (or other authorities) and also enables a further layer
of monitoring to be done (manually or automated) by a Techni-
cal Administrator (discussed next). It is important to note that the
Forensic Node must be separate from the ITA Harness, in that the
Forensic Node is more concerned with Data Logging, as opposed
to the monitoring in relation to boundaries.
Technical Administrator. A Technical Administrator, a form of a
service provider appointed by the AI-ITA to act as the final safe-
guard for the system, must be appointed and in place at all times.
The Technical Administrator must be able to intervene, if required
to do so by the MDIA, another authority or legally (such as in the
event of a breach of law by the AI-ITA), to limit further impact to
the users and where necessary limit or reverse losses. For example,
consider an AI system that utilises reinforced learning and which,
after a period of time, starts to exhibit discriminatory bias that goes
against the principles laid down in the ethical AI framework and/or
against the requirements of any laws or rules it must abide by. In
this case the Technical Administrator must be able to halt the op-
eration of the system to prevent further damage and revert to an
older model (as may be mandated by a legal judgement). As such,
this also imposes an indirect requirement for the AI-ITA to provide
mechanisms to enable the Technical Administrator to conduct their
actions as may be necessary (e.g. by ensuring regular snapshots of
the machine learning models are kept to revert back to).
English Description and Consumer Protection. The system be-
ing certified is checked by the systems auditors who, amongst other
things, ensure that its functionality matches that described in the
blueprint in human-readable form (in English). If, post-deployment,
the system exhibits behaviour contrary to this description against
which it was certified, the Innovative Technology Arrangements
and Services Act specifies that the English version prevails legally.
Auditing of Design and Development Processes. Systems Au-
dits include oversight of the design and development process of
the system-under-audit. Not only does such oversight cover tradi-
tional software engineering principles, but for systems including
an element of AI also includes assurances that certain foundational
principles have been taken into consideration in the process.
Build on a human-centric approach. The systems auditors ensure
that the AI system was designed in a manner to support and assist
humans without overriding the user into taking any unwanted
decisions and the manner with which it operates musts be equitable
and inclusive across different segments of society.
Adherence to applicable laws and regulations. It is crucial that be-
haviour induced by the system, including parts driven by AI, will
be designed in a manner that adheres to the law.
Maximise benefits of AI systems while preventing and minimising
their risks. It is crucial that any risks induced through the use of AI
are identified and mitigated accordingly, including the setting up
of controls to ensure fairness, transparency and resiliency to new
AI-specific attack vectors.
Aligned with emerging international standards and norms around AI
ethics. As the world is increasingly becoming globalised through
technology, and which may be further amplified through the prolif-
eration of AI systems, this objective was laid down to ensure that
Regulating Artificial Intelligence: A Technology Regulator’s Perspective
Malta’s ethical framework is aligned with similar ethical guidelines
by the EU commission9 and OECD10 .
The framework further builds on these principles by delineat-
ing a number of principles (such as Human Autonomy, Fairness,
Prevention of Harm and Explicability) and proposes 63 controls of
how these can be tested. While not all of these controls apply to all
AI-ITAs, the AI-ITA must show that it has taken them into consid-
eration and justify in the Blueprint (and ultimately top the users)
those controls which do not apply.
5 RELATED WORK
The work presented herein is complementary and orthogonal to a
number of different areas which we will now provide an overview.
Regulatory Bodies. The European Parliament had proposed for
the setting up of an European Agency for robotics and artificial
intelligence address technical, ethical and regulatory aspects [5],
mostly driven from the need for transparency of automated sys-
tems handling personal data and the often impossibility of doing so
due to trade secret protection [15]. A solution proposed was to al-
low for a trusted third party to undertake an audit of the system in
question. On similar lines, to avoid differing domestic approaches,
the need for an International Artificial Intelligence Organisation
was highlighted [8]. Indeed, this would be a step in the right di-
rection, however it is the opinion of the authors that the need for
providing regulation should not wait for such an organisation to
emerge, yet national authorities (such as the MDIA) could work to-
gether towards harmonisation and adapt to eventual international
standards and guidance as it emerges.
International Standards. Whilst global software regulatory bod-
ies do not (yet) exist, global standards do. The International Organi-
sation for Standardization (ISO) has developed a number of different
standards for use within the software domain11 . Whilst, such soft-
ware focused standards can be useful for global recognition within
the framework described herein local national standards were re-
quired to be developed for the following reasons: (i) standards avail-
able to date do not provide guidelines or comprehensive control
objectives specific for the artificial intelligence domain; (ii) mecha-
nisms and roles for ensuring continuous monitoring and interven-
tion are not defined [2, 4];and (iii) the authority is ultimately re-
sponsible and empowered through the MDIA Act to ensure audit in-
tegrity and quality whilst at the same time able to propose changes
to legislation and guidelines. Once international standards adequate
to adopt are developed national guidelines may be updated to make
use of them (if deemed to meet the national requirements). That
said, the authority has adopted and requires that audits are under-
taken following the ISAE 3000 [12] standard which specifies generic
(i.e. not software nor AI related) principles for quality management,
ethical behaviour and performance for use in non-financial areas.
Other Non-technology Assurance Related Aspects. There is a
large body of work looking at regulating the application of tech-
nology which is orthogonal to the to approach presented in this pa-
per, including how to handle issues of liability, intellectual property
and copyright, sector specific regulation (e.g. health, autonomous
9 https://ec.europa.eu/digital-single-market/en/news/ethics-guidelines-trustworthy-
ai
10 https://www.oecd.org/going-digital/ai/principles/
11 https://www.iso.org/ics/35.080/x/
194
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
vehicles, finance, etc.), privacy and data protection, fundamental
rights, profiling and anti-discrimination issues, competition law,
and legal personality. Quite a number of ethical frameworks have
been proposed— it suffices to note that “at least 63 public-private
initiatives have produced statements describing high-level principles,
values and other tenets to guide the ethical development, deployment
and governance of AI” [14]. Within the framework proposed herein
an ethical framework is referenced to, however the scope of such a
discussion would warrant a paper of its own.
6 CONCLUSIONS
Moving towards a more efficient digital world has its benefits, how-
ever it brings various risks that need to be mitigated through ade-
quate levels of technology assurances. In this paper we have high-
lighted the need for an AI technology assurance regulatory frame-
work which is implemented in a manner that both promotes tech-
nology and does not stifle innovation, yet at the same time enforces
assurances where required, and we have presented an implementa-
tion of a national AI regulatory framework that is overseen by a
technology regulator. The guiding principle was that mandatory
regulation of AI should be avoided, but other national regulators
(e.g. finance, health, communications, etc.) can then work together
with the technology-centric regulator to identify whether manda-
tory assurances are required.
REFERENCES
[1] Isaac Asimov. 1950. I, robot. Gnome Press.
[2] Olivier Boiral. 2011. Managing with ISO systems: lessons from practice. Long
Range Planning 44, 3 (2011), 197–220.
[3] Nick Bostrom and Eliezer Yudkowsky. 2014. The ethics of artificial intelligence.
The Cambridge handbook of artificial intelligence 1 (2014), 316–334.
[4] Francois Coallier. 1994. How ISO 9001 fits into the software world. IEEE Software
11, 1 (1994), 98–100.
[5] M Delvaux. 2016. Motion for a European Parliament resolution: with recommenda-
tions to the commission on civil law rules on robotics. Technical Report. Technical
Report (2015/2103 (INL)), European Commission.
[6] Joshua Ellul, Jonathan Galea, Max Ganado, Stephen Mccarthy, and Gordon J Pace.
2020. Regulating Blockchain, DLT and Smart Contracts: a technology regulator’s
perspective. In ERA Forum. Springer, 1–12.
[7] Joshua Ellul, Stephen McCarthy, Trevor Sammut, Juanita Brockdorff, Matthew
Scerri, and Gordon J. Pace. 2021. A Pragmatic Approach to Regulating Artificial
Intelligence: A Technology Regulator’s Perspective. CoRR abs/2105.06267 (2021).
arXiv:2105.06267 http://arxiv.org/abs/2105.06267
[8] Olivia J Erdélyi and Judy Goldsmith. 2018. Regulating artificial intelligence:
Proposal for a global solution. In Proceedings of the 2018 AAAI/ACM Conference
on AI, Ethics, and Society. 95–101.
[9] Government of Malta. 2018. The Innovative Technology Arrangements and
Services Act (Chapter 592 of the Laws of Malta).
https://legislation.mt/eli/cap/592/eng/pdf.
[10] Government of Malta. 2018. The Malta Digital Innovation Act (Chapter 591 of
the Laws of Malta). https://legislation.mt/eli/cap/591/eng/pdf.
[11] Gonenc Gurkaynak, Ilay Yilmaz, and Gunes Haksever. 2016. Stifling artificial
intelligence: Human perils. Computer Law & Security Review 32, 5 (2016), 749–758.
[12] IAASB. 2013. ISAE 3000 (revised), assurance engagements other than audits or
reviews of historical financial information. (2013).
[13] Keith Kirkpatrick. 2016. Battling algorithmic bias: How do we ensure algorithms
treat us fairly?
[14] Inioluwa Deborah Raji, Andrew Smart, Rebecca N White, Margaret Mitchell,
Timnit Gebru, Ben Hutchinson, Jamila Smith-Loud, Daniel Theron, and Parker
Barnes. 2020. Closing the AI accountability gap: defining an end-to-end frame-
work for internal algorithmic auditing. In Proceedings of the 2020 Conference on
Fairness, Accountability, and Transparency. 33–44.
[15] Sandra Wachter, Brent Mittelstadt, and Luciano Floridi. 2017. Why a right to
explanation of automated decision-making does not exist in the general data
protection regulation. International Data Privacy Law 7, 2 (2017), 76–99.
[16] Roman V Yampolskiy. 2013. Artificial intelligence safety engineering: Why ma-
chine ethics is a wrong approach. In Philosophy and theory of artificial intelli-
gence. Springer, 389–396.
 Making Intelligent Online Dispute Resolution Tools available to
Self-Represented Litigants in the Public Justice System∗
Towards and Ethical use of the AI technology in the administration of Justice
Fernando Esteban de la Rosa†
University of Granada
Granada, Spain
festeban@ugr.es
ABSTRACT
Over the last decade online dispute resolution (ODR) has moved
from merely e-commerce litigation to widespread use in court sys-
tems. Two phenomena have led to this situation: the rise of Self-
Represented Litigants and Courts moving beyond their traditional
focus, allowing parties, for instance, to file a claim, formulate their
arguments, obtain legal information or even a receive a forecast
about the resolution of the case. AI tools have mainly been used to
enable legal professionals (lawyers, mediators) to better perform
their tasks. Today some jurisdictions have begun to provide justice
users with truly useful intelligent- user centric ODR systems in-
corporating assessment and diagnosis AI tools. These tools may
provide information about a possible outcome. This paper analy-
ses the use being made by some jurisdictions of combined Online
Dispute Resolution and Artificial Intelligence tools and aims to
promote the debate on the ethical governance of making these
tools available to unrepresented litigants. The evaluation follows a
European perspective on the ethical governance of the use of AI in
the Justice System.
KEYWORDS
Justice Systems, Intelligent-user centric Online Dispute Resolution
Systems, Self-Represented Litigants, Ethics
ACM Reference Format:
Fernando Esteban de la Rosa and John Zeleznikow. 2021. Making Intelligent
Online Dispute Resolution Tools available to Self-Represented Litigants in
the Public Justice System: Towards and Ethical use of the AI technology
in the administration of Justice. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.3466077
∗ Lodder and Zeleznikow [8] indicate, whilst there is no generally accepted definition
of ODR, we can think of it as using the Internet to perform Alternative Dispute Resolution
(ADR).
† Both authors contributed equally to this research.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466077
195
John Zeleznikow†
La Trobe University Law School
Melbourne, Australia
J.Zeleznikow@latrobe.edu.au
1 INTRODUCTION: THE NEW COMBINED USE
OF ONLINE DISPUTE RESOLUTION AND
ARTIFICIAL INTELLIGENCE TOOLS IN THE
PUBLIC JUSTICE SYSTEM
Advances in Computing and Information and Communication Tech-
nology have opened new possibilities for implementing traditional
models of justice systems. The development of the COVID19 pan-
demic has further enhanced the rise of Online Dispute Resolution
(ODR) and led to the incorporation of a wide range of technological
tools into dispute resolution. Developments in Artificial Intelligence
(AI) hold promises for improving efficiency and quality in the pro-
vision of access to justice, leading to improved transparency and
standardisation of case-law.
Tania Sourdin [12] has suggested that there are three primary
ways in which technology has already restructured the justice
system. First, and at the most rudimentary level, are “supportive”
technologies – these technologies aim to inform, support and ad-
vise individuals involved in the justice system and include, for
example, online legal applications (apps). At the second level are
“replacement” technologies – these technologies replace the roles
and activities traditionally conducted by humans and include, inter
alia, e-filing processes and online mediation services. Finally, and at
the most advanced level, are “disruptive” technologies – these tech-
nologies fundamentally alter the way in which legal professionals
work and include, for example, AI judges or other algorithm-based
decision-making programs that may reshape the judicial role.
Whilst there are numerous e-courts (for instance in the United
Kingdom, USA and Australia), not many jurisdictions have taken
the step forward of proposing a combined use of ODR tools designed
to provide information in such a way that the parties may obtain a
diagnosis or a prediction of the outcome so that disputants can be
better prepared to deal with a direct negotiations conducted online.
Manifestations of this combination of tools are found today in the
Dutch Platform Rechtwijzer, the new Internet Courts in China and
the Civil Resolution Tribunal (CRT) in British Columbia (Canada).
There are also projects with the same aim in Singapore and in
Estonia, in the latter case this includes the creation of a “robot
judge”.
This paper analyses the combined use of ODR and AI tools in
courts, with a special focus on cases where AI tools aim to support
Self Represented Litigants (SRL)s in reaching settlements. This
combination of tools makes use of the notion of Bargaining in the
Shadow of the Law developed by Mnookin and Kornhauser [9]. The
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
paper describes the underlying reasons that have triggered the rise
of SRLs and includes proposals to ensure their fair treatment.
2 EXAMPLES INCORPORATING ARTIFICIAL
INTELLIGENCE INTO ONLINE DISPUTE
RESOLUTION SYSTEMS
2.1 The Dutch platform Rechtwijzer
Rechtwijzer1 (Roadmap to Justice) was designed for couples, with
children who are separating. The aim of Rechtwijzer was ‘to em-
power citizens to solve their problems by themselves or together
with his or her partner. If necessary, it refers people to the assistance
of experts.’ Couples pay €100 for access to Rechtwijzer, which starts
by asking each partner for information such as their age, income,
education, whether they want the children to live with only one
parent or part time with each, then guides them through questions
about their preferences.
The platform had a diagnosis phase; an intake phase for the
initiating party; and then invited the other to join and undertake
the same intake process. Once intake is completed, the parties start
working on agreements. The dispute resolution model is that of
integrative (principled) negotiation[5]. The parties are informed of
rules such as those for dividing property, child support and stan-
dard arrangements for visiting rights so that they could agree on
the basis of informed consent. Agreed agreements are reviewed by
a neutral lawyer. If the proposed solutions are not accepted, then
couples can employ the system to request a mediator for an addi-
tional €360, or a binding decision by an adjudicator. Rechtwijzer is
voluntary and non-binding up until the point where the parties seek
adjudication. Rechtwijzer had aimed to be self-financing through
user contributions. This has not occurred.
2.2 The British Columbia Civil Resolution
Tribunal
The British Columbia Civil Resolution Tribunal [11] is the most
significant current widely available ODR system that comes closest
to providing a full suite of dispute resolution services. It commences
by diagnosing the dispute, by providing a decision tree, and provides
legal information and tools such as customized letter templates.
If this action does not resolve the dispute, one can then apply
to the Civil Resolution Tribunal for dispute resolution. The sys-
tem directs the user to the appropriate application forms. Once the
application is accepted, the user enters a secure and confidential
negotiation platform, where the disputants can attempt to resolve
their dispute. If the parties cannot resolve the dispute, a facilitator
will assist. Agreements can be turned into enforceable orders. If
negotiation or facilitation does not lead to a resolution, an indepen-
dent member will make a determination about the dispute.
Currently, the Civil Resolution Tribunal deals with Motor vehi-
cle injury disputes, Small claims disputes, Strata property disputes,
Societies and cooperative associations disputes and Shared accom-
modation and some housing disputes. For some of these domains
potential litigants can only use the Civil Resolution Tribunal.
To assist digitally disadvantaged litigants, technical support is
provided in accessing the Internet. One of the major reasons that
1 https://rechtwijzer.nl/ last viewed 5 February 2021.
196
F. Esteban de la Rosa and J. Zeleznikow
the Civil Resolution Tribunal has been so successful, is that British
Columbia residents are mandated to use the system when dealing
with issues listed above. Whilst such an approach may be seen novel
and discriminatory, it does ensure that the system is used, with
relative ease, quickly and at minimal cost. In most cases parties are
to represent themselves, even if representation and legal assistance
is allowed.
2.3 The Internet Courts in China
Between 2017 and 2018 China created three new Courts: the Hangzhou
Internet Court, the Beijing Internet Court and the Guangzhou In-
ternet Court. These courts only have material jurisdiction over
internet-related cases. The online platform makes an intelligent
litigation risk assessment system available to the user and can pro-
vide a report synthesising the litigants’ case and the corresponding
risk based on the analysis of court data and similar cases. Litigation
risk assessment aims to help the party without legal knowledge
to identify and exclude common litigation risks, thereby reduc-
ing unnecessary losses. Meanwhile, the assessment can make the
party aware that litigation is risky and costly and guide the parties
to choose ADR or diversified dispute resolution. The system can
automatically generate a complaint letter by simply selecting the
suitable response options. [4]
2.4 Projects in Estonia and Singapore
In July 2019 the Estonian Ministry of Justice launched a project
developing AI software to hear and resolve small economic dis-
putes by eliminating human intervention [10]. The “robot judge”
is configured to decide disputes of up to 7,000 euros. According
to the project, the disputing parties would have to upload their
documents and relevant information to a judicial platform. The AI
machine renders a decision that can be appealed to a human judge.
The project limits its scope to contractual disputes.
Singapore has been committed to digital justice since 2000. In
recent years it has been developing a more ambitious online system,
initially only for injuries arising from motor vehicle accidents. An
outcome simulator will provide guidance to potential claimants,
prior to the commencement of proceedings, helping them to de-
cide on offers from insurance companies. The aim is for parties
to first use the technology to reach amicable settlements without
professional legal advisors [14].
3 A FRAMEWORK FOR BUILDING ONLINE
DISPUTE RESOLUTION TOOLS FOR
SELF-REPRESENTED LITIGANTS
An increasing phenomenon in Common Law countries is the grow-
ing number of pro se (or self-represented) litigants. Landsman [7]
argues that pro se cases pose inherent problems: they can cause de-
lays, increase administrative costs, undermine the judges’ ability to
maintain impartiality and can leave the often-unsuccessful litigant
feeling as though she has been treated unfairly.
Research conducted in the Family Court of Australia shows that
there are a range of reasons why people represent themselves,
such as funding cuts and changes in eligibility to legal aid [3].
Other contributing factors include changes in technology, cultural
shifts towards self-help and self-representation, and changes in
Making Intelligent Online Dispute Resolution Tools available to Self-Represented Litigants in the Public Justice System
legislation. The experiences of self-representation in Australian
law has generally accepted that SRLs are at a disadvantage in legal
proceedings and their experience of the legal system may indeed
be negative. The lack of knowledge or skills of SRLs means that
some are not able to access fair and equal justice in a system often
geared towards legal representation.
In England litigants can go to court without legal aid, in practice
the technical and formal nature of proceedings, with the exception
of the small claims procedure (for claims up to £10,000), makes
legal aid necessary. Its lack has led to public dissatisfaction but
also frustration among judges, faced with the need to inform lay
litigants about the technicalities of the process without being able
to cross the line between information and legal aid. This situation
led to a considerable increase in the time and cost spent on each
judicial decision, even doubling it2 . While some SRLs can present
their case competently, most research suggests that SRLs struggle
with substantive law and procedure [6].
Recent experiences, such as the online court established in Utah,
are demonstrating that ODR has the potential to transform the way
the American legal system deals with pro se litigants and access
to justice issues at large. Although it may seem counterintuitive
to bridge the justice gap by precluding people from appearing in
court, requiring certain types of claims to begin online will actually
provide quicker and more accessible legal solutions. As long as the
programming and administration of ODR technology are conducted
with attention to legal and ethical concerns, pro se litigants will
benefit from having their claims resolved online [2]. For this aim
access to justice is helped by the use of intelligent-user centric ODR
systems incorporating assessment and diagnosis AI tools [15].
Stranieri et al. [13] approach for providing advice about the dis-
tribution of marital property following divorce in Australia was to
use machine learning to provide advice about BATNAs (a BATNA is
used to inform disputants of the likely outcome if the dispute were
to be decided by decision-maker e.g. judge, arbitrator or ombuds-
man). Despite using Machine Learning, it involved the development
of 94 Toulmin argument structures [16] to model the domain as it
existed in 1995. Twenty-five years later, the theoretical principles
behind machine learning software have not changed. But computer
hardware is now much cheaper and data can be much more easily
stored. This has led to the development of ‘quicker; systems’, which
the community has seen as ‘more intelligent’3 .
Whilst the Split-Up system provides advice about BATNAs, the
Family Winner System [1] provided advice to disputing parents
on how they could best negotiate trade-offs. The disputing parties
were asked to indicate how much they valued each item in dispute.
Using logrolling, parties obtained what they most desired.
Zeleznikow [20] discusses how it is possible to build ODR sys-
tems that support self-represented litigants and what skills do self-
represented litigants require to use such systems. Zeleznikow [21]
considers how we can construct such systems with user centric
computing. So, what are the various types of ODR systems and
how can self-represented litigants use them? Having regard to the
2 JUSTICE, “Delivering Justice in an Age of Austerity” (April 2015). Available in https:
//justice.org.uk/justice-age-austerity-2/ last viewed 19 April 2021
3 See for example amica.gov.au which uses machine learning to advise upon property
distribution amongst separating couples in Australia.
197
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
vulnerable position of the unrepresented litigant a truly helpful
ODR system should provide the following facilities:
(1) Case management: the system should allow users to enter
information, ask them for appropriate data and provide for
templates to initiate the dispute. Self-represented litigants
should be able to initiate the dispute, enter their pertinent
data and also track what is happening during the dispute
as well as being aware of what documents are required at
specific times;
(2) Triaging: the system should provide information on how
important it is to act in a timely manner and where to send
the dispute. This may be particularly important in cases of
domestic abuse or where there is a potential for children
to be kidnapped. Triaging systems are vital for expediting
action in high risk cases;
(3) Advisory tools: the system should provide tools for reality
testing: these could include, books, articles, reports of cases,
copies of legislation and videos; there would also be calcu-
lators (such as to advise upon child support) and BATNA
advisory; systems (to inform disputants of the likely out-
come if the dispute were to be decided by decision-maker
(e.g. judge, arbitrator or ombudsman). Advisory tools, as
suggested by Zeleznikow [19] are a vital cog in supporting
self-represented litigants. An important associated question
is how can we design advisory tools that self-represented
litigants can gainfully use? Are the legal concepts behind
these tools too difficult for amateurs to understand? How do
we construct suitable user interfaces?
(4) Communication tools – for negotiation, mediation, con-
ciliation or facilitation. This could involve shuttle mediation
if required. For many ODR providers, the provision of com-
munication tools is their main goal;
(5) Decision Support Tools – if the disputants cannot resolve
their conflict, software using game theory or AI can be used
to facilitate trade-offs. Professionals (such as lawyers) can
provide useful advice re trade-offs. In their absence, suitable
decision support tools are vital;
(6) Drafting software: if and once a negotiation is reached,
software can be used to draft suitable agreements. Drafting
plans (such as parenting plans) once there is an in-principle
agreement for a resolution of a dispute, is a non-trivial task.
No single dispute is likely to require all six processes. However, the
development of such a hybrid ODR system would be very signifi-
cant. A total system would require us to construct the appropriate
systems 1 to 6, and the ultimate solution is to make sure that all the
systems are capable of communicating with each other.
4 ETHICAL ISSUES RELATING TO THE
PROVISION OF ARTIFICIAL
INTELLIGENCE-BASED TOOLS TO
SELF-REPRESENTED LITIGANTS BY THE
PUBLIC JUSTICE SYSTEM: A EUROPEAN
PERSPECTIVE
Neither the recent official documents of the European Union deter-
mining how AI should be used in the field of the administration of
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
justice4 nor the European Ethical Charter (EEC) on the use of AI
in judicial Systems and their environment adopted in 2018 by the
European Commission for the Efficiency of Justice of the Council
of Europe deal directly with the admission of AI tools aimed at
enabling the parties to assess their legal position. Because SRLs
generally lack legal skills and in view of the objective to encour-
age negotiation we submit that this use of technology for these
purposes should be considered high-risk.
The EEC points out the inherent risks in these technologies may
even transcend the act of judging and affect essential functioning
elements of the rule of law and judicial systems. These include
principles such as the primacy of law. These tools could create
a new form of normativity, which could supplement the law by
regulating the sovereign discretion of the judge, and potentially
leading, in the long term, to a standardisation of judicial decision
based no longer on case-by-case reasoning by the courts, but on
a pure statistical calculation linked to the average compensation
previously awarded by other courts. That is why the report submits
a need to consider whether these solutions are compatible with
the individual rights enshrined in the European Convention on
Human Rights (ECHR). These would include the rights to a fair
trial (particularly the right to a natural judge established by law, the
right to an independent and impartial tribunal and equality of arms
in judicial proceedings) and, where insufficient care has been taken
to protect data communicated in open data, the right to respect for
private and family life. Thus the EEC considers that applications
of predictive justice should be assigned to the field of research and
further development in order to ensure that they fully tie in with
actual needs before contemplating use on a significant scale in the
public sphere.
The European Commission (EC) recognises that the use of AI
applications can bring many benefits, such as making use of infor-
mation in new and highly efficient ways, and improve access to
justice, including by reducing the duration of judicial proceedings.
At the same time it is aware that the opacity or biases embedded in
certain AI applications can also lead to risks and challenges for the
respect of and effective enforcement of fundamental rights, includ-
ing in particular the right to an effective remedy and a fair trial.
The EC recognises as a possible high-risk a use case using the tech-
nology as part of decision-making processes with significant effects
on the rights of people. However, it also considers that the pro-
posed requirements in the White Paper on increased transparency,
human oversight, accuracy and robustness of these systems aim
to facilitate their beneficial use, while ensuring that fundamental
rights including non-discrimination based on sex, racial or ethnic
4 Among the last official documents are the Proposal for a Regulation Laying down
Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act) and Amending
Certain Union Legislative Act of 21.4.2021 COM (2021) 206 final; the Communication
from the Commission to the European Parliament, the Council, the European Economic
and Social Committee and the Committee of the regions called “Digitalisation of justice
in the European Union. A toolbox of opportunities”, COM (2020) 710 final, of 2.12.2020;
the European Parliament Resolution of 20 October 2020 with recommendations to
the Commission on a framework of ethical aspects of artificial intelligence, robotics
and related technologies (2020/2012 (INL); White Paper on Artificial Intelligence - A
European approach to excellence and trust, COM(2020) 65 final of Brussels, 19.2.2020;
the European e-Justice Strategy 2019-2023 of 13 March 2019 (2019/C 96/04) Council
2019-2023; the Digital Revolution in view of Citizens’ Needs and Rights. Opinion of
the European Economic and Social Committee of 20.02.2019.
198
F. Esteban de la Rosa and J. Zeleznikow
origin, religion or belief, disability, age or sexual orientation) are
respected and rule of law and due process principles upheld.
In order to understand the European position it is also relevant to
know the criterion followed by the new proposal for a Regulation of
April 2021. AI systems intended for the administration of justice are
not listed among the prohibited practices (art. 5) but among the high-
risk AI systems (point 40 of the preamble). The new proposal for a
Regulation separates two kinds of judicial activities: it is considered
as high-risk the systems intended to assist judicial authorities in
researching and interpreting facts and the law and in applying the
law to a concrete set of facts. Such qualification is not extended to AI
systems intended for purely ancillary administrative activities that
do not affect the actual administration of justice in individual cases.
The proposed Regulation does not establish the definitive answer
as any use of AI must continue to occur solely in accordance with
the applicable requirements resulting from the European Charter
of Fundamental Rights, the rest of European Law and the national
law.
We submit that in view of the beneficial impact it may have on
the functioning of the judicial system, it is necessary to identify the
real possibilities, technical limits and safeguards to be met by the
machines offered by the public justice system to SRLs.
For specific areas of administrative law it is possible to develop
legal rules as code providing useful information and support for
SRLs. The use of code as rules in combination with User Centric
ODR Tools using decision trees, may have success promoting ac-
cess to justice for SRLs. The CRT in the British Columbia is an
example of success. The design of AI rule-based systems does not
exhibit the difficulties arising from the lack of transparency and
the creation of biases that may arise employing ML induction al-
gorithms. Deductive AI tools (the so called Experts Systems) allow
transparency and the monitoring of the machine output is facili-
tated to be able to rectify what is necessary in case any errors in the
programming are discovered. Programming is, however, a delicate
process and if not done well can lead to unfair treatment when the
algorithm doesn’t match reality. This can occur when a one-size-
fits-all rule is implemented in a complex environment. A recent
example is Australia’s Centrelink “robodebt” debacle5 . In that case,
welfare payments made on the basis of self-reported fortnightly
income were cross-referenced against an estimated fortnightly in-
come, taken as a simple average of annual earnings reported to
the Australian Tax Office, and used to auto-generate debt notices
without any further human scrutiny or explanation. This assump-
tion is at odds with how Australia’s highly casualised workforce is
actually paid. For example, a graphic designer who was unable to
find work for nine months of the financial year but earned A$12,000
in the three months before June would have had an automated debt
raised against her. This is despite no fraud having occurred, and
this scenario constituting exactly the kind of hardship Centrelink
is designed to address.
Rules as codes requires alterations to be introduced in case of
legislative changes. Although it will not be possible to attain the
quality of advice offered by a legal expert, we submit that the infor-
mation provided to SRLs through machines makes a contribution to
5 See https://tinyurl.com/y3dqe6mg last viewed 19 April 2019
Making Intelligent Online Dispute Resolution Tools available to Self-Represented Litigants in the Public Justice System
improving access to justice for those who cannot afford legal assis-
tance. Regarding the quality of advice provided by these machines,
it seems reasonable that the proposals of the European Commission
about requirements concerning possible testing of applications and
the need to provide relevant documentation on their purposes and
functionalities. It seems also reasonable to require maintaining the
possibility to correct errors and providing information to the user
that the answer given by the machine may not necessarily match
the answer that would be given by a judge hearing the case.
Two of the disadvantages of the use of Machine Learning systems
are that they are not transparent, and the data and the software on
which they are based may be manipulated. There is also a concern
that the use of Machine Learning in the legal system will worsen
biases against minorities or deepen the divide between those who
can afford quality legal assistance and those who cannot [17]. Algo-
rithms will continue to perform existing biases against vulnerable
groups because the algorithms are largely copying and amplifying
the decision-making trends embedded in the legal system. There is
already a class divide in legal access – those who can afford high
quality legal professionals will always have an advantage. The de-
velopment of intelligent support systems can partially redress this
power imbalance by providing users with important legal infor-
mation that was previously unavailable to them. Difficulties may
stem from biases. One example is COMPAS, a decision support
system designed to help parole boards in the United States [18]
decide which prisoners to release early, by providing a probability
score of their likelihood of reoffending. Rather than rely on a simple
decision rule, the algorithm used a range of inputs, including demo-
graphic and survey information, to derive a score. The algorithm
did not use race as an explicit variable, but it did embed systemic
racism by using variables that were shaped by police and judicial
biases.
What can be done is to ensure the traceability and cleanliness
of the data with which the machine operates, and to introduce
elements of weighting. But as Richard Susskind considers, an ethical
programming is not feasible. It is not at all clear, either technically
or philosophically, what is meant when it is proposed that ethics
should be embedded in Machine Learning. Nor it is clear what
is meant when it is demanded that software engineers program
Machine Learning systems to provide intelligent explanations. To
think so is to misunderstand the difference between the inductive
processes inherent in Machine Learning and the kind of argument
we expect when we ask for an explanation [14].
A different issue is the use of AI tools by judges to decide a case.
We share the European Commission’s view that it is important that
judgments are delivered by judges who fully understand the AI
applications and all information taken into account therein that
they might use in their work (AI not to replace but as Augmented
Intelligence), on the understanding that the use of AI applications
must not prevent any public body from giving explanations for its
decisions. As for the machine being able to decide the case on its
own, as the Estonian project poses, this should not be completely
ruled out. However, we are not at that stage yet! In the current
state of the art, machines can neither motivate nor explain the
decisions and predictions they make [14]. Legal arguments require
persuasion that does not depend on predictable variables.
199
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
5 CONCLUSION
One of the latest trends in the incorporation of technology in the
administration of justice is the provision by public justice systems
to support SRLs by the use of a combination of AI and ODR tools.
These allow SRLS to have a diagnosis of the case, which influences
the parties either to determine a dismissal of the action or how
to negotiate. This combination of tools shows great potential in
reducing the level and duration of litigation. The paper submits that
this use of the technology must be considered as high risk as it may
function as a replacement of judicial activities. However, it is still
possible to obtain positive results from this technology by inserting
some safeguards, as is beginning to emerge from the European legal
sphere. The debate is now about what safeguards are necessary
to ensure that the use of high-risk artificial intelligence tools in
the field of justice is fully compatible with the rule of law. The
implementation and use of this technology should be preceded by
the detection and diagnosis of the functioning of justice in specific
sectors, so that the efforts are made in the areas with most pressing
needs.
REFERENCES
[1] Emilia Bellucci and John Zeleznikow. 2006. Developing Negotiation Decision
Support Systems that support mediators: a case study of the Family_Winner
system. Journal of Artificial Intelligence and Law 13, 2 (2006), 233–271.
[2] Julianne Dardanes. 2021. When Accessing Justice Requires Absence from the
Courthouse: Utah’s Online Dispute Resolution Program and the Impact it Will
Have on Pro Se Litigants. Pepperdine Dispute Resolution Law Journal 21, 1 (2021).
[3] John Dewar, Barry W. Smith, and Cate Banks. 2000. Litigants in Person in the
Family Court of Australia – Research Report No 20, Family Court of Australia.
Vol. Research Report No 20. Family Court of Australia Canberra.
[4] Xuhui Fang. 2018. Recent Development of Internet Courts in China. International
Journal on Online Dispute Resolution 5 (2018), 1–2, 49–55.
[5] Roger Fisher and William Ury. 1981. Getting to yes. Penguin Group, New York.
[6] Hazel Genn and Yvette Genn. 1989. The effectiveness of representation at tribunals.
Lord Chancellor’s Department.
[7] Stephan Landsman. 2009. The growing challenge of pro se litigation. 13 Lewis &
Clark L. Rev. 439 (2009).
[8] Arno Lodder and John Zeleznikow. 2010. Enhanced dispute resolution through the
use of information technology. Cambridge University Press.
[9] Robert H. Mnookin and Lewis Kornhauser. 1979. Bargaining in the shadow of
the law: The case of divorce. The Yale Law Journal 88, 5 (1979), 950.
[10] Eric Niiler. 2019. Can AI be a Fair Judge in Court? Estonia Thinks So. https:
//www.wired.com/story/can-ai-be-fair-judge-court-estonia-thinks-so/
[11] Shannon Salter and Darin Thompson. 2016. Public-Centred Civil Justice Redesign:
A Case Study of the British Columbia Civil Resolution Tribunal”. McGill Journal
of Dispute Resolution 3 (2016), 113.
[12] Tania Sourdin. 2018. Judge v. Robot: Artificial Intelligence and Judicial Decision-
Making. UNSWLJ 41 (2018), 1114.
[13] Andrew Stranieri, John Zeleznikow, Mark Gawler, and Bryn Lewis. 1999. A hybrid
rule–neural approach for the automation of legal reasoning in the discretionary
domain of family law in Australia. Artificial intelligence and Law 7, 2-3 (1999).
[14] Richard E. Susskind. 2019. Online courts and the future of justice. Oxford University
Press.
[15] Darin Thompson. 2015. Creating New Pathways to Justice Using Simple Artificial
Intelligence and Online Dispute Resolution. International Journal of Online
Dispute Resolution 4 (2015).
[16] S. Toulmin. 1958. The Uses of Argument. Cambridge University Press, Cambridge.
[17] Peter K. Yu. 2020. The Algorithmic Divide and Equality in the Age of Artificial
Intelligence, 72 FLA. L. REV 331 (2020).
[18] Monika Zalnieriute, Lyria Bennett Moses, and George Williams. 2019. The rule
of law and automation of government decision-making. The Modern Law Review
82, 3 (2019), 425–455.
[19] J. Zeleznikow. 2002. Using Web-based Legal Decision Support Systems to Improve
Access to. Justice Information and Communications Technology Law 11, 1 (2002).
[20] John Zeleznikow. 2020. The challenges of using Online Dispute Resolution to
support Self Represented Litigants. Journal of Internet Law 3 23, 7 (2020).
[21] John Zeleznikow. 2021. Using Artificial Intelligence to provide Intelligent Dispute
Resolution Support. Group Decision and Negotiation (2021). https://doi.org/10.
1007/s10726-021-09734-1
 Plum2Text: A French Plumitifs–Descriptions Data-to-Text
Dataset for Natural Language Generation
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel
Laval University, Computer Science Department and Faculty of Law
Québec, Canada
nicolas.garneau@ift.ulaval.ca,eve.gaumond@observatoire-ia.ulaval.ca
luc.lamontagne@ift.ulaval.ca,pierre-luc.deziel@fd.ulaval.ca

ABSTRACT
In this paper, we introduce a new French Data-to-Text (D2T) dataset
in the legal domain: Plum2Text 1 . It is made out of plumitifs (docket
files) – descriptions pairs that are derived from publicly available
documents issued by Canadian criminal courts. The development of
Plum2Text is primarily intended to train statistical natural language
generation algorithms, in order to make the plumitifs more easily
understandable for Canadian citizens. The inputs and outputs of
the dataset are unique: on the data side, the values of the table
contain long pieces of textual utterance, and on the text side (or
reference), it most often consists of a paraphrase of the table values.
We describe how we curated the plumitif–description associations
by introducing an annotation tool and a methodology specific to
the D2T natural language generation task. We do so by using simple
yet efficient text classifiers to help the annotator leverage annotated
examples during the annotation process. As a matter of privacy, we
also illustrate how we are decontextualizing the descriptions.
data (e.g. a table 𝑡) and generates a textual utterance (namely the
hypothesis ℎ) that is syntactical and semantically faithful to both the
structured input, and one or possibly several textual references 𝑟 . In
this paper, we propose Plum2Text, a new French D2T dataset in the
legal domain, rather different from those previously introduced in
the literature. The dataset was built from Quebec’s plumitifs which
are legal documents that lie in the same family as court dockets.
The plumitifs are textual summaries, written in a structured format,
which present all the steps of a judicial case. They also provide
information about parties’ identity, the jurisdiction in charge of
administering the case, and some information relating to the nature
and the course of the preceding, as illustrated in Figure 12 .
Figure 1: A Plumitif example illustrating the accused and
plaintiff personal information along with charges and asso-
ciated pleas, decisions and penalty.

CCS CONCEPTS
• Applied computing → Law; Annotation; • Computing method-
ologies → Language resources; Natural language generation;

KEYWORDS
Legal Language Resource, Natural Language Generation, Annota-
tion Methodology
ACM Reference Format:
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel. 2021.
Plum2Text: A French Plumitifs–Descriptions Data-to-Text Dataset for Natu-
ral Language Generation. In Eighteenth International Conference for Artificial
Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM,
New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.3466148

1 INTRODUCTION
D2T generation [12, 18] is a specialized task of natural language
generation (NLG) where a model takes as input semi-structured
1 https://nicolas.nlp.quebec/files/icail_2021_plum2text.jsonl It has been shown that the plumitifs lack intelligibility [1, 19].
Indeed, these files are written by clerks and contain several abbrevi-
Permission to make digital or hard copies of all or part of this work for personal or ations and references to the Criminal Code’s provisions, making it
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation difficult for litigants to understand. Beauchemin et al. [1] attempted
on the first page. Copyrights for components of this work owned by others than ACM to generate plumitifs descriptions using a rule-based system. They
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
found out however that their model could hardly generalize to
fee. Request permissions from permissions@acm.org. plumitifs from other judicial districts or having slight differences
ICAIL’21, June 21–25, 2021, São Paulo, Brazil in the way they were organized. But having only the plumitifs in
© 2021 Association for Computing Machinery.
2 All
examples (except screenshots) in this paper have been translated from French to
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466148 English to facilitate understanding

200
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel

Table 1: Statistics about the different D2T datasets intro- summaries paired with their corresponding statistic tables. The
duced in Section 2.1, WikiBio, WebNLG, RotoWire, Ro- dataset provides 4.9K examples containing on average 630 records
toWire augmented and purified by Wang [21], E2E, ToTTo, with longer associated text of 340 tokens, on average. Wang [21]
and our dataset, Plum2Text. For each dataset, we present extended RotoWire, namely RotoWire-FG, adding 50% more data
the number of examples (NE), average number of input at- and enriched input tables. Wang [21] observed that only 60% of
tributes (NA) along with the average number of tokens for the summary contents can be grounded to the table records and
the inputs (A-Avg) and the references (R-Avg). thus proposed a purified version of RotoWire. Thomson et al. [20]
refined even further Rotowire-FG by providing more attributes
Datasets NE NA A-Avg R-Avg across multiple dimensions, increasing the content overlap between
WikiBio 728K 20 3 26 statistic tables and reference texts. Dušek et al. [8] proposed the
WebNLG 22K 1-7 5 25 E2E challenge dataset, a crowdsourced dataset of 50K examples in
RotoWire 4.9K 630 1 340 the restaurant domain, as well as a cleaned version of it [7]. Inputs,
RotoWireW+ 7.5K 630 1 206 using the Meaning Representation format, contain 3-8 attributes
E2E 50K 3-8 15 20 and an average of 20 tokens per reference text. Finally, Parikh et al.
ToTTo 120K 4 2 20 [15] proposed an open-domain D2T dataset, ToTTo, covering a wide
range of topics. It contains over 120K examples of Wikipedia tables
Plum2Text 2.5K 2-9 61 50
along with one-sentence descriptions. They extend their dataset
with highlighted cells offering better control for generation. On
hand, they could not train a statistical model for natural language average, a given input contains 4 attributes and the reference text
generation in order to alleviate this problem. We wish to solve this has 20 tokens. While this list of D2T datasets is not exhaustive, it
issue by working with court judgments. One way to describe judge- illustrates the diverse nature of the data which is tightly coupled
ments is that they are in a sense a longer version of the plumitif; to the task of D2T NLG. We aggregated the datasets’ statistics in
they thoroughly explain what is written in the plumitif. By pairing Table 1.
plumitifs with their court judgments, we created Plum2Text, a D2T
dataset that can be used to train statistical language generation
model. It also allowed us to reframe the problem into one that can
2.2 Datasets in the Legal Domain
be solved by D2T natural language solutions. The input and output Legal–NLP is an emerging field and so is the creation of datasets
components of Plum2Text are very unique on their own; the values supporting the recent advances. While the following list of re-
of the table contain long pieces of text (e.g. paragraphs from the sources is not exhaustive, we selected the ones most related to our
law) and the references are mostly paraphrases of the table content. contribution. Kano et al. [11] introduced in 2018 the COLIEE legal
In the next sections, we compare Plum2Text with other standard case retrieval task along with a dataset drawn from an existing
D2T datasets known to the community. We also position our con- collection of case law primarily from the Federal Court of Canada.
tribution amongst other datasets in the legal field. We provide an In 2019, Rabelo et al. [17] introduced the statute law competition
at-length explanation of the annotation process we used to create data corpus where the questions (for the question answering task)
Plum2Text and present the tools we designed for this task. We hope were drawn from Japanese Legal Bar exams. They also proposed
that our methodology will encourage the creation of many other three new tasks: legal case entailment, statute law retrieval, and
datasets not only in the legal field but also in different D2T domains. legal question answering.
Xiao et al. [24] introduced the Chinese AI and Law challenge
2 RELATED WORK dataset (CAIL2018), the first large-scale Chinese legal dataset for
Since we are introducing a new D2T dataset in the legal domain, we judgment prediction. Still in Chinese, Duan et al. [6] introduced
first introduce standard D2T datasets on general purpose domains the Chinese judicial reading comprehension (CJRC) dataset. More
(e.g. Wikipedia) known to the community and further describe recently, Chalkidis et al. [3] proposed a large-scale multi-label text
what has been proposed in the legal field regarding different natural classification dataset on the European Union legislation. In the
language processing tasks. French Canadian spectrum, Westermann et al. [22] introduced two
datasets drawn from 1 million written judgments from the Régie du
2.1 Data-to-Text Datasets logement du Québec where they used factors to predict and analyze
landlord–tenant judgments and to create a chatbot out of it. Cumyn
There is a handful of D2T datasets and each of them has its specifics
et al. [4] proposed an annotated set of 2,500 judgments using a
that we further expose in this section. Lebret et al. [13] introduced
faceted scheme with the objective of improving the performance
WikiBio, a large-scale fact table of biographical sentences extracted
of legal search engines. Closer to the NLG task, Bhattacharya et al.
from Wikipedia. This dataset has 728K examples, each contain-
[2] studied several summarization techniques on a large set of
ing on average 20 records and 26 description tokens. In a similar
Indian Supreme Court judgments. Ye et al. [25] proposed a dataset
vein, Gardent et al. [9] introduced the WebNLG dataset consisting
where a generative model learns to generate court views from fact
of 22K RDF–text pairs extracted from DBPedia 3 . An input may
descriptions in Chinese. They framed the generation problem as
contain up to seven RDF triples and the average text length is 25
text summarization. From that point of view, Plum2Text is a rather
tokens. Wiseman et al. [23] introduced RotoWire, BasketBall game
unique new dataset in the legal field since it tackles a new task (i.e.
3 https://wiki.dbpedia.org/ D2T generation) and is in French.

201
Plum2Text: A French Plumitifs–Descriptions Data-to-Text Dataset for Natural Language Generation ICAIL’21, June 21–25, 2021, São Paulo, Brazil

3 THE PLUM2TEXT DATASET

In this section, we first introduce the two source datasets used
to create Plum2Text. We further describe the methodology and
annotation tools we developed to rapidly obtain such a dataset. We
then explain how we deidentified4 the resulting dataset for a matter
of privacy, and then conclude with a set of statistics.
Figure 2: An excerpt listing the paragraph (1) containing the
accusations associated to the plumitif in Figure 1, an accusa-
tion of conspiracy (art. 465 (1) c from the Criminal Code of
Canada).

3.1 Source Datasets

In this section, we introduce the two source datasets, namely the
Quebec’s plumitifs and their associated Court Judgments.
3.1.1 Quebec’s plumitifs. Quebec’s plumitifs are essentially used
as summaries presenting all the steps of a court case. Information
concerning parties’ identity, jurisdiction in charge of administering
the case, and some information regarding the nature and the course
of the preceding are available through the plumitifs. In the context
of criminal proceedings, they contain relevant information for liti-
gants about the plaintiff, the defendant, and the various charges he’s
convicted of along with associated penalties (if applicable), as illus-
trated in Figure 1. Beauchemin et al. [1], who previously introduced
the plumitif corpus for NLG, showed that there is a need to make
the plumitifs more understandable. They proposed a template-based
generation architecture which, given a raw plumitif text, generates text classifiers in the annotation pipeline in order to rapidly cre-
its description in an unsupervised way due to the lack of annotated ate a D2T dataset. On average, the process of annotating pairs of
pairs. They concluded that a generalization challenge stems from Plumitif-Judgment with the help of classifiers takes approximately
the usage rules to generate a plumitif ’s description. two to five minutes per pair.
3.1.2 Court Judgments. We found out recently that a large pro- 3.2.1 The Plumitif Setting. The plumitif part is decomposed
portion of plumitifs have an associated court judgment (document) into 3 sections; the defendant, the plaintiff, and the list of charges.
that describes at-length the content of a given criminal case, hence In this work, we focus on generating the charges sections of a
being a textual description of the plumitif. We selected the same plumitif since it conveys an interesting fact verification challenge.
criminal plumitifs introduced in [1], added some plumitifs from 21 More precisely, each charge is composed of three components; the
other districts, and fetched their associated judgment (in French, if provision, the verdict, and the plea. Similarly, a judgment associ-
available) heard by the Criminal and penal division of the court of ated with the plumitif can be broken down into a set of relevant
Québec. With that in mind, we deemed it possible to create such paragraphs. The main goal of the annotator is to find the mapping
a D2T dataset to train a statistical NLG architecture, hoping to between the set of charges’ components with their respective rel-
solve the generalization problem raised by Beauchemin et al. [1]. evant paragraphs. This association step refines the context table
With the plumitifs and judgments both gathered from the SOQUIJ 𝑡, in a similar way Dhingra et al. [5] did by highlighting cells in
website5 , we created a total of 1,289 plumitif –judgment pairs from Wikipedia tables.
29 different districts. Judgments, often spanning over several pages,
3.2.2 Annotation Application. The annotation application is
are composed of multiple paragraphs containing detailed contex-
composed of one main interface; the plumitif–judgment association
tual information about the case6 . However, most of the document’s
interface as depicted in Figure 3. The interface contains on the left-
content is irrelevant to the generation of the plumitif ’s description
hand side the set of accusations of a given plumitif, and on the right-
(e.g. non-essential factual details). To align the plumitif ’s compo-
hand side, the list of paragraphs that compose the judgment. The
nents with its corresponding judgment’s sections, we mainly focus
annotator can then create new plumitif–description annotations by
on paragraphs relating to the charges, verdict, and plea. We explain
selecting which paragraph is referring to which accusation. Once
in the next section how we retrieve these paragraphs from the
the annotator has skimmed through the whole decision document,
judgment in order to obtain plumitifs’ descriptions, and we provide
he can move onto the next plumitif.
in Figure 2 an example of the kind of passage we are looking for.
However, documents can be very long with many useless para-
3.2 Annotation Methodology graphs to the task of plumitif description generation. To this end,
we train a Relevant Paragraph Classifier (RPC) that learns if a para-
In this section, we present our methodology and the main annota- graph is relevant or not. This classifier facilitates the work of the
tion interface. We illustrate the benefits of using simple yet efficient annotator by presenting beforehand paragraphs that potentially
4 What we mean by deidentifying is that one cannot reasonably be expected either, by refer to the plumitif. Given the list of relevant paragraphs and
itself or when combined with other information available, to enable the reidentification some parts of the plumitif that are left without associations, the
of individuals, as Ohm [14] explained on page 1744.
5 https://soquij.qc.ca/ annotator must nevertheless skim through the whole judgment to
6 On average, a judgment has 160 paragraphs. make sure the classifier did not miss something. To construct the

202
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 3: The accusation–paragraphs association interface. The set of accusations is displayed on the left and the whole deci-
sion on the right. The relevant paragraph classifier scores every paragraph and those with a high probability of being relevant
are displayed at the top of the interface, in the Relevant section.
RPC, we trained a standard Binary Text Classifier from the Spacy
library [10]. We used 1,000 relevant paragraphs and 1,000 irrelevant
ones randomly sampled from the judgments. On the test set, the
paragraph classifier obtains 85% accuracy.
3.3 Decontextualization of the Paragraphs
In this section, we present how we decontextualize the paragraphs
(also known as delexicalization, see [18]). Paul Panenghat et al. [16]
showed that delexicalizing a dataset not only removes bias but also
improves out-of-domain portability, for instance in our case from
criminal to civil law. Our decontextualization process is motivated
by the underlying task Plum2Text is designed for; i.e. generating
an intelligible description of a plumitif. With that in mind, we
remove not only personal information but also any information in
the paragraph that cannot be found in the plumitif.
3.3.1 Automatic Depersonalization. Following the argumenta-
tion of Beauchemin et al. [1], we deemed it essential to remove any
personal information that the judgment’s paragraphs may hold,
especially in the case of releasing this dataset. While the main con-
cern here is privacy, it also greatly reduces the vocabulary size by
removing rare tokens, and hopefully will improve the performance
of NLG models. The first step of our decontextualisation process
is to automatically remove all names as well as all information
describing places or organisations (e.g. police department). A pre-
trained named entity recognition (NER) model [10] is used for this
purpose. We also replace dates using regular expressions.
3.3.2 Information Specific to the Case. To further decontextu-
alize the paragraphs, we manually remove information specific to
the case. More concretely, we went through every paragraph and
to make sure that people’s privacy is well protected;
203
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel
(1) Replace names, locations, and organizations that the NER
model may have missed. For example, to preserve the privacy
of certain parties according to the order restricting publi-
cation (486.4 (1) C.Cr.), names are often elided as such; J...
D... for John Doe. A pre-trained NER model does not catch
this kind of elision and leaves unnecessary noise within the
dataset. It had to be done manually.
(2) Remove contextual information specific to the crime that was
perpetrated. For example, one accused may have assaulted
his neighbor, which is very specific to the case. Furthermore,
we remove information that is not supported by the table 𝑡
such as “after a trial of five days”.
(3) Remove gender and numbers7 related to the accused. As
such, we replace the French feminine version of “accusée”
(accused) with “accusé”. One example of an accusation in-
volving several defendants is conspiracy whereas “Person 1
and Person 2 conspired...”. We thus replace “Person 1 and
Person 2” with “The accused”.
(4) Remove information describing the victims such as “Per-
son X, only 9 years old” or “Person X, a woman”. We also
normalize amounts, such as X$ or X kilos of cocaine, re-
garding crimes committed against the controlled drugs and
substances act.
Decontextualization reduced vocabulary size from 2,144 to 1,464
token types. We should also mention that we retrieve the sections
text from their corresponding law. Compared to the D2T datasets
introduced in Section 2.1, Plum2Text is particularly interesting as
several values from the table are sentences or even text paragraphs
(the laws), as illustrated in Figure 4. Also the resulting annotated
7 One case may be regarding several defendants.
Plum2Text: A French Plumitifs–Descriptions Data-to-Text Dataset for Natural Language Generation
dataset contains on average 61 tokens per input value, whereas a
typical D2T dataset usually has 1 to 5 tokens, as depicted in Table 1.
Furthermore, a table has on average 5 associated references with
overlapping table values. These characteristics pose some challenge
for both the generation of a description and its evaluation.
Figure 4: A translated example of a Plumitif–Paragraph pair
from our dataset. The highlighted information in the boxes
illustrates the presence of paraphrasing within Plum2Text.
4 CONCLUSION
In this paper, we introduced a new French D2T dataset in the le-
gal field, Plum2Text. We thoroughly present how we created and
annotated this dataset, by introducing a methodology that we be-
lieve will help the research community. The creation of the dataset
presented in this paper is a stepping stone in the development of
a web application aiming at making plumitifs – a legal document
providing a summary of a given judicial case – more easily un-
derstandable. As explained by [1], enhancing the intelligibility of
plumitifs foster the right to access judicial information which is a
hallmark of Canadian democracy (and that of many other countries
as well). In future works, we plan to train a statistical NLG model
on Plum2Text and enable intelligible description of plumitif at scale.
Acknowledgements
We thank the reviewers for their insightful comments. This research was
funded by the Natural Sciences and Engineering Research Council of Canada
and the Social Sciences and Humanities Research Council of Canada.
REFERENCES
[1] David Beauchemin, Nicolas Garneau, Eve Gaumond, Pierre-Luc Déziel, Richard
Khoury, and Luc Lamontagne. 2020. Generating Intelligible Plumitifs Descrip-
tions: Use Case Application with Ethical Considerations. In Proceedings of the
13th International Conference on Natural Language Generation. Association for
Computational Linguistics, Dublin, Ireland, 15–21. https://www.aclweb.org/
anthology/2020.inlg-1.3
[2] P. Bhattacharya, Kaustubh Hiware, Subham Rajgaria, Nilay Pochhi, Kripabandhu
Ghosh, and S. Ghosh. 2019. A Comparative Study of Summarization Algorithms
Applied to Legal Case Judgments. In ECIR.
[3] Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, and Ion Androut-
sopoulos. 2019. Large-Scale Multi-Label Text Classification on EU Legislation.
In Proceedings of the 57th Annual Meeting of the Association for Computational
Linguistics. Association for Computational Linguistics, Florence, Italy, 6314–6322.
https://doi.org/10.18653/v1/P19-1636
[4] Michelle Cumyn, Günter Reiner, S. Mas, and David Lesieur. 2019. Legal Knowl-
edge Representation Using a Faceted Scheme. Proceedings of the Seventeenth
International Conference on Artificial Intelligence and Law (2019).
204
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
[5] Bhuwan Dhingra, Manaal Faruqui, Ankur Parikh, Ming-Wei Chang, Dipanjan
Das, and William Cohen. 2019. Handling Divergent Reference Texts when
Evaluating Table-to-Text Generation. In Proceedings of the 57th Annual Meeting
of the Association for Computational Linguistics. Association for Computational
Linguistics, Florence, Italy, 4884–4895. https://doi.org/10.18653/v1/P19-1483
[6] X. Duan, Baoxin Wang, Ziyue Wang, Wentao Ma, Yiming Cui, D. Wu, S. Wang, T.
Liu, Tianxiang Huo, Z. Hu, Heng Wang, and Z. Liu. 2019. CJRC: A Reliable Human-
Annotated Benchmark DataSet for Chinese Judicial Reading Comprehension.
ArXiv abs/1912.09156 (2019).
[7] Ondřej Dušek, David M. Howcroft, and Verena Rieser. 2019. Semantic Noise
Matters for Neural Natural Language Generation. In Proc. of the 12th Interna-
tional Conference on Natural Language Generation. Association for Computa-
tional Linguistics, Tokyo, Japan, 421–426. https://doi.org/10.18653/v1/W19-8652
arXiv:1911.03905
[8] Ondřej Dušek, Jekaterina Novikova, and Verena Rieser. 2020. Evaluating the
State-of-the-Art of End-to-End Natural Language Generation: The E2E NLG
Challenge. Computer Speech & Language 59 (Jan. 2020), 123–156. https://doi.org/
10.1016/j.csl.2019.06.009 arXiv:1901.11528
[9] Claire Gardent, Anastasia Shimorina, Shashi Narayan, and Laura Perez-
Beltrachini. 2017. Creating Training Corpora for NLG Micro-Planners. In Proceed-
ings of the 55th Annual Meeting of the Association for Computational Linguistics
(Volume 1: Long Papers). Association for Computational Linguistics, Vancouver,
Canada, 179–188. https://doi.org/10.18653/v1/P17-1017
[10] Matthew Honnibal and Ines Montani. 2017. spaCy 2: Natural language under-
standing with Bloom embeddings, convolutional neural networks and incremen-
tal parsing. (2017). To appear.
[11] Yoshinobu Kano, Miyoung Kim, M. Yoshioka, Yao Lu, J. Rabelo, Naoki Kiyota,
R. Goebel, and K. Satoh. 2018. COLIEE-2018: Evaluation of the Competition on
Legal Information Extraction and Entailment. In JSAI-isAI Workshops.
[12] K. Kukich. 1983. Design of a Knowledge-Based Report Generator. In ACL.
[13] Rémi Lebret, David Grangier, and Michael Auli. 2016. Neural Text Generation
from Structured Data with Application to the Biography Domain. In Proceedings
of the 2016 Conference on Empirical Methods in Natural Language Processing.
Association for Computational Linguistics, Austin, Texas, 1203–1213. https:
//doi.org/10.18653/v1/D16-1128
[14] Paul Ohm. 2009. Broken promises of privacy: Responding to the surprising failure
of anonymization. UCLA l. Rev. 57 (2009), 1701.
[15] Ankur Parikh, Xuezhi Wang, Sebastian Gehrmann, Manaal Faruqui, Bhuwan
Dhingra, Diyi Yang, and Dipanjan Das. 2020. ToTTo: A Controlled Table-To-Text
Generation Dataset. In Proceedings of the 2020 Conference on Empirical Methods in
Natural Language Processing (EMNLP). Association for Computational Linguistics,
Online, 1173–1186. https://doi.org/10.18653/v1/2020.emnlp-main.89
[16] Mithun Paul Panenghat, Sandeep Suntwal, Faiz Rafique, Rebecca Sharp, and
Mihai Surdeanu. 2020. Towards the Necessity for Debiasing Natural Language
Inference Datasets. In Proceedings of the 12th Language Resources and Evaluation
Conference. European Language Resources Association, Marseille, France, 6883–
6888. https://www.aclweb.org/anthology/2020.lrec-1.850
[17] J. Rabelo, Miyoung Kim, R. Goebel, M. Yoshioka, Yoshinobu Kano, and K. Satoh.
2019. A Summary of the COLIEE 2019 Competition. In JSAI-isAI Workshops.
[18] Ehud Reiter and R. Dale. 1997. Building applied natural language generation
systems. Nat. Lang. Eng. 3 (1997), 57–87.
[19] Sandrine Prom Tep, Florence Millerand, Alexandra Parada, Alexandra Bahary,
Pierre Noreau, and Anne-Marie Santorineos. 2019. Legal Information in Digital
Form: the Challenge of Accessing Computerized Court Records. IJR 8 (2019).
[20] Craig Thomson, Ehud Reiter, and Somayajulu Sripada. 2020. SportSett: Basketball
- A robust and maintainable dataset for Natural Language Generation. https:
//intellang.github.io/ IntelLanG : Intelligent Information Processing and Natural
Language Generation ; Conference date: 07-09-2020 Through 07-09-2020.
[21] Hongmin Wang. 2019. Revisiting Challenges in Data-to-Text Generation with
Fact Grounding. In Proceedings of the 12th International Conference on Natural
Language Generation. Association for Computational Linguistics, Tokyo, Japan,
311–322. https://doi.org/10.18653/v1/W19-8639
[22] Hannes Westermann, V. Walker, K. Ashley, and Karim Benyekhlef. 2019. Using
Factors to Predict and Analyze Landlord-Tenant Decisions to Increase Access
to Justice. Proceedings of the Seventeenth International Conference on Artificial
Intelligence and Law (2019).
[23] Sam Wiseman, Stuart Shieber, and Alexander Rush. 2017. Challenges in Data-
to-Document Generation. In Proceedings of the 2017 Conference on Empirical
Methods in Natural Language Processing. Association for Computational Linguis-
tics, Copenhagen, Denmark, 2253–2263. https://doi.org/10.18653/v1/D17-1239
[24] Chaojun Xiao, Haoxi Zhong, Z. Guo, Cunchao Tu, Zhiyuan Liu, M. Sun, Yansong
Feng, Xianpei Han, Z. Hu, Heng Wang, and J. Xu. 2018. CAIL2018: A Large-Scale
Legal Dataset for Judgment Prediction. ArXiv abs/1807.02478 (2018).
[25] Hai Ye, Xin Jiang, Zhunchen Luo, and W. Chao. 2018. Interpretable Charge
Predictions for Criminal Cases: Learning to Generate Court Views from Fact
Descriptions. ArXiv abs/1802.08504 (2018).
 Anonymization of German Legal Court Rulings
Ingo Glaser Tom Schamberger
Technical University of Munich Technical University of Munich
Garching bei München, Germany Garching bei München, Germany
ingo.glaser@tum.de tom.schamberger@tum.de
ABSTRACT
In the legal domain, many legal documents such as court decisions
and contracts are regularly anonymized. This process requires text
sequences with high sensitivity to be identified and neutralized
to secure sensitive information from third parties. Usually, this
process is performed manually by trained employees. Therefore,
anonymization is generally considered an expensive and inefficient
process. This work proposes a machine learning approach for the
automatic identification of sensitive text elements in German le-
gal court decisions and provides an implementation. For this task,
different deep neural network architectures based on generally pre-
trained contextual embeddings as well as trained word embeddings
are evaluated. Because of the lack of non-anonymized data sets, an
approach to create pseudonymized data sets is proposed as well.
CCS CONCEPTS
• Applied computing → Law; • Information systems → Expert
systems; • Computing methodologies → Neural networks.
ACM Reference Format:
Ingo Glaser, Tom Schamberger, and Florian Matthes. 2021. Anonymization
of German Legal Court Rulings. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.3466087
1 INTRODUCTION
In the age of digitization, an increasing amount of documents is cre-
ated in a digital manner, but more importantly also made available
online. As a result, personal information must be removed from
such documents. This applies particularly to the legal domain. The
rising of legal technology is highlighted by the increasing num-
ber of digitized legal documents, in particular legal contracts and
court decisions [9]. Legal documents such as contracts or court de-
cisions are regularly being anonymized, in order to be published or
handed out to third parties. Usually, such documents are manually
anonymized by skilled employees. This is crucial, as the detection
of personal information depends on the context and cannot be
considered as trivial. For this reason, many organizations consider
anonymization an expensive and inefficient process.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466087
205
Florian Matthes
Technical University of Munich
Garching bei München, Germany
florian.matthes@tum.de
In recent years, this has led to a scarcity of publicly available
legal datasets. On the other side, court decisions in particular repre-
sent vital base material for legal professions, academic researchers,
journalists and private companies [12]. Especially lawyers and law
firms require those documents for case research, because cases
are usually evaluated by means of related cases. Furthermore, the
lack of publicly available datasets inhibits legal innovation utilizing
machine learning or big data techniques.
As a result in this paper, after describing the process of anonymiza-
tion in the German legal domain, we propose an approach to auto-
matically anonymize German legal court rulings.
2 RELATED WORK
Classical anonymization approaches, such as [11] and [10], make
use of rule-based named entity recognition (NER) systems that
rely on dictionary look-ups and regular expressions to classify
individual words. On one hand, rule-based methods usually suffer
from weak robustness against rare occurrences and outliers such as
spelling mistakes as well as words that lie outside of the vocabulary,
because manually defined rules usually are too simple to match the
high variety of different named entity representations. On the other
hand, less data is necessary compared to ML-based NER systems [2].
Modern approaches make use of ML-based named entity recog-
nition for anonymization [2]. Current ML-based NER systems use
recurrent neural networks (RNNs) like stacked Long Short-Term
Memory (LSTM) RNNs [3] in combination with conditional ran-
dom fields (CRFs), in order to classify tokens in the IOB (Inside,
Outside, Beginning) tagging scheme [6]. This tagging scheme is
used, since named entities may span over multiple tokens. Thereby,
the CRF guarantees compliance with the scheme guidelines. Due
to the small amounts of publicly available NER training datasets,
state-of-theart NER architectures make use of pre-trained word
embeddings such as GloVe [6, 8].
Applied to the legal domain, the downside of all related methods
presented is that all detected entities are considered to be sensitive.
As a result, insensitive information is unnecessarily neutralized
as long as it contains any entity. In legal documents, entities such
as dates and locations must also be considered, because they may
reveal sensitive information if combined with additional informa-
tion from the document. Because many dates and locations are
insensitive, but yet essential to understand the meaning of the text,
the above methods would discard vital information.
3 ANONYMIZATION
Anonymization is defined as the task of identifying and neutraliz-
ing sensitive references within a given document or a set of docu-
ments [7]. Sensitivity is a binary measure determining whether or
not a particular reference, if publicly disclosed, might potentially
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
cause harm or engender undesirable personal or legal repercus-
sions [7].
In order to anonymize documents, sensitive information has to be
identified first, before the detected entities can be pseudonymized
or neutralized. According to the definition of sensitivity, this infor-
mation must contain at least one direct reference to an object or a
juristic person outside the context of the document. In order to pre-
serve the meaning of the text, it is necessary to replace references
referring to the same real object or person in such a way that the
connection within the text stays intact.
The most challenging aspect of anonymization is the identifica-
tion step. In this work, the anonymization problem is reduced to a
text sequence classification task referred to as contextual sensitivity
prediction. It is not sufficient to only detect references, because
in legal documents such as court decisions, many references are
insensitive. Instead, the sensitivity of references in legal documents
is assumed to mainly depend on the textual context. For instance,
court names reference real-world objects, but they are insensitive
and contain useful information for reviewers. On the other hand,
references to expert witnesses must not be exposed as they are
highly sensitive.
In order to underline the importance of supporting the anonymiza-
tion of German legal court rulings, we conducted expert interviews.
In knowledge-intensive and thus highly individual domains such
as the legal domain, it is indispensable to understand the current,
non digital, process at first. Therefore, we also used the interviews
to capture the current state of the anonymization.
We contacted 18 courts with a request for a respective expert
interview. Courts from various jurisdictions were deliberately in-
cluded. Eight courts agreed to support our research and provided
appropriate contacts. These were one financial court, one social
court, one administrative court, and five courts from the ordinary
jurisdiction. The latter ones were distributed at all instances (district
court to higher regional court). Furthermore, we also interviewed
the Bavarian Ministry of Justice.
One of the major conclusions which emerged during the inter-
views was the fact, that only two courts actually utilize standardized
guidelines for the anonymization process. It is important to note
that the other courts do not, of course, anonymize indiscriminately,
but rely much more on subject matter expertise. Therefore, it was
important in these cases to understand the exact processes involved
in anonymization, particularly the crucial entity types.
Another interesting aspect, which bases upon the fact of non-
existing guidelines, is the fact that these entities vary not only
between the different jurisdictions, but also within the ordinary
courts. While it seems obvious that decisions of different court
types require other entity types to be anonymized, it may not
be obvious at first glance why this applies even within ordinary
courts. However, such courts consist of multiple senates handling
different topics. In general, it is important that a judgment is still
understandable after anonymization and, above all, that the judicial
decision is comprehensible. We therefore hypothesized that an ML
model must be specifically trained for the jurisdiction at hand and
cannot be utilized as a general purpose model.
The replacement of sensitive information is a crucial part of the
anonymization method. Therefore, it was important to understand
the used methods in practice. However, it turned out that, while
206
Glaser et al.
the applied methods differ, the actual usage is flexible. While the
courts use different software to manage trials and store the actual
documents, including court rulings, all courts utilize Microsoft
Word for the actual anonymization. In fact, the "tool" at hand is
the search-and-replace functionality. This makes us belief that an
approach, as proposed in this paper, will be highly beneficial.
4 METHODOLOGY
4.1 Anonymization of German Legal Court
Rulings
We introduce a anonymization method for court rulings solely
trained on anonymized data. This approach is based on the idea
that the sensitivity of a entity does not depend on the specific entity
itself, but on its type and context, only. In this paper, we developed
a machine learning approach, which classifies the sensitivity of
entities in legal documents based on their context. These entities
have been pre-selected by general purpose NER.
The dataset used for model training consists of 1400 German
anonymized court rulings of the state court in Munich Landgericht
Munich that have been published in recent years. 1 In each ruling,
each sensitive reference has previously been removed and replaced
with a placeholder for publication at the courts. So, the anonymized
entities, i.e. placeholders, could be detected using a rule-based algo-
rithm, as discussed below.
Table 1 summarizes the most important information about the
pre-processed training corpus without the test set of 180 documents.
In sum, those documents contain about 35.000 anonymization place-
holders that are positively labeled for model training. However,
due to the large variety of different reference types and different
token lengths, the amount of data is considerably small. This espe-
cially impacts rare references like authorizations and bank accounts.
Anonymization mistakes in documents aggravate this issue even
further.
Document count 1.220
Word count 4.181.266
Placeholder count 33.779
Average tokens per word 1.8
Table 1: Information summary of the pre-processed training
corpus
The placeholder detection is a rule-based classification algorithm,
which takes paragraphs of anonymized legal documents and labels
the anonymization placeholders within the paragraph. The algo-
rithm scans words using a sliding window of 3 consecutive text
elements. Each triplet of text elements consists of a predecessor, a
anonymization candidate and a successor. Two different types of
placeholders are distinguished using regular expressions: Obvious
and potential placeholders. Obvious placeholders e.g. "Xxxx" meet
strong criteria and represent placeholders that are specially marked
by the author, e.g. ’"E."’. Potential placeholders may be interpreted
as placeholders if viewed outside the context, but may alternatively
possess one of the following meanings: (1) omission within cites (e.g.
testimonies), (2) abbreviation (e.g. "i.d.R."), (3) reference to pages or
1 Source: https://www.gesetze-bayern.de
Anonymization of German Legal Court Rulings
Annotated Tokens
Benutzer der klägerischen Internetseite
Masked Token
Benutzer [MASK] klägerischen Internetseite
BERT
Bi-LSTM Layer
Weighted Binary Cross Entropy Loss
Prediction
A nA nA
Figure 1: Anonymization Model Architecture
appendices (e.g. "siehe Anhang A 34"), and (4) reference to laws or
regulations (e.g. "§ 8 Abs. 3 Ziffer 2 UWG").
In order to overcome the limitation that anonymized document
entities are latent to the model during training, we chose a state-
of-the-art pre-trained masked language model called Bidirectional
Encoder Representations from Transformer (BERT) [1] as the core
model architecture. Tokenization has been done using the Senten-
cePiece tokenizer 2 and partitioned into paragraphs of at most 512
tokens. Then, the previously detected anonymized entities together
with randomly chosen alternative passages are masked using spe-
cial tokens. The masked paragraphs are fed into a pre-trained BERT
model [1]. This yields high-dimensional embeddings. Subsequently,
a bidirectional LSTM layer is used to classify corresponding embed-
ding vectors, whether they belong to the anonymized passages or
the randomized ones. For that, a fully connected layer is applied
to each output vector of the LSTM layer [4], which yields a final
single output per input token. Finally, we use a weighted binary
cross entropy loss function to transform the one-dimensional input
to a binary classification per input token. Figure 1 visualizes this ar-
chitecture using a simple example. For each document part (in this
case each word), the model classifies its sensitivity (A = sensitive,
nA = insensitive).
In the end, only classifications for masked tokens are considered,
since non-masked tokens are never sensitive. Therefore, in order to
apply the model to original data, candidate tokens must be chosen
in the documents to be masked before being fed into the model.
Otherwise, no tokens in the document would have been masked,
so that each token would be classified insensitive. Thereby, the
number of candidates is restricted, because masked tokens hide
original information so that the model has less data available, upon
which a classification can be made.
Table 2 shows the different variants of this architecture which
were evaluated, along with number of layer and weights. The num-
bers denote the parameter count for each layer - ’d’, ’c’ and ’r’ mean
’dense’, ’convolutional’ and ’biLSTM’ layers accordingly.
Variant Layer 1 Layer 2 Layer 3 Layer 4 Layer 5
RNN1 r256 r256 d128 d64 - terpart are exactly the sensitive entities that has to be excluded
RNN2 r128 r128 r128 d128 d64
RNN3 r512 r512 d256 d128 d64
Table 2: Model architecture variants on top of BERT embed-
dings
2 SentencePiece tokenization: https://github.com/google/sentencepiece
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
The model was implemented using the Tensorflow framework
ANONYM können
using Python and the NumPy library. For model training, we used
the Adam optimizer [5] with a decaying learning rate. The training
[MASK] können
dataset was split into a training set of 1220 documents and a test
set of the remaining 180 documents. Then, the model was trained
for 4 epochs, i.e. the whole training dataset has been completely
iterated 4 times. Beyond the fourth epoch, the models performance
on the test set decreased, hence reaching over-fitting.
nA A nA
For application to original documents, we used the general pur-
pose named entity recognizer from the SpaCy framework, which
detects the entity types person, organization, miscellaneous and lo-
cation in documents so that they can be masked for classification.
Because the entity types date, website and number were not sup-
ported, we extended the general named entity recognizer with a
rule-based algorithm that supports the additional named entities.
The performance of this algorithm has been manually evaluated
using randomly extracted text passages from the dataset, which
contain different instances of each entity type. Using the extended
named entity recognizer, entities in original documents are marked
as candidates and masked before being classified by the model.
The automated anonymization process needs to be accessible
to courts through a simple user interface, in order to assess the
practicality and quality of the approach. Therefore, we developed a
web interface, which allows the direct upload of court rulings (in
Microsoft Word DocX format) by incorporated staff at the courts to
a locally deployed server. On the server side, the extended named
entity recognition labels and masks the candidate entities, before
the document is applied to the model and the resulting classification
is sent to the user. A screenshot of the classification result shown
to the user is available on GitHub 3 . Finally, the document can be
downloaded as a Word document with sensitively classified text
passages replaced with customisable placeholders.
4.2 Pseudonymization of German Legal Court
Rulings
The performance of anonymization of legal court rulings has been
limited by the availability of original legal datasets. One way to
overcome this limitation is to utilize model training methods, which
do not rely on original data. Another way is to provide courts with a
convenient way to locally pre-process original data at the courts, so
that privacy and data protection issue do not play a role anymore.
Therefore next to our anonymization method purely based on
non-anonymized data, we introduce an automatic pseudonymiza-
tion method that can be applied locally at courts to pre-process pairs
of original and anonymized court rulings. This method yields to pro-
duce pairs of pseudonomized/anonymized documents, which con-
tain no sensitive information and can be used to train anonymiza-
tion models that usually require original and anonymized document
pairs. The main idea behind this process is that the difference be-
tween manually anonymized court rulings and their original coun-
from the original document in order to leave the court. If the type
of those entities is already determined at the courts and included
with the published documents, the removed entities can later be
replaced by randomized instances by the same entity type. This
3 https://github.com/sebischair/verlyze-anonymization
207
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Glaser et al.

Entity type Example Method Model Evaluation set Precision Recall

date 15.09.2012 rule-based Variant
website www.max.de rule-based RNN1 Munich district court 68.9% 79.1%
email max@email.de rule-based RNN2 Munich district court 68.2% 72.4%
street Birketweg 12 rule-based RNN3 Munich district court 68.4% 74.4%
district 80000 München rule-based RNN1 Munich financial court 64.7% 54.6%
number DE 0123 4567 rule-based RNN2 Munich financial court 62.8% 52.3%
amount 15.6 m rule-based RNN3 Munich financial court 63.1% 53.9%
person Max Müller SpaCy NER Table 4: Evaluation results of the anonymization procedure
location München SpaCy NER
organization Futter GmbH SpaCy NER
were returned to the court for inspection before being utilized in
miscellaneous Haribo SpaCy NER
our evaluation. Because the anonymization decision is assumed to
unknown undetected
be consistent for different entities of the same type, the resulting
Table 3: Extended named entity types sorted by priority (de- dataset of 123 unique anonymized entities (358 in total) is consid-
scending) ered to closely resemble original court ruling document bodies for
results in pseudonymized legal datasets that closely resemble the evaluation purposes. Nevertheless, because the document origi-
original data without sacrificing data protection. nated from the published format, original headers and footers of
As a first step, the original and anonymized documents pair is documents were missing and the format has been changed slightly
divided into tokens. Then a NER extracts entities in both documents. by the publication process.
If an entity is present only in the original document, it is assumed to The second evaluation dataset (Munich financial court consists
be sensitive and is replaced by a placeholder with a type hint. After of 33 original court rulings of the financial court in Munich. Each
all sensitive entities have been replaced, the remaining tokens of original document includes original header and footer and has been
the original document are compared to the anonymized document provided with its anonymized counterpart, together containing 122
tokens in order to find the entities undetected by the named entity unique anonymized entities (329 in total).
recognizer. Hence, the remaining entities are replaced by placehold-
ers with the type annotation unknown. Because those unknown 5.2 Anonymization Based on Contextual
entities can usually be considered to occur rarely for advanced NER, Information
those exceptional types can still be manually inferred from their The anonymization approach has been evaluated on both evaluation
context. datasets. The precision and recall metrics presented in Table 4 refer
In order to infer the type of the entities being detected in the to the classification of entities as (in-)sensitive. Thereby, the recall
difference between anonymized and original documents, a NER is metric is considered more important than the precision metric,
used to find named entities in both documents. Because general pur- since, in practice, the anonymization of an insensitive entity yields
pose NER only support very limited entity types such as person and lesser harm than the revelation of a sensitive entity.
location, we extended the NER similar to the extension described in It is important to underline the fact that the type II error is only
4.1 but with different and more supported types. Because the types partially introduced by the model, but also by the named entity
are not always distinct, the type of an entity is chosen according to recognizer pre-selecting the entities to be classified. The NER misses
the priorities as listed in table 3. 19.8 % of sensitive entities in the Munich district court dataset, while
The pseudonymization process is provided as a web interface4 . missing 28.5 % in the financial court dataset. This strictly limits
Compressed directories can be uploaded with document pairs as- the model performance, since detection during pre-processing is a
signed to each other by naming convention. Then the document precondition to be masked and possibly classified.
pairs are processed and the final list of pseudonymized documents While we have trained and evaluated many different approaches,
is shown. Court employees can either randomly review the docu- we report here only on the best performing models. Particularly,
ments or show each of the resulting documents to ensure that no the reported models with a NER-based candidate selection out-
sensitive information is still present. Finally, the pseudonymized performed other approaches. Without NER-based candidates, the
documents paired with the corresponding anonymized documents models have to overcome the "validation-test gap". In the train-
can be downloaded. ing dataset, the anonymization placeholders are masked during
training, because the real references behind the placeholder are
5 EVALUATION unknown. In this paper, this issue is referred to as the "validation-
5.1 Evaluation Data test gap", because the results on the validation dataset (masked
placeholders) during training have partly been much better than
Based on court rulings of the district court of Bavaria Landgericht
on the test set (manually replaced placeholders) during evaluation.
München published on the Bavarian legal platform ’Recht.Bayern’,
This is mitigated by our NER-based methods.
we created an evaluation set of 13 document pairs by manual re-
As shown in Table 4, the variant RNN1 outperformed the larger
placement of anonymized document parts with randomized named
models by a small margin. However, all models yield much more
entities of matching type (Munich District Court). These documents
promising results on the court rulings from the district court than
4 https://github.com/sebischair/verlyze-pseudonymization on the decisions of the financial court. This can be explained by

208
Anonymization of German Legal Court Rulings
the high variety of different anonymization standards practiced by
different types of courts. Entities such as financials or process dates
are often classified insensitive by district court, but frequently sen-
sitive by financial court. Due to lack of published financial rulings,
the training corpus has been mainly composed of more general
rulings, which explains the drop in evaluation performance ( recall).
During interviews at the courts, the anonymization web tool
was demonstrated and results from randomly sampled court rulings
were presented to court employees, being specialized in the manual
anonymization of court documents. The feedback has been consis-
tently positive, since most employees considered the anonymization
of legal documents as an unpleasant task. Nevertheless, most in-
terviewees criticized that the error rates were not high enough
to enable unsupervised automation. A small portion of the inter-
viewed workers also faulted that no manual configuration of the
algorithm could be done, after the model had been trained. How-
ever, this seems to be a rather less important concern as denoted
by the court director of the same court.
One important module of the anonymization approach presented
is the detection of placeholders in the anonymized training data.
Poor performance of this rule-based algorithm may lead to low
model performance. The algorithm has been evaluated using the
Munich district court dataset as used for the model evaluation. The
placeholder detection module performed with a precision of 99.9
%, a recall of 98.0 % and an accuracy of 99.9 %. These results show
that rule-based algorithms as described in Section 4.1 are capa-
ble of delivering sufficient performance to be used to pre-process
anonymized legal data for anonymization models.
5.3 Pseudonymization
The pseudonymization procedure was evaluated using the Munich
financial court rulings, since the contained documents have been
in the original format including headers and footers. The evalua-
tion revealed that all sensitive entities were filtered as intended
(100 %). Thereby, the specialized named entity recognizer correctly
recognized 72.8 % of the types of sensitive entities, which enables
automatic pseudonymization for a majority of entities.
Considering that the remaining entity types may be manually re-
covered using the anonymized documents, the proposed pseudonymiza-
tion approach provides a considerable basis for the creation of large
datasets, resembling original legal data without the necessity for
privacy considerations. However, the ratio of correctly identified
entity types can still be further improved by more advanced models.
The pseudonymization web tool has been demonstrated dur-
ing interviews at the courts with samples of court rulings from
the financial court in Munich. The samples had manually been
pseudonymized already, because the rulings contained original
data with us. The feedback was generally very positive and courts
finally agreed to share data pseudonymized using this tool. Most
court employees found the tool to be simple to work with and ap-
preciated the feature to upload whole archives instead of favoring
individual document uploads, which requires more manual work.
6 CONCLUSION & OUTLOOK
After identifying the problem of anonymization of legal documents,
we verified that automatic legal anonymization is a highly desirable
209
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
tool in order to increase the quantity of German legal data sets
being publicly available by means of expert interviews.
For this purpose, multiple different deep learning architectures
were trained using state-of-the-art generally pre-trained contextual
embeddings. Furthermore, a rule-based placeholder detection algo-
rithm was developed and validated, in order to label anonymization
placeholders in anonymized legal documents. Due to the difference
between training and evaluation data, the "validation-test gap" issue
has been introduced, which has caused a drop in model performance
on the non-anonymized test set. This issue was resolved using reg-
ularization methods such as input masking and dropout layers. The
models were evaluated on a manually created test document cor-
pus. Thereby, we found out that purely contextual classification
cannot distinguish between named entities and entities that refer to
named entities within the document. No model reached both, high
recall and high precision, metrics on the direct sequence classifi-
cation task. Nevertheless, in combination with a generally trained
NER model, the feature-based BERT tuning approach using stacked
biLSTM-RNN delivered promising results, but a specialized NER
model supporting more reference types is required.
In order to achieve high performances on the task of anonymiza-
tion, it is inevitable to utilize original and non-anonymized training
data. Hence, we also proposed a system to automatically pseudonymize
legal court ruling datasets that produces pairs of original and
anonymized court rulings.
To conclude, the anonymization of German legal documents re-
mains a complex problem and more data is required in order to build
fully autonomous anonymization systems. Nonetheless, contextual
sensitivity classification represents an important foundation for
future anonymization systems.
REFERENCES
[1] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding.
https://arxiv.org/abs/1810.04805
[2] Francisco Manuel Carvalho Dias. 2016. Multilingual automated text anonymiza-
tion. Instituto Superior Técnico of Lisboa (2016).
[3] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
[4] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long Short-term Memory. https:
//www.researchgate.net/publication/13853244_Long_Short-term_Memory
[5] Diederik Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Optimiza-
tion. https://arxiv.org/pdf/1412.6980.pdf
[6] Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami,
and Chris Dyer. 2016. Neural architectures for named entity recognition. arXiv
preprint arXiv:1603.01360 (2016).
[7] Ben Medlock. 2006. An Introduction to NLP-based Textual Anonymisation.. In
LREC. Citeseer, 1051–1056.
[8] Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove:
Global vectors for word representation. In Proceedings of the 2014 conference on
empirical methods in natural language processing (EMNLP). 1532–1543.
[9] Manavalan Saravanan, Balaraman Ravindran, and Shivani Raman. 2009. Im-
proving legal information retrieval using an ontological framework. Artificial
Intelligence and Law 17, 2 (2009), 101–124.
[10] Latanya Sweeney. 1996. Replacing personally-identifying information in medical
records, the Scrub system.. In Proceedings of the AMIA annual fall symposium.
American Medical Informatics Association, 333.
[11] Amund Tveit, Ole Edsberg, TB Rost, Arild Faxvaag, O Nytro, T Nordgard, Mar-
tin Thorsen Ranang, and Anders Grimsmo. 2004. Anonymization of general
practioner medical records. In Proceedings of the second HelsIT Conference.
[12] Marc van Opijnen, Ginevra Peruginelli, Eleni Kefali, and Monica Palmirani. 2017.
On-line publication of court decisions in the eu: Report of the policy group of the
project ‘building on the european case law identifier’. Available at SSRN 3088495
(2017).
Enhancing a Recidivism Prediction Tool With Machine Learning:
Effectiveness and Algorithmic Fairness
Marzieh Karimi-Haghighi
Universitat Pompeu Fabra
marzieh.karimihaghighi@upf.edu
ABSTRACT
This paper addresses a key application of Machine Learning (ML)
in the legal domain, studying how ML may be used to increase the
effectiveness of a criminal recidivism risk assessment tool named
RisCanvi, without introducing undue biases. The two key dimen-
sions of this analysis are predictive accuracy and algorithmic fair-
ness. ML-based prediction models obtained in this study are more ac-
curate at predicting criminal recidivism than the manually-created
formula used in RisCanvi, achieving an AUC of 0.76 and 0.73 in pre-
dicting violent and general recidivism respectively. However, the
improvements are small, and it is noticed that algorithmic discrimi-
nation can easily be introduced between groups such as national vs
foreigner, or young vs old. It is described how effectiveness and al-
gorithmic fairness objectives can be balanced, applying a method in
which a single error disparity in terms of generalized false positive
rate is minimized, while calibration is maintained across groups.
Obtained results show that this bias mitigation procedure can sub-
stantially reduce generalized false positive rate disparities across
multiple groups. Based on these results, it is proposed that ML-
based criminal recidivism risk prediction should not be introduced
without applying algorithmic bias mitigation procedures.
CCS CONCEPTS
• Computing methodologies → Supervised learning by clas-
sification.
KEYWORDS
criminal recidivism, risk assessment, algorithmic fairness
ACM Reference Format:
Marzieh Karimi-Haghighi and Carlos Castillo. 2021. Enhancing a Recidivism
Prediction Tool With Machine Learning: Effectiveness and Algorithmic
Fairness. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY,
USA, 5 pages. https://doi.org/10.1145/3462757.3466150
1 INTRODUCTION
Risk assessment is a necessary process in many important decisions
such as public health, information security, project management,
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466150
210
Carlos Castillo
Universitat Pompeu Fabra
carlos.castillo@upf.edu
auditing, and criminal justice. Since the 1920s, violence risk as-
sessment tools have been progressively used in criminal justice by
probation and parole officers, police, and psychologists to assess
the risk of harm, sexual, criminal, and violent offending in more
than 44 countries [22, 32]. In comparison to traditional prediction
methods and unstructured clinical judgments, risk assessment tools
offer superior accuracy and performance [18]. In this regard, factors
such as the availability of large databases, inexpensive computing
power, and developments in statistics and computer science have
brought an increase in the accuracy and applicability of these struc-
tured tools [3]. Such advances have effectively increased the use
of tools based on Machine Learning (ML) in criminal justice deci-
sions for risk forecasting [4, 7, 8]. Today, various semi-structured
protocols for assessing risk of recidivism can be found in different
countries including the U.S. [16], the U.K. [21], Canada [24], Aus-
tria [30], and Germany [13]. In Spain, among current violence risk
assessment tools including SAVRY, PCL-R, HCR-20, SVR-20, and
SARA, RisCanvi is a relatively new tool for risk assessment of recidi-
vism. It was originally developed in 2009 in response to concerns of
Catalan prison system officials regarding violent recidivism among
offenders after their sentences.
Research contribution. In this study, the effectiveness and algo-
rithmic fairness of RisCanvi risk assessment tool are evaluated in
comparison to ML models such as logistic regression, perceptron,
and support-vector machines, in violent and general recidivism
prediction. The effectiveness of the ML models are evaluated and
compared to RisCanvi in terms of various metrics including AUC,
Generalized False Positive (GFPR), and Generalized False Negative
(GFNR). Also, potential algorithmic bias introduced by the ML meth-
ods is evaluated in both violent and general recidivism prediction.
Given that model learning may lead to unfairness [11, 12, 34], the
impact of the obtained ML models is compared along nationality
(national origin vs foreign origin) and age (young vs old). Then
some differences are addressed through a mitigation procedure [29],
which try to equalize GFPR across nationality and age groups while
preserving the calibration in each group.
The rest of this paper is organized as follows. Section 2 outlines
related work. In Section 3, the RisCanvi risk assessment tool and the
dataset used in this study are described. The methodology including
the ML models and algorithmic fairness analysis are presented in
Section 4. Results are given in Section 5, and a procedure to mitigate
algorithmic discrimination is used in Section 6. Finally, the results
are discussed and the paper is concluded in Section 7.
2 RELATED WORK
The introduction of algorithms for risk assessment in criminal
justice is a controversial topic, and perhaps one that has motivated
a great deal of research on algorithmic fairness.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
In seminal research done by investigative journalism organiza-
tion ProPublica [2, 25] it was concluded that a widely-used program
named Correctional Offender Management Profiles for Alternative
Sanctions (COMPAS) is biased against African American defendants.
A follow-up study [19] found that COMPAS outcomes systemat-
ically over-predict risk for women, thereby indicating systemic
gender bias. However, the findings of the ProPublica study were
rejected by Northpointe (COMPAS developer), claiming their al-
gorithm is fair because it is well calibrated [17]. Moreover, in this
report it is shown that the COMPAS risk scales exhibit accuracy
equity and predictive parity.
In contrast to the case of COMPAS, other studies have shown
that other risk assessment tools such as the Post Conviction Risk
Assessment (PCRA), the Structured Assessment of Violence Risk in
Youth (SAVRY) and the Youth Level of Service/Case Management
Inventory (YLS/CMI) do not exhibit racial bias in the recidivism
prediction [28, 33]. In a more recent study focused on SAVRY [26,
34], it is shown that although machine learning models could be
more accurate than the simple summation used to compute SAVRY
scores, they would introduce discrimination against some groups
of defendants.
There are many different definitions of algorithmic fairness [27],
some of which are incompatible with one another. It is impossible to
satisfy all of them simultaneously except in pathological cases (such
as a perfect classifier), and in general it is impossible to maximize
algorithmic fairness and accuracy at the same time [5, 6]. Hence,
there are necessary trade-offs between different metrics [6, 10, 23].
In this regard, some studies [20, 36, 37] try to mitigate potential
algorithmic discrimination by satisfying equalized odds or in other
words avoiding disparate mistreatment along different sensitive
groups. In addition, due to the importance of the calibration in
risk assessment tools [6, 17], some previous work has also tried to
minimize error disparity across groups while maintaining calibrated
probability estimates [29].
The most closely related previous work is Pleiss et al. [29], where
algorithmic bias in a machine learned risk assessment (COMPAS)
is minimized by equalizing generalized false positive rates along
different races, finding this equalization to be incompatible with
calibration. In contrast, in the work presented on this paper, we
start from an expert-based risk assessment method, which is not
machine learned, and propose a new machine learning model to
replace it, describing the effects of algorithmic bias mitigation on
both the original and the machine learned model. Additionally, we
find that in RisCanvi equalization along nationality and age groups
is not entirely incompatible with calibration.
3 RISCANVI DATASET
3.1 The RisCanvi Risk Assessment Tool
RisCanvi was introduced as a multi-level risk assessment protocol
for violence prevention in the Catalan prison system in 2009 [1].
This protocol is applied multiple times during an inmate’s period in
prison; the official recommendation is to do so every six months or
at the discretion of the case manager. RisCanvi is not a questionnaire.
Instead, each inmate is interviewed by professionals. In the original
RisCanvi protocol, risk is determined for each inmate relative to
four possible outcomes: self-directed violence, violence in the prison
211
Marzieh Karimi-Haghighi and Carlos Castillo
facilities, committing further violent offenses, and breaking prison
permits. A fifth risk score was introduced more recently for general
recidivism [31].
Two versions of the RisCanvi protocol were created, an abbre-
viated one of 10 items for screening (RisCanvi-S), and a complete
one of 43 items (RisCanvi-C). Risk items can be categorized into
five different categories: Criminal/Penitentiary, Biographical, Fam-
ily/Social, Clinical, Attitudes/ Personality. These items can also
be divided into static factors (such as “criminal history of family”
and “age of starting violent activity”) and dynamic factors (such
as “member of socially vulnerable groups” and “pro-criminal or
antisocial attitudes”).
3.2 Dataset
The anonymized dataset used on this research comprises 7,239 of-
fenders who first entered the prison between 1989 and 2012 and who
were evaluated with the RisCanvi protocol between 2010 and 2013.
Only offenders for which nationality information was recorded
were kept that comprises 2,634 offenders. The result population
was filtered in terms of their violent/general recidivism, freedom
and last RisCanvi evaluation dates considering the following con-
ditions: inmates who were released at most 9 months after their
last RisCanvi evaluation, and for which violent/general recidivism
(or its absence) was recorded at most two years after their release.
Finally, samples with the size of 2,027 (out of 2,634) were reached.
Among this population, 146 committed a violent offence (violent
recidivism) and 310 committed a violent or non-violent offence
(general recidivism) after being released. The data includes all of
the information for the two RisCanvi versions (RisCanvi-S and
RisCanvi-C). This study is focused on the RisCanvi-C protocol
which is the complete version done after RisCanvi-S and it consists
of more risk factors which results in three risk levels (low, medium,
and high).
3.3 Violent and General Recidivism
This work is focused on RisCanvi protocol to assess Violent Re-
cidivism (“REVI” in the RisCanvi manual) and General Recidivism
(“REGE” in the RisCanvi manual) risks in sentenced inmates. REVI
and REGE risks are outcomes predicted using two different sub-sets
of risk factors. REVI risk is obtained using 23 items out of the 43 risk
factors of the RisCanvi-C version plus two demographic features
(gender and nationality) and to compute REGE risk, 14 items (out of
43 risk factors of the RisCanvi-C version) are used. In RisCanvi-C,
each of the REVI and REGE scores has been computed by applying
the summation of their related features in a hand-crafted formula,
then using two cut-offs, obtaining three risk levels (details in [1]).
The distribution of REVI and REGE risk scores in the last RisCanvi
evaluation is compared by nationality and age groups. Grouping
by gender is not considered as the number of women in the sample
is too small to draw robust conclusions. The comparison shows
that recidivism risk scores have approximately similar distributions
along nationality and age group except for the REVI score in nation-
ality group which shows that foreigners tend to have lower REVI
risk scores compared to Spaniards (Figures are omitted for brevity).
For age groups, 30 years old is used as a cut-off, as criminology
research suggests that the types of offense and context are different
Enhancing a Recidivism Prediction Tool With Machine Learning: Effectiveness and Algorithmic Fairness
for people under 30 and over 30 (see, e.g., [35]). This age is also used
as a cut-off for young and old people in the design of the RisCanvi
protocol. In the present dataset, the majority of the population are
Spanish nationals (70%) and older than 30 years old (74%).
According to the average violent and general recidivism rates for
nationality and age groups, it can be seen that in general, foreigners
and older offenders have a lower recidivism rate.
4 METHODOLOGY
The goal of this study is to compare the effectiveness and fairness
of Machine Learning (ML) models and the RisCanvi risk assessment
tool in the prediction of violent and general recidivism.
4.1 ML-based Models
Different ML methods, such as logistic regression, multi-layer per-
ceptron (MLP), and support vector machines (SVM) are used. The
ground truth is the violent/general recidivism, which is recorded at
most two years after the inmate’s release.
Different sub-sets of features are tested as input to the ML mod-
els, such as 43 RisCanvi-C items, Violent Recidivism (REVI)/General
Recidivism (REGE) risk items, and a set of features selected from 43
risk items using a feature selection method. In addition, three demo-
graphic features (gender, nationality, and age) are used as general
input features. Finally, the average of REVI/REGE risk scores over
all of the RisCanvi evaluations from the first to the last evaluation
is added.
The split of the two sets is done k times using stratified k-fold
cross-validation, reporting average results.
4.2 Algorithmic Fairness
Algorithmic fairness is evaluated by comparing the impact of the
risk prediction method across nationality and age groups.
As it is known, model calibration is a necessary condition, espe-
cially in criminal justice risk assessments [6, 17]. If the risk tool is
not calibrated with respect to different groups, then the same risk es-
timate carries different meanings and cannot be interpreted equally
for different groups. Furthermore, creating parity in the error rates
of different groups (“equalized odds”) is a well-established method
to mitigate algorithmic discrimination in automatic classification.
Previous work has also emphasized the importance of this algo-
rithmic fairness metric for this particular application [20, 36, 37].
Hence, to mitigate potential algorithmic discrimination, a relaxation
method [29] is used in this paper which seeks to satisfy equalized
odds or parity in the error rates (generalized false positive rate and
generalized false negative rate) while preserving calibration in each
sub-group of nationality and age. In most cases, calibration and
equalized odds are mutually incompatible goals [10, 23], so in this
method it is sought to minimize only a single error disparity across
groups while maintaining calibration probability estimates.
Generalized False Positive Rate (GFPR) and Generalized False
Negative Rate (GFNR) are the standard notions of false-positive and
false-negative rates that are generalized for use with probabilistic
classifiers [29]. If variable 𝑥 represent an inmate’s features vector,
𝑦 indicates whether or not the inmate recidivists, 𝐺 1 , 𝐺 2 are the
two different groups, and ℎ 1 , ℎ 2 are binary classifiers which classify
samples from 𝐺 1 , 𝐺 2 respectively, GFPR and GFNR are defined as
212
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 1: Effectiveness of models in violent and general re-
cidivism prediction
Risk Violent Recidivism General Recidivism
Model AUC GFNR GFPR AUC GFNR GFPR
LR 0.76 0.82 0.06 0.73 0.75 0.14
RisCanvi_score 0.72 0.87 0.07 0.70 0.79 0.14
follows [29]: the GFPR of classifier ℎ𝑡 for group 𝐺𝑡 is 𝑐 𝑓 𝑝 (ℎ𝑡 ) =
E (𝑥,𝑦)∼𝐺𝑡 [ℎ𝑡 (𝑥) | 𝑦 = 0]. GFPR is the average probability of being
recidivist that the classifier estimates for people who actually do
not recidivate. Conversely, the GFNR of classifier ℎ𝑡 is 𝑐 𝑓 𝑛 (ℎ𝑡 ) =
E (𝑥,𝑦)∼𝐺𝑡 [(1−ℎ𝑡 (𝑥)) | 𝑦 = 1]. So the two classifiers ℎ 1 and ℎ 2 show
probabilistic equalized odds across groups 𝐺 1 and 𝐺 2 if 𝑐 𝑓 𝑝 (ℎ 1 ) =
𝑐 𝑓 𝑝 (ℎ 2 ) and 𝑐 𝑓 𝑛 (ℎ 1 ) = 𝑐 𝑓 𝑛 (ℎ 2 ).
Classifier ℎ𝑡 is said to be well-calibrated if ∀𝑝 ∈ [0, 1], P (𝑥,𝑦)∼𝐺𝑡
[𝑦 = 1 | ℎ𝑡 (𝑥) = 𝑝] = 𝑝. To prevent the probability scores from
carrying group-specific information, both classifiers ℎ 1 and ℎ 2 are
calibrated with respect to groups 𝐺 1 and 𝐺 2 [6, 17].
5 RESULTS
5.1 Effectiveness Evaluation
Among logistic regression (LR), multi-layer perceptron (MLP) and
support vector machines, the best results were obtained using LR
for both violent and general recidivism predictions. Hence, the non-
LR based models are omitted for brevity. The final set of features
used for the model consists of a sub-set of the 43 risk items of
the RisCanvi evaluation selected using a feature selection method
(based on a linear model with L1-based penalization to yield sparse
coefficients), the average Violent Recidivism (REVI)/General Recidi-
vism (REGE) score (from the first to the last RisCanvi evaluation),
gender, nationality, and age at the time of the last evaluation.
Results in terms of AUC-ROC, GFNR, and GFPR are presented
and compared with the existing RisCanvi protocol in Table 1 for
both violent and general recidivism prediction. These results are
compared against RisCanvi_score, which is a number resulting
from the application of the RisCanvi formula.
In both violent and general recidivism prediction, LR yields better
results than RisCanvi in terms of all metrics. However, the results
are close to RisCanvi. In general, the LR model is more accurate
than RisCanvi, although by a small amount, which is surprising
considering that RisCanvi was not computationally optimized for
predictive accuracy.
5.2 Algorithmic Fairness Evaluation
The results for the analysis of algorithmic fairness in all metrics
along nationality (national and foreigner), and age groups (young
and old inmates) are shown in Table 2 for violent and general
recidivism prediction. In the LR_calibrated model, the predictions
have been calibrated with respect to each of the two sub-groups in
nationality and age.
For violent recidivism, all models show a bias against nationals
in terms of GFPR. The difference is less noticeable in RisCanvi. In
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Marzieh Karimi-Haghighi and Carlos Castillo

Table 2: Effectiveness of models in violent and general recidivism prediction per group

Risk Violent Recidivism General Recidivism

Model LR LR_Calibrated RisCanvi LR LR_Calibrated RisCanvi
Group/Metrics AUC GFNR GFPR AUC GFNR GFPR AUC GFNR GFPR AUC GFNR GFPR AUC GFNR GFPR AUC GFNR GFPR
National 0.81 0.77 0.07 0.81 0.81 0.06 0.76 0.85 0.08 0.78 0.70 0.15 0.77 0.73 0.13 0.72 0.78 0.14
Foreigner 0.85 0.87 0.05 0.85 0.85 0.04 0.72 0.91 0.05 0.68 0.80 0.11 0.72 0.77 0.11 0.59 0.83 0.13
National
Foreigner (Ratio) (0.95) (0.88) (1.64) (0.95) (0.95) (1.50) (1.05) (0.93) (1.44) (1.14) (0.87) (1.30) (1.07) (0.95) (1.20) (1.22) (0.94) (1.08)
Young 0.84 0.78 0.08 0.84 0.83 0.06 0.79 0.86 0.07 0.67 0.74 0.17 0.72 0.75 0.15 0.58 0.82 0.14
Old 0.83 0.78 0.06 0.83 0.83 0.06 0.76 0.85 0.07 0.78 0.71 0.12 0.75 0.74 0.11 0.75 0.78 0.14
Young
Old (Ratio) (1.02) (1.00) (1.26) (1.02) (1.01) (1.11) (1.04) (1.00) (1.03) (0.85) (1.04) (1.38) (0.96) (1.01) (1.37) (0.77) (1.06) (1.03)

LR model, we can also observe higher GFPR for young inmates
compared to old offenders. In general, LR_calibrated and RisCanvi
models lead to more algorithmically fair results along both nation-
ality and age in terms of all metrics, except for the metrics in which
all the models show discrimination.
The results for general recidivism prediction show higher AUC
for nationals compared to foreigners in RisCanvi. In terms of GFPR,
the LR and LR_calibrated models show discrimination against na-
tional group. In age group, LR and LR_calibrated models show
higher GFPR along young compared to old group. In terms of AUC,
we can see more discrimination against young inmates in RisCanvi
compared to other models. As a result, LR_calibrated model shows
better algorithmic fairness properties across nationality and more
balanced values can be observed along age group in RisCanvi.
6 EQUALIZED ODDS AND CALIBRATION
In this section, it is tried to achieve parity along nationality and age
groups in terms of two fairness metrics simultaneously. For this
purpose, the method introduced by Pleiss et al. [29] is used that
seeks parity in Generalized False Positive Rate (GFPR) or Gener-
alized False Negative Rate (GFNR) while preserving calibration in
each sub-group of nationality and age. The conclusion from the
previous section based on the results obtained per group in Table 2,
is that in both violent and general recidivism predictions, machine
learning models show inequality in terms of GFPR along nationality
and age. RisCanvi also shows an imbalance in GFPR values along
nationality groups in violent recidivism prediction.
Hence, it is tried to create parity in this metric while preserving
calibration in each group. The results after bias mitigation is pre-
sented in Table 3 for violent and general recidivism prediction. The
obtained models are referred to in the following as LR-Equalized,
LR_Calibrated-Equalized, and RisCanvi-Equalized.
By comparing the results before and after this bias mitigation
(Table 2 and Table 3 respectively) in violent recidivism, it can be
seen that the discrimination in GFPR has decreased in the order
of 0.08-0.26 and 0.06-0.09 along nationality and age groups respec-
tively. Also, comparing the results before and after bias mitigation
in general recidivism shows that there are reductions in GFPR dis-
parity in the orders of 0.03-0.04 and 0.16-0.19 along nationality
and age groups respectively. However, in both violent and general
recidivism prediction, the decline in GFPR bias is obtained at the
expense of further inequalities in other metrics.
7 DISCUSSION AND CONCLUSIONS
The effectiveness and fairness of Machine Learning (ML) models
in violent and general recidivism prediction were compared to the
RisCanvi risk assessment tool, an in-use model created by experts.
ML models were generated with AUC of 0.76 and 0.73 in violent and
general recidivism prediction respectively which shows slightly
better results compared to the AUC of RisCanvi protocol which is
0.72 and 0.70. It is noteworthy that in this type of task, predictions
are not very accurate in general (existing recidivism prediction
tools typically have AUC in the range of 0.57-0.74 [9, 14, 15]), and
it is found that a hand-crafted formula created by experts is quite
comparable to a machine-learned one. Although the improvement
in accuracy by ML would be insufficient on its own to support its
introduction as a risk assessment tool, a key element of ML models
is their flexibility. An ML model can be re-trained with newer data,
and incorporate new factors as the population of inmates changes
and more data on recidivism becomes available.
By studying differential treatment of RisCanvi and ML models
across different groups, it can be stated that depending on the
desired metric and groups, machine learning and human expert
can lead to different but comparable results. An advantage of ML
models is that the emphasis on different metrics can be changed
during the modeling as legal or policy changes are introduced.
In this study, results in Table 2 showed that in both violent and
general recidivism predictions, there is an inequality in terms of
Generalized False Positive Rate (GFPR) metric along nationality
and age groups. So using a relaxation method [29], it was tried to
set parity in GFPR while preserving calibration in each sub-group
of nationality and age. The results after bias mitigation (in Table 3)
showed that GFPR disparity in violent and general recidivism has
been respectively decreased at most 0.26 and 0.04 along nationality
and 0.09 and 0.19 along age, however, in exchange for inequalities
in some other metrics.
A robust conclusion from this work is that in a context in which
predictive factors neither determine nor yield a clear signal of
low/medium/high recidivism risk, ML cannot be considered a silver
bullet. At the very least, improvements in accuracy need to be
carefully contrasted with potential issues of algorithmic fairness

213
Enhancing a Recidivism Prediction Tool With Machine Learning: Effectiveness and Algorithmic Fairness
Table 3: Equalized GFPR while preserving calibration in violent and general recidivism prediction
Risk Violent Recidivism
Model LR-Equalized LR_Calib-Equalized
Group/Metrics AUC GFNR GFPR AUC GFNR GFPR
National 0.81 0.77 0.07 0.81 0.81 0.06
Foreigner 0.64 0.92 0.05 0.61 0.91 0.05
National
Foreigner (Ratio) (1.27) (0.83) (1.38) (1.32) (0.89) (1.42)
Young 0.84 0.78 0.08 0.71 0.86 0.06
Old 0.62 0.88 0.07 0.60 0.89 0.06
Young
Old (Ratio) (1.36) (0.89) (1.17) (1.19) (0.97) (1.05)
when introducing ML, and calibration and some bias mitigation
method (such as equalized odds in this study) needs to be used.
ACKNOWLEDGMENTS
This work has been partially supported by the HUMAINT pro-
gramme (Human Behaviour and Machine Intelligence), Centre for
Advanced Studies, Joint Research Centre, and European Commis-
sion. The project leading to these results has received funding
from “la Caixa” Foundation (ID 100010434), under the agreement
LCF/PR/PR16/51110009.
REFERENCES
[1] Antonio Andrés-Pueyo, Karin Arbach-Lucioni, and Santiago Redondo. 2018. The
RisCanvi: a new tool for assessing risk for violence in prison and recidivism.
Recidivism Risk Assessment: A Handbook for Practitioners (2018), 255–268.
[2] Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner. 2016. Machine bias.
ProPublica, May 23 (2016), 2016.
[3] Richard Berk. 2012. Criminal justice forecasts of risk: A machine learning approach.
Springer Science & Business Media.
[4] Richard Berk. 2017. An impact assessment of machine learning risk forecasts
on parole board decisions and recidivism. J. of Experimental Criminology 13, 2
(2017), 193–216.
[5] Richard Berk. 2019. Accuracy and fairness for juvenile justice risk assessments.
J. of Empirical Legal Studies 16, 1 (2019), 175–194.
[6] Richard Berk, Hoda Heidari, Shahin Jabbari, Michael Kearns, and Aaron Roth.
2018. Fairness in criminal justice risk assessments: The state of the art. Sociological
Methods & Research (2018), 0049124118782533.
[7] Richard Berk and Jordan Hyatt. 2015. Machine learning forecasts of risk to inform
sentencing decisions. Federal Sentencing Reporter 27, 4 (2015), 222–228.
[8] Richard A Berk, Susan B Sorenson, and Geoffrey Barnes. 2016. Forecasting
domestic violence: A machine learning approach to help inform arraignment
decisions. J. of Empirical Legal Studies 13, 1 (2016), 94–115.
[9] Tim Brennan, William Dieterich, and Beate Ehret. 2009. Evaluating the predictive
validity of the COMPAS risk and needs assessment system. Criminal Justice and
Behavior 36, 1 (2009), 21–40.
[10] Alexandra Chouldechova. 2017. Fair prediction with disparate impact: A study
of bias in recidivism prediction instruments. Big data 5, 2 (2017), 153–163.
[11] Alexandra Chouldechova and Aaron Roth. 2018. The frontiers of fairness in
machine learning. arXiv preprint arXiv:1810.08810 (2018).
[12] S. Corbett-Davies and S. Goel. 2018. The measure and mismeasure of fairness: A
critical review of fair machine learning. arXiv preprint arXiv:1808.00023 (2018).
[13] K.P. Dahle, J. Biedermann, R.J. Lehmann, and F. Gallasch-Nemitz. 2014. The devel-
opment of the Crime Scene Behavior Risk measure for sexual offense recidivism.
Law and human behavior 38, 6 (2014), 569.
[14] Matthew DeMichele, Peter Baumgartner, Michael Wenger, Kelle Barrick, Megan
Comfort, and Shilpi Misra. 2018. The public safety assessment: A re-validation
and assessment of predictive utility and differential prediction by race and gender
in kentucky. Available at SSRN 3168452 (2018).
[15] S.L. Desmarais, K.L. Johnson, and J.P. Singh. 2016. Performance of recidivism
risk assessment instruments in US correctional settings. Psychological Services
13, 3 (2016), 206.
[16] Sarah Desmarais and Jay Singh. 2013. Risk assessment instruments validated
and implemented in correctional settings in the United States. (2013).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
General Recidivism
RisCanvi-Equalized LR-Equalized LR_Calib-Equalized
AUC GFNR GFPR AUC GFNR GFPR AUC GFNR GFPR
0.76 0.85 0.08 0.78 0.70 0.15 0.67 0.78 0.14
0.62 0.92 0.06 0.61 0.81 0.12 0.53 0.88 0.12
(1.23) (0.93) (1.28) (1.28) (0.86) (1.27) (1.26) (0.89) (1.16)
- - - 0.67 0.74 0.17 0.72 0.75 0.15
- - - 0.63 0.78 0.14 0.53 0.86 0.13
- - - (1.06) (0.95) (1.22) (1.35) (0.88) (1.18)
[17] W. Dieterich, C. Mendoza, and T. Brennan. 2016. COMPAS risk scales: Demon-
strating accuracy equity and predictive parity. Northpoint Inc (2016).
[18] William M Grove, David H Zald, Boyd S Lebow, Beth E Snitz, and Chad Nelson.
2000. Clinical versus mechanical prediction: a meta-analysis. Psychological
assessment 12, 1 (2000), 19.
[19] Melissa Hamilton. 2019. The sexist algorithm. Behavioral sciences & the law 37, 2
(2019), 145–157.
[20] M. Hardt, E. Price, and N. Srebro. 2016. Equality of opportunity in supervised
learning. In Advances in neural information processing systems. 3315–3323.
[21] Philip D Howard and Louise Dixon. 2012. The construction and validation of the
OASys Violence Predictor: Advancing violence risk assessment in the English and
Welsh correctional services. Criminal Justice and Behavior 39, 3 (2012), 287–307.
[22] Danielle Leah Kehl and Samuel Ari Kessler. 2017. Algorithms in the criminal
justice system: Assessing the use of risk assessments in sentencing. (2017).
[23] J. Kleinberg, S. Mullainathan, and M. Raghavan. 2016. Inherent trade-offs in the
fair determination of risk scores. arXiv preprint arXiv:1609.05807 (2016).
[24] Carolin Kröner, Cornelis Stadtland, Matthias Eidt, and Norbert Nedopil. 2007.
The validity of the Violence Risk Appraisal Guide (VRAG) in predicting criminal
recidivism. Criminal Behaviour and Mental Health 17, 2 (2007), 89–100.
[25] Jeff Larson, Surya Mattu, Lauren Kirchner, and Julia Angwin. 2016. How we
analyzed the COMPAS recidivism algorithm. ProPublica (5 2016) 9 (2016).
[26] Marius Miron, Songül Tolan, Emilia Gómez, and Carlos Castillo. 2020. Evaluating
causes of algorithmic bias in juvenile criminal recidivism. Artificial Intelligence
and Law (2020), 1–37.
[27] Arvind Narayanan. 2018. 21 fairness definitions and their politics. presenterad
på konferens om Fairness, Accountability, and Transparency 23 (2018).
[28] Rachael T Perrault, Gina M Vincent, and Laura S Guy. 2017. Are risk assess-
ments racially biased?: Field study of the SAVRY and YLS/CMI in probation.
Psychological assessment 29, 6 (2017), 664.
[29] Geoff Pleiss, Manish Raghavan, Felix Wu, Jon Kleinberg, and Kilian Q Weinberger.
2017. On fairness and calibration. In Advances in Neural Information Processing
Systems. 5680–5689.
[30] M. Rettenberger, M. Mönichweger, E. Buchelle, F. Schilling, and R. Eher. 2010. The
development of a screening scale for the prediction of violent offender recidivism.
Monatsschrift für Kriminologie und Strafrechtsreform 93, 5 (2010), 346–360.
[31] J.P. Singh, D.G. Kroner, J.S. Wormith, S.L. Desmarais, and Z. Hamilton. 2018.
Handbook of recidivism risk/needs assessment tools. John Wiley & Sons.
[32] Jay P Singh, Sarah L Desmarais, Cristina Hurducas, Karin Arbach-Lucioni, Car-
olina Condemarin, Kimberlie Dean, Michael Doyle, Jorge O Folino, Verónica
Godoy-Cervera, Martin Grann, et al. 2014. International perspectives on the
practical application of violence risk assessment: A global survey of 44 countries.
Int. J. of Forensic Mental Health 13, 3 (2014), 193–206.
[33] Jennifer L Skeem and Christopher T Lowenkamp. 2016. Risk, race, and recidivism:
Predictive bias and disparate impact. Criminology 54, 4 (2016), 680–712.
[34] Songül Tolan, Marius Miron, Emilia Gómez, and Carlos Castillo. [n.d.]. Why
Machine Learning May Lead to Unfairness: Evidence from Risk Assessment for
Juvenile Justice in Catalonia. In Proc. of ICAIL’19.
[35] Jeffrey Todd Ulmer and Darrell J Steffensmeier. 2014. The age and crime rela-
tionship: Social variation, social explanations. In The nurture versus biosocial
debate in criminology: On the origins of criminal behavior and criminality. SAGE
Publications Inc., 377–396.
[36] B. Woodworth, S. Gunasekar, M.I. Ohannessian, and N. Srebro. 2017. Learning
non-discriminatory predictors. arXiv preprint arXiv:1702.06081 (2017).
[37] Muhammad Bilal Zafar, Isabel Valera, Manuel Gomez Rodriguez, and Krishna P
Gummadi. 2017. Fairness beyond disparate treatment & disparate impact: Learn-
ing classification without disparate mistreatment. In Proc. of the 26th Int. Conf.
on WWW. 1171–1180.
214
 Towards compliance checking in reified I/O logic via SHACL
Livio Robaldo
livio.robaldo@swansea.ac.uk
Legal Innovation Lab Wales - Swansea University
Swansea, Wales, UK
ABSTRACT
Reified Input/Output logic [29] has been recently proposed to han-
dle natural language meaning in Input/Output logic [17]. So far,
the research in reified I/O logic has focused only on KR issues,
specifically on how to use the formalism for representing contex-
tual meaning of norms (see [28]). This paper is the first attempt to
investigate reasoning in reified I/O logic, specifically compliance
checking. This paper investigates how to model reified I/O logic for-
mulae in Shapes Constraint Language (SHACL) [2], a recent W3C
recommendation for validating and reasoning with RDFs/OWL.
KEYWORDS
reified I/O logic, SHACL, RDFs/OWL
ACM Reference Format:
Livio Robaldo. 2021. Towards compliance checking in reified I/O logic via
SHACL. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY,
USA, 5 pages. https://doi.org/10.1145/3462757.3466065
1 INTRODUCTION
Reified Input/Output logic [29] is Input/Output logic [17] enriched
with reification. The introduction of reification in I/O logic enhances
the expressivity of the I/O logic formulae without substantially
affecting the I/O logic constructs that implement deontic reasoning.
Reification is a formal mechanism that associates instantiations
of high-order predicates and operators with FOL terms [13], [27],
[26]. The latter can be then directly inserted as arguments of other
FOL predicates, which may be in turn reified again into new FOL
terms. In other words, reified I/O logic associates norms with ex-
plicit terms, e.g., constants or variables, and not only with truth-
conditional symbols such as predicates or (second-order) deontic
operators. These terms can be then inserted as parameter of sepa-
rated meta-properties.
Reified I/O logic is grounded on a specific reification-based ap-
proach for Natural Language Semantics: the framework in [12].
The main insight of [12] is to massively use reification in order to
transform every second-order operator, including boolean connec-
tives, into a FOL predicate applied to FOL terms. The final resulting
formulae are then flat conjunctions of atomic FOL predicates.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466065
215
As shown in [28], the formal simplicity and the modular structure
of reified I/O logic facilitate the implementation of user-friendly
interfaces to encode large knowledge bases of norms in reasonable
time. [28] presents the DAPRECO knowledge base (D-KB), a repos-
itory of 966 formulae in reified I/O formulae that translates norms
from the GDPR. The D-KB was built in four months via a special
JavaScript editor implemented to this purpose.
While past research in reified I/O logic has focused on how
building formulae associated with norms in natural language, this
paper represents the first attempt to investigate how these formulae
can be implemented and used for compliance checking, i.e., to infer
which obligations have been violated in a given state of affairs and
with respect to a given set of norms.
Compliance checking has never been really studied in I/O logic.
Most past literature in I/O logic has focused on deontic reasoning,
and, recently, normative reasoning [15].
Deontic reasoning is to reason about what is obligatory and
permitted, while dealing with contrary-to-duty reasoning, deontic
paradoxes, ethical/moral conflicts, etc. Reasoning about obligations
and permissions is of course orthogonal to what agents really do,
i.e., whether they did or did not violate their obligations or whether
they did or did not perform what they were permitted to do.
Compliance checking does not involve deontic reasoning. Still,
compliance checking could not be so simple to handle, e.g., because
norms might include exceptions that lead to defeasible reasoning.
This paper proposes a formalization of non-deontic inferences
in reified I/O logic via SHACL [2]. While recent literature offered
solutions for compliance checking implemented in RDFs/OWL, e.g.,
[6], only preliminary works use SHACL to this end, e.g., [21].
2 BACKGROUND - REIFIED I/O LOGIC
2.1 Input/Output logic
I/O logic was originally introduced in [17]. I/O logic is a family of
logics, just like modal logic is a family of systems K, S4, S5, etc.
However, while modal logic uses possible world semantics, I/O
logic uses norm-based semantics, in the sense of [11]: I/O systems
are families of if-then rules (𝑎, 𝑏), such that when 𝑎 is given in input,
𝑏 is returned in output. 𝑎 and 𝑏 are formulae in another logic, called
“the object logic”. It has been argued that norm-based reasoning
features some advantages over reasoning based on possible world
semantics, first of all a lower computational complexity [30].
I/O logic neatly decouples deontic and non-deontic inferences.
I/O logic is indeed a meta-logic wrapped around another logic (e.g.,
[12], in case of reified I/O logic) called “the object logic”. The meta-
logic implements deontic inferences while the object logic imple-
ments the non-deontic ones. In I/O systems for legal reasoning,
rules (𝑎, 𝑏) can be obligations, permissions, and constitutive rules.
These are clustered within three distinct sets 𝑂, 𝑃, and 𝐶 such that
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
∀(𝑎, 𝑏)∈𝑂 reads as “given 𝑎, 𝑏 is obligatory”, ∀(𝑎, 𝑏)∈𝑃 reads as “given
𝑎, 𝑏 is permitted”, and ∀(𝑎, 𝑏)∈𝐶 reads as “given 𝑎, 𝑏 holds”.
Most past research on I/O logic has focused on theoretical inves-
tigations in the meta-logic, for modeling deontic reasoning. Since
the focus was on studying the meta-logic, the object logic was al-
ways kept as simple as possible, i.e., it was always propositional
logic. Reified I/O logic is perhaps the most relevant proposal so far
in the I/O logic literature that employs an alternative (first-order)
object logic: the logical framework in [12].
In I/O logic, inferences in the meta-logic are achieved by impos-
ing axioms and constraints on the sets of if-then rules. Different
combinations of axioms and constraints trigger different inferences.
For instance, [17] defines the basic axioms in (1), where the
symbol ‘⊢’ is the entailment relation of the object logic. Variants of
these axioms have been further investigated in [23] and [22].
(1) • SI: from (𝑎, 𝑥) to (𝑏, 𝑥) whenever 𝑏 ⊢ 𝑎.
• OR: from (𝑎, 𝑥) and (𝑏, 𝑥) to (𝑎 ∨ 𝑏, 𝑥).
• WO: from (𝑎, 𝑥) to (𝑎, 𝑦) whenever 𝑥 ⊢ 𝑦.
• AND: from (𝑎, 𝑥) and (𝑎, 𝑦) to (𝑎, 𝑥 ∧ 𝑦).
• CT: from (𝑎, 𝑥) and (𝑎 ∧ 𝑥, 𝑦) to (𝑎, 𝑦).
By imposing axioms SI, WO, and AND, we obtain a specific derivation
system called deriv1 . Adding OR to deriv1 gives deriv2 . Adding
CT to deriv1 gives deriv3 . The five axioms together give deriv4 .
Each derivation system is sound and complete with respect to a
different (norm-based) semantics and can therefore trigger different
inferences (see [17] for further discussion and details).
Given a derivation system, we may further constrain its sets of
if-then rules, by considering only subsets that do not yield outputs
conflicting with given inputs. This is needed to handle contrary-to-
duty reasoning, i.e., to determine which obligations are detached
in a situation that already violates some among them [16].
This paper is not concerned with the meta-level of I/O logic.
Rather, it will focus on the object logic and non-deontic inferences,
including defeasible ones to handle exceptions in legal reasoning.
2.2 Adding reification to I/O logic
Reification is a well-known technique used in linguistics and com-
puter science for representing abstract concepts. These are associ-
ated with explicit objects, e.g., FOL terms (see below in this section)
or RDF resources (see §3 below), on which we can assert properties.
These assertions can be recursively reified again into new terms.
Both [12] and RDFs/OWL recursively reify assertions until the
knowledge is represented in terms of a flat list of atomic predi-
cates applied to terms. In RDFs/OWL, these flat lists are made of
triples “(subject, predicate, object)”, while [12] also allows
predicates with higher arity; however, any n-ary predicate can be
transformed into an equivalent conjunction of RDF triples.
In [12] and in reified I/O logic, both the antecedent and the con-
sequent of the if-then rule are conjunctions of predicates. Universal
and existential quantifiers are added to bound the free variables
occurring in the formulae. Universals that outscope the whole if-
then rules are used to “carry” individuals from the antecedent to
the consequent. Formal details and definitions are available in [29].
A simple example from the D-KB [28] is shown in (2). (2) encodes
in reified I/O logic part of Art.5(1)(a) of the GDPR. The if-then rule
216
Robaldo et al.
belongs to the set 𝑂 (note “∈ 𝑂” in (2)): it is an obligation requiring
each personal data processing to be lawful.
(2) ∀𝑒𝑝 ( ∃𝑡1 ,𝑧,𝑤,𝑦,𝑥 [ (𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇𝑖𝑚𝑒 𝑒𝑝 𝑡 1 ) ∧
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎 𝑧 𝑤) ∧ (𝐷𝑎𝑡𝑎𝑆𝑢𝑏 𝑗𝑒𝑐𝑡 𝑤) ∧
(𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑟 𝑦 𝑧) ∧ (𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑜𝑟 𝑥) ∧ (𝑛𝑜𝑚𝑖𝑛𝑎𝑡𝑒𝑠 𝑦 𝑥) ∧
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ 𝑒𝑝 𝑥 𝑧) ],
(𝑖𝑠𝐿𝑎𝑤 𝑓 𝑢𝑙 𝑒𝑝 ) ) ∈ 𝑂
Formulae in reified I/O logic employ two kind of predicates: primed
predicates such as 𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ and non-primed pred-
icates such as 𝐷𝑎𝑡𝑎𝑆𝑢𝑏 𝑗𝑒𝑐𝑡. The former are obtained by reifying
the latter; the first argument of primed predicates is the reification
of the non-primed counterpart, i.e., a FOL term.
We should not reify all predicates, but only those we need. For
instance, we do need to reify (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔 𝑥 𝑧) into
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ 𝑒𝑝 𝑥 𝑧), where 𝑒𝑝 explicitly refers to the
action of processing, because we need to assert a property on this ac-
tion: in the consequent of the obligation, we require it to be lawful,
i.e., to satisfy the 𝑖𝑠𝐿𝑎𝑤 𝑓 𝑢𝑙 predicate. Note that in (2), in order to
“carry” the variable 𝑒𝑝 from the antecedent to the consequent, a uni-
versal quantifier outscoping the if-then rule has been inserted. All
other variables are existentially quantified within the antecedent.
The other predicate that 𝑒𝑝 is required to satisfy is 𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇 𝑖𝑚𝑒.
This is a special predicates used to assert which reifications “really
exist” at a certain time. 𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇𝑖𝑚𝑒 parallels the well-known
predicate 𝐻𝑜𝑙𝑑𝑠𝐴𝑡 used in Event Calculus [14].
Thus, formula (2) reads: “for every personal data processing 𝑒𝑝
of some personal data 𝑧, owned by a data subject 𝑤, controlled by
a controller 𝑦, and processed by a processor 𝑥 (nominated by 𝑦), it
is obligatory for 𝑒𝑝 to be lawful.
2.3 Adding defeasibility to reified I/O logic
It is common in legislation that some rules override others in re-
stricted contexts. These more specific rules are seen as exceptions
of the general rules, as penguins may be seen as exceptions of birds
with respect to the ability of flying.
In line with the literature, e.g., [10], reified I/O logic models
exceptions via special predicates “Ex” that are false by default. This
is achieved via negation-as-failure (naf). “naf(Ex)” is true if either
“Ex” is false or it is unknown. On the other hand, when “Ex” holds,
“naf(Ex)” is false, and the general rule is blocked. An example, taken
from [28], is given by the following rules:
(a) If the data subject has given consent to processing, then the
processing is lawful.
(b) If the age of the data subject is lower than the minimal age for
consent of his member state, (a) is not valid.
(c) In case of (b), if the holder of parental responsibility has given
consent to processing, then the processing is lawful.
(a)-(c) are formalized as the following constitutive rules:
(3) ∀𝑒𝑝 ( ∃𝑡,𝑧,𝑤,𝑦,𝑥 [ (𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇𝑖𝑚𝑒 𝑒𝑝 𝑡) ∧ (𝐷𝑎𝑡𝑎𝑆𝑢𝑏 𝑗𝑒𝑐𝑡 𝑤) ∧
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ 𝑒𝑝 𝑥 𝑧) ∧ (𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑟 𝑦 𝑧) ∧
(𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑜𝑟 𝑥) ∧ (𝑛𝑜𝑚𝑖𝑛𝑎𝑡𝑒𝑠 𝑦 𝑥) ∧ (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎 𝑧 𝑤) ∧
(𝐺𝑖𝑣𝑒𝐶𝑜𝑛𝑠𝑒𝑛𝑡𝑇𝑜 𝑤 𝑒𝑝 ) ∧ 𝑛𝑎𝑓 ((𝑒𝑥𝑐𝑒𝑝𝑡𝑖𝑜𝑛𝐴𝑔𝑒𝐷𝑆 𝑒𝑝 )) ],
(𝑖𝑠𝐿𝑎𝑤 𝑓 𝑢𝑙 𝑒𝑝 ) ) ∈ 𝐶
Towards compliance checking in reified I/O logic via SHACL
(4) ∀𝑒𝑝 ( ∃𝑡,𝑧,𝑤,𝑦,𝑥,𝑠 [ (𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇 𝑖𝑚𝑒 𝑒𝑝 𝑡) ∧ (𝐷𝑎𝑡𝑎𝑆𝑢𝑏 𝑗𝑒𝑐𝑡 𝑤) ∧
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ 𝑒𝑝 𝑥 𝑧) ∧ (𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑟 𝑦 𝑧) ∧
(𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑜𝑟 𝑥) ∧ (𝑛𝑜𝑚𝑖𝑛𝑎𝑡𝑒𝑠 𝑦 𝑥) ∧ (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎 𝑧 𝑤) ∧
(𝑆𝑡𝑎𝑡𝑒𝑂 𝑓 𝑠 𝑤) ∧ (< 𝑎𝑔𝑒𝑂 𝑓 (𝑤) 𝑚𝑖𝑛𝐶𝑜𝑛𝑠𝑒𝑛𝑡𝐴𝑔𝑒𝑂 𝑓 (𝑠)) ],
(𝑒𝑥𝑐𝑒𝑝𝑡𝑖𝑜𝑛𝐴𝑔𝑒𝐷𝑆 𝑒𝑝 ) ) ∈ 𝐶
(5) ∀𝑒𝑝 ( ∃𝑡,𝑧,𝑤,𝑦,𝑥,𝑠,ℎ [ (𝑅𝑒𝑥𝑖𝑠𝑡𝐴𝑡𝑇 𝑖𝑚𝑒 𝑒𝑝 𝑡) ∧ (𝐷𝑎𝑡𝑎𝑆𝑢𝑏 𝑗𝑒𝑐𝑡 𝑤) ∧
(𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑖𝑛𝑔’ 𝑒𝑝 𝑥 𝑧) ∧ (𝐶𝑜𝑛𝑡𝑟𝑜𝑙𝑙𝑒𝑟 𝑦 𝑧) ∧
(𝑃𝑟𝑜𝑐𝑒𝑠𝑠𝑜𝑟 𝑥) ∧ (𝑛𝑜𝑚𝑖𝑛𝑎𝑡𝑒𝑠 𝑦 𝑥) ∧ (𝑃𝑒𝑟𝑠𝑜𝑛𝑎𝑙𝐷𝑎𝑡𝑎 𝑧 𝑤) ∧
(𝑆𝑡𝑎𝑡𝑒𝑂 𝑓 𝑠 𝑤) ∧ (< 𝑎𝑔𝑒𝑂 𝑓 (𝑤) 𝑚𝑖𝑛𝐶𝑜𝑛𝑠𝑒𝑛𝑡𝐴𝑔𝑒𝑂 𝑓 (𝑠)) ∧
(ℎ𝑎𝑠𝐻𝑜𝑙𝑑𝑒𝑟𝑂 𝑓 𝑃𝑟 ℎ 𝑤) ∧ (𝐺𝑖𝑣𝑒𝐶𝑜𝑛𝑠𝑒𝑛𝑡𝑇𝑜 ℎ 𝑒𝑝 ) ],
(𝑖𝑠𝐿𝑎𝑤 𝑓 𝑢𝑙 𝑒𝑝 ) ) ∈ 𝐶
3 COMPLIANCE CHECKING IN RDFS/OWL
RDFs/OWL is nowadays the W3C standard language for the Seman-
tic Web [1]. RDFs/OWL represents knowledge via flat sets of triples
“(subject, predicate, object)”, in which the predicate is
an rdf:Property while the subject and the object can be any
rdfs:Resource, including other rdf:Property(s). In other words,
RDFs/OWL allows to treat rdf:Property(s) as first-order terms on
which separately asserting other (meta-)properties.
It is then evident that reification is, in essence, the very same
mechanism used to represent knowledge in RDFs/OWL, thus the
idea of implementing reified I/O logic in the W3C standard.
Some proposals have been done to implement compliance check-
ing in RDFs/OWL, e.g., [9] and [6]. In these approaches compliance
checking is achieved by enriching the ontology with classes re-
ferring to sets of individuals compliant with the norms and by
enforcing “is-a” inferences on these classes.
For instance, the OWL ontology used in [9] includes a class
Supplier including individuals that supply consumers with some
goods. Since suppliers are obliged to communicate their contractual
conditions to their consumers (rule R1), the corresponding class in-
cludes a boolean datatype property hasCommunicatedConditions
which is true for those suppliers that has complied with their obli-
gation and false otherwise. The ontology includes then a class
SupplierR1compliant defined as to include only individuals in
Supplier for which hasCommunicatedConditions is true. Com-
pliance checking is enforced by applying simple “is-a” inferences.
In the same spirit, [6] encodes in a fragment of OWL2 selected
norms from Artt. 6, 7, 15, 23, and 30 of the GDPR, which concern
data usage policies. Compliance on these policies is again imple-
mented via “is-a” inferences.
While [9] and [6] are of course important contributions towards
the same direction of research advocated here, it is not clear how to
model exceptions in those frameworks. Furthermore, adding explicit
classes specifically devoted to “collect” the individuals compliant
with the norms, as well as introducing new ones to properly handle
exceptions, does not appear to be an easy and intuitive solution.
The rest of the paper proposes to use SHACL as an alternative
of the accounts in [9] and [6].
4 COMPLIANCE CHECKING IN SHACL
This paper proposes and makes initial investigations to encode
legal rules in a formal language different from RDFs/OWL. This
217
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
formal language is SHACL [2], proposed by W3C precisely for
validation and inferences on RDFs/OWL graphs. The use of SHACL
is currently a matter of ongoing research in the Semantic Web
community (see [7], [24], among others).
SHACL appears to be the right formal language for modelling
compliance checking, although so far it has been scarcely investi-
gated to this end, preliminary works being [20], [21], and [8].
SHACL was originally proposed to define special conditions on
RDFs/OWL graphs, called “SHACL shapes”, more expressive than
standard OWL cardinality and quantifier restrictions. RDFs/OWL
graphs can be then validated against a set of such SHACL shapes.
However, SHACL “may be used for a variety of purposes beside
validation, including user interface building, code generation and data
integration” (cit. [2]). This paper adds a new use cases for SHACL
in that it proposes to use it for serializing reified I/O logic formulae
fit to check compliance.
In order to enhance the expressivity and the flexibility of the
standard, a current W3C Working Group Note proposes to enrich
SHACL shapes with advanced features1 such as “SHACL rules” to
derive inferred triples from asserted ones, prior to validation.
As explained in [25], SHACL rules can trigger ontological or
non-ontological inferences. Ontological inferences derive facts that
can be added to the model. On the other hand, non-ontological
inferences have the sole purpose of aggregating data, without nec-
essarily asserting them in the model, in order to facilitate validation.
5 SERIALIZING REIFIED I/O LOGIC IN SHACL
This paper represents the first attempt to investigate how to serialize
reified I/O formula modeling obligations as SHACL shapes and
reified I/O formula modeling constitutive rules as SHACL rules.
(6) shows the SHACL shape that serializes (2) above. Both require
every personal data processing to be lawful.
(6) CheckLawfulness
rdf:type sh:NodeShape;
sh:targetClass PersonalDataProcessing;
sh:property [ sh:path is-lawful;
sh:hasValue "true"ˆˆxsd:boolean; ];
In (6), “sh:” is SHACL namespace prefix. (6) is a sh:NodeShape
requiring each individual of the sh:targetClass to satisfy the
sh:property. The latter constrains the individuals reached from
the sh:targetClass through the sh:path to satisfy sh:hasValue.
On the other hand, PersonalDataProcessing, is-lawful, and
all other RDFs/OWL resources used in this paper are associated
1:1 with the predicates used in the reified I/O logic formulae such
as (2), in the same way as the predicates occurring in the D-KB
[28] are associated with RDFs/OWL resources from the PrOnto
ontology [19], an OWL ontology proposed to conceptualize the
data protection domain. Space constraints avoid to provide further
details about the 1:1 mapping between reified I/O logic predicates
and RDFs/OWL resources.
SHACL shapes refer to constraints, a solution that appears to be
more intuitive and economical than overpopulating the ontology
with extra classes as suggested in [9] and [6].
The validation facts, as well as new individuals, derived through
SHACL are not mandatorily inserted in the ontology. The SHACL
1 See https://www.w3.org/TR/shacl-af
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
rules to model the reified I/O logic formulae in (3), (4), and (5)
represent non-ontological inferences, in the sense explained in [25]:
these rules are only functional to infer the truth value of is-lawful
before the SHACL shape in (6) is validated.
(3), (4), and (5) are serialized in the SHACL rules in (7), (8), (9),
and, below, (10).
(7) sh:rule [rdf:type sh:TripleRule; sh:order 0;
sh:subject sh:this;
sh:predicate has-min-consent-age;
sh:object [sh:path
(has-theme has-personal-data
is-personal-data-of has-member-state
has-min-consent-age);]; ];
(8) sh:rule [rdf:type sh:TripleRule; sh:order 1;
sh:condition [
sh:property [sh:path has-min-consent-age;
sh:minCount 1;];
sh:property [
sh:path (has-agent has-age);
sh:lessThan has-min-consent-age;]; ];
sh:subject [sh:path has-theme;];
sh:predicate rdf:type;
sh:object exceptionAgeDS; ];
(9) sh:rule [rdf:type sh:TripleRule; sh:order 2;
sh:condition [
sh:not [sh:property [sh:path has-theme;
sh:class exceptionAgeDS;]; ]; ];
sh:subject [sh:path has-theme;];
sh:predicate is-lawful;
sh:object "true"ˆˆxsd:boolean; ];
The sh:targetClass of all these SHACL rules is GiveConsent.
Rules are executed according to the sh:order, from the lowest to
the highest value. Each rule in (7)-(9) makes a new assertion: the
rdf:Property specified in the sh:predicate of the rule is asserted
between the two RDFs/OWL resources in the sh:subject and
the sh:object. The sh:subject and the sh:object may be the
sh:targetClass itself (keyword “sh:this”), a resource reachable
from the sh:targetClass through a path specified in sh:path,
any other resource in the ontology, or a literal.
(7) is executed as first because its sh:order is “0”. This rule sets
the value of the property has-min-consent-age for each individ-
ual in the class GiveConsent. This values is set to the integer value
reachable from the sh:path defined on sh:object in (7)). Specifi-
cally, this is the minimal consent age (has-min-consent-age) of
the Member State (has-member-state) of the data subject owning
the personal data (has-personal-data is-personal-data-of) in-
volved in the personal data processing occurring as the theme of
the GiveConsent instances (has-theme).
It is important to understand that has-min-consent-age will
not be asserted on the individuals of GiveConsent also in the ref-
erence ontology, but only in the derived one. In other words, (7) is
a non-ontological inference rule that collects/aggregates this value
in GiveConsent for validation purposes only. After the validation,
these values will be discharged.
Rule (8) compares the minimal consent age of the agents’ Member
State, just asserted by (7) on GiveConsent’s instances, with the
agents’ age. The two rules are then executed in a pipeline, thanks to
218
Robaldo et al.
SHACL command sh:order. Mirroring these inferences in native
RDFs/OWL seems to be more difficult in that the formalism does
not allow to specify a priority between the inference rules.
When the agents’ age has been specified (sh:minCount 1) and it
is lower than (sh:lessThan) the minimal consent age of the Mem-
ber State previously asserted by (7), rule (8) asserts the individual
of PersonalDataProcessing in the has-theme property of the in-
dividual of GiveConsent as member of the class exceptionAgeDS
(see rdf:type in sh:predicate).
Finally, (9) sets as true the property is-lawful of the instances
of PersonalDataProcessing that do not (sh:not) belong to class
exceptionAgeDS. (9) implements the reified I/O logic formula shown
above in (3) and the SHACL operator sh:not the negation-as-failure
(predicate 𝑛𝑎𝑓 ) occurring therein. sh:not is in fact true when the
ontology does not include any specific assertion of the personal
data processing as member of the class exceptionAgeDS. In other
words, since the close world assumption hold for both RDFs/OWL
and SHACL, sh:not is true when it is either false or unknown
whether the personal data processing belongs to this class.
Finally, (10) implements the reified I/O logic formula (5) above:
(10) sh:rule [rdf:type sh:TripleRule; sh:order 2;
sh:condition [
sh:property [sh:path
(has-theme; has-personal-data
is-personal-data-of has-age);
sh:lessThan has-min-consent-age; ];
sh:property [sh:path (has-theme
has-personal-data
is-personal-data-of
has-holder-of-pr);
sh:equals has-agent;]; ];
sh:subject [sh:path has-theme;];
sh:predicate is-lawful;
sh:object "true"ˆˆxsd:boolean; ];
If the age of the data subject (has-age) who owns the personal
data of the processing (has-personal-data is-personal-data)
that is the theme of a GiveConsent individual (has-theme) is lower
than (sh:lessThan) the minimal consent age of his/her Member
State and the agent of this GiveConsent individual is the holder
of the data subject’s parental responsibility (has-holder-of-pr),
then the boolean is-lawful is again set to true.
6 CONCLUSIONS
Reified I/O logic is a recent deontic logical framework explicitly
designed to handle natural language semantics, i.e., to represent
norms occurring in existing legislation such as the GDPR.
So far, the research in reified I/O logic has focused only on knowl-
edge representation issues, specifically on how to use the formalism
for representing contextual meaning of norms [3].
On the other hand, this paper is the first attempt to investi-
gate computational issues in reified I/O logic, specifically how to
represent the reified I/O logic if-then rules in a computable machine-
readable format fit to enforce compliance checking.
This paper proposed to model regulative rules as SHACL shapes
and constitutive rules as SHACL rules. SHACL shapes and rules are
applied to RDFs/OWL models that describe states of affairs.
Towards compliance checking in reified I/O logic via SHACL
The solution proposed here is alternative to some recent ap-
proaches that model compliance checking on RDFs/OWL ontolo-
gies, e.g., [9] and [6].
On the other hand, the present work only represents the first step
of a research endevour aiming at developing a full inference engine
for reified I/O logic that implements and integrates all components
involved in normative reasoning. Much further work needs to be
done in order to obtain a formally well-defined framework, tested
on existing industrial use cases.
Further directions of research include the automatic or semi-
automatic generation of RDFs/OWL or SHACL assertions from
legal texts, possibly via NLP (cf. [4], [5], [18]).
ACKNOWLEDGMENTS
This research has been supported by the Legal Innovation Lab Wales
operation within Swansea University’s Hillary Rodham Clinton
School of Law. The operation has been part-funded by the European
Regional Development Fund through the Welsh Government.
REFERENCES
[1] 2012. Web Ontology Language (OWL). Technical Report. W3C. https://www.w3.
org/OWL
[2] 2017. Shapes constraint language (SHACL). Technical Report. W3C. https:
//www.w3.org/TR/shacl
[3] Cesare Bartolini, Andra Giurgiu, Gabriele Lenzini, and Livio Robaldo. 2016.
Towards Legal Compliance by Correlating Standards and Laws with a Semi-
automated Methodology. In BNCAI (Communications in Computer and Informa-
tion Science, Vol. 765). Springer, 47–62.
[4] G. Boella, L. di Caro, L. Humphreys, L. Robaldo, and L. van der Torre. 2012. NLP
Challenges for Eunomos, a Tool to Build and Manage Legal Knowledge. Proceed-
ings of the International Conference on Language Resources and Evaluation.
[5] Guido Boella, Luigi Di Caro, Daniele Rispoli, and Livio Robaldo. 2013. A System
for Classifying Multi-label Text into EuroVoc. In Proceedings of the Fourteenth
International Conference on Artificial Intelligence and Law (Rome, Italy) (ICAIL
’13). ACM, New York, NY, USA, 239–240.
[6] Piero A. Bonatti, Luca Ioffredo, Iliana M. Petrova, Luigi Sauro, and Ida Sri Rejeki
Siahaan. 2020. Real-time reasoning in OWL2 for GDPR compliance. Artificial
Intelligence 289 (2020).
[7] Julien Corman, Juan L. Reutter, and Ognjen Savkovic. 2018. Semantics and Vali-
dation of Recursive SHACL. In The Semantic Web - ISWC 2018 - 17th International
Semantic Web Conference, Monterey, CA, USA, October 8-12, 2018, Proceedings,
Part I (Lecture Notes in Computer Science, Vol. 11136), Denny Vrandecic, Kalina
Bontcheva, Mari Carmen Suárez-Figueroa, Valentina Presutti, Irene Celino, Marta
Sabou, Lucie-Aimée Kaffee, and Elena Simperl (Eds.). Springer, 318–336.
[8] Christophe Debruyne, Harshvardhan J. Pandit, Dave Lewis, and Declan
O’Sullivan. 2019. Towards Generating Policy-Compliant Datasets. In 13th IEEE
International Conference on Semantic Computing, ICSC 2019, Newport Beach, CA,
USA, January 30 - February 1, 2019. IEEE, 199–203.
[9] Enrico Francesconi and Guido Governatori. 2019. Legal Compliance in a Linked
Open Data Framework. In Legal Knowledge and Information Systems - JURIX
2019: The Thirty-second Annual Conference, Madrid, Spain, December 11-13, 2019
(Frontiers in Artificial Intelligence and Applications, Vol. 322), Michal Araszkiewicz
and Víctor Rodríguez-Doncel (Eds.). IOS Press, 175–180.
[10] G. Governatori, F. Olivieri, A. Rotolo, and S. Scannapieco. 2013. Computing
Strong and Weak Permissions in Defeasible Logic. Journal of Philosophical Logic
6, 42 (2013), 799–829.
[11] Jörg Hansen. 2014. Reasoning about permission and obligation. In David Makinson
on Classical Methods for Non-Classical Problems, S. O. Hansson (Ed.). Outstanding
Contributions to Logic Volume 3, Springer, 287–333.
[12] J.R. Hobbs and A.S. Gordon. 2017. A formal theory of commonsense psychology,
how people think people think. Cambridge University Press.
[13] J. R. Hobbs. 2008. Deep Lexical Semantics. In Proc. of the 9th International
Conference on Intelligent Text Processing and Computational Linguistics (CICLing-
2008). Haifa, Israel.
[14] R Kowalski and M Sergot. 1986. A Logic-based Calculus of Events. New Generation
Computing 4, 1 (1986), 67–95.
[15] Tomer Libal and Alexander Steen. 2020. Towards an Executable Methodology for
the Formalization of Legal Texts. In Logic and Argumentation - Third International
Conference, CLAR 2020, Hangzhou, China, April 6-9, 2020, Proceedings (Lecture
219
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Notes in Computer Science, Vol. 12061), Mehdi Dastani, Huimin Dong, and Leon
van der Torre (Eds.). Springer, 151–165.
[16] David Makinson and Leendert van der Torre. 2001. Constraints for input/output
logics. Journal of Philosophical Logic 30, 2 (2001), 155–185.
[17] David Makinson and Leendert W. N. van der Torre. 2000. Input/Output Logics.
Journal of Philosophical Logic 29, 4 (2000), 383–408.
[18] Rohan Nanda, Luigi Di Caro, Guido Boella, Hristo Konstantinov, Tenyo Tyankov,
Daniel Traykov, Hristo Hristov, Francesco Costamagna, Llio Humphreys, Livio
Robaldo, and Michele Romano. 2017. A Unifying Similarity Measure for Auto-
mated Identification of National Implementations of European Union Directives.
In Proceedings of the 16th Edition of the International Conference on Articial Intelli-
gence and Law (ICAIL 2017). Association for Computing Machinery.
[19] Monica Palmirani, Michele Martoni, Arianna Rossi, Cesare Bartolini, and Livio
Robaldo. 2018. PrOnto: Privacy Ontology for Legal Compliance. In Proceedings
of the 18𝑡ℎ European Conference on Digital Government (ECEG).
[20] Harshvardhan Jitendra Pandit, Declan O’Sullivan, and Dave Lewis. 2018. Explor-
ing GDPR Compliance Over Provenance Graphs Using SHACL. In Proc. of the
Posters and Demos Track of the 14th International Conference on Semantic Systems
co-located with the 14th International Conference on Semantic Systems (SEMAN-
TiCS 2018), Vienna, Austria, September 10-13, 2018 (CEUR Workshop Proceedings,
Vol. 2198), Ali Khalili and Maria Koutraki (Eds.).
[21] Harshvardhan J. Pandit, Declan O’Sullivan, and Dave Lewis. 2019. Test-Driven
Approach Towards GDPR Compliance. In Semantic Systems. The Power of AI and
Knowledge Graphs, Maribel Acosta, Philippe Cudré-Mauroux, Maria Maleshkova,
Tassilo Pellegrini, Harald Sack, and York Sure-Vetter (Eds.). Springer International
Publishing, 19–33.
[22] Xavier Parent and Leon van der Torre. 2014. Aggregative Deontic Detachment for
Normative Reasoning. In Principles of Knowledge Representation and Reasoning:
Proceedings of the Fourteenth International Conference, KR 2014, Vienna, Austria,
July 20-24, 2014.
[23] Xavier Parent and Leendert van der Torre. 2014. “Sing and Dance!”. In Deontic
Logic and Normative Systems, Fabrizio Cariani, Davide Grossi, Joke Meheus, and
Xavier Parent (Eds.). Springer International Publishing, 149–165.
[24] Paolo Pareti, George Konstantinidis, Fabio Mogavero, and Timothy J. Norman.
2020. SHACL Satisfiability and Containment. In The Semantic Web - ISWC 2020 -
19th International Semantic Web Conference, Athens, Greece, November 2-6, 2020,
Proceedings, Part I (Lecture Notes in Computer Science, Vol. 12506), Jeff Z. Pan,
Valentina A. M. Tamma, Claudia d’Amato, Krzysztof Janowicz, Bo Fu, Axel
Polleres, Oshani Seneviratne, and Lalana Kagal (Eds.). Springer, 474–493.
[25] Paolo Pareti, George Konstantinidis, Timothy J. Norman, and Murat Sensoy. 2019.
SHACL Constraints with Inference Rules. In The Semantic Web - ISWC 2019 -
18th International Semantic Web Conference, Auckland, New Zealand, October
26-30, 2019, Proceedings, Part I (Lecture Notes in Computer Science, Vol. 11778),
Chiara Ghidini, Olaf Hartig, Maria Maleshkova, Vojtech Svátek, Isabel F. Cruz,
Aidan Hogan, Jie Song, Maxime Lefrançois, and Fabien Gandon (Eds.). Springer,
539–557.
[26] L. Robaldo. 2010. Independent Set readings and Generalized Quantifiers. The
Journal of Philosophical Logic 39(1) (2010), 23–58.
[27] L. Robaldo. 2011. Distributivity, Collectivity, and Cumulativity in terms of
(In)dependence and Maximality. The Journal of Logic, Language, and Information
20(2) (2011), 233–271.
[28] L. Robaldo, C. Bartolini, M. Palmirani, A. Rossi, M. Martoni, and G. Lenzini. 2020.
Formalizing GDPR provisions in reified I/O logic: the DAPRECO knowledge base.
The Journal of Logic, Language, and Information 29 (2020). Issue 4.
[29] L. Robaldo and X. Sun. 2017. Reified Input/Output logic: Combining Input/Output
logic and Reification to represent norms coming from existing legislation. The
Journal of Logic and Computation 7 (2017). Issue 8.
[30] X. Sun and L. Robaldo. 2017. On the complexity of Input/Output logic. The
Journal of Applied Logic 25 (2017), 69–88.
 Modelling Legal Procedures
Antonino Rotolo
Alma AI, University of Bologna
Bologna, Italy
antonino.rotolo@unibo.it
ABSTRACT
A legal procedure in court proceedings is a sequence of actions
in which the last action is (the creation of) a(n individual) norm,
where the court settles that it is obligatory in the interest of some
agents that other agents bring about a certain state of affairs. This
paper models legal procedures by using a variant of Propositional
Dynamic Logic (PDL) enriched with a preference operator for pri-
oritising procedural actions. The key reason towards the usage of
PDL is that, in procedural law, claims and resolutions resemble
programs to be executed. Requests are organised in a preference
order and resolutions have their own dynamics of execution (either
spontaneously by the one obliged and/or by force of law).
CCS CONCEPTS
• Applied computing → Law, social and behavioral sciences;
• Theory of computation → Logic.
KEYWORDS
Legal procedures, PDL, preferences
ACM Reference Format:
Antonino Rotolo and Clara Smith. 2021. Modelling Legal Procedures. In Eigh-
teenth International Conference for Artificial Intelligence and Law (ICAIL’21),
June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 5 pages.
https://doi.org/10.1145/3462757.3466089
1 INTRODUCTION
A legal procedure in civil court proceedings is the formal way in
which civil proceedings are conducted [6]. The practice and proce-
dures of a court are regulated by rules usually stated in a procedure
rule corpus (or made by a committee) possibly supplemented, e.g.,
by practice directions and pre-action protocols. A legal procedure is
usually defined as a chain of consecutive actions which has as (final)
goal the decision/solution of a conflict. Thus, a legal procedure is a
(finite) sequence of actions in which the last action is (the creation
of) a(n individual) norm, the judgment, where the court settles that
it is obligatory in the interest of some agents that other agents bring
about a certain state of affairs.
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466089
220
Clara Smith
Law Faculty, University of La Plata
La Plata, Argentina
claritasmith@gmail.com
Although we may identify several technical differences between
a process and a procedure1 , in this work we will generically refer
to a procedure as a part of the process.
One peculiar aspect of civil proceedings is that some types of
procedure in the process are prioritised. Priorities indeed derive
from individual preferences of the parties in the process, but they
can also follow from objective ordering requirements from civil
procedures. For example, in the Italian procedure [4], we have the
following procedural headings:
• in via preliminare, as a preliminary procedural matter – As a
preliminary procedural matter, the petitioner tries to raise
new issues in her reply and advances affidavits which were
not included in the petition;
• in via istruttoria, as an interlocutory matter;
• in via pregiudiziale di merito, plea on the merits;
• in via subordinata di merito, alternatively/in the alternative;
• in via esecutiva, as an enforcement matter;
• in estremo subordine, as a further alternative.
For capturing procedural priorities, the operator ⊗ is introduced
here, which is to be understood as a preference operator [1] that
allows for a compact representation of the execution of alternative
(or supplementary or backup) actions in social domains. The in-
tended reading of an expression of the form a ⊗ b ⊗ c is that a is
preferred, but if a is not carried out then b is to be the case, and if b
is not carried out then c is to be applied.
Technically, in this paper we combine this preference logic with
a propositional dynamic logic (PDL) to put procedures, preferences,
courses of actions, and obligations to work all together.
The layout of the paper is a follows. Section 2 offers some dis-
cussion on the components of legal procedure. Section 3 illustrates
syntax and semantics of the formal framework. Section 4 discusses
some axioms and non-logical principles for modelling procedures.
Section 5 sketches some patterns for reasoning and computational
methods for handling procedures. A summary concludes the paper.
2 THE LEGAL PROCEDURE
A legal procedure is guided by procedural principles which give
to procedures their particular structure and logic. This implies
that procedures have a particular way of moving forward from its
beginning towards the decision.
A procedure starts with the very first claim of the plaintiff, who
usually sues some agent(s). In a regular procedure, the plaintiff
exposes a state of affairs, and presents her/his request to the court.
Intuitively, we can imagine the ideal structure of a procedure as a
kind of directed graph, with one entry node (the plaintiff’s claim)
and one final node, the decision, i.e., a specific obligation imposed
1 Forexample, any procedure typically concerns the enforcement of individuals’ rights
during the process [6].
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
to the parties by the authority of the court. In between, there are
multiple possible and intermediate claims proposed by the plaintiff
and multiple possible defences presented by the defendant, and,
of course, many intermediate decisions taken by the court2 . Some
of these court decisions have a preclusive nature, i.e. they do not
allow plaintiff and/or defendant to come back to previous points in
the procedural flow. Given this ideal structure, the real procedure,
i.e. the actions effectively ordered by the authority, corresponds to
a path (a subgraph) in the bigger, ideal graph.
2.1 Structure of Claims
Plaintiff and defendant are both parties in the procedure. A claim
is the only available tool a party has for communicating with the
court within a procedure, asking for an action or a fact to be con-
sidered. Petitions rarely consist in one single request. They are
usually organised in the form of one main (preferred) request, plus
subsidiary requests. This structure constitutes a common guideline.
The main reason for this preferred/subsidiary organisation of ac-
tions resides in the fact that plaintiff and defendant know that the
court may not decide in their favour regarding their main request.
Therefore, each party presents to the court a menu of wills for the
court’s consideration, altogether and at the same instant.
Petitions can be seen, therefore, as prioritised goals. This prefer-
ence structure outlines a strategic move: I ask the court to order to
bring about my preferred state of affairs, if this would or could not
be the case, I ask the court to order to bring about this other state
of affairs, and so on. Indeed, the goal pursued by the parties is that
her/his main request becomes the content of the decision of the court
(if not possible, then one of the subsidiaries, as given).
2.2 Resolutions
When the court receives a claim of a party it analyses it, and chooses
among the menu of proposed options.The court’s chosen option
implies that the court discards those options appearing before the
one chosen, and that the options after the one chosen are not taken
into account, at least at the time. The normal course of the procedure
usually indicates that the plaintiff presents the claim as described,
then the court resolves, next the defendant defends himself with his
own claim, then the court resolves. To each request of the parties
the court produces an answer. We call this response a resolution.
The claim/resolution chain repeats from the very first claim of the
plaintiff to the judgment.
A resolution always has a performative nature even if it is a
low impact decision (e.g. “Take this fact into account for later in
this procedure”); it is an order of the authority. So far, in a sense,
resolutions have an executable nature: the court declares through
a resolution which actions are to be done. The resolution, thus,
has to be executed. Suppose I am sued and I defend myself with
the claim of paying with no interest, and, subsidiary, to pay with
minimum interest. (My claim operates as the input to the court’s
decision.) Suppose next that the court sentences me to pay with
the minimum interest. Then I, the obliged agent, have to comply
with the judgment by effectively and spontaneously paying. With
my payment the judgment is considered to be “executed” (and
2 We will restrict ourselves to procedures including only these agents: plaintiff, defen-
dant, and the decision-maker (e.g. the judge, the court, a mediator).
221
“consumed” in a technical sense w.r.t. that tort.) But, if I do not pay,
by the force of law I may be compelled to do so through a brief
secondary process known as judgement execution.
3 FORMALISING LEGAL PROCEDURES
We will use for representing legal procedures a multi-agent variant
of basic core of Propositional Dynamic Logic (PDL) [3] enriched
with the preference operator ⊗ for denoting preferences among
procedural actions [1]. The key reason towards the usage of PDL is
that, in the procedural law domain, claims and resolutions indeed
resemble programs to be executed. Requests or proposed actions
are organised in a preference order. Resolutions have their own
dynamics of execution (either spontaneously by the one obliged
and/or by force of law).
3.1 Syntax
Let Ag be a set of agents. The language L consists of a set PROP =
{A, B, C, . . . } of countably many proposition symbols, a set P =
{α i |i ∈ Ag} of countably many atomic programs which we call
atomic procedural actions or atomic procedures, the usual boolean
operators, the program constructors ;, ∪, and ⊗ = {⊗i |i ∈ Ag} and
the modality [Π] for any procedure Π. A procedure Π is ⊗-free iff
⊗ does not occur in it.
So, formally, expressions of language are defined as follows:
p ::= A|¬p|p1 ∧ p2 |⟨Π⟩p with Π ::= α i |Π1 ; Π 2 |Π 1 ∪ Π 2 |
Πi 1 ⊗i · · · ⊗i Πi n
where A ∈ PROP, α i ∈ P, and Πi 1 , . . . , Πi n are ⊗-free.
We usually deal with at least three types of agents, let’s denote
them p, d, k for representing the plaintiff, the defendant, and the
court. Propositional letters denote as usual states of affairs. Complex
formulas are built using classical boolean connectives as expected.
As usual, we also have an infinite collection of [Π] operators where
Π is a (lawful) procedure. In the simplest case, we may have [α i ],
an atomic procedure for the agent i; hence, [α i ] A is a formula that
reads “every execution of α by i from the present state leads to a
state where A is true”. The dual assertion ⟨α i ⟩A—such that ⟨α i ⟩A ≡
¬[α i ]¬A—states that “the execution of α by i from the present state
leads to a state where A is true”. Complex procedures are intuitively
defined from fixed basic atomic procedures as follows:
(Sequence) if Π 1 and Π2 are procedures then Π1 ; Π 2 (“do Π 1
followed by Π2 ”) is a procedure,
(Choice) if Π 1 and Π2 are procedures then Π1 ∪ Π2 (“do Π 1 or
Π 2 , non-deterministically”) is a procedure,
(Preference) if Π1 and Π 2 are procedures, so are (Π1 ⊗p Π 2 ),
(Π 1 ⊗d Π2 ), and (Π 1 ⊗k Π2 ), meaning “agent p prefers doing
α, but if not then p prefers doing β” (resp. for d, k).
Notice that we do not use the usual program constructor ∗, which
models in PDL the execution for a program of a nondeterministically
chosen finite number of times. Although it is smooth to use it here,
we can ignore this constructor for our specific purpose.
The crucial variation w.r.t. the original use of ⊗ in [1] is that, in
their work, the authors interpret an expression [a ⊗ b]A as a being
the most preferred state of affairs, and if a is not the case then b
is preferred. In this present work we interpret ⊗ as a preference
operator among procedural actions.
Modelling Legal Procedures ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Remark 3.1. We read α p as “α as proposed by p”. For example, the Definition 3.7. Let M a procedural model M = ⟨W , R Π , R O, ≺
formula [α p ⊗k β d ]A is to be read: “the court prefers that p’s pro- , V ⟩ and Ri (w) := {v ∈ W | wRi v}, ∥A∥V := {w ∈ W | |=V w A}. The
posal is done, if that is not procedurally possible then d’s proposal valuation function for M is a follows:
must be carried out (leading to a state-of-affairs where A holds)”. • usual for atoms and boolean conditions,
• w |= OA iff ∃R ∈ R O such that R(w) = ∥A∥V [2];
Remark 3.2. Resolutions are expressions in which the procedure
• w |= [Π]A iff ∀v ∈ W if wR Πv then v ∈ ∥A∥V ;
subexpression is one relativised to the court (⊗k ). Within a (lawful)
• w |= [Π 1 ⊗i · · · ⊗i Πn ] A iff w |= [Π 1 ∪ · · · ∪ Πn ] A iff there
procedure, obligations are always imposed by the agent that holds
exists a preference path R Π1 ≺i · · · ≺i R Πn .
the legal power (e.g. the court, the third neutral.) When the court
speaks, it speaks in the form of an obligation, which is added as a To sum-up, this semantics combines the one for classical modal
milestone to the procedure. A resolution has the force of law. E.g.: logics proposed in [2] and the standard one for PDL, plus ranking
[α p ⊗k β p ]A implies the deontic expression O[α p ⊗k β p ] A3 . accessibility relations for procedures.
Note that the “ideal” graph intuitively described in the first sec-
In this work we state that obligations such as OA follow from tion is a multigraph. Vertices are states of affairs and arcs are pro-
resolutions. Later in this paper we address the detachment of O- cedures, relativised to agents. For example, suppose that starting
expressions from court resolutions. from the state-of-affairs v, procedure α p leads us to state w 1 , proce-
Example 3.3. The formulas [α p ⊗p β p ]A and [α p ⊗k β p ]A are read: dure µ d leads us to state w 2 , procedure β p leads us to state w 3 and
“The plaintiff proposes α, subsidiary he asks for β” and “The court procedure γ p to state w 4 . We may e.g. write the plaintiff’s request
as [α p ⊗p β p ⊗p γ p ]A, the defendant’s defence as µ d ; and a court’s
resolves that α as proposed by p is to be done, subsidiary β as

proposed by p is to be done” (leading to a state of affairs where A decision as e.g. [α p ⊗k µ p ]A (or α p ⊗k µ p A).
holds), respectively.
αp / w1 A
The formula [(α p ∪ β p ) ⊗k γ p ]A has in its turn the intuitive read-
ing: “the court decides that either α p or β p are to be performed,
being γ p subsidiary (leading to a state of affairs where A holds)”4 . µd
v / w2 A
3.2 Semantics βp
Let us now present an adequate semantics for our logic. The idea is '
to extend standard semantics for PDL with a relational version of w3 A
the one for ⊗. Multi-relational frames for ⊗-logics are based on the γp
idea of directly ranking relations.
, w4 A
Definition 3.4 (Procedural frame). A procedural frame is a struc-
ture F = ⟨W , R Π , R O, ≺⟩, where We assume that ≺ is a collection of strict partial orders, i.e., which
are irreflexive, transitive and asymmetric: one cannot validate a
• W is a non empty set of  possible worlds,
formula such as [Π1 ⊗i Π1 ] A. Transitivity and antisymmetry are
• R Π is a countable set R α i |α i ∈ P of binary relations over
adopted as expected, and this it ensures the validity of
W ; we inductively extend R Π , for each non-atomic procedure
Π, as follows: [Π1 ⊗i · · · ⊗i Πn ] A ≡ (Contraction)
– wR Π1 ;Π2 v iff there exists a world z such that wR Π1 z and [Π1 ⊗i · · · ⊗i Πk−1 ⊗i Πk+1 · · · ⊗i Πn ] A
zR Π2 v; where Π j = Πk , j < k
– wR Π1 ∪Π2 v iff wR Π1 v or wR Π2 v;
• R O is countable set of binary relations over W , Lemma 3.8. The axiom (Contraction) is valid in the class of proce-
• ≺= {≺i |i ∈ Ag} is a collection of a strict partial orders over dural frames.
RΠ.
4 AXIOMS AND PRINCIPLES
Definition 3.5 (Procedural model). A procedural model is a struc-
4.1 Consistency
ture M = ⟨F , V ⟩, where
First of all, the fact that the court is the authority and that their
• F is a procedural frame, and resolutions have the force of law can be formalised as follows:
• V is a valuation function, V : Prop → 2W .

Definition 3.6 (Preference path). Let R Π1 , . . . , R Πn ⊆ R Π . We [Π1 ⊗k · · · ⊗k Π j ⊗k · · · ⊗k Πm ⊗k · · · ⊗k Πn ] A →
write R Π1 ≺i · · · ≺i R Πn to express that, for each j where 1 ≤ j < n, (Consistency)
R Π j ≺i R Π j+1 . We call R Π1 ≺i · · · ≺i R Πn a i-preference path from → ¬[Π1 ⊗x · · · ⊗x Πm ⊗x · · · ⊗x Π j ⊗x · · · ⊗x Πn ] A
R Π1 to R Πn of length n.
with x ∈ Ag
3 “O” is minimal non-normal deontic operator for representing obligations [2]. In the simplest case,
4 We can informally refer to procedures either by mentioning or not the state-of-affairs
they lead us to.
[Π 1 ⊗k Π2 ]A → ¬([Π2 ⊗x Π 1 ]A).

222
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
This axiom settles procedural consistency with respect to prefer-
ences, i.e., that a resolution always prevails over any preferences
and cannot imply a preference on the contrary by any other agent in
the process, not even by means of a court’s contradictory resolution.
For example, if the court resolves action α i is to be done (and,
subsidiary, βk ) then no other agent should prefer them in a converse
order, not even the court itself.
Although this axiom stands for a consistency axiom, its structure
cannot be imposed between plaintiff’s and defendant’s respective
strategies, as we will see next.
4.2 Procedural Bilaterality
Generally speaking, bilaterality in court procedures implies the
necessary existence of two confronted parties in a procedure.Two
confronting parties is the essence of a tort.
A lawful procedure implies a duality of arguing agents and there-
fore a corresponding duality of confronting goals and confronting
strategies. All these hold, at least in the very beginning, when the
procedure starts (it should be clear that, during the course of the
process, parties can engage in an agreement). Such a duality does
not mean, in a legal sense, irreconcilable positions.
That is why formulas such as the following can be true:
[Π1 ⊗p Π 2 ]A ∧ ([Π 2 ⊗d Π1 ]A) (1)
 model is M2 reaching an agreement from v is unclear; this because
Indeed, it is legally admissible that both α ⊗p β A and ⟨β ⊗d α⟩ A
hold at a given time during the process, even when both formulas
reflect opposite preferences. Parties are needed as opponents during
the process, and the final word would be the court’s.
4.3 Discussion
A court’s analysis of particular “instantiations” of the contradiction
principle during a process may lead the court members to the
detection of possible points of agreement between parties.
For the sake of simplicity, assume to work with atomic proce-
dures and suppose that we have the following model M1 where, at
world v, the proposals α and β (which are respectively plaintiff’s
and defendant’s) lead us both to world w, where A holds
M1 : αp
2 wM A
v βd
If α p ≺p β d , α p ≺k β d , and β d ≺d α p , it is easy to see that
M1, v |= [α p ⊗p β d ]A M1, v |= [β d ⊗d α p ]A
M1, v |= [α p ⊗k β d ]A
We may state that, semantically speaking, the three expressions
are somehow equivalent because they all lead us to the same state
and give us the same result in M1 . Nonetheless, the three expres-
sions are indeed different and highly expressive: they allow us to
write exactly what we want to express in three different expres-
sions: plaintiff’s and defendant’s strategies (as distinct or opposite
one from the other), and the court’s resolution.
Moreover, note that M1 depicts a peaceful transition within the
bi-simulation  [3, 5].
For example,
process. This fact may help the court to see that both parties are
223
seeking or willing the same state-of-affairs (and e.g. the situation
may be used by the court as a basis to analyse a call for agreement.)
Suppose now that we have the following model M2 :
w 1A,C
?
αp
/ w A,D
v 2
βd
M1 and M2 are indeed different models (they are indeed different
graphs). Even both models are different, if α p ≺k β d we have
M2, v |= [α p ⊗k β d ]A
as it holds for M1 . But we have that
M2, v |= [α p ]C M2, v |= [β d ]D
M2, v |= [α p ]A M2, v |= [β d ]A.
Now assume that the court is interested in that plaintiff and
defendant reach an agreement based on the fact that A holds. If
our model is M1 , the possibility of reaching an agreement while
in v seems quite straightforward; it is easy to automatically detect
that in a graph like M1 both parties propose actions that lead to
the same world where the same state-of-affairs hold. But if our
when trying to reach A we may also be conducted to other (possibly
unwanted) states-of-affairs, e.g. C or D.
However, if the court considers C and D as irrelevant, we can
argue that both models M1 and M2 are equivalent.
A different discussion raises when we focus on choosing be-
tween α and β. Such a choice may be either important or useless,
according to what actions α and β actually are. There is quite an
extra amount of procedural reasoning to be performed when mak-
ing the decision on whether to choose α or β. One well known
definition of equivalence between programs in PDL says that two
programs are equivalent if, provided the same input, they always
produce the same output. This definition helps to solve the question
on the procedural equivalence of the two graphs above, they are
equivalent because they e.g. lead us to A. But if we focus on the
second model, suppose that d wants to solve the tort by paying
with one cow (β), and p wants cash (α). Even when both actions
help to bring about the intended outcome (A), they are not the same
actions (for example, there may be a later need of a collateral or
incidental action to the process, for example, to sell out the cow.)
4.4 Principle of Procedural Economy
We mentioned that procedures are finite chains of actions. The
principle of economy says that if procedures are shorter, it is better
for all the agents involved. This because there is, e.g., a less amount
of resources used, especially, and possibly most important, less
procedural steps involved and therefore less time (and possibly less
effort) spent.
A form of procedural thinking involving the principle of econ-
omy can be carried out somehow easily through techniques
for reducing the complexity of graphs and comparing them us-
ing
 suppose that we have
(α p ; β p ) ⊗p γ p A and θ d ⊗d γ p A with α p , β p , γ p and θ d single
Modelling Legal Procedures
actions. The court should decide θ d ⊗k dp A as it leads to the
expected state-of-affairs A, but with less court activity.
A variant of the Principle of Procedural Economy is the Concen-
tration Principle, which means that agents ought to present together
all what it can be done in one step. The presentation of parties’
proposals as goal preferences, i.e. as a main goal and subsidiary
goals (as in [α n ⊗p β p ⊗p ⊗γ p ]A), is an example of the application
of this principle.
5 SPECIFIC RULES
5.1 Detachment of Obligations
From a resolution we can derive an obligation. Such an inference
rule reflects the lawful reading of expressions such as [Πi ⊗k Π 2 ]A.
If we have α ⊗p β, then α is obligatory in achieving A and β is to
be done in case α is not possible, because the authority imposes so.
Let us use O as a non-normal minimal modal deontic operator.
Technically then, from a ⊗k -expression we can derive a formula
which is in the scope of O. The following general form for this
high-level reasoning rule is:
! ! !!
Ì
n Ì
m Û
n
′ ′′
A → O[Π′ ]A
k Π i ⊗k Π ⊗k k Πi A ∧ ( ⟨Πi ⟩ ¬A)
i =1 i =1 i =1
(O-detachment)
which should be intuitively understood as: “The court’s preference
that holds is obligatory”, reflecting the intuitive reading of the ⊗k .
5.2 Consistency Check w.r.t. Court Resolutions
Suppose that we have the following court resolution:
[α ⊗k β ⊗k γ ]A. Assume that, later, the court also resolves
that [β ⊗k ϕ]A.
By application of the O-detachment rule we get O[α] A from the
first resolution, and also we get O[β] A from the second resolution.
Both detachments give us O[α] A ∧ O[β] A. From the procedural
point of view this is not what it is expected because, according
to the first resolution given, we should get O[α] A, but not O[β] A
(unless β is to be done because α cannot). This consistency conflict
raises when we analyse the resolutions in the framework of a lawful
procedure. The intuition behind the solution to this (lawful) incon-
sistency is that there is form of procedural reasoning that consists
on the temporal ordering of court resolutions and, following, the
alignment of the forthcoming resolutions with respect to the first
resolution given. Suppose the given resolutions are:
[α ⊗k β ⊗k γ ]A [β ⊗k ϕ]A [ϕ ⊗k ψ ]A.
Note that all three lead to the same state of affairs. We set the
first one as the conductor resolution, then align the rest of them
according to the preferred option in each resolution, as follows:
(Conductor) Resolution 1: [α ⊗k β ⊗k γ ]A
Resolution 2: [β ⊗k ϕ]A
Resolution 3: [ϕ ⊗k ψ ]A
Resultant Resolution 4: [α ⊗k β ⊗k (γ ∪ ϕ) ⊗k ψ ]A
We call this rule preference alignment. The intuition behind is the
following: indeed, court resolutions have a temporal ordering. Ac-
cording to this ordering, and starting with the first given resolution,
the remaining resolutions are aligned to the conductor resolution
with respect to their respective preferred option. That is, we put
224
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
the preferred option of every resolution in the column where the
same option appears in the conductor resolution. Following, all
the given resolutions are rejoined in one resulting main resolution,
column by column. In this course, component options in a column
are gathered together with the ∪ (choice) PDL operator.
A preference alignment algorithm is as follows:
procedure preference_alignment
(list_of_resolutions):result
begin
align resol_1 as the conductor resolution
for i:=2 to n do
if (preferred action in resol_i is in resol_1) then
position:= position in resol_1 of that preferred
action
align resol_i below resol_1 starting from position
end;
i := 1
result := false
while (there are actions in column i) do
begin
union := connect with ∪ all actions from resol_1 to
resol_n in column i
result := result ⊗k union
i := i+1
end;
end;
6 SUMMARY
A legal procedure in court proceedings is the formal way in which
civil proceedings are conducted. A legal procedure was defined
as a chain of consecutive actions which has as (final) goal the
decision/solution of a conflict, i.e., as a finite sequence of actions
in which the last action is (the creation of) a(n individual) norm,
usually, an obligation.
We argued that one peculiar aspect of proceedings is that some
types of procedure in the process are prioritised. Priorities indeed
derive from individual preferences of the parties in the process, or
they can also follow from objective ordering requirements from
procedures. In order to model legal procedures, in this paper we
technically added obligations and a preference operator for proce-
dural actions to a multi-agent version of PDL.
This paper is a preliminary research. For example, complexity
features and a full investigation on the effective application of the
proposed machinery are a matter of future work.
REFERENCES
[1] Erica Calardo, Guido Governatori, and Antonino Rotolo. Sequence semantics for
modelling reason-based preferences. Fundam. Inform., 158(1-3):217–238, 2018.
[2] Erica Calardo and Antonino Rotolo. Variants of multi-relational semantics for
propositional non-normal modal logics. J. Appl. Non Class. Logics, 24(4):293–320,
2014.
[3] David Harel, Jerzy Tiuryn, and Dexter Kozen. Dynamic Logic. MIT Press, Cam-
bridge, MA, USA, 2000.
[4] P. Luiso. Diritto processuale civile. Giuffre, 2019.
[5] Johan van Benthem. Program constructions that are safe for bisimulation. Stud
Logica, 60(2):311–330, 1998.
[6] Joachim Zekoll. Comparative civil procedure. In Mathias Reimann and Rein-
hard Zimmermann, editors, The Oxford Handbook of Comparative Law. Oxford
University Press, 2012.
Automatic Extraction of Amendments from Polish Statutory Law
Aleksander Smywiński-Pohl
Mateusz Piech
Zbigniew Kaleta
{apohllo,mpiech,zkaleta}@agh.edu.pl
AGH University of Science and Technology
Kraków, Poland
ABSTRACT
The article discusses the problem of automatic detection of amend-
ments found in the Polish statutory law. We treat the problem as a
token-classification task and we introduce a scheme constructed
by analysis of more than 200 amending bills. We apply recent neu-
ral architectures such as BERT and BiRNN to the task of token
classification. The achieved results of all models are very high as
micro average F1 score ranges from 96.3% to 98.2% for BiRNN. The
presented solution is a first step towards fully automatic structuring
and application of amendments in the Polish statutory law.
CCS CONCEPTS
• Computing methodologies → Information extraction; • Ap-
plied computing → Law.
KEYWORDS
amendment extraction, information extraction, named entity recog-
nition, legal information system, Polish statutory law
ACM Reference Format:
Aleksander Smywiński-Pohl, Mateusz Piech, Zbigniew Kaleta, and Krzysztof
Wróbel. 2021. Automatic Extraction of Amendments from Polish Statutory
Law. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY,
USA, 5 pages. https://doi.org/10.1145/3462757.3466141
1 INTRODUCTION
The legal system in Poland is based on statutory law passed by
the Polish parliament. The laws are published in the Journal of
Laws of the Republic of Poland and are available via ISAP 1 website
which distributes the laws as PDF files with metadata available
as HTML pages. Although the primary law is linked with all its
amendments, the user can only list all laws (but not amendments
specific to that law) that modified the given document. In the case
of the most important laws such as civil and criminal codes, the
consolidated text is also published yearly in the form of PDF files,
with specific regulations marked as recently amended. All in all, if
1 http://isap.sejm.gov.pl
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466141
225
Krzysztof Wróbel
krzysztof@wrobel.pro
Jagiellonian University
Kraków, Poland
Table 1: The types of entities annotated in the amendments.
Amendment type
add_content remove_content change_content
add_unit remove_unit change_unit change_id
Identifier
new_id amended_id preceding_id
Content
new_content old_content preceding_content
the user tries to use that website to track the changes of any law,
which underwent a number of amendments, the website is not very
useful.
Taking into account the fact that the source texts of the amend-
ments are weakly structured (i.e. PDF files) and the fact that the
number of amendments is tremendous, we investigate the possi-
bility to apply machine learning approach to the problem of the
automatic structuring of the text that would allow for converting
amending bills into structured data.
The contribution of the article is as follows. We start by casting
the problem of amendment extraction as token classification. We
introduce a scheme devised for detecting the amendments based
on an analysis of a large number of Polish amending laws. Then we
present three approaches to the problem of amendment extraction:
one based on rules and two others based on machine learning. We
pay special attention to the pre-processing of the data since it has
a huge impact on the obtained results. We discuss the related work
in Section 5. We conclude the article with prospects for the future
research.
2 ANNOTATION CORPUS
For this research we have created a corpus of 242 bills of Polish
statutory law from the years 1993-2018. Table 1 shows the types
of the textual units identified by analyzing the sample of bills. The
primary element used to define the amendment is its type. The set
of required elements and in some cases their exact meaning are
dependant on the type of amendment.
The amendments related to units (with unit suffix) are the basic
ones. Thus they require the id of the amended unit, the new content
(in the case of addition and change), and the id of the preceding
unit (in the case of addition).
The amendments related to the change in content (with content
suffix) are much more sophisticated, since they change the content
of the unit only partly. The most common case is addition, sub-
stitution or removal of a short phrase or a punctuation mark. But
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 2: The counts of tokens for different types of annota-
tions.
Annotation type Count
Amendment type 5 377
Identifier 9 747
Content 171 937
No annotation 230 300
Total 417 361
this action might be much more sophisticated, such as substitution
of certain phrase en masse, e.g. replacing „minister of sport” with
„minister competent for sport”.
The last type of amendment – change_id – is related to the
specific case when a part of text receives a new identifier: mainly
when an article containing one, unnumbered paragraph is extended
with a new paragraph, thus the existing paragraph receives an id.
Regarding the identifiers, we distinguish between the id of the
element when it is changed or removed (amended_id) and when
it is added, since in the second case the id of the preceding unit
has to be indicated and we have to distinguish between them. It
was assumed that only the most specific part of the identifier is
annotated.
The annotation was performed by two independent annotators
and then reviewed by a super-annotator with the help of Inforex
system [11]. The corpus contains roughly 4 hundred thousand
tokens and its size is comparable with the manually annotated
part of the National Corpus of Polish (1.2 million) [14]. The Table 2
summarizes the counts for different types of tokens. The dominating
class are the tokens without annotation (230 thousand tokens).
But still almost half of the corpus contains amendment-related
annotations. This is not a surprise, since the majority of the enacted
bills are amendments to the existing laws. The second most popular
annotation types are related to content (more than 170 thousand
tokens). This is expected, since these tokens constitute the actual
contents of the new and changed regulations. The total number of
individual amendments – partly reflected by the first group, since
amendment type is usually indicated by two tokens – is roughly
2.5 thousand (this is an estimation, since there are cases when one
amendment type relates to changes of more than one element).
We have divided the annotated documents into three subsets:
the training (~80%), the development (~10%) and the test (~10%) set.
We imposed two requirements when splitting the documents. First
of all, since the structure of the bill is crucial for the extraction of
amendments, we have split the whole documents rather than their
fragments, such as individual provisions. As a second requirement,
we have paid special attention to keep similar distribution of tags
in each of the subsets. It wasn’t straightforward, thanks to the first
requirement, but we managed to keep each type of tag in 62%–85%,
8%–23% and 6%–23% range for the training, the development and
the test subsets respectively.
3 AMENDMENT DETECTION
The detection of the amendments – treated as a NER-like problem
– was tested using the following approaches. The first one was
226
A. Smywiński-Pohl, M. Piech, Z. Kaleta, K. Wróbel
a rule-based approach employing regular expressions (a baseline,
henceforth called rules), the second one was an approach based on
the transformer architecture [19], with bi-directional encoding [5]
(henceforth called BERT ) and the third one was an approach based
on a bidirectional recurrent neural network (BiRNN) [16] with Long
short-term memory cells [8] (henceforth called BiRNN ). We have
tested three transformer models pre-trained on different corpora:
HerBERT on a large Polish corpus[15], RoBERTa on an English
corpus and XLMR on multilingual corpus [4]. The HuggingFace’s
Transformers library [21] was used in that setting. The bidirectional
RNN employed character-level language model (cLM) an was pre-
trained on Polish texts from the law domain. We have chosen the
FLAIR framework [2], for the pre-training and fine-tuning of that
model.
3.1 Rules
In order to fairly compare the system with the other approaches, we
have created the rules looking only at the training and development
datasets. The size of the “model” compared to the size of both RNN
and HerBERT is very favorable for that approach. We were unable
to define the regular expression for the preceding_content tag
(lack of common phrases in the examples), thus it is not detected
by this approach. A tag associated with the regex is assigned to
text span matched with that regex and the O tag is assigned to the
rest of text. In the case of multiple regular expressions matching
overlapping text spans, the more popular class, according to training
and development sets, is assigned. The matching was run from left
to right, thus it is continued right after the phrase matching the
assigned tag.
3.2 BERT
The second approach uses a transformer model for Polish called
HerBERT [15]. The model was trained on a large number of Polish
texts and it uses the RoBERTa pre-training optimizations [10]. The
model achieves SOTA results on KLEJ benchmark [15]2 , a collection
of Polish Natural Language Processing (NLP) tasks resembling the
GLUE benchmark [20]. The results on that benchmark are very
similar to those of the XLMR model [4], yet HerBERT is better
suited for Polish since the previous model was trained on a very
large number of languages (100), while HerBERT was trained only
on one language. As a result, HerBERT is much smaller and the
fine-tuning is faster compared to XLMR. Comparing to the previous
approach, we have not used any domain-specific texts to reduce
the perplexity of the model on texts coming from the law domain.
In the reported experiments we have used the large variant of the
model.
To measure the impact of language-specific pre-training, we
have included two other models: Roberta (large) [10] and the largest
multilingual model, i.e. XLMR [4]. Both of these models were trained
using the same approach, but they were trained for English and 100
languages respectively.
3.3 BiRNN
The last approach reflects a pretty recent SOTA model for NER.
Bidirectional RNNs with LSTM cells were very popular until the
2 Klej in Polish means glue in English.
Automatic Extraction of Amendments from Polish Statutory Law
transformer model dominated the NLP landscape. Also, the RNNs
have one important advantage compared to BERT – they accept
inputs of arbitrary length. BERT architectures are limited to 512
subtokens due to memory and time complexity which is quadratic
to the length of the input. This is important in the context of amend-
ment detection since we would prefer a model not requiring a so-
phisticated text segmentation (at least in the stage of amendment
detection), but rather accepting a complete text of the bill. The Flair
architecture assumes usage of a character-level language model pre-
trained on texts from a given domain. The model was pre-trained
on two corpora – the full National Corpus of Polish [13, 14] (approx.
2 billion tokens) and corpus containing the Polish statutory law
and the judgments of Polish courts (approx. 4 billion tokens).
4 EXPERIMENTS
We start the discussion of the experiments by presenting the re-
sults for the Rules model. For most of the types of elements the
rules work perfectly, yet there is a group of types (4 – new_id, 5 –
amended_id, 7 – preceding_id) that are confused for each other.
This is the reason why the results for the rules approach (cf. Table
3) gives low scores both in terms of micro, macro and weighted
average F1 measure. This shows that the identification of the ids is
problematic for this approach. We have to stress, that we have ex-
cluded from the results the false positives related to the O tag, since
that would yield a very large value of incorrectly determined identi-
fiers since they appear in a large number of contexts. Moreover, we
were unable to come up with a rule for detecting the 6th category
(i.e. preceding_content). We argue that the approach based on
rules is not universal enough to easily differentiate between the
various types of elements appearing in the amendments.
As a second approach, we tested the family of BERT models.
These models have a hard limit3 on the length of the input, which
amounts to 512 subtokens. It is obvious that the length of bills is
longer than that limit, but it might be the case that during training
it is not necessary to perform any pre-processing since all phenom-
ena are present at the beginning of the document4 . We have run
two preliminary experiments using the HerBERT model, in order
to determine if the pre-processing is needed. In the first, we have
submitted the full document as the input. This approach results in
a large number of truncated text (more than 60% in the training
set). In the second approach as input we took individual sentences
(obtained from the Inforex system that uses MorphoDiTa library
[18] for sentence boundary detection). This approach yielded com-
plete coverage of the annotated tokens since the detected sentences
were always shorter than the 512 limit.
The first approach resulted in 91.58% micro-average F1 score and
76.03% macro-average F1 score. The low macro-average score is a re-
sult of complete ignorance of the types of tags preceding_content
and remove_content, which were very rare both in the training
and the testing corpora. This outcome – as expected – falsifies the
assumption that the pre-processing is not needed.
Regarding the second approach, we observed that it resulted
in a larger number of examples (3-fold increase in the case of the
3 The limit is hard in the sense that the quadratic memory complexity makes longer
inputs prohibitively slow to process with the current available hardware.
4 That would also be a waste of a large number of annotations, yet we believe that it is
good to test even the simplest approach.
227
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 3: The results of the detection of amendments.
Model Micro F1 Macro F1 Weighted F1 Support
Rules 69.58 82.03 72.07 1335
HerBERT 97.96 90.81 97.92 1174
RoBERTa 97.69 97.68 97.70 1148
XLMR 96.72 84.62 96.73 1174
BiRNN 1 90.81 80.72 90.60 1773
BiRNN 2 98.20 98.90 98.19 1174
training set). Thanks to that it gave a much better macro-average
F1 score (94%), but the micro-average F1 score of 84.36% was far
below expectation. That was primarily due to low outcome for the
new_content class: 74.84% F1 score. This result is easy to explain,
since the new content usually contains a number of sentences which,
classified individually, lack the context required to determine if
they are part of the amending or the amended bill (they are quoted
in the second case, but the quotation spans a large number of
sentences). In fact, the 75% score seems to be very high if we take
that phenomenon into account.
As a result, we have decided to perform a more elaborate pre-
processing. The algorithm iterated through the lines in the CONLL
document. If a new sentence (i.e. an empty line) was detected, it
broke the sentence only if the preceding tag was O (i.e. the sentence
break was not inside an annotated span). In the other case, it broke
the sentence when the first O tag was detected. Such approach does
not ensure that the context is always preserved, but it plays well
with the new_content type, since these tokens always form long
spans of consecutive tokens.
Yet, during inference such a procedure is impossible to apply be-
cause we do not have the tags, we are going to predict. To overcome
that problem we implemented a procedure based on the structure
of the bill. We treated textual content of provisions at any level as
independent inputs for the algorithm. In the case the provision of
the lowest level was longer than 512 subtokens, it was truncated
and the tag assigned to the last token was assigned to the remaining
tokens besides the last one (which in most cases is a quotation).
Table 3 summarizes the results of an experiment conducted using
the models belonging to the BERT family using the more sophisti-
cated pre-processing. All of them were trained with the same set
of hyper-parameters (batch size: 8, epochs: 10, learning rate: 5e-6,
eps: 1e-8, maximum gradient: 1.0, weight decay: 0.0, max sequence
length: 512, seed: 0 and F1 score on the development set used for
early stopping). The results are reported for the test set.
The first observation relates to the number of testing examples
(support) available for each approach. As it was explained, there
were cases when the input had to be truncated. The different number
of examples stems from the fact that the models yield a different
number of subtokens5 , but the differences are small.
The second observation is that the performance of the models
is very good, especially if we look at the micro average scores.
XLMR – which is the worst according to that metric – is still almost
5 HerBERT uses a tokenizer trained on Polish texts, RoBERTa – on English texts and
XLMR on texts from 100 languages. Since the vocabulary size is limited, the Polish
tokenizer model may yield the lowest number of tokens, since HerBERT and RoBERTa
use the same vocabulary size. XLMR uses a dictionary which is 4 times larger.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
perfect, yielding 96.72% micro average F1 score. According to that
metric, the best results are obtained by HerBERT which yields 1.2
percentage point better result. Yet its macro average result (90.81%
F1 score) is pretty low, compared to the result given by RoBERTa
(97.78% F1 score). That result is a bit surprising, especially if we
recall that this model is pre-trained for English, while we analyze
texts in Polish. A closer inspection of the results shows that the
low outcome of HerBERT was due to lower scores for the detec-
tion of the content-related tags ({add,remove,change}_content).
If we take into account the fact that RoBERTa gives results only
0.27 pp. lower than HerBERT in terms of micro average F1 score,
we may conclude that for that problem it seems to be the best
option among the family of BERT models, even though it is not
pre-trained on the Polish texts. This result is particularly inter-
esting since it shows that we could leverage a more recent
model pre-trained only for English, especially since the pre-
training is a very costly procedure.
As the last type of approach, we present the results for BiRNN
model. Although RNN do not have a hard limit on the length, to
fairly compare their performance with the previous approach, we
have used the same pre-processing strategy. We have tested the
approach when the sentences are provided by MorphoDiTa (version
1) and when they are provided by the optimized version of the input
pre-processing (version 2). We have trained the model with the
following hyper-parameters: hidden layer size: 256, max epochs:
150, learning rate: 0.1, mini-batch size: 32, word embeddings: pl-
wiki-fasttext-300d-1M [12]. The results of the experiments are given
in Table 3. The comparison of the results produced by HerBERT
and BiRNN on the input provided by MorphoDiTa shows that – as
expected – the number of testing (and training) examples is much
higher for BiRNN, since the input is not truncated. The recurrent
architecture receives almost 3-times more examples. Interestingly
the model achieves a low value (80.72%) for the macro average F1
score (much lower than HerBERT). Inspection of the individual
classes showed that this is caused by 0.0% scores for new_id and
change_id tags, which are rare in the training set. The results
for new_content and old_content were also lowering the result,
since both of them were below 90%. Yet the results are surprising
compared to the BERT model since that model had a much larger
number of training instances.
The most interesting result is for the second setting, where the
customized pre-processing was applied. Since the input was trun-
cated, the number of training and testing examples was lowered by
approx. 33%. Yet this approach gave the best results overall – com-
paring both the simplified pre-processing and to the family of BERT
models. Both micro (98.2%) and macro (98.90%) average F1 scores
were almost perfect, meaning that the model was able to learn all
classes. That result shows that even though the approach based
on BERT has dominated the landscape of NLP problems, for the
amendment extraction problem – at least for the Polish language –
RNNs still might be a good alternative.
5 RELATED WORK
The approach used in this article – BIO tagging – is often used for
named entity recognition. One of the most commonly used methods
for NER is the bidirectional LSTM often used with word embeddings
228
A. Smywiński-Pohl, M. Piech, Z. Kaleta, K. Wróbel
and alike models. For example, the authors of [2] propose the use
of a character level language model to generate contextual string
embeddings. A bidirectional neural model is trained for the task of
predicting the next character in the sequence. The hidden states of
the forward and backward parts of the model, respectively after and
at the beginning of the analyzed word are concatenated to create
context-aware string embedding. In later work – [1] – they extend
this vector with second part, which is a function (e.g. element-wise
minimum, maximum, or mean) of all embeddings for the same
string including the newest one. One of our solutions (BiRNN) is
directly based on that work.
The authors of [22] have created LUKE – a language model
based on RoBERTa, where they train the embeddings for entities
alongside the embeddings for words. The input for training LUKE
is the concatenation of tokenized sentence and the list of all entities
present in this sentence. The training process is similar to BERT
and other masked language models – randomly selected parts of
the input (in this case words and entities alike) are masked and the
model is trained to predict those masked fragments. BERT training
is fully unsupervised, so it requires just a plain, unannotated corpus,
but LUKE needs an entity-annotated corpus. The authors used the
corpus from Wikipedia with good results. LUKE also extends the
self-attention mechanism normally used in transformers so that it
is entity-aware. Although that approach yields SOTA results for the
NER task, it does not apply to our problem, since the entities in our
approach are much different from typical entities in NER: they are
either short, constant phrases (e.g. is removed, is added), identifiers
(e.g. art. 5a, letter b) and long spans of text (the new content of the
amended provisions). Besides the first type, which is very easy to
detect, they belong to an open set of text phrases and cannot have
their own – learned – representations.
The work on automatic processing of legal amendments dates as
far as the work of Timothy Arnold-Moore in 1995[3]. There are two
main approaches to the problem of amendment processing. One
is to take two versions of the same legal act and compare them in
a diff-like manner. The second approach, presented herein, is to
generate an amended version of the legal act by using its previous
text and the text of the amending act.
The most common approach is to use syntactic and semantic
parsing, using a rule-based system. E.g. authors of [17] use a shal-
low syntactic parsing (chunking) using a battery of finite state
automata and a semantic analysis using a compiler based on spe-
cialized grammar. The system also contains an automatic classifier
that recognizes three kinds of amending provisions and discards
not amending provisions from further processing. In our work, we
expand the set of provisions by providing more fine-grained dis-
tinctions and introducing the new type related to the introduction
of an identifier.
The authors of [9] treat the task as a slots filling problem, where
the correct frame is chosen based on the verb and its dependents
using IF-THEN rules. In the case of multiple solutions, a heuristic is
used to pick the best one. They also address the problem of idioms:
some complex phrases common in legal documents are rewritten
using hand-crafted rules into a form that is easier to process by
further stages of the system.
In [7] the authors extend the number of types of modificatory pro-
visions and the article focuses on temporal modifications – changes
Automatic Extraction of Amendments from Polish Statutory Law
to either force or efficacy time. Unlike previous works in this field
they do not process all sentences of the amending act, but instead
filter them using regular expressions to increase both accuracy and
performance. They also introduce a sliding window to better handle
long and complex sentences
All of the aforementioned systems were devised for Italian legal
acts (and in result Italian language) and use NIR (Norme in Rete)
XML representation of the data.
The authors of [6] describe an attempt to make an automated
system for the consolidation of Greek legal acts, based on officially
published PDFs. This system uses regular expressions on multiple
stages of the process, including recognizing the amendment type
(addition, substitution, or deletion of a text portion) and extracting
the required data (filling the slots). If the amendment concerns
whole structural units, such as paragraphs, the change is applied to
the structure of the XML file containing the act. On the subpara-
graph level, the Python NLTK model is used to break down the
paragraph into units of an appropriate level.
6 CONCLUSIONS
We have presented a novel algorithm for detecting amendments
found in Polish statutory law. The primary difference presented
in this article and the previous work is the application of neural
models to the detection of amendment constituents. By treating
it as a token classification problem it is possible to use the most
recent SOTA models from the BERT family and a bit older BiRNNs.
It turns out that using such models provides very accurate results.
This observation supported by experiments conducted with the
most recent neural models: HerBERT, RoBERTa, XLMR, and BiRNN.
All of them yielded at least 96% for the micro average measure.
Yet, contrary to our expectations the best performance among the
family of BERT models was thanks to RoBERTa (a model trained
for English), while the best results were achieved by BiRNN whose
score was above 98% for each weighting scheme. This opens the
possibility for the automation of the amendment extraction in Pol-
ish.
Still, there are problems that have to be addressed to complete
that goal. First of all our solution is the first step in the automa-
tion pipeline, since the detected tokens has to be converted into
meaningful amendment representations. The solution thus requires
a proper structuring of the bill as well as automatic detection of
the references to the provisions. Both of these problems may be
resolved following the same approach, but they were out of the
scope of this article. A connected problem is the detection and
processing of temporal expressions that determine the application
date of a specific amendment. We will address these issues in the
forthcoming research.
7 ACKNOWLEDGMENTS
This work was supported by the Polish National Centre for Research
and Development – LIDER Program under Grant LIDER/27/0164/L-
8/16/NCBR/2017 titled “Lemkin – intelligent legal information sys-
tem” and in part by the PLGrid Infrastructure.
229
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
REFERENCES
[1] Alan Akbik, Tanja Bergmann, and Roland Vollgraf. 2019. Pooled Contextualized
Embeddings for Named Entity Recognition. In Proceedings of the 2019 Conference
of the North American Chapter of the Association for Computational Linguistics:
Human Language Technologies, Volume 1 (Long and Short Papers). Association for
Computational Linguistics, Minneapolis, Minnesota, 724–728. https://doi.org/10.
18653/v1/N19-1078
[2] Alan Akbik, Duncan Blythe, and Roland Vollgraf. 2018. Contextual string embed-
dings for sequence labeling. In Proceedings of the 27th International Conference
on Computational Linguistics. 1638–1649.
[3] Timothy Arnold-Moore. 1995. Automatically processing amendments to legis-
lation. In Proceedings of the 5th international conference on Artificial intelligence
and law. 297–306.
[4] Alexis Conneau, Kartikay Khandelwal, Naman Goyal, Vishrav Chaudhary, Guil-
laume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer,
and Veselin Stoyanov. 2019. Unsupervised Cross-lingual Representation Learning
at Scale. arXiv:cs.CL/1911.02116
[5] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
Pre-training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805 (2018).
[6] John Garofalakis, Konstantinos Plessas, and Athanasios Plessas. 2016. A Semi-
Automatic System for the Consolidation of Greek Legislative Texts. In Proceedings
of the 20th Pan-Hellenic Conference on Informatics (Patras, Greece) (PCI ’16).
Association for Computing Machinery, New York, NY, USA, Article 1, 6 pages.
https://doi.org/10.1145/3003733.3003735
[7] Davide Gianfelice, Leonardo Lesmo, Monica Palmirani, Daniele Perlo, and
Daniele P Radicioni. 2013. Modificatory provisions detection: a hybrid NLP
approach. In Proceedings of the Fourteenth International Conference on Artificial
Intelligence and Law. 43–52.
[8] Sepp Hochreiter and Jürgen Schmidhuber. 1997. Long short-term memory. Neural
computation 9, 8 (1997), 1735–1780.
[9] Leonardo Lesmo, Alessandro Mazzei, Monica Palmirani, and Daniele Radicioni.
2013. TULSI: an NLP system for extracting legal modificatory provisions. Artificial
Intelligence and Law 21 (05 2013). https://doi.org/10.1007/s10506-012-9127-6
[10] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A
Robustly Optimized BERT Pretraining Approach. arXiv:cs.CL/1907.11692
[11] Michał Marcińczuk, Marcin Oleksy, and Jan Kocoń. 2017. Inforex—A collaborative
system for text corpora annotation and analysis. In Proceedings of the international
conference recent advances in natural language processing, RANLP. INCOMA
Shoumen, 473–482.
[12] Agnieszka Mykowiecka, Małgorzata Marciniak, and Piotr Rychlik. 2017. Testing
word embeddings for Polish. Cognitive Studies 17 (2017).
[13] Piotr Pęzik. 2012. Wyszukiwarka PELCRA dla danych NKJP. In Narodowy
Korpus Języka Polskiego, Adam Przepiórkowski, Mirosław Bańko, Rafał Górski,
and Barbara Lewandowska-Tomaszczyk (Eds.). Wydawnictwo Naukowe PWN,
253–279.
[14] Adam Przepiórkowski, Mirosław Bańko, Rafał Górski, and Barbara Lewandowska-
Tomaszczyk. 2012. Narodowy Korpus J ezyka Polskiego. Wydawnictwo Naukowe
PWN.
[15] Piotr Rybak, Robert Mroczkowski, Janusz Tracz, and Ireneusz Gawlik.
2020. KLEJ: Comprehensive Benchmark for Polish Language Understanding.
arXiv:cs.CL/2005.00630
[16] Mike Schuster and Kuldip K Paliwal. 1997. Bidirectional recurrent neural net-
works. IEEE Transactions on Signal Processing 45, 11 (1997), 2673–2681.
[17] PierLuigi Spinosa, Gerardo Giardiello, Manola Cherubini, Simone Marchi, Giulia
Venturi, and Simonetta Montemagni. 2009. NLP-Based Metadata Extraction for
Legal Text Consolidation. In Proceedings of the 12th International Conference on
Artificial Intelligence and Law (Barcelona, Spain) (ICAIL ’09). Association for
Computing Machinery, New York, NY, USA, 40–49. https://doi.org/10.1145/
1568234.1568240
[18] Milan Straka and Jana Straková. 2014. MorphoDiTa: Morphological Dictionary
and Tagger. (2014).
[19] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All
You Need. arXiv:cs.CL/1706.03762
[20] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R
Bowman. 2018. GLUE: A multi-task benchmark and analysis platform for natural
language understanding. arXiv preprint arXiv:1804.07461 (2018).
[21] Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue,
Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie
Brew. 2019. HuggingFace’s Transformers: State-of-the-art Natural Language
Processing. CoRR abs/1910.03771 (2019). arXiv:1910.03771 http://arxiv.org/abs/
1910.03771
[22] Ikuya Yamada, Akari Asai, Hiroyuki Shindo, Hideaki Takeda, and Yuji Matsumoto.
2020. LUKE: Deep Contextualized Entity Representations with Entity-aware
Self-attention. arXiv:cs.CL/2010.01057
 A Dataset for Evaluating Legal Question Answering on Private
International Law
Francesco Sovrano Monica Palmirani
francesco.sovrano2@unibo.it monica.palmirani@unibo.it
University of Bologna - DISI University of Bologna - CIRSFID
Italy Italy
Salvatore Sapienza
salvatore.sapienza@unibo.it
University of Bologna - CIRSFID
Italy
ABSTRACT
International Private Law (PIL) is a complex legal domain that
presents frequent conflicting norms between the hierarchy of le-
gal sources, legal domains, and the adopted procedures. Scientific
research on PIL reveals the need to create a bridge between Euro-
pean and national laws. In this context, legal experts have to access
heterogeneous sources, being able to recall all the norms and to
combine them using case-laws and following the principles of inter-
pretation theory. This clearly poses a daunting challenge to humans,
whenever Regulations change frequently or are big-enough in size.
Automated reasoning over legal texts is not a trivial task, because
legal language is very specific and in many ways different from a
commonly used natural language. When applying state-of-the-art
language models to legalese understanding, one of the challenges is
always to figure how to optimally use the available amount of data.
This makes hard to apply state-of-the-art sub-symbolic question
answering algorithms on legislative texts, especially the PIL ones,
because of data scarcity. In this paper we try to expand previous
works on legal question answering, publishing a larger and more
curated dataset for the evaluation of automated question answering
on PIL.
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Reasoning about belief and knowledge; Information extrac-
tion.
KEYWORDS
Legal Question Answering, Private International Law, Knowledge
Graph Extraction
ACM Reference Format:
Francesco Sovrano, Monica Palmirani, Biagio Distefano, Salvatore Sapienza,
and Fabio Vitali. 2021. A Dataset for Evaluating Legal Question Answering
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466094
230
Biagio Distefano
biagio.distefano@univie.ac.at
Universität Wien
Austria
Fabio Vitali
fabio.vitali@unibo.it
University of Bologna - DISI
Italy
on Private International Law. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.3466094
1 INTRODUCTION AND BACKGROUND
The International Private Law (PIL) is a complex legal domain that
presents frequent conflicting norms between the hierarchy of legal
sources (e.g., national vs. European level), between legal domains
(e.g., consumer law vs. labour law), between the adopted procedures
(e.g., alternative dispute resolution vs. litigation). Scientific research
on PIL reveals the need to create a bridge between European and
national laws on this domain by accessing heterogeneous legal
sources. The European project Interlex1 intended to investigate this
domain and to use technology to fill the gap between different legal
sources. This need to rely on technology is due to the complexity
of the PIL domain. In fact, in this context, legal experts have to
access heterogeneous sources, being able to recall all the norms and
to analyse them using case-law2 and following the principles of
interpretation theory. This poses a daunting challenge to humans,
whenever Regulations change frequently or are big-enough in size.
In fact, searching within thousands and thousands of pages of legal
documents from different sources and jurisdictions is undoubtedly
a task requiring human effort and specialised expertise. This is prob-
ably one of the reasons why researchers, governments and industry
have for long looked for a way to build “intelligent” machines ca-
pable of helping humans in detecting the relevant legal provisions
over such complex corpora [11]. In literature we may find at least
two distinct main approaches to reasoning and artificial intelligence.
The first approach is more symbolic and formal, capable to model
legal knowledge into a formal representation. For example, the
legal ontology modelling method [4, 8] is a relevant instrument
for defining the legal concepts and relationships included in legal
texts (e.g., hard law, judgement, soft law, etc.) but it is extremely
expensive, it depends on the hermeneutic approach adopted by
each scholar or community (e.g., common law vs. civil law), it is
influenced by a strong localisation due to the local jurisdiction (e.g.,
domestic regulation and local court action), by the cultural and so-
cial norms (e.g., concept of gender) and, furthermore, modifications
in the legal framework (e.g., new legislation) require a refinement
or (even worse) a whole extension of the ontology is required. The
1 http://www.interlexproject.eu/index.html
2 http://www.interlexproject.eu/del/Deliverable2dot3.pdf
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
second approach is the most recent and in many ways the most
versatile, but sub-symbolic and opaque. A sub-symbolic approach
is said to be more data-oriented and it follows the recent success
of Deep Neural Networks (DNNs) on natural language processing
and understanding. Current state-of-the-art on natural language
understanding is heavily based on this data-centred approach, and
many models specifically applied to legalese have already been
published. For example in 2018 [3] published a framework for natu-
ral language processing and information extraction for legal and
regulatory texts. In 2019 [5] proposed one of the first models for
legal word embeddings. While, in 2015 Kim et al. [10] presented
one of the very first algorithms based on DNNs for Legal Question
Answering (reasoning) applied to a dataset of Boolean questions
from Japanese legal bar exams, then followed up by [7] and others
[9, 12]. In 2020, Sovrano et al. [13] proposed a novel and hybrid ap-
proach for legal question answering on PIL, using a legal ontology
based on Ontology Design Patterns (like agent, role, event, tempo-
ral parameter, action) in order to mirror the legal significance of
the relationships within and among the provisions. More generally,
automated reasoning over legal texts (not just the PIL’s ones) is
not a trivial task, due to the fact that the legal jargon (legalese) is
less frequent and more ambiguous than commonly-used natural
language. This is probably the reason why some works have de-
cided to focus on corpora, such as privacy policies [12], with a legal
language that would be more similar to its natural counterpart, or to
focus on more argumentative texts (e.g. sentences, procedural docu-
ments, cross-examinations, parliamentary court reports) instead of
legislative texts or contracts. Anyway, this challenge makes hard to
apply state-of-the-art sub-symbolic question answering algorithms
on legislative texts, especially the PIL ones, because of data scarcity
or novel topics introduced for the first time in the legal system (e.g.,
no historical series).
With this work we are interested into advancing on automated
answering to questions written in legalese and on PIL legislative
texts. Our goal is to be able to properly evaluate canonical question
answering techniques for PIL. This is why we try to expand the
work presented by Sovrano et al. in [13], publishing a larger and
more curated dataset extracted from Regulations such as: Rome
I Regulation EC 593/2008; Rome II Regulation EC 864/2007; and
Brussels I bis Regulation EU 1215/2012.
In Section 2 we describe our dataset, and the methodology we
followed to design it. While in Section 3 we analyse the results
obtained by re-running the experiment of [13] on the new dataset,
pointing to future work in Section 4.
2 A DATASET FOR EVALUATING LEGAL
QUESTION ANSWERING ON PIL
In this Section, we explain how we expanded the dataset presented
in [13], doubling its size. We improved over [13], publishing a larger
and more curated dataset for the evaluation of automated question
answering on PIL.
Both the old and the new dataset were extracted from the fol-
lowing Regulations, in English:
• Rome I Regulation EC 593/2008;
• Rome II Regulation EC 864/2007;
• and Brussels I bis Regulation EU 1215/2012.
231
Sovrano, et al.
These regulations are, respectively, on the law applicable to con-
tractual obligations; on the law applicable to non-contractual obli-
gations; on jurisdiction and the recognition and enforcement of
judgements in civil and commercial matters. These Regulations aim
to provide a tool for identifying the applicable law and the jurisdic-
tion in cases when two or more legal systems connect and generate
complex relationships (e.g. a sale of goods contract between an
Italian and a German citizen regarding commodities situated in
Spain).
It is important to highlight the fact that, for the construction of
the new dataset, we decided to inherit some methodological choices
from [13], considering PIL as a subject simply from the point of view
of these three EU Regulations, as a self-contained environment, i.e.,
excluding references to other international conventions and general
principles. So that it is possible to evaluate Q&A techniques with
respect to their ability to handle the general principles in the recitals,
the scope of application in the initial articles, and the specific cases
(e.g. exceptions) in the other articles. The methodological choices
we kept raised some issues with regard to the formulation of the
questions and their relevance. Conceptual questions (e.g. “What is a
non-contractual obligation?”) cannot be fully answered by relying
solely on these 3 Regulations, as the goal of this legislation - when
considered atomistically - is limited to discipline conflict of law and
conflict of jurisdiction cases. While the Regulations, as with any
other piece of legislation, rely somewhat on external definitions
and legal concepts, including those derived from jurisprudence
and opinions from commentators, they also define intrinsically and
specifically for their own purposes, key concepts (e.g. “judgement”
in Art. 2 of Reg. Brussels I-bis). Therefore we decided to exclude
any conceptual question but those involving key concepts defined
within the Regulations.
The legal question answering tools we are interested in evaluat-
ing are meant to be used by practising lawyers, with reasonable -
yet, not expert - knowledge of PIL to:
• explore the contents of the Regulations;
• get support in the reasoning concerning large Regulations.
The dataset for evaluating such tools shall comprise a set of ques-
tions for each of which there is also a set of expected answers in
the form of Articles, Recitals or Commission Statements3 . Recitals
are considered beside Articles because the user persona could be
interested in prima facie interpretive tools emerging from the text
itself, let alone the debated bindingness of Recitals. The dataset
published in [13], was designed following a methodology that is
similar to the one we are going to use for this extension. For the
selection of the questions and the identification of the expected
answers we adapted to our case a specific methodology encoded by
Ashley and others in their works [1, 2, 6] during the last years. This
methodology is common to other works in the field and it is meant
to validate the experiment also from a legal perspective. In our case,
the questions were selected by two legal experts, while other two
independent legal experts matching our intended user persona were
responsible for identifying the expected answers by relying solely
on the verbatim information that can be found in the Regulations.
3 Rome
II Regulation contains three Commission Statements meant to bind the EU
Commission to publish studies on selected topics
A Dataset for Evaluating Legal Question Answering on Private International Law
Therefore, legal experts were instructed to prevent case-law, gen-
eral principles or scholar opinions from influencing their answers,
as well as requested to avoid interaction with each other. As stated
above, the research wants to model only the neutral legislative
information from the three Regulations without any interpretation
other than the literal one. The inclusion of other knowledge will be
left to further research. First, experts read the three Regulations and
answered to the questions without any assistance from auxiliary
sources, including tools and previous knowledge. Then, they were
allowed to compare their answers with those provided by the tool
for legal question answering, selecting tool-assisted correct answers
and missing replies to be used to calculate performance scores in
the later stages. Despite the efforts to draft interpretation-neutral
questions, each independent expert has a certain margin of appre-
ciation both when providing her/his answers and when assessing
the correctness of the tool-provided answers. Therefore, another
intervention was necessary when divergences in their evaluation
occurred. When identifying the expected answers, the aggrega-
tion kept into account only theoretical replies that were common
between the two independent experts. This aggregation was con-
ducted by one legal expert who dispose of a higher level of expertise
in comparison to the independent evaluators, yet relying on the
same criterion, i.e. literal interpretation only.
At the end of the process we got the 9 new questions shown in
Table 2.
The questions were chosen with the following criteria: they had
to be sufficiently specific to find adequate answer in the Regula-
tions (we avoided too broad or excessively conceptual questions);
the questions needed not to be focused on specific cases but with
a reasonable level of abstraction (e.g., instead of “Where can an
employee that carries out their work in Spain sue an employee
located in Spain, if they had not agreed on the jurisdiction?”, a
question such as “Where can an employee sue their employer?”);
the questions needed to be sufficiently different from one another
(i.e., not asking repetitive questions such as “What is the applicable
law in contracts of carriage?” and “What is the applicable law in
insurance contracts?”). Some of the questions in the dataset are
relatively similar to one another, with some of them being the a
more correct specification of another, such as “Which parties of a
contract should be protected by conflict-of-law rules?” vs “What is
the applicable rule to protect the weaker party of a contract?”
Questions in the dataset are not speculative or de iure condendo
and are agnostic to elements that are placed outside the Regulation
(e.g. jurisprudence, general principles, etc.). As such, they are not
meant to nudge towards forms of interpretations other than the
literal one (e.g. analogy, principle-based reasoning, lex specialis,
etc.)
Furthermore, in order to be able to further analyse the results
of any evaluation based upon our dataset, we decided to pick an
heuristic for classifying questions, that is the context specificity
(Low, Normal, High), and we applied it also to the old dataset we
extended. Context specificity is a subjective concept and it is highly
dependant on each jurist. For this reason, we opted to use a cri-
terion that would ensure an acceptable level of objectivity. Thus,
specific questions whose answer is exactly in the domain of the
Regulations were labelled as Highly specific (e.g., “Can the parties
232
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
choose a different applicable law for different parts of the con-
tract?”); questions whose answer falls in part within the scope of
the Regulations, but somewhat relies on external concepts were la-
belled as Normally specific (e.g., “Which parties of a contract should
be protected by conflict-of-law rules?”); finally, broad questions
whose answer requires the significant use of external legal concepts
and resources and whose answer is found through an articulate
combinations of articles and recital were labelled as having Low
specificity (e.g., “How should a contract be interpreted according
to this regulation?”).
Of the 17 questions that compose the new extended dataset:
29.41% have a Low specificity; 35.29% have a Normal specificity;
35.29% have a High specificity.
3 DATASET ANALYSIS
In order to understand the behaviour of existing question answer-
ing tools on the new dataset, we repeated on it the experiment
described in [13] ,changing the metrics used for the evaluation.
Considering that we are not interested in the order answers are
ranked, as metric for estimating the performance of the algorithm
we chose: top5-recall, top5-precision and top5-F1, defined as follows.
Let 𝑚 be the number of strictly-correct answers that are produced
as output by the algorithm, let |𝐸| be the number of expected an-
swers for a question, let |𝐴| be the number of given answers to
a question, then the top5-recall is given by 𝑚𝑖𝑛 𝑚 ( |𝐸 |,5) , while the
top5-precision is given by 𝑚𝑖𝑛 (𝑚|𝐴 |,5) . The top5-recall is a measure
of how many relevant answers are selected by the algorithm in
the top five answers, while the top5-precision is a measure of how
many selected answers in the top five are relevant. Knowing the
top5-recall and the top5-precision, it is easy to compute the top5-F1
score by following the formula.
After running the experiment we computed the average top5-
F1 for all the questions in the dataset presented in Section 2 (that
is the old dataset of [13] plus our new extension). The results on
the whole dataset are a Top5-Recall of 37.58%, a Top5-Precision of
45.17% and a Top5-F1 of 38.05%.
We also performed an error analysis taking under consideration
how top5-F1 scores vary when the context specificity change, ex-
pecting that questions with low context specificity are harder to
answer correctly.
Results partly confirmed our expectations. In fact, we can observe
a trend where top5-F1 scores increase proportionally to the context
specificity. Our expectations were based on the fact that:
• the specificity of a question is low when it asks something
that is not closely related to the Regulations;
• multi-hop reasoning is usually required to answer questions
with a low specificity, but the baseline is not equipped for
that kind of reasoning (yet).
For example, the question “How should a contract be interpreted
according to this regulation?” has a very low specificity and it would
probably require to pinpoint both recitals and articles for a proper
answer, therefore more distinct and distant paragraphs. Probably,
most of the speculative questions would require a broader view
on the subject matter, having a low specificity to the Regulations,
therefore requiring multi-hop reasoning.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Sovrano, et al.

Table 1: First block of answers (ordered by the pertinence to the question estimated by the tool) given by the baseline to the
questions in [13]. “B” stands for Brussels, “RI” for Rome I and “RII” for Rome II. “Rec.” stands for Recital, “Art.” for Article, and
“Stat.” for Commission Statement. For each answer, the top5 scores (precision, recall, F1) are shown. In the “scores“ columns:
“P” stands for Precision and “R” stands for Recall. In the “Specificity“ column: “L” stands for Low, “N” stands for Normal and
“H” stands for High.

Question Speci Expected Answers Baseline’s Top5 Baseline’s

ficity Scores
R: 25%
Who determines disputes under a con- L B Art. 7.1, B Art. 8.3, B Art. 8.4, B RI Rec. 12, B Art.17.2, RI Rec. P: 33%
tract? Art. 17 24 F1: 28.44%
R: 25%
What factors should be taken into account N B Art. 7.1, B Art. 17, B Art. 20, B Art. RI Rec. 12, B Art.25, B P: 40%
for conferring the jurisdiction to deter- 25 Art.25.5, B Rec.15, RI Rec. 21 F1: 30.76%
mine disputes under a contract?
R: 25%
Which parties of a contract should be pro- N RI Rec. 23, RI Art. 6, RI Art. 8, RI RI Rec. 23, B Rec.18, RI Rec. P: 20%
tected by conflict-of-law rules? Art. 13 24, RI Art.25.1, RI Rec. 27 F1: 22.22%
R: 33%
In which case claims are so closely con- H B Art. 8, B Art. 30, B Art. 34 B Art. 8.1 P: 100%
nected that it would be better to treat them F1: 49.62%
together in order to avoid irreconcilable
judgments?
R: 20%
What kind of agreement between parties L B Rec. 6, B Rec. 10, B Rec. 12, B Art. B Art.73.3, B Rec. 12, B Rec. P: 20%
are regulated by these Regulations? 1, RI Rec. 7, RI Art. 1 36, B Art.71.2, B Art. 71.1 F1: 20%
R: 66%
In which court is celebrated the trial in H B Art. 21, B Art. 22, B Art. 23 B Art. 21.1, B Art.22.1, B Art. P: 60%
case the employer is domiciled in a Mem- 21.2, B Art. 20.1, B Art. 20.2 F1: 62.85%
ber State?
R: 0%
How should a contract be interpreted ac- L RI Rec. 22, RI Rec. 12, RI Rec. 26, RI RI Art. 10.1, RI Rec.17 P: 0%
cording to this regulation? Rec. 29, RI Art. 12 F1: 0%
R: 60%
Which law is applicable to a non- N RII Rec. 17, RII Rec. 18, RII Rec. RI Art. 8.1, RII Art.15, RII P: 60%
contractual obligation? 26, RII Rec. 27, RII Rec. 31, RII Art. Art.16, RII Art.8.1, RII Rec. F1: 60%
4-20 22

4 CONCLUSIONS ontology5 for better handling: obligations, permissions, exceptions,

With this paper we extended the work presented by Sovrano et al.
in [13], proposing a larger and more curated dataset for the evalua-
tion of automated question answering on PIL. In the future we will
use these datasets for evaluating new algorithms for question an-
swering, exploiting Akoma Ntoso XML4 models of the Regulations,
for better capturing the relationships between different portions of
the legal hierarchy (e.g. recitals connected via metadata to articles)
and also for reusing as much as possible other legal metadata like:
i) temporal legal information concerning modifications occurred
over time, ii) life-cycle information concerning the history of the
regulations, iii) normative references (citations). We also intend
to make the question answering tool “aware” of the LegalRuleML
4 http://docs.oasis-open.org/legaldocml/akn-core/v1.0/akn-core-v1.0-part1-
vocabulary.html
derogations, prohibitions.
ACKNOWLEDGEMENTS
This paper is was conducted with the contribution of CIRSFID-
Alma AI and DISI University of Bologna (Interlex Project Grant
Agreement Number 800839 and LAILA PRIN2017). The questions
were selected by Biagio Distefano, Salvatore Sapienza, while Pier
Giorgio Chiara and Noemi Conditi picked the expected answers,
separately. All of them are PhD candidates at the "Law, Science and
Technology" International PhD program.
REFERENCES
[1] Kevin D Ashley. 2017. Artificial intelligence and legal analytics: new tools for law
practice in the digital age. Cambridge University Press.
[2] Trevor Bench-Capon, Michał Araszkiewicz, Kevin Ashley, Katie Atkinson, Floris
Bex, Filipe Borges, Daniele Bourcier, Paul Bourgine, Jack G Conrad, Enrico
5 http://docs.oasis-open.org/legalruleml/legalruleml-core-spec/v1.0/cs02/rdfs/

233
A Dataset for Evaluating Legal Question Answering on Private International Law ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 2: Second block of expected answers and answers given by the baseline. See the caption of Table 1 for more details about
how to read this table.

Questions Speci Expected Answers Baseline’s Top5 Baseline’s

ficity Scores
Can the parties choose the applicable law H RI Rec. 11, RI Rec. 25, RI Rec. 27, RI B Art. 18.2, B Art. 18.1, RI R: 25%
in consumer contracts? Art. 6 Rec. 28, RI Art. 5.2, RI Art. P: 20%
6.2 F1: 22.22%
What factors should be taken into account N B Rec. 18, B Art. 17, B Art. 18, B Art. RI Rec. 12, RI Rec. 24, B Art. R: 40%
for conferring the jurisdiction to deter- 19, B Art. 26 19, B Art. 17.1, B Art. 25.5 P: 40%
mine disputes under a consumer contract? F1: 40%
Can the parties choose a different applica- L RI Rec. 11, RI Art. 3.1 RI Art. 3.1, RI Art. 5.2, RI Art. R: 50%
ble law for different parts of the contract? 7.3, RII Art. 25.2, RI Art. 22.2 P: 20%
F1: 28.57%
What non-contractual obligations fall into H RII Rec. 10, RII Rec. 11, RII Art. RII Stat. 1, RI Rec. 7 R: 0%
the scope of Regulation Rome II? 1, RII Art. 2 P: 0%
F1: 0%
What is the applicable rule to protect the N RI Rec. 23, B Rec. 18 RI Rec. 23, B Rec. 18 R: 100%
weaker party of a contract? P: 100%
F1: 100%
What is the applicable law to determine L RI Art. 3.5, RI Art. 10, RI Art. 11, RI RI Art. 3.5, RI Art. 10.2, RI R: 50%
the validity of consent? Art. 13 Art. 10.1, B Rec. 20 P: 75%
F1: 60%
When are two actions to be considered re- N B Rec. 21, B Art. 30.3 R: 0%
lated according to the Regulation Brussels P: 0%
I Bis? F1: 0%
What court has jurisdiction in case of a N B Art. 8.3, B Art. 14.2, B Art. 18.3, B B Art. 18.3, B Art. 14.2, B Art. R: 100%
counter-claim? Art. 22.2 22.2, B Art. 8, B Art. 24 P: 80%
F1: 88.88%
Where can an employee sue their em- H B Rec. 14, B Rec. 18, B Art. 21.1, B B Art. 21.1 R: 20%
ployer? Art. 22.1, B Art. 23 P: 100%
F1: 33.33%

[4] Pompeu Casanovas, Monica Palmirani, Silvio Peroni, Tom Van Engers, and Fabio
Figure 1: Average top5-F1 scores for each class of con- Vitali. 2016. Semantic web for the legal domain: the next step. Semantic Web 7, 3
(2016), 213–227.
text specificity: Low, Normal, High. Scores are respectively: [5] Ilias Chalkidis and Dimitrios Kampas. 2019. Deep learning in law: early adaptation
27.40%, 42.16%, 42.81% and legal word embeddings trained on large corpora. Artificial Intelligence and
Law 27, 2 (2019), 171–198.
0.5 [6] Jack G Conrad and John Zeleznikow. 2013. The significance of evaluation in
AI and law: a case study re-examining ICAIL proceedings. In Proceedings of the
0.42 Fourteenth International Conference on Artificial Intelligence and Law. 186–191.
[7] Phong-Khac Do, Huy-Tien Nguyen, Chien-Xuan Tran, Minh-Tien Nguyen, and
Minh-Le Nguyen. 2017. Legal question answering using ranking SVM and deep
convolutional neural network. arXiv preprint arXiv:1703.05320 (2017).
Top5-F1

0.27 [8] Meritxell Fernández-Barrera and Giovanni Sartor. 2011. The legal theory perspec-
tive: doctrinal conceptual systems vs. computational ontologies. In Approaches
to Legal Ontologies. Springer, 15–47.
[9] Nils Holzenberger, Andrew Blair-Stanek, and Benjamin Van Durme. 2020. A
Dataset for Statutory Reasoning in Tax Law Entailment and Question Answering.
arXiv preprint arXiv:2005.05257 (2020).
[10] Mi-Young Kim, Ying Xu, and Randy Goebel. 2015. A convolutional neural network
0 in legal question answering. In JURISIN Workshop.
L N H [11] Friedrich V Kratochwil. 1991. Rules, norms, and decisions: on the conditions
Specificity of practical and legal reasoning in international relations and domestic affairs.
Number 2. Cambridge University Press.
[12] Abhilasha Ravichander, Alan W Black, Shomir Wilson, Thomas Norton, and
Norman Sadeh. 2019. Question answering for privacy policies: Combining com-
putational and legal perspectives. arXiv preprint arXiv:1911.00841 (2019).
Francesconi, et al. 2012. A history of AI and Law in 50 papers: 25 years of
[13] Francesco Sovrano, Monica Palmirani, and Fabio Vitali. 2020. Legal Knowledge
the international conference on AI and Law. Artificial Intelligence and Law 20, 3
Extraction for Knowledge Graph Based Question-Answering. In Legal Knowledge
(2012), 215–319.
and Information Systems: JURIX 2020. The Thirty-third Annual Conference, Vol. 334.
[3] Michael J Bommarito II, Daniel Martin Katz, and Eric M Detterman. 2018. LexNLP:
IOS Press, 143–153.
Natural language processing and information extraction for legal and regulatory
texts. arXiv preprint arXiv:1806.03688 (2018).

234
 Discovering the Rationale of Decisions:
Towards a Method for Aligning Learning and Reasoning
Cor Steging Silja Renooij
c.c.steging@rug.nl s.renooij@uu.nl
Bernoulli Institute of Mathematics, Department of Information and
Computer Science and Artificial Computing Sciences,
Intelligence, University of Groningen Utrecht University
ABSTRACT
In AI and law, systems that are designed for decision support should
be explainable when pursuing justice. In order for these systems
to be fair and responsible, they should make correct decisions and
make them using a sound and transparent rationale. In this paper,
we introduce a knowledge-driven method for model-agnostic ratio-
nale evaluation using dedicated test cases, similar to unit-testing
in professional software development. We apply this new quantita-
tive human-in-the-loop method in a machine learning experiment
aimed at extracting known knowledge structures from artificial
datasets from a real-life legal setting. We show that our method
allows us to analyze the rationale of black box machine learning
systems by assessing which rationale elements are learned or not.
Furthermore, we show that the rationale can be adjusted using
tailor-made training data based on the results of the rationale eval-
uation.
CCS CONCEPTS
• Applied computing → Law.
KEYWORDS
Learning knowledge from data, Explainable AI, Responsible AI,
Machine Learning
ACM Reference Format:
Cor Steging, Silja Renooij, and Bart Verheij. 2021. Discovering the Rationale
of Decisions: Towards a Method for Aligning Learning and Reasoning.
In Eighteenth International Conference for Artificial Intelligence and Law
(ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA,
5 pages. https://doi.org/10.1145/3462757.3466059
1 INTRODUCTION
In AI and Law, explainability is a key requirement in system design,
due to the need for the justification of decisions. For machine-
supported decisions, this is encoded in the GDPR’s right to ex-
planation. Four types of explanations can be distinguished, all of
which have been applied to AI and Law [Atkinson et al. 2020a]:
contrastive explanations [Ashley 1990; Rissland and Ashley 1987;
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466059
235
Bart Verheij
bart.verheij@rug.nl
Bernoulli Institute of Mathematics,
Computer Science and Artificial
Intelligence, University of Groningen
Verheij 2003a], selective explanations [Atkinson et al. 2020b; Ver-
heij 2003b], probabilistic explanations [Vlek et al. 2016] and social
explanations [Atkinson et al. 2020b; Gordon 1995; Hage et al. 1993].
This requirement of explainability is problematic for the ap-
plication of central machine learning techniques in law. Neural
networks, for example, are known to perform well, but behave like
a black box algorithm. Hence, explanation techniques have been
developed to ‘open the black box’ (cf. LIME [Ribeiro et al. 2016],
SHAP [Lundberg and Lee 2017]). Even in the domain of vision
(where the successes of deep learning are especially significant),
the necessity of such methods is underpinned by studies regarding
adversarial attacks that show that slight perturbations of images,
invisible to the human observer, can radically change the outcome
of a classifier [Goodfellow et al. 2015].
In this paper, we therefore evaluate black box machine learning
methods with a focus on proper explainability, and not only in
terms of accuracy as in the standard machine learning protocol. We
are in particular interested in evaluating the discovered rationale
underlying decisions, where the rationale is the knowledge struc-
ture that can justify a decision, such as the rule applied. We aim
to measure the quality of rationale discovery, with an eye on the
possibility of improving rationale discovery.
Our work builds on a study investigating whether neural net-
works are able to tackle open texture problems [Bench-Capon 1993]
(also investigated in [Možina et al. 2005; Wardeh et al. 2009]). To
measure and possibly improve rationale discovery, we create dedi-
cated test datasets, on which a machine learning system can only
perform well if it has learned a particular component of the knowl-
edge structure that defined the data. The idea is similar to how unit
testing works in professional software development: we define a
set of cases, targeting a specific component, in which we know
what the answer should be, and compare that to the output that
the system gives.
In order to focus on what is methodologically feasible, we do not
use natural language corpora (such as conceptual retrieval [Grab-
mair et al. 2015], argument mining [Mochales Palau and Moens
2009; Wyner et al. 2010] or case prediction [Ashley 2019; Brüning-
haus and Ashley 2003; Medvedeva et al. 2019]). Instead we work
with datasets of artificial decisions with known underlying gener-
ating rationale. Other earlier discussions of neural networks in law
are [Hunter 1999; Philipps and Sartor 1999; Stranieri et al. 1999].
2 REPLICATION EXPERIMENT
The first step towards developing our method for rationale evalua-
tion was replicating the study by Bench-Capon [1993]. This was
done using modern, widely-used neural network methods and with
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Steging et al.

significantly larger datasets, in order to reaffirm that the claims

made in 1993 still hold today. The study introduces a fictional legal
domain, where the eligibility for a welfare benefit for elderly citi-
zens is determined by the conjunction of six independent conditions.
Artificial datasets were generated specifying personal information
of elderly citizens with their eligibility for the welfare benefit. Mul-
tilayer perceptrons were trained and tested on these datasets, and
managed to achieve high accuracy scores (above 98%).
Using special test datasets, it was shown that the neural net-
works were unable to properly learn the first and the last of the
six conditions. Furthermore, the networks performed significantly
worse when ineligibility was caused by the failure of only a sin-
gle condition. The training data was therefore altered such that
ineligible people only failed on a single condition, rather than on
multiple conditions as in the original training dataset. By making
these adjustments to the training dataset, the neural networks were
able to learn conditions more adequately, while maintaining similar
accuracy scores. However, even after adjustment, the conditions
that defined the data were not learned perfectly.
In our replication, we discovered that even with more data and
modern, commonly used neural networks, the nets are still un- Figure 1: Arguments and attacks (A) and their elementary
able to learn all six conditions that define eligibility, despite high propositions (B) in Dutch tort law [Verheij 2017].
accuracies (99%). Using two dedicated datasets (as also defined
in [Bench-Capon 1993]), it was shown that the nets still did not 𝑑𝑢𝑡 (𝑥) ⇐⇒ 𝑐 1 (𝑥) ∧ 𝑐 2 (𝑥) ∧ 𝑐 3 (𝑥) ∧ 𝑐 4 (𝑥) ∧ 𝑐 5 (𝑥)
learn condition the first and last condition. The rationale of the 𝑐 1 (𝑥) ⇐⇒ cau(𝑥)
nets is therefore not sound, despite high accuracies. Just as in the 𝑐 2 (𝑥) ⇐⇒ ico(𝑥) ∨ ila(𝑥) ∨ ift (𝑥)
original study, adjusting the training data based on expert knowl- 𝑐 3 (𝑥) ⇐⇒ vun(𝑥) ∨ (vst (𝑥) ∧ ¬jus(𝑥)) ∨ (vrt (𝑥) ∧ ¬jus(𝑥))
edge of the domain significantly improves the rationale of the net 𝑐 4 (𝑥) ⇐⇒ dmg(𝑥)
without seriously impacting the accuracy. 𝑐 5 (𝑥) ⇐⇒ ¬(vst (𝑥) ∧ ¬prp(𝑥))
Additionally, we created a simplified version of the domain, con-
where the elementary propositions are provided alongside an argu-
taining only the first and last condition, to see how well the net-
mentative model of the law in Figure 1 [Verheij 2017], and condi-
works are able to extract a simplified rationale. In this domain, the
tions 𝑐 2 and 𝑐 3 capture the legal notions of unlawfulness (unl) and
networks were able to learn both conditions.
imputability (imp) respectively.
The methods and results of the replication experiment as well
Compared to the fictional welfare domain in [Bench-Capon 1993]
as its variations, can be found in detail in [Steging et al. 2021].
and our replication variations [Steging et al. 2021], the Dutch tort
law domain is captured in 5 conditions for duty to repair (dut), based
3 TORT LAW: DOMAIN AND DATASETS
upon 10 Boolean features. Each condition is a disjunction of one
Following up on the fictional welfare benefit domain, we study or more features, possibly with exceptions. The feature capturing
a non-fictional legal setting, namely Dutch tort law. This domain a violation of a statutory duty (vst) is present in both condition 𝑐 3
uses only Boolean variables, but allows for exceptions to underlying and 𝑐 5 , rendering these dependent.
rules. This section describes the underlying knowledge structure
of the Tort Law domain using logic, from which we will generate 3.2 Datasets
datasets to train a series of neural networks. These networks will
We generate four different types of datasets, each for different
subsequently be analysed using a method we propose for assessing
purposes.1 For most types of datasets, the generating process is
the quality of their rational discovery. To this end we need two types
at least partly stochastic and repeated for every repetition of an
of datasets for the purpose of testing. The first are standard test
experiment. Using the same type of dataset, for example in training
sets sampled from the complete domain to evaluate the accuracy
and testing a neural network, does therefore not mean that the
of the networks. The second type is a dedicated test set designed
exact same dataset was used in both training and testing. Table 1
to target a specific aspect of the domain knowledge. This section
shows an overview of the datasets of the tort law domain.
describes all datasets we use.
With 10 Boolean features there are 210 = 1024 possible unique
cases that can be generated from the argumentation structure of
3.1 Domain
the tort law domain in Figure 1. Each case has a corresponding
Our domain concerns Dutch tort law: articles 6:162 and 6:163 of the outcome for dut, indicating whether or not there is a duty to repair
Dutch civil code that describe when a wrongful act is committed someone’s damages.
and resulting damages must be repaired. This ‘duty to repair’ (𝑑𝑢𝑡)
1 The
can be formalised as follows: Jupyter notebooks used for generating the data can be found in the following
Github repository: https://github.com/CorSteging/DiscoveringTheRationaleOfDecisions

236
Discovering the Rationale of Decisions
Table 1: An overview of the tort law datasets. Datasets
marked with an asterisk are used for testing purposes only.
For each type of dataset, the size and label distribution is
given.
T/F label
Dataset Size distribution
Regular 5,000/500 50%/50%
Unique* 1024 10.94%/89.06%
Unlawfulness* 168 66.67%/33.33%
Imputability* 128 87.5%/12.5%
The unique dataset contains these 1024 unique instances for
the 10 features plus the label. In this dataset, there are 912 instances
where dut is false and 112 instances where dut is true (10.94%).
The regular type datasets are generated such that dut is true
in exactly half of the instances. The sets are regular in the sense
that balanced label distributions are common in machine learning
problems. These regular datasets are generated by sampling uni-
formly from the subset of cases from the unique dataset, such that
each possible case is represented equally within the 50/50 label
distribution. In a typical machine learning experiment, only a sub-
set of the possible cases is typically available and presented to a
network, upon which the network will have to learn to generalize
to all possible cases. In addition to generating regular type datasets
with 5,000 cases, we therefore also generate smaller regular type
datasets with only 500 instances; the latter contains 35.35% of the
unique instances.
In the tort law domain we focus on the notions of unlawfulness
(𝑐 2 ) and imputability (𝑐 3 ) to assess whether the networks are able
to discover conditions in the data. For each of the two conditions,
we create a dedicated dataset.
The Unlawfulness dataset is the subset of the unique dataset in
which the features for the unlawfulness condition 𝑐 2 can take on
any of their values, while the other features have values that are
guaranteed to satisfy the remaining conditions. Whether or not
there is a duty to repair is therefore solely determined by whether
or not condition 𝑐 2 is satisfied. All combinations of values of the
other features are considered. The Unlawfulness dataset therefore
consists of 168 unique instances, of which 66.66% have a positive
𝑑𝑢𝑡 value.
The Imputability dataset is a similar subset of the unique dataset,
but now the features for the imputability condition (𝑐 3 ) can take on
any value, except that the value of vst must be such that condition
𝑐 5 is satisfied. The value of 𝑑𝑢𝑡 (𝑥) is now completely dependent on
whether or not condition 𝑐 3 evaluates to true. Due to the interde-
pendency of conditions 𝑐 3 and 𝑐 5 , the Imputability dataset only has
128 unique instances, 87.5% of which have a positive 𝑑𝑢𝑡 value.
4 EXPERIMENTAL SETUP AND RESULTS
In this section we describe and motivate the experiments we per-
formed for the tort law domain and report on their results.
4.1 Experiments
We decide to use neural networks like in [Bench-Capon 1993].
The method is model-agnostic, however, meaning that it can be
237
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
applied to any other machine learning model as well. We assume
that assessing and improving rationale discovery is relevant only
for models that perform well on their respective task. Our first step,
after training the above mentioned neural networks, is therefore
to evaluate their performance on typical test sets in terms of the
standard accuracy measure. Subsequently we will evaluate the
performance of the networks on the dedicated, knowledge-driven
test sets that were specifically designed for assessing the networks’
quality of rationale discovery.
4.1.1 Neural network architectures. Similar to the original experi-
ments, three multilayer perceptrons were used with one, two and
three hidden layers, respectively [Bench-Capon 1993]. The nets
all have 10 input nodes, corresponding to the number of features
and a single output node, representing duty to repair. The node
configuration (i.e. number of nodes per layer) of each network is as
follows:
• One hidden layer network: 10-12-1
• Two hidden layer network: 10-24-6-1
• Three hidden layer network: 10-24-10-3-1
We use the MLPClassifier of the scikit-learn package [Pedregosa
et al. 2011], the sigmoid function as the activation function, the
Adam stochastic gradient-based optimizer [Kingma and Ba 2015],
with a constant learning rate of 0.001. A total of 50,000 training
iterations are used with a batch size of 50. Recall that the focus
of this study is not on creating the best possible classifier, but to
assess rationale discovery.
4.1.2 Training and performance testing. The three types of neural
networks are trained and tested on all combinations of different
datasets from Table 1. Every combination of training dataset and
testing dataset is evaluated in terms of the accuracy of the resulting
network on the test data. Because some of the datasets are stochas-
tic (each generated dataset is slightly different), the whole process
of data generation, training and testing is repeated 50 times. The
mean classification accuracies along with their standard deviations
are reported. To assess the rationale discovery capabilities of all the
trained networks, we study their performance on the dedicated test
sets for unlawfulness and imputability conditions. Performance is
measured both quantitatively, using standard accuracy, and qual-
itatively by a more detailed comparison of actual and expected
outcomes.
4.2 Results
Table 2 shows the mean classification accuracies over 50 runs, to-
gether with their standard deviations, for the different combinations
of training and testing sets in the tort law domain. The table includes
the quantitatively measured performance on the two dedicated test
sets.
We can evaluate how well conditions 𝑐 2 (unlawfullness) and 𝑐 3
(imputability) are learned. For these conditions, the network should
output 1 in cases from the Unlawfulness dataset where the case is
unlawful (𝑐 2 ), or in the Imputability dataset where the case can be
imputated to a person (𝑐 3 ); otherwise the output should be 0. The
mean output of the 3 layer network over 50 runs for the two training
sets on the Unlawfulness and Imputability datasets is presented in
Table 3.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Steging et al.

Table 2: The accuracies obtained by the neural networks in the tort law domain.

Trained on all instances Trained on smaller dataset

General Unique Unlawfulness Imputability General Unique Unlawfulness Imputability
1 hidden layer 100±0 100±0 100±0 100±0 98.45±0.5 97.24±0.89 92.8±3.47 91.22±4.04
2 hidden layers 100±0 100±0 100±0 100±0 99.03±0.44 98.27±0.78 95.71±3.1 94.38±3.84
3 hidden layers 99.86±0.37 99.76±0.66 99.67±1.83 99.5±1.56 98.23±0.72 96.83±1.28 92.96±5.33 91.45±3.51

Table 3: Mean network output on the Unlawfulness and Im-
putability datasets versus the logical evaluation of the un-
lawfulness resp. imputability conditions.
Trained on all instances Trained on smaller dataset
Unlawfulness Output Unlawfulness Output
False 0 False 0.018
True 1 True 1 the Unlawfulness dataset and 91-94% for the Imputability dataset.
Imputability Output Imputability Output
False 0 False 0.875
True 1 True 1 Table 3 shows that networks still perform perfectly on cases in
5 DISCUSSION
5.1 Standard Accuracy
Standard accuracy is measured to see whether the learned models
are able to solve the classification problem, regardless of whether
or not they discovered the rationale underlying the data. We find
accuracies of 100% or near 100% for networks trained on all in-
stances (see Table 2). When presented with all unique instances,
the networks with one and two hidden layers are able to perfectly
predict the outcome from Dutch tort law, and the network with
three hidden layers can create a very close approximation.
Presenting a neural network with all available cases is in practice
often infeasible. If it is possible, then a simple lookup table rather
than a neural network would most likely suffice. For this reason,
we also trained the networks on a subset of only around 35% of the
unique instances (see Table 2). As expected, the accuracies of the
networks on the general test sets drop, but only slightly (to 98-99%).
Even on the unique test set, accuracies remain around 96%. This
suggests that it is possible for the models to approximate tort law
with a small subset of the unique cases.
5.2 Rationale Discovery
Looking at the performance of the networks on the dedicated test
sets partially exposes how well the rationale is captured by the
network. We designed these test sets such that each one targets a
single condition from the domain. In addition to considering the
accuracy on these dedicated test sets, we qualitatively evaluate the
rational discovery capabilities of the networks by comparing their
outputs with the actual outputs we would ideally expect for the
different domains.
Recall that on the Imputability dataset, networks should output 1
if the act is imputable to the person, and 0 otherwise; on the Unlaw-
fulness dataset, the networks should output 1 if the case is unlawful,
and 0 otherwise. Table 3 shows how well the networks were able
to internalize the notions of unlawfulness and imputability. When
trained on all instances, the mean output of the networks is 0 if the
logical evaluation of unlawfulness is false, and 1 if it is true, which is
exactly what it should do. Networks trained on all instances attain
a perfect score on the Imputability dataset as well. This can also
be seen in Table 2, where the networks score 100% accuracy on
the Unlawfulness and Imputability datasets after training on all
instances.
With less data, however, accuracies drop to around 92-95% for
This accuracy may still seem high, but we should take into account
the label distributions (66.67-33.33% and 87.5-12.5%, respectively).
which the unlawfulness and imputability conditions evaluate to true.
When the conditions are false, however, mistakes are made. The
average output of networks on the Unlawfulness dataset increases
to 0.018, which should be 0, meaning that networks classify some
lawful cases as unlawful. In the Imputability dataset, the mean
output increased more drastically to 0.875 when imputability is
false, meaning that in 87.5% of the instances in which the act is
not imputable to a person, the network incorrectly decided that it
should be. This means that despite high accuracy on the general
test set, the networks largely ignored the concept of imputability.
5.3 A Method for Rationale Evaluation
Although our experiments and discussion focused on specific ex-
ample domains and neural networks, our approach for rationale
evaluation can be interpreted as a general method independent of
the machine learning algorithm applied. Building on the results of
this paper, we therefore proposes a knowledge-driven method for
model-agnostic rationale evaluation, consisting of three distinct
steps:
(1) Measure the accuracy of a trained system, and proceed if the
accuracy is sufficiently high;
(2) Design dedicated test sets for rationale evaluation targeting
selected rationale elements based on expert knowledge of
the domain;
(3) Evaluate the rationale through the performance of the trained
system on these dedicated test sets.
The first step is based on the assumption that efforts for assessing
and possibly improving the rationale discovery capabilities of a
learned model are only taken if the general performance of the
model is already considered good enough. Here we assume per-
formance is measured using accuracy, but other measures can be
employed as well and the threshold of what is considered good
enough may vary per domain and application.
The second step in our method depends on domain knowledge.
Hence the method effectively is a quantitative human-in-the-loop
solution for rationale evaluation.

238
Discovering the Rationale of Decisions
In the third step, performance is again evaluated, by now not
only considering accuracy but also examining model output and
expected output in terms of the dedicated test sets.
The method does not currently specify how the dedicated test
sets are constructed. We aim to further operationalize the rationale
evaluation method by using information about the knowledge in
the domain, and the distribution of examples, for instance building
on Bayesian networks. Subsequently, the information gained by
using this rationale evaluation method can be used to improve the
rationale of the system by adjusting the training data accordingly,
such as in [Bench-Capon 1993] and our replication variants [Steg-
ing et al. 2021], effectively allowing us to impose sound rationale
discovery.
6 CONCLUSION
The work in this paper was inspired by Bench-Capon’s 1993 paper
that investigated whether neural networks are able to tackle open
texture problems. The conclusions were that trained networks can
perform very well in terms of accuracy, even though some condi-
tions from the domain are not learned [Bench-Capon 1993]. Similar
results were found when we repeated the experiments with larger
training datasets, in order to ensure that the original conclusions
about conditions that were not learned are not due to a lack of data.
The idea of constructing test cases to test specific conditions
inspired us to propose a method for assessing rationale discovery
capabilities by designing dedicated test datasets and to evaluate per-
formance on these knowledge-driven test sets, combining quantita-
tive and qualitative evaluation elements in a hybrid way. Adjusting
the training dataset based on this evaluation methods demonstrates
that the rationale can be improved using knowledge-driven tailor
made training sets [Bench-Capon 1993; Steging et al. 2021].
In the real life tort law domain, with a non-fictional knowledge
structure and different characteristics, a similar pattern can be
observed as before: the networks failed to learn the independent
condition that defines imputability, despite high accuracies on the
general test set.
This study therefore reaffirms the conclusions from previous
work, while simultaneously introducing a model-agnostic method
for assessing rationale discovery capabilities of machine-learned
black box models, using dedicated test datasets designed with ex-
pert knowledge of the domain. In future research, we aim to further
detail and extend our method such that by employing it, the sound-
ness of the rationale underlying system decisions becomes tangible,
and its quality can be asserted. Based on this evaluation, the train-
ing data of the black-box systems can be altered to improve their
rationale. Further expanding upon this design method will bring
us closer to AI that is both explainable and responsible.
ACKNOWLEDGMENTS
This research was funded by the Hybrid Intelligence Center, a
10-year programme funded by the Dutch Ministry of Education,
Culture and Science through the Netherlands Organisation for
Scientific Research, https://hybrid-intelligence-centre.nl.
239
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
REFERENCES
K. D. Ashley. 1990. Modeling Legal Arguments: Reasoning with Cases and Hypotheticals.
The MIT Press, Cambridge (Massachusetts).
K. D. Ashley. 2019. A brief history of the changing roles of case prediction in AI and
law. Law in Context 36, 1 (2019), 93–112.
K. Atkinson, T. Bench-Capon, F. Bex, T. F. Gordon, H. Prakken, G. Sartor, and B. Verheij.
2020b. In memoriam Douglas N. Walton: the influence of Doug Walton on AI and
law. Artificial Intelligence and Law (2020), 1–46.
K. Atkinson, T. Bench-Capon, and D. Bollegala. 2020a. Explanation in AI and law: Past,
present and future. Artificial Intelligence 289 (2020), 103387.
T. Bench-Capon. 1993. Neural networks and open texture. In Proceedings of the 4th
International Conference on Artificial Intelligence and Law (ICAIL ’93). ACM, New
York, 292–297.
S. Brüninghaus and K. D. Ashley. 2003. Predicting outcomes of case based legal
arguments. In Proceedings of the 9th International Conference on Artificial Intelligence
and Law (ICAIL 2003). ACM, New York (New York), 233–242.
I. J. Goodfellow, J. Shlens, and C. Szegedy. 2015. Explaining and harnessing adversarial
examples. In Proceedings of International Conference on Learning Representations.
T. F. Gordon. 1995. The Pleadings Game: An Artificial Intelligence Model of Procedural
Justice. Kluwer, Dordrecht.
M. Grabmair, K. D. Ashley, R. Chen, P. Sureshkumar, C. Wang, E. Nyberg, and V. R.
Walker. 2015. Introducing LUIMA: an experiment in legal conceptual retrieval of
vaccine injury decisions using a UIMA type system and tools. In Proceedings of the
15th International Conference on Artificial Intelligence and Law. ACM, New York
(New York), 69–78.
J. C. Hage, R. Leenes, and A. R. Lodder. 1993. Hard cases: a procedural approach.
Artificial intelligence and law 2, 2 (1993), 113–167.
D. Hunter. 1999. Out of their minds: Legal theory in neural networks. Artificial
Intelligence and Law 7, 2 (1999), 129–151.
D. P. Kingma and J. Ba. 2015. Adam: A method for stochastic optimization. In Proceed-
ings of 3rd International Conference on Learning Representations.
S. M. Lundberg and S. Lee. 2017. A unified approach to interpreting model predictions.
In Advances in Neural Information Processing Systems 30. Curran Associates, Inc.,
4765–4774.
M. Medvedeva, M. Vols, and M. Wieling. 2019. Using machine learning to predict
decisions of the European Court of Human Rights. Artificial Intelligence and Law
(2019), 1–30.
R. Mochales Palau and M. F. Moens. 2009. Argumentation mining: the detection, classi-
fication and structure of arguments in text. In Proceedings of the 12th International
Conference on Artificial Intelligence and Law (ICAIL 2009). ACM Press, New York
(New York), 98–107.
M. Možina, J. Žabkar, T. Bench-Capon, and I. Bratko. 2005. Argument based machine
learning applied to law. Artificial Intelligence and Law 13, 1 (2005), 53–73.
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel,
P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau,
M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: machine learning in
Python. Journal of Machine Learning Research 12 (2011), 2825–2830.
L. Philipps and G. Sartor. 1999. Introduction: from legal theories to neural networks
and fuzzy reasoning. Artificial Intelligence and law 7, 2 (1999), 115–128.
M. T. Ribeiro, S. Singh, and C. Guestrin. 2016. "Why should I trust you?": Explaining the
predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA. 1135–
1144.
E. L. Rissland and K. D. Ashley. 1987. A case-based system for trade secrets law. In
Proceedings of the 1st International Conference on Artificial Intelligence and Law
(ICAIL ’87). ACM, New York, NY, USA, 60–66.
C. Steging, S. Renooij, and B. Verheij. 2021. Discovering the Rationale of Decisions:
Experiments on Aligning Learning and Reasoning. arXiv:2105.06758 [cs.AI]
A. Stranieri, J. Zeleznikow, M. Gawler, and B. Lewis. 1999. A hybrid rule–neural
approach for the automation of legal reasoning in the discretionary domain of
family law in Australia. Artificial Intelligence and Law 7, 2-3 (1999), 153–183.
B. Verheij. 2003a. Artificial Argument Assistants for Defeasible Argumentation. Artifi-
cial Intelligence 150, 1–2 (2003), 291–324.
B. Verheij. 2003b. Dialectical argumentation with argumentation schemes: An approach
to legal logic. Artificial intelligence and Law 11, 2-3 (2003), 167–195.
B. Verheij. 2017. Formalizing arguments, rules and cases. In Proceedings of the 16th
International Conference on Artificial Intelligence and Law (ICAIL ’17). ACM, New
York, 199–208.
C. S. Vlek, H. Prakken, S. Renooij, and B. Verheij. 2016. A method for explaining
Bayesian networks for legal evidence with scenarios. Artificial Intelligence and Law
24, 3 (2016), 285–324.
M. Wardeh, T. Bench-Capon, and F. Coenen. 2009. Padua: a protocol for argumentation
dialogue using association rules. Artificial Intelligence and Law 17, 3 (2009), 183–215.
A. Wyner, R. Mochales-Palau, M. F. Moens, and D. Milward. 2010. Approaches to text
mining arguments from legal cases. In Semantic Processing of Legal Texts. Springer,
Berlin, 60–79.
 Process Mining-Enabled Jurimetrics
Analysis of a Brazilian Court’s Judicial Performance in the Business Law Processing
Adriana Jacoto Unger
José Francisco dos Santos Neto
Marcelo Fantinato
Sarajane Marques Peres
{ajacoto,jose.francisco.neto,m.fantinato,sarajane}@usp.br
University of Sao Paulo
Sao Paulo, Brazil
ABSTRACT
Improving judicial performance has become increasingly relevant
to guarantee access to justice for all, worldwide. In this context,
technology-enabled tools to support lawsuit processing emerge as
powerful allies to enhance the justice efficiency. Using electronic
lawsuit management systems within the courts of justice is a wide-
spread practice, which also leverages production of big data within
judicial operation. Some jurimetrics techniques have arisen to eval-
uate efficiency based on statistical analysis and data mining of data
produced by judicial information systems. In this sense, the pro-
cess mining area offers an innovative approach to analyze judicial
data from a process-oriented perspective. This paper presents the
application of process mining in a event log derived from a dataset
containing business lawsuits from the Court of Justice of the State
of Sao Paulo, Brazil – the largest court in the world – in order to
analyze judicial performance. Although the results show these law-
suits have an ad hoc sequence flow, process mining analysis have
allowed to identify most frequent activities and process bottlenecks,
providing insights into the root causes of inefficiencies.
CCS CONCEPTS
• Applied computing → Law; Business process management; Busi-
ness intelligence; • Information systems → Data mining.
KEYWORDS
Process mining, jurimetrics, judicial performance, administration of
justice, legal informatics, business process management, procedural
law, business law.
ACM Reference Format:
Adriana Jacoto Unger, José Francisco dos Santos Neto, Marcelo Fantinato,
Sarajane Marques Peres, Julio Trecenti, and Renata Hirota. 2021. Process
Mining-Enabled Jurimetrics: Analysis of a Brazilian Court’s Judicial Per-
formance in the Business Law Processing. In Eighteenth International Con-
ference for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466137
240
Julio Trecenti
Renata Hirota
{jtrecenti,rhirota}@abj.org.br
Brazilian Jurimetrics Association
Sao Paulo, Brazil
Paulo, Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/
3462757.3466137
1 INTRODUCTION
Within the administration of justice, electronic lawsuits and man-
agement information systems emerged to control lawsuit process-
ing in the courts of justice. These systems allow the practice of
procedural acts by magistrates and other participants in lawsuits in
a fully digital environment. Since 2015, the Court of Justice of the
State of Sao Paulo (TJSP1 ), Brazil, is 100% digital, i.e., all new law-
suits born digital. TJSP is considered the largest court in the world in
terms of volume of lawsuits2 , with over 19 million ongoing lawsuits,
managed by the Justice Automation System (e-SAJ3 ). Data-centric
analysis techniques, such as jurimetrics [11] and machine learning
[14], take advantage of the increasing availability of judicial data.
Jurimetrics supports statistics-based analysis on justice big data. In
this sense, process mining [20] emerges as an approach to bridge
data mining to Business Process Management (BPM), providing a
process-oriented perspective which turns out to be more valuable
when analyzing phenomena distinguished by procedural behaviour,
such as in lawsuit processing ruled by procedural law.
This paper presents an innovative application of process mining
to analyze judicial performance based on a lawsuit dataset, as a
proof of concept for process mining-enabled jurimetrics. The ex-
periment was restricted to business law, as it is an area of law that
commonly presents poor judicial performance and difficulties in
diagnosis due to its high heterogeneity. This study raises and high-
lights contributions of process mining showing the benefits of a
process-oriented approach to analyzing legal data for performance
diagnostics. The paper is divided as follows: background, present-
ing the theoretical framework of this work; related work, pointing
out prior research related to this topic; research method, describing
the stages of this research study; results, showing results of the
process mining application; analysis of the results, presenting pro-
cess mining-enabled analysis of lawsuit processing; and conclusion,
which highlights the research findings.
2 BACKGROUND
This section introduces the main theoretical concepts related to
this work, contextualized in the Brazilian setting.
1 Tribunal
de Justiça de São Paulo (in Portuguese)
2 http://www.tjsp.jus.br/QuemSomos
3 Sistema de Automação da Justiça (in Portuguese)
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
2.1 Judicial Performance
Judicial performance is a field of study dedicated to evaluate per-
formance of judicial systems and courts of justice [8]. Some inter-
national organizations conducted studies to measure and compare
judicial performance worldwide [10, 17]. In Brazil, the National
Council of Justice provides the Justice in Numbers annual report
[5], offering an overview of the judiciary productivity.
Current models for judicial performance evaluation contributed
substantially to the continuous measurement, monitoring and com-
parison of judicial performance. However, such evaluation models
have done little to guide improvement actions. Identifying the root
causes of inefficiencies is often limited due to the asymmetry of
information among the courts’ information systems, as well as the
lack of detailed data on the activities of lawsuit processing [1].
2.2 Jurimetrics
Jurimetrics [11] is defined as ‘statistics applied to the law’. Although
it emerged decades ago, recent advances in computing and data
storage capabilities have enabled alternative ways of observing
patterns in data-based and hence statistics-based court decisions.
In the USA, the application of statistics to law has been developed
under alternative nomenclature, such as Empirical Legal Studies
[7] and, more recently, Judicial Analytics [4]. In Brazil, jurimetrics
has received growing interest [12].
When analyzing alternative ways of managing the lawsuit pro-
cessing, the analysis of the lawsuit throughput time using jurimet-
rics techniques may present quality issues, as it eventually considers
inadequate time intervals for the object of analysis [3]. Procedural
viscosity [16], defined as “a set of structural characteristics of a
lawsuit that is able to affect the speed of its processing”, may apply.
Specifically in business law, processing of lawsuits may require
about twice as much effort as a common lawsuit [2]. The same
authors suggest “future research focused on process flow analysis,
i.e., the study of the stochastic process that generates all events and
timestamps of lawsuit processing”.
2.3 Process Mining
Process mining [20] emerged as a set of techniques for mining
business process-related information from event data logged by
information systems. A business process is a chain of activities that
produces an outcome that adds value to an organization and its
customers [6]. Business process models play a dominant role during
BPM life cycle, leading in achieving organizational improvement
goals, including reducing costs, lead times, and error rates. By using
real event data to discover process models, process mining leverages
data mining to understand operational processes in organizations.
Table 1 shows the mapping of the basic elements of event logs
from the process mining perspective to their counterparts in the
procedural law domain.
3 RELATED WORK
Process mining should be seen as an analytical tool naturally suit-
able for lawsuits due to their inherently procedural nature. It is
cited as a promising approach to suggest improvements for lawsuit
processing [15]. Lawsuit digitization is presented as not being an
241
Unger et al.
Table 1: Process mining in procedural law
Process mining Procedural law
Activity Name of the lawsuit procedural movement.
Event Occurrence of a procedural movement, i.e., oc-
currence of a lawsuit’s procedural progress.
Timestamp Date of occurrence of the lawsuit’s procedural
movement.
Case Lawsuit.
Case ID Lawsuit identifier.
Event attributes General attributes related to the movement.
Case attributes General attributes related to the lawsuit.
end in itself, so that process mining can enhance the way the judi-
ciary treats its digital data through the application of algorithms for
process discovery, compliance and predictive analysis [9]. A tech-
nical report with suggested actions to improve judicial efficiency
highlights the use of process mining as one of these actions [3].
Nevertheless, few studies on this topic were found in the liter-
ature. Empirical studies were performed using lawsuit data from
Brazilian courts[19, 22] though focused on comparison of lawsuit
throughput times. Attempts to apply data mining to extract infor-
mation from lawsuit data were made [13, 18], but these studies do
not directly address process mining on that data.
4 RESEARCH METHOD
This study applied the Process Mining Project Methodology [21] to
guide the application of process mining to analyze judicial perfor-
mance in a specific context. Only the first five stages of the method
were carried out: planning, extraction, data processing, mining and
analysis, and evaluation. The project was finished with the insights
generated by the evaluation stage. Besides restricting the scope to
business law, a period of analysis was defined to consider only the
TJSP’s lawsuits distributed between January 1, 2018 and July 21,
2020. All lawsuits with a procedural movement published in that
time interval were considered. Full progress data for each lawsuit
was retrieved until July 31, 2020, including lawsuits opened before
and lawsuits not yet closed.
The data were extracted in two steps. First, the identifiers of
lawsuits of interest were obtained. For this, all issues of the TJSP’s
Electronic Journals of Justice (DJE4 ) were downloaded from the DJE
website5 , considering the defined analysis period. DJE publishes in-
formation on provisional or final decisions for all ongoing lawsuits
at TJSP, daily. An automated scraping of these files was carried out
using keywords associated with business law litigation. Second,
the lawsuit identifiers obtained were used to retrieve data from
the e-SAJ website6 , where information on lawsuits is published. A
web scraping was carried out to retrieve information on lawsuits
attributes and progress events including their respective dates. In
TJSP, there are four filing court departments dedicated to business
4 Diário da Justiça Eletrônico (in Portuguese)
5 http://www.dje.tjsp.jus.br
6 http://esaj.tjsp.jus.br
Process Mining-Enabled Jurimetrics ICAIL’21, June 21–25, 2021, São Paulo, Brazil

law; as a result, the lawsuits not filed at these four court depart-
ments were discarded. Data from both DJE and e-SAJ websites are Figure 2: Histogram of procedural movements by lawsuit
publicly available7 .
The process knowledge transfer with domain experts was carried
out, resulting in a mapping between the elements of the dataset
and the concepts of event log used by process mining, presented
in Table 2. Event data from lawsuits were used to create the event
log to be used in process mining. The lawsuit dataset was filtered
to remove columns with missing values or data not relevant to the
scope of this study. The judge column was made anonymous for
protecting personal data. The additional column order was added
to the movement database, and hence to the event log, to allow the
process mining discovery algorithm to identify the correct order of
activities within a case occurring on the same date.

5 RESULTS
The resulting event log contains data on lawsuits referring to 4,795 Figure 3: Process map based on average duration metrics
cases and 266,834 events, with procedural movements dating back
to 2008, and 10 case attributes, as described in Table 2. The event log
file8 was imported using the EverFlow9 process mining tool, which
produced the business process maps and the main process metrics,
such as number of cases, number of events, and average duration,
as presented in Figure 1 and Figure 3. Process map views are user-
interactive, so that activities, transitions, and the time interval can
be selected and filtered for drill-down analysis. Detailed views on
specific metrics are presented on dedicated dashboards and panels,
as shown in Figure 2, Figure 4, Figure 5 and Figure 6.

Figure 1: Process map of the lawsuit processing

Figure 4: Slow transition analysis panel

6 ANALYSIS OF THE RESULTS

As a proof of concept for process mining-enabled jurimetrics, the
results were analyzed considering each of the following process
mining perspectives: control-flow, time, resource, and case [20].

6.1 Control-flow Perspective
The control flow is the main perspective considered in process
mining discovery, adding process-oriented value to the data mining
7 The right of access to judicial data in Brazil is guaranteed by the constitutional
principle of judicial publicity, except in lawsuits protected by the secrecy of justice.
8 The event log file is available in the repository: https://doi.org/10.4121/14593857
9 http://everflow.ai
analysis of the event log, as shown in Figure 1. The complexity
and procedural viscosity of lawsuits in business law can be verified
by the process metrics, i.e., average rate of 55.6 events per case
and average case duration of 334 days. As shown in Figure 2, over
10% of lawsuits have over 100 events per case, which corresponds
to procedural movements during lawsuit processing. In addition,
one can verify that lawsuit processing in business law has an ad

242
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Unger et al.

Table 2: Judicial data mapped to process mining event log

Dataset column Data description Event log role

lawsuit_id Lawsuit identifier. Case ID.
movement Name of the lawsuit procedural movement. Event activity.
date Date of occurrence of the lawsuit’s procedural movement. Event timestamp.
order Sequential order of occurrence of the lawsuit’s procedural movement. Event attribute.
area Law area to which the lawsuit refers. Case attribute.
claim_amount Lawsuit claim amount. Case attribute.
class Procedural class (refers to specific law or claim reason). Case attribute.
control Internal control number. Case attribute.
court_department Filing court department. Case attribute.
digital It shows whether the process is digital (born digital or scanned) or on paper. Case attribute.
distribution_date Date and time when the lawsuit was distributed. Case attribute.
judge Name of the ruling judge for the lawsuit. Case attribute.
status Lawsuit status. Case attribute.
subject_matter Main topic of the lawsuit. Case attribute.

6.2 Time Perspective

Figure 5: Lawsuits by judge dashboard The analysis of the lawsuit throughput time using process mining
techniques is favored by the easy handling of incomplete cases,
considering the activities carried out in the sequence flow. Since
the full progress data for each lawsuit was retrieved, there is no
issue related to the beginning of the process (i.e., lawsuit opening).
However, the lawsuit processing time evaluation can be affected by
cases that did not ended (i.e., lawsuits not yet closed). The Definitely
Archived activity is the most frequent final activity, which confirms
the natural behaviour of lawsuit processing within the court. It is
the final activity in 20% of the cases, which represents the number of
closed lawsuits in the dataset. By filtering all cases that go through
this activity, the lawsuit throughput time for closed lawsuits can
be calculated as 312 days.
Furthermore, based on the process map in Figure 3, one can
identify highlighted slow transitions and activity bottlenecks. These
Figure 6: Comparative analysis between on-paper and digi- transitions are also listed in the slow transition analysis panel
tal lawsuits in Figure 4, which offers a detailed diagnostics on each of them.
Comments on some slow transitions of the lawsuit processing-
related business process are presented in the following:

• Transition between Issued Publication Certificate and Issued

hoc sequence flow, characterized by the large number of process
variants. The most frequent control-flow is shared among 155 cases,
which represents only 3% of all lawsuits. The extended process map
in Figure 1 shows the process control-flow considering only 4.19%
of cases. Figure 1 also highlights the most frequent activities and
event counting for each transition.
Judicial Office Certificate activities, with duration of 18 days
and effort of 121 years and 209 days, affecting 38% of cases:
this is a clear bottleneck within the court department. After
the publication certificate is issued, the judicial office publi-
cation certificate must be issued. However, there is a long
waiting time due to the lack of human resources on the court.
This bottleneck has a high impact on lawsuit processing time
and the use of court resource, suggesting that this activity is
a good candidate for automation solutions.
• Transition between Issued Publication Certificate and Suspen-
sion of the Term activities, with duration of 21 days and effort
of 164 years and 258 days, affecting 42% of cases: bottleneck
possibly due to litigant issues, such as lack of required doc-
umentation. Figure 4 shows that this transition causes the
average duration of the case to increase by 77%.

243
Process Mining-Enabled Jurimetrics
• Transition between Suspension of the Term and Conclusions
for Decision activities, with duration of 13 days and effort of
19 years and 298 days, affecting 10% of cases: this bottleneck
is related to the previous one, possibly due to the judge being
unable to evaluate the case for proper decision making.
6.3 Resource Perspective
The resource perspective analysis allows process maps and metrics
to be evaluated considering the resources (people or devices) that
execute the process activities. In addition, the values of the process
execution effort can be evaluated based on the average duration
and event counting for each activity. Performance and comparative
analysis of lawsuit processing from the resource perspective were
carried out based on the judge attribute, as shown in Figure 5.
6.4 Case Perspective
The case perspective allows performance diagnostics to be carried
out considering specific case attributes. Considering the digital case
attribute, a side-by-side comparison was carried out to analyze how
this attribute impacts process flow and metrics. Surprisingly, Figure
6 shows that digital lawsuits are 6% slower than the on-paper ones.
Some considerations on these findings are as follows:
• The definition of the digital attribute includes both the born
digital lawsuits and the scanned ones, although the scanned
ones usually require more time and effort to be processed.
It is not possible to distinguish them by the value of this
attribute, but further investigation on procedural movements
may confirm this assumption, if scan activities are logged.
• Procedural viscosity in business law might hide inefficiencies
usually associated with on-paper lawsuits, which means
that the main process bottlenecks in business law might
be related to internal activities that are independent of the
digital nature of lawsuits. However, this statement requires
further investigation, since digital lawsuit management is
often associated with performance gains.
• A data selection bias may have occurred due to the analysis
period, as it is not uncommon for dedicated task forces to
occur within court departments to reduce the backlog of
on-paper lawsuits.
7 CONCLUSION
The analysis of the results revealed a comprehensive set of diag-
nostic metrics, insights into the root causes of inefficiencies, and
ideas for improvement, which would hardly be discovered without
a process-oriented analysis approach. The prospects for using pro-
cess mining in the Brazilian judiciary include the use of process
mining tools to provide online dashboards for judicial performance
monitoring by the State Internal Affairs Divisions of Justice, which
monitors the performance on the provision of jurisdictional ser-
vices. These dashboards can be used to identify inefficiencies in
near real time and define targets for resolving them.
ACKNOWLEDGMENTS
The authors thank the EverFlow company for kindly supporting
this research.
244
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
REFERENCES
[1] Bruna Armonas Colombo, Pedro Buck, and Vinicius Miana Bezerra. 2017. Chal-
lenges When Using Jurimetrics in Brazil—A Survey of Courts. Future Internet 9,
4 (10 2017), 68. https://doi.org/10.3390/fi9040068
[2] Associação Brasileira de Jurimetria. 2016. Estudo sobre varas empresariais na
Comarca de São Paulo. Technical Report. Associação Brasileira de Jurimetria. 20
pages. https://abj.org.br/pdf/ABJ_varas_empresariais_tjsp.pdf (in Portuguese).
[3] Associação Brasileira de Jurimetria. 2020. Justiça Pesquisa - Formas Al-
ternativas de Gestão Processual: a especialização de varas e a unificação de
serventias. Technical Report. Conselho Nacional de Justiça, Brasília. 126
pages. https://www.cnj.jus.br/wp-content/uploads/2020/08/Justica-Pesquisa_
Relatorio_ABJ_2020-08-21_1.pdf (in Portuguese).
[4] Daniel L. Chen. 2019. Judicial analytics and the great transformation of American
Law. Artificial Intelligence and Law 27, 1 (3 2019), 15–42. https://doi.org/10.1007/
s10506-018-9237-x
[5] Conselho Nacional de Justiça. 2020. Justiça em Números 2020: ano-
base 2019. Technical Report. Conselho Nacional de Justiça, Brasília. 267
pages. https://www.cnj.jus.br/wp-content/uploads/2020/08/WEB-V3-Justiça-
em-Números-2020-atualizado-em-25-08-2020.pdf (in Portuguese).
[6] Marlon Dumas, Marcello La Rosa, Jan Mendling, and Hajo A. Reijers. 2018. Fun-
damentals of Business Process Management (2 ed.). Springer, Berlin. 527 pages.
https://doi.org/10.1007/978-3-662-56509-4
[7] Theodore Eisenberg. 2011. The origins, nature, and promise of empirical legal
studies and a response to concerns. University of Illinois Law Review 2011, 5
(2011), 1713–1738. https://doi.org/10.2139/ssrn.1727538
[8] Adalmir de Oliveira Gomes and Tomás de Aquino Guimarães. 2013. Desempenho
no judiciário. Conceituação, estado da arte e agenda de pesquisa. Revista de
Administracao Publica 47, 2 (3 2013), 379–401. https://doi.org/10.1590/S0034-
76122013000200005 (in Portuguese).
[9] Bráulio Gusmão. 2021. Mineração de processos e gestão de casos no judiciário.
In Inteligência Artificial e Direito Processual: os impactos da virada tecnológica no
direito processual (2 ed.), Dierle Nunes, Paulo Henrique dos Santos Lucon, and
Erik Navarro Workart (Eds.). JusPodivm, Salvador, 589–594. (in Portuguese).
[10] International Consortium for Court and Excellence. 2020. Global Measures of
Court Performance. Technical Report. Secretariat for the International Consortium
for Court Excellence, Sydney, Australia. 112 pages. http://www.courtexcellence.
com/resources/global-measures
[11] Lee Loevinger. 1948. Jurimetrics–The Next Step Forward. Minnesota Law Review
33, 5 (1948), 455–493.
[12] Marcos Maia and Cicero Aparecido Bezerra. 2020. Bibliometric analysis of scien-
tific articles on jurimetry published in Brazil. Revista Digital de Biblioteconomia e
Ciência da Informação 18 (2020). https://doi.org/10.20396/RDBCI.V18I0.8658889
[13] Oleg Metsker, Egor Trofimov, Sergey Sikorsky, and Sergey Kovalchuk. 2019. Text
and data mining techniques in judgment open data analysis for administrative
practice control. In Communications in Computer and Information Science, Vol. 947.
Springer, Cham, 169–180. https://doi.org/10.1007/978-3-030-13283-5_13
[14] Tom M. Mitchell. 1997. Machine Learning. McGraw-Hill Sci-
ence/Engineering/Math. 432 pages.
[15] Dierle Nunes. 2021. A technological shift in procedural law (from automation
to transformation): Can legal procedure be adapted through technology? In
Inteligência Artificial e Direito Processual: os impactos da virada tecnológica no
direito processual (2 ed.), Dierle Nunes, Paulo Henrique dos Santos Lucon, and
Erik Navarro Workart (Eds.). JusPodivm, Salvador, 55–78.
[16] Marcelo Guedes Nunes. 2019. Jurimetria: como a estatística pode reinventar o
Direito (2 ed.). Revista dos Tribunais, São Paulo. 192 pages. (in Portuguese).
[17] Giuliana Palumbo, Giulia Giupponi, Luca Nunziata, and Juan Mora-Sanguinetti.
2013. Judicial performance and its determinants: a cross-country perspec-
tive. OECD Economic Policy Papers 5 (2013), 1–38. https://doi.org/10.1787/
5k44x00md5g8-en
[18] Shahmin Sharafat, Zara Nasar, and Syed Waqar Jaffry. 2019. Data mining for
smart legal systems. Computers and Electrical Engineering 78 (9 2019), 328–342.
https://doi.org/10.1016/j.compeleceng.2019.07.017
[19] Universidade de São Paulo. 2019. Justiça Pesquisa - Mediações e Conciliações
Avaliadas Empiricamente: jurimetria para proposição de ações eficientes. Technical
Report. Conselho Nacional de Justiça, Brasília. 193 pages. https://www.cnj.
jus.br/wp-content/uploads/2011/02/e1d2138e482686bc5b66d18f0b0f4b16.pdf (in
Portuguese).
[20] Wil M. P. van der Aalst. 2016. Process mining: Data science in Action (2 ed.).
Springer, Berlin. 467 pages. https://doi.org/10.1007/978-3-662-49851-4
[21] Maikel L. van Eck, Xixi Lu, Sander J. J. Leemans, and Wil M. P. van der Aalst.
2015. PM2: A process mining project methodology. In Lecture Notes in Computer
Science, Vol. 9097. Springer, Cham, 297–313. https://doi.org/10.1007/978-3-319-
19069-3_19
[22] Caio Castelliano de Vasconcelos, Eduardo Watanabe de Oliveira, Henrique Pais da
Costa, and Tomas de Aquino Guimaraes. 2018. Tempo de Processos Judiciais
na Justiça Federal do Brasil. In XLII Encontro da ANPAD. Curitiba, 1–16. http:
//www.anpad.org.br/abrir_pdf.php?e=MjQ1NzA= (in Portuguese).
 Using Transformers to Improve Answer Retrieval
for Legal Questions
Andrew Vold
Thomson Reuters
TR Labs Research
andrew.vold@thomsonreuters.com
ABSTRACT
Transformer architectures such as BERT, XLNet, and others are
frequently used in the field of natural language processing. Trans-
formers have achieved state-of-the-art performance in tasks such as
text classification, passage summarization, machine translation, and
question answering. Efficient hosting of transformer models, how-
ever, is a difficult task because of their large size and high latency.
In this work, we describe how we deploy a RoBERTa Base ques-
tion answer classification model in a production environment. We
also compare the answer retrieval performance of a RoBERTa Base
classifier against a traditional machine learning model in the legal
domain by measuring the performance difference between a trained
linear SVM on the publicly available PRIVACYQA dataset. We show
that RoBERTa achieves a 31% improvement in F1-score and a 41%
improvement in Mean Reciprocal Rank over the traditional SVM.
CCS CONCEPTS
• Information systems → Information Retrieval; Retrieval
Tasks and Goals; Question Answering; Information Retrieval;
Retrieval Models and Ranking; Language Models; Information
Retrieval; Evaluation of retrieval results; Relevance assessment.
KEYWORDS
Question Answering, Legal Applications, Deep Learning, Language
Models, BERT Engines, Evaluation
ACM Reference Format:
Andrew Vold and Jack G. Conrad. 2021. Using Transformers to Improve
Answer Retrieval for Legal Questions. In Eighteenth International Conference
for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo,
Brazil. ACM, New York, NY, USA, 5 pages. https://doi.org/10.1145/3462757.
3466102
1 INTRODUCTION
Historically, when legal professionals performed natural language
search, they would be required to sift through exhaustive lists of
results, ranked by probability of relevance, in order to identify
materials relevant to their search [18]. The task could be a time
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466102
245
Jack G. Conrad
Thomson Reuters
TR Labs Research
jack.g.conrad@thomsonreuters.com
consuming and laborious effort. Over time, we began to see an in-
terest in more focused question answering systems taking the place
of traditional information retrieval systems. In the field of AI and
Law, Quaresma and Rodrigues were among the first to implement
a question answering system for legal documents [13], one that
focused on Portuguese legal decisions. More recently, however, de-
velopments in deep learning-based approaches for tasks like open
domain question answering have resulted in major gains in answer
rate performance. They have also been responsible for comparable
advances in closed domain question answering in fields such as
Legal QA [1]. Such progress has resulted in performance gains for
both factoid and non-factoid question answering.
Transformer architectures have delivered impressive perfor-
mance gains over baselines for standard natural language process-
ing (NLP) tasks. Open domain language modeling as a pretraining
step, followed by domain specific fine-tuning on another domain
has delivered state-of-the-art performance for tasks in a specific
domain, including the legal domain. One should thus expect to see
significant performance gains in legal question answer retrieval
by utilizing the output of a transformer based classifier which has
been fine-tuned on legal QA pairs.
It has been well observed that transformers are highly perfor-
mant at answering factoid questions which typically have answers
of one or a few words [5]. Transformer based research in the Legal
domain has evolved toward more complex non-factoid questions
which are more nuanced and may require several sentences to pro-
vide context and elaboration in order to answer the legal question at
hand, for example, "When is a party entitled to a protective order?"
The current work extends this research by processing a publicly
available non-factoid QA dataset in an application workstream,
while addressing the challenges of performance quality, speed and
scale.
2 PRIOR WORK
The primary approaches employed to improve question answering
search results fall into three categories: document-centric, query-
centric, and ranking-centric (e.g., neural approaches). The works
described below generally fall into one or more of these categories.
2.1 Open-Domain Question Answering
Open domain question answering is a task that answers factoid
questions using large collections of documents [19]. Historically,
retrieval in open domain QA was usually conducted using tf.idf
or BM25 approaches, which match keywords with an inverted in-
dex, and represent the question and content in high-dimensional,
sparse vectors [16]. In their 2017 report, Chen et al. propose us-
ing Wikipedia for open domain question answering for factoid
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
questions [5]. The task is one of machine reading at scale, which
addresses the challenges of document retrieval and machine com-
prehension (identifying text spans containing the answer). Their
approach combines a search component based on bigram hashing
and tf.idf matching with a multi-layer recurrent neural network
model trained to detect answers in Wikipedia paragraphs. They use
the SQuAD dataset for training and three other datasets for testing
[14]. They obtain an F-score of 79%, which was within a point of
the top performing method at the time.
In their work on dense passage retrieval for open domain ques-
tion answering, Karpukhin et al. show that retrieval can be ef-
fectively implemented using dense representations alone, where
embeddings are learned from a small number of questions and
passages via a simple dual encoder framework [9]. It has outper-
formed traditional QA baselines (top-20 results) by 9%-19%, while
establishing new end-to-end baseline performance levels.
In their earlier work on Bidirectional Encoder Representations
from Transformers (BERT), Devlin et al. introduced a new language
representation model which is designed to pre-train deep bidirec-
tional representations from unlabeled text by jointly conditioning
on both left and right context in all layers [7]. BERT consequently
can be fine-tuned with just one additional output layer to create
state-of-the-art, highly performant models for a wide range of tasks,
including question answering.
As an extension to BERT, Liu et al. developed a "robustly op-
timized" pretraining approach to BERT known as RoBERTa [10].
They found that BERT was significantly undertrained. In their repli-
cation study of BERT, they carefully measured the impact of many
key hyperparameters and training data size. They showed how
hyperparameter choices have a major impact on final results. Their
best model achieved state of the art results against such standard
collections as GLUE, RACE, and SQuAD.
Because pre-trained language models are usually computation-
ally expensive, and it is difficult to execute them on resource limited
devices, researchers like Jiao et al. have focused on transformer
model distillation methods and proposed a novel method that was
specially designed for knowledge distillation (KD). By leveraging
their new KD method, while focusing on the knowledge already
preserved in larger models like RoBERTa, they discovered that such
knowledge could be transferred to a smaller TinyBert model [8].
The new framework captured in TinyBert performs transformer
distillation at both the pre-training and task specific learning stages.
They have shown that their framework ensures that TinyBert cap-
tures the general knowledge and task specific knowledge preserved
in BERT.
In contrast with factoid question answering, Zhu et al. pursued
non-factoid question answering where the answers tend to be
longer passages [22]. In this work, the authors determine that by
generating synthetic training data of arbitrary volume and with
well understood properties, the learning capacity of Knowledge
Graph architectures can be better understood and characterized.
Whether a given neural architecture for KGQA will train a model to
generalize rather than memorize may depend on dataset properties.
2.2 Legal Domain Question Answering
In a recent work, the authors address a due diligence topic where
lawyers review documents for indication of risk due to the prospect
246
Andrew Vold and Jack G. Conrad
of a merger or acquisition [6]. They claim that what is novel in
their approach is that the proposed system explicitly handles the
imbalance in the data, by generating synthetic instances of the
minority answer categories, using the Synthetic Minority Oversam-
pling Technique [4]. This ensures that the number of instances in
all the classes are roughly equal to each other, thus leading to more
accurate and reliable classification. They use conditional random
fields as their text selection algorithm. Each sentence in the contract
under consideration is featurized into a tf.idf vector and fed into the
CRF algorithm. The authors found a 13% improvement in accuracy
due to the imbalance handling.
The recently published work on Legal BERT has reported on
performance gains on an assortment of downstream NLP tasks
[3]. The authors compare the performance of out of the box BERT
with a version that benefits from additional pre-training with legal
domain data, and finally with a version where the pre-training with
legal domain data starts from scratch. The legal domain training
data consists of UK and EU legislation, European Court of Justice
and Court of Human Rights cases, and finally U.S. court cases as
well as U.S. contracts. The authors show that the best strategy
to transfer BERT to a new domain may vary, but that one may
consider either further pre-training or pre-training from scratch
on data from the new domain. Legal BERT achieved state-of-art
results in three end-tasks, and, most notably, the performance gains
were stronger for the most challenging end-tasks (i.e., multi-label
classification in ECHR-cases and contract header & lease details
in Contracts-NER) where in-domain (legal) knowledge is arguably
the most important. The authors also released a version of Legal
BERT-SMALL, which is 3 times smaller than Legal BERT, but quite
competitive performance-wise to the other versions of Legal BERT.
Reports on question answering systems have also recently been
published by researchers at Thomson Reuters and LexisNexis [2,
11]. The current work demonstrates the robustness of a Legal QA
system deployed in a multi-stage workstream where the engine is
fine-tuned on an application-specific dataset. The application and
dataset are discussed below. The system is shown to significantly
outperform the baseline using contemporary neural techniques.
3 METHODOLOGY
Transformer models have achieved state-of-the-art performance
in many NLP applications such as text classification, text summa-
rization, and question answering. Though transformers are highly
performant, their generally large size make them difficult to deploy
in production systems. Successful transformer model hosting in
a production environment would be a major advance in natural
language applications. For this reason, we developed a high perfor-
mance question answering (QA) system based on the RoBERTa base
architecture, but other transformer architectures could be used as
well [10, 12]. The challenges and our strategies for handling these
problems will be discussed in this section.
QA system researchers do not frequently have access to evalu-
ated QA pairs that are broad, balanced, and comparable to what
a user would ask. Open sourced QA pairs tend to be either very
general or belong to a niche domain. If one is fortunate to have
access to labeled QA pairs in the working domain, it is unlikely that
there is enough data for broad topic coverage. To address this issue,
subject matter experts (SMEs) can be assigned to procure quality
Using Transformers to Improve Answer Retrieval for Legal Questions
Figure 1: QA System Development Cycle
QA pairs. Yet SMEs often experience fatigue when producing nu-
merous examples, even if the queries originate from user query
logs. This phenomenon often manifests itself in the form of weak
question-answer pair generation where examples differ by only
a few words. To address such limitations, natural language user
queries are identified, run through the classifier, and the highest
scoring QA pairs are evaluated. The resulting data can then be used
to train the model, yielding a cyclic data curation, model training
process as seen in Figure 1.
Given the QA system that we developed was intended for ap-
plication to sets of in-house legal documents, many of which are
not freely available to the general public, for the purposes of this
research report, we have opted to apply our techniques to the pub-
licly available legal questioning collection described in section 3.2.
Though it covers a subdomain of the legal space, it is nonetheless a
broad ranging and complex dataset that contains an array of top-
ics, question and answer lengths and types. It is a nuanced and
challenging set of data which is indicative of the kinds of question
and answer types one can expect to see in the legal domain. The
findings we obtain apply specifically to the PRIVACYQA dataset,
but are also representative of the kinds of issues and challenges
one encounters with wider-ranging legal datasets as well.
3.1 Training Targets
In order to assess the performance of the QA classifier, natural
language user log queries and their retrieved answers are presented
to an SME. The SME then must determine whether or not the top
answers returned by the classifier satisfy what was being asked.
The grade by the SME can be a binary "pass/fail", a letter grade,
or even a score on a continuous scale. In our case, the grade is
converted into a label or regression target to be used for model
fine-tuning.
For our internal QA classifier, we utilized a multi-label grading
criteria which determined whether or not the answer satisfies the
requirements and to what degree it answers the given question. In
order to avoid grader bias, we have two SMEs grade each QA pair,
and the average is taken. Disagreements of more than one grade
may be adjudicated by a senior SME. A similar approach was used
by the creators of the PRIVACYQA dataset, which will be explained
below.
3.2 Data
The dataset used in these experiments comes from the PRIVACYQA
dataset described by Ravichander et al. in [15]. It is a corpus con-
sisting of 1,750 questions about privacy policies associated with
247
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
mobile applications [20], and more than 3,500 relevant answers that
have been annotated by experts. From the data provided, we have
obtained approximately 130K passages for our training set, of which
about 25% was used in our validation set. The goal of the collection
was to achieve broad coverage across a spectrum of application
types. The researchers collected privacy policies from 35 mobile
applications representing different categories in the Google Play
Store [17]. Another goal of the creators was to include both policies
from well-known applications, which are likely to have carefully-
constructed privacy policies, and lesser-known applications with
smaller install bases, whose policies might be considerably less
sophisticated. They set a threshold of 5 million installs to ensure
each category includes applications with installs on both sides of
the threshold. All policies in the corpus are in English, and were
collected before April 1, 2018, predating many companies’ GDPR-
focused revisions.
3.2.1 Answer Identification. In order to identify legally valid an-
swers, seven subject matter experts with legal training were re-
cruited to formulate answers to the Amazon Mechanical Turk ques-
tions. They indicated relevant material within the given privacy
policy in addition to supplying relevant metadata regarding the
question’s relevance, subjectivity, OPP-115 category [21], and how
likely any policy is to containing the answer to the question.
Table 1 presents aggregate statistics for the PRIVACYQA dataset.
1750 questions are posed to an imaginary privacy assistant over 35
mobile applications and their associated privacy documents.
Dataset Train Test All
No. of Questions 1350 400 1750
No of Policies 27 8 35
No. of Sentences 3704 1243 4947
Avg. Q Length 8.42 8.56 8.46
Avg. Doc. Length 3121.3 3629.13 3237.37
Avg. Ans. Length 123.73 153.44 139.62
Table 1: Statistics of the PRIVACYQA Dataset
4 EXPERIMENTS
To demonstrate the quality of answer retrieval performance of a
transformer in comparison with traditional ML models, we fine-
tune an open domain pretrained RoBERTa classifier and train a
linear SVM with tf.idf features on the PRIVACYQA dataset. Training
models on this dataset is challenging for several reasons. First, the
dataset is largely unbalanced with negative examples occurring 25
times more often than positive examples. In addition, there exists
considerable noise in both the queries and the answers. Finally, the
number of unique questions and answers are far fewer than the
total counts of QA pairs in the dataset.
Class imbalance is a common problem in real world machine
learning applications. For this reason, there are many methods to
effectively combat the adverse effects of training on an imbalanced
dataset. These can include over/under sampling, class weighting
on the loss, external or generated training data augmentation, and
more. For our experiments, we apply a simple class weighting
scheme to give more weight to the underrepresented positive class.
The PRIVACYQA data is quite noisy. The queries and answers are
riddled with misspellings, URLs, improper grammar, fragmented
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
sentences, lack of punctuation, and more. In order to have the
data resemble the data existing in our internal system, significant
data cleaning and filtering is applied. This includes capitalizing
sentence beginnings, removing URLs, removing queries or answers
with more than 4 non-english words, and additional cleaning and
filtering steps. Even after all of this data preprocessing, the data
remains far from perfect, but is sufficient to meet the requirements
of our experimental conditions.
The original PRIVACYQA paper split the training and testing
datasets by privacy category, rather than by unique queries (Table
1). The original PRIVACYQA dataset thus contains data leakage.
Several queries from the test set can also be found in the training
set. In order to rectify this, we identify the queries which exist in
the test set, and reassign those QA pairs as training data, which can
be seen in Table 2.
Set Positives Negatives Total
Train 6,950 152,903 159,487
Test 5,276 45,493 50,720
Table 2: Dataset Split Statistics
We perform tf.idf fitting on the unigrams and bigrams from
the corpus of unique answers, and use it to vectorize the QA pairs
which are then used as inputs to a linear SVM. The hyperparameters
of the SVM are found by performing 5-fold cross validation via
grid searching with maximizing the validation set F1-score as the
objective. This process leads to optimal hyperparameters for the
SVM model and a consistent training-validation split to be used for
training RoBERTa.
Due to the large number of parameters in RoBERTa, it is trained
by gradually unfreezing the layers, starting with the classification
head. The learning rate and the batch size are decreased as layers
are unfrozen, as to avoid overloading the CUDA memory. After
each epoch, the validation F1-score is measured until a plateau is
reached, at which point the model loses generalizability.
5 RESULTS
After training both RoBERTa and SVM classifiers, the models are
run over the test set to determine the performance differences when
using a transformer based QA classification engine. The results can
be seen in Table 3.
Metric SVM RoBERTa
Precision 0.212 0.470
Recall 0.480 0.326
F1-score 0.294 0.385*
MRR 0.074 0.105**
Table 3: Classifier Performance on the Test Set
As seen in the table, RoBERTa outperforms the SVM for all
metrics except recall. This makes sense because the SVM looks for
exact token matches between the query and answer to assign a
positive label. RoBERTa, however, uses the latent representation of
the tokens to identify potential answers. In any QA application, it
is important to serve an expansive set of quality answers; for this
reason, RoBERTa is preferable to the SVM for its 31% improvement
in F1-score over the SVM (* 𝑝 < 1 × 10−5 ).
248
Andrew Vold and Jack G. Conrad
One of the most important metrics for QA classification systems
is Mean Reciprocal Rank (MRR). This simple metric is the average
inverse position of the first true labeled examples in the answer
ranks. MRR is a useful metric for ensuring that the highest quality
answers make it to the highest rank in the list. This is especially
important for applications like question answering which may
return a few or even one answer for a particular query. Due to
the importance of MRR, RoBERTa is the better choice for a QA
model with a 41.4% improvement in MRR over the SVM baseline
(** 𝑝 < 1 × 10−5 ).
It is interesting to see that a simple, traditional ML model like
an SVM, operating on sparse word vectors, achieves performance
relatively similar to that of a transformer. One explanation for this
is that the data is very messy and lacks uniqueness. A simple ML
model doesn’t get distracted by nuances of this dataset such as
fragmented sentences, misspellings, and the frequent use of URLs
and company names. A simple ML model is also less prone to over-
fitting than a transformer, especially considering the redundancy
of the text in the dataset. Overfitting was a challenge during exper-
imentation. For this reason, one can expect even higher RoBERTa
performance if the experiments are repeated with a more sophisti-
cated strategy for combatting overfitting. A major lesson learned
from running this experiment is to ensure that the data used for
training a transformer QA classifier is clean and without redun-
dancy. In addition, more careful domain adaptation could be applied
before fine-tuning on the experimental dataset.
6 APPLICATION PIPELINE
Developing a strong QA classifier is only one piece of deploying
a scalable QA application. It is not feasible to simply concatenate
all passages from a corpus to a user’s query and sequentially feed
them to a classifier. Instead, there needs to be a way to quickly filter
out obvious negative passages, yielding a smaller pool of potential
answers to be fed to the classifier. An additional challenge of using a
transformer based classifier like RoBERTa is its size and latency. In
order to address these challenges, we propose a solution consisting
of a parallel cluster for candidate retrieval (Stage 1) and RoBERTa
operating on a GPU endpoint (Stage 2). In addition, in order not to
overwhelm the user of the application, we typically return the top
n answers as predicted by RoBERTa, where n is small.
One of the most important requirements for a powerful QA clas-
sification engine is to have a sufficiently large corpus of passages
against which a query can be compared. Oftentimes, this can be
on the scale of hundreds of thousands to millions of passages. The
overwhelmingly vast number of passages is irrelevant to a particu-
lar query, and these are not difficult to identify. For this reason, it is
advisable to have a computationally efficient method of removing
the clearly irrelevant passages before performing any QA infer-
encing. In addition, due to the scale of the data, it is imperative to
perform this filtering in parallel. To accomplish this, we employ a
parallel data cluster in the cloud with our data spanning several
nodes (See Figure 2). The cluster functions by serving up the top-n
most relevant passages as determined by properties such as term
overlap between the query and passages. It is up to the application
designer to determine the appropriate number of passages to in-
clude in a candidate pool. Typically a candidate pool size between
100 and 1000 will suffice. Increasing the number of nodes decreases
Using Transformers to Improve Answer Retrieval for Legal Questions
Figure 2: QA Application Pipeline
latency but increases cost, so application engineers must decide in
advance on how many nodes to include in their cluster.
After a satisfactory candidate pool has been retrieved, the QA
pairs are then tokenized, pushed to the CUDA device, and fed to
the classifier. The classifier returns a list of prediction scores of the
relevance of the passage to the answer. The passages associated
with these predictions are then sorted, and the top-n are returned,
where n is determined by the application development team. One
may also wish to apply a RoBERTa score threshold, so that very low
predictions, which are very often negative, are not shown to the
user. If executed properly on the appropriate hardware, the entire
answer serving process can take a second or less to perform.
7 CONCLUSIONS
Question answering is a challenging task which has been in devel-
opment for many years. Question answering can take on different
forms such as answer generation, answer snippet retrieval, and
question answer classification. We propose an end-to-end pipeline
which combines the speed of a parallel data retrieval mechanism
with the classification power of a fine-tuned RoBERTa Base classi-
fier. Our observations from our internal data and the data discussed
in this paper indicate that transformer architectures can achieve
greater classification performance than traditional machine learn-
ing methods in legal QA classification tasks.
We have discussed the efficacy of transformer models in text clas-
sification tasks. We observe a significant increase in F1-score and
MRR of a RoBERTa classifier over a linear SVM on the PRIVACYQA
dataset. Our experiment has shown that transformer models can
achieve superior performance over traditional machine learning
techniques in legal question answer classification.
We have also also discussed some of the challenges and solu-
tions associated with developing and operating a transformer based
question answer classification system. With a large set of content,
subject matter experts, and sufficient computing power, it is pos-
sible to train and operate a transformer based system in a cost
effective manner.
REFERENCES
[1] S. Badugu and R. Manivannan. A study on different closed domain question
answering approaches. Int. J. Speech Technol., 23(2):315–325, 2020.
[2] Z. Bennett, T. Russell-Rose, and K. Farmer. A scalable approach to legal question
answering. In Proceedings of the 16th Edition of the International Conference on
Artificial Intelligence and Law, ICAIL ’17, pages 269–270, New York, NY, USA,
2017. Association for Computing Machinery.
249
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
[3] I. Chalkidis, M. Fergadiotis, P. Malakasiotis, N. Aletras, and I. Androutsopoulos.
Legal-bert: The muppets straight out of law school, 2020.
[4] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. Smote: Synthetic
minority over-sampling technique. J. Artif. Int. Res., 16(1):321–357, June 2002.
[5] D. Chen, A. Fisch, J. Weston, and A. Bordes. Reading wikipedia to answer open-
domain questions, 2017.
[6] R. Chitta and A. K. Hudek. A reliable and accurate multiple choice question
answering system for due diligence. In Proceedings of the Seventeenth International
Conference on Artificial Intelligence and Law, ICAIL ’19, pages 184–188, New York,
NY, USA, 2019. Association for Computing Machinery.
[7] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep
bidirectional transformers for language understanding, 2019.
[8] X. Jiao, Y. Yin, L. Shang, X. Jiang, X. Chen, L. Li, F. Wang, and Q. Liu. Tinybert:
Distilling BERT for natural language understanding. CoRR, abs/1909.10351, 2019.
[9] V. Karpukhin, B. Oğuz, S. Min, P. Lewis, L. Wu, S. Edunov, D. Chen, and W. t. Yih.
Dense passage retrieval for open-domain question answering, 2020.
[10] Y. Liu, M. O., N. Goyal, J. Du, M. Joshi, D. Chen, O. L., M. Lewis, L. Zettlemoyer,
and V. Stoyanov. Roberta: A robustly optimized bert pretraining approach, 2019.
[11] G. McElvain, G. Sanchez, S. Matthews, D. Teo, F. Pompili, and T. Custis. West-
search plus: A non-factoid question-answering system for the legal domain. In
Proceedings of the 42nd International ACM SIGIR Conference on Research and
Development in Information Retrieval, SIGIR 2019, Paris, France, July 21-25, 2019,
pages 1361–1364. ACM, 2019.
[12] M. Ott, S. Edunov, A. Baevski, A. Fan, S. Gross, N. Ng, D. Grangier, and M. Auli.
fairseq: A fast, extensible toolkit for sequence modeling, 2019.
[13] P. Quaresma and I. Rodrigues. A question-answering system for portuguese
juridical documents. In Proceedings of the 10th International Conference on Artifi-
cial Intelligence and Law, ICAIL ’05, pages 256–257, New York, NY, USA, 2005.
Association for Computing Machinery.
[14] P. Rajpurkar, J. Zhang, K. Lopyrev, and P. Liang. Squad: 100,000+ questions for
machine comprehension of text, 2016.
[15] A. Ravichander, A. W. Black, S. Wilson, T. B. Norton, and N. M. Sadeh. Question
answering for privacy policies: Combining computational and legal perspectives.
CoRR, abs/1911.00841, 2019.
[16] S. Robertson and H. Zaragoza. The probabilistic relevance framework: Bm25 and
beyond. Found. Trends Inf. Retr., 3(4):333–389, apr 2009.
[17] P. Story, S. Zimmeck, and N. Sadeh. Which apps have privacy policies? In M. Med-
ina, A. Mitrakas, K. Rannenberg, E. Schweighofer, and N. Tsouroulas, editors,
Privacy Technologies and Policy, pages 3–23, Cham, 2018. Springer International
Publishing.
[18] H. R. Turtle. Text retrieval in the legal world. Artif. Intell. Law, 3(1-2):5–54, 1995.
[19] E. M. Voorhees. The trec-8 question answering track report. In Proceedings of
TREC-8, pages 77–82, 1999.
[20] D. Weissenborn, G. Wiese, and L. Seiffe. Making neural QA as simple as possible
but not simpler. In Proceedings of the 21st Conference on Computational Natural
Language Learning (CoNLL 2017), pages 271–280, Vancouver, Canada, Aug. 2017.
Association for Computational Linguistics.
[21] S. Wilson, F. Schaub, A. A. Dara, F. Liu, S. Cherivirala, P. Giovanni Leon,
M. Schaarup Andersen, S. Zimmeck, K. M. Sathyendra, N. C. Russell, T. B. Norton,
E. Hovy, J. Reidenberg, and N. Sadeh. The creation and analysis of a website
privacy policy corpus. In Proceedings of the 54th Annual Meeting of the Association
for Computational Linguistics (Volume 1: Long Papers), pages 1330–1340, Berlin,
Germany, Aug. 2016. Association for Computational Linguistics.
[22] M. Zhu, A. Ahuja, D. Juan, W. Wei, and C. K. Reddy. Question answering with
long multiple-span answers. In T. Cohn, Y. He, and Y. Liu, editors, Proceedings of
the 2020 Conference on Empirical Methods in Natural Language Processing: Findings,
EMNLP 2020, Online Event, 16-20 November 2020, pages 3840–3849. Association
for Computational Linguistics, 2020.
 Toward Summarizing Case Decisions via Extracting Argument
Issues, Reasons, and Conclusions
Huihui Xu Jaromir Savelka
Intelligent Systems Program School of Computer Science
University of Pittsburgh Carnegie Mellon University
USA USA
huihui.xu@pitt.edu jsavelka@andrew.cmu.edu
ABSTRACT
In this paper, we assess the use of several deep learning classifica-
tion algorithms as a step toward automatically preparing succinct
summaries of legal decisions. Short case summaries that tease out
the decision’s argument structure by making explicit its issues, con-
clusions, and reasons (i.e., argument triples) could make it easier
for the lay public and legal professionals to gain an insight into
what the case is about. We have obtained a sizeable dataset of
expert-crafted case summaries paired with full texts of the deci-
sions issued by various Canadian courts. As the manual annotation
of the full texts is prohibitively expensive, we explore various ways
of leveraging the existing longer summaries which are much less
time-consuming to annotate. We compare the performance of the
systems trained on the annotations that are manually ported to
the full texts from the summaries to the performance of the same
systems trained on annotations that are projected from the sum-
maries automatically. The results show the possibility of pursuing
the automatic annotation in the future.
CCS CONCEPTS
• Information systems → Information retrieval; Retrieval mod-
els and ranking; Similarity measures; • Applied computing →
Law; Annotation.
KEYWORDS
Information retrieval, argument mining, legal analysis, relevant
sentences, summarization
ACM Reference Format:
Huihui Xu, Jaromir Savelka, and Kevin D. Ashley. 2021. Toward Summariz-
ing Case Decisions via Extracting Argument Issues, Reasons, and Conclu-
sions. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY,
USA, 5 pages. https://doi.org/10.1145/3462757.3466098
1 INTRODUCTION
The ability to automatically prepare succinct summaries of legal
decisions could contribute to making legal source materials more
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466098
250
Kevin D. Ashley
Intelligent Systems Program,
University of Pittsburgh
USA
ashley@pitt.edu
accessible to the lay public. This depends, however, on whether the
summaries capture the gist of the argument in the decision. In prior
work, we proposed that such case summaries could be generated
by extracting legal argument triples (IRC triples) including: 1) the
major issues a court addressed in the case, 2) the court’s conclusion
with respect to each issue, and 3) the court’s reasons for reaching
the conclusion.
In [23], we evaluated whether a machine learning (ML) model
can identify the components of legal argument triples in summaries
prepared by legal professionals. We applied traditional ML algo-
rithms (random forest variations) and deep neural network models
(LSTM, CNN and FastText) to identify the sentence components of
IRC triples in legal summaries and to the task of binary classifying
sentences (IRC vs. non-IRC) in the summaries and corresponding
full text decisions. While the performance on the summaries was
promising, the performance on the full texts was quite poor.
In this work, we have substantially increased the size of the
annotated data set of full case texts compared to the prior work. We
focus on applying deep learning algorithms (LSTM, CNN), includ-
ing pre-trained transformer models (RoBERTa, CNN-BERT) with
different loss functions to deal with the continuing challenge of
data imbalance in our training set. We report the results of apply-
ing the different kinds of neural models on cases’ full texts after
training with manually-mapped human-annotated sentences from
the summaries and analyze the effects of using different loss func-
tions and embeddings. We also report results of a proof-of-concept
experiment that applied automatically mapped human-annotated
sentences from the summaries to the full-texts in order to classify
argument triples. If this succeeds, we would not need to manually
annotate the full texts. It would suffice to manually annotate the
summaries, automatically map those summary annotations to the
full texts, and train a model directly on the full texts.
2 RELATED WORK
Argument mining research in the legal domain has focused on ex-
tracting propositions, premises, conclusions, and nested argument
structures [16], argument schemes such as by example [6], rhetori-
cal and other roles that sentences play in legal arguments [1, 19],
stereotypical fact patterns that strengthen a side’s claim (i.e., legal
factors) in domains like trade secret law [4], reasons or warrants in
arguments citing facts or principles [21], functional parts of legal
decisions such as analysis or conclusions [20], and segments by
topic [13] or by linguistic analysis [5, 7, 22].
We aim to identify legal argument triples and employ them
to succinctly summarize case summaries. Yamada, et al. [24] have
summarized Japanese judgments in terms of issues, conclusions, and
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Huihui Xu, Jaromir Savelka, and Kevin D. Ashley

framings. The legal argument triples we seek to employ are simpler

types, which have not been tailored to Japanese legal judgements.
In addition, we make use of a set of case summaries prepared by
human experts to assist with extracting argument triples from the
full case texts, a resource not employed in the cited work.

3 DATA SET
We defined the components of legal argument triples as follows:

(1) Issue – Legal question which a court addressed in the case.

(2) Conclusion – Court’s decision for the corresponding issue.
(3) Reason – Sentences that elaborate on why the court reached
the Conclusion.

All non-annotated sentences are treated as non-IRC sentences.

Two paid third-year law school students annotated sentences
from the human-prepared summaries to identify the issues, reasons,
and conclusions. Both students annotated 574 randomly selected
pairs from the 28,733 case/summary pairs that were available to
us. The total number of sentences from the corresponding full
texts is 120,707, which is significantly more than the summaries’
7,484 sentences. The total number of sentences from full texts are
significantly more than sentences from summaries.
Both annotators followed an 8-page Annotation Guide prepared
by the third author, a law professor, in order to mark-up instances
of IRC sentence types in the summaries. Using the Gloss annota-
tion environment of the second author, the annotators worked on
successive batches of summaries during a series of weeks. After an-
notating each batch, the annotators resolved any coding differences
Figure 1: Distribution of annotated IRC type sentences in
in regular Zoom meetings attended by the first and third authors.
summaries (top) and in full texts (bottom).
The procedure for annotating the full texts of cases differs from
annotating the summaries. For each annotated sentence in the
summary, the Annotation Guide instructs annotators to search the
full text of the case for those sentences that are most similar to the Cohen’s 𝜅 [2] is used to measure the degree of agreement be-
summary sentence and to assign them the same label (i.e., Issue, tween two annotators after their independent annotations of each
Conclusion, or Reason) as in the summary. Annotators may pick batch of summaries. The mean of Cohen’s 𝜅 coefficients across all
terms or phrases from the annotated summary sentences and search types for summaries is 0.709, and the mean for full texts is 0.827. Ac-
for corresponding sentences in the full texts. If the annotators find cording to [11], both scores indicate substantial agreement between
corresponding sentences, they do not need to read the full text of annotators about the sentence type. For the summary annotation,
the case. The Guide warns the annotators that there may not be the mean of Reason agreement is the lowest and Issue and Conclu-
an exact correspondence between the annotated sentences in the sion’s are the highest. Reasons are more challenging since they tend
summary and those in the full text of the case. This makes sense; to include descriptions of case facts. The agreement scores of full
having selected sentences in the full case texts to include in the texts are higher than the summaries’ scores. Since the full text anno-
summary, the summarizers probably edited them, for example, by tation took place after reconciling disagreements for the summary
combining some short sentences. annotation, the full text scores were expected to be higher.
By using the summaries’ annotations as anchors to target cor- Figure 1 reports the distributions of final consensus labels from
responding sentences in the full text, we attempted to leverage summaries and full texts. The most frequent label is the non-IRC
the summarizers’ work in selecting important sentences and the label for both summaries and full texts. The second most frequent
annotators’ work in marking up some of those sentences as issues, label is the Reason label for both summaries and full texts. The
conclusions, or reasons. We developed this strategy to expedite label distribution is aligned with our observation: Reasons tend to
the process of full text annotation which would be much more be more elaborated than Issues and Conclusions.
time-consuming and costly if performed directly on the full texts.
The strategy is based on the observation that sentences of sum- 4 EXPERIMENTS
maries stem from those in the full texts. The strategy also helps us We aim to leverage state-of-the-art deep learning models to identify
to confirm the mapping relationship between summaries and full the role a sentence may play in a case as an Issue, Conclusion, or
texts. This in turn helps us to develop the heuristic for automatic Reason. For our experiments, we used 80% of the full texts as the
mapping. training set, 10% as the validation set and 10% as the test set.

251
Toward Summarizing Case Decisions via Extracting Argument Issues, Reasons, and Conclusions
In this section, we present the details of the models, including
convolutional [10] and recurrent neural networks [14], BERT-based
neural networks [12] and a hybrid neural model combining a con-
volutional neural network with BERT embedding [8]. We also ex-
periment with different loss functions for those neural networks,
including cross entropy loss, F1 loss and focal loss.
4.1 Model Architectures
Convolutional Neural Networks. Convolutional neural networks
(CNN) utilize convolutional filters to extract local features. Origi-
nally applied to computer vision tasks, CNNs have also achieved a
high level of performance in sentence classification tasks. [9].
In our study, we use the settings of hyperparameters of filters
from [9]: filter sizes of 3, 4, and 5 with 100 for each size of filter.
In other words, the models are looking for tri-grams, 4-grams and
5-grams in sentences.
Long Short-Term Memory Networks. Long Short-Term Memory (LSTM)
networks, a different RNN architecture, overcomes the vanishing
gradient problem by employing a cell to control removing or adding
information throughout the whole training process [14].
GloVe [17] is an unsupervised learning algorithm for obtaining
vector representations for words1 . We used “glove.6B.100d” as pre-
trained word embeddings to feed into the LSTM model, where the
vectors were trained on 6-billion tokens and have 100 dimensions.
Dropout is also adapted to the LSTM model.
BERT-based Neural Networks. Google AI Language introduced Bidi-
rectional Encoder Representations from Transformers (BERT) in
2018 [3]. Instead of using single word embedding like GloVe, BERT
takes the context into account by using bidirectional pre-training
for language representations. This pre-training method is intended
to better grasp contextual meaning of a language than single-
directional pre-training.
RoBERTa [12] replicates BERT training using an improved train-
ing methodology with more data and computational resources. For
our study we used RoBERTa in its default configuration.
Convolutional Neural Network with BERT embedding. CNN with
BERT embedding takes BERT-pretrained embeddings as input and
feeds them into a CNN model for classification. Unlike GloVe pre-
trained word embedding, BERT-pretrained embedding is not a static
embedding. As sentences are fed in, it produces the word embed-
dings in real time.
We combine the two models: a BERT-based model and a CNN
classification model. The encoded text passes through the BERT
model first and produces BERT embeddings. The dimension of the
BERT embedding (768) is higher than that of GloVe pre-trained
word embedding (100). Other hyperparameters remain the same as
for the CNN-only model.
4.2 Automatic mapping
As mentioned in Section 3, the human annotation of full texts uti-
lizes manual mapping: annotators used key words from annotated
sentences in original longer summaries to find corresponding sen-
tences in full texts. Even without actually reading the full case,
1 https://nlp.stanford.edu/projects/glove/
252
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Table 1: Comparison of manually annotated IRC summary
sentences with top 1 automatically ranked full-text sen-
tences (Sentence-BERT embedding with cosine similarity)
Issue
Manual Damage to both vehicles exceeded the insurance deductibles and both parties
claim damages against each other for the amount of the deductibles.
Rank 1 The damages to both the truck and the car exceeded the $500.00 insurance
deductible. [. . . ]
Reason
Manual The plaintiff should have taken more appropriate measures to avoid the acci-
dent
Rank 1 Even if Schmidt concluded that Henry was going to proceed into his path, he
had more appropriate alternatives than locking his brakes and turning to the
right.
Conclusion
Manual Fault for this accident was attributed 10% to the defendant and 90% to the
plaintiff.
Rank 1 I attribute 10% of the fault in this accident to Henry and 90% to Schmidt.
annotators still need to read contextual information around a sen-
tence to confirm the mapping and IRC type.
We undertook a proof-of-concept experiment to assess if a strat-
egy of automatic mapping could make the process more efficient
in the future. The idea is to employ sentence embedding to map
annotated summary sentences to full texts. Sentence embedding
can represent a sentence and capture semantic information as vec-
tors. Cosine similarity is used to examine the degree of similarity
between sentences in annotated summaries and full texts.
Sentence-BERT Embedding. Sentence embedding techniques rep-
resent the entire sentence and semantic information as vectors.
Sentence-BERT is a modification of the BERT neural model that
uses siamese and triplet networks to produce semantically mean-
ingful sentence embeddings [18]. Sentence-BERT has achieved high
levels of performance in measuring the similarity of sentential ar-
guments in [15]. Considering the size of our data set, we chose
to use the BERT base model for sentence embeddings with 768
dimensions.
We calculated the cosine similarity score for annotated sentences
from a summary to every sentence in its corresponding full text.
All the similarity scores are ranked in descending order, and only
the top 5 sentences are selected as useful. The remaining sentences
are marked as non-IRC type sentences.
There are reasons to believe that the automatically mapped sen-
tences bear useful similarities to manually mapped ones. Table 1
shows examples comparing manually mapped IRC sentences and
automatically mapped sentences. The top ranked sentences often
include the same key words as the manual sentences do. How-
ever, the Reason sentences that the algorithm prefers have fewer
overlapping keywords than Issue and Conclusion.
5 RESULTS
Table 2 reports scores for the classification on the test split of the
full texts. The left side of the table reports the results of training
on the full text sentences corresponding to the manually mapped
human-annotated sentences from the summaries.
We tested LSTM, CNN, RoBERTa and CNN-BERT with three
different loss functions: cross-entropy loss, focal loss and F1 loss.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Huihui Xu, Jaromir Savelka, and Kevin D. Ashley

Table 2: Scores for test set on both manually mapped full text sentences and automatically mapped full text sentences by
using LSTM, CNN, RoBERTa, CNN-BERT. The abbreviations of Issue, Reason and Conclusion are I, R, and C. The suffixes are
-P(precision), -R(recall). Ave-F1 stands for the average of class-wise F1 scores.

Training on manually mapped data Training on automatically mapped data

I-P I-R I-F1 R-P R-R R-F1 C-P C-R C-F1 Ave-F1 I-P I-R I-F1 R-P R-R R-F1 C-P C-R C-F1 Ave-F1
LSTM(cross-entropy) 0.72 0.49 0.58 0.35 0.09 0.15 0.72 0.42 0.53 0.42 0.19 0.41 0.26 0.28 0.16 0.20 0.10 0.23 0.14 0.20
LSTM(focal) 0.75 0.43 0.54 0.38 0.11 0.17 0.72 0.35 0.47 0.39 0.17 0.34 0.22 0.16 0.07 0.09 0.16 0.42 0.23 0.18
LSTM(F1) 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.22 0.49 0.30 0.06 0.43 0.11 0.08 0.44 0.13 0.18
CNN(cross-entropy) 0.91 0.34 0.49 0.58 0.08 0.14 0.65 0.36 0.46 0.36 0.27 0.23 0.25 0.25 0.01 0.02 0.19 0.17 0.18 0.15
CNN(focal) 1.00 0.28 0.44 0.44 0.14 0.22 0.73 0.28 0.40 0.35 0.30 0.30 0.30 0.13 0.11 0.12 0.17 0.37 0.23 0.22
CNN(F1) 0.73 0.39 0.50 0.43 0.19 0.27 0.73 0.40 0.52 0.43 0.18 0.50 0.26 0.10 0.23 0.14 0.12 0.49 0.19 0.20
RoBERTa(cross-entropy) 0.35 0.38 0.36 0.08 0.01 0.01 0.71 0.09 0.16 0.18 0.21 0.47 0.29 0.16 0.23 0.19 0.19 0.37 0.25 0.24
RoBERTa(focal) 0.36 0.27 0.31 0.00 0.00 0.00 0.35 0.08 0.12 0.14 0.24 0.41 0.30 0.15 0.20 0.17 0.16 0.31 0.21 0.23
RoBERTa(F1) 0.26 0.63 0.36 0.20 0.22 0.21 0.35 0.51 0.41 0.33 0.17 0.63 0.27 0.14 0.28 0.18 0.30 0.44 0.35 0.27
CNN-BERT(cross-entropy) 0.11 0.43 0.18 0.54 0.10 0.17 0.96 0.13 0.23 0.19 0.10 0.59 0.17 0.45 0.07 0.12 0.68 0.28 0.39 0.23
CNN-BERT(focal) 0.12 0.39 0.18 0.03 0.09 0.05 0.46 0.29 0.36 0.20 0.18 0.48 0.26 0.63 0.09 0.15 0.76 0.21 0.33 0.25
CNN-BERT(F1) 0.57 0.50 0.53 0.37 0.18 0.24 0.85 0.27 0.41 0.39 0.13 0.61 0.22 0.09 0.21 0.12 0.44 0.18 0.26 0.20

All the models were trained for 30 epochs. We stored models’ check-
points after each epoch and evaluated on the separate validation set.
The models with lowest loss value on the validation were selected
for the classification on the test set. The performance on the test
set is shown in the table. Average F1 is the average of class-wise F1
scores.
We found that LSTM with the cross-entropy loss function achieved
the highest F1 scores on identifying Issues and Conclusions which
are, 0.58 and 0.53, respectively. Both CNN with the F1 loss function
and CNN-BERT with cross-entropy loss have the highest F1 scores
(0.27) on identifying Reasons. On average F1, CNN with the F1 loss
function reached 0.43 which is highest among all the models.
The right side of of the table reports the performance of the
models that were trained on the full text sentences corresponding
to the automatically mapped human-annotated sentences from the
summaries. Both LSTM with F1, CNN with focal and RoBERTa
with focal tied in their performance on Issues (0.30). LSTM with
cross-entropy has the highest F1 (0.20) on Reasons. For Conclu-
sion classification, CNN-BERT with cross-entropy loss achieves the
highest performance (0.39). Finally, RoBERTa with F1 loss achieves
the highest score in terms of average F1.
Surprisingly, RoBERTa and CNN-BERT with cross-entropy and
focal losses perform better on automatically mapped data than
manually mapped data in terms of average F1. However, LSTM and
CNN do not show the same pattern. The automatically mapped data
are selected by sentence similarity scores with respect to BERT-
Sentence embedding. RoBERTa and CNN-BERT somehow take the
advantage of information contained in the sentence embedding
to make a better classification. We are not sure how it affects the
performance and will investigate it further. We also observed that
models tend to perform better on Issue and Conclusion than Reason
despite the type of training set. Since Reasons frequently include
case facts, it is harder for models to classify them.
Despite the relative comparable F1 scores between training on
manually mapped data and automatically mapped data, the preci-
sion of all types of sentences drops significantly in most cases when
models are trained on automatically mapped data. This means that
models have lower probability of making correct classifications.
When we compare the performance among the same models with
different loss functions, the model with F1 loss function always has
the highest average F1 score when trained on manually mapped
data except for LSTM. The same pattern does not hold for the
automatically mapped data. Each loss function has its own strength
in terms of training set and model selection.
6 DISCUSSION AND ERROR ANALYSIS
6.1 Discussion
The results for classifying full text sentences trained on automati-
cally mapped data, the right side of Table 2, are significantly higher
than trained on annotated summaries in prior work. There the
highest F1 scores for full texts trained on annotated summaries
were Issue (0.27), Reason (0.14), and Conclusion (0.24). We attribute
this improvement to using manually-mapped training sentences in
the full texts, the higher numbers of annotated data, and the use
of deep learning algorithms plus transformer models (LSTM, CNN,
RoBERTa, and CNN-BERT).
As noted, we tried different kinds of neural models paired with
different loss functions. We confirmed that the F1 loss function im-
proved the performances of CNN and RoBERTa: RoBERTa with F1
loss yielded 0.21 on Reason and 0.41 on Conclusion while RoBERTa
without F1 loss produced only 0.01 on Reason and 0.16 on Conclu-
sion. When a loss function is aligned with the evaluation metric, it
is likely to improve model performance. LSTM, however, did not
perform well with the F1 loss function: on the manually mapped
data , LSTM(F1) yielded 0.0 on all IRC types.
Those models each have advantages for certain sentence types.
LSTM(cross-entropy) yielded the highest F1 scores on Issues and
Conclusions. CNN(F1) and CNN-BERT(cross-entropy) performed
best on identifying Reasons. In general, models have difficulty iden-
tifying Reason sentences, since Reasons have more complex se-
mantic meanings. As noted, Reasons are intertwined with facts,
which can easily be classified as the non-IRC type. The annotators

253
Toward Summarizing Case Decisions via Extracting Argument Issues, Reasons, and Conclusions
confirmed that Issues and Conclusions are easier to catch. They
employ distinct keywords such as “issue”, “conclusion”, etc.
LSTM has the ability to detect temporal information about a
sequence and can handle arbitrary input lengths. Meanwhile, CNN
can only accept fixed size input. We think the ability to handle se-
quential information and longer lengths make LSTM more suitable
for Issues and Conclusions, since these involve plainer language
than Reasons. CNN has the upper hand on spotting Reasons. The lit-
eral composition of Reasons is more diverse than that of Issues and
Conclusions; the convolutional features can capture this diversity.
6.2 Error Analysis
With respect to the right side of Table 2, the proof-of-concept study
training models on automatically mapped data, the results suggest
that classifying argument triples is feasible, but less effective than
with the manually mapped data when taking precision and recall
into account. We are particularly interested in the errors that the
models made classifying Reasons. As noted, targeting the Reasons
correctly is harder since they tend to be more complex and diverse
than Issues and Conclusions.
Some of the misclassifications involved phrases attributing an
expressed view to the judge. This is a positive sign, in that such
self-referential judicial sentences are relatively less frequent in a
case opinion and indicate sentences where the judge is more likely
to assert that something is an Issue, Conclusion, or Reason. On
the other hand, such self-referential attribution phrases do not
necessarily discriminate among the three classifications.
7 FUTURE WORK
We plan to continue to annotate new cases in order to increase the
size of the training set. Currently, the corpus includes 574 anno-
tated summary / full text pairs. The size of the data set is still not
large enough for adequately training more complex neural network
models. The data set is sufficiently large, however, to allow us to
continue to explore models and identify some challenges. The ex-
perience helps us to improve the quality of data as well as informs
our intuitions about how human summarizers do their work. We
expect that the more annotated data we collect the more interesting
properties we will be able to observe in this process.
As noted, prior work explored different sampling strategies for
dealing with imbalanced data to improve model performance. Dif-
ferent sampling methods have their merits in terms of their effects
on training sets and model types. In this study, we briefly inves-
tigated a different method of adding augmented data to improve
the performance of the models. Although the results were not as
we expected, we observed that it had some positive effect on iden-
tifying Reasons from full texts. We will continue to explore other
methods to deal with our imbalanced data.
We also plan to test whether a pre-trained legal language model
improves performance over a generic language model.
ACKNOWLEDGMENTS
This work has been supported by grants from the Autonomy through
Cyberjustice Technologies Research Partnership at the University
of Montreal Cyberjustice Laboratory and the National Science Foun-
dation, grant no. 2040490, FAI: Using AI to Increase Fairness by
254
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Improving Access to Justice. The Canadian Legal Information In-
stitute provided the corpus of paired legal cases and summaries.
Computation resources are provided by the Center for Research
Computing at the University of Pittsburgh.
REFERENCES
[1] A. Bansal, Z. Bu, B. Mishra, S. Wang, K. Ashley, and M. Grabmair. 2016. Document
Ranking with Citation Information and Oversampling Sentence Classification in
the LUIMA Framework.
[2] J. Cohen. 1960. A coefficient of agreement for nominal scales. Educational and
psychological measurement 20, 1 (1960), 37–46.
[3] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. Bert:
Pre-training of deep bidirectional transformers for language understanding. arXiv
preprint arXiv:1810.04805 (2018).
[4] M. Falakmasir and K. Ashley. 2017. Utilizing Vector Space Models for Identifying
Legal Factors from Text.. In JURIX. 183–192.
[5] A. Farzindar and G. Lapalme. 2004. Legal text summarization by exploration of
the thematic structure and argumentative roles. In Text Summarization Branches
Out. 27–34.
[6] V. Feng and G. Hirst. 2011. Classifying arguments by scheme. In Proceedings of
the 49th annual meeting of the association for computational linguistics: Human
language technologies. 987–996.
[7] C. Grover, B. Hachey, and C. Korycinski. 2003. Summarising legal texts: Sen-
tential tense and argumentative roles. In Proceedings of the HLT-NAACL 03 Text
Summarization Workshop. 33–40.
[8] Changai He, Sibao Chen, Shilei Huang, Jian Zhang, and Xiao Song. 2019. Us-
ing convolutional neural network with BERT for intent determination. In 2019
International Conference on Asian Language Processing (IALP). IEEE, 65–70.
[9] Yoon Kim. 2014. Convolutional Neural Networks for Sentence Classification.
CoRR abs/1408.5882 (2014). arXiv:1408.5882 http://arxiv.org/abs/1408.5882
[10] Kamran Kowsari, Kiana Jafari Meimandi, Mojtaba Heidarysafa, Sanjana Mendu,
Laura Barnes, and Donald Brown. 2019. Text classification algorithms: A survey.
Information 10, 4 (2019), 150.
[11] J. Landis and G. Koch. 1977. The measurement of observer agreement for cate-
gorical data. Biometrics (1977), 159–174.
[12] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A
robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692
(2019).
[13] Qi. Lu, J. Conrad, K. Al-Kofahi, and W. Keenan. 2011. Legal document clustering
with built-in topic segmentation. In Proc. 20th ACM int’l conf. Info. and knowledge
management. 383–392.
[14] S. Minaee, N. Kalchbrenner, E. Cambria, N. Nikzad, M. Chenaghlu, and J. Gao.
2020. Deep learning based text classification: A comprehensive review. arXiv
preprint arXiv:2004.03705 (2020).
[15] Amita Misra, Brian Ecker, and Marilyn A Walker. 2017. Measuring the similarity
of sentential arguments in dialog. arXiv preprint arXiv:1709.01887 (2017).
[16] R. Mochales and M. Moens. 2011. Argumentation mining. Artificial Intelligence
and Law 19, 1 (2011), 1–22.
[17] Jeffrey Pennington, Richard Socher, and Christopher D Manning. 2014. Glove:
Global vectors for word representation. In Proceedings of the 2014 conference on
empirical methods in natural language processing (EMNLP). 1532–1543.
[18] Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings
using Siamese BERT-Networks. In Proceedings of the 2019 Conference on Em-
pirical Methods in Natural Language Processing. Association for Computational
Linguistics. http://arxiv.org/abs/1908.10084
[19] M. Saravanan and B. Ravindran. 2010. Identification of rhetorical roles for
segmentation and summarization of a legal judgment. Artificial Intelligence and
Law 18, 1 (2010), 45–76.
[20] J. Savelka and K. Ashley. 2018. Segmenting U.S. Court Decisions into Functional
and Issue Specific Parts. In Proceedings, 31st Int. Conf. on Legal Knowledge and
Information Systems, Jurix. 111–120.
[21] O. Shulayeva, A. Siddharthan, and A. Wyner. 2017. Recognizing cited facts
and principles in legal judgements. Artificial Intelligence and Law 25, 1 (2017),
107–126.
[22] A. Wyner, R. Mochales-Palau, M. Moens, and D. Milward. 2010. Approaches to
text mining arguments from legal cases. In Semantic processing of legal texts.
Springer, 60–79.
[23] Huihui Xu, Jaromír Šavelka, and Kevin D Ashley. 2020. Using Argument Mining
for Legal Text Summarization. Legal Knowledge and Information Systems JURIX
(2020), 184–193.
[24] H. Yamada, S. Teufel, and T. Tokunaga. 2019. Building a corpus of legal argumenta-
tion in Japanese judgement documents: towards structure-based summarisation.
Art. Int. and Law 27, 2 (2019), 141–170.
 Part III

Extended Abstracts
 CriminelBART: A French Canadian Legal Language Model
Specialized in Criminal Law∗
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel
Laval University, Computer Science Department and Faculty of Law
Québec, Canada
nicolas.garneau@ift.ulaval.ca,eve.gaumond@observatoire-ia.ulaval.ca
luc.lamontagne@ift.ulaval.ca,pierre-luc.deziel@fd.ulaval.ca
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Neural networks;
KEYWORDS
Criminal Law, Language Model, Cloze Tests, Text Generation
ACM Reference Format:
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel. 2021.
CriminelBART: A French Canadian Legal Language Model Specialized in
Criminal Law. In Eighteenth International Conference for Artificial Intelligence
and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York,
NY, USA, 2 pages. https://doi.org/10.1145/3462757.3466147
1 INTRODUCTION
Learning language representations is a key component in many
natural language processing tasks, and their usefulness is most often
challenged by specialized target domains and vocabulary. We have
witnessed several neural causal language models (CLM) that learn
contextual representations such as ELMo [8]. More recently, the
Transformer architecture [10] has tremendously improved language
representation learning, giving birth to new architectures such as
BERT [4], a masked language model, pushing the state-of-the-art
of natural language understanding to an unprecedented level of
performance on standard benchmarks. Moreover, it has been found
that Transformer-based CLM, such as GPT [9], are excellent feature
extractors as well as being impressive text generators. BART [7],
an architecture combining the backbone of both BERT and GTP
proved to be particularly effective at generating text while being
competitive in comprehension tasks. BARThez, the French version
of BART, was recently introduced as a pre-trained model on a very
large monolingual French corpus [6]. In this paper, we introduce
CriminelBART, a fine-tuned version of BARThez specialized for
criminal law using a French Canadian corpus of legal judgments,
and we evaluate its performance on different tasks.
∗ By “Specialized” we mean a language model that knows a little more than the average
citizen. CriminelBART is far from having the competencies of a criminal lawyer.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466147
256
2 CRIMINELBART
There has been few attempts to fine-tune pre-trained language
models to make them effective on the legal domain. For exam-
ple, Chalkidis et al. [3] proposed to fine-tune BERT on English law
texts scraped from publicly available resources, yielding LegalBERT.
These recent advances are all characterized by the same underlying
source language that is English. We thus propose in this paper a
new French language model specialized in Canadian Criminal law,
a field gaining more attention recently [1].
One important issue for this model is the choice of vocabulary
that differs in important ways between private and public law.
The legal remedies are not the same. Indeed, criminal cases are
mainly about fines and prison whereas civil cases are mainly about
damages. Also, the lawyer and their clients are not called the same
whether it is a criminal trial or a civil matter, etc. We thus deemed
it essential to instill this criminal specialization into a language
model. Just as a lawyer could not be an expert in every domain, we
thought that it would be ill-advised to develop a generalist legal
model to perform very specific tasks in the criminal field such as
plumitifs description generation[1].
The dataset we are working with is a French legal corpus specific
to the criminal law. It describes several offenses from different
laws such as the Criminal Code and the Controlled Drugs and
Substances Act. This dataset, comprising around 9000 judgments
from the last 10 years has been gathered from the Criminal and
Penal Chamber and was extracted from the Société Québécoise
d’Information Juridique (SOQUIJ) website1 .
To construct a specialized model from this corpus, we propose
to further pre-train BARThez [5] according to BART’s pre-training
objectives (i.e. denoising auto-encoder) using the fairseq library2
for 250,000 steps achieving a train and validation perplexity of 1.95
and 1.92 respectively. We then assess the “legal comprehension” of
CriminelBART with three different Cloze tests regarding the predic-
tion of criminal charges, legal provisions, and individuals. Given a
textual utterance and deleted passages, a model undergoing a Cloze
test tries to fill in the blanks with the best possible answers. It thus
requires the ability to understand the context and the vocabulary.
Formally, given a textual utterance of length 𝑛, a model successfully
pass the Cloze Test by maximizing the following probability;
𝑃 (𝑤𝑖 |𝑤 1, 𝑤 2, . . . , 𝑤𝑖−1, 𝑤𝑖+1, . . . , 𝑤𝑛−1, 𝑤𝑛 ) (1)
where 𝑤𝑖 could be a word or a phrase.
1 https://soquij.qc.ca/
2 https://github.com/pytorch/fairseq
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Criminal charges. In order to determine if CriminelBART grasped
the accusations’ distribution from the corpus, we ask the model to
fill in the following sentence (translated in English):
The defendant is accused of <Mask> under the Criminal Code.
The top 5 predicted passages (with semantic duplicates being
removed3 ) comprises “driving under the influence”, “aggravated
assault”, “possession of narcotics”, “dangerous driving”, and “hit-
and-run”. As expected, CriminelBART suggests crimes related to
the driving of motor vehicles and some other infractions related
to controlled drugs and substances which constitute the main ac-
cusations in the corpus (up to 50%). Nonetheless, CriminelBART
predicts various crimes while BARThez, a more generic model,
mainly returns crimes restricted to premeditated murder.
Legal provisions. We also probe CriminelBART for legal provi-
sions such that, given a context, it predicts which provision it is
associated to. To this end, we created 84 cloze tests4 on 28 provi-
sions, where CriminelBART achieves an accuracy of 64% over all
provisions. Unsurprisingly, BARThez achieves 0% by predicting ran-
dom tokens. Here is an example regarding provision 4, “Possession
of substance”, from the controlled drugs and substances act;
PER is put on trial on charges of possession of cannabis,
thus committing the criminal act provided in article <Mask>.5
By analyzing the predictions, we can see that CriminelBART
demonstrates some semantic understanding of the different crim-
inal offenses. For example, section 5 (“Trafficking of substance”)
and section 7 (“Production of substance”) are both predicted in
this previous context along with section 4. It is relevant as these
criminal offenses are most likely to appear with one another and
are semantically related. Interestingly, section 95 (“Possession of
prohibited or restricted firearm with ammunition”) is not within the
top 10 predicted sections for this context. This syntactically good
candidate, i.e the “possession” of something illegal, is rejected by the
model as not a semantically good one. These results exhibit Crim-
inelBART capability to establish semantic relationships between
certain provisions due to their collocation or similar contexts.
Privacy. When it comes to neural language models, there is
inevitably some privacy concerns that arise. Numerous papers such
as Carlini et al. [2] report on the leakage of training data from these
kinds of models. In the context of CriminelBART, we feared that it
would be possible to reidentify someone from the corpus simply
by querying the model with the right questions; e.g. would it be
possible to reidentify one defendant by asking the model who was
accused of one particular offense. We thus designed a cloze test
where the context contains a specific accusation (e.g. fraud) and
ask CriminelBART to predict the name of the defendant;
Mr./Ms. <mask> is accused of fraud under the criminal code.
We did so for the top 10 provisions contained in the judgment
corpus and found out that the individual was not changing for
different accusations. Indeed, the name most often predicted (either
male or female) was the name of a judge appearing most often
in the corpus. From the other predicted names, we could not find
any association between an individual and who committed the
3 We consider “impaired driving” and “driving while impaired” as semantic duplicates.
4 Availablehere: https://bit.ly/3edmM4N
5 We anonymized the defendant’s name with the generic token PER on purpose.
257
Nicolas Garneau, Eve Gaumond, Luc Lamontagne, Pierre-Luc Déziel
crimes. Similar experiments with a generic BARThez model always
result with a few unrelated names such as“Mr. Gagné”. While this
experiment does not expose a clear bias in CriminelBART, the sole
possibility, how small could it be, that defendants’ names may be
coming out of this model is a privacy matter that cannot be ignored.
Even though the identity of judges is deemed public, this is not to
be taken lightly. In a context where judges are already reluctant to
the idea that their work might be subject to analyses conducted by
AI, it is important not to scare them with uses of technology that
may be prejudicial to them. A more in-depth analysis should be
conducted before releasing this language model at scale, which is
kept as future work. This decision is also highly motivated by the
recent critics on language model being stochastic parrots and the
different attacks performed on them in order to extract the training
set. In our future works, we wish to leverage CriminelBART as a
textual description generator of plumitifs, short legal documents
known to be unintelligible [1].
Acknowledgements. We thank the reviewers for their insightful com-
ments. This research was funded by both the Natural Sciences and Engi-
neering & Social Sciences and Humanities Research Councils of Canada.
REFERENCES
[1] David Beauchemin, Nicolas Garneau, Eve Gaumond, Pierre-Luc Déziel, Richard
Khoury, and Luc Lamontagne. 2020. Generating Intelligible Plumitifs Descrip-
tions: Use Case Application with Ethical Considerations. In Proceedings of the
13th International Conference on Natural Language Generation. Association for
Computational Linguistics, Dublin, Ireland, 15–21. https://www.aclweb.org/
anthology/2020.inlg-1.3
[2] N. Carlini, Florian Tramèr, Eric Wallace, M. Jagielski, Ariel Herbert-Voss, K.
Lee, A. Roberts, Tom Brown, D. Song, Úlfar Erlingsson, Alina Oprea, and Colin
Raffel. 2020. Extracting Training Data from Large Language Models. ArXiv
abs/2012.07805 (2020).
[3] Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras,
and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of
Law School. In Findings of the Association for Computational Linguistics: EMNLP
2020. Association for Computational Linguistics, Online, 2898–2904. https:
//doi.org/10.18653/v1/2020.findings-emnlp.261
[4] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proceedings of the 2019 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Volume 1 (Long and
Short Papers). Association for Computational Linguistics, Minneapolis, Minnesota,
4171–4186. https://doi.org/10.18653/v1/N19-1423
[5] M. K. Eddine, Antoine J.-P. Tixier, and M. Vazirgiannis. 2020. BARThez: a Skilled
Pretrained French Sequence-to-Sequence Model. ArXiv abs/2010.12321 (2020).
[6] Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin
Lecouteux, Alexandre Allauzen, Benoit Crabbé, Laurent Besacier, and Didier
Schwab. 2020. FlauBERT: Unsupervised Language Model Pre-training for French.
In Proceedings of the 12th Language Resources and Evaluation Conference. European
Language Resources Association, Marseille, France, 2479–2490. https://www.
aclweb.org/anthology/2020.lrec-1.302
[7] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman
Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART:
Denoising Sequence-to-Sequence Pre-training for Natural Language Generation,
Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of
the Association for Computational Linguistics. Association for Computational
Linguistics, Online, 7871–7880. https://doi.org/10.18653/v1/2020.acl-main.703
[8] Matthew Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark,
Kenton Lee, and Luke Zettlemoyer. 2018. Deep Contextualized Word Repre-
sentations. In Proceedings of the 2018 Conference of the North American Chapter
of the Association for Computational Linguistics: Human Language Technologies,
Volume 1 (Long Papers). Association for Computational Linguistics, New Orleans,
Louisiana, 2227–2237. https://doi.org/10.18653/v1/N18-1202
[9] A. Radford. 2018. Improving Language Understanding by Generative Pre-
Training.
[10] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All
You Need. In NIPS.
 Applying Decision Tree Analysis to Family Court Decisions:
Factors Determining Child Custody in Taiwan
Sieh-Chuen Huang Hsuan-Lei Shao∗
College of Law East Asia Studies Dept.
National Taiwan University National Taiwan Normal University
Taiwan Taiwan
schhuang@ntu.edu.tw hlshao2@gmail.com
CCS CONCEPTS
• Information systems applications → Data mining; • Decision
support systems → Expert systems.
KEYWORDS
child custody, best interests of the child, legal factor, machine learn-
ing, decision tree.
1 INTRODUCTION
The doctrine of “best interests of the child” has guided courts in
determining post-divorce child custody cases in Taiwan since the
amendment of the Taiwan Civil Code in 1996, which overturned
previous patriarchal practices at least as a matter of law. Previous
empirical studies have adopted descriptive statistics in analyzing
court cases to determine which factors, as set out in Article 1055-1,
are the ones judges tend to consider. However, these approaches
do not clarify which factors judges consider primary.
This study collects Taiwanese family court decisions from 2012
to 2017. The study employs decision tree analysis, a commonly used
machine learning technology. This appears to be the first published
application worldwide of machine learning to analysis of family
court decisionmaking.
The study concludes that the three most significant factors con-
sidered by judges in Taiwan are first, which parent is the child’s
current primary caregiver, followed by the wishes of the child
and the judge’s assessment of the relative quality of each parent’s
parent-child interaction. This result runs counter to widely held
beliefs that parental gender and parents’ occupations and economic
resources are still prime factors in judges’ contemplation.
2 BACKGROUND OF TAIWAN CHILD
CUSTODY WHEN PARENTS DIVORCE
Before the 1996 amendment, Taiwan’s Civil Code had stipulated
that, in both consensual and judicial divorce, the custody of children
∗ Corresponding
author
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466076
258
Robert B Leflar
School of Law, University of Arkansas
United States (retired 2020)
College of Law, National Taiwan
University, Taiwan (since 2020)
rbleflar@uark.edu
belonged to the father unless either it had been agreed otherwise
in a consensual divorce (Article 1051) or the court had decided
otherwise (Article 1055). The 1996 amendment repealed Article 1051
and amended Article 1055, replacing paternal preference with the
“best interests of the child” doctrine and recognizing joint custody
and non-custodial parents’ visitation rights. New Article 1055-1
lists several factors that judges must consider, such as the age, sex,
birth order, health condition, and the wishes of the child, and the
age, occupation, character, economic ability and lifestyle of the
parents. Such a broad standard gives judges considerable discretion
in deciding what is in the best interests of the child.
Our goal is to provide a clear answer to the long-standing debate
on what “most important factors” in fact influence judges’ custody
decisions. To accomplish this goal, we applied a machine learning
algorithm to determine what those factors are. We established a
dataset which we carefully labeled according to our research con-
cerns (weighing normative “factors”). Our model predicts outcomes
accurately, and also provides insights on “most important factors”
in a way useful to parents, their lawyers, and legal scholars.
3 RESEARCH DESIGN
3.1 Data Collecting
All cases except juvenile and sexual assault cases decided by district
courts in Taiwan since 2000 are open to the public on the official
website of the Judicial Yuan. We focus on child custody decisions of
first instance decided by family and district courts. Using carefully
chosen causes of action, keywords and decision dates, we iden-
tified 3,028 child custody decisions between January 1, 2012 and
December 31, 2017.
We then limited our sample to cases in which both parents were
Taiwanese and both sought to acquire custody. This is because when
one parent (usually the defendant) does not come to court or keeps
silent, the other (the plaintiff) is very likely to receive custody. We
excluded these cases from our dataset. Among the 3,028 cases, the
2,096 cases in which one of the parents did not express any opinion
regarding custodian and the 97 cases of transnational marriage were
therefore excluded. The remaining 835 cases contain 1,290 children.
Among them, 1,126 children were under sole custody (87.3%), 159
children were under joint custody (12.3%), and 5 children were
under third party guardianship (0.4%)
3.2 Dataset Construction/ Annotation Labels
We created a model that predicts the value of a dependent variable:
custody granted to father (labeled “1”) or mother (labeled “0”). As
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Huang, et al.

Table 1: Factors Considered by Judges the current caregiver of the child is in the first place. If the caregiver
Character Factor is the mother (represented by the number “3”) or both mother and
Sex father (represented by the number “2”), the model will follow the
Age line (branch) on the left side to the next node, childWill, meaning
Child
Child Willing
that the model predicts that judges will, secondarily, consider the
Emotional feelings between the other persons living
together and the child
child’s wishes. If the child prefers the mother or has no preference
Health condition (>= 2), the model again follows the line on the left side coming to
Character (drug use, alcohol consumptions. . . ) the next end node, which is labeled with a probability distribution
Economy indicating that in the validation set, custody has a very high prob-
Willing ability (95%) of going to mother (the label “0”). On the contrary,
Undue behavior (domestic violence towards the child) even if the mother is the primary caregiver, in the case that the
Parent Parenting time child prefers the father (< 2), the probability that the mother gets
Parenting environment custody becomes fairly low (17% in the validation set).
Friendly parent
Primary caretaker (caregiver) 4.2 Model Efficiency
Understanding the child
Parenting plan The model’s accuracy is 96.5% in its test set and the F1 score is
Parent-child interaction 0.9783, indicating that the model is quite satisfactory.
Both
Current residence
Support system Table 2: Confusion Matrix of Child Custody Model Test Set
Others
Social worker’s report Results predicted by the machine
Test_set_N=226
Positive (for mother) Negative (for father)
independent variables, we manually defined 19 factors from Article
Actual Positive True positive, TP=181 False negative, FN=7
1055-1, social workers’ evaluation items and previous literature results (for mother)
(Table 1 below). Negative False positive, FP=1 True negative, TN=37
(for father)
3.3 Decision Tree Learning
This study adopts the CHAID (Chi-squared Automatic Interaction 5 CONCLUSION
Detector) algorithm to make strategic splits. This algorithm’s pre-
Among the numerous factors stipulated in Article 1055-1 of the Tai-
dictive power performs better than statistical techniques such as
wan Civil Code, “primary caregiver,” “child’s wishes” and “parent-
regressions and it shows the nodes clearly, making the outcomes
child interaction” are the three most significant factors contem-
explainable in terms of the 19 factors listed above.
plated by judges. The pattern of Taiwanese judges’ decision-making
regarding child custody appears to be relatively constant and stable
4 RESEARCH RESULTS
during the six-year period studied.
4.1 Model Demonstration Furthermore, in custody disputes addressed by judicial decisions,
the mother seems to have overwhelming supremacy: in our dataset,
Figure 1: Decision Tree of Child Custody Cases in Taiwan mothers had a 75% likelihood of receiving sole custody.
This research should help legal scholars identify particular cus-
tody cases as either typical or exceptional by comparing the machine-
predicted results to actual judgments. A case may be exceptional,
worth further exploration and research, if the two outcomes are
inconsistent. In addition, this study should assist parents and their
lawyers to preliminarily evaluate their possibilities of acquiring
custody. And when the outcome of litigation can be predicted in
advance by the parties, the likelihood of cases going to court will
fall and the likelihood of settlement will increase. Thus, machine
learning helps us predict what the “law” is and hence contributes
to legal certainty.

ACKNOWLEDGMENTS
Hsuan-Lei Shao, “From Knowledge Genealogy to Knowledge Map-
China Studies in Big Data and Machine Learning” (107-2410-H-003-
058-MY3, Ministry of Science and Technology), Taiwan.

Sieh-Chuen Huang, “Exploring Implications of Legal Analytics

along with AI Technologies” (108-2628-H-002-005-MY2, Ministry
The root node of the tree—the feature that best divides the data—is of Science and Technology), Taiwan.
Caregiver, indicating that judges in Taiwan primarily consider who

259
 Sentence Classification for Contract Law Cases: A Natural
Language Processing Approach
Jonathan R. Mok Wai Yin Mok
Norris Injury Lawyers PC University of Alabama in Huntsville
jonrexmok@gmail.com mokw@uah.edu
1 INTRODUCTION
Legal practice can gain much from advances in machine learning
and A.I. technology. Common law countries such as the United
States, the United Kingdom, Canada, and Australia, rely on judicial
precedent to decide what law governs new cases and how that law
is applied. Judicial precedent is the body of case law that observes
how past courts have decided similar cases. Case law is composed of
individual case decisions by courts that resolve a particular dispute.
When resolving new cases, courts look back on judicial precedent to
structure how to decide said new cases. Thus, an extensive knowl-
edge base of how prior cases have been decided under what law is
needed in each field of legal practice. Legal research is the practice
of identifying relevant case law for a particular set of facts and is
essential for any practitioner seeking to make the most persuasive
argument for clients. Reviewing case law and identifying relevant
cases can be time consuming and imperfect; there is always risk
that practitioners overlook relevant and/or new case law that has
changed the legal landscape of their practice. Thus practitioners are
not only required to retain an extensive knowledge base of relevant
case law, but also to keep abreast of changes in case law.
Machine learning and A.I. technology has potential to assist
practitioners and the judiciary by automating legal research and
analysis for new cases. A fully realized system would streamline
legal workflow and may be able to analyze inputs raised in new
cases, infer legal issues, identify relevant case law, draft arguments
for and against each issue, consider jurisdictional variations, and
generate a predictive outcome with a corresponding percent cer-
tainty. Such a system could not only be used by practitioners to
assist new case analysis, but also judges to assist them to resolve
pending cases without present bias.
However, there are tremendous barriers to developing such a
system. Firstly, not all case law is relevant to every new case. The
practice of law is subcategorized into many different fields of law
that govern each type of dispute. In the United States, the first
major difference in law arises when distinguishing whether a case
is governed by civil law or criminal law. Civil law governs disputes
among private parties that mainly deal with monetary damages;
however, the government may be involved in civil suits that it is
not immune from. Criminal law governs conduct deemed harmful
to the state or society that is prosecuted by the government and
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466074
260
Rachel V. Mok
Independent Researcher
rmok57@gmail.com
can result in the loss of personal liberties, such as imprisonment,
and/or fines. Civil law case law is considered different to criminal
law case law with little overlap. Moreover, civil law and criminal
law are subcategorized into further specialized fields of law; civil
law includes contract law, property law, tort law, family law, em-
ployment law, patent law, etc. and criminal law includes case law
on crimes against a person’s body and/or mind, crimes against prop-
erty, crimes that violate a statute, crimes of financial fraud and/or
impropriety, etc. Each field of law listed is based on a body of case
law that is considered different to any other with little overlap and
can involve years of practice to become intimately familiar with.
That said, a universal system that could analyze any type of case
brought before it would need to rely upon a vast knowledge base
of case law, be able to distinguish relevant facts from a new case,
and pair them with relevant case law to even begin an analysis, let
alone complete the remaining steps in a typical legal workflow.
This paper proposes to take tentative steps towards a self-learning
system for legal research with initial steps being to classify the in-
tended knowledge base, case law, into machine learnable informa-
tion. This involves classifying individual court opinions into legally
useful sentence types, a common practice among law students and
practitioners alike. This paper will focus on case law in the United
States, specifically contract law case law in the state of Alabama,
and may explore other fields of law in later research.
Translating case law into machine learnable information requires
understanding case law first. Case law is composed of individual
court opinions written by judges. A single case focuses on a par-
ticular set of facts that gave rise to a dispute(s). Each case involves
at least two parties, but can involve more. Case law generally fol-
lows patterns to how each opinion is written. An opinion can be
classified into five main types of sentences for purposes of legal
understanding: fact sentences, issue sentences, holding sentences,
reasoning sentences, and law sentences.
Fact sentences identify relevant facts to the case such as, but
not limited to, party actions leading up to the dispute; background
information such as practices, scientific principles, etc. needed to
understand the dispute; and damages sustained as a result of the
dispute. Bear in mind any case can have a myriad of facts surround-
ing it; the facts that are included in a judicial opinion are those
facts as understood by the court and deemed relevant to resolve
the dispute.
Issue sentences frame the nature of the dispute and can be further
classified into two types: court issue sentences and party issue
sentences. Court issue sentences are statements made by the court
itself that define the nature of the dispute; a court emphasizing
clarity may even have headers identifying court issue sentences, but
this is not expected in every judicial opinion. Party issue sentences
are statements made by each party attempting to frame the dispute
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
in their terms and are reiterated by the court; these sentences will
typically begin by stating ABC party argues XYZ point. Party issue
sentences are less significant than court issue sentences, unless the
court explicitly agrees with a particular party issue sentence, which
at that point elevates it to become a court determination known as
a court holding.
A holding sentence states a court conclusion that resolves a
dispute by applying relevant law to the present set of facts in a
case. A holding sentence is the most important type of sentence in
any judicial opinion; it is considered precedent and binding on any
future case brought within the issuing court’s jurisdiction.
Reasoning sentences help explain how a court has reached its
conclusion; it is analogous to a mathematical proof, but not nearly
as exact. Reasoning sentences can be difficult to definitively classify;
seasoned practitioners can mistake reasoning sentences for holding
sentences and vice versa. Any court determination not based on the
present facts of the case is not considered a holding sentence, but
likely a reasoning sentence; any hypothetical statement, such as an
example or analogy, made by the court is not a holding sentence,
but likely a reasoning sentence; and any commentary on law or
fact not regarding the present case, such as on a past case, is not a
holding sentence, but likely a reasoning sentence.
Law sentences are statements of law made by the court and are
always preceded or followed by a citation, or the citation may be
in the sentence itself. A law sentence may be the court restating a
law as it appears in the code of law, the court restating a holding
from a prior case, or the court stating its own interpretation of a
law or a prior court’s holding.
Another type of sentence found in court opinions are proce-
dural history sentences, which describe how a case has progressed
through the court system to be evaluated before the present court.
Almost all cases found in case law begin in a court that makes both
factual and legal determinations such as which party’s account
of the dispute is considered correct and what legal principles are
applied to resolve the case; such a court is called a court of first
instance or trial court. All losing parties in civil and criminal mat-
ters, bar the criminal prosecution, have a right to appeal to a higher
court called an appellate court for additional review of mainly legal
principles. Appellate courts rarely engage in fact finding and do
so only under specific circumstances, such as when new relevant
evidence arises in a case and is properly presented. In actuality, the
majority of case law is written by appellate courts. Consequently,
appellate courts will describe in procedural history sentences how
a case was decided by a court of first instance and which party
appealed the case on what grounds to bring the case before it. Such
information is of some value to practitioners, but can essentially be
disregarded when forming a knowledge base for machine learnable
information for the purposes of this paper.
Thus far, sentence classification according to fact sentences, is-
sue sentences further classified by court or party, law sentences,
reasoning sentences, holding sentences, and procedural history
sentences can be applied to every court opinion. Every court opin-
ion also contains references that are valuable to practitioners, but
can essentially be disregarded when forming a knowledge base
for machine learnable information for the purposes of this paper.
Citations are classified as reference sentences in this paper.
261
Mok et al.
As previously stated, this paper will explore classification of
legally useful sentence types found in contract law case law in
the state of Alabama. Due to the difficulty of classifying reasoning
sentences at this stage of the research, such classification has been
excluded, but will be examined in later research. As a result, this
paper will only focus on seven sentence types: Fact (FCT), Court
Issue (CTI), Party Issue (PTI), Holding (HLD), Law (LAW), Procedural
History (PRH), and Reference (REF).
2 KNOWLEDGE BASE
(a) Testing cases, 1063 total sen- (b) Training cases, 646 total sen-
tences. tences.
Figure 1: Confusion matrices of the testing and training cases.
The elements along the diagonal show the number of sen-
tences that were identified correctly, and each off-diagonal
element show the number of sentences that were identified
incorrectly and in what label they were identified as.
spaCy [1] has been chosen to parse and process court cases.
Twelve Alabama contract cases were downloaded from JUSTIA, of
which five are designated as training cases and seven as testing
cases. Two different approaches are adopted for classifying the
sentences: a rule-based approach for the reference sentences and
a knowledge-based approach for the other types of sentences. Be-
cause fact sentences are the default, the knowledge base contains
fragments of sample sentences that are neither reference nor fact
sentences. The fragments of the sample sentences are chosen to rep-
resent the essential parts of the sample sentences from which they
were extracted, but are general enough so that sentences whose
types are to be determined may contain similar fragments. The sim-
ilarity of a sentence and the sentence fragments in the knowledge
base is calculated by the spaCy’s similarity function, which is based
on word vectors [2, 3]. Shown in the above figure are the results of
our algorithm. Although our algorithm is still in an early stage of
development, it yields a classification accuracy rate of 68.67% on
the testing cases, demonstrating the validity of our approach.
REFERENCES
[1] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. 2020.
spaCy: Industrial-strength Natural Language Processing in Python. https://doi.org/
10.5281/zenodo.1212303
[2] Tomás Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient
Estimation of Word Representations in Vector Space. In 1st International Con-
ference on Learning Representations, ICLR 2013, Scottsdale, Arizona, USA, May
2-4, 2013, Workshop Track Proceedings, Yoshua Bengio and Yann LeCun (Eds.).
http://arxiv.org/abs/1301.3781
[3] Tomás Mikolov, Ilya Sutskever, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013.
Distributed Representations of Words and Phrases and their Compositionality.
CoRR abs/1310.4546 (2013). arXiv:1310.4546 http://arxiv.org/abs/1310.4546
 Constraint Answer Set Programming as a Tool to Improve
Legislative Drafting
A Rules as Code Experiment
Jason Morris
jmorris@smu.edu.sg
Singapore Management University Centre for Computational Law
Singapore
CCS CONCEPTS
• Computer systems organization → Embedded systems; Re-
dundancy; Robotics; • Networks → Network reliability.
KEYWORDS
legal knowledge representation and reasoning, constraint answer
set programming, rules as code
ACM Reference Format:
Jason Morris. 2021. Constraint Answer Set Programming as a Tool to Im-
prove Legislative Drafting: A Rules as Code Experiment. In Eighteenth
International Conference for Artificial Intelligence and Law (ICAIL’21), June
21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 2 pages. https:
//doi.org/10.1145/3462757.3466084
1 RULES AS CODE
"Rules as Code" in this paper is used to refer to a proposed method-
ology of legislative and regulatory drafting.1 That legislation can
be represented in declarative code for automation has long been
recognized [6], as has the opportunity for improving the quality of
legal drafting with the techniques of formal representation [1].
Rules as Code further proposes that both drafting and automation
would be improved by initially co-drafting statute law in both
natural and computer languages simultaneously [4].
Knowledge acquisition bottlenecks and roadblocks associated
with statutory interpretation are largely avoided. The co-drafted
encoding need only reflect what the legislation says, and not what
the legislators meant. Legislative intent is instead encoded as tests
by people with authoritative knowledge of the intent, the drafters.
In this way, failed tests can be used in the drafting process to signal
issues with the natural language draft. When the drafting process is
complete an authoritative encoding consistent with the legislative
intent already exist. This encoding can be used by regulators and
regulated entities to automate services and compliance tasks.
1 The phrase "Rules as Code" is also often used to refer to legal knowledge representation
and reasoning generally.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466084
262
2 S(CASP)
s(CASP) is a stable-model constraint answer set programming lan-
guage, implemented in the Ciao programming language [3]. s(CASP)
was selected for use in this experiment because of its ability to gen-
erate natural language explanations for answer sets [2], and its
ability to perform both deductive and abductive reasoning from
the same encoding with minor adjustments [3]. Justifications for
automated conclusions have long been recognized as useful in legal
applications both for end-users and as a development tool [5].
3 RULE 34, LEGAL PROFESSION
(PROFESSIONAL CONDUCT) RULES OF
SINGAPORE, 2015
Singapore’s Legal Profession Act, (Cap. 161) (the “Act”) governs the
legal profession in Singapore. Part VI of the Act establishes the
Professional Conduct Council, which in section 71(2) of the Act
is given broad authority for drafting rules governing the practice,
conduct, etiquette, and discipline of legal practitioners in Singapore.
The Professional Conduct Council has enacted the Legal Profession
(Professional Conduct) Rules (S 706/2015) ("the Rules"). In 2015 the
Rules were significantly amended with a new Rule 34 setting out
restrictions on lawyers accepting executive appointments outside
of their legal practice.
4 EXPERIMENTAL DESIGN
Our interdisciplinary team undertook to assess the strengths and
weaknesses of s(CASP) as a tool for improving legislative drafting
in a Rules as Code approach. The author encoded a literal interpre-
tation of Rule 34 in s(CASP), and separately encoded the author’s
expectations of the behaviour of Rule 34 as a set of tests. Test fail-
ures that the author attributed only or primarily to issues with
the natural language drafting of Rule 34 were raised with legally-
trained team members to confirm whether the expected behaviour
was reasonable, and whether the cause of the test failure was a legal
drafting issue.
The discovery of such issues would demonstrate the feasibility of
using s(CASP) to detect legislative drafting issues. Any issues aris-
ing would be in the context of non-authoritative opinions as to the
expected behaviour of the text. The experiment cannot, therefore,
be used to diagnose issues with the Rule as enacted.
5 EXPERIMENTAL RESULTS
The code written to implement this experiment is available at https:
//github.com/smucclaw/r34_scasp.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
A set of 25 tests were encoded, and there were 4 test failures
not explained by errors encoding the Rule or the tests. These four
failures were investigated by the author by performing ’why not’
queries and reviewing the justifications provided by s(CASP).
This process revealed that the failing tests were encoded on the
basis of an expectation that the word "business" in Rule 34(1)(b)
referred to a legal practitioner’s activities. But Rule 34(9) defines
"business" to refer to a general category of undertaking. Setting out
a test in which Rule 34(1)(b) applied, while also using the defined
meaning of "business", required making statements that did not
have clearly meaningful real world equivalents. This suggested
that Rule 34(1)(b) might also use the word "business" in a way
inconsistent with the defined meaning, which would be a drafting
issue.
That issue was raised with the rest of the research team, who
confirmed that Rule 34(1)(b) had been faithfully encoded, that the
expectations of the failing tests were reasonable, and that Rule
34(1)(b) required the use of an interpretations of the word "business"
that is inconsistent with the defined meaning of the word in order
to give effect to that expectation, or to give it any clear meaning at
all.
The research team seriously considered the possibility that there
might be a different interpretation of other aspects of the Rule that
would make Rule 34(1)(b) more clearly meaningful. The team was
unable to find an interpretation that would have had that effect and
would not also make Rule 34(1)(b) redundant to other portions of
the Rule. The research team therefore concluded that it would be
more correct if Rule 34(1)(b) referred not to businesses but to the
holding of an executive appointment.
The researchers agreed on the following proposed replacement
for Rule 34(1)(b):
(1A) A legal practitioner must not accept any execu-
tive appointment that materially interferes with —
(i) the legal practitioner’s primary occupation of
practising as a lawyer;
(ii) the legal practitioner’s availability to those
who may seek the legal practitioner’s services as
a lawyer; or
(iii) the representation of the legal practitioner’s
clients.
The proposed amendment was encoded, and the tests re-run. All
25 tests passed.
6 CONCLUSIONS
Our experiment demonstrates the use of the Rules as Code method-
ology to detect a drafting issue in a proposed statutory text, and to
verify the effect of a proposed amendment. The issue discovered in
this experiment is the type of issue that Rules as Code is intended
to address early: one that if left unaddressed negatively affects the
degree to which the statutory text can be automated.
With regard to s(CASP)’s strengths and weaknesses for this task,
the access to "why not" queries and natural language justifications
was extremely valuable both in the encoding of the Rule, and in
the analysis of test failures. s(CASP)’s abductive reasoning over
constraints, and the fact that it returned answer sets rather than
bindings, allowed the author to test the encoding against a wide
263
Jason Morris
variety of fact scenarios simultaneously, quickly providing a deep
level of insight into the behaviour of the encoding. s(CASP) also
facilitated the use of a version of defeasibility that allowed defeating
relations of both the "subject to" and "despite" types to be encoded
where they appear in the text, enhancing maintainability of the
code [7].
s(CASP)’s abductive queries slow down considerably with the
complexity of the code, and so it may not be an appropriate ap-
proach for real-time applications of abductive reasoning. However,
its performance on deductive reasoning tasks was very efficient,
completing the 25 tests in this experiment in an average of less
than 1 second each, which suggests it can also be used to answer
legal questions with complicated fact scenarios and complicated
rules in a user-facing application.
ACKNOWLEDGMENTS
I owe a debt of gratitude to all my colleagues at the SMU Centre
for Computational Law, and in particular our Principal Investigator
Meng Weng Wong, Industry Director Alexis Chun, and Professors
Lim How Khang and Jerrold Soh, all of whom contributed greatly to
the legal analysis. Professors Gopal Gupta of University of Texas at
Dallas and Joaquín Arias at Universidad Rey Juan Carlos provided
valuable assistance on the effective use of s(CASP). The feedback
of the reviewers has also improved the paper and is gratefully
acknowledged.
This research is supported by the National Research Founda-
tion (NRF), Singapore, under its Industry Alignment Fund – Pre-
Positioning Programme, as the Research Programme in Computa-
tional Law. Any opinions, findings and conclusions or recommen-
dations expressed in this material are those of the author(s) and do
not reflect the views of National Research Foundation, Singapore.
REFERENCES
[1] L. Allen and C. R. Engholm. 1978. Normalized Legal Drafting and the Query
Method. Journal of Legal Education 29 (1978), 380–412.
[2] Joaquín Arias, Manuel Carro, Zhuo Chen, and Gopal Gupta. 2020. Justifica-
tions for Goal-Directed Constraint Answer Set Programming. arXiv preprint
arXiv:2009.10238 (2020).
[3] Joaquin Arias, Manuel Carro, Elmer Salazar, Kyle Marple, and Gopal Gupta. 2018.
Constraint answer set programming without grounding. Theory and Practice of
Logic Programming 18, 3-4 (2018), 337–354.
[4] Organization for Economic Cooperation and Development Observatory for Public
Sector Innovation. [n.d.]. Cracking the Code: Rulemaking for humans and ma-
chines. Accessed February 28, 2021, at https://oecd-opsi.org/wp-content/uploads/
2020/10/Rules-as-Code_Highlights_Final_HighRes.pdf.
[5] D. Merritt. 2017. Expert Systems in Prolog. Independently Published. https:
//books.google.com.sg/books?id=6IQGyQEACAAJ
[6] Marek J. Sergot, Fariba Sadri, Robert A. Kowalski, Frank Kriwaczek, Peter Ham-
mond, and H. Terese Cory. 1986. The British Nationality Act as a logic program.
Commun. ACM 29, 5 (1986), 370–386.
[7] Hui Wan, Benjamin Grosof, Michael Kifer, Paul Fodor, and Senlin Liang. 2009. Logic
Programming with Defaults and Argumentation Theories. In Logic Programming,
Patricia M. Hill and David S. Warren (Eds.). Springer Berlin Heidelberg, Berlin,
Heidelberg, 432–448.
 Predicting Legal Proceedings Status: Approaches Based on
Sequential Text Data
Felipe Maia Polo Itamar Ciochetti
felipemaiapolo@gmail.com itamar@tikal.tech
University of São Paulo, Brazil Tikal Tech, Brazil
Advanced Institute for AI (AI2), Brazil
ACM Reference Format:
Felipe Maia Polo, Itamar Ciochetti, and Emerson Bertolo. 2021. Predict-
ing Legal Proceedings Status: Approaches Based on Sequential Text Data.
In Eighteenth International Conference for Artificial Intelligence and Law
(ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA,
2 pages. https://doi.org/10.1145/3462757.3466138
1 OBJECTIVE AND PRACTICAL
IMPORTANCE OF THIS WORK
The objective of this work, which is fully given by Polo et al. [8], is
to develop predictive models to classify Brazilian legal proceedings
in three possible classes of legal status: (i) archived proceedings, (ii)
active proceedings, and (iii) suspended proceedings. Each proceed-
ing is made up of a chronological sequence of short texts called
“motions” written by the courts’ administrative staff. The motions
relate to the proceedings, but not necessarily to their legal status.
Moreover, the proceedings’ labels are decided by the courts to orga-
nize their workflow. This problem’s resolution is intended to assist
public and private institutions in managing large portfolios of legal
proceedings, providing gains in scale and efficiency.
2 RELATED WORK
Despite researchers’ efforts to create applications in the legal field,
we were unable to find an attempt to solve a problem like ours in
the literature. The issues closest to ours we could relate in litera-
ture are those of identifying the parties in legal proceedings [7],
classification of legal documents according to their administrative
labels [2] or predicting the area a proceeding belongs to [11]. This
paper has a different application that can be useful when looking
for efficiency in legal systems, especially in developing countries.
Unlike previous work, we consider sequences of texts explicitly
in our modeling, which has not yet been observed in Law and AI
literature by us.
3 DATA
Our data is composed of two datasets: a dataset of 3 · 106 unlabeled
motions and a dataset containing 6449 legal proceedings, each
with an individual and a variable number of motions, but which
have been labeled by lawyers. Among the labeled data, 47.14% is
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466138
264
Emerson Bertolo
emerson@tikal.tech
Tikal Tech, Brazil
classified as archived (class 1), 45.23% is classified as active (class 2),
and 7.63% is classified as suspended (class 3). The datasets we use
are representative samples from the first and third most significant
Brazilian state courts (São Paulo and Rio de Janeiro).
In this work, we split at random our labeled dataset into three
parts: training set (70%) for training models, validation set (10%)
for hyperparameter tuning, and test set (20%) for final assessment.
4 METHODOLOGY
We used four approaches to extract features from the legal texts
and three base classifiers to create our predictive models to classify
legal proceedings, i.e., text sequences. A more detailed explanation
of our methodology can be found in Polo et al. [8].
4.1 Classifiers
The first classifier we use is a many-to-one long short-term memory
network (LSTM). The inputs are given by 𝑇 vectors representing
the 𝑇 most recent texts in chronological order. The outputs will be
predicted probabilities for each of the three classes, returned by the
Softmax function. The second classifier is a multilayer perceptron
neural network (MLP) with one hidden layer and ReLU activation
functions. The MLP input is the concatenation of the feature vectors
of the last 𝑇 texts. The third classifier is given by a XGBoost [1] tree
ensemble. We feed the last classifier with same inputs used for the
MLP. All classifiers make classification choosing the most probable
class and we fix 𝑇 = 5, using zero-padding vectors when necessary.
Details of the hyperparameter tuning phase can be found in Polo
et al. [8].
4.2 Feature Extraction
We use four different approaches to extract features from texts:
Word2Vec (W2V) [5], Doc2Vec (D2V/PV-DM) [4], TFIDF [9], and a
Brazilian Portuguese BERT-Base [10]. All of them are completely
unsupervised or self-supervised methods. For all approaches, we
had text preprocessing and hyperparameter setting steps, which
are detailed in Polo et al. [8]. Word2Vec, Doc2Vec, and TFIDF repre-
sentations are fully trained using the mass of 3 · 106 texts/motions
from unlabeled proceedings, while the BERT-Base is fine-tuned
in the same dataset, making use of the Masked Language Model
(MLM) objective. One thing that is worth mentioning is that we
use a method proposed by Mikolov et al. [6] in order to identify
presence words (2 to 4) that should be considered as unique tokens
when working with W2V, D2V, and TFIDF.
Given Word2Vec creates representations for tokens and not for
entire texts, we use two different approaches to that end. One
of them is used in conjunction with the LSTM classifier, and the
other is used in conjunction with the MLP and XGBoost classifiers.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Felipe Maia Polo, Itamar Ciochetti, and Emerson Bertolo

Macro averaging Weighted averaging

Classifier Feature extraction Accuracy F1 Score Precision Recall F1 Score Precision Recall
W2V (using CNN) 0.93 ± 0.01 0.88 ± 0.01 0.92 ± 0.01 0.85 ± 0.02 0.92 ± 0.01 0.93 ± 0.01 0.93 ± 0.01
Doc2Vec 0.82 ± 0.01 0.76 ± 0.02 0.77 ± 0.02 0.75 ± 0.02 0.82 ± 0.01 0.82 ± 0.01 0.82 ± 0.01
LSTM
TFIDF 0.90 ± 0.01 0.85 ± 0.01 0.85 ± 0.01 0.85 ± 0.02 0.90 ± 0.01 0.90 ± 0.01 0.90 ± 0.01
BERT 0.93 ± 0.01 0.89 ± 0.01 0.92 ± 0.01 0.87 ± 0.02 0.93 ± 0.01 0.93 ± 0.01 0.93 ± 0.01
W2V 0.92 ± 0.01 0.87 ± 0.01 0.92 ± 0.01 0.84 ± 0.02 0.92 ± 0.01 0.92 ± 0.01 0.92 ± 0.01
Doc2Vec 0.87 ± 0.01 0.83 ± 0.01 0.89 ± 0.01 0.79 ± 0.02 0.87 ± 0.01 0.88 ± 0.01 0.87 ± 0.01
XGBoost
TFIDF 0.92 ± 0.01 0.88 ± 0.01 0.93 ± 0.01 0.84 ± 0.02 0.92 ± 0.01 0.93 ± 0.01 0.92 ± 0.01
BERT 0.92 ± 0.01 0.86 ± 0.01 0.92 ± 0.01 0.83 ± 0.02 0.92 ± 0.01 0.92 ± 0.01 0.92 ± 0.01

Table 1: Evaluation of classification approaches (scores ± bootstrap std. errors). We combine three basic classifiers (LSTM, MLP,
and XGBoost) and four approaches for extracting features (Word2Vec, Doc2Vec, TFIDF, and BERT). In this extended abstract,
we omit MLP’s results since they are not better than LSTM’s and XGBoost’s.

When using LSTM networks as classifiers, we first create a text

matrix, each row being given by a token embedding. Secondly,
we apply 𝐾 convolutional filters [3], given by one-dimensional
matrices, to extract the desirable information from the texts. The
filters are trained in conjunction with the LSTM weights, while
embeddings are frozen. Due to the last detail, we also refer to the
W2V/LSTM classification approach as W2V/CNN/LSTM, where the
"CNN" stands for convolutional neural networks. On the other hand,
to represent texts when working with MLP and XGBoost classifiers,
we use the average vector of each text’s tokens embeddings.
5 CLASSIFICATION RESULTS
Given that all classifiers make classification choosing the most prob-
able class, we now compare their performance according to a few
key metrics presented in Table 1. In this extended abstract, we omit
MLP’s results since they are not better than LSTM’s and XGBoost’s.
In general, our classifiers have an excellent performance; however,
the Doc2Vec approaches performed relatively worse than the oth-
ers. A pertinent point that can be observed in the "Recall" and "F1
Score" columns of the macro averages is that our classifiers perform
relatively worse when finding examples from the "Suspended" mi-
nority class, which add up to about 7.63%. This is not necessarily a
problem though.
6 INTERPRETABILITY
We present interpretable insights of the W2V/CNN/LSTM approach.
To better understand the patterns extracted by the neural network’s
convolutional layer, let us look at the embedding representations
of tokens in our vocabulary that have the closest representations to
the filters according to cosine similarity. In this extended abstract,
we focus on two specific filters (6 and 7) from a total of nine. Table
2 shows which tokens (translated from Portuguese to English) most
closely resemble our filters after they are learned.
One can see that the patterns sought by the neural network do
have to do with the classifications we want to make, especially
when looking at filters 6 and 7. For example, the expressions "final
storage of docket" and "final remittance to origin" indicate archiving
of proceedings (class 1), and the expression "temporarily stored
docket" may indicate suspension (class 3). Polo et al. [8] goes deeper
in the analysis and presents a method to connect patterns in Table
2 with the classifier’s output.
Filters Tokens cos(𝜃 )
"final storage of docket" 0.46
6 "final remittance to origin" 0.45
"remittance to origin" 0.42
"final storage of docket" 0.47
"temporarily stored docket" 0.43
7
"final remittance to origin" 0.42
Table 2: Similarity between filters and their most similar to-
kens. It is possible to check what kind of information the
filters seek in a text excerpt.
REFERENCES
[1] Tianqi Chen and Carlos Guestrin. 2016. Xgboost: A scalable tree boosting system.
In Proceedings of the 22nd acm sigkdd international conference on knowledge
discovery and data mining. 785–794.
[2] N Correia da Silva, FA Braz, DB Gusmão, FB Chaves, DB Mendes, DA Bezerra,
GG Ziegler, LH Horinouchi, MHP Ferreira, PHG Inazawam, et al. 2018. Docu-
ment type classification for Brazil’s supreme court using a Convolutional Neural
Network. (2018).
[3] Yoon Kim. 2014. Convolutional neural networks for sentence classification. arXiv
preprint arXiv:1408.5882 (2014).
[4] Quoc Le and Tomas Mikolov. 2014. Distributed representations of sentences and
documents. In International conference on machine learning. 1188–1196.
[5] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient
estimation of word representations in vector space. arXiv preprint arXiv:1301.3781
(2013).
[6] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013.
Distributed representations of words and phrases and their compositionality. In
Advances in neural information processing systems. 3111–3119.
[7] Truong-Son Nguyen, Le-Minh Nguyen, Satoshi Tojo, Ken Satoh, and Akira Shi-
mazu. 2018. Recurrent neural network-based models for recognizing requisite
and effectuation parts in legal texts. Artificial Intelligence and Law 26, 2 (2018),
169–199.
[8] Felipe Maia Polo, Itamar Ciochetti, and Emerson Bertolo. 2021. Predict-
ing Legal Proceedings Status: Approaches Based on Sequential Text Data.
arXiv:2003.11561 [cs.CL]
[9] Gerard Salton and Michael J McGill. 1986. Introduction to modern information
retrieval. (1986).
[10] Fabio Souza, Rodrigo Nogueira, and Roberto Lotufo. 2019. Portuguese Named
Entity Recognition using BERT-CRF. arXiv preprint arXiv:1909.10649 (2019).
http://arxiv.org/abs/1909.10649
[11] Octavia-Maria Sulea, Marcos Zampieri, Shervin Malmasi, Mihaela Vela, Liviu P
Dinu, and Josef Van Genabith. 2017. Exploring the use of text classification in
the legal domain. arXiv preprint arXiv:1710.09306 (2017).

265
 Pathways to Legal Dynamics in Robotics
Antonino Rotolo Luciano H. Tamargo
University of Bologna Universidad Nacional del Sur
Bologna, Italy Bahia Blanca, Argentina
antonino.rotolo@unibo.it lt@cs.uns.edu.ar
ACM Reference Format:
Antonino Rotolo, Luciano H. Tamargo, and Diego C. Martínez. 2021. Path-
ways to Legal Dynamics in Robotics. In Eighteenth International Conference
for Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo,
Brazil. ACM, New York, NY, USA, 2 pages. https://doi.org/10.1145/3462757.
3466146
1 RESEARCH CHALLENGES
Normative concepts can play a crucial role in modelling the be-
haviour and interaction of artificial agents. Investigations are still
relatively underdeveloped in robotics while interesting ideas come
from related fields, such as multi-agent systems (MAS).
We outlines some research challenges in this domain for which
we can use existing models of legal change developed in AI&Law.
Three challenges for MAS can be adapted for robotics [1]:
Challenge 1. Explain which of the following choices should be
made in robotics: (a) norms must be explicitly represented in robots
in a declarative way, or (b) norms must be explicitly represented in
the overall system specification.
Option (a) must be preferred if we should avoid trivialising the
notion of norm, which is a risk when we see any specification
requirement as a norm that the system has to comply with [1]. In
addition, since legal norms change, maintenance would be probably
easier. However, (b) is more suitable to address this problem:
Problem 1. How we can check whether a robot complies with
norms applicable to it? How can we design a robot such that it complies
with a given set of norms?
Addressing Challenge 1 requires the preliminary clarification of
norm features that we need to embed within robots. In particular,
temporal aspects are especially relevant for legal dynamics [3],
since legal norms can be qualified by temporal properties, such as:
(1) the time when the norm comes into existence and belongs to
the legal system, (2) the time when the norm is in force, (3) the time
when the norm produces legal effects (it is applicable), and (4) the
time when the normative effects hold.
A norm is a kind of system constraint. While hard constraints
are restricted to preventive control systems in which violations
are impossible (we call this mechanism regimentation), soft con-
straints are used in detective control systems where violations can
be detected (we call this mechanism regulation). This justifies the
following challenge:
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466146
266
Diego C. Martínez
Universidad Nacional del Sur
Bahia Blanca, Argentina
dcm@cs.uns.edu.ar
Challenge 2. Make explicit why your norms are a kind of (soft)
constraint that deserve special analysis.
With hard constraints the problem is how to reconfigure robots’
behaviour in presence of norm change. If the norms are represented
as soft constraints, then the problem is to check if the process of
monitoring violations is correctly managed. For example, it may be
the case that violations are not detected often enough.
Whatever model we adopt for legal norms on robotics, we need
a formal model handling norm change:
Challenge 3. Why and how can norms be changed at runtime?
Many legal issues can be raised in regard to robots [2]. Example
1 illustrates how norm change can impact on robotics.
Example 1. The Italian penal code states the following:
Art. 111 Italian of Penal Code – Procuring a person for an
offence who is not indictable or not punishable. Anyone
who has procured a person for a criminal offence who is not
indictable or is not punishable on the basis of a personal condi-
tion or quality is liable for that offence which was committed
by this person, and an increased penalty is applied.
Imagine Mr. Smith induces a robot to threaten Mr. Jones, and the
robot is bound to that goal (to threat Mr. Jones) but is equipped with
autonomy in achieving it. Can we apply art. 111 to this case? It should
be noted that the provision cover cases where the procured person
is not legally capable (she is not in full possession of her faculties)
and this makes more serious the offence committed by the procuring
person. However, robots, though intelligent, are not indictable and the
principle of legality in criminal law does not allow the provision to be
applied by analogy when the crime is committed by a robot.
Suppose that the legislator enacts at the 1st of January 2005 a new
version of art. 111 (denoted as ‘Art.111-n’.):
Art. 111 [Amended] Italian of Penal Code – Procuring a person—
who is not indictable or not punishable—or an intelligent
machine for an offence. Anyone who has procured a person—
who is not indictable or is not punishable on the basis of a
personal condition or quality—or an intelligent machine for a
criminal offence is liable for that offence which was committed
by this person or machine, and an increased penalty is applied.
The AI&law community proposed several frameworks for norm
change, two of them focused on temporal models: one tem-
poralised rule-based system [3] and one extending belief revi-
sion techniques [5]. Both view a legal system as a time-series
LS(t 1 ), LS(t 2 ), . . . , LS(t j ) of its versions, where each version is ob-
tained from previous ones through entering new norms, or by mod-
ification or repeal of existing norms: each LS(ti ) is the snapshot of
the norms in the legal system at time ti .
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
2 NORM CHANGE IN ROBOTICS
Temporal models are crucial to dynamically regiment robots by
design or to activate countermeasures for compensating norm vi-
olations. We illustrate how the temporal machinery conceptually
operates by using the specification of the MAS proposed by [4] for
sociotechnical systems and adapt it to a robot interacting with hu-
mans. Consider Figure 1. For the sake of simplicity, we can imagine
Figure 1: Static normative interaction
that a robot and a human interact in a given environment at time
t. The legal system at t is denoted by LS(t). The technical layer
is composed of software components supporting robot’s actions.
Norms in LS(t) can govern robot-human interactions in two ways:
(1) by regimentation of robot’s behaviour via access control to soft-
ware components, thus preventing specific robot’s actions, (2) by
regulation, in such a way that norms can be violated but we check
accountability and possibly on resort to sanctions.
An interesting case is how legal dynamics impact on the above
framework. Figure 2 depicts the case in Example 1.
Figure 2: Example 1: static
Assume to operate in 2005, so the version in force of art. 111 of the
Italian of Penal Code is the amended one in Example 1 and denoted
by Art.111 − n, where intelligent machines are also considered.
Suppose now we are in 2007 and 1) Mr. Smith induces twice the
robot to threaten Mr. Jones, a first time in 2005 and a second time in
2007; 2) Art.111 − n is abrogated in 2007 and Art.111 is reinstated.
This is depicted in Figure 3. This case simply embeds legal dynamics
within the scenario of Figure 2, since abrogations are not retroactive
[3]. We are in 2007 and norm Art.111 − n applies in 2007 to the case
of Mr. Smith inducing in 2005 the robot to threaten Mr. Jones (by
the regulation mechanism because we are in 2007). If Art.111 − n
would not be abrogated, then we would apply in 2007 to the case of
Mr. Smith inducing in 2007 the robot to threaten Mr. Jones (either
267
Antonino Rotolo, Luciano H. Tamargo, and Diego C. Martínez
Figure 3: Example 1: dynamic non-retroactive
by the regulation mechanism or by preventing it via regimentation);
however, this is not the case because the norm is abrogated. Of
course, the initial version of the norm. i.e., Art.111 does not cover
intelligent machines and so the case of Mr. Smith inducing in 2007
the robot to threaten Mr. Jones is not relevant.
Suppose now we are in 2007 and 1) Mr. Smith induces twice the
robot to threaten Mr. Jones, a first time in 2005 and a second time
in 2007; 2) Art.111 − n is annulled in 2007 (thus repealing the norm
since 2005) and Art.111 is reinstated. This is depicted in Figure 4.
Figure 4: Example 1: dynamic retroactive
This case is different, since annulments are retroactive [3]. We
are in 2007 and norm Art.111 − n no longer exists as if it were never
enacted, so it does not cover the case of Mr. Smith inducing in 2005
and 2007 the robot to threaten Mr. Jones. (Unless Mr. Smith’s action
was prevented by regimentation in 2005; in this case, we do not
have an issue regarding it.) Finally, Art.111 does not cover the case
of Mr. Smith inducing in 2007 the robot to threaten Mr. Jones.
REFERENCES
[1] T. Balke et al. Norms in MAS: Definitions and Related Concepts. In Normative Multi-
Agent Systems. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik, Dagstuhl,
Germany, 2013.
[2] M. Corrales, M. Fenwick, and N. Forgó. Robotics, AI and the Future of Law. Springer,
2018.
[3] G. Governatori and A. Rotolo. Changing legal systems: legal abrogations and
annulments in defeasible logic. Logic Journal of the IGPL, 18(1):157–194, 2010.
[4] Ö. Kafali, N. Ajmeri, and M. P. Singh. DESEN: specification of sociotechnical
systems via patterns of regulation and control. ACM Trans. Softw. Eng. Methodol.,
29(1):7:1–7:50, 2020.
[5] L. H. Tamargo, D. C. Martinez, A. Rotolo, and G. Governatori. An axiomatic
characterization of temporalised belief revision in the law. Artificial Intelligence
and Law, pages 1–21, 2019.
 Labels distribution matters in performance achieved in legal
judgment prediction tasks
Olivier Salaün Philippe Langlais
salaunol@iro.umontreal.ca felipe@iro.umontreal.ca
RALI, DIRO, University of Montréal RALI, DIRO, University of Montréal
Montréal, Québec, Canada Montréal, Québec, Canada
CCS CONCEPTS
• Applied computing → Law; • Computing methodologies →
Natural language processing; Neural networks.
KEYWORDS
legal judgment prediction, multilabel text classification, legal arti-
cles
ACM Reference Format:
Olivier Salaün, Philippe Langlais, and Karim Benyekhlef. 2021. Labels dis-
tribution matters in performance achieved in legal judgment prediction
tasks. In Eighteenth International Conference for Artificial Intelligence and
Law (ICAIL’21), June 21–25, 2021, Sao Paulo, Brazil. ACM, New York, NY,
USA, 2 pages. https://doi.org/10.1145/3462757.3466144
1 INTRODUCTION
In recent years, transformer [4] and BERT models [1] have been
widely used in plain NLP tasks with the assumption that models
first pretrained on massive corpora then fine-tuned on the dataset
of a given task may suffice to achieve significant improvements.
At the intersection of machine learning and law, legal judgment
prediction (LJP) is a task that aims at predicting the outcome of a
lawsuit based on a representation of the case. Such task is usually
formalized in NLP as a text classification with different classes or
labels corresponding to the verdicts. One specificity of court rulings
is that their decisions are based on the application of legal articles
to the facts described by the two parties (applicant and defendant).
In this work, we designed a LJP multilabel classification task
based on a corpus of landlord-tenant disputes in French from Que-
bec, Canada [3, 5] in which a model must predict the verdicts labels
on the basis of a truncated extract from the decision made by the
tribunal. We applied CamemBERT [2], a BERT model pretrained
on French material, in order to assess to what extent the use of a
pretrained model can handle a LJP task. We also injected article-
based input features with the hope that adding knowledge specific
to housing law domain could improve classification performance.
Although such an approach allows better results, labels distribution
must be taken into account when analyzing coarse and label-specific
scores.
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored.
For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06.
https://doi.org/10.1145/3462757.3466144
268
Karim Benyekhlef
karim.benyekhlef@umontreal.ca
Cyberjustice Laboratory, Faculty of
Law, University of Montréal
Montréal, Québec, Canada
2 PREPROCESSING OF THE DATASET
The decisions of the corpus we used here come from a court of Que-
bec in Canada that deals with all legal disputes occurring between
landlords and tenants. From 2001 to 2018, it issued 667,305 decisions
in French, some of which are freely available at SOQUIJ (Société
québécoise d’information juridique) portal1 . Decisions have a mean
and a median lengths of 307 and 235 tokens respectively while the
standard deviation is 371, indicating a high variability in length
across the documents.
A first step in preparing the dataset consisted in extracting the
text of each decision then split it in two parts thanks to syntax-
based heuristics: the pre-verdict text and the verdict. The former
contains the description of the dispute and is used as text input
for the text classification task. We also extracted the housing law
articles cited in the pre-verdict text from which we retained 445
that are cited in Book Five - Obligations in Civil Code of Quebec that
are specifically related to property lease and thus more relevant to
our task. The verdict text is further processed in order to generate
several targets labels that cover the diversity of the verdicts decided
by the judges. Thanks to regular expressions and the like, plus
some expert knowledge of housing law, we pseudo-automatically
annotated the verdicts with 23 cumulative labels. Eventually, we
excluded all decisions for which no relevant article or verdict label
was identified. All in all, the instances of the corpus amount to
544,857 documents with an average of 3.3 labels and 2 cited articles,
and are randomly split into training, validation and test sets with a
60-20-20 ratio.
3 MODELS
Within the framework of this multilabel classification, we chose
as a baseline a One-Versus-Rest Logistic Regression with the in-
put text as character-based TF-IDF vectors spanning 2-grams to
8-grams (the top 100k most frequent n-grams are kept). We also
use a CamemBERT model that we fined-tuned to our task during
10 epochs with a batch size of 32 and a learning rate of 10-5 with
the Adam optimizer and binary cross-entropy as loss function. The
maximum sequence length amounts to 128 tokens for all of our
models. Moreover, we also propose a model that leverage both
CamemBERT and cited articles by concatenating the BERT output
(vector corresponding to the [CLS] token) with a 445-dimensional
one-hot vector that embeds which articles are cited in the decision.
Then, the concatenation is sent to two fully connected layers as
shown in Figure 1
1 https://soquij.qc.ca/
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Figure 1: Architecture combining CamemBERT with a one-
hot vector for cited articles
4 RESULTS
The metrics chosen for evaluation are exact match (EM) and F1
scores. The former implies that all labels of an instance must be ex-
actly predicted by the model in order to be considered as correctly
classified. The latter is an unweighted average of the F1 scores
obtained across the 23 labels. As shown on Table 1, CamemBERT
favoured models in this task, especially when BERT output is com-
bined with articles as such an approach outperforms the sole BERT
architecture by 3.2 points and 2.2 points for F1 macro-average and
EM scores respectively. Still, such coarse results must be viewed
with precaution as they do not take into account the high imbalance
among labels.
For instance, the top three most frequent labels cover more than
half of the corpus while the sixteen least ones have a support be-
low 5%. Such biases have repercussions on the F1 scores obtained
on each label as shown on Figure 2. The F1 results obtained for
labels with a support below 5% are spread out between 0 (none of
the three least frequent labels could be correctly predicted) and
90%. Whenever the support of a label is around or above 40%, the
corresponding F1 score has a minimum value at 75%. Another obser-
vation that can be drawn is that although the addition of a one-hot
vector helps in improving BERT scores, such improvement is only
significant for certain verdict labels, suggesting that the inclusion
of domain-related knowledge is only significant in some cases. For
instance, improvement is very little or non significant for the three
most frequent labels while it seems more noticeable for those with
lower support.
5 CONCLUSION
Within the framework of a LJP task formalized as multilabel text
classification that uses a corpus in French of landlord-tenant dis-
putes, we extended a CamemBERT model with a one-hot vector of
cited articles. This led to better overall results with respect to a sole
Logistic CamemBERT
CamemBERT
regression + one-hot
F1 (macro-avg.) 53.5 58.4 61.6
Exact match 58.6 63.7 65.9
Table 1: F1 (macro-average) and exact match scores achieved
by each model on the test set.
269
Salaün et Langlais
Figure 2: F1 scores obtained for each label (x-axis is in loga-
rithmic scale) relative to label support
BERT approach. Still the results obtained must be considered with
caution. Firstly, the higher the support of a label, the more likely it
will be accurately predicted. Secondly, including articles into the
model does not improve performance results uniformly across all
labels.
As future works, we plan on investigating further in what con-
ditions articles allow classification improvements and how such
knowledge could be used as a way to make predictions more suit-
able for interpretability.
ACKNOWLEDGMENTS
We would like to thank the Cyberjustice Laboratory at Université de
Montréal, the LexUM Chair on Legal Information and the Autonomy
through Cyberjustice Technologies (ACT) project for their support
of this research.
REFERENCES
[1] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
NAACL-HLT (1). 4171–4186. https://aclweb.org/anthology/papers/N/N19/N19-
1423/
[2] Louis Martin, Benjamin Muller, Pedro Javier Ortiz Suárez, Yoann Dupont, Laurent
Romary, Éric Villemonte de la Clergerie, Djamé Seddah, and Benoît Sagot. 2019.
CamemBERT: a Tasty French Language Model. arXiv preprint arXiv:1911.03894
(2019).
[3] Olivier Salaün, Philippe Langlais, Andrés Lou, Hannes Westermann, and Karim
Benyekhlef. 2020. Analysis and Multilabel Classification of Quebec Court Decisions
in the Domain of Housing Law. In International Conference on Applications of
Natural Language to Information Systems. Springer, 135–143.
[4] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N
Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In
Advances in neural information processing systems. 5998–6008.
[5] Hannes Westermann, Vern R Walker, Kevin D Ashley, and Karim Benyekhlef.
2019. Using Factors to Predict and Analyze Landlord-Tenant Decisions to Increase
Access to Justice. In Proceedings of the Seventeenth International Conference on
Artificial Intelligence and Law. 133–142.
 A simple mathematical model
for the legal concept of balancing of interests
Frederike Zufall∗ Rampei Kimura∗ Linyu Peng∗
zufall@coll.mpg.de rampei@aoni.waseda.jp l.peng@mech.keio.ac.jp
Max Planck Institute for Research on Waseda Institute for Advanced Study Keio University
Collective Goods Tokyo, Japan Yokohama, Japan
Bonn, Germany

ACM Reference Format: 2 FROM LEGAL CRITERIA TO

Frederike Zufall, Rampei Kimura, and Linyu Peng. 2021. A simple math- MATHEMATICAL LEGAL PARAMETERS
ematical model for the legal concept of balancing of interests. In Eigh-
teenth International Conference for Artificial Intelligence and Law (ICAIL’21), We implement these legal criteria as parameters in our models,
June 21–25, 2021, São Paulo, Brazil. ACM, New York, NY, USA, 2 pages. called legal parameters.
https://doi.org/10.1145/3462757.3466140
Status of the person. We understand a person’s status as an in-
dicator for the degree of relevance the person is assigned for the
This work investigates the extent to which a mathematical model public discourse. We operationalise this criterion as the following
is able to stand in for a legal assessment performed by a lawyer. We parameter taking values between 0 and 1:
propose two simple mathematical models for the legal concept of 𝛼𝑝 ∈ [0, 1]: status of the person.
the balancing of interests by transforming legal criteria into input We take the following data points as examples to create data:
arguments of the models.
A person that ...
𝛼𝑝 = 0.01 .. is publicly unknown
1 THE LEGAL CONCEPT 𝛼𝑝 = 0.25 .. is relatively unknown to the public
A recurring concept in legal systems is the resolution of conflicts 𝛼𝑝 = 0.50 .. is to a certain degree known to the public
between competing interests through balancing. In order to develop 𝛼𝑝 = 0.75 .. is largely known in public
a mathematical model for this concept, we build on the conflict 𝛼𝑝 = 0.95 .. is known to nearly anyone
between Art. 7, Art. 8 of the EU Charter of Fundamental Rights
(EUCh) (right to privacy and right to the protection of personal data) Sphere of the information. Given the more or less private nature
and Art. 11 EUCh (right to freedom of expression and information). of information, we assume a sphere-model, starting from an inner
The typical example is the disclosure of personal data on the internet circle containing the most private information followed by such
as an act of free expression or subject to the right of access to relating to family and friends to the social sphere at the outer
information. circle. We operationalise the sphere of information by the following
As the law ultimately cannot foresee every possible situation parameter with values between 0 and 1:
in which these interests might collide, the balancing of interests 𝛼𝑠 ∈ [0, 1]: sphere of the information.
provides the legal instrument to take into consideration the partic-
𝛼𝑠 = 0.05 (e.g., health data)
ularities of each case. These particularities can be generalised to
𝛼𝑠 = 0.25 (e.g., family and friends)
legal criteria affecting the balancing. For the balancing between
𝛼𝑠 = 0.50
the right to the protection of personal data and access to informa-
𝛼𝑠 = 0.75 (e.g., professional misconduct)
tion these criteria may be the person’s social status or role in public 𝛼𝑠 = 0.95 (e.g., committing a major crime)
life, the sphere from which the relevant information originated or
the time that had passed since the occurrence of the underlying Time. We also consider the passage of time as a legal criterion
facts of that information. that affects the balancing. Following the European Court of Justice’s
case law on the “the right to be forgotten" (Case C-131/12 - Google
∗ All
Spain), we assume that the more time has passed, the more the
authors contributed equally to this research.
Full paper available at https://dx.doi.org/10.2139/ssrn.3834333. balancing leans increasingly towards the right to data protection.
Time 𝑡 is nondimensionalised as a legal parameter 𝛼𝑡 .
𝛼𝑡 = 𝑇𝑡 ∈ (−∞, 0]: a rescaling of time 𝑡 ≤ 0 with a properly
Permission to make digital or hard copies of part or all of this work for personal or chosen large number 𝑇 > 0.
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation The legal decision is made on facts that just occurred (𝛼𝑡 = 0), or
on the first page. Copyrights for third-party components of this work must be honored. on facts that occurred in the past:
For all other uses, contact the owner/author(s).
𝛼𝑡 = −𝑚 m years ago (In this paper, we choose 𝑚 = 0, 1, 3, 6, 8, 10.)
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06. Outcome. We denote the competing rights as (𝑖 1 ) privacy of infor-
https://doi.org/10.1145/3462757.3466140 mation and (𝑖 2 ) access to information. Their indices are respectively

270
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Zufall, Kimura and Peng

denoted as 𝑢 1 and 𝑢 2 such that This time-independent model (1) does not take 𝛼𝑡 as an input
𝑢𝑘 ∈ [0, 1], 𝑘 = 1, 2, subject to 𝑢 1 + 𝑢 2 = 1, argument and only captures each point in time separately.

making it sufficient to use a single parameter 𝑢 ∈ [0, 1] as the 3.2 A time-dependent mathematical model
balancing outcome. In order to model time continuously and not just as intermittent
Data coding. Based on the above criteria, a dataset (150 sets of points, we propose the following time-dependent model for the
outcomes 𝑢) is hand-coded by a fully-qualified lawyer, to serve as outcome function
training data for the models. The data points are based on standards 𝑐 00 + 𝑐 10 𝛼𝑝 + 𝑐 01 𝛼𝑠 + 𝑐 20 𝛼𝑝2 + 𝑐 11 𝛼𝑝 𝛼𝑠 + 𝑐 02 𝛼𝑠2
inferred from the relevant case law. 𝑢 (𝛼𝑝 , 𝛼𝑠 , 𝛼𝑡 ) = ,
𝑎(log (|𝛼𝑡 | + 1)) 2 + 𝑏 log (|𝛼𝑡 | + 1) + 1
(4)
3 THE MATHEMATICAL MODELS where the model parameters 𝑎, 𝑏, 𝑐 00, 𝑐 01, . . . are to be determined
For any given piece of information, the purpose of the models is by using the data. Unlike the time-independent model, this function
to determine whether (𝑖 1 ) privacy of information outweighs (𝑖 2 ) takes the legal parameter 𝛼𝑡 as an argument and reduces to the time-
access to information or vice versa. To summarize, the parameters independent model (1) at a given time. We impose the following
are defined as follows: assumptions:
𝛼𝑝 ∈ [0, 1] status of the person 𝑢 (𝛼𝑝 , 𝛼𝑠 , −∞) = 0 for all 𝛼𝑝 , 𝛼𝑠 ,
(5)
𝛼𝑠 ∈ [0, 1] sphere of the information 𝑢 (0, 0, 𝛼𝑡 ) = 0 for all 𝛼𝑡 , 𝑢 (1, 1, 0) = 1,
𝛼𝑡 = 𝑇𝑡 ∈ (−∞, 0] nondimensionalised time
yielding 𝑐 00 = 0 and
𝑢𝑘 ∈ [0, 1], subject to 𝑢 1 + 𝑢 2 = 1 index for (𝑖𝑘 ), 𝑘 = 1, 2
𝑐 10 + 𝑐 01 + 𝑐 20 + 𝑐 11 + 𝑐 02 = 1. (6)
The constraint 𝑢 1 + 𝑢 2 = 1 allows us to define one single index Again, we use Mathematica to derive optimal values of the model
to fulfill the task. This is the outcome 𝑢, which is a function of parameters using the method of least squares:
the legal parameters 𝛼𝑝 , 𝛼𝑠 and 𝛼𝑡 . The final decision, namely
𝑎 ∗ = 0.165792, 𝑏 ∗ = −0.212271, ∗
𝑐 01 = 0.529979,
whether (𝑖 1 ) privacy of information or (𝑖 2 ) access to the information (7)
dominates, is made via the comparison with a prior given threshold
∗
𝑐 10 = −0.0110422, ∗
𝑐 02 = −0.0559473, ∗
𝑐 11 = 0.295508.
value 𝑢 0 ∈ [0, 1]. Without loss of generality, we assume that when As an illustration, Figure 1 shows surface of the fitted time-
𝑢 ≤ 𝑢 0 , (𝑖 1 ) dominates, and otherwise, (𝑖 2 ) dominates. dependent outcome function 𝑢 in comparison to our data points at
−3 year.
3.1 A time-independent mathematical model
For simplicity, we first propose a simple quadratic model for each
(rescaled) year 𝛼𝑡 respectively as follows
𝑢 (𝛼𝑝 , 𝛼𝑠 ) = 𝑐 00 + 𝑐 10 𝛼𝑝 + 𝑐 01 𝛼𝑠 + 𝑐 20 𝛼𝑝2 + 𝑐 11 𝛼𝑝 𝛼𝑠 + 𝑐 02 𝛼𝑠2, (1)
where 𝑐 00, 𝑐 01, . . . are to be determined using the given dataset for
each year separately. Note that in the mathematical model, the legal
parameters 𝛼𝑝 , 𝛼𝑠 and 𝛼𝑡 are model arguments while 𝑐 00, 𝑐 10, . . .
serve as model parameters. We impose the reasonable assumptions
𝑢 (0, 0) = 0 for all 𝛼𝑡 , 𝑢 (1, 1) = 1 for all 𝛼𝑡 , (2)
leading to that for all 𝛼𝑡 ,
Figure 1: −3 year for the time-dependent model (4).
𝑐 10 + 𝑐 01 + 𝑐 20 + 𝑐 11 + 𝑐 02 = 1. (3)
The proposed model can be regarded as a linear optimisation 3.3 Evaluation
problem for which the coded data can be used to determine the
Chi-square test. To evaluate the fitted function for our time-
above coefficients, i.e. model parameters. Thus, we fit this function
dependent model in comparison to the whole dataset, we use the
with the coded data by using Mathematica; the algorithm is based on
chi-square test where 𝑁 is the number of data in the dataset; here
the theory of linear least squares. In Table 1, the optimal coefficients
𝑁 = 150. This gives the reduced chi-square
(denoted by 𝑐 ∗ ), e.g., model parameters, are listed for each year.
∑︁
𝑁
(𝑢 data − 𝑢) 2 𝜒2
𝛼𝑡 (year) 𝜒 2 := and = 0.0343305. (8)
∗
𝑐 01 ∗
𝑐 10 ∗
𝑐 02 ∗
𝑐 20
𝑢 𝑁
0 0.756269 0.218749 -0.144324 0.181876 𝑖=1
-1 0.655165 0.0286861 -0.088864 0.301803 It implies that the fitted function can describe the original dataset
-3 0.429315 -0.159663 0.00774652 0.390121 with sufficient accuracy.
-6 0.184965 -0.174577 0.15114 0.253607
-8 0.129208 -0.241208 0.163708 0.30786 Cross-validation. In order to evaluate the time-dependent model
-10 0.0662971 -0.295998 0.185813 0.364145 in terms of predictability, we use leave-on-out cross-validation and
Table 1: Fitted model parameters for model (1). calculate the mean absolute error:
𝑀𝐴𝐸 = 0.0728038. (9)

271
 Part IV

Demonstrations
 Interactive System for Arranging Issues
based on PROLEG in Civil Litigation
Ken Satoh Kazuko Takahashi
National Institute of Informatics Kwansei Gakuin University
Chiyoda, Tokyo, Japan Sanda, Hyogo, Japan
ksatoh@nii.ac.jp ktaka@kwansei.ac.jp
1 INTRODUCTION
In Japan, we have the procedure of “arranging issues” in civil liti-
gation where we clarify which facts are in dispute and what kind
of evidence action should be made for these issues. Currently, IT
technology is used only for online meeting for arranging issues
and more sophisticated method is expected by a full use of IT/AI
technology.
We proposed a method for formalizing Japanese presupposed
Ultimate Fact theory (JUF theory, in short, Youken-jijisturon, in
Japanese) and converting it into logic programming and developed
a system called PROLEG (PROlog-based LEGal reasoning support
system)[1]. JUF formalises which party should give certain facts to
get a desired legal effect for the party, in other words, formalizes
which party has a burden of proof for these facts. Then, given these
facts, PROLEG simulates reasoning by a judge to make a final con-
clusion and present such process in a directed tree structure called
“block diagram.”
In this work, we modify the PROLEG system to support arrang-
ing issues in civil litigation. In the PROLEG system, they assume
that all the facts are given before simulation of judge’s reasoning
so there are no interaction between both parties of plaintiff and de-
fendant during the simulation. On the other hand, given a desired
effect requested by one party, our interactive system (which we call
int-PROLEG) automatically calculates possible justifications for the
desired effect based on JUF theory stored in the system. After the
party chooses a justification and int-PROLEG asks for the existence
of necessary facts to the party to satisfy the justification. Then,
int-PROLEG asks whether the other party agrees on alleged facts
and also calculates possible counter-arguments against the chosen
justification and provides them to the other party. We iterate this
process until no further (counter-)arguments are presented. When
this process is finished, disagreed facts are issues for which a judge
decides the truth value.
Most interactive argument systems are mainly for construct-
ing arguments manually by a user or evaluating arguments con-
structed by a user. A notable exception would be the Carneades
system which has a funcition of argument invention using argu-
mentation schemes [2]. On the other hand, in our system, a user
does not construct arguments from the scratch but chooses a pat-
tern of legal arguments provided by int-PROLEG based on the JUF
Permission to make digital or hard copies of part or all of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full cita-
tion on the first page. Copyrights for third-party components of this work must be
honored. For all other uses, contact the owner/author(s).
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s).
273
Tatsuki Kawasaki
educe Co.,Ltd
Chiyoda, Tokyo, Japan
sktk40829@gmail.com
theory and specifies concrete facts to make concrete legal argu-
ments for a specific case in a civil litigation. Moreover, from our
experience working with lawyers, we noticed that if we introduced
a system with sophisticated but complex reasoning mechanisms,
they would be very reluctant to use the system. So, our purpose
of this work is to identify the simplest function to arrange issues
which would be easily understood by lawyers. In a sense, we ex-
tract a useful part of Carneades for arrangement of issues in civil
litigation.
2 PROLEG
We firstly review the PROLEG system[1]. A program of PROLEG
consists of a rulebase and a factbase. A rulebase consists of a set of
general rules of the form
𝐻 ⇐ 𝐵 1, ..., 𝐵𝑛 .
where 𝐻 (called head or conclusion), 𝐵 1, ..., 𝐵𝑛 (called body) are first-
order atom, and a set of exception rules of the form
𝑒𝑥𝑐𝑒𝑝𝑡𝑖𝑜𝑛(𝐻, 𝐸)
where 𝐻 and 𝐸 are the head of some general rules. We call 𝐸 excep-
tion. A factbase consists of set of the following expression 𝑓 𝑎𝑐𝑡 (𝑃)
where 𝑃 is an atom which is never the head of any general rule.
We call 𝑃 fact predicate.
A rule represents a general default rule meaning that if 𝐵𝑖 in 𝑅
are all proved then in general 𝐻 is true except there is an exception
rule 𝑒𝑥𝑐𝑒𝑝𝑡𝑖𝑜𝑛(𝐻, 𝐸) such that 𝐸 is proved.
Given a PROLEG program, we can construct a proof tree of a
given goal which is the root of the tree and the child nodes are
conditions of general rules of the conclusion and the exceptions of
the conclusion.
3 EXTENSION TO ARRANGE ISSUES
In this section, we show how to modify the PROLEG system into
int-PROLEG.
3.1 Indexing a level for PROLEG literals
(1) First, we define the dependency on the atomic formula that
appears in the general rule. Among the conclusions of the
general rule, the conclusion that does not appear in the body
of any general rule or is not an exception of any exception
rule is called 0-level conclusion. Then, when making a top-
down proof tree from the 0-level conclusion using only the
general rules, we end up the fact predicates. We call the fact
predicates finally visited 0-level facts and the 0-level con-
clusion and the intermediate visited atomic formulas called
0-level atomic formulas.
(2) Suppose that the 𝑖-level atomic formulas and the 𝑖-level facts
are decided. For exception rules that conclude with the 𝑖-
level atomic formula, the collection of the atomic formulas
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
of the exceptions of the such exception rules are called (𝑖 +
1)-level exception rules. When making a top-down proof tree
from the (𝑖 + 1)-level exception using only the general rules,
we end up the atomic formulas. We call the fact predicates
finally visited (𝑖 +1)-level facts and the (𝑖 +1)-level exception
and the intermediate visited atomic formulas called (𝑖 + 1)-
level atomic formulas.
3.2 Interaction Process in int-PROLEG
Let 𝐷 be a defendant and 𝑃 be a plaintiff.
(1) Let 𝑃 choose one of the following 0-level conclusion. This
is 𝑃’s claim.
(2) We make a top-down proof tree from 0-level conclusion and
when we encounter the 0-level fact, we ask 𝑃 if the 𝑃 claims
that fact. If 𝑃 claims it, then we let 𝑃 to instantiate variables
in 0-level facts and added instantiated facts to the fact base
and we reflect the relevant part of the proof tree with such
instantiation. If 𝑃 does not claim it, we delete the path re-
lated with the fact in a proof tree.
(3) If all the concrete facts are entered, the modified proof tree
is displayed.
(4) Suppose that a user finished to enter 𝑡-level facts. We make a
proof tree for (𝑡 +1)-level exception and when we encounter
the (𝑡 + 1)-level fact,
• if (𝑡 +1) is odd, we ask 𝐷 if 𝐷 claims that fact. If 𝐷 claims it,
then we let 𝐷 to instantiate variables in (𝑡 + 1)-level facts
and added instantiated facts to the fact base and we reflect
the relevant part of the proof tree with such instantiation.
If 𝐷 does not claim it, we delete the path related with the
fact in a proof tree. We also ask 𝐷 whether 𝐷 admits 𝑡𝑢𝑟𝑛-
level fact or not, and if 𝐷 does not admit it, we make a
label of the fact as an “issue to be determined”.
• if (𝑡 + 1) is even, we ask 𝑃 if 𝑃 claims that fact. If 𝑃 claims
it, then we let the 𝑃 to instantiate variables in (𝑡 + 1)-
level facts and added instantiated facts to the fact base and
we reflect the relevant part of the proof tree with such
instantiation. If 𝑃 does not claim it, we delete the path
related with the fact in a proof tree. We also ask 𝑃 whether
𝑃 admits 𝑡𝑢𝑟𝑛-level fact or not, and if 𝑃 does not admit it,
we make a label of the fact as an “issue to be determined”.
(5) If all the concrete facts are entered, the modified proof tree
is displayed.
(6) Continue above until there are no new argument raised.
The final proof tree will be all the arguments raised by the both
parties and fact nodes labelled as an “issue to be determined” are
issues for which a judge decide the truth value.
4 DEMONSTRATION
We show a demonstration how a plaintiff and a defendant interact
each other to arrange issues in the following claims about claiming
payment in a purchase contract1 . In the example case, the plaintiff
(Alice) requested the defendant (Bob) for the payment of goods ac-
cording to the purchase contract with Bob. Bob claimed that Alice
threatened Bob to buy the goods for a counter-argument, but Alice
did not admit the threatening action. The system works as follows.
1 http://research.nii.ac.jp/~ksatoh/PROLEGdemo/IssueArrangmentDemo.mp4
274
Ken Satoh, Kazuko Takahashi, and Tatsuki Kawasaki
Figure 1: Final Diagram
(1) Firstly, Alice claims the payment and the system tells a nec-
essary fact to establish the claim according to the Civil Code,
which is an establishment of a purchase contract. Alice gives
a specific fact for the purchase contract (called “contract1”).
(2) Secondly, the system computes counter arguments against
Alice’s claim according to the Civil Code and provides two
counter arguments (“already paid” argument and “menaced
to establish the contract” argument). Bob chooses “menaced
to establish the contract” argument and gives relevant fact
for the argument. (A block conntected with a dotted arrow
in Fig.1 represents a counter argument.).
(3) Thirdly, the system computes counter arguments agaist Bob’s
claim but founds no counter arguments. So the only thing
Alice can do is to deny some facts. In this example, Alice
denies the facts related with Bob’s claim of being menaced
so these facts become issues for which a judge decides the
truth value. (Leaf blocks with ’u’ in the bottom in Fig.1 are
issues.).
5 CONCLUSION
We present how to extend the PROLEG system to arrange issues.
This system helps for lawyers to prevent missing claims and for
lay person to make correct legal claims without lawyers. For a fu-
ture research, we need an evaluation of the system and investigate
a method of automatic extraction of relevant legal facts directly
from claims in natural language.
Acknowledgements: This work was supported by JSPS KAKENHI,
JP17H06103 and JP19H05470 and JST AIP, JPMJCR20G4.
REFERENCES
[1] K. Satoh et al. 2012. PROLEG: An Implementation of the Presupposed Ultimate
Fact Theory of Japanese Civil Code by PROLOG Technology. In New Frontiers
in Artificial Intelligence: JSAI-isAI 2010 Workshops, Revised Selected Papers, LNAI
6797. 153–164.
[2] D. Walton and T. F. Gordon. 2017. Argument Invention with the Carneades Argu-
mentation System. Scripted 14, 2 (2017), 168–207. https://script-ed.org/?p=3391
 Live Demonstration Of A Working Collaborative
eNegotiation System (Smartsettle Infinity)
Ernest Thiessen Graham Ross
Founder & President Head of International Marketing
Smartsettle Resolutions Inc. Smartsettle Resolutions Inc.
Vancouver, BC, Canada Vancouver, BC, Canada
ernest.thiessen@smartsettle.com Board Member
International Council for Online Dispute Resolution

KEYWORDS
negotiation, multivariate visual blind bidding, eNegotiation, col- Figure 1: Smartsettle Multivariate Visual Blind Bidding
laborative, conflict resolution, alternative dispute resolution, ADR,
online dispute resolution, ODR, algorithms

The Smartsettle1 process, using anonymised proposals, Visual

Blind Bidding, machine learning of disputant preferences, ma-
chine generated proposals, and intelligent algorithms, not only pro-
motes and increases fair outcomes that go “beyond win-win” but,
by rewarding good negotiating behaviour, creates a revolutionary
challenge to status quo adversarial methods of negotiation. Eight al-
gorithms power a collaborative negotiation software product called
Smartsettle Infinity2 , which is designed to support multiple deci-
sion makers with conflicting objectives and perspectives in complex
multi-party negotiations.
Infinity models all aspects of complex cases[3], including any
number of quantitative and qualitative issues, and related prefer-
ences of negotiators. With preferences well-represented, parties (3) Each party has a minimum level of satisfaction that they
can confidently compare complete settlement packages in a non- would be willing to accept. Supplier achieves 100% satisfac-
emotional and objective way. tion at the top left while buyer’s goal is at the bottom ght, but
The context in the example case, depicted by Figure 1, is a nego- Smartsettle’s goal is to find a fair solution for both parties
tiation between a Supplier and a Buyer. The following dozen steps on the Efficiency Frontier.
that refer to Figure 1 invoke all eight Smartsettle algorithms. Parties (4) The area bounded by the minimum satisfaction of each party
would normally encounter these steps in an iterative process that and the Efficiency Frontier is the Zone of Possible Agree-
is designed to reward good negotiating behaviour. ment.
(5) Parties are encouraged to exchange optimistic proposals (red
(1) With the negotiation problem well-represented by a Single
stars).
Negotiating Framework, parties model their problem within
(6) The Visual Blind Bidding algorithm is depicted by Fig-
the Smartsettle platform.
ure 1 for a two-party negotiation. Visual Blind Bidding
(2) Parties use Comprehensive Preference Analysis to define
enables parties to request Suggestions (black stars) based
how they become satisfied on each of the issues. A secure
on party preferences and that fall between the optimistic
neutral server guards confidential information.
proposals. This process virtually eliminates the tedious nego-
tiation dance that characterizes most ordinary negotiations.
1 Smartsettle is disruptive technology that is emerging from nearly three decades of Another unique feature of Visual Blind Bidding is the
research and development. This research grew out of a challenge put to Ernest Thiessen
in 1990 by his major advisor, Professor Emeritus D. Peter Loucks at Cornell University. ability of either party or a mediator to make anonymous pro-
2 A simplified version of this system called Smartsettle ONE is described in a longer posals. They do this by requesting Smartsettle to camouflage
version of this paper[5] their proposal so that it appears no different than any other
Suggestion generated by the machine.
Permission to make digital or hard copies of part or all of this work for personal or (7) Parties motivated to collaborate secretly accept proposals
classroom use is granted without fee provided that copies are not made or distributed that fall in the Zone of Possible Agreement, which in turn
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for third-party components of this work must be honored. moves them quickly toward a fair resolution to their conflict.
For all other uses, contact the owner/author(s). (8) If both parties have each accepted more than one of the same
ICAIL’21, June 21–25, 2021, São Paulo, Brazil packages by the end of a Session (inside the red circle), Re-
© 2021 Copyright held by the owner/author(s).
ACM ISBN 978-1-4503-8526-8/21/06. ward for Early Effort determines which of those packages
https://doi.org/10.1145/3462757.3466091 becomes the Baseline agreement (orange star).

275
ICAIL’21, June 21–25, 2021, São Paulo, Brazil E. Thiessen et al.

(9) If no mutually accepted packages exist, the Automatic Deal- research[4] and live case work have demonstrated that the magni-
Closer can be invoked to avoid impasse due to a small gap. tude of value left behind in ordinary negotiations can be around 16%.
(10) Having reached a Baseline quickly, parties have energy left Shell’s research[2] concluded that negotiators that are subjected
for Smartsettle’s signature algorithm, Maximize the Mini- to a tedious negotiation dance become exhausted and have little
mum Gain, to uncover any remaining hidden value and gen- energy left to go “beyond win-win”® in a search for hidden value.
erate an Improvement that distributes the additional value The Smartsettle Visual Blind Bidding process not only conserves
fairly to all parties. This algorithm is foundational to all the the energy of negotiators but makes it very easy to uncover hidden
others and endorsed3 by experts in the field. value.
(11) Fairness Enhancing Normalization distributes additional
benefits fairly among all the parties, and this brings them to Table 1: Rewards for Good Negotiating Behaviour
an optimal solution on the Efficiency Frontier (green star).
Smartsettle allows parties to represent their preferences with Objective Behaviour Reward
any scale that is convenient to them. But under the covers, Acceptance of a fair A timely win-win
Smartsettle employs a proprietary method for normalization Fairness outcome outcome
that effectively neutralizes the efforts of any party to inflate Early movement to Bigger portion of
the benefits of optimization for themselves. Zone of Agreement the overlap
(12) In the background is the Expert Neutral Deal-Closer, which Agreement to Ex- Guaranteed agree-
the parties may fall back on in case of a large gap, and just pert Neutral Deal- ment
the existence of this remedy results in it rarely being used. Closer
Efficiency Secure Honesty Uncovered hidden
The process described above works for almost any formal negotia-
Truthfulness value
tion between any number of parties and can be depicted graphically
by adding more dimensions to Figure 1. Maximize the Minimum Peace Collaboration Improved relation-
Gain as described by US ICANS Patent US 5495412A has been ships
slightly modified to deal better with multi-party negotiations.
In high-value cases, Infinity’s ability to uncover both tangible In traditional negotiations, parties will often tend to hide or even
and intangible hidden value that improves the outcome for all misrepresent their true preferences. However, with Smartsettle
parties even after a settlement package has been agreed and thereby Infinity, the temptation to misrepresent preferences is eliminated.
enhances further the inter-party relationships, is by far its greatest Skilled facilitators help parties understand that it is actually counter-
asset. The greatest benefits are achieved when well-trained parties productive to use any sort of deception as a negotiating strategy
use the system collaboratively. Rather than usurp control from the and that truthfulness is rewarded. All of these good behaviours
user or the work of the mediator or negotiation advisor, the process together represent the fifth behaviour of collaboration, and result
benefits from human input that best understands and assesses the in improved relationships.
preferences and interests of all parties.
Table 1 summarizes how Smartsettle intelligently rewards good REFERENCES
negotiating behaviour. Whether this is artificial intelligence (AI) or [1] Howard Raiffa. 1996. Lectures on Negotiation Analysis. Program on Negotiation at
Harvard Law School.
intelligence augmented (IA) we leave for the reader to decide. [2] G. Richard Shell. 1999. Bargaining for Advantage: Negotiation Strategies for Rea-
Acceptance of a fair outcome is the first prerequisite for achieving sonable People. Penguin.
a result that benefits all parties. Smartsettle enables this behaviour [3] Ernest Thiessen, Peter Holt, Graham Ross, and Diana Wallis. 2017. Brexit 2.0
Negotiation Simulation with Smartsettle Infinity. International Journal of Online
in a process where parties can place secret bids on packages. When Dispute Resolution 2, 4 (2017).
a Zone of Agreement occurs, the party who made the smallest [4] Ernest M. Thiessen and D. Pete. Loucks. 1992. Computer-Assisted Negotiation
last move is rewarded with a bigger portion of the overlap. An of Multi-objective Water Resources Conflicts. Water Resources Bulletin, American
Water Resources Association 28, 1 (February 1992), 163–177.
agreement is ensured if parties agree to the Expert Neutral Deal- [5] Ernest M. Thiessen and Graham L. Ross. 2021. Using AI & IA to Reward Good
Closer in Final Session, and in fact is more likely to happen without Negotiating Behaviour. (on smartsettle.com).
the need of outside intervention. These first three behaviours all
contribute to quickly achieving a fair outcome and are applicable
to all negotiations, whether simple or complex.
In more complex multivariate cases, the importance of coming to
an early agreement is even greater. In addition to time savings, nego-
tiators also have the opportunity of uncovering hidden value with
the fourth behaviour of secure honesty and truthfulness. Thiessen’s

3 Harvard Professor Emeritus Howard Raiffa published [1] a preference for Maximize
the Minimum Gain (MMG) over Nobel Laureate’s algorithms Maximize the Utility
Product (MUP). Raiffa said that MMG was more intuitive and (in his opinion) produced
better outcomes in certain hypothetical illustrations. He did admit however that the
difference between these two algorithms would be insignificant in most real world
applications.

276
 Part V

COLIEE Papers
 BERT-based Ensemble Methods with Data Augmentation for
Legal Textual Entailment in COLIEE Statute Law Task
Masaharu Yoshioka
yoshioka@ist.hokudai.ac.jp
Faculty of Information Science and Technology, Hokkaido
University
Graduate School of Information Science and Technology,
Hokkaido University
Sapporo-shi, Hokkaido, Japan
ABSTRACT
The Competition on Legal Information Extraction/Entailment (COL-
IEE) statute law legal textual entailment task (task 4) is a task to
make a system judge whether a given question statement is true
or not by provided articles. In the last COLIEE 2020, the best per-
formance system used bidirectional encoder representations from
transformers (BERT), a deep-learning-based natural language pro-
cessing tool for handling word semantics by considering their con-
text. However, there are problems related to the small amount of
training data and the variability of the questions. In this paper, we
propose a BERT-based ensemble method with data augmentation
to solve this problem. For the data augmentation, we propose a
systematic method to make training data for understanding the
syntactic structure of the questions and articles for entailment. In
addition, due to the nature of the non-deterministic characteristics
of BERT fine-tuning and the variability of the questions, we pro-
pose a method to construct multiple BERT fine-tuning models and
select an appropriate set of models for ensemble. The accuracy of
our proposed method for task 4 was 0.7037, which was the best
performance among all submissions.
CCS CONCEPTS
• Computing methodologies → Information extraction; En-
semble methods.
KEYWORDS
Textual entailment, Data augmentation, BERT, Ensemble method
ACM Reference Format:
Masaharu Yoshioka, Yasuhiro Aoki, and Youta Suzuki. 2021. BERT-based
Ensemble Methods with Data Augmentation for Legal Textual Entailment
in COLIEE Statute Law Task. In Eighteenth International Conference for
Artificial Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil.
ACM, New York, NY, USA, 7 pages. https://doi.org/10.1145/3462757.3466105
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466105
278
Yasuhiro Aoki
Youta Suzuki
yasu-a_01@eis.hokudai.ac.jp
suzuki@eis.hokudai.ac.jp
Graduate School of Information Science and Technology,
Hokkaido University
Sapporo-shi, Hokkaido, Japan
1 INTRODUCTION
The Competition on Legal Information Extraction/Entailment (COL-
IEE) [3, 4, 10, 11, 15] serves as a forum to discuss issues related to
legal information retrieval (IR) and entailment. There are two types
of tasks in COLIEE. One is a task using case law (tasks 1 and 2),
and the other is a task using Japanese statute law with Japanese bar
exam questions (tasks 3, 4, and 5). Task 3 is an IR task that aims to
retrieve (a) relevant law article(s) to judge whether the statement of
the question is true, task 4 is an entailment task that judges whether
a given relevant article entails a given question statement and task
5 is a combination of tasks 3 and 4.
Because a part of bar exam questions are based on real use cases,
it is important to have a mechanism for semantic matching to dis-
cuss the relevance of words in the articles and those in the questions
for entailment. At an earlier stage, machine-readable thesauruses
such as WordNet[7] and distributed representation of words such
as Word2Vec[6] were used. Recently, the deep learning-based natu-
ral language processing tool bidirectional encoder representations
from transformers (BERT) [1] was introduced. One of the character-
istics of BERT is that it provides a general semantic analysis system
that can be fine-tuned for a particular task. For the last COLIEE
[10], the best performance systems for tasks 3 [12] and 4 [9] used
BERT as a core component of the system.
In this paper, we propose a method to use BERT-based ensem-
ble methods for task 4. This method utilizes BERT with data aug-
mentation that increases training examples by making article-and-
question pairs systematically using sentences in the statute law
articles. We also propose a system that ensembles the results from
multiple BERT-based system outputs. The accuracy of the system
for task 4 was 0.7037, which was the best performance among all
the submitted runs for task 4 at COLIEE 2021.
2 RELATED WORKS
Because bar exam questions include questions about real use cases
of articles, it is necessary to discuss the correspondence between
the concepts used in the articles and real use cases. In the early
stage of COLIEE, several attempts were made to utilize resources
for discussing such semantic matching, such as a machine-readable
thesaurus and data for the distributed representation of the terms.
For example, Mi-Young et al. [5] used Word2Vec [6] as a resource
for distributed representation, and Taniguchi et al. [14] proposed
a method to utilize WordNet [7] as a machine-readable thesaurus.
However, because those methods cannot handle the context to
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
estimate the meaning of such terms, they are not as effective for
utilizing such resources.
Recently, Devlin et al. [1] proposed BERT, a deep learning-based
natural processing tool pretrained for solving general tasks that re-
quire semantic information with larger corpora (such as the whole
contents of Wikipedia). Based on this training process, BERT can
handle the meaning (distributed representation) of words in a sen-
tence by considering the context. In addition, BERT can be used
for various tasks by employing a fine-tuning process that utilizes
comparatively small numbers of training data. Because a pretrained
model of BERT contains rich information about the semantics of
the words, the fine-tuned models may be able to handle semantic
information even though the words themselves are not included in
the training data.
At COLIEE 2020, the BERT-based system achieved the best per-
formance for legal textual entailment tasks (JNLP [9]). In that paper,
they proposed a lawfulness classification approach that classified
the appropriateness of legal statements by using many legal sen-
tences that include bar exam questions provided by organizers
without considering given relevant articles. This approach worked
well for COLIEE 2020 because of the large number of training data.
In addition, they also pointed out that it was difficult to select an
appropriate model using validation data for the unseen questions
because of the significant variability of the questions.
To increase the size of the training data, the data augmentation
approach is widely used in the field of image recognition [13].
However, few studies related to the data augmentation method
have been conducted for the legal textual entailment task. Min
et al. [8] proposed a syntactic-based data augmentation method
to increase the robustness of natural language inferences. They
proposed a systematic method to create positive and negative data
from the correct inference sentence by a syntactic operation such
as passivation and the inversion of subject and object. Evans et al.
[2] proposed a method for data augmentation for logical entailment.
In this framework, their method increased negative and positive
data by modifying logical inference rules using symbolic vocabulary
permutation, which includes an operation to make implication rules
that share the same contents for the condition and derived parts.
Those approaches are useful to design data augmentation methods
for legal textual entailment.
3 BERT-BASED ENSEMBLE LEGAL TEXTUAL
ENTAILMENT SYSTEM
Based on a discussion of the previous best performance system
(JNLP [9]), we propose a system with the following characteristics.
(1) Textual entailment approach with data augmentation
We assume that the reason why the lawfulness classification
approach outperformed the textual entailment one in the
last COLIEE is the size of the training data. Therefore, when
we provide larger training data by data augmentation, the
textual entailment approach may outperform the lawfulness
classification approach because it uses the most important
information (relevant articles).
(2) Ensemble results of multiple BERT-based model outputs
As discussed, it is difficult to select appropriate models for
the task by only evaluating the validation model. From our
279
M. Yoshioka, Y. Aoki and Y. Suzuki
preliminary experiment (the details are discussed in Section
3.3), we confirmed that the characteristics of the fine-tuned
BERT-based model are different and that the accuracy of the
validation data is not directly related to that of the test data.
We assume that this result reflects the different character-
istics of each model and that the appropriate selection of
the generated models for ensemble may improve the perfor-
mance for the unseen questions.
3.1 Data augmentation using articles
In the deep learning framework, it is common to enlarge training
data by modifying the existing data (data augmentation). However,
it is important to define the appropriate data augmentation method
to obtain better results. Related to the legal textual entailment task,
data augmentation methods have been used for natural language
and logical inference, as introduced in Section 2. However, it is
difficult to apply these methods to this legal textual entailment
data.
In this research, we assume that there are two types of errors to
judge whether an article entails a given question. One is semantic
mismatch, and the other is logical mismatch (the appropriateness
of the judicial decision).
For example, let us discuss the example of training data using
the following article (a part of article 9): “A juridical act performed
by an adult ward is voidable.”
(1) “A juridical act performed by an adult is voidable.”
The article does not entail this question because of semantic
matching (“adult” is not “adult ward” ).
(2) “A juridical act performed by an adult ward is not voidable.”
The article does not entail this question because of the
inappropriateness of the juridical decision (“voidable” and
“not voidable”)
(3) “A juridical act performed by an adult is not voidable.”
We cannot judge whether this question is true (it may re-
quire another article). However, a given article cannot entail
the question.
For the semantic matching case (1), it is difficult to select appro-
priate pairs (“adult” and “adult ward”) for replacement to make such
a semantic mismatch sentence. For both cases (3), it is also difficult
to make the data and to use these data for negative examples to
identify types of errors to judge the entailment results.
By contrast, if we make the pair of correct answers with logical
mismatch cases (2), the examples may help to explain the impor-
tance of comparisons between the judicial decision of the article
and that of the question.
Based on this assumption, we create training data that charac-
terize the logical mismatch between the articles and questions.
The procedures to make this augmented data are as follows.
(1) Extraction of (a) judicial decision parts from the articles.
If there are multiple decisions in an article, the sentences
in the article are split into smaller sentences that contain
one judicial decision (Figure 1). When the split sentence
explains an exceptional case, a flipped judicial decision is
complemented for the split sentence (underlined part of the
split sentences). When the sentence does not contain any
BERT-based Ensemble Methods with Data Augmentation for Legal Textual Entailment in COLIEE Statute Law Task ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Input article Split sentence

!"#$%& ' ! ()*+,-.()/0123456.78(2
()*+,-.()/0123456.78(2*78 *78()*9:;:(2)06<.=>?@3
()*9:;:(2)06<.=>?@34A90 4A90BCD1)*.EFGHA9+IJKLM
BCD1)*.EFGHA9+IJKLM If the lessee lawfully subleases a
4N38O:?@:PQ8(2*+()*:RSTU leased thing, the lessor may not duly
V.=H?@W0X3DL49J78A:Y>IKLM assert against the sublessee the
Translation: cancellation by agreement of the lease
Article 613 (3) with the lessee.
If the lessee lawfully subleases a
leased thing, the lessor may not duly 4"#5(6*,7#8+3*&*92#(,0#,27(&*92#
assert against the sublessee the 0(&(#-'+:#&;2#32,&2,)23
cancellation by agreement of the lease
with the lessee; provided, however, that Negative pair
this does not apply if, at the time of ! Question:
the cancellation, the lessor has a right ()*+,-.()/0123456.78(2
to cancel due to non-performance on the *78()*9:;:(2)06<.=>?@3
part of the lessee. 4A90BCD1)*.EFGHA9+IJHM
If the lessee lawfully subleases a
leased thing, the lessor may duly
!"#$%&'()&*+,#+-#./'*0*)(1# assert against the sublessee the
02)*3*+,#32,&2,)23 cancellation by agreement of the lease
with the lessee.
Split sentences
! Article:
! ()*+,-.()/0123456.78(2
()*+,-.()/0123456.78(2
*78()*9:;:(2)06<.=>?@3
*78()*9:;:(2)06<.=>?@3
4A90BCD1)*.EFGHA9+IJKLM
4A90BCD1)*.EFGHA9+IJKLM
If the lessee lawfully subleases a
If the lessee lawfully subleases a
leased thing, the lessor may not duly
leased thing, the lessor may not duly
assert against the sublessee the
assert against the sublessee the
cancellation by agreement of the lease
cancellation by agreement of the lease
with the lessee
with the lessee.
! O:?@:PQ8(2*+()*:RSTUV.
=H?@W0X3DL49J78(2*78()
*9:;:(2)06<.=>?@34A90
BCD1)*.EFGHA9+IJHM Figure 2: Example of splitting sentences (Article 613(3))
If the lessor has a right to cancel
due to non-performance on the part of
the lessee at the time of the Based on this process, we construct 3,331 (positive: 1,677, nega-
cancellation, the lessor may duly tive: 1,654) training examples for data augmentation.
assert against the sublessee the
cancellation by agreement of the lease 3.2 BERT-based entailment system
with the lessee.
We implemented a BERT-based entailment system using the ordinal
BERT fine-tuning process proposed in [1]. We concatenated the
Figure 1: Example of splitting sentences (Article 613(3)) question and article using a sentence-separator token ([𝑆𝐸𝑃]) and
fed it into the BERT model to estimate whether the article entails
a question (positive:1) or not (negative:0). We use the BERT-based
model of BERT-Japanese 1 .
judicial decision (e.g., definitions of terms), those sentences Training and validation data are constructed from the training
are used only to generate positive pairs. bar exam questions from H18–30 (13 years of data with 695 ques-
(2) Make positive and negative data by using extracted sen- tions) by randomly splitting 90% (625) for training and 10% (70)
tences. for validation. All augmented data are merged with training data,
We use a sentence extracted from the step 1 as an article part and we use 3,956 examples for training and 70 for validation. We
and the same sentence text as a question for a positive exam- also made training sets without using augmented data (625 training
ple pair. We made a sentence that is generated by flipping the and 70 validation examples) for comparing system performance
judicial decision. For most of the cases, we add “ない” (not) without augmented data.
or remove “ない” (not) for the judicial decision verb. We The fine-tuning of the BERT model is done using a max sequence
also use the antonym dictionary to make flipped sentences. length of 256, adam as an optimizer, a training batch size of 12, and
Pairs of the original sentence and flipped sentence are used a learning rate of 1e-5. The validation loss is calculated at each
as negative example pairs. Figure 2 shows an example of
making negative pairs. 1 https://github.com/cl-tohoku/bert-japanese

280
ICAIL’21, June 21–25, 2021, São Paulo, Brazil M. Yoshioka, Y. Aoki and Y. Suzuki

Table 1: Evaluation results of the 10 models Table 2: Evaluation results of the ensemble models

Model Validation Validation Test Model used Accuracy

No. Loss Accuracy Accuracy (1, 2, 3, 4, 5, 6, 7) 0.694
1 0.6935 0.4857 0.5946 (1, 2, 4) 0.689
2 0.7247 0.5286 0.6667 (1, 2, 3) 0.685
3 0.7566 0.6286 0.6486 (1, 2, 3, 4, 7) 0.682
4 0.6822 0.6286 0.6486 (1, 2, 6) 0.676
5 0.7347 0.5143 0.6486 (1, 2, 5) 0.676
6 0.7745 0.6143 0.6396 (1, 2, 3, 4, 5, 7) 0.676
7 0.6913 0.5429 0.6126 (1, 2, 3, 4, 5, 6) 0.676
8 0.7123 0.5857 0.6486 (1, 2, 3, 4, 5) 0.676
9 0.7504 0.6286 0.6486 ··· ···
10 0.7735 0.5857 0.6396 (1, 2, 3, 4, 5, 6, 7, 8, 9, 10) 0.622

Table 3: Evaluation results of the ensemble models

epoch and stop-training process when the validation loss increases.
We use a model with minimal validation loss.
Fine-tuned models accept a pair of question statements and (an)
article(s) as input and return whether the article(s) entail(s) the state-
ment (positive) or not (negative), which is decided by comparing
the score for the probability of being positive or negative.
3.3 Preliminary experiment
To evaluate the performance of the proposed BERT-based entail-
ment system, we conducted a preliminary experiment using the R01
data (1 year of data with 111 questions) for evaluation. To discuss
the effect of the variability of the training and validation questions
set, we made 10 models using the same procedures. Because of
the non-deterministic characteristics of the BERT fine-tuning pro-
cess and different training sets selected randomly, we expected
the system to construct different models that use different features
for analyzing the texts. Table 1 shows the evaluation results for
the validation and test data. As shown in the table, the validation
accuracy and loss were not closely related to the test accuracy. We
assume that these results reflect the variability of the question set.
We also made another 10 models without using augmented data.
The average accuracy of these 10 models was 0.5108 (best: 0.5946,
worst: 0.4505). From this comparison, we confirmed that data aug-
mentation is effective to improve the performance of the BERT
training process.
In this inference process, we can estimate the confidence of the
BERT model output by comparing the probability of positive or
negative. When the probability of positive is almost equal to 1(0), the
system outputs positive (negative) results with higher confidence.
By contrast, when the probability is close to 0.5, the system output
can be interpreted as less confidence.
When we checked the distribution of such confidence for each
question, the tendency of such confidence was not consistent among
these models. From this observation, we assumed that these models
may use different features to estimate the entailment results and
that their characteristics may differ, even though we used the same
model architecture for training. In such a case, there is a probability
to increase the performance of the overall system by ensembling
the results of different models.
Submission ID Model used
HUKB-1 (1, 2, 3, 4, 5, 6, 7)
HUKB-2 (1, 2, 4)
HUKB-3 (1, 2, 4, 7 , 8)
To discuss the appropriate settings for making the ensemble
model, we calculated the performance of the ensemble model using
these 10 models. For making the ensemble model, we used the
average probability of positive and negative from the target models.
Table 6 shows the evaluation results of the ensemble models by
accuracy. In this table, all combinations of the ensemble models are
used (selecting three to 10 models from those introduced in Table
1).
There were large differences between the accuracies of the en-
semble cases. The best accuracy system ensembled seven models,
and the worst used all models. Most of the cases that used three
to five models were adequate to estimate the results with better
accuracy.
All of the highest rank sets contained the best performance
system model 2. In addition, they also used model 1, even though
the accuracy of model 1 was the lowest among these 10 models.
This suggests that it is important to use a complementary set of
models that have different characteristics to improve the overall
performance of the ensemble models.
3.4 Submitted results
Based on the results of the preliminary experiments, we submitted
the following three results that used different model sets for the
ensemble.
HUKB-1 and HUKB-2 were the best and second-best perfor-
mance systems using R1 data as a kind of validation data. HUKB-3
selected the five best models using validation loss information.
Table 4 shows the final evaluation results of all submission runs,
among which, HUKB-2 achieved the highest accuracy.
3.5 Discussion
To understand the effect of the ensemble method, we compared
the performance of the ensemble results with one of each model.

281
BERT-based Ensemble Methods with Data Augmentation for Legal Textual Entailment in COLIEE Statute Law Task ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Table 4: Final evaluation results Table 6: Number of questions classified by the ensemble re-
sults
Submission ID Correct Accuracy
BaseLine No 43/All 81 0.5309 Submission Agree Majority Other
HUKB-2 57 0.7037 ID Correct Wrong Correct Wrong Correct Wrong
HUKB-1 55 0.6790 HUKB-1 17 4 27 14 11 8
HUKB-3 55 0.6790 HUKB-2 28 9 23 11 6 4
UA_parser 54 0.6667 HUKB-3 19 9 17 12 19 9
JNLP.Enss5C15050 51 0.6296
JNLP.Enss5C15050SilverE2E10 51 0.6296 Table 7: Topic difficulty analysis based on the number of cor-
JNLP.EnssBest 51 0.6296 rect runs
OVGU_run3 48 0.5926
TR-Ensemble 48 0.5926 No. of No. of No. of correct
TR-MTE 48 0.5926 correct runs questions answers by HUKB-2
OVGU_run2 45 0.5556 1–3 7 0
KIS1 44 0.5432 4–6 11 1
KIS3 44 0.5432 7–9 12 9
UA_1st 44 0.5432 10–12 19 15
KIS2 43 0.5309 13–15 19 19
UA_dl 43 0.5309 16–18 13 13
TR_Electra 41 0.5062
OVGU_run1 36 0.4444
than that of HUKB-2. However, the accuracy of HUKB-3 was lower
Table 5: Evaluation results of the 10 models for the test data than that of HUKB-2, which suggests that selecting an appropriate
set of models for the ensemble is also effective for maintaining the
Model No. Accuracy accuracy of the “Agree” questions.
1 0.6790 Second, we analyze the characteristics of our system based on
2 0.6666 the difficulty estimated by the number of runs that return the cor-
3 0.5185 rect answer provided by organizers. Table 7 shows the number of
4 0.5555 questions corresponding to the number of correct runs from the 18
5 0.6666 submitted runs (Table 4). Questions with a smaller number of cor-
6 0.5308 rect runs may be common difficult problems among all submitted
7 0.6790 methods.
8 0.5925 From this table, we confirm that our method answers the easy
9 0.5555 questions consistently. These characteristics may come from our
10 0.5308 ensemble method reducing the effect of variability of the training
data sets.
By contrast, our system performs poorly for difficult questions,
Table 5 shows the evaluation results of the 10 models. This year, the suggesting common problems that nearly all submitted systems
basic model performed well and the best performance systems were cannot handle at this moment.
almost equivalent to the ensemble ones. However, the appropriate We would like to discuss the characteristics of such difficult
selection of the models (HUKB-2) made the ensemble results better questions using examples.
than one for each model. The following question (Figure 3) is a difficult question that only
These results justify the appropriateness of using the ensemble one run can answer correctly. Because the main terms appear in
method by selecting an appropriate ensemble set using validation both the question and the first sentence, the systems tend to say
data. positive (entail) for this question. However, it also matches the last
Table 6 shows the number of questions classified by agreement sentence that explains an exceptional case of the articles. As a result,
level among the models used. “Agree,” “Majority,” and “Other” rep- the given article does not entail the question.
resent “all models return the same results,” “final results are same Because our data augmentation method splits the sentences and
as majority voting,” and others, respectively. From these results, we only handles flipped negative cases, as introduced in Section 3.2,
can confirm that the average calculation ensemble method is better our system cannot answer this question correctly either. However,
than majority voting because the number of correct questions for because several articles have such exceptional cases, it may be better
others is larger than the number of wrong ones. For the “Agree” to propose a data augmentation method to handle such articles.
questions, the best performance system (HUKB-2) had the largest The following failure example (Figure 4: one run can answer
numbers because of the small number of used models (three), but correctly) is also related to the logical expression (quantifier). The
the accuracy of HUKB-1 (using seven models) for “Agree” was better article says “together with the obligee” (more than two), but the

282
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Question: R02-25-E
賃借人が適法に賃借物を転貸し，その後，賃貸
人が賃借人との間の賃貸借を合意により解除し
た場合，賃貸人は，その解除の当時，賃借人の
債務不履行による解除権を有していたときであ
っても，その合意解除をもって転借人に対抗す
ることはできない。
If the lessee lawfully subleases a leased thing, the
lessor may not duly assert against the sublessee the
cancellation by agreement of the lease with the lessee
even if the lessor has a right to cancel due to non-
performance on the part of the lessee at the time of
the cancellation.
Article for entailment (answer is No)
第六百十三条３　賃借人が適法に賃借物を転貸
した場合には、 賃貸人は、賃借人との間の賃貸
借を合意により解除したことをもって転借人に
対抗することができない。 ただし、その解除の
当時、賃貸人が賃借人の債務不履行による解除
権を有していたときは、この限りでない。
Article 613 (3) If the lessee lawfully subleases a leased
thing, the lessor may not duly assert against the sub-
lessee the cancellation by agreement of the lease with
the lessee; provided, however, that this does not apply
if, at the time of the cancellation, the lessor has a right
to cancel due to non-performance on the part of the
lessee.
Figure 3: Example of the failure of a difficult question
question says “independently” (single). For this case, it is not so
easy to make a simple data augmentation method for handling this
type of logical mismatch.
The following failure example (Figure 5: three runs can answer
correctly) is related to the logical expression and semantic mismatch.
The article says “the other party to the contract gives consent”
(require consent), but the question says “regardless of whether A
consents.” Because there are no patterns for handling such logical
mismatches in augmented data, it is comparatively difficult for the
system to identify this type of logical mismatch. In addition, the
vocabulary used for representing related persons is totally different;
“A”, “B”, and “E” are used for the questions and “one of the party”,
“the other party” and “the third party”, are used in the article. It
is also difficult for the system to estimate the relationship among
them.
4 SUMMARY
In this paper, we introduced our system for participating in task
4 (legal textual entailment) of COLIEE 2021. This system uses a
BERT-based entailment system with data augmentation by flipping
the judicial decision of the article sentences. We also proposed a
method to make various BERT models and selected an appropriate
ensemble model set using a validation data set. The effectiveness of
the proposed system was evaluated by COLIEE 2021 task 4 (textual
entailment task), and the accuracy of our system was 0.7037, which
283
M. Yoshioka, Y. Aoki and Y. Suzuki
Question: R02-19-I
保証人は，被担保債権の一部を弁済したが残債
務がある場合，その弁済をした価額の限度にお
いて，代位により取得した被担保債権及びその
担保権を単独で行使することができる。
If a guarantor has partially paid a secured claim but
there is a remaining obligation, the guarantor may
independently exercise the secured claim and the secu-
rity right acquired through subrogation in proportion
to the value of the subrogee’s performance.
Article for entailment (answer is No)
第五百二条　債権の一部について代位弁済があ
ったときは、 代位者は、 債権者の同意を得て、
その弁済をした価額に応じて、債権者とともに
その権利を行使することができる。
Article 502　(1)　If performance by subrogation oc-
curs with respect to one part of a claim, the subro-
gee, with the consent of the obligee, may exercise
the rights of the subrogee together with the obligee
in proportion to the value of the subrogee’s perfor-
mance.
Figure 4: Example of the failure of a difficult question (2)
Question: R02-23-U
ＡＢ間においてＡの所有する中古の時計甲の売
買契約が締結された場合、 Ｂが，Ｅとの間で，
売買契約における買主たる地位をＥに譲渡する
旨の合意をした場合，Ａの承諾の有無にかかわ
らず，買主たる地位はＥに移転する。
If B makes an agreement with E to transfer contrac-
tual status of the buyer to E, regardless of whether A
consents, the status of the buyer is transferred to E.
Article for entailment (answer is No)
第五百三十九条の二　契約の当事者の一方が第
三者との間で契約上の地位を譲渡する旨の合意
をした場合において、その契約の相手方がその
譲渡を承諾したときは、 契約上の地位は、 その
第三者に移転する。
Article 539-2 If one of the parties to a contract made an
agreement with a third party to transfer that party’s
contractual status to that third party, and the other
party to the contract gives consent to the transfer, the
contractual status is transferred to the third party.
Figure 5: Example of the failure of a difficult question (3)
was the best among all runs. We also discussed the characteristics
of the failure of our system for future development.
ACKNOWLEDGMENT
We thank the organizers of the COLIEE for their efforts in con-
structing this test data. This work was partially supported by JSPS
KAKENHI Grant Number 18H0333808.
BERT-based Ensemble Methods with Data Augmentation for Legal Textual Entailment in COLIEE Statute Law Task
REFERENCES
[1] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proceedings of the 2019 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Volume 1 (Long and
Short Papers). Association for Computational Linguistics, Minneapolis, Minnesota,
4171–4186. https://doi.org/10.18653/v1/N19-1423
[2] Richard Evans, David Saxton, David Amos, Pushmeet Kohli, and Edward Grefen-
stette. 2018. Can Neural Networks Understand Logical Entailment?. In 6th In-
ternational Conference on Learning Representations, ICLR 2018, Vancouver, BC,
Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net.
https://openreview.net/forum?id=SkZxCk-0Z
[3] Yoshinobu Kano, Mi-Young Kim, Randy Goebel, and Ken Satoh. 2017. Overview
of COLIEE 2017. In COLIEE 2017. 4th Competition on Legal Information Extraction
and Entailment (EPiC Series in Computing, Vol. 47), Ken Satoh, Mi-Young Kim,
Yoshinobu Kano, Randy Goebel, and Tiago Oliveira (Eds.). EasyChair, 1–8.
[4] Mi-Young Kim, Randy Goebel, Yoshinobu Kano, and Ken Satoh. 2016. COLIEE-
2016: Evaluation of the Competition on Legal Information Extraction and Entail-
ment. In The Proceedings of the 10th International Workshop on Juris-Informatics
(JURISIN2016). Paper 11.
[5] Mi-Young Kim, Ying Xu, and Randy Goebel. 2017. Applying a Convolutional
Neural Network to Legal Question Answering. In New Frontiers in Artificial
Intelligence, Mihoko Otake, Setsuya Kurahashi, Yuiko Ota, Ken Satoh, and Daisuke
Bekki (Eds.). Springer International Publishing, Cham, 282–294.
[6] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. 2013.
Distributed representations of words and phrases and their compositionality. In
Advances in neural information processing systems. 3111–3119.
[7] George A. Miller. 1995. WordNet: A Lexical Database for English. Commun. ACM
38, 11 (1995), 39–41.
[8] Junghyun Min, R. Thomas McCoy, Dipanjan Das, Emily Pitler, and Tal Linzen.
2020. Syntactic Data Augmentation Increases Robustness to Inference Heuristics.
In Proceedings of the 58th Annual Meeting of the Association for Computational
284
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
Linguistics. Association for Computational Linguistics, Online, 2339–2352. https:
//doi.org/10.18653/v1/2020.acl-main.212
[9] Ha-Thanh Nguyen, Hai-Yen Thi Vuong, Phuong Minh Nguyen, Binh Tran Dang,
Quan Minh Bui, Sinh Trong Vu, Chau Minh Nguyen, Vu Tran, Ken Satoh, and
Minh Le Nguyen. 2020. JNLP Team: Deep Learning for Legal Processing. In The
Proceedings of the 14th International Workshop on Juris-Informatics (JURISIN2020).
The Japanese Society of Artificial Intelligence„ 195–208.
[10] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2020. COLIEE2020:Methods for Legal Document Retrieval
and Entailment. In The Proceedings of the 14th International Workshop on Juris-
Informatics (JURISIN2020). The Japanese Society of Artificial Intelligence„ 114–
127.
[11] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2020. A Summary of the COLIEE 2019 Competition. In
New Frontiers in Artificial Intelligence, Maki Sakamoto, Naoaki Okazaki, Koji
Mineshima, and Ken Satoh (Eds.). Springer International Publishing, Cham, 34–
49.
[12] Hsuan-Lei Shao, Yi-Chia Chen, and Sieh-Chuen Huang. 2020. BERT-based Ensem-
ble Model for The Statute Law Retrieval and Legal Information Entailment. In The
Proceedings of the 14th International Workshop on Juris-Informatics (JURISIN2020).
The Japanese Society of Artificial Intelligence„ 223–234.
[13] Connor Shorten and T. Khoshgoftaar. 2019. A survey on Image Data Aug-
mentation for Deep Learning. Journal of Big Data 6 (2019), 1–48. https:
//doi.org/10.1186/s40537-019-0197-0
[14] Ryosuke Taniguchi, Reina Hoshino, and Yoshinobu Kano. 2019. Legal Question
Answering System Using FrameNet. In New Frontiers in Artificial Intelligence,
Kazuhiro Kojima, Maki Sakamoto, Koji Mineshima, and Ken Satoh (Eds.). Springer
International Publishing, Cham, 193–206.
[15] Masaharu Yoshioka, Yoshinobu Kano, Naoki Kiyota, and Ken Satoh. 2018.
Overview of Japanese Statute Law Retrieval and Entailment Task at COLIEE-
2018. In The Proceedings of the 12th International Workshop on Juris-Informatics
(JURISIN2018). The Japanese Society of Artificial Intelligence„ 117–128.
 Legal Norm Retrieval with Variations of the BERT Model
Combined with TF-IDF Vectorization
Sabine Wehnert Viju Sudhi
sabine.wehnert@gei.de Shipra Dureja
Georg Eckert Institute Libin Kutty
Leibniz Institute for International
Textbook Research
Saijal Shahania
<firstname>.<lastname>@st.ovgu.de
Germany
Otto von Guericke University
Otto von Guericke University
Magdeburg, Germany
Magdeburg, Germany
ABSTRACT
In this work, we examine variations of the BERT model on the
statute law retrieval task of the COLIEE competition. This includes
approaches to leverage BERT’s contextual word embeddings, fine-
tuning the model, combining it with TF-IDF vectorization, adding
external knowledge to the statutes and data augmentation. Our
ensemble of Sentence-BERT with two different TF-IDF representa-
tions and document enrichment exhibits the best performance on
this task regarding the F2 score. This is followed by a fine-tuned
LEGAL-BERT with TF-IDF and data augmentation and our third
approach with the BERTScore. As a result, we show that there
are significant differences between the chosen BERT approaches
and discuss several design decisions in the context of statute law
retrieval.
CCS CONCEPTS
• Applied computing → Law; • Information systems → Docu-
ment representation; Language models; Similarity measures; Rel-
evance assessment; • Computing methodologies → Neural net-
works.
KEYWORDS
contextual word embeddings, document enrichment, data augmen-
tation, legal information retrieval
ACM Reference Format:
Sabine Wehnert, Viju Sudhi, Shipra Dureja, Libin Kutty, Saijal Shahania,
and Ernesto W. De Luca. 2021. Legal Norm Retrieval with Variations of the
BERT Model Combined with TF-IDF Vectorization. In ICAIL ’21: Interna-
tional Conference on Artificial Intelligence and Law, June 21–15, 2021, online.
ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3462757.3466104
1 INTRODUCTION
In this paper, we describe our approach for the retrieval task 3 of
the COLIEE competition, based on the English version of the Japan-
ese Civil Code. Legal Statute Retrieval is a challenging task due to
the short, abstract nature of law articles and at times very specific
Publication rights licensed to ACM. ACM acknowledges that this contribution was
authored or co-authored by an employee, contractor or affiliate of a national govern-
ment. As such, the Government retains a nonexclusive, royalty-free right to publish or
reproduce this article, or to allow others to do so, for Government purposes only.
ICAIL ’21, June 21–15, 2021, online
© 2021 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466104
285
Ernesto W. De Luca
deluca@gei.de
Georg Eckert Institute
Leibniz Institute for International
Textbook Research
Germany
Otto von Guericke University
Magdeburg, Germany
scenarios described in a query. Having a requirement of high recall
and reasonable precision values at the same time, most state-of-
the-art retrieval approaches are not reliable enough for real-life
scenarios. Since their first use in the Competition on Legal Informa-
tion Extraction/Entailment (COLIEE), BERT (Bidirectional Encoder
Representations from Transformers) approaches received criticism
regarding their explainability compared to other traditional ma-
chine learning methods that have been employed. A term frequency
- inverse document frequency (TF-IDF) document representation
may appear more reliable to a legal practitioner, since ranking based
on term statistics and its side effects are interpretable. Neverthe-
less the ongoing success of BERT-based methods may justify their
continued use. A big part in the performance of BERT models is
attributed to the rigorous pre-training on huge corpora and the
large model capacity, for example with 768 embedding dimensions
in the base model. A pre-trained language model has encountered
a big diversity of text and has been therefore exposed to enough
examples to form a decent representation for the contextual use
of words. Hence, substantial knowledge from a distributional se-
mantics point of view appears to be helpful in this task. However,
using a standard BERT model on the plain training data is not suffi-
cient. The COLIEE data for task 3 is quite limited in its size due to
massive human effort behind its creation. This poses a challenge in
the training of deep learning models in addition to the legal jargon,
which differs from the language the models may be pre-trained on.
We investigate whether data manipulations such as augmentation
and the decomposition of relevant articles helps in this issue.
In previous COLIEE editions, machine learning models have also
benefited from other types of external knowledge, for example by
using ontologies for information extraction. This motivates us to
also enrich the training data and to encode the additional content
jointly with the original text by using a BERT model. In addition,
several BERT approaches have emerged, with variants specializing
on our domain, such as LEGAL-BERT [3]. Aside from the model
selection, we can also choose between using the BERT model in a
supervised setting as a relevance classifier or to extract contextual
word embeddings and use them directly for similarity scoring.
We use our three runs in the competition to test a few BERT
variations based on those considerations. The contributions of our
three runs are:
• We combine Sentence-BERT with modifications on TF-IDF
vectorization and document enrichment strategies
ICAIL ’21, June 21–15, 2021, online
• We perform dataset manipulations to train a BERT classifier
for retrieval and combine it with TF-IDF vectorization
• We test similarity scores obtained from the BERTScore
The remainder of this work is organized in the following way:
In Section 2, we describe approaches using TF-IDF, BERT models
and their various use in the past editions of the COLIEE compe-
tition. Section 3 contains conceptual descriptions for each of our
three runs: Sentence-BERT Embedding with TF-IDF, LEGAL-BERT
with TF-IDF, and BERTScore. Section 4 consists of more details on
our experimental setup, results and a following discussion. In the
final section we conclude our results and indicate future research
potential.
2 RELATED WORK
In this section, we describe related research of retrieval methods we
used. In particular, we investigate past uses of the respective method
within the COLIEE competition. First, we briefly review the TF-IDF
(term frequency - inverse document frequency) vectorization and
its place within the competition. Second, we collect approaches
which are similar to our methods which are using BERT (Bidirec-
tional Encoder Representations from Transformers) and also make
a distinction between our methods for the three runs we submitted
and the existing work.
2.1 Retrieval with TF-IDF
TF-IDF vectorization gives an idea of how relevant a particular term
is within a document and within the document collection. TF-IDF
vectors represent a document by assigning a higher weight to terms
which appear relatively frequent in few documents - compared to
their usual occurrence in the rest of the corpus - by discounting
the term frequency with the inverse document frequency. As Beel
et al. [1] comment, the TF-IDF vectorization scheme is the most
widely used approach for content-based filtering for recommender
systems and related text mining domains. In the COLIEE competi-
tion multiple teams in the previous years used TF-IDF vectors with
or without other representation methods to retrieve the relevant
articles given a query [8, 10]. In legal information retrieval, TF-IDF
only is still a valuable baseline model because its results are easy to
interpret for domain experts. However, in previous editions of the
competition, a mere TF-IDF approach could not reliably achieve
winning scores. When used in conjugation with any other em-
bedding techniques, competitive results were attainable. One such
approach has been employed by Rabelo et al. [9] to address the case
law entailment task. They employ two different cosine similarity
approaches and a confidence score from BERT [4] to improve the
extraction/entailment results. We adopt a similar approach in our
second run by combining TF-IDF similarity scores with the softmax
scores obtained from fine-tuned BERT models. However, we also
differ in the way of choosing documents to calculate similarity and
in thresholding for the retrieval task.
2.2 Retrieval with BERT
Nowadays, many pre-trained deep learning-based language models
are available, coming from neural network architectures for Natu-
ral Language Processing (NLP) with significant improvements for
various downstream tasks, such as single sentence classification,
286
Wehnert et al.
question answering tasks, sentence tagging tasks and paraphrase
identification. BERT introduced by Devlin et al. [4] is currently a
common choice for such downstream tasks, replacing various tradi-
tional NLP pipelines. Following this, there has been an exhaustive
study about the applications of BERT and experiments to inves-
tigate different fine-tuning methods for these pre-trained models
by Sun et al. [12]. They present various fine-tuning strategies for
BERT on a text classification task, providing a general solution
for achieving state-of-the-art results on a variety of text classifi-
cation datasets. We follow a few of these best practices for better
performance with our selected models, such as:
(1) the use of the right combination of different hyperparameters
that directly affect the learning,
(2) the importance of the selection of the correct value of warm-
up steps,
(3) how concentrating on the decay rate can help to converge
towards the minima and when the learning rate decay should
start, and
(4) the right combination of batch-size with the number of
epochs and warm-up steps.
2.2.1 Fine-tuning BERT. When Devlin et al. proposed the BERT
model, they described its use on downstream tasks in two phases:
a pre-training and a fine-tuning phase [4]. Therefore, its intended
use for any further task is to first fine-tune it in order to achieve the
desired performance. Nowadays, BERT is not always fine-tuned,
sometimes the pre-trained model and its embeddings perform well
enough, if the domain is not substantially changed compared to
what the model was pre-trained on. However, for the legal domain it
can be worthwhile to adapt an existing BERT model to the different
use of vocabulary in that context. This can be also observed in the
past COLIEE competitions. For the task on statute law retrieval,
Nguyen et al. [8] use an ensemble of BERT models. The publicly
available bert-base-uncased 1 model is pre-trained on the English
language Wikipedia and BooksCorpus [14] and then fine-tuned by
Nguyen et al. on the COLIEE training data. The model is combined
with another special bert-base-uncased model that is further trained
with the masked language model (MLM) on the entire COLIEE data
(BERT-CC) and fine-tuned on training data to obtain a measure
of relevance. This ensemble of BERT achieved the best F2 score
for the validation data. As their BERT-CC focuses on legal domain
knowledge, we reviewed further special BERT models. We found
RoBERTa (Robustly Optimized BERT pre-training Approach) [7]
with its variants, and LEGAL-BERT [3] as promising models for
task 3. RoBERTa [7] is optimized with some alterations to essential
hyperparameters in BERT and trained with relatively bigger batches
over a large training data size. It also excludes BERT’s next-sentence
prediction task, allowing it to improve on MLM over BERT. This
leads to a better performance on various baseline NLP downstream
tasks [7]. Similarly, LEGAL-BERT [3] is an adaption of BERT in the
legal domain where pre-training is carried out on a collection of
several fields of English legal text, such as contracts, court cases, and
legislation. This special BERT model has been performing better
than the original version of BERT on legal domain-specific tasks [3].
1 https://huggingface.co/bert-base-uncased
Legal Norm Retrieval with Variations of the BERT Model Combined with TF-IDF Vectorization
2.2.2 Contextual Embeddings from BERT. Aside from further train-
ing the whole BERT language model and using it on a classification
task, we can also use contextual word embeddings from a pre-
trained BERT model to determine semantic similarity of the query
and the article(s). Contextual word embeddings are computed at
runtime. In particular, we obtain different vectors for the same
word, when it is used in another context or position in a sentence.
In that way, we can also distinguish homonyms when they are
accompanied by enough words in the appropriate context. Since
the contextual embedding type is quite recent, there is no final
consensus in the research community of how to compute the dis-
tance between two contextual word embedding sequences. The
most common methods are: using the [CLS] token which is often
seen as a representation of a whole sentence, using the individual
word embeddings or averaging all individual word embeddings in
a sentence and then computing a similarity score using the Word
Mover’s Distance [6] or cosine similarity. In the experiments by
Reimers et al. [11], using the mean of the individual contextual
word embeddings outperformed the approach with the [CLS] token.
A recent approach related to this is the BERTScore [13]. After com-
puting the pairwise cosine similarity of all token-wise contextual
embeddings from two sentences, BERTScore selects token pairs
between the two sentences which have the highest cosine similarity.
Those similarities are summed up and discounted by the words in
the sentence to obtain precision, recall and the according F1-score.
Optionally, the BERTScore can also incorporate IDF weighting. We
employ the BERTScore in our third run to test whether the mere
embeddings of BERT can also capture enough context in the train-
ing data, compared to using document enrichment or fine-tuning
on a BERT model for relevance classification.
For us, it is particularly interesting that there are recent ap-
proaches for fine-tuning a language model specifically to obtain
meaningful sentence embeddings [2, 11]. In the previous COLIEE
edition, the cyber team achieved the best performance among all
teams using the universal sentence encoder, TF-IDF and a support
vector machine for the case law retrieval task [10]. Hence, we as-
sume that TF-IDF combined with sentence embeddings could also
work well on the statute law retrieval task. A new advancement
on sentence embeddings has been made by Reimers et al. [11] who
introduce Sentence-BERT. It outperforms the existing embedding
methods and is found useful for multiple downstream tasks. It is
based on a Siamese network architecture which ties the weights of
two BERT models (one for each input sentence) that are updated
during fine-tuning. As a default, the mean is used to pool the ob-
tained contextual word embeddings from each BERT model. Then,
the two resulting sentence embeddings are concatenated with their
element-wise difference, so that the final softmax layer can pre-
dict a class. We have made use of this state-of-the-art embedding
approach to create a richer and more meaningful numeric represen-
tation of each article and query pair in our first run. In a previous
COLIEE edition, Kim et al. [5] have employed a Siamese Deep Con-
volutional Neural Network for the entailment task, which results
in better performance compared to regular Convolutional Neural
Networks. They attribute their success to the Siamese architecture
which requires less parameters due to the weight sharing mecha-
nism and a lower risk of overfitting. For this reason, we assume
that a sentence embedding based on a similar architecture may be a
287
ICAIL ’21, June 21–15, 2021, online
Table 1: Methods for each run for task 3
Run Name Method
OvGU_run1 Sentence-BERT + TF-IDF + data enrichment
OvGU_run2 LEGAL-BERT + TF-IDF + data augmentation
OvGU_run3 BERTScore
good fit for the COLIEE data and also perform well on the retrieval
task.
Overall, TF-IDF vectorization and BERT-based approaches have
already been tried in course of the past COLIEE editions. Never-
theless, there are many options to employ both methods, while
fine-tuning can affect the outcome substantially.
3 STATUTE RETRIEVAL TASK
This section describes in detail the three different methods we
proposed and implemented for task 3 in COLIEE 2021, as mentioned
in Table 1. While the first method exploits Sentence-BERT coupled
with TF-IDF vectors and data enrichment, the second method uses
LEGAL-BERT with TF-IDF vectors and data augmentation. The
third method applies the BERTScore to solve the problem at hand.
3.1 Sentence-BERT Embedding with TF-IDF
The first run involves a combination of 2-stage TF-IDF vectorization
with Sentence-BERT embeddings. This run was the best out of all
the runs submitted for task 3 in COLIEE 2021. An overview of the
approach is depicted in Figure 1 and described in the following.
We start by enriching the training data with multiple adjustments
as described in Table 2. This enrichment helps us to create vectors
for each article in the Civil Code which are more unique than those
the training data itself could deliver. A concrete example of the
enrichment process for Article 177 can be found in Table 3. We
enrich each article in the training data as follows:
• Metadata: We add structural information using the section
titles in the Civil Code. In that way, hierarchical relations
between articles within the same Part, Chapter, Section and
even Subsection are modeled.
• Crawled data: We crawl Japanese open source commentary
on the Civil Code articles and thereby potentially enrich
the original article text with general remarks, corner cases,
previous versions, related articles and a reasoning for the
relation.
• Relevant queries from training data: We parse the train-
ing data labels of task 4 (entailment) to enrich our training
data of task 3 with queries that have a positive entailment
relationship. With a positive entailment relationship we can
be sure that the queries added correspond to the meaning of
the article and can help in determining relevance, too.
After data enrichment, we encode the enriched texts with the
TF-IDF vectorizers and the Tokenizer2 for our Sentence-BERT and
progress to the final relevance score calculation with the following
steps:
2 https://huggingface.co/distilroberta-base
ICAIL ’21, June 21–15, 2021, online Wehnert et al.

𝑣𝑎

TF-IDF Stage 1
Article with metadata +
Article with relevant queries
cos

𝑣𝑞

𝑣𝑎
Query
TF-IDF Stage 2
Article with metadata Í
+ cos Normalization Thresholding

𝑣𝑞
Article(s)
𝑣𝑎
𝑣 𝑎 - article vectors (for all articles)
Sentence Embedding 𝑣𝑞 - article vectors (for a given query)
Article with metadata +
cos Í
Article with relevant queries + cosine
cos sum
Article with crawled queries similarity
𝑣𝑞

Figure 1: Overview of the approach using Sentence-BERT embeddings with TF-IDF.

Table 2: Data enrichment for the statute retrieval task - 𝑡 𝑓𝑡,𝑑 is the term frequency - frequency of term 𝑡 in document
𝑑. Here, documents are the individual articles of the Civil
Description Code.
- 𝑁 is the total number of documents in the collection.
Articles with metadata training data + details
- 𝑑 𝑓𝑡 is the document frequency - frequency of term 𝑡 in the
regarding Part, Chapter, Section
collection.
and Subsection.
- 𝑤𝑡,𝑑 is the weight which is the product of term frequency
Articles with crawled data training data + translated
and inverse document frequency.
crawled data from the website
The vectors after a single stage of TF-IDF vectorization
https://ja.wikibooks.org/
yielded significant precision-recall trade-offs reflected in
Articles with relevant queries training data + queries from
the relatively lower F2 scores. This prompted us to provide
training data if the entailment
a different, but unique representation of the articles, which
label is Y
ended up in a second stage of TF-IDF where query and article
for the respective article.
vectors have been created considering only the Articles with
metadata enrichment. The combination of both stages acts
as a counter-balance in the trade-off.
(1) TF-IDF vectors are computed for queries and articles to- (2) Sentence-BERT embeddings for each article are created with
gether as a two-stage process. In the first stage, we rely the enrichment described in Table 2. We rely on the imple-
on sub-linear term frequency scaling and L2 normalization mentation3 by Reimers et al. [11] and use the pre-trained
while computing the vectors. Articles are enriched by a com- paraphrase-distilroberta-base-v1 model to create the article
bination of Articles with relevant queries and Articles with and query embeddings. We select the aforementioned para-
metadata. phrase model because it was trained on millions of para-
⃗⃗
The vectors 𝒗 are computed by the following equations 1 - 4, phrase examples and is reportedly performing well on natu-
ral language inference tasks4 .
𝑡 𝑓𝑡,𝑑 = (1 + 𝑙𝑜𝑔(𝑡 𝑓𝑡,𝑑 )) (1) (3) Finally, for each query-article pair we compute the cosine
similarity to determine the relevance of each article for the
𝑁 respective query. For each pair, we obtain three different
𝑖𝑑 𝑓𝑡 = 𝑙𝑜𝑔( ) (2)
1 + 𝑑 𝑓𝑡 similarity scores from the first stage TF-IDF, second stage
TF-IDF and Sentence-BERT embeddings. The sum of these
𝑤𝑡,𝑑 = 𝑡 𝑓𝑡,𝑑 ∗ 𝑖𝑑 𝑓𝑡 (3) scores is then normalized and we empirically determine a
threshold to filter out the best relevant articles for each test
⃗⃗⃗ query.
⃗⃗ 𝒘
𝒗 = qÍ (4)
2
𝑖 𝑤𝑖,𝑑
3 https://github.com/UKPLab/sentence-transformers

where, 4 https://www.sbert.net/docs/pre-trained_models.html#paraphrase-identification

288
Legal Norm Retrieval with Variations of the BERT Model Combined with TF-IDF Vectorization ICAIL ’21, June 21–15, 2021, online

Table 3: Example data enrichment for Article 177 of the Civil Code

training data
Article 177: Acquisitions of, losses of and changes in real rights on immovables .. and other laws regarding registration.
Metadata
Part: II Real Rights Chapter: I General Provisions Section: 3 Extinctive Prescription
Subsection: Requirements of Perfection of Changes in Real Rights on Immovables
Crawled data
Comprehensive succession - The range of changes in property rights that require registration has been determined ..
Legal evidence theory - What kind of person is referred to as "a person who has a legitimate interest ..
.. (161 unique words in total, shortened to conserve space)
Relevant queries from training data
– H19-11-3: In a case where A bought a registered building owned by B .. his/her acquisition of ownership of that building.
– H21-24-E: If a mortgage creation contract has the agreement of the mortgagee .. there is no registration of its creation.
– R01-6-A: In cases A sold Land X belonging to A and B sold it to C, C may be asserted .. for sales without security.

For developing an explanatory dialogue in a real setting, the
additional text we gained in the enrichment steps can be marked in
a different font style. Then, we can highlight important keywords
based on the scores of each TF-IDF stage. Since we did not apply any
weighting during the cosine similarity computation of the Sentence-
BERT embeddings, the similarity between the word vectors of query
and article can be visualized using a heatmap.
3.2 LEGAL-BERT with TF-IDF
For our run 2, we treat task 3 as a sentence-pair classification task to
predict the relevance of 1 if the given article is related to the query
and 0 otherwise. Considering its good performance on previous
retrieval tasks, we choose to work with a BERT model. A variety of
BERT models that are pre-trained on different datasets can be used
for addressing domain-specific tasks with fine-tuning.
3.2.1 BERT configuration. Following this convention with fine-
tuning, we initially used bert-base-uncased which has 12 hidden
layers with 768 hidden units in each layer 12 attention heads. A
classification head is added on top of the base model consisting of a
single layer of fully connected linear neurons. We use the softmax
function to get a probability distribution for the two labels and use
cross-entropy loss with Adam optimizer to fine-tune the model.
We split the training dataset into two parts for fine-tuning (∼ 85 %
training) the model and use the rest of the dataset for validation
(all queries starting with id "R01-*").
3.2.2 Data Pre-processing. An overview of the pre-processing for
the LEGAL-BERT with TF-IDF approach is illustrated in Figure
2. We pre-process both the training and validation splits in the
following manner:
(1) Data Decomposition: This is performed to extract each
relevant article for a given query to form separate instances.
For every query, there is one or more than one article associ-
ated and relevant to it. We take individual articles to create
a new instance in the training dataset so that the query can
be divided into multiple instances against all of its relevant
articles. An example is shown in the Table 4 for the query
with the Pair ID "H27-22-4":
Query Q: "In the contract for deposit for value, if the perfor-
mance of the obligation to return deposited Thing has become
impossible due to reasons not attributable to the depositary,
he/she may not claim remuneration from the depositor, with
respect to the period after the impossibility of performance of
the agreed duration."
After achieving better results with data decomposition than
with the original dataset, we further extract referenced arti-
cles from each relevant article of the query using regular ex-
pressions and append that as well to form multiple instances
of query-article pairs for each query. The same example is
extended further for Approach 2 in Table 5.
However, this extensive decomposition of referenced arti-
cles did not optimize our recall further. We assume this is
plausible as these articles are supporting articles to the rele-
vant article content but are not directly relevant to the query.
We compare the results with and without data decomposi-
tion and summarize them in Table 6. We decided to go with
Approach 1 where we have a better recall score.
(2) Data Augmentation: We use the non-relevant articles to
reduce data imbalance. For this, we enriched this decom-
posed dataset using the top 50 non-relevant articles for each
query instance. These non-relevant articles are based on the
highest cosine similarity between TF-IDF vectors of the rel-
evant article to all the articles excluding the other relevant
ones for the respective query. This approach is similar to the
implementation by Nguyen et al. [8], where they considered
query-article similarity. However, we assume that article-
article similarity is better suited than query-article similarity
since we find that articles are more related to each other
in terms of cosine similarity than they are to the queries.
Based on the cosine similarity, we select only the top 50
non-relevant articles as training examples, since we did not
intend to reintroduce the data imbalance that we attempted
to overcome with augmentation.

289
ICAIL ’21, June 21–15, 2021, online Wehnert et al.

Table 4: Approach 1 - Data decomposition of multiple articles for each query into multiple instances

Queries Articles
Before Pre-processing
Query Q Article 665 The provisions of Articles 646 through 648, Article 649, and Article 650, paragraphs ...
Article 648 (1) In the absence of any ... (2) ... the provisions of Article 624 ... (3) ... course of performance.
Article 536 (1) If the performance ... (2) ... obligee for the benefit.
After Pre-processing
Query Q Article 665 The provisions of Articles 646 through 648, Article 649, and Article 650, paragraphs ...
Query Q Article 648 (1) In the absence of any ... (2) ... the provisions of Article 624 ... (3) ... course of performance.
Query Q Article 536 (1) If the performance ... (2) ... obligee for the benefit.

Table 5: Approach 2 - Data decomposition of multiple articles and their referenced articles for each query into multiple in-
stances

Queries Articles
Before Pre-processing
Query Q Article 665 The provisions of Articles 646 through 648, Article 649, and Article 650, paragraphs ...
Article 648 (1) In the absence of any ... (2) ... the provisions of Article 624 ... (3) ... course of performance.
Article 536 (1) If the performance ... (2) ... obligee for the benefit.
After Pre-processing
Query Q Article 665 The provisions of Articles 646 through 648, Article 649, and Article 650, paragraphs ...
Query Q Article 646 (1) A mandatary must deliver to the mandator monies and other things received during ...
Query Q Article 647 If the mandatary has personally consumed monies that were to be delivered to the mandator ...
Query Q Article 648 (1) In the absence of any special agreements, the mandatary may not claim remuneration ...
Query Q Article 649 If costs will be incurred in administering the mandated business, the mandator must ...
Query Q Article 650 (1) If the mandatary has expended costs found to be necessary for the administration ...
Query Q Article 624 (1) An employee may not demand remuneration until the work the employee promised ...
Query Q Article 536 (1) If the performance ... (2) ... obligee for the benefit.

Table 6: Results on the validation set for different data pre-
processing approaches of run 2
Model Prec Recall
bert-base-uncased without decomposition 0.1392 0.3973
bert-base-uncased with Approach 1 0.2529 0.4421
bert-base-uncased with Approach 2 0.1179 0.4300
(3) Augmenting the Original Dataset: To ensure that the
original data could still influence the model, we also append
the original data. In other words, for each query without any
data decomposition, all relevant articles are processed in an
instance as they are given in the dataset. This increases the
number of relevant articles for each query at the cost of gen-
erating some duplicates, since for queries which have only
one relevant article, those are already obtained at the step of
data decomposition. Overall, the three pre-processing steps
increase the number of training instances by the factor 10.
3.2.3 Fine-Tuning. On comparing the results with the legal domain-
specific pre-trained BERT, bert-base-uncased was outperformed by
legal-bert-base-uncased and legal-RoBERTa on similar hyperparam-
eters. We finally choose legal-bert-base-uncased as it indicated the
most satisfactory results to further test with different experimen-
tal setups in Section 4.1. To extract relevant articles for a given
query during testing, we combine each query with all the articles.
For LEGAL-BERT, we applied the softmax function to the logits
predicted from our model. For each query-article pair, we obtain
two softmax probability values, indicating the non-relevance and
relevance of the article to the query. We only consider the softmax
probabilities of the relevance column. To avoid the underflow of
softmax probabilities of top relevant articles, we max-normalize
these scores. At the same time, we also calculate the query-article
cosine similarity of all the articles for each query. The similarity
scores are also max-normalized for the same reasons as stated above.
We ultimately compute an average of these two normalized scores.
To select the top-n relevant articles we use a threshold value se-
lected based on the precision-recall trade-off for the validation set.
The time for training the LEGAL-BERT model increases from 2
minutes on the original dataset to 2 hours on the fully enriched
dataset5 . The larger amount of text in the enriched data does not
have a significant impact during the test phase. At runtime, we
directly process the new query and all pre-stored enriched articles
5 We used an NVIDIA Quadro RTX 8000 to accelerate training.

290
Legal Norm Retrieval with Variations of the BERT Model Combined with TF-IDF Vectorization ICAIL ’21, June 21–15, 2021, online

Query
Training
Split
Original +
Pair
Relevant
Article(s) TF-IDF LEGAL-BERT

cos 𝜎
Query
Max Max
Normalization Normalization
Copies

Decomposed Training cos cosine

+ similarity
Pairs Split 𝜇
Individual
Articles 𝜎 softmax for
relevant class
Thresholding 𝜇 arithmetic
Relevant mean
Article(s)

Figure 3: Overview of the approach using LEGAL-BERT with

TF-IDF.
Query
+

50 Training
Non-Relevant Data
Augmentation Article(s)
Pairs
cos
BERTScore

Corpus of Individual
all Articles Article
F1

cos cosine Thresholding F1 F1 Score

similarity

Figure 2: Pre-processing for LEGAL-BERT with TF-IDF. Figure 4: Overview of the approach using BERTScore.

(3) For the test data, we take the average K of all the BERTScore
with the already trained language model, so that the prediction is
thresholds from the training data.
not causing any noticeable delay in the system’s response time.

3.3 BERTScore 4 EVALUATION

In this section, we evaluate the previously presented methods. First,
In the third run of the retrieval task, we use the BERTScore [13].
we describe details of the experimental setup. Second, we proceed
In Zhang et al.’s implementation6 , BERTScore outputs precision,
to show the respective results on the competition task. Third, we
recall, and F1 measure. We use the F1 score as the main value for
discuss our findings. We evaluate our runs for the statute retrieval
further analysis. To decide how many articles should be retrieved
task on our validation split using variations in hyperparameters for
per query, the following steps are used to determine the BERTScore
training different models. The evaluation contributes a quantitative
threshold (K):
and qualitative analysis of how the runs perform with different
(1) We calculate the BERTScore for each article given a query, hyperparameter settings - if applicable - and if they are comparable
and then rank the result set in descending order. to each other. We discuss a few of the experiments below.
(2) For each query in the training data we select the top n doc-
uments, with a BERTScore value for each n. From this, we 4.1 Experimental Setup
select the BERTScore value K as a threshold where the F2
4.1.1 Sentence-BERT Embedding with TF-IDF. To enrich the articles
score is maximized, since the task performance is evaluated
further, we make use of the crawled content7 . We extract all the
on the F2 score.
7 https://ja.wikibooks.org/wiki/民法第<id>条 , where <id> stands for the Article ID of
6 https://github.com/Tiiiger/bert_score the different articles in the Civil Code.

291
ICAIL ’21, June 21–15, 2021, online Wehnert et al.

Table 7: Two stages of TF-IDF counter-balancing the Table 8: Results on validation set for run 2 candidates
precision-recall trade-off with Sentence-BERT.
Model Prec Recall
F2 Prec Recall
bert-base-uncased 0.2529 0.4421
Validation data legal-bert-base-uncased 0.3447 0.5357
legal-RoBERTa 0.2205 0.4866
with 1st stage TF-IDF 54.67 50.16 61.98
with 2nd stage TF-IDF 53.74 49.54 60.27
with both stages 56.52 52.60 63.39 Table 9: Task 3 Results for COLIEE 2021
COLIEE 2021 test data
with 1st stage TF-IDF 72.98 66.77 78.40 Position Run F2 Prec Recall R_30
with 2nd stage TF-IDF 73.02 66.28 79.63 1 OvGU_run1 0.7302 0.6749 0.7778 0.8515
with both stages 73.02 67.49 77.78 9 OvGU_run2 0.6717 0.4857 0.8025 0.9010
18 OvGU_run3 0.3016 0.1570 0.7006 0.7030

paragraph tags (<p>) and the list tags (<ol>, <ul>, <dl>, <li>) to get
relevant information about the articles. This is motivated by the
team TRC3 in the previous year of COLIEE [10], where they used the
content in Japanese itself. However, we translate the fetched content
to English using the google-trans-new package 8 . To vectorize these
enriched articles and queries we used the TfidfVectorizer from
scikit-learn.
To address the problem of the precision-recall trade-off, we use
two-stages of TF-IDF, which is motivated from previous experi-
ments we conducted on queries starting with the id "R01-*", as
shown in Table 7. It is evident on the validation data that the two-
stage TF-IDF can counter-balance the classical trade-off between
precision and recall, considering the improved F2 scores. For the
COLIEE 2021 test data the second stage has a positive effect on the
F2 score as well, though it is not as significant as we found it for
our validation split.
The threshold value to filter out the top n relevant articles was
found empirically. After normalizing the sum of the scores from the
two stages of TF-IDF and Sentence-BERT embeddings, we consid-
ered the top 4 articles. This was purely based on our validation data,
where none of the queries had relevant articles exceeding a count of
four. This is true with COLIEE 2021 test data as well, where none of
the articles have more than 4 relevant articles. To find a threshold
for the scores of these articles, an index-based threshold was found
to be better than a single value for the whole set. Accordingly, we
take the article in the 1st index (with a score of 1.0) and then set a
threshold of 0.91 or higher for the articles if found in the 2nd index
and a threshold of 0.85 or above if found in the subsequent indices.
4.1.2 LEGAL-BERT with TF-IDF. To decide among the three al-
ternative models that we selected as candidates for our run 2 as
discussed in Section 3.2, we validate them on various hyperpa-
rameter settings and observe that the default hyperparameters
of the Adam Optimizer with a selective change in the learning
rate ranging from 1e−03 to 1e−06 , 1e−05 achieve the best results
among all three models (see Table 8) when trained on 3 epochs
for batch-size 16. Considering the highest recall score, we select
legal-bert-base-uncased to be our final choice for run 2.
8 https://github.com/lushan88a/google_trans_new
Further, we experiment with warm-up steps, introduce a de-
cay rate and did some hyperparameter tuning to optimize our re-
sults. We notice that with 3500 warm-up steps and a decay rate of
0.1 (1+𝑒𝑝𝑜𝑐ℎ) , we achieve the best performance. We then perform
further training on the validation set with an increased batch size
of 24. We then create an ensemble of legal-bert-base-uncased and
TF-IDF vectors, both with max-normalized similarity scores for
the article-query pairs, assigning equal weights to both the scores.
Finally, we fetch the articles that are above the threshold value
of 0.5.
4.1.3 BERTScore. For the BERTScore, we use the model type bert-
base-uncased, 9 layers and no re-weighting with IDF. This setup
was determined based on the performance on our validation data
which we also used before (queries starting with the id "R01-*"). The
text is processed with the regular Tokenizer of BERT and we pass
query and article(s) without further modification to the scorer of
the original BERTScore implementation. Our thresholding strategy
for this run results in a threshold value of 0.63331205.
4.2 Results
Our first run, OvGU_run1 obtained the first position for its F2 score
in the overall task evaluation for COLIEE 2021. OvGU_run2 also
has the best recall sharing the position with the run
JNLP.CrossLMultiLThreshold, closely followed by OvGU_run1. While
considering Recall at 30, our runs have the third best (for OvGU_run2)
and the fifth best (for OvGU_run1) scores. The results for our runs
are summarized in Table 9. Values in bold are the best scores for
the corresponding metric.
4.3 Discussion
We assume that our first run provides reliable results because of
the combination of contextual Sentence-BERT embeddings with
the TF-IDF vectors. This is supported by the test query R02-1-A:
"The family court may decide to commence an assistance also in re-
spect of a person whose capacity to appreciate their own situation is
extremely inadequate due to a mental disorder.",
as shown in Table 10. For this query, only Sentence-BERT embed-
dings could retrieve the most relevant Article 15 which was not
retrieved in either stage of TF-IDF vectorization. The Article 15 has

292
Legal Norm Retrieval with Variations of the BERT Model Combined with TF-IDF Vectorization
Table 10: Comparison of results for query R02-1-A
Method Retrieved articles
With TF-IDF stages Article 11
With Sentence-BERT Article 15, Article 7, Article 11
With combination Article 11, Article 15
Relevant articles Article 15, Article 11
the following content:
"(Decisions for Commencement of Assistance)
Article 15 (1) The family court may decide to commence an as-
sistance in respect of a person whose capacity to appreciate
their own situation is inadequate due to a mental disorder,
at the request of the person in question, that person’s spouse, that
person’s relative within the fourth degree of kinship, the guardian,
the guardian’s supervisor, the curator, the curator’s supervisor, or a
public prosecutor; provided, however, that this does not apply to a
person with respect to whom there are grounds as prescribed in Article
7 or the main clause of Article 11. (2) The issuance of a decision for
commencement of assistance at the request of a person other than the
person in question requires the consent of the person in question. (3)
A decision for commencement of assistance must be made concurrent
with a decision as referred to in Article 17, paragraph (1) or a decision
as referred to in Article 876-9, paragraph (1)."
It turns out that for this query-article pair, we have a significant
term overlap, which may be diluted by the whole article length. In
that way, sentence-based approaches in general may work well for
this query. Only our run OvGU_run1 and TR_HB have a 100% F2
score for this query.
On comparing our different runs, we find interesting similarities
in the articles retrieved by each of them. This might possibly be
because of the common TF-IDF coupling in the first two runs. We
did not expect that embeddings from a pre-trained model (in run 1)
could give more or less comparable results with those from a model
further trained on the COLIEE dataset (in run 2).
Another insight from the results is how thresholding plays a
significant role in the retrieval task. For example, the test query
R02-24-U:
"A donor shall assume a duty to retain the subject matter exercising
care identical to that he/she exercises for his/her own property until
the completion of such delivery.",
retrieved only one relevant article with run 1 but both relevant
articles with run 2. This is described in Table 11. Drawing conclu-
sions from this query - out of the many similar queries, we are not
surprised to see the fine-tuned model of run 2 retrieve 74 candidate
articles and run 1 retrieving only 70 candidate articles from a total
of 101. This results in run 2 with the overall best recall of 0.8025.
With BERTScore, the interesting query to analyse is the test
query R02-17-I:
"In the case that D manifests the intention to release another obligor
(C) from the obligation to D, even if neither D nor B manifests a
particular intention, D may not claim the payment of 600,000 to
another obligor (A)."
We are able to retrieve 3 out of 4 articles (Article 439, 440 and
441) using this technique which was the highest number when
293
ICAIL ’21, June 21–15, 2021, online
Table 11: Comparison of results for query R02-24-U
Method Top retrieved articles
With Run 1 Article 563, Article 566, Article 567
With Run 2 Article 563, Article 565, Article 567,
Article 562
Relevant articles Article 562, Article 563
compared to other teams for this query. However, this result can
be attributed to the threshold we selected, with high recall and
lower precision. The ranking of the articles by the BERTScore is
only average even for this query, considering the Mean Average
Precision (MAP). The MAP score is only 0.0509 for run 3, while
run 1 gets 0.0299 and run 2 achieves 0.1250. The best MAP score
for this query with a value of 0.2309 was obtained by the team
JNLP with their run called JNLP.CrossLBertJP. For assessing the
final ranking performance of run 3, we can compare its MAP score
to other teams. Also here we observe that the BERTScore with a
MAP score of 0.5557 is the fourth-lowest performing run in the
competition, whereas our run 1 achieves 0.7496 and run 2 has the
highest overall MAP score among our runs of 0.7571. The best
MAP score of 0.7947 was achieved by the team JNLP with their run
JNLP.CrossLMultiLThreshold. This leads us to the conclusion that
the standard BERTScore without IDF-reweighting or any further
combined methods may not be sufficient to solve this task, at least
with the the query type distribution of this year’s test dataset. We
also observe how thresholding influences our F2 score in run 1, so
that our method scores higher than a run by the JNLP team which
has a better ranking performance.
From the results and discussion above, our main takeaways from
this COLIEE edition for task 3 are:
(1) Contextual embeddings can significantly enhance retrieval
performance when coupled with TF-IDF vectors.
(2) Adding external knowledge to the articles in the form of
structural information, entailed queries or definitions can
help to make them more unique.
(3) Data augmentation techniques are useful to train a BERT
classifier for a retrieval task.
(4) An intelligent or rather more effective thresholding mech-
anism should be devised to further improve precision and
maintain a decent F2 score.
5 CONCLUSION AND FUTURE WORK
In this work, we study variations of the language model BERT for
task 3 of the COLIEE competition on statute law retrieval. We find
a benefit in combining the BERT model with TF-IDF vectorization
and in working on a sentence level with contextual embeddings.
Furthermore, it is helpful to test different pre-trained models and
fine-tuning, as well as adding external knowledge and data aug-
mentation techniques. Our winning approach is an ensemble of
Sentence-BERT and two different TF-IDF representations with dif-
ferent extents of document enrichment. In the second run, we
fine-tune a BERT classifier for retrieval based on an augmented
dataset. The third run is similarity scoring using the BERTScore
ICAIL ’21, June 21–15, 2021, online Wehnert et al.

with thresholding. Future enhancements can consist of an improved IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile,
thresholding mechanism and of encoding other types of external December 7-13, 2015. IEEE Computer Society, 19–27. https://doi.org/10.1109/
ICCV.2015.11
knowledge, for example named entities.

REFERENCES
[1] Jöran Beel, Bela Gipp, Stefan Langer, and Corinna Breitinger. 2016. Research-
paper recommender systems: a literature survey. Int. J. Digit. Libr. 17, 4 (2016),
305–338. https://doi.org/10.1007/s00799-015-0156-0
[2] Daniel Cer, Yinfei Yang, Sheng-yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St
John, Noah Constant, Mario Guajardo-Cespedes, Steve Yuan, Chris Tar, et al.
2018. Universal Sentence Encoder for English. In Proceedings of the 2018 Con-
ference on Empirical Methods in Natural Language Processing, EMNLP 2018: Sys-
tem Demonstrations, Brussels, Belgium, October 31 - November 4, 2018, Eduardo
Blanco and Wei Lu (Eds.). Association for Computational Linguistics, 169–174.
https://doi.org/10.18653/v1/d18-2029
[3] Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras,
and Ion Androutsopoulos. 2020. LEGAL-BERT: The Muppets straight out of Law
School. CoRR abs/2010.02559 (2020). arXiv:2010.02559 https://arxiv.org/abs/2010.
02559
[4] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
Pre-training of Deep Bidirectional Transformers for Language Understanding. In
Proceedings of the 2019 Conference of the North American Chapter of the Associa-
tion for Computational Linguistics: Human Language Technologies, NAACL-HLT
2019, Minneapolis, MN, USA, June 2-7, 2019, Volume 1 (Long and Short Papers), Jill
Burstein, Christy Doran, and Thamar Solorio (Eds.). Association for Computa-
tional Linguistics, 4171–4186. https://doi.org/10.18653/v1/n19-1423
[5] Mi-Young Kim, Yao Lu, and Randy Goebel. 2017. Textual Entailment in Legal Bar
Exam Question Answering Using Deep Siamese Networks. In New Frontiers
in Artificial Intelligence - JSAI-isAI Workshops, JURISIN, SKL, AI-Biz, LENLS,
AAA, SCIDOCA, kNeXI, Tsukuba, Tokyo, Japan, November 13-15, 2017, Revised
Selected Papers (Lecture Notes in Computer Science, Vol. 10838), Sachiyo Arai,
Kazuhiro Kojima, Koji Mineshima, Daisuke Bekki, Ken Satoh, and Yuiko Ohta
(Eds.). Springer, 35–48. https://doi.org/10.1007/978-3-319-93794-6_3
[6] Matt J. Kusner, Yu Sun, Nicholas I. Kolkin, and Kilian Q. Weinberger. 2015. From
Word Embeddings To Document Distances. In Proceedings of the 32nd International
Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015 (JMLR
Workshop and Conference Proceedings, Vol. 37), Francis R. Bach and David M. Blei
(Eds.). JMLR.org, 957–966. http://proceedings.mlr.press/v37/kusnerb15.html
[7] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer
Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. RoBERTa: A
Robustly Optimized BERT Pretraining Approach. CoRR abs/1907.11692 (2019).
arXiv:1907.11692 http://arxiv.org/abs/1907.11692
[8] Ha-Thanh Nguyen, Hai-Yen Thi Vuong, Phuong Minh Nguyen, Tran Binh Dang,
Quan Minh Bui, Vu Trong Sinh, Chau Minh Nguyen, Vu D. Tran, Ken Satoh,
and Minh Le Nguyen. 2020. JNLP Team: Deep Learning for Legal Processing in
COLIEE 2020. CoRR abs/2011.08071 (2020). arXiv:2011.08071 https://arxiv.org/
abs/2011.08071
[9] Juliano Rabelo, Mi-Young Kim, and Randy Goebel. 2019. Combining Sim-
ilarity and Transformer Methods for Case Law Entailment. In Proceedings
of the Seventeenth International Conference on Artificial Intelligence and Law,
ICAIL 2019, Montreal, QC, Canada, June 17-21, 2019. ACM, 290–296. https:
//doi.org/10.1145/3322640.3326741
[10] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2020. COLIEE 2020: Methods for Legal Document Retrieval
and Entailment. https://sites.ualberta.ca/~rabelo/COLIEE2021/COLIEE2020_
summary.pdf. Accessed: 2021-05-09.
[11] Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embed-
dings using Siamese BERT-Networks. In Proceedings of the 2019 Conference
on Empirical Methods in Natural Language Processing and the 9th International
Joint Conference on Natural Language Processing, EMNLP-IJCNLP 2019, Hong
Kong, China, November 3-7, 2019, Kentaro Inui, Jing Jiang, Vincent Ng, and
Xiaojun Wan (Eds.). Association for Computational Linguistics, 3980–3990.
https://doi.org/10.18653/v1/D19-1410
[12] Chi Sun, Xipeng Qiu, Yige Xu, and Xuanjing Huang. 2019. How to Fine-Tune
BERT for Text Classification?. In Chinese Computational Linguistics - 18th China
National Conference, CCL 2019, Kunming, China, October 18-20, 2019, Proceedings
(Lecture Notes in Computer Science, Vol. 11856), Maosong Sun, Xuanjing Huang,
Heng Ji, Zhiyuan Liu, and Yang Liu (Eds.). Springer, 194–206. https://doi.org/10.
1007/978-3-030-32381-3_16
[13] Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q. Weinberger, and Yoav Artzi.
2020. BERTScore: Evaluating Text Generation with BERT. In 8th International
Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April
26-30, 2020. OpenReview.net. https://openreview.net/forum?id=SkeHuCVFDr
[14] Yukun Zhu, Ryan Kiros, Richard S. Zemel, Ruslan Salakhutdinov, Raquel Urtasun,
Antonio Torralba, and Sanja Fidler. 2015. Aligning Books and Movies: Towards
Story-Like Visual Explanations by Watching Movies and Reading Books. In 2015

294
 To Tune or Not To Tune?
Zero-shot Models for Legal Case Entailment
Guilherme Moraes Rosa
NeuralMind, Brazil
University of Campinas (Unicamp), Brazil
Roberto de Alencar Lotufo
NeuralMind, Brazil
University of Campinas (Unicamp), Brazil
ABSTRACT
There has been mounting evidence that pretrained language models
fine-tuned on large and diverse supervised datasets can transfer
well to a variety of out-of-domain tasks. In this work, we investigate
this transfer ability to the legal domain. For that, we participated
in the legal case entailment task of COLIEE 2021, in which we
use such models with no adaptations to the target domain. Our
submissions achieved the highest scores, surpassing the second-best
submission by more than six percentage points. Our experiments
confirm a counter-intuitive result in the new paradigm of pretrained
language models: given limited labeled data, models with little or
no adaption to the target task can be more robust to changes in the
data distribution than models fine-tuned on it. Code is available
at https://github.com/neuralmind-ai/coliee.
KEYWORDS
Legal NLP, Legal Case Entailment, Deberta, T5, Zero-shot
ACM Reference Format:
Guilherme Moraes Rosa, Ruan Chaves Rodrigues, Roberto de Alencar Lotufo,
and Rodrigo Nogueira. 2021. To Tune or Not To Tune? Zero-shot Models for
Legal Case Entailment. In Eighteenth International Conference for Artificial
Intelligence and Law (ICAIL’21), June 21–25, 2021, São Paulo, Brazil. ACM,
New York, NY, USA, 6 pages. https://doi.org/10.1145/3462757.3466103
1 INTRODUCTION
An ongoing trend in natural language processing and information
retrieval is to use the same model with small adaptions to solve a va-
riety of tasks. Pretrained transformers, epitomized by BERT [7], are
the state-of-the-art in question answering [15], natural language
inference [12, 33], summarization [3, 17], and ranking tasks [11, 21].
Although these tasks are diverse, current top-performing models
in each of them have a similar architecture to the original Trans-
former [40] and are pretrained on variations of the masked language
modeling objective used by Devlin et al. [7].
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permitted. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior specific permission and/or a
fee. Request permissions from permissions@acm.org.
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
© 2021 Association for Computing Machinery.
ACM ISBN 978-1-4503-8526-8/21/06. . . $15.00
https://doi.org/10.1145/3462757.3466103
295
Ruan Chaves Rodrigues
NeuralMind, Brazil
Federal University of Goiás (UFG), Brazil
Rodrigo Nogueira
NeuralMind, Brazil
University of Campinas (Unicamp), Brazil
University of Waterloo, Canada
Zero-shot and few-shot models are becoming more competitive
with models fine-tuned on large datasets. For instance, few-shot
results of GPT-3 [4] sparked an interest in prompt engineering
methods, which are now an active area of research [20, 36, 38]. The
goal of these methods is to find input templates such that the model
is more likely to give the correct answer.
In information retrieval, pretrained models fine-tuned only on
a large dataset have also shown strong zero-shot capabilities [39].
For example, the same multi-stage pipeline based on T5 [33] was
the best or second best-performing system in 4 tracks of the TREC
2021 [27], including specialized tasks such as Precision Medicine [34],
and TREC-COVID [48]. A remarkable feature of this pipeline is that,
for most tasks, the models are fine-tuned only on a general-domain
ranking dataset, i.e., they do not use in-domain data.
However, to date, there has not been strong evidence that zero-
shot models transfer well to the legal domain. Most state-of-the-art
models need adaptations to the target task. For example, the top-
performing system on the legal case entailment task of COLIEE
2020 [31] uses an interpolation of BM25 [35] scores and scores from
a BERT model fine-tuned on the target task [24].
In this work, we show that, for the legal case entailment task of
COLIEE, pretrained language models without any fine-tuning on
the target task perform at least equivalently or even better than
models fine-tuned on the task itself. Our approach is characterized
as zero-shot since the model was only fine-tuned on annotated data
from another domain. Our result confirms in the legal domain a
counter-intuitive recent finding in other domains: given limited
labeled data, zero-shot models tend to perform better on held-out
datasets than models fine-tuned on the target task [27, 32].
2 RELATED WORK
It is a common assumption among NLP researchers that models
developed using nonlegal texts would lead to unsatisfactory perfor-
mance when directly applied to legal tasks [9, 50]. To overcome this
issue, general-purpose techniques are adapted to the legal domain.
For example, Chalkidis and Kampas [5] pre-trained legal word em-
beddings using word2vec [22, 23] over a large corpus comprised
of legislation from multiple countries. Zhong et al. [51] created
a question answering dataset in the legal domain, collected from
the National Judicial Examination of China and evaluated different
models on it, including Transformers. Elnaggar et al. [8] applied
multi-task learning to minimize the problems related to data scarcity
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Rosa et al.

in the legal domain. The models were trained in translation, sum- We separate 80% of the 2020 training set for training and the
marization, and multi-label classification tasks, and achieved better remaining for validation, which yields 260 and 65 positive examples
results than single-task models. for training and validation sets, respectively. Negative examples
Pretrained transformer models have only begun to be adopted in are all candidates not labeled as positive.
legal NLP applications more broadly [2, 10, 16, 37, 44]. In some tasks,
they marginally outperform classical methods, especially when 2020 2021
training data is scarce. For example, Zhong et al. [50] showed that Train Test Train Test
a BERT-based model performs better than a tf-idf similarity model Examples (base cases) 325 100 425 100
on a judgment prediction task [43], but is slightly less effective than Avg. # of candidates / example 35.52 36.72 35.80 35.24
an attention-based convolutional neural network [47]. Avg. positive candidates / example 1.15 1.25 1.17 1.17
In some cases, they outperform classical methods, but at the Avg. of tokens in base cases 37.72 37.03 37.51 32.97
expense of using hand-crafted features or by being fine-tuned on Avg. of tokens in candidates 100.16 112.65 103.14 100.83
the target task. For example, the best submission to task 2 of COLIEE
Table 1: Statistics of COLIEE’s Task 2.
2019 was a BERT model fed with hand-crafted inputs and fine-tuned
on in-domain data [29].
Peters et al. [26] demonstrate that fine-tuning on the target task
may not perform better than simple feature extraction from a pre- The micro F1-score is the official metric in this task:
trained model if the pretraining task and the target task belong to F1 = (2 × 𝑃 × 𝑅)/(𝑃 + 𝑅), (1)
highly different domains. These findings lead us to consider zero-
where 𝑃 is the number of correctly retrieved paragraphs for all
shot approaches while investigating how general domain Trans-
queries divided by the number of retrieved paragraphs for all
former models can be applied to legal tasks.
queries, and 𝑅 is the number of correctly retrieved paragraphs
Although zero-shot approaches are relatively novel in the legal
for all queries divided by the number of relevant paragraphs for all
domain, our work is not the first to apply zero-shot Transformer
queries.
models to domain-specific entailment tasks where limited labeled
data is available. Yin et al. [45] have transformed multi-label clas- 3 METHOD
sification tasks into textual entailment tasks, and then evaluated
the performance of a BERT model fine-tuned on mainstream entail- We experiment with the following models: BM25, monoT5-zero-
ment datasets. Yin et al. [46] also performed similar experiments shot, monoT5, and DeBERTa. We also evaluate an ensemble of our
while transforming question answering and coreference resolution monoT5 and DeBERTa models.
tasks into entailment tasks. We are not the first to use zero-shot
techniques on the legal case entailment task. For instance, Rabelo
3.1 BM25
et al. [28] used a BERT fine-tuned for paraphrase detection com- BM25 is a bag-of-words retrieval function that scores a document
bined with two transformer-based models fine-tuned on a generic based on the query terms appearing in it. We use the BM25 imple-
text entailment dataset and features generated by a BERT model mented in Pyserini [18], a Python toolkit that supports replicable
fine-tuned on the COLIEE training dataset. However, we are the information retrieval research. We use its default parameters.
first to show that zero-shot models can outperform fine-tuned ones We first index all paragraphs in datasets of tasks 1 and 2. Having
on this task. more paragraphs from task 1 improves the term statistics (e.g.,
document frequencies) used by BM25. Task 1 dataset is composed
of long documents, while task 2 is composed of paragraphs. This
difference in length may degrade BM25 scores for task 2 paragraphs
2.1 The Legal Case Entailment Task because the average document length will be higher due to task 1
The Competition on Legal Information Extraction/Entailment (COL- documents. We address this problem by segmenting each document
IEE) [13, 14, 30, 31] is an annual competition whose aim is to eval- into several paragraphs using a context window of 10 sentences
uate automatic systems on case and statute law tasks. with overlapping strides of 5 sentences.
Among the five tasks of the 2021 competition, we submitted The entailed fragment might be comprised of multiple sentences.
systems to task 2, called legal case entailment, which consists of Here we treat each of its sentences as a query and compute a BM25
identifying paragraphs from existing cases that entail a given frag- score for each sentence and candidate paragraph pair independently.
ment of a base case. The final score for each paragraph is the maximum among its sen-
Training data consists of a set of decision fragments, its respec- tence and paragraph pair scores. We then use the method described
tive candidate paragraphs that could be relevant or not to the frag- in Section 3.5 to select the paragraphs that will comprise our final
ment and a set of labels containing the number of the paragraphs answer.
by which the decision fragment is entailed. Test data includes only
decision fragments and candidate paragraphs, but no labels. As 3.2 monoT5-zero-shot
shown in Figure 1, the input to the model is a decision fragment At a high level, monoT5-zero-shot is a sequence-to-sequence adap-
Q of an unseen case and the output should be a set of paragraphs tation of the T5 model [33] proposed by Nogueira et al. [25] and
𝑃 = [𝑃 1, 𝑃2, ..., 𝑃𝑛 ] that are relevant to the given decision 𝑄. In further detailed in Lin et al. [19]. This ranking model is close to
table 1, we show the statistics of the 2020 and 2021 datasets. or at the state-of-the-art in retrieval tasks such as Robust04 [42],

296
Zero-shot Models for Legal Case Entailment
Figure 1: COLIEE’s Task 2 example.
TREC-COVID, and TREC 2020 Precision Medicine and Deep Learn-
ing tracks. Details of the model are described in Nogueira et al. [25];
here, we only provide a short overview.
In the T5 model, all target tasks are cast as sequence-to-sequence
tasks. For our task, we use the following input sequence template:
Query: 𝑞 Document: 𝑑 Relevant:
where 𝑞 and 𝑑 are the query and candidate texts, respectively. In
this work, 𝑞 is a fragment, and 𝑑 is one of the candidate paragraphs.
The model estimates a score 𝑠 quantifying how relevant a candi-
date text 𝑑 is to a query 𝑞. That is:
𝑠 = 𝑃 (Relevant = 1|𝑑, 𝑞).
The model is fine-tuned to produce the tokens “true” or “false”
depending on whether the candidate is relevant or not to the query.
That is, “true” and “false” are the “target tokens” (i.e., ground truth
predictions in the sequence-to-sequence transformation). The suffix
“Relevant:” in the input string serves as hint to the model for the
tokens it should produce.
We use a T5-large model fine-tuned on MS MARCO [1], a dataset
of approximately 530k query and relevant passage pairs. We use a
checkpoint available at Huggingface’s model hub that was trained
with a learning rate of 10−3 using batches of 128 examples for 10k
steps, or approximately one epoch of the MS MARCO dataset.1
In each batch, a roughly equal number of positive and negative
examples is sampled. We refer to this model as monoT5-zero-shot.
Although fine-tuning for more epochs leads to better perfor-
mance on the MS MARCO development set, Nogueira et al. [25]
showed that further training degrades a model’s zero-shot perfor-
mance on other datasets. We observed similar behavior in our task
and opted to use the model trained for one epoch on MS MARCO.
At inference time, to compute probabilities for each query-candidate
pair, a softmax is applied only on the logits of the tokens “true” and
“false”. The final score of each candidate is the probability assigned
to the token “true”.
3.3 monoT5
We further fine-tune monoT5-zero-shot on the 2020 task 2 training
set following a similar training procedure described in the previous
section.
Fragments are mostly comprised of only one sentence, while
candidate paragraphs are longer, sometimes exceeding 512 tokens in
length. Thus, to avoid excessive memory usage due to the quadratic
1 https://huggingface.co/castorini/monot5-large-msmarco-10k
ICAIL’21, June 21–25, 2021, São Paulo, Brazil
memory cost of Transformers with respect to the sequence length,
we truncate inputs to 512 tokens during both training and inference.
The model is fine-tuned with a learning rate of 10−3 for 80 steps
using batches of size 128, which corresponds to 20 epochs. Each
batch has the same amount of positive and negative examples. We
refer to this model as monoT5.
(2)
(3) 3.4 DeBERTa
Decoding-enhanced BERT with disentangled attention (DeBERTa)
improves on the original BERT and RoBERTa architectures by in-
troducing two techniques: the disentangled attention mechanism
and an enhanced mask decoder [12]. Both improvements seek to
introduce positional information to the pretraining procedure, both
in terms of the absolute position of a token and the relative position
between them.
The COLIEE 2021 Task 2 dataset has very few positive examples
of entailment. Therefore, for fine-tuning DeBERTa on this dataset,
we found appropriate to artificially expand the positive examples.
As fragments take up only a small portion of a base case paragraph,
we expand positive examples by generating artificial fragments
from the same base case paragraph in which the original fragment
has occurred. This is done by moving a sliding window, with a
stride that is half the size of the original fragment, over the base
case paragraph. Each step of this sliding window is taken to be an
artificial fragment, and such artificial fragments are assigned the
same labels as the original fragment.
Although the resulting dataset after these operations is several
times larger than the original Task 2 dataset, we achieved better
results by fine-tuning DeBERTa on a small sample taken from this
artificial dataset. After experimenting with distinct sample sizes,
we settled for a sample of twenty thousand fragment and candidate
paragraph pairs, equally balanced between positive and negative
entailment pairs.
In order to find the best hyperparameters for fine-tuning a De-
BERTa Large model, we perform a grid search over the hyperpa-
rameters suggested by He et al. [12] while early stopping always
at the second epoch. The best combination of hyperparameters is
used to fine-tune the model for ten epochs. The checkpoint with
the best performance on the 2020 test set is selected to generate
our predictions for the 2021 test set.
297
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Rosa et al.

2020 2021
Description Submission name F1 Prec Recall F1 Prec Recall 𝛼, 𝛽, 𝛾
(1a) Median of submissions - 0.5718 - - 0.5860 - - -
(1b) Best of 2020 [24] JNLP.task2.BMWT 0.6753 0.7358 0.6240 - - - -
(1c) 2nd best of 2021 UA_reg_pp - - - 0.6274 - - -
(2) BM25 - 0.6046 0.7222 0.52 0.6009 0.6666 0.5470 0.07, 2, 0.99
(3) DeBERTa DeBERTa 0.7094 0.7614 0.6640 0.6339 0.6635 0.6068 0, 2, 0.999
(4) monoT5 monoT5 0.6887 0.7155 0.660 0.6610 0.6554 0.6666 0, 3, 0.995
(5) monoT5-zero-shot - 0.6577 0.7400 0.5920 0.6872 0.7090 0.6666 0, 3, 0.995
(6) Ensemble of (3) and (4) DebertaT5 0.7217 0.7904 0.6640 0.6912 0.7500 0.6410 0.6, 2, 0.999
(7) Ensemble of (3) and (5) - 0.7038 0.7592 0.6560 0.6814 0.7064 0.6581 0.6, 2, 0.999
Table 2: Test set results on Task 2 of COLIEE 2020 and 2021. Our best single model F1 for each year is in bold.

3.5 Answer Selection
The models described above estimate a score for each (fragment,
candidate paragraph) pair. To select the final set of paragraphs for
a given fragment, we apply three rules:
• Select paragraphs whose scores are above a threshold 𝛼;
• Select the top 𝛽 paragraphs with respect to their scores;
• Select paragraphs whose scores are at least 𝛾 of the top score.
We use exhaustive grid search to find the best values for 𝛼, 𝛽, 𝛾
on the development set of the 2020 task 2 dataset. We swept 𝛼 =
[0, 0.1, ..., 0.9], 𝛽 = [1, 2..., 10], and 𝛾 = [0, 0.1, ..., 0.9, 0.95, 0.99, 0.995,
..., 0.9999]. The best values for each model can be found in Table 3.
Note that our hyperparameter search includes the possibility of
not using the first or third strategies if 𝛼 = 0 or 𝛾 = 0 are chosen,
respectively.
3.6 DeBERTa + monoT5 Ensemble (DebertaT5)
Ensemble methods seek to combine the strengths and compensate
for the weaknesses of the models in order that the final model has
better generalization performance.
We use the following method to combine the predictions of
monoT5 and DeBERTa (both fine-tuned on COLIEE 2020): We con-
catenate the final set of paragraphs selected by each model. We
remove duplicates, preserving the highest score. Then, we apply
again the grid search method explained in the previous section
to select the final set of paragraphs. It is important to note that
our method does not combine scores between models. It ensures
that only individual answers with a certain degree of confidence
are maintained in the final answer, which generally leads to an
increase in Precision. The final answer for each test example can be
composed of individual answers from one model or both models.
4 RESULTS
We present our main result in Table 2. Our baseline BM25 method
scores above the median of submissions in both COLIEE 2020 and
2021 (row 2 vs. 1a). This confirms that BM25 is a strong baseline
and it is in agreement with results from other competitions such as
the Health Misinformation and Precision Medicine track of TREC
2020 [27].
Our pretrained transformer models (rows 3, 4 and 5) score above
BM25, the best submission of 2020 [24], and the second-best submis-
sion of 2021. Likewise, our ensemble method effectively combines
DeBERTa and monoT5 predictions, achieving the best score among
all submissions (row 6). However, the performance of monoT5-
zero-shot decreases when combined with DeBERTa (row 5 vs. 7),
showing that monoT5-zero-shot is a strong model.
The most interesting comparison is between monoT5 and monoT5-
zero-shot (rows 4 and 5). In the 2020 test data, monoT5 showed
better results than monoT5-zero-shot. Hence, we decided to submit
only the fine-tuned model to the 2021 competition. After the release
of ground-truth annotations of the 2021 test set, our evaluation of
monoT5-zero-shot showed that it performs better than monoT5.
A similar “inversion” pattern was found for DeBERTa vs. monoT5
(rows 3 and 4). DeBERTa was better than monoT5 on the 2020 test
set, but the opposite happened on the 2021 test set.
One explanation for these results is that we overfit on the test
data of 2020, i.e., by (unintentionally) selecting techniques and
hyperparameters that gave the best result on the 2020 test set as
experiments progressed. However, this is unlikely to be the case
for our fine-tuned monoT5 model, as our hyperparameter selection
is fully automatic and maximized on the development set, whose
data is from COLIEE competitions before 2020.
Another explanation is that there is a significant difference
between the annotation methodologies of 2020 and 2021. Conse-
quently, models specialized in the 2020 data could suffer from this
change. However, this is also unlikely since BM25 performed simi-
larly in both years. Furthermore, we cannot confirm this hypothesis
since it is difficult to quantify differences in the annotation process.
Regardless of the reason for the inversion, our main finding is
that our zero-shot model performed at least comparably to fine-
tuned models on the 2020 test set and achieved the best result of a
single model on 2021 test data.
4.1 Ablation of the Answer Selection Method
In Table 3, we show the ablation result of the answer selection
method proposed in Section 3.5. Our baseline answer selection
method, which we refer to as “no rule” in the table, uses only the
paragraph with the highest score as the final answer set, i.e., 𝛼 = 𝛾 =

298
Zero-shot Models for Legal Case Entailment ICAIL’21, June 21–25, 2021, São Paulo, Brazil

Model F1 Prec Recall 𝛼, 𝛽, 𝛾 Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario
Amodei. 2020. Language Models are Few-Shot Learners. In Advances in Neural In-
monoT5-zero-shot (no rule) 0.6517 0.7373 0.584 0, 1, 0 formation Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan,
monoT5-zero-shot 0.6577 0.74 0.592 0, 3, 0.995 and H. Lin (Eds.), Vol. 33. Curran Associates, Inc., 1877–1901. https://proceedings.
neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf
monoT5 (no rule) 0.6755 0.7600 0.608 0, 1, 0 [5] Ilias Chalkidis and Dimitrios Kampas. 2019. Deep learning in law: early adaptation
monoT5 0.6887 0.7155 0.6640 0, 3, 0.995 and legal word embeddings trained on large corpora. Artificial Intelligence and
Law volume 27, pages171–198(2019) (2019).
DeBERTa (no rule) 0.6933 0.7800 0.6240 0, 1, 0
[6] Sanyuan Chen, Yutai Hou, Yiming Cui, Wanxiang Che, Ting Liu, and Xiangzhan
DeBERTa 0.7094 0.7614 0.6640 0, 2, 0.999 Yu. 2020. Recall and Learn: Fine-tuning Deep Pretrained Language Models with
DebertaT5-zero-shot (no rule) 0.6875 0.7777 0.6160 0, 1, 0 Less Forgetting. arXiv:2004.12651 [cs.CL]
[7] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT:
DebertaT5-zero-shot 0.7038 0.7592 0.6560 0.6, 2, 0.999 Pre-training of Deep Bidirectional Transformers for Language Understanding. In
DebertaT5 (no rule) 0.7022 0.7900 0.6320 0, 1, 0 Proceedings of the 2019 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies, Volume 1 (Long and
DebertaT5 0.7217 0.7904 0.6640 0.6, 2, 0.999
Short Papers). 4171–4186.
[8] Ahmed Elnaggar, Christoph Gebendorfer, Ingo Glaser, and Florian Matthes. 2018.
Table 3: Ablation on the 2020 data of the answer selection Multi-Task Deep Learning for Legal Document Translation, Summarization and
method presented in Section 3.5. Multi-Label Classification. AICCC ’18: Proceedings of the 2018 Artificial Intelligence
and Cloud Computing Conference December 2018 Pages 9–15 (2018).
[9] Ahmed Elnaggar, Bernhard Waltl, Ingo Glaser, Jörg Landthaler, Elena Scepankova,
and Florian Matthes. 2018. Stop Illegal Comments: A Multi-Task Deep Learning
Approach. AICCC ’18: Proceedings of the 2018 Artificial Intelligence and Cloud
0 and 𝛽 = 1. For all models, the proposed answer selection method Computing Conference December 2018 Pages 41–47 (2018).
[10] Emad Elwany, Dave Moore, and Gaurav Oberoi. 2019. BERT Goes to Law School:
gives improvements of 0.6 to two F1 points over the baseline. Quantifying the Competitive Advantage of Access to Large Legal Corpora in
Contract Understanding. In Workshop on Document Intelligence at NeurIPS 2019.
[11] Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. Rethink Training of BERT
5 CONCLUSION Rerankers in Multi-Stage Retrieval Pipeline. arXiv preprint arXiv:2101.08751
We confirm a counter-intuitive result on a legal case entailment task: (2021).
[12] Pengcheng He, Xiaodong Liu, Jianfeng Gao, and Weizhu Chen. 2020. DeBERTa:
that models with little or no adaptation to the target task can have Decoding-enhanced BERT with Disentangled Attention. arXiv:2006.03654 [cs.CL]
better generalization abilities than models that have been carefully [13] Yoshinobu Kano, M. Kim, R. Goebel, and K. Satoh. 2017. Overview of COLIEE
fine-tuned to the task at hand. Domain adversarial fine-tuning [41] 2017. In COLIEE 2017 (EPiC Series in Computing, vol. 47). 1–8.
[14] Yoshinobu Kano, Mi-Young Kim, Masaharu Yoshioka, Yao Lu, Juliano Rabelo,
and changes to the Adam optimizer [6] [49] have been proposed Naoki Kiyota, Randy Goebel, and Ken Satoh. 2018. COLIEE-2018: Evaluation of
as valid approaches for fine-tuning Transformer models on small the competition on legal information extraction and entailment. In JSAI Interna-
domain-specific datasets. However, whether these techniques could tional Symposium on Artificial Intelligence. 177–192.
[15] Daniel Khashabi, Sewon Min, Tushar Khot, Ashish Sabharwal, Oyvind Tafjord,
successfully be applied to the legal case entailment task to make Peter Clark, and Hannaneh Hajishirzi. 2020. UnifiedQA: Crossing Format Bound-
models fine-tuned on target task data perform better than zero-shot aries With a Single QA System. In Proceedings of the 2020 Conference on Empirical
Methods in Natural Language Processing: Findings. 1896–1907.
approaches remains an open question. [16] Spyretta Leivaditi, Julien Rossi, and Evangelos Kanoulas. 2020. A Benchmark for
Therefore, although domain-specific language model pretraining Lease Contract Review. arXiv preprint arXiv:2010.10386 (2020).
and adjustments to the fine-tuning process are promising directions [17] Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman
Mohamed, Omer Levy, Veselin Stoyanov, and Luke Zettlemoyer. 2020. BART:
for future research, we believe that zero-shot approaches should Denoising Sequence-to-Sequence Pre-training for Natural Language Generation,
not be ignored as strong baselines for such experiments. Translation, and Comprehension. In Proceedings of the 58th Annual Meeting of
It should also be noted that our research has implications for the Association for Computational Linguistics. 7871–7880.
[18] Jimmy Lin, Xueguang Ma, Sheng-Chieh Lin, Jheng-Hong Yang, Ronak Pradeep,
future experiments beyond the scope of legal case entailment tasks. and Rodrigo Nogueira. 2021. Pyserini: An Easy-to-Use Python Toolkit to Support
Based on previous work by Yin et al. [45, 46], it is possible that Replicable IR Research with Sparse and Dense Representations. arXiv preprint
arXiv:2102.10073 (2021).
other legal tasks with limited labeled data, such as legal question [19] Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. 2020. Pretrained transformers
answering, may benefit from our zero-shot approach. for text ranking: Bert and beyond. arXiv preprint arXiv:2010.06467 (2020).
[20] Jiang Lu, Pinghua Gong, Jieping Ye, and Changshui Zhang. 2020. Learning from
Very Few Samples: A Survey. arXiv preprint arXiv:2009.02653 (2020).
REFERENCES [21] Xinyu Ma, Jiafeng Guo, Ruqing Zhang, Yixing Fan, Xiang Ji, and Xueqi Cheng.
[1] Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong 2021. PROP: Pre-training with Representative Words Prediction for Ad-hoc
Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Retrieval. In Proceedings of the 14th ACM International Conference on Web Search
Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. and Data Mining. 283–291.
MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. [22] Tomas Mikolov. 2013. Distributed representations of words and phrases and their
arXiv:1611.09268v3 (2018). compositionality. arXiv preprint arXiv:1310.4546 (2013).
[2] Purbid Bambroo and Aditi Awasthi. 2021. LegalDB: Long DistilBERT for Legal [23] Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient Esti-
Document Classification. In 2021 International Conference on Advances in Elec- mation of Word Representations in Vector Space. arXiv preprint arXiv:1301.3781
trical, Computing, Communication and Sustainable Technologies (ICAECT). IEEE, (2013).
1–4. [24] Ha-Thanh Nguyen, Hai-Yen Thi Vuong, Phuong Minh Nguyen, Binh Tran Dang,
[3] Hangbo Bao, Li Dong, Furu Wei, Wenhui Wang, Nan Yang, Xiaodong Quan Minh Bui, Sinh Trong Vu, Chau Minh Nguyen, Vu Tran, Ken Satoh, and
Liu, Yu Wang, Songhao Piao, Jianfeng Gao, Ming Zhou, and Hsiao-Wuen Minh Le Nguyen. 2020. JNLP Team: Deep Learning for Legal Processing in
Hon. 2020. UniLMv2: Pseudo-Masked Language Models for Unified Lan- COLIEE 2020. arXiv preprint arXiv:2011.08071 (2020).
guage Model Pre-Training. ArXiv. https://www.microsoft.com/en- [25] Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. 2020. Document
us/research/publication/unilmv2-pseudo-masked-language-models-for- Ranking with a Pretrained Sequence-to-Sequence Model. In Proceedings of the
unified-language-model-pre-training/ 2020 Conference on Empirical Methods in Natural Language Processing: Findings.
[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Pra- 708–718.
fulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, [26] Matthew E Peters, Sebastian Ruder, and Noah A Smith. 2019. To Tune or Not to
Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Tune? Adapting Pretrained Representations to Diverse Tasks. In Proceedings of
Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, the 4th Workshop on Representation Learning for NLP (RepL4NLP-2019). 7–14.
Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark,

299
ICAIL’21, June 21–25, 2021, São Paulo, Brazil Rosa et al.

[27] Ronak Pradeep, Xueguang Ma, Xinyu Zhang, Hang Cui, Ruizhou Xu, Rodrigo [51] Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and
Nogueira, and Jimmy Lin. [n.d.]. H2oloo at TREC 2020: When all you got is Maosong Sun1. 2020. JEC-QA: A Legal-Domain Question Answering Dataset.
a hammer... Deep Learning, Health Misinformation, and Precision Medicine. Proceedings of the AAAI Conference on Artificial Intelligence, 34(05), 9701-9708.
Corpus 5, d3 ([n. d.]), d2. (2020).
[28] J. Rabelo, M.Y. Kim, and R. Goebel. 2020. Application of text entailment techniques
in COLIEE 2020. International Workshop on Juris-informatics (JURISIN) associated
with JSAI International Symposia on AI (JSAI-isAI) (2020).
[29] Juliano Rabelo, Mi-Young Kim, and Randy Goebel. 2019. Combining similarity and
transformer methods for case law entailment. In Proceedings of the Seventeenth
International Conference on Artificial Intelligence and Law (ICAIL ’19). 290–296.
[30] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2019. A Summary of the COLIEE 2019 Competition. In
JSAI International Symposium on Artificial Intelligence. 34–49.
[31] Juliano Rabelo, Mi-Young Kim, Randy Goebel, Masaharu Yoshioka, Yoshinobu
Kano, and Ken Satoh. 2020. COLIEE 2020: Methods for Legal Document Retrieval
and Entailment. (2020).
[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh,
Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark,
et al. 2021. Learning transferable visual models from natural language supervision.
arXiv preprint arXiv:2103.00020 (2021).
[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang,
Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the Limits
of Transfer Learning with a Unified Text-to-Text Transformer. Journal of Machine
Learning Research 21, 140 (2020), 1–67. http://jmlr.org/papers/v21/20-074.html
[34] Kirk Roberts, Dina Demner-Fushman, E. Voorhees, W. Hersh, Steven Bedrick,
Alexander J. Lazar, and S. Pant. 2019. Overview of the TREC 2019 Precision
Medicine Track. The ... text REtrieval conference : TREC. Text REtrieval Conference
26 (2019).
[35] Stephen E Robertson, Steve Walker, Susan Jones, Micheline M Hancock-Beaulieu,
Mike Gatford, et al. 1995. Okapi at TREC-3. Nist Special Publication Sp 109 (1995),
109.
[36] Timo Schick and Hinrich Schütze. 2020. Exploiting cloze questions for few-shot
text classification and natural language inference. arXiv preprint arXiv:2001.07676
(2020).
[37] Shohreh Shaghaghian, Luna Yue Feng, Borna Jafarpour, and Nicolai Pogreb-
nyakov. 2020. Customizing Contextualized Language Models for Legal Docu-
ment Reviews. In 2020 IEEE International Conference on Big Data (Big Data). IEEE,
2139–2148.
[38] Derek Tam, Rakesh R Menon, Mohit Bansal, Shashank Srivastava, and Colin
Raffel. 2021. Improving and Simplifying Pattern Exploiting Training. arXiv
preprint arXiv:2103.11955 (2021).
[39] Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna
Gurevych. 2021. BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of
Information Retrieval Models. arXiv preprint arXiv:2104.08663 (4 2021). https:
//arxiv.org/abs/2104.08663
[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is All
you Need. In NIPS.
[41] Giorgos Vernikos, Katerina Margatina, Alexandra Chronopoulou, and Ion An-
droutsopoulos. 2020. Domain Adversarial Fine-Tuning as an Effective Regularizer.
arXiv:2009.13366 [cs.LG]
[42] Ellen M. Voorhees. 2004. Overview of the TREC 2004 Robust Track. Proceedings
of the Thirteenth Text REtrieval Conference, TREC 2004, Gaithersburg, Maryland,
November 16-19, 2004 (2004).
[43] Chaojun Xiao, Haoxi Zhong, Zhipeng Guo, Cunchao Tu, Zhiyuan Liu, Maosong
Sun, Yansong Feng, Xianpei Han, Zhen Hu, Heng Wang, and Jianfeng Xu. 2018.
CAIL2018: A Large-Scale Legal Dataset for Judgment Prediction. arXiv:1807.02478
(2018).
[44] Chin Man Yeung. 2019. Effects of inserting domain vocabulary and fine-tuning
BERT for German legal language. Master’s thesis. University of Twente.
[45] Wenpeng Yin, Jamaal Hay, and Dan Roth. 2019. Benchmarking Zero-
shot Text Classification: Datasets, Evaluation and Entailment Approach.
arXiv:1909.00161 [cs.CL]
[46] Wenpeng Yin, Nazneen Fatema Rajani, Dragomir Radev, Richard Socher, and
Caiming Xiong. 2020. Universal Natural Language Processing with Limited An-
notations: Try Few-shot Textual Entailment as a Start. arXiv:2010.02584 [cs.CL]
[47] Wenpeng Yin, Hinrich Schütze, Bing Xiang, and Bowen Zhou. 2016. ABCNN:
Attention-based convolutional neural network for modeling sentence pairs. Trans-
actions of the Association for Computational Linguistics 4 (2016), 259–272.
[48] Edwin Zhang, Nikhil Gupta, Rodrigo Nogueira, Kyunghyun Cho, and Jimmy
Lin. 2020. Rapidly Deploying a Neural Search Engine for the COVID-19 Open
Research Dataset. In Proceedings of the 1st Workshop on NLP for COVID-19 at ACL
2020.
[49] Tianyi Zhang, Felix Wu, Arzoo Katiyar, Kilian Q. Weinberger, and Yoav Artzi.
2021. Revisiting Few-sample BERT Fine-tuning. arXiv:2006.05987 [cs.CL]
[50] Haoxi Zhong, Chaojun Xiao, Cunchao Tu, Tianyang Zhang, Zhiyuan Liu, and
Maosong Sun. 2020. How Does NLP Benefit Legal System: A Summary of Legal
Artificial Intelligence. arXiv:2004.12158 (2020).

300
Index of Authors
Adler, Rachel F., 119
Aguiar, Derek, 99
Alexander, Charlotte S., 129
Almasian, Satya, 2
Amariles, David Restrepo, 129
Anderson, Brandon R., 159
Aoki, Yasuhiro, 278
Araszkiewicz, Michał, 129
Ash, Elliott, 109
Ashley, Alexandra, 129
Ashley, Kevin D., 129, 250
Atkinson, Katie, 12, 170
Aumiller, Dennis, 2
Barrie, Cameron, 119
Bench-Capon, Trevor, 12, 170
Benyekhlef, Karim, 129, 268
Bertolo, Emerson, 264
Bex, Floris, 175
Bhattacharya, Paheli, 22
Borges, Georg, 32
Branting, Karl, 129
Brockdorff, Juanita, 190
Buckley, Joshua, 139
Calegari, Roberta, 180
Castillo, Carlos, 210
Ceross, Aaron, 185
Ciochetti, Itamar, 264
Collenette, Joe, 170
Conrad, Jack G., 245
De Luca, Ernesto W., 285
de Souza, Edelcio G., 89
Distefano, Biagio, 230
dos Santos Neto, Jose Francisco, 240
Dureja, Shipra, 285
Dzehtsiarou, Kanstantsin, 170
Déziel, Pierre-Luc, 200, 256
Einarsson, Alexander, 119
EL HAMDANI, Rajaa, 40
El Hamdani, Rajaa, 129
Ellul, Joshua, 190
Esteban de la Rosa, Fernando, 195
Falduti, Mattia, 129
Fantinato, Marcelo, 240
Fungwacharakorn, Wachara, 50
Garneau, Nicolas, 200, 256
Gaumond, Eve, 200, 256
Gertz, Michael, 2
Ghosh, Kripabandhu, 22
Ghosh, Saptarshi, 22
Gipp, Bela, 109
Glaser, Ingo, 205
Governatori, Guido, 69, 139
Grabmair, Matthias, 79, 129
Grant, Jayla C., 129
Grossi, Davide, 149
Guha, Neel, 159
Górski, Łukasz, 60
Hammond, Kristian, 119
Harašta, Jakub, 129
Henderson, Peter, 159
Hirota, Renata, 240
Ho, Daniel E., 79, 159
Huang, Sieh-Chuen, 258
Huang, Zihan, 79
Huggins, Anna, 139
Johnson, Shiwanni, 129
Kaleta, Zbigniew, 225
Karimi-Haghighi, Marzieh, 210
Kawasaki, Tatsuki, 273
Kimura, Rampei, 270
Krasnashchok, Katsiaryna, 40
Krass, Mark S., 79
Kutty, Libin, 285
Lackner, Sebastian, 2
Lamontagne, Luc, 200, 256
Langlais, Philippe, 268
Leflar, Robert B, 258
Li Zhao, Andong L., 119
Lotufo, Roberto de Alencar, 295
Low, Charles, 79
Maranhão, Juliano, 89
Martínez, Diego C., 266
Matthes, Florian, 205
McCarthy, Stephen, 190
McConnell, Devin J., 99
Meeùs, Sébastien, 40, 129
Mok, Jonathan R., 260
Mok, Rachel V., 260
Mok, Wai Yin, 260
Moreno-Schneider, Julian, 109
Morris, Jason, 262
Mustapha, Majd, 40
Nogueira, Rodrigo, 295
Novotná, Tereza, 129
Olivieri, Francesco, 69
Ostendorff, Malte, 109
Pace, Gordon, 190
Pack, Harper, 119
Pah, Adam, 119
Paley, Andrew, 119
Palmirani, Monica, 230
302
Pandya, Sachin, 99
Peng, Linyu, 270
Peres, Sarajane Marques, 240
Piech, Mateusz, 225
Poddar, Soham, 22
Polo, Felipe Maia, 264
Prakken, Henry, 175
Ramakrishna, Shashishekar, 60
Rehm, Georg, 109
Renooij, Silja, 235
Restrepo Amariles, David, 40
Riveret, Régis, 180
Robaldo, Livio, 215
Rodrigues, Ruan Chaves, 295
Rosa, Guilherme Moraes, 295
Ross, Graham, 275
Rotolo, Antonino, 220, 266
Ruas, Terry, 109
Rudra, Koustav, 22
Salaün, Olivier, 268
Sammut, Trevor, 190
Sapienza, Salvatore, 230
Sartor, Giovanni, 89, 180
Satoh, Ken, 50, 273
Savelka, Jaromir, 129, 250
Scerri, Matthew, 190
Schamberger, Tom, 205
Schwartz, David, 119
Servantez, Sergio, 119
Shahania, Saijal, 285
Shao, Hsuan-Lei, 258
Smith, Clara, 220
Smywiński-Pohl, Aleksander, 225
Sovrano, Francesco, 230
Steging, Cornelis Cor, 235
Sterbentz, Marko, 119
Sudhi, Viju, 285
Suzuki, Youta, 278
Takahashi, Kazuko, 273
Tamargo, Luciano H., 266
Teng, Mengqiu, 79
Thiessen, Ernest, 275
Tippett, Elizabeth, 129
Trecenti, Julio, 240 Wróbel, Krzysztof, 225
Troussel, Aurore, 40, 129
Tsushima, Kanae, 50 Xu, Huihui, 250

Unger, Adriana Jacoto, 240 Yoshioka, Masaharu, 278

Verheij, Bart, 149, 235 Zeleznikow, John, 195

Vitali, Fabio, 230 Zhang, Hongyi, 79
Vold, Andrew, 245 Zheng, Heng, 149
Zheng, Lucia, 159
Wehnert, Sabine, 285 Zhu, James, 99
Westermann, Hannes, 129 Zhu, Tingting, 185
Witt, Alice, 139 Zufall, Frederike, 270

303