Download as pdf or txt
Download as pdf or txt
You are on page 1of 1

Legal Argument Recommendation and Precedent

Prediction using Graph Neural Networks

Abstract—The task of recommending legal arguments for In the work by Wu et al. [3], the task of precedence re-
new cases and identifying relevant precedents is critical in trieval for judgement prediction was done using unsupervised
judicial decision-making. This paper presents a novel approach contrastive learning. However, this approach did not leverage
leveraging advanced graph learning techniques to enhance legal
argument recommendations based on Facts and Issues at hand the graphical nature of legal citations, potentially limiting its
as well as precedent prediction and precedent ranking. We utilize effectiveness. Similarly, AutoLAW (Robert Zev Mahari) [4]
Retrieval-Augmented Generation (RAG) models for argument utilized a BERT model combined with a three-layered feed-
recommendation, effectively incorporating facts and issues into forward neural network for legal text processing; yet again, it
the argument generation using argument templates. For prece- did not exploit the citation graph structure.
dent prediction, we exploit the graphical structure of legal cita-
tions using a Relational Graph Convolutional Network (RGCN), To date, work has yet to successfully integrate the graphical
specifically designed for heterogeneous graphs, to predict the nature of legal citations with argument recommendation and
most relevant precedents that align with the current case’s generation. Our approach aims to fill this gap by leveraging
context. Our methodology demonstrates significant improvements the structured relationships inherent in legal citations.
in argument quality and the accuracy of precedent prediction,
thereby providing robust support for legal practitioners. III. N OVEL A PPROACH
Index Terms—RAG, Argument Recommendation, RGCN,
Precedent Prediction and Ranking Our proposed framework consists of two main components:
Argument Recommendation: We aim to recommend the
I. I NTRODUCTION facts and issues that can be used to make an argument. We
Automating legal citing, reasoning, and argumentation is employ a RAG model to generate the arguments using a large
of growing interest, particularly with the advent of advanced language model and argument templates.
graph learning techniques. In judicial decision-making, Precedent Prediction: To predict the most relevant prece-
recommending legal arguments for new cases and identifying dents for a given case, we utilize a RGCN based model for
relevant precedents are crucial tasks. Despite various efforts, heterogeneous graphs. This model exploits the structure of
existing methods often fall short of effectively capturing the the citation network to identify which precedents are most
complex, interconnected graphical nature of legal documents similar to the current case and likely to influence the outcome
and the nuanced relationships between them. favourably. By capturing the complex dependencies and rela-
tionships between legal documents, our graph-based approach
Our work is grounded in a newly created dataset, Indian significantly enhances the accuracy of precedent identification.
Legal Judgements (ILJ), compiled by scraping judgments Our approach not only addresses the tasks of legal argument
from the public legal website indiankanoon.org. The ILJ recommendation and precedent prediction but also introduces
dataset encompasses judgments from the 1900s to 2023, a novel way of leveraging the graphical structure of legal
forming a comprehensive resource for training and evaluating citations. Experimental results on the ILJ dataset demonstrate
our models. We can effectively model the precedence citation that our method outperforms existing techniques, highlighting
network by creating a graph using citations as edge links in its potential for practical application in the legal domain.
the heterogeneous graph. R EFERENCES
II. R ELATED W ORK [1] Paul, Shounak, Goyal, Pawan, Ghosh, Saptarshi. ”LeSICiN: A Heteroge-
neous Graph-based Approach for Automatic Legal Statute Identification
Previous research has explored various methods to address from Indian Legal Documents.” Proceedings of the 36th AAAI Confer-
the challenges of precedent prediction. Prior approaches pri- ence on Artificial Intelligence (AAAI), 2022.
[2] Bhattacharya, Paheli, Ghosh, Kripabandhu, Pal, Arindam, and Ghosh,
marily relied on models such as BERT and GPT to understand Saptarshi. ”Legal case document similarity: You need both network and
and generate legal text, improving context comprehension and text.” Information Processing & Management, vol. 59, no. 6, 2022, pp.
argument coherence. In LeSICiN (Shounak Paul et al.) [1] 103069.
[3] Yiquan Wu, Siying Zhou, Yifei Liu, Weiming Lu, Xiaozhong Liu, Yating
and in Legal Case Document Similarity (Paheli Bhattacharya Zhang, Changlong Sun, Fei Wu, and Kun Kuang. 2023. Precedent-
2022 et al.) [2], the precedence citation graph was utilized to Enhanced Legal Judgment Prediction with LLM and Domain-Model
understand the relationships between cases, but these studies Collaboration. In Proceedings of the 2023 Conference on Empirical
Methods in Natural Language Processing, pages 12060–12075, Singa-
did not address the task of ranking and predicting which pore. Association for Computational Linguistics.
sections of the precedents should be cited in judgments for [4] Mahari, Robert Zev. ”AutoLAW: Augmented Legal Reasoning through
an unseen case document. Legal Precedent Prediction.” arXiv:2106.16034 [cs.CL], June 2021.

You might also like