SMGNN: an entity alignment method based on subgraph matching and graph neural network

Abstract

With the wide application of knowledge graph such as recommendation system and text analysis, it is particularly important to create high-quality knowledge graph, it requires precise knowledge graph fusion. As a key part of knowledge graph fusion, entity alignment can provide more prior knowledge for knowledge graph and improve its usability. In order to obtain a more global graph structure feature, this paper designs a subgraph matching based method for entity alignment, named SMGNN. It based on the two features of map structure information and local relation semantics, and captures relationships between entities through GNN capture. Firstly, the entity is encoded by the subgraph information of the target node through the GNN based on two entity-aligned knowledge graphs. Secondly, the subgraph is the graph composed of the target node and all neighboring nodes connected to the node. Then, the alignment between the two graphs is regarded as a mapping on the hyperplane, and TransH model is used for alignment. Finally, we do experiments on DBP15K, a crosslanguage entity alignment dataset, the results show that SMGNN can effectively improve the alignment accuracy of knowledge graphs.