融合局部上下文的双图文档级关系抽取方法  

Bi-graph-based Document-level Relation Extraction with Local Context Fusion

作　　者：闻克妍 纪婉婷 宋宝燕[1] WEN Keyan;JI Wanting;SONG Baoyan(College of Information,Liaoning University,Shenyang 110036,China)

机构地区：[1]辽宁大学信息学部,沈阳110036

出　　处：《小型微型计算机系统》2025年第3期535-541,共7页Journal of Chinese Computer Systems

基　　金：教育部产学合作协同育人项目(230701160261310)资助;辽宁省应用基础研究计划项目(2022JH2/101300250)资助;国家重点研发计划项目(2023YFC3304900)资助;辽宁省自然科学基金项目博士启动项目(2023-BS-085)资助.

摘　　要：文档级关系抽取是一项复杂的自然语言处理任务,旨在识别出文档中存在的实体,并预测实体之间的关系.相较于句子级关系抽取任务,文档级关系抽取面临更大的挑战,因为它需要考虑整个文档的语义信息和句子间的逻辑关系.针对这一任务,提出了一种融合局部上下文信息的双图推理方法(BRM)用于文档级关系抽取.该方法首先识别文档中的实体提及,并构造了一个提及级别的异构图来表示这些提及以及它们之间的关系.在获得提及级别的表示后,方法进一步构建了一个实体级别的推理图,通过聚合提及级别的信息来形成实体级别的表示,以判断实体之间的关系.该方法在文档级关系抽取公开数据集DocRED上进行了实验.实验结果表明,与现有的文档级关系抽取方法相比,该方法能够更准确地识别实体并预测它们之间的关系.Document-level relation extraction is a complex natural language processing task,which aims to identify the entities exit in a document and predict the relations between these entities.Compared with sentence-level relation extraction task,document-level relation extraction faces more challenges,since it needs to consider the semantic information of the whole document and the logical relationships between sentences.Aiming at this task,this paper proposes a bi-graph reasoning method(BRM)with local context fusion for document-level relation extraction.The proposed method first identifies entity mentions in a document and constructs a mention-level heterogeneous graph to represent these mentions and the relationships between them.After obtaining the mention-level representation,the method further constructs an entity-level reasoning graph to form the entity-level representation by aggregating the mention-level information to judge the relationships between entities.The proposed method is evaluated on the public document-level relation extraction dataset DocRED.Experimental results show that,compared with the existing document-level relation extraction methods,the proposed method can more accurately identify entities and predict the relations between them.

关 键 词：文档级关系抽取 局部上下文 双图推理 数据集成 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]