基于实体层次结构的文档级别关系抽取  

Document level relation extraction based on entity hierarchical structure

作　　者：李强 李实[1] LI Qiang;LI Shi(College of information and Computer Engineering,Northeast Forestry University,Harbin 150040,China)

机构地区：[1]东北林业大学信息与计算机工程学院,黑龙江哈尔滨150040

出　　处：《计算机工程与设计》2023年第4期1081-1087,共7页Computer Engineering and Design

基　　金：中央高校基本科研业务费基金项目(2572019BH03)。

摘　　要：针对目前文档级别关系抽取主要关注实体间的逻辑推理,未充分利用实体间的层次语义信息问题,提出一种基于实体层次结构的文档级别关系抽取模型。考虑多句文本中实体间的交互,将实体构建为文档图并使用图卷积神经进行信息传播;通过实体间的上下位关联构建实体层次树,使用注意力机制将层次语义信息融入实体;为降低模型对实体表面信息的关注,使用实体类型对实体词进行替换。实验结果表明,在大规模文档级别关系抽取数据集上实体语义信息增强的方案能够有效提高文档级别关系抽取的效果。Aiming at the problem that the current document level relation extraction mainly focuses on the logical reasoning between entities and does not make full use of the hierarchical semantic information between entities,a document level relation extraction model based on entity hierarchy was proposed.Considering the interaction between entities in multi sentence text,the entities were constructed as document graph,and the graph convolution neural network was used for information dissemination.The hierarchical tree of entities was constructed through the upper and lower association between entities,and the hierarchical semantic information was integrated into entities using the attention mechanism.To reduce the attention of the model to the entity surface information,entity types were used to replace entity words.Experimental results show that the scheme of entity semantic information enhancement on large-scale document level relation extraction dataset can effectively improve the perfor-mance of document level relation extraction.

关 键 词：文档级别 关系抽取 实体层次结构 逻辑推理 注意力机制 文档图 信息融合 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]