基于预训练模型的法律文本类案匹配研究  

Research on Legal Text Matching Based on Pre-training Model

作　　者：余传明[1] 江一帆 YU Chuanming;JIANG Yifan(School of Information and Safety Engineering,Zhongnan University of Economics and Law,Wuhan 430073)

机构地区：[1]中南财经政法大学信息与安全工程学院,武汉430073

出　　处：《科技情报研究》2023年第3期13-25,共13页Scientific Information Research

基　　金：湖南省科技创新人才计划项目“湖南科技情报服务创新研究群体”(编号:2021RC5029);国家自然科学基金项目“面向跨语言观点摘要的领域知识表示与融合模型研究”(编号:71974202);中南财经政法大学中央高校基本科研业务费专项资金资助项目(编号:2722021AJ011)。

摘　　要：[目的/意义]文章旨在解决传统的短文本匹配模型难以适用于法律类案检索等长文匹配的问题。[方法/过程]针对法律案例匹配任务,本文提出了一种基于预训练的法律文本类案匹配LTMR模型。在编码层通过RoFormer和法律特征提取器对法律案例进行编码,在推理层利用交互注意力和自注意力机制来进一步提取长文本的上下文和交互信息。将所提出的模型应用到CAIL2019-SCM数据集上,开展实证研究。[结果/结论]相对于基线方法,LTMR模型在CAIL2019数据集上均取得了最优结果,研究结果对于促进法律案例匹配的应用具有借鉴意义。[Purpose/significance]This study aims to solve the problem of traditional short text matching models being difficult to apply to long text matching tasks such as legal case retrieval.[Method/process]For the task of legal case matching,this paper proposes a Legal Text Matching model based on RoFormer(LTMR).In the coding layer,the legal case is encoded through the RoFormer model and the legal feature extractor.In the reasoning layer,the context and interactive information of long text are further extracted by using interactive attention and self-attention mechanisms.We conducted the empirical research by applying the proposed model to the CAIL2019-SCM dataset.[Result/conclusion]Compared to the baseline methods,the LTMR model achieved the best results.The research sheds light on promoting the application of legal case matching.

关 键 词：类案匹配 预训练模型 法律文本 文本语义匹配 

分 类 号：G358[文化科学—情报学]