基于双向对齐与属性信息的跨语言实体对齐  

Cross-language Entity Alignment Based on Bidirectional Alignment and Attribute Information

作　　者：车超[1] 刘迪 CHE Chao;LIU Di(Key Laboratory of Advanced Design and Intelligent Computing,Ministry of Education,Dalian University,Dalian,Liaoning 116622,China)

机构地区：[1]大连大学先进设计与智能计算省部共建教育部重点实验室,辽宁大连116622

出　　处：《计算机工程》2022年第3期74-80,共7页Computer Engineering

基　　金：国家自然科学基金面上项目(62076045);辽宁省自然科学基金(2019-ZD-0569)。

摘　　要：实体对齐表示在不同的知识图谱中查找引用相同现实身份的实体。目前主流的基于图嵌入的实体对齐方法中的对齐实体通常具有相似的属性,有效利用属性信息可提升实体对齐效果,同时由于不同知识图谱之间的知识分布差异,仅考虑单个方向的对齐预测会导致预测结果出现偏差。针对上述问题,提出一种改进的跨语言实体对齐方法。利用融合属性信息的双向对齐图卷积网络模型,将前馈神经网络编码实体对应的属性信息与初始的实体嵌入相结合,得到联合属性信息的实体表示,并使用双向对齐机制实现跨语言的实体对齐预测。在3个跨语言数据集上的实验结果表明,该方法通过融合更多的知识图谱信息增强了实体表示能力,并且利用双向对齐机制缓解了数据分布差异问题,相比基于图嵌入的实体对齐方法整体性能更优。Entity alignment is to find the entities that refer to the same real identity in different knowledge graphs.The alignment entities in most of the existing graph embedding-based entity alignment methods share similar attributes,which means utilizing attribute information can improve entity alignment performance. At the same time,due to the knowledge distribution differences between different knowledge graphs,alignment prediction that considers only a single direction will lead to deviation in alignment results. In response to the above problems,this paper proposes an improved cross-language entity alignment method. The method uses a Bidirectional alignment Graph Convolutional Network model with Attribution information(BiGCN-A),and combines the attribute information corresponding to the coded entity with the initial entity embedding through a feed-forward neural network to obtain an entity representation of joint attribute information. A bidirectional alignment mechanism is also used to realize cross-language entity alignment prediction. Experimental results on three cross-language datasets show that the proposed method enhances the entity representation ability by fusing more knowledge graph information.It also uses a bidirectional alignment mechanism to alleviate the problem of data distribution differences. Compared with the entity alignment method based on graph embedding,the proposed method displays better overall performance.

关 键 词：实体对齐 知识图谱 属性信息 双向对齐 图卷积网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]