An Entity Alignment Method Based on Graph Attention Network with Pre-classification

Abstract

Entity alignment is the process of identifying entities that point to the same object in different knowledge graphs. Entity alignment is a key step in building knowledge graphs, and the result of entity alignment directly affects the quality of the knowledge graphs. Most of the current entity alignment methods learn the feature vectors of entities or entity attributes based on representation learning. While the study of the interaction learning between entity attributes and entity relationship features is not sufficient, and the data division is not specific enough to confuse the feature information of different categories of entities. That reduces the quality of the finally learned entity vectors. To solve that problem, we propose an entity alignment method that uses graph attention mechanism after entity classification. Firstly, we classify the entities according to their semantics of the source data, and then classify the entities with different entity attributes to complete the dual entity classification. Based on the dual entity classification, the graph attention mechanism is used to complete the aggregation of the feature information of different categories of entities, which is used to learn the final entity vector representation for the calculation of similarity in the entity alignment task. On the general dataset for entity alignment, DBP15K, our model achieves hits@1 scores 80.82, 80.17 and 93.13 on its three subdatasets. The results show our method are better than the compared methods.