Learning Short-Term Differences and Long-Term Dependencies for Entity Alignment

Abstract

We study the problem of structure-based entity alignment between knowledge graphs (KGs). The recent mainstream solutions for it apply KG embedding techniques to map entities into a vector space, where the similarity between entities could be measured accordingly. However, these methods which are mostly based on TransE and its variants treat relation triples in KGs independently. As a result, they fail to capture some advanced interactions between entities that are implicit in the surrounding and multi-hop entities: One is the differences between the one-hop and two-hop neighborhood of an entity, which we call as short-term differences, while the other is the dependencies between entities that are far apart, which we call as long-term dependencies. Based on the above observations, this paper proposes a novel approach learning to capture both the short-term differences and the long-term dependencies in KGs for entity alignment using graph neural networks and self-attention mechanisms respectively. Our empirical study conducted on four couples of real-world datasets shows the superiority of our model, compared with the state-of-the-art methods.