针对长距离实体的双图路径推理模型  

Double Graph Path Inference Model for Long-Distance Entities

作　　者：祝涛杰 卢记仓[1] 周刚[1] 皮乾坤 丁肖摇 ZHU Taojie;LU Jicang;ZHOU Gang;PI Qiankun;DING Xiaoyao(Information Engineering University,Zhengzhou 450001,China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《信息工程大学学报》2024年第3期272-277,共6页Journal of Information Engineering University

基　　金：河南省自然科学基金(222300420590)。

摘　　要：文档中句间实体关系往往无法直接获取,现有方法通常利用语法知识及共指、邻接、共现等方式将文档构建为文档图,捕获实体之间的交互。然而图节点和图边数量及类型较多,极大地限制了模型的推理能力。因此,提出一种结构简单且推理效果更好的双图模型。首先,采用启发式规则提取提及交互和证据句,并基于此构建基于证据句的提及图和实体图;其次,利用注意力机制捕获实体图中实体节点之间的推理路径;最后,根据推理路径,采用合适的评分函数预测实体关系事实。在文档级通用领域数据集DocRED中的实验表明,所提模型取得了较好的效果。The entities relation between sentences in documents are often not directly obtainable.Existing approaches usually use syntactic knowledge and co-reference,adjacency,co-occurrence,etc.to construct documents as the document graph and capture the interactions between entities.However,the large number and types of graph nodes and graph edges greatly limit the inference ability of the model.A bi-graph model with a simple structure and better inference effect is proposed in this paper.Firstly,heuristic rules are used to extract mention interactions and evidence sentences,and based on this,the mention graph and entity graph based on evidence sentences are constructed.Secondly,the attention mechanism is utilized to capture the inference paths between entity nodes in the entity graph.Finally,according to the inference paths,a suitable scoring function is used to predict entity relationship facts.Experiments on DocRED show that the model proposed in this paper achieves good results.

关 键 词：文档级关系抽取 图神经网络 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]