Legal Element-oriented Modeling with Multi-view Contrastive Learning for Legal Case Retrieval Wang Zhaowei12

2025-05-02 0 0 419.91KB 10 页 10玖币

侵权投诉

Legal Element-oriented Modeling with Multi-view

Contrastive Learning for Legal Case Retrieval

Wang, Zhaowei1,2*

1Department of Computer Science and Engineering

The Hong Kong University of Science and Technology

Hong Kong Special Administrative Region, China

2Beijing Laiye Network Technology Co., Ltd

Beijing, China

Email: zwanggy@connect.ust.hk

Abstract—Legal case retrieval, which aims to retrieve relevant

cases given a query case, plays an essential role in the legal

system. While recent research efforts improve the performance

of traditional ad-hoc retrieval models, legal case retrieval is still

challenging since queries are legal cases, which contain hundreds

of tokens. Legal cases are much longer and more complicated

than keywords queries. Apart from that, the deﬁnition of legal

relevance is beyond the general deﬁnition. In addition to general

topical relevance, the relevant cases also involve similar situations

and legal elements, which can support the judgment of the

current case. In this paper, we propose an interaction-focused

network for legal case retrieval with a multi-view contrastive

learning objective. The contrastive learning views, including case-

view and element-view, aim to overcome the above challenges.

The case-view contrastive learning minimizes the hidden space

distance between relevant legal case representations produced by

a pre-trained language model (PLM) encoder. The element-view

builds positive and negative instances by changing legal elements

of cases to help the network better compute legal relevance.

To achieve this, we employ a legal element knowledge-aware

indicator to detect legal elements of cases. We conduct extensive

experiments on the benchmark of relevant case retrieval. Eval-

uation results indicate our proposed method obtains signiﬁcant

improvement over the existing methods.

Index Terms—legal case retrieval, contrastive learning, legal

elements enhancement

I. INTRODUCTION

Legal case retrieval, a specialized information retrieval

task, has drawn increasing attention rapidly, as legal artiﬁcial

intelligence can liberate legal professionals from a maze of

paperwork [1]. In countries with a Common Law system, like

the United States, searching previous relevant cases is essential

since judicial judgments are made according to past preceden-

tial judgments of relevant cases. In countries with a Civil Law

system, professionals could also use previous similar cases as

references. With the rapid growth of digitized legal documents,

it takes great effort to search for relevant cases manually.

The development of natural language processing (NLP) and

information retrieval (IR) has brought new opportunities to

the law system for searching in a deluge of legal materials.

Legal case retrieval is the research problem about applying

NLP and IR techniques to searching legal domain documents.

*The work is partially done during internship at Beijing Laiye Network

Technology Co., Ltd

Case A: In 2010, PersonX borrowed 10k dollars from PersonY at

10% interest rate. PersonY urged the PersonX for the principal and

interest, but the PersonX has not paid back so far.

Case B: In 2010, PersonX borrowed 10k dollars from PersonY at

40% interest rate. PersonY urged the PersonX for the principal and

interest, but the PersonX has not paid back so far.

Different legal element: different interest rates (10% vs. 40%)

TABLE I

AN EXAMPLE OF LEGAL ELEMENT. EVEN THOUGH THE TWO CASES ARE

SIMILAR,THEY ARE IRRELEVANT IN LEGAL CASE RETRIEVAL DUE TO THE

LEGAL ELEMENT:DIFFERENT RATES OF INTEREST.

Many works [2, 3, 4] recently tried to build legal retrieval

engines using links or citations between case documents and

legal statutes. Since the recent success of BERT [5], pre-

trained language models have drawn great attention in the ﬁeld

of legal case retrieval. The BERT-PLI model [6], a BERT-

based model with paragraph-level interactions, is designed for

legal documents. However, it suffers from the lack of long-

distance attention and limits the model performance [7] due

to breaking text into paragraphs. At the same time, many

models were proposed for the general IR task, but most of

these efforts [8, 9, 10] focused on matching between short

and long texts, which will cause complexity issues when being

applied between long texts. There is little literature for ﬁnding

methods to handle long text matching [11, 12]. Thus, matching

between long document pairs, which the legal case retrieval

belongs to, was less explored, requiring more research work.

Whether designed for the general or legal domain, existing

methods still face a few main challenges when applied to legal

case retrieval [6]:

(1) both the query and candidate cases involve long texts.

Cases in CAIL2019-SCM, a legal case retrieval dataset, con-

tain around 650 words on average. Long texts cause several

issues [12], including superﬁcial text understanding and high

memory consumption.

(2) the deﬁnition of legal domain relevance is beyond the

general deﬁnition. Relevant cases are not only related to

content but also can support judgments of each other [13].

General IR models only capture whether two pieces of texts

contain similar content but fail to consider their judgments [6].

Legal elements are crucial situations in a case (e.g. interest

arXiv:2210.05188v1 [cs.CL] 11 Oct 2022

rate, whether debit note exists, whether borrowers refuse to

repay the debt, and so on). They can help the model understand

legal cases’ judgments well [6]. They also play a pivotal role

in affecting experts’ judgment of legal relevance in manual

search. An example of legal element is shown in Table I. In

general IR, the two cases are considered as relevant texts due to

many shared words and phrases. In legal case retrieval, they are

irrelevant and have totally different judgments because Case B

has a surprisingly high rate of interest and becomes usury. We

can observe that legal elements have a signiﬁcant impact on

judgments and legal relevance. In practice, even though they

are essential for the deﬁnition of legal relevance, it is difﬁcult

to model legal elements in current neural networks [1].

To tackle the above challenges, we propose an interaction-

focused network with a Multi-View Contrastive Learning

objective. The whole model is entitled MVCL. The proposed

model builds word-level interactions based on pre-trained lan-

guage models (PLMs) and then aggregates these interactions

to get the matching representation between a query and a

candidate case. We also introduce a multi-view contrastive

learning objective, including case-view and element-view, to

tackle the challenges mentioned earlier. The case-view con-

trastive learning can help the model understand long texts

more thoroughly (Challenge 1). The element-view contrastive

learning builds positive and negative instances by changing

legal elements of cases. In this way, legal elements are

integrated into our neural network. We also introduce a legal

element knowledge-aware indicator to detect elements in legal

cases. Through legal element-oriented modeling (i.e., element-

view contrastive learning), the model can better understand

the query and candidate cases, decide whether two cases’

judgments can support each other, and compute legal relevance

(Challenge 2). Therefore, the MVCL model can learn legal

relevance instead of the general deﬁnition.

To evaluate our model, we conduct experiments with bench-

marks of legal case retrieval in CAIL2019-SCM. We achieved

state-of-the-art results in terms of automatic evaluation. In

summary, our main contributions include:

•We propose a multi-view contrastive learning objective

for legal case retrieval, which can effectively use the cases

and elements to align representations of legal cases.

•We ﬁrst introduce a legal element knowledge-aware indi-

cator as an external knowledge module of legal elements

to improve the model’s understanding of legal relevance.

•We empirically verify the effectiveness of MVCL on the

CAIL2019-SCM dataset and surpass 711 teams in the

China AI&Law Challenge1.

II. RELATED WORK

A. Traditional Ad-hoc Retrival

Ad-hoc retrieval has received a lot of attention in the past

few decades. Traditional models are built on the bag-of-words

model and measure the relevance between texts by comparing

1http://cail.cipsc.org.cn:2019/

tokens in texts, including BM25 [14] and VSM [15]. In-

evitably, they face the problem of sparse and high-dimensional

representation.

The recent development of deep learning has also inspired

applications of neural models in IR. Typically, most ap-

proaches can be categorized into two types: representation-

focused and interaction-focused [16]. To measure the relevance

for documents, models that are representation-focused (i.e.,

Siamese models) learn the representations of a query and a

document separately and then compare two representations by

computing distance with functions like cosine, dot product,

bilinear, or Euclidean distance. The advantage of these ap-

proaches is that shared parameters make Siamese architecture

smaller and far more efﬁcient. Nevertheless, representing an

entire text using a single vector is not sufﬁcient to capture

all crucial information. On the other hand, interaction-focused

approaches ﬁrst compute a query-document word-by-word

similarity matrix to build local interactions and then aggregate

these word-level matching results to generate the ﬁnal rele-

vance score. Well-known methods include MatchPyramid [17],

Match-SRNN [18], DRMM [8], and HCRN [19]. These mod-

els get better performance in traditional ad-hoc retrieval tasks

due to the detailed comparison of queries and documents. Ad-

ditionally, there are neural models that combine representation-

focused and interaction-focused approaches [10, 20].

Regarding the length of query and document, those works

mentioned above mainly concentrate on matching between

short texts, or between short and long texts. When adopting to

matching between long texts, they don’t work well and lead

to huge memory cost. Matching between long texts was less

explored. In this line of work, SMASH RNN [11] is proposed

to learn representation using a hierarchical architecture of

RNN to model multiple abstraction levels of the document

structure, including words, sentences, and paragraphs. Yang

et al. [12] proposed a Transformer-based hierarchical archi-

tecture to better model document structure and to increase

maximum input text length of general Transformer-based

models.

B. Legal Case Retrieval

Legal case retrieval is still a challenging task, due to its con-

cept of relevance [13], professional terms and expressions, text

length [6] and logical structure behind natural languages [21].

Many works tried to build a retrieval engine using links or

citations between case documents and legal statutes [2, 3, 4].

Recently, Shao et al. [22] and Liu et al. [23] studied user

behaviors and practical search interactions between users and

systems and observed signiﬁcant differences between legal

case retrieval and general web search.

To attain the purpose of computing semantic content,

deep learning methods has been applied to Legal Case Re-

trieval. Tran et al. [24] builds an CNN-based model, which also

utilizes summarization information and lexical features. Zhong

et al. [1] adopted a few deep learning models to get a better

view of the current progress of Legal Case Retrieval, like

BiDAF [25], SMASH RNN [11] and so on. Shao et al. [6]

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

LegalElement-orientedModelingwithMulti-viewContrastiveLearningforLegalCaseRetrievalWang,Zhaowei1;2*1DepartmentofComputerScienceandEngineeringTheHongKongUniversityofScienceandTechnologyHongKongSpecialAdministrativeRegion,China2BeijingLaiyeNetworkTechnologyCo.,LtdBeijing,ChinaEmail:zwanggy@connect.ust...

展开>> 收起<<

Legal Element-oriented Modeling with Multi-view Contrastive Learning for Legal Case Retrieval Wang Zhaowei12.pdf

共10页,预览2页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Legal Element-oriented Modeling with Multi-view Contrastive Learning for Legal Case Retrieval Wang Zhaowei12

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: