A semi-supervised neighborhood matching model for global entity alignment

As an important way to integrate knowledge graphs, entity alignment is widely used in the field of natural language processing. Entity alignment is to find entities that exist in different knowledge graphs but have the same real-world meaning. Recently, most entity alignment models consider only the one-hop neighborhood node information of candidate entity alignment pairs, without considering the relations connected to neighboring nodes. Relations are critical to determining whether two entities can be aligned when they have the same neighborhood structure. Therefore, besides using the structural information of entities, we also use relation semantics to enhance entity alignment. The larger the number of training seed set, the better the model performance. Based on the premise, we use the semi-supervised bidirectional nearest neighbor iteration strategy to expand the size of the training seed set without manual labeling. Furthermore, to ensure the stability of the entity alignment results, we consider the dependencies between alignment decisions and perform global entity alignment from a comprehensive perspective. We evaluate the performance of our model on three publicly available cross-lingual datasets, and the experimental results demonstrate the effectiveness of our model.