DSEA: Dynamic Semantic Entity Alignment in Knowledge Graphs

Entity alignment aims to accurately identify corresponding entities in different knowledge graphs, and is the cornerstone of building comprehensive network knowledge graphs. Traditional methods rely on labeling information to achieve accurate matching, but self-supervised learning in recent years has inspired us to find that simultaneously pulling positive entity pairs and pushing away unlabeled negative entity pairs in entity alignment learning is more effective than pulling positive entity pairs only. This suggests that unlabeled entities also contain latent information. Based on this finding, the DSEA strategy is proposed, a dynamic multi-attention mechanism is designed to deeply mine the rich information and semantic features in the text, and NTXentLoss is proposed to make the DSEA strategy more precise, so that it can optimize the entity alignment objective without label supervision, which improves the accuracy and efficiency. To verify the accuracy of the model, we conduct experiments on the public datasets DWY100K. The experimental results show that our method outperforms existing methods and is able to effectively achieve entity alignment through self-supervised learning.