Augmenting Legal Judgment Prediction with Contrastive Case Relations

Proceedings of the 29th International Conference on Computational Linguistics, pages 2658–2667 October 12–17, 2022. 2658 Augmenting Legal Judgment Prediction with Contrastive Case Relations Dugang Liu1,2, Weihao Du1,2, Lei Li3, Weike Pan1,2,*, Zhong Ming1,2,* 1Shenzhen University, Shenzhen, China 2Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China 3Hong Kong Baptist University, Hong Kong, China dugang.ldg@gmail.com, {panweike,mingz}@szu.edu.cn Abstract Existing legal judgment prediction methods usually only consider one single case fact de- scription as input, which may not fully utilize the information in the data such as case rela- tions and frequency. In this paper, we propose a new perspective that introduces some con- trastive case relations to construct case triples as input, and a corresponding judgment pre- diction framework with case triples modeling (CTM). Our CTM can more effectively uti- lize beneficial information to refine the encod- ing and decoding processes through three cus- tomized modules, including the case triple mod- ule, the relational attention module, and the category decoder module. Finally, we conduct extensive experiments on two public datasets to verify the effectiveness of our CTM, includ- ing overall evaluation, compatibility analysis, ablation studies, analysis of gain source and visualization of case representations. 1 Introduction As an important component of legal intelligence in civil law systems, legal judgment prediction (LJP) has received a lot of attention and research in recent years (Chalkidis et al., 2019; Zhong et al., 2020). Given a case fact description, LJP usually includes three sub-tasks, i.e., law article prediction, charge prediction and terms of penalty prediction for this case (Xiao et al., 2018), and an example of LJP is shown on the left side of Figure 1. As an auxiliary tool to serve legal practitioners and people without professional knowledge in law, a more accurate method for LJP is necessary. The existing legal judgment prediction methods mainly include two lines of single-task modeling and multi-task modeling. The former usually fo- cuses on targeted modeling of a certain sub-task, such as introducing some more advanced network architectures (Chen et al., 2019a; Le et al., 2020) or *Co-corresponding authors more sources of information (Luo et al., 2017; Hu et al., 2018; Chen et al., 2019b). The latter takes multiple sub-tasks as a whole and uses a multi-task learning (MTL) framework for unified modeling. The most representative methods in this line aim to design different decoding structures, including MTL (Zhong et al., 2018) that ignores the inter-task dependency, TopJudge (Zhong et al., 2018) that considers unidirectional topological dependency among sub-tasks, and MPBFN (Yang et al., 2019) that considers bidirectional topological dependency. In this paper, we focus on the line of multi-task learning because it is more aligned with practical applications. Although the existing methods have shown promising results, as shown on the left side of Fig- ure 1, most of them only consider the fact descrip- tion of one single case as input when modeling. This form of modeling ignores the full utilization of the beneficial information contained in the data, such as the case relation and frequency …

Content truncated to 500 words