基于邻域聚合的实体对齐方法

Entity Alignment Method Based on Neighborhood Aggregation

实体对齐旨在判断来自不同知识图谱的实体是否为指向真实世界的同一个对象。然而,知识图谱间的结构异质性往往会影响实体对齐的准确性。提出一种基于邻域聚合匹配网络(NAMN)模型的实体对齐方法。根据每跳邻居对中心实体重要性不同的特点,采用分层的思想区别处理每跳邻域信息,通过门控机制进行聚合以学习图结构的表征。在此基础上,为每个实体构建邻域局部子图进行跨图邻域匹配,并将匹配阶段的输出与通过门控机制所学习到的图结构表征进行联合编码,生成最终面向匹配的表征。采用DBP15K数据集进行实验,结果显示,Hits@1的所有值均在75%以上,Hits@10的所有值均在85%以上,最高可达到97%,平均倒数排名均高于80%,表明NAMN模型能够有效提高实体的匹配准确度。

Entity Alignment(EA)aims to judge whether entities from different Knowledge Graph(KG)are the same object pointing to the real world.However,the structural heterogeneity between KG often affects the accuracy of EA.Hence,an EA method based on a Neighborhood Aggregation Matching Network(NAMN)model is proposed.Based on the different importance of each hop neighbor to the central entity,a hierarchical idea is applied to process the neighborhood information of each hop differently,and the gating mechanism is used to perform aggregation to learn the representation of a graph structure.Subsequently,a neighborhood local subgraph is constructed for each entity for cross graph neighborhood matching,and the output of the matching stage is jointly encoded with the graph structure representation learned through the gating mechanism to generate the final matching oriented representation.The experiment is performed using the DBP15K dataset.The experimental results show that all values of Hits@1 exceed75%,all values of Hits@10 are between 85% and 97%,and the Mean Reciprocal Rank(MMR)exceeds 80%,indicating that the NAMN model can effectively improve the matching accuracy of entities.