Professional Documents
Culture Documents
Artificial Intelligence IJCAI 2019 International Workshops Macao China August 10 12 2019 Revised Selected Best Papers Amal El Fallah Seghrouchni
Artificial Intelligence IJCAI 2019 International Workshops Macao China August 10 12 2019 Revised Selected Best Papers Amal El Fallah Seghrouchni
https://textbookfull.com/product/advances-in-computer-games-16th-
international-conference-acg-2019-macao-china-
august-11-13-2019-revised-selected-papers-tristan-cazenave/
https://textbookfull.com/product/euro-par-2019-parallel-
processing-workshops-euro-par-2019-international-workshops-
gottingen-germany-august-26-30-2019-revised-selected-papers-
ulrich-schwardmann/
https://textbookfull.com/product/web-and-big-data-apweb-
waim-2019-international-workshops-kgma-and-dsea-chengdu-china-
august-1-3-2019-revised-selected-papers-jingkuan-song/
Security Protocols XXVII 27th International Workshop
Cambridge UK April 10 12 2019 Revised Selected Papers
Jonathan Anderson
https://textbookfull.com/product/security-protocols-xxvii-27th-
international-workshop-cambridge-uk-april-10-12-2019-revised-
selected-papers-jonathan-anderson/
https://textbookfull.com/product/theoretical-computer-
science-37th-national-conference-nctcs-2019-lanzhou-china-
august-2-4-2019-revised-selected-papers-xiaoming-sun/
https://textbookfull.com/product/agents-and-artificial-
intelligence-11th-international-conference-icaart-2019-prague-
czech-republic-february-19-21-2019-revised-selected-papers-jaap-
van-den-herik/
https://textbookfull.com/product/computer-engineering-and-
technology-23rd-ccf-conference-nccet-2019-enshi-china-
august-1-2-2019-revised-selected-papers-weixia-xu/
Amal El Fallah Seghrouchni
David Sarne (Eds.)
LNAI 12158
Artificial Intelligence
IJCAI 2019 International Workshops
Macao, China, August 10–12, 2019
Revised Selected Best Papers
123
Lecture Notes in Artificial Intelligence 12158
Series Editors
Randy Goebel
University of Alberta, Edmonton, Canada
Yuzuru Tanaka
Hokkaido University, Sapporo, Japan
Wolfgang Wahlster
DFKI and Saarland University, Saarbrücken, Germany
Founding Editor
Jörg Siekmann
DFKI and Saarland University, Saarbrücken, Germany
More information about this series at http://www.springer.com/series/1244
Amal El Fallah Seghrouchni •
Artificial Intelligence
IJCAI 2019 International Workshops
Macao, China, August 10–12, 2019
Revised Selected Best Papers
123
Editors
Amal El Fallah Seghrouchni David Sarne
Sorbonne University – Sciences Bar-Ilan University
Paris, France Ramat Gan, Israel
This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Preface
chairs, and the IJCAI local organization team for their invaluable contribution and
support for the success of the IJCAI 2019 workshops.
Workshops Chairs
Amal El Fallah Seghrouchni Sorbonne University, LIP6, France
David Sarne Bar-Ilan University, Israel
Web Master
Hongming Zhang The Hong Kong University of Science
and Technology, Hong Kong, China
Workshops Coordinator
Derek Fai Wong University of Macau, Macau, China
Publication Chair
Arthur Casals Sorbonne University, LIP6, France
Workshops Organizers
Eunika Mercier-Laurent International Federation for Information Processing
(IFIP)
Mieczyslaw Lech Owoc Wroclaw University of Economics, Poland
Chung-Chi Chen National Taiwan University, Taiwan
Hen-Hsen Huang National Chengchi University, Taiwan
Hiroya Takamura AIST, Japan
Hsin-Hsi Chen National Taiwan University, Taiwan
Matthew Klenk Palo Alto Research Center, USA
Diedrich Wolter University of Bamberg, Germany
Ruben Glatt Lawrence Livermore National Laboratory, USA
Felipe Leno da Silva Advanced Institute for AI, Brazil
Denis Steckelmacher Vrije Universiteit Brussel, Belgium
Patrick MacAlpine Microsoft Research, USA
Sheng Li University of Georgia, USA
Yaliang Li Alibaba Group, USA
Jing Gao University at Buffalo, USA
Yun Fu Northeastern University, USA
Douglas Lange Naval Information Warfare Center Pacific, USA
Luke Marsh Defence Science and Technology Group, Australia
Takayuki Ito Nagoya Institute of Technology, Japan
Minjie Zhang University of Wollongong, Australia
viii Organization
Qualitative Reasoning
Humanizing AI
1 Summary
1 Introduction
Being the largest bank in Southeast Asia1 , DBS Bank has tremendous customer data
from various sources such as demographics, payments, fund transfers and so on. As
most of the data is inter-related, graph is a promising way to store the data, analyze the
relations and extract useful insights.
Graphs storing knowledge base such as customer data are often considered as
knowledge graph, where facts in the knowledge base are represented by links in the
graph (edges between nodes). The nodes in a knowledge graph represent entities in
knowledge base and the edges represent relations. Additionally, each node and edge
could be associated with a type. For example, in the context of consumer banking, the
type of a node in the graph could be customer, merchant, or building. A link (fact) in
knowledge graph is a triplet h, r, t, where h is a node representing the subject of the
fact, t is a node representing the object, and r is the edge between the two nodes repre-
senting the relationship of the fact. As in the consumer banking example, the triplet
Bob, transf er to, Carlo is a transaction between two customers, and the triplet
Alice, with address, Orchard represents the fact that the customer Alice lives in
Orchard district.
As a benefit of the graphical structure, knowledge graph can capture interrelations
and interactions across tremendous types of entities more effectively than traditional
1
https://en.wikipedia.org/wiki/List of largest banks in Southeast Asia.
c Springer Nature Switzerland AG 2020
A. El Fallah Seghrouchni and D. Sarne (Eds.): IJCAI 2019 Workshops, LNAI 12158, pp. 5–17, 2020.
https://doi.org/10.1007/978-3-030-56150-5_1
6 D. Shao and R. Annam
methods, and hence has drawn lots of attention recently. For example, Google has
been using its Google Knowledge Graph [6] containing millions of facts and entities
to enhance its search engine’s result. DBpedia [8] is a knowledge graph extracted from
data in Wikipedia useful for question answering and fact checking.
Among the various applications of knowledge graph, link prediction is one of the
most widely used and researched. The aim of link prediction is to predict new links
between nodes based on the existing graph. It comes from the situation that many
knowledge graphs are incomplete. For example, a customer of a bank may not have
credit cards in this bank. But he or she may have cards in other banks, which cannot
be captured by the graph within the bank. Using link prediction, this customer could be
linked with a similar card issued by the bank. As link prediction can uncover unobserved
facts and intentions, it is highly useful in community detection, customer recommen-
dation, anti-money laundering, and many more. For the same reason, link prediction is
also known as knowledge graph completion.
To solve the link prediction problem, embedding-based models inspired by word
embedding [10] in text mining have been increasingly popular recently. In the embed-
ding framework, nodes and edges in a graph are all embedded into some continuous
vector space. So in a triplet h, r, t, h, r and t are represented by vectors or matri-
ces. Then a score function is proposed to measure the probability that this triplet is a
fact in the graph. Hence the embeddings are obtained by minimizing a loss function
over all relations in the graph. With the learnt embeddings, the probability that h and
t are linked by r can be estimated by the score of h, r, t. Most previous works take
the same embedding framework with different scoring function and loss function. One
straightforward model [4] is to use vector addition to composite h and r to h + r and
use Lp -norm of h + r − t to score the triplet. More models [5, 9, 13–15, 18, 19] have
been proposed with various complicated scoring functions.
Experiments of previous works have been conducted and compared on public
datasets such as WN18 about lexical relation between words and FB15k about gen-
eral facts. However, it has been pointed out [17] that for many of the triplets in these
two dataset, its reversion also belongs to the same dataset. For example, WN18 has two
relations with large number of instances: “has part” and “parts of ”. So every triplet
h, has part, t corresponds to t, part of, h in the same dataset. In the consumer
banking data, such inverse relation property only applies to a small portion of the data
such as fund transfer. More importantly, each link in the banking knowledge graph is
explicitly associated with certain types of nodes only. Hence irrelevant nodes should be
excluded in the link prediction for a given link as pointed out in [14].
Knowledge Graph Completion in Consumer Banking 7
In this paper, we will construct a knowledge graph using real consumer banking
data from DBS Bank, run experiments of popular models with node types and link
types introduced, conduct comprehensive comparison of the performance drilled down
to relation level, and conclude the most suitable model for different relation types.
Moreover, use cases of link prediction for consumer banking will be provided.
To the best of our knowledge, this is the first work to benchmark knowledge graph
completion models using real consumer data and propose a guideline to select models
for different relation types.
The rest of this paper is organized as follows. We will review relevant literatures
in Sect. 2 and explore the details of popular models in Sect. 3. The performance of
different models on our consumer banking data will be summarized in Sect. 4. Section 5
will show some use cases and the paper will be concluded in Sect. 6.
2 Related Work
Most of the models proposed for link prediction in knowledge graph leverage embed-
ding framework. Given a triplet h, r, t and embeddings of nodes into a vector space
or matrix space, there are different approaches to score the probability of the existence
of the triplet. Popular models are summarized in Table 1 with complexity.
The structured embedding model (SE) [2] embeds edges into two matrices Lh ∈ Rn
and Lt ∈ Rn , where Lh is to transform h and Lt is to transform t. Then ||Lh ·h−Lt ·t||p
is employed as the score function, where || · ||p is the Lp -norm. The idea of this model
is that if two nodes are in the same triplet, then their embedding vectors should be close
in some subspace depending on the edge. The TransE model [4] took a more direct way.
Instead of mapping an edge into two matrices, TransE embeds each edge into Rn , hence
the head node h can be translated by adding the edge vector r to obtain a candidate tail
vector h + r. Then Lp -norm is adopted to measure the distance between h + r and t.
Although SE model has a better expressiveness, it is outperformed by TransE model.
Following the idea of TransE that the tail node t is a translation from the head node
h via the edge r in the same embedding space, many models have been proposed by
extending TransE from different aspects. TransH [19] first projects the embedding vec-
tors of h and t to a hyperplane that is perpendicular to the edge embedding vector, and
then uses Lp -norm to score the projected triplet. TransR [9] generalizes TransH in the
sense that the node embedding vectors could be projected to arbitrary edge-related sub-
space. It finds a matrix Mr to each edge r and uses Lp -norm the score the transformed
difference Mr h + r − Mr t.
Instead of translating from head node to tail node via the edge, bilinear approaches treat
the triplets as a 3D binary tensor and score a triplet by edge-related bilinear functions
8 D. Shao and R. Annam
whose inputs are embedding vectors of h and t. RESCAL model [12] represents the
edge r by a matrix Mr in Rn×n , and scores the triplet by hT Mr t.
DistMult model [20] extends RESCAL model by setting Mr to be a diagonal matrix,
so the scoring function is hT diag(r)t, where diag(r) is the diagonal matrix whose
main diagonal is r. Hence DistMult reduces the number of learning parameters sig-
nificantly compared with RESCAL, and also outperforms RESCAL. However, neither
of these two models have considered the difference of positioning one node as head
and tail. In other words, the triplet h, r, t and the triplet t, r, h will be given the
same score. To overcome this issue, ComplEx model [18] proposes to embed nodes and
edges into complex vector space Cn instead of real space, and uses the conjugation of
the embedding vector of a node when it is in tail position, so the scoring function is
Re(hT diag(r)t̄).
3 Methodology
A knowledge graph is a directed graph G = (V, E) with node type set TV and edge
type set TE . Given the set of vertices V and the set of edges E, the node type set TV
is a cover of V so that each node is associated with at least one type in TV . The edge
type set TE is a subset of the Cartesian product P (TV ) × P (TV ) where P (TV ) is the
powerset of TV . So, each edge is associated with one type connecting from one node
type to another node type. For example, the edge type with address connects from
customer nodes to district nodes.
Different models solve the problem adopting different scoring function and different
loss function. In this paper, we focus on the following three models with the lowest
time complexity and space complexity.
TransE. TransE model [4] embeds all nodes and edges into Rn , and assumes that
h + l ≈ t if h, r, t is in the graph. So the scoring function is:
where p is either 1 or 2.
The embeddings are learnt by minimizing the following loss function over the train-
ing set D:
max(γ + ||h + r − t||p − ||h + r − t ||p , 0),
h,r,t∈D h ,r,t ∈Dh,r,t
where γ is a margin hyperparameter and Dh,r,t is the corruption set of h, r, t defined
by:
Dh,r,t = {h , r, t|h ∈ V } ∪ {h, r, t |t ∈ V }.
DistMult. DistMult model [20] embeds all nodes and edges into Rn , and adopts the
following scoring function:
ComplEx. ComplEx model [18] embeds all nodes and edges into the complex vector
space Cn . In order to differentiate the head embedding and tail embedding of the same
node, it adopts an asymmetrical scoring function:
Given the training set D, a negative set D is generated as a set of false facts. Then
the embeddings are learnt by minimizing the following loss function:
log(1 + exp(−φ(h, r, t, Θ)))+
h,r,t∈D
log(1 + exp(φ(h, r, t, Θ))) + λ||Θ||22 ,
h,r,t∈D
Suppose Θ has been learnt from some model whose scoring function is φ. When
scoring a new candidate triplet h, r, t, refine φ to be negative infinity if h ∈ H(r) or
t ∈ T (r). In other words, if the type of h or t does not match with the type of r, then
this potential link should be ignored in the knowledge graph completion process.
4 Experiments
A knowledge graph of customers in DBS Bank has been constructed where the eval-
uation of the three most efficient models will be conducted. The models are evaluated
with different performance measurements and drilled down to relation level.
4.1 Datasets
To evaluate the models on real data, the knowledge graph constructed with DBS cus-
tomers involves money transfer between customers, demographics, credit cards and
merchants.
Specifically, there are 6 types of node and 6 types of relation in the knowledge
graph. Node types are listed in Table 2 with description. Table 3 describes the relations
with corresponding percentage in the graph, showing that 60% of the links are from
relation with mcc.
The data source of the knowledge graph covers the first half year of 2018. In other
words, only transactions happened between Jan 2018 and Jun 2018 are captured in the
graph, and all demographics information are updated until Jun 2018.
Most of the previous works were evaluated on FB15K and WN18 datasets pro-
vided by [4], containing 14,951 and 40,943 nodes respectively. FB15K is a collection
Knowledge Graph Completion in Consumer Banking 11
of general facts retrieved from Freebase [1] and WN18 is a collection of lexical relation
between words retrieved from WordNet [11]. To compare the performance on datasets
with comparable scale, our knowledge graph is built on a subset of customers and split
into training, validation and testing datasets on a ratio of 8:1:1. The metadata is sum-
marized in Table 4.
triplet xi , define its head corruption set by replacing its head node with any other node
in the graph:
C h (xi ) = {h, ri , ti |(h ∈ V ) ∧ (h = hi )}.
And define its tail corruption set by replacing its tail node with any other node in
the graph:
C t (xi ) = {hi , ri , t|(t ∈ V ) ∧ (t = ti )}.
Hence a good model is supposed to score xi higher than its corruptions. Let φ be
the scoring function, then compute the rank of xi in its corruptions using φ as follows:
rankih = 1 + I(φ(xi ) < φ(x)),
x∈C h (xi )
and
rankit = 1 + I(φ(xi ) < φ(x)),
x∈C t (xi )
and HITS@k is defined as the portion of test triplets whose rank is no larger than k:
1
HIT S@k = [I(rankih ≤ k) + I(rankit ≤ k)].
2|T |
xi ∈T
Since the target of the models is to give higher score to links existing in the graph, it
should not be considered as a mistake to score a test triplet xi lower than its corruption
x if x is also valid in the graph. For this reason, a filtered version of MRR and HITS@k
is defined using refined corruptions by removing valid triplets:
For each model in the experiment2 , we test with different initial learning rates among
{0.05, 0.01, 0.005, 0.001} and different embedding dimensions among {50, 100, 150,
200, 250, 300}, and the best combination will be selected according to MRR (Raw).
Then the performance will be compared across different models with and without
type restriction at both overall level and relation level for “Raw” and “Filtered” mea-
surements.
4.4 Comparison
The performance of TransE, DistMult and ComplEx models with and without type
restriction is shown in Table 5, which is much lower than the performance on WN18
and FB15K datasets as reported in [18].
For models without type restriction, ComplEx gives the best performance on MRR
and HITS@1 for both Raw and Filtered and HITS@3 Filtered, and TransE is the best for
HITS@10 for both Raw and Filtered and HITS@3 Raw. It shows that without type restric-
tion, ComplEx is better at identifying only few most possible triplets. In other words, if
the aim is to obtain a small number of new links, ComplEx should be used because it has
the best HITS@1 performance. On the other hand, with a larger number of new links to
pick out, TransE model should be used because it has the best HITS@10 performance.
It is obvious that the performance has been uplifted significantly by type restriction.
With type restriction, TransE model outperforms DistMult and ComplEx for almost all
measurements except Filtered HITS@3 and Filtered HITS@10.
The relation level Filtered MRR performance for models with type restric-
tion is shown in Table 6. ComplEx model has the best performance on relation
2
The implementation uses public code at https://github.com/mana-ysh/knowledge-graph-
embeddings with reduced running time.
14 D. Shao and R. Annam
with occupation, with address and with card, while TransE model performs the
best for relations with mcc and transf er to where TransE model is the best.
As in the graph each customer has only one occupation, one address and one age,
so with occupation, with address and with card are Many-to-One relations. On the
other hand, as one customer could have more than one merchant type and transfer to
more than on customer, and one merchant type could be linked to multiple customers,
with mcc and transf er to are Many-to-Many relations. Hence, ComplEx is more
suitable for Many-to-One relations and TransE is more suitable for Many-to-Many rela-
tions. Since the relation with mcc is dominating the graph as shown in Table 3, TransE
has the best overall performance.
It is worth pointing out that transf er to is the only relation in the graph whose head
node and tail node are from the same type, and ComplEx performs much worse than the
other two models on this relation. So, ComplEx should be avoided for graphs with large
portion of Many-to-Many relations or relations connecting nodes with the same type.
5 Use Case
The link prediction on a knowledge graph has many useful applications in consumer
banking sector. It uncovers relations which is unobserved now but has high probability
to be a fact.
Generally, there are two types of unobserved links. One type is links existing beyond
DBS data. For example, there is no fund transfer between two customers within DBS,
but they may have fund transfer through other banks. The other type is links to be true.
For example, a customer does not have DBS credit card. But based on the connections
Knowledge Graph Completion in Consumer Banking 15
from this customer to other customers and products, there is some card suitable for
this customer. With the knowledge graph completed using link prediction, we can get a
better understand about what our customers intend to do.
Suppose we have a sample knowledge graph completed using link prediction as
illustrated by Fig. 1. Solid edges in the graph are existing ones and dashed edges are
obtained from link prediction.
5.2 Member-Get-Member
Member-Get-Member (MGM) is an important channel for the bank to acquire new trea-
sures customers. The targeting customers of MGM are customers who have treasures
products and have many first-degree customer connections.
As in the example, Bob has treasures investment and he connects to Alice. Without
link prediction, he has only 1 first-degree customer connection which is Alice. However,
he has high probability to connect Carlo given by link prediction. There may be transfers
from Bob to Carlo through other banks. So, it is more reasonable to count Bob’s number
of first-degree customer connections as 2.
Targeting customers with more first-degree customer connections including pre-
dicted links, it can capture intended connections of customers, making MGM more
profitable.
6 Conclusions
Extensive experiments have been conducted using DBS datasets for different models.
The performance is much lower than that on the public datasets.
Type restriction can improve the model performance significantly. With type restric-
tion, ComplEx model performs best on Many-to-One relations while TransE model has
the best performance on Many-to-Many relations. Additionally, ComplEx model has
very low performance on relations linking nodes with the same type.
16 D. Shao and R. Annam
References
1. Bollacker, K.D., Evans, C., Paritosh, P., Sturge, T., Taylor, J.: Freebase: a collaboratively
created graph database for structuring human knowledge. In: Proceedings of the ACM SIG-
MOD International Conference on Management of Data, pp. 1247–1250 (2008)
2. Bordes, A., Weston, J., Collobert, R., Bengio, Y.: Learning structured embeddings of knowl-
edge bases. In: Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence
(2011)
3. Krompaß, D., Baier, S., Tresp, V.: Type-constrained representation learning in knowledge
graphs. In: International Semantic Web Conference, pp. 640–655 (2015)
4. Bordes, A., Usunier, N., Garcı́a-Durán, A., Weston, J., Yakhnenko, O.: Translating embed-
dings for modeling multi-relational data. In: Advances in Neural Information Processing
Systems, NIPS, pp. 2787–2795 (2013)
5. Dettmers, T., Minervini, P., Stenetorp, P., Riedel, S.: Convolutional 2D knowledge graph
embeddings. In: Proceedings of the Thirty-Second AAAI Conference on Artificial Intelli-
gence (2018)
6. Dong, X., et al.: Knowledge vault: a web-scale approach to probabilistic knowledge fusion.
In: The 20th ACM SIGKDD International Conference on Knowledge Discovery and Data
Mining, pp. 601–610 (2014)
7. Ebisu, T., Ichise, R.: TorusE: knowledge graph embedding on a lie group. In: Proceedings of
the Thirty-Second AAAI Conference on Artificial Intelligence (2018)
8. Lehmann, J., et al.: DBpedia - a large-scale, multilingual knowledge base extracted from
Wikipedia. Semant. Web 6(2), 167–195 (2015)
9. Lin, Y., Liu, Z., Sun, M., Liu, Y., Zhu, X.: Learning entity and relation embeddings for
knowledge graph completion. In: Proceedings of the Twenty-Ninth AAAI Conference on
Artificial Intelligence, pp. 2181–2187 (2015)
10. Mikolov, T., Sutskever, I., Chen, K., Corrado, G.S., Dean, J.: Distributed representations of
words and phrases and their compositionality. In: Advances in Neural Information Process-
ing Systems, NIPS, pp. 3111–3119 (2013)
11. Miller, G.A.: Wordnet: a lexical database for English. Commun. ACM 38(11), 39–41 (1995)
12. Nickel, M., Tresp, V., Kriegel, H.-P.: A three-way model for collective learning on multi-
relational data. In: Proceedings of the 28th International Conference on Machine Learning,
ICML, pp. 809–816 (2011)
13. Nickel, M., Rosasco, L., Poggio, T.A.: Holographic embeddings of knowledge graphs. In:
Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, pp. 1955–1961
(2016)
14. Qian, W., Fu, C., Zhu, Y., Cai, D., He, X.: Translating embeddings for knowledge graph
completion with relation attention mechanism. In: Proceedings of the Twenty-Seventh Inter-
national Joint Conference on Artificial Intelligence, IJCAI, pp. 4286–4292 (2018)
15. Shi, B., Weninger, T.: Proje: embedding projection for knowledge graph completion. In:
Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence, pp. 1236–1242
(2017)
16. Socher, R., Chen, D., Manning, C.D., Andrew, Y.N.: Reasoning with neural tensor networks
for knowledge base completion. In: Advances in Neural Information Processing Systems,
NIPS, pp. 926–934 (2013)
17. Toutanova, K., Chen, D.: Observed versus latent features for knowledge base and text infer-
ence. In: Proceedings of the 3rd Workshop on Continuous Vector Space Models and their
Compositionality, pp. 57–66 (2015)
18. Trouillon, T., Welbl, J., Riedel, S., Gaussier, É., Bouchard, G.: Complex embeddings for
simple link prediction. In: Proceedings of the 33nd International Conference on Machine
Learning, ICML, pp. 2071–2080 (2016)
Knowledge Graph Completion in Consumer Banking 17
19. Wang, Z., Zhang, J., Feng, J., Chen, Z.: Knowledge graph embedding by translating on
hyperplanes. In: Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intel-
ligence, pp. 1112–1119 (2014)
20. Yang, B., Yih, W., He, X., Gao, J., Deng, L.: Embedding entities and relations for learning
and inference in knowledge bases. CoRR, abs/1412.6575 (2014)
Financial Technology and Natural
Language Processing
Financial Technology and Natural Language
Processing
Abstract. The 1st FinNLP workshop was held in conjunction with IJCAI-
2019. There were 13 oral presentations, 8 posters, and 71 registered partici-
pants from both academia and industry. In this paper, we will introduce the
topics in the workshop and illustrate the future research directions for NLP in
FinTech applications.
Applying artificial intelligence (AI) to improve the efficiency of humans is one of the
trends recently. In this workshop, we focus on the problems of the financial industry.
The followings are the topics discussed in this year FinNLP. Many works dealt with
market information prediction via end-to-end models. However, if machines plan to
capture the fine-grained information embedded in the documents, more in-depth works
are needed. Explainable AI is one of the rise topics in the hype cycle for AI. Extracting
the cause-effect is a potential approach for explaining the outcomes of the models. The
best paper, entitled “Economic Causal-Chain Search using Text Mining Technology”,
is related to this topic. Evaluating the rationales behind the decisions is an additional
issue after extracting the reason. For instance, different rationales may cause different
influence to the results.
Legal issues are also important for enterprises, especially, for the department of
compliance. Leveraging machine learning techniques, the enterprises can predict the
possible litigations based on the multi-dimensional information of the companies. The
fine-grained analysis of the contracts or other legal instruments could help understand
the background knowledge of the models and thus further improve the performance.
Know your customers (KYC) is also an important topic for enterprises, particularly, for
the bank. Evaluating the credit and the trustworthiness of the customers is a routing
work in the bank. With machine learning and natural language understanding
methodologies, the accuracy of the evaluation could be improved. Anti-money
Another random document with
no related content on Scribd:
Papa gave him a little counsel in a low tone of voice, but I do not
think he listened very attentively. He was a boy without a bit of
sentiment or tender regard. He merely sang out—“Good-bye, old
chap,” to Louis, and though he thanked us for our kindness, it was
only from a gentlemanly instinct. Then he sprang into the stage and
was off.
“I do not know whether I should like to have such a son or not,”
mamma said slowly, as we entered papa’s study. “He is bright and
manly and entertaining, but he leaves you with a feeling that out of
sight is out of mind.”
“I have tried to sow a little good seed;” yet papa shook his head
gravely.
“But you are afraid it is in sandy ground;” Fan added, with a touch
of comforting sweetness in her voice. “I haven’t much faith in its
bearing fruit, and yet I do believe he has come to have more
consideration for Louis. He has not tormented him half as much
lately. That would be one point gained.”
“Yes. After all, I have more hope of Louis. The struggle will be
much harder, for his temperament and his health are against him,
but he will be steadier in anything he undertakes. I have become
deeply interested in both of them, and I do not feel as if it was going
to end here.”
It seemed as if the day was to be rendered memorable for us. In
the midst of the talk came a sudden hard ring. I answered it and
found Mr. Fairlie’s man with a frightened look in his face.
“If you please, Miss—is your father in?” he asked.
“Papa!”
He came at the summons.
“The master is—very bad, sir. They want to see you right away.
Mister Dick is taking it very hard.”
“Mr. Fairlie!” exclaimed papa in amaze.—“Why, I saw him
yesterday, and well.”
“He’s been rather queer in his head for two or three days. It was
the sun or something.—And about midnight he was taken. The
Doctor has given him up now.”
“Yes,” said papa, bewildered. “I’ll be there directly.”
“I’ve the wagon here for you, sir.”
He just kissed mamma and went without another word. Such calls
left no room for discussions.
“It cannot be possible!” ejaculated Fan.
“Mrs. Fairlie and Kate away!” said mamma. “How very sad.”
We had not the heart to talk about it and separated for our
morning’s employment. School had begun again, so I made the
children ready. Nelly had just entered the Seminary. Then I put my
rooms in order while Fan assisted in the kitchen. Tabby came up
stairs followed by her small gray and white kitten, who was a puffy
ball of frolic. She glanced around the room in a curious, complacent
fashion.
“Yes, Tabby,” I said, “the plague of your life has departed. Mrs.
Whitcomb will be here next, and you know she is fond of you, so
your troubles are ending. I don’t believe we have learned to like boys
so very much, after all.”
“No,” returned Tabby, with a grave whisk of the tail, while the kitten
made a vigorous attack on the bits of sunshine quivering through the
great sycamore leaves.
I went down stairs and sewed awhile in the nursery. Dinner came,
but no papa. Louis had returned from his drive and looked very
cheerful. We could not wait on account of the children, and
unconsciously his prolonged absence gave us a little hope.
It was dashed down presently. The church bell began to toll. We
glanced at each other in a startled way.
“Poor Dick!” said Fan, turning her head, and I knew her eyes were
full of tears. I could not help a curious thought. What if this sorrow
should bring them together?
Miss Churchill made us a nice long call in the afternoon, and
before she had gone papa returned. Dick had begged him to stay
and go to the station for Mrs. Fairlie who had just come, and do
several other special errands for him. The ladies had stopped on
their homeward way at the house of a cousin in Bridgeport, and were
thus easily reached by telegraph.
“What a terrible shock!” exclaimed Miss Churchill. “A man in
almost perfect health, too; though Dr. Hawley I believe mentioned his
having some trouble with his heart. Was that the cause?”
“I have no doubt it helped materially. He had complained of a dull,
heavy headache for two or three days, and yesterday he was out in
the sun which appeared to affect him a good deal. At midnight he
was taken with paralysis. But brief as the time was it found him
ready. He seemed to have gleams of consciousness and knew me at
intervals. His trust was staid upon God, and there was no fear, no
shrinking.”
“He has been a good, upright man. Kenton always esteemed him
highly.”
“He was more than that, Miss Churchill, he was an earnest
Christian. If the household had been of one mind, workers in the
vineyard, he would have lived a fuller and more joyous christian life.
But we are to work our way through hindrances. God gave him grace
and strength and perfected him in good deeds. I feel as if I had lost
my mainstay in the church. He was not a man of many words, but
you could rely upon him to the uttermost. And though I shall grieve
for a true and staunch friend, I shall also rejoice that he has gone to
his reward, better far than any earthly happiness.”
“You loved him very much,” said Miss Churchill, deeply moved.
“I did indeed.”
“The loss is dreadful to his family.”
“My heart ached for Richard. He and his father were tender
friends, and the watching through long hours, the not being able to
give him up, was agonizing in the extreme. Mrs. Fairlie was stunned
by the suddenness.”
“I wonder if I could be any—comfort to her?” Miss Churchill
questioned slowly.
“I wish you would call to-morrow,” said papa. “I don’t know but I
shall have to come to you and your brother now.”
“I am sure I should be glad to give you any assistance in my
power. I have been thinking lately that we live quite too much for
ourselves.”
“For the night cometh in which no man can work,” said papa
solemnly.
An awe fell over us all. One and another dropped in to wonder at
the occurrence. Sudden deaths always shock a community greatly.
Even the children did not want to play but sat on the porch steps and
looked into vacancy. Louis went up stairs directly after supper, but I
heard him pacing his room restlessly. I had put the little ones to bed
and was going down stairs when he called.
“Did you want anything?” I inquired.
“No—that is—are you busy?”
“Not especially.”
“I am going away so soon;” he said apologetically.
“And if I can do anything for you, I shall be glad to,” I made answer
cheerfully. “Shall I come in and read?”
“Thank you—I don’t care about that, I am in an odd, inconsequent
mood to night. Suppose you talk to me? I believe your voice has a
soothing effect.”
“Let us go down on the porch. It is cooler.”
“Where are the others?”
“Papa and Fanny have gone for a call. Mamma is in the nursery.”
“O, I wanted only you.”
“Come down then.”
I brought an easy chair out on the porch, and dropped into my own
small rocker. Tabby came along and crawled in my lap, turning round
three times and settling herself regardless of the welfare of her small
child, though I dare say she was asleep in some one’s slipper. The
moon was nearly at its full and made silvery shadows through the
interstices of the vines. The dewy air was fragrant and the night
musical with chirp and hum of countless insects.
“It is quite a relief to be rid of Stuart,” he began presently. “And
when I am gone you will doubtless feel still more comfortable.”
“I think you are quite comfortable to get on with now;” I said
cheerfully.
“Which implies—there was a time. Miss Endicott, do you think I
have improved any?”
“I do not know as it would be hardly fair to judge you by the first
week or two. You were on the eve of a severe illness, with your
nervous system completely disorganized.”
“But since then—be honest?”
“I think you have been pleasanter, more considerate, not so easily
ruffled;” I answered slowly.
“Please don’t fancy me fishing for compliments.”
“Compliments from me would not be so very flattering to one’s
vanity. They do not carry weight enough.”
“You believe that one could overcome—any fault?” after a pause
between the words.
“With God’s help—yes.”
“Without God’s help—what then?”
I was always so afraid of going astray in these talks. I could feel
what I meant, but I could not explain it clearly.
“‘Every good and perfect gift cometh from God,’” I made answer.
“And the desire to be better or stronger, to overcome any fault, must
proceed from Him.”
“Then why doesn’t he make Christians perfect?”
“God gives us the work to do. He says, ‘My grace shall be
sufficient for thee.’ Therefore we are to strive ourselves. He shows
us the right way, but if we seek out other paths, or if we sink into
indolence waiting for an angel to come and move our idle hands or
stir up our languid wills, can we reasonably blame Him?”
“I had not thought of that, I must confess. I had a fancy that—
religion did all these things for you.”
“What then is the Christian warfare? You know that grand old St.
Paul had to fight to the last, that he might not be a cast-away. Yet I
think no one ever doubted the genuineness of his conversion.”
“But if a man of his own determination, resolved, he could do a
great deal.”
“I should be weak to deny it. People have achieved heroic
victories, suffered pain and shame and death bravely for pride, or
some chosen idea. Only when it is done for the sake of Him who
saved us, it becomes so much the more noble. It is obeying Him.”
“Is it an easy thing to be good, Miss Endicott?”
“Not for every one,” I said.
“You admit that natures are different?”
“I do, cheerfully. Some people have very little self-control, others a
great deal. But it is strengthened by use, like a limb.”
“I have very little?”
“I did not say that.”
“But you know I have.”
“Papa said your temperament and your health were against you!”
“Did he say that?” was the eager question. “Well there are a
hundred things—I sometimes have such headaches that I can hardly
tell where I am, and if anything bothers me I feel as if I could stamp
on it, crush it out of existence. And if it is a person—”
“Oh,” I cried, “don’t please! That is murder in one’s heart.”
“And when any one annoys Stuart he laughs at him, flings, jeers
and exasperates. It is his way, yet every one thinks he has a lovely
temper. He makes others angry. I have seen him get half a class by
the ears, and in such a mess that no one knew what was the matter.
—I do not believe I ever in my life set about making another person
angry. But I cannot stand such things. They stir up all the bad blood
in me.”
“So you need patience, first of all.”
“But I can’t stop to think.”
“Ah, that is just it. Stopping to think saves us. And when we have
our great Captain to remember, and are endeavoring to walk in the
path He marked out for us, it makes it easier. We are trying for the
sake of one we love.”
“What else do I want?”
“Don’t ask me, please,” I entreated.
“Yes. I shall not let you evade me. Write me some copies to take
with me. Patience—what next?”
“Cheerfulness;” seeing that he compelled me to it. “Your nature is
morbid and melancholy. Just try to think that people will like you.
“But they do not.”
“Then you must give them something to like. Suppose we all hid
away our brightness?”
He laughed.
“It would be a rather blue world. But to try for admiration.”
“You don’t try for admiration. You give freely of the very best you
have. You remember about the little boy who hid his cake away until
it was mouldy and spoiled?”
“I believe you always give of the best here. And you never seem to
have any lack.”
“Did you ever break off a sprig of lemon verbena? Three new
shoots come in its place. When I was a little girl mamma explained it
to me, and said that if you nipped off one bit of pleasantness for a
friend or neighbor, something grew instantly for the next one. You
never give away all your joy and good feeling.”
He sighed a little, and said slowly—
“I believe I shall begin with my temper. I have always known that it
was bad, and expected to keep it all my life, but if it could be made a
little more reasonable!”
“I am sure it can, if you will try. It is hard work to be fighting
continually, to be on your guard against surprises, and sometimes to
have your best efforts misunderstood, yet it seems to me a grand
thing to gain a victory over one’s self.”
“You make it so;” he replied in a half doubtful tone.
“I wish you could be good friends with papa. He is so much wiser,
and can explain the puzzles. When you came to know him well you
would like him, you couldn’t help it.”
“Sometime—when I want such a friend;” he answered a trifle
coldly.
The voices sounded on the walk just then, and in a few moments
they came up. We had no special talk after that.
Mamma went over to Mrs. Fairlie’s the next day and met Miss
Churchill there. Kate had been in violent hysterics all night. They
appeared so utterly helpless. What should they do about black?
There wasn’t any thing decent in Wachusett! And could Mrs. Fairlie
find a long widow’s veil any where? There would not be time to send
to the city.
“I am quite sure that Mrs. Silverthorne has one. Hers was very
beautiful and she never wore it but a little; and a plain bonnet will
do.”
“Thank you, Miss Churchill. How kind you are. But I cannot
understand why this grief should come upon me.”
“God’s ways are not as our ways;” said mamma.
“But Mr. Fairlie was needed so much. I don’t know how I can live
without him!”
Mamma and Miss Churchill soothed and tried to comfort. Each
took a few orders on leaving.
“My objection to mourning is just this,” said Miss Churchill, when
they were seated in her basket phaeton. “In the midst of your grief
you have to stop and think wherewithal you shall be clothed. Dress-
makers and milliners are your constant care for the first month.”
“The fashion of this world;” mamma replied a little sadly.
That afternoon Louis received a telegram from his cousin. He
would meet him the next noon at the station in a through train, that
there might be no lost time. He only packed a valise, as his trunk
would be sent to Wilburton. We said our good-byes in quite a friendly
fashion. He appeared really grateful and sorry to leave us. Papa
went to the station with him and returned in an unusually grave
mood.
We kept up to the tense point of excitement until after Mr. Fairlie’s
funeral. It was largely attended, and very solemn and affecting.
Indeed, nearly every heart ached for Kate and her mother.
“But I do believe Dick suffers the most;” Fanny said. “I never saw
any one so changed in a few days.”
Afterward the will was read. The farm was bequeathed to Richard.
Stocks, bonds and mortgages were divided between Mrs. Fairlie and
Kate, who were thus made quite rich women. They could go to
Europe now.
I found myself wondering a little what Mr. Fairlie’s life would have
been with different surroundings. The Fairlies in their way were as
old and as good a family as the Churchills, only they did not happen
to settle at the West Side, and had gone a little more into active
business. But they did not lay claim to any special position or
grandeur. This had always seemed to mortify Mrs. Fairlie somewhat.
“Mr. Fairlie is so old-fashioned,” she would say complainingly. “There
was no getting him out of the one groove.” She wanted to make a
show, to have people admit that she was somebody.—She went to
church regularly and would have been much offended not to have
been considered an important member. She gave to the Christmas
and Easter feasts and adornings, but for the poor or the needy sick
she rarely evinced any sympathy. Her duty stopped at a certain
point, the rest of her time, money, and interest was distinctly her
own. So the husband and wife lived separate lives, as it were.
Would Richard’s fate repeat the same confused and tangled
story? No doubt his mother would desire him to marry well in worldly
point of view. She might even object to Fan on the score of money.
Would he have the courage to suit himself? For what he needed was
a sweet, domestic woman with the culture that did not disdain every
day matters. His tastes were simple and homelike, yet he was by no
means dull. He wanted a woman to honor him, to put him in his true
position as head of the family.
Would Providence bring him happiness, or discipline only?
CHAPTER XII.