基于语义图的法律文本多跳阅读理解研究  

Semantic Graph Based Multi-hop Reading Comprehension of Legal Text

作　　者：林志强 杨帆 苏劲松[3] 吴旭阳[4] LIN Zhiqiang;YANG Fan;SU Jinsong;WU Xuyang(Department of Automation,School of Aerospace Engineering,Xiamen University,Xiamen,Fujian 361005,China;Xiamen Key Laboratory of Big Data Intelligent Analysis and Decision,Xiamen,Fujian 361005,China;Department of Artificial Intelligence,School of Informatics,Xiamen University,Xiamen,Fujian 361005,China;School of Law,Xiamen University,Xiamen,Fujian 361005,China)

机构地区：[1]厦门大学航空航天学院自动化系,福建厦门361005 [2]厦门市大数据智能分析与决策重点实验室,福建厦门361005 [3]厦门大学信息学院人工智能系,福建厦门361005 [4]厦门大学法学院,福建厦门361005

出　　处：《中文信息学报》2024年第11期91-101,共11页Journal of Chinese Information Processing

基　　金：国家自然科学基金(62173282);厦门市自然科学基金(3502Z20227180);科技部科技创新2030“新一代人工智能”重大项目(2021ZD0112600)。

摘　　要：目前基于图神经网络的机器阅读理解模型难以有效建模法律文书中的复杂关系。为此,该文提出了基于语义图的法律文书机器阅读理解方法(Semantic Graph Based Reader,SGB Reader)。其核心思想是通过语义依存分析构建以法律关键实体为中心的语义图和句子关系图,然后通过图嵌入来学习法律文书中的复杂关系。除此之外,SGB Reader还设计了两阶段的答案片段预测模块和答案类型联合预测模块来进一步提升模型的性能。实验结果表明,SGB Reader显著优于已有的图网络模型,在CJRC和CJRC 2.0数据集上分别取得了76.97%和65.39%的Joint F1分数。Current machine reading comprehension models based on graph neural networks cannot effectively model complex relationships in legal text.Therefore,this paper proposes the Semantic Graph Based Reader(SGB Reader)for legal text machine reading comprehension.The core idea is to construct semantic graphs and sentence relation graphs centered on legal key entities through semantic dependency parsing,and then learn the complex relationships in legal text through graph embedding.In addition,SGB Reader designs a two-stage answer span prediction module and an answer type joint prediction module to further improve the model's performance.Experimental results show that SGB Reader significantly outperforms other graph neural network models and EBF Reader,achieving Joint F1 scores of 76.97%and 65.39%on CJRC and CJRC2.0 datasets,respectively.

关 键 词：机器阅读理解 法律文本 图注意力网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]