基于图神经网络文档相似度的实体与关系层次匹配方法  

A Hierarchical Matching Method of Entity and Relation Based on Graph Neural Network for Document Similarity

作　　者：赵文彬 王佳琦 吴峰[3] 任雁[3] 安寅生 ZHAO Wenbin;WANG Jiaqi;WU Feng;REN Yan;AN Yinsheng(School of Information Science and Technology,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;Daqin Railway Co.,Ltd,Taiyuan 030024,China;Hebei Science and Technology Information Processing Laboratory,Hebei Institute of Science and Technology Information,Shijiazhuang 050021,China)

机构地区：[1]石家庄铁道大学信息科学与技术学院,河北石家庄050043 [2]大秦铁路股份有限公司,山西太原030024 [3]河北省科学技术情报研究院、河北省科技信息处理实验室,河北石家庄050021

出　　处：《郑州大学学报（理学版）》2023年第6期8-14,共7页Journal of Zhengzhou University:Natural Science Edition

基　　金：国家自然科学基金项目(61373160);河北省自然科学基金项目(F2021210003);河北省教育厅青年基金项目(QN2020197)。

摘　　要：针对文本中深层语义难以计算的问题,提出了基于句法依存关系的多头图注意力实体关系联合抽取模型和融合层次类型的文档相似性匹配。首先通过多头图注意力网络对文本进行实体关系抽取,然后设计融合层次类型的词移距离相似性计算方法以及基于图相似的文档相似性计算模型,利用文档中的实体和关系构建图结构,根据图级特征进行相似性计算。最后,通过对比实验验证了所提方法在文档相似性计算、图相似度计算和图分类任务中的有效性。Aiming at the difficulty to mine deep semantics in text,a multi-head graph attention entityrelation joint extraction model based on syntactic dependencies and a fusion hierarchical type of document similarity matching were proposed.Firstly,the entity relation extraction was carried out on the text through the multi-head graph attention network.Then,the word shift distance similarity calculation method of fusion hierarchical type and the document similarity calculation model based on graph similarity were designed,and the graph structure was constructed by using the entities and relations in the document.Thus,the features representing the graph level were obtained for similarity calculation.Finally,the effectiveness of the proposed method in document similarity calculation,graph similarity calculation and graph classification tasks was verified by comparative experiments.

关 键 词：实体关系抽取 相似性 层次类型 图神经网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]