基于多任务学习的民事案件判决预测方法  

CIVIL CASE JUDGMENT PREDICTION METHOD BASED ON MULTI-TASK LEARNING

作　　者：余连玮 马志柔[1] 刘杰[1] 叶丹[1] Yu Lianwei;Ma Zhirou;Liu Jie;Ye Dan(Software Engineering Center,Institute of Software,Chinese Academy of Sciences,Beijing 100190,China;University of Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]中国科学院软件研究所软件工程技术研发中心,北京100190 [2]中国科学院大学,北京100190

出　　处：《计算机应用与软件》2024年第1期18-25,共8页Computer Applications and Software

基　　金：国家重点研发计划项目(2018YFC0831302)。

摘　　要：针对民事案件预测中法律法规组合多样化的问题,提出一种基于多任务学习的民事案件判决预测方法。该方法采用多CNN融合、阈值设定等多种策略,利用案由和法条之间的依赖关系,实现民事案件的案由纠纷和相关法律法规的联合判决预测。基于中国裁判文书公开网的民事案件,构造了10万篇民事文书进行判决预测实验。实验结果表明,相比于传统的预测模型,该方法针对有依赖关系的预测任务更加合理和有效。Aiming at the problem of multiple combinations of laws and regulations for civil case prediction,this paper proposes a civil case judgment prediction method based on multi-task learning.It used a variety of strategies such as CNN model fusion and threshold setting,and used the dependence between legal disputes and legal articles to realize the joint judgment prediction of legal disputes and legal articles in civil cases.Based on the civil cases of China Judgment Online,a dataset of 100,000 civil cases was constructed,and multiple sets of experiments were performed on the dataset.Experimental results show that compared with traditional prediction models,this method is more reasonable and effective for the prediction task with dependency.

关 键 词：纠纷预测 法条预测 多任务学习 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]