共指消解是确定上下文中的代词或名词短语所指的具体对象或实体，是自然语言处理(Natural Language Processing，NLP)的基本任务之一，对理解文本语义具有重要意义。现有的方法主要集中在一般领域的代词、所有格和名词短语的解析上，针对法律领域的研究较少。为了更好地学习法律文本中的知识，并消除共同指代现象，提出一种基于图神经网络的法律文本共指消解模型(Graph Neural Network for Coreference Resolution，CR-GNN)。所提CR-GNN可以促进法律文本挖掘中的一系列后续任务。利用预训练语言模型和双向门控循环单元(Bidirectional Gate Recurrent Unit，BiGRU)对法律文本进行编码;使用基于元任务的动 态图卷积网络(Meta Dynamic Graph Convolutional Network，MDGCN)整合实体之间的引用关系;使用前馈神经网络(Feed-Forward Neural Network，FFNN)和Biaffine模型为候选对进行加权评估。CR-GNN可以有效识别实体之间的引用关系，并对实体依赖关系进行建模。在法庭记录文件数据集上进行大量实验，结果表明所提CR-GNN模型达到89。76％的F1分数，均高于现有基准模型。
Coreference Resolution Model for Legal Texts Based on Graph Neural Network
Coreference resolution is to determine the specific object or entity that pronoun or noun phrase in context refers to.It is one of the basic tasks of Natural Language Processing(NLP),and it is of great significance to the understanding of text semantics.Existing methods mainly focus on the analysis of pronouns,possessive and noun phrases in generic research areas,but there is little research in the legal area.In order to better learn the knowledge from legal texts and eliminate the phenomenon of coreference,a model for legal texts based on Graph Neural Network for Coreference Resolution(CR-GNN)is proposed.The proposed CR-GNN can facilitate a series of subsequent tasks in legal text mining.Firstly,pre-trained language model and Bidirectional Gate Recurrent Unit(BiGRU)are used to encode legal texts.Secondly,Meta Dynamic Graph Convolutional Network(MDGCN)is used to integrate reference relations between entities.Lastly,the Feed-Forward Neural Network(FFNN)and Biaffine model are used for weighted evaluation of candidate pairs.CR-GNN can effectively identify reference relationships between entities and model the dependency relationships of entities.A large number of experiments are conducted on a court record dataset,and results show the proposed CR-GNN model achieves 89.76％Fl-score,which is higher than existing baseline models.
NLPcoreference resolutionlegal textpre-trained language modelGNN
国家科技期刊平台
NSTL
万方数据
