Enhancing Knowledge Graph Attention by Temporal Modeling for Entity Alignment with Sparse Seeds

Abstract

As a fundamental task of knowledge graph integration, entity alignment (EA) matches equivalent entities across knowledge graphs (KGs). Temporal knowledge graphs (TKGs) enhance static KGs with temporal information. Traditional EA approaches tackle alignments of static KGs, but cannot effectively deal with TKGs. Therefore, temporal EA solutions are called for. To this end, we propose a time-aware graph attention network for entity alignment (TGA-EA). Generally, we learn high-quality temporal-relational entity embeddings for alignments by systematically integrating temporal information into KG embeddings. We propose three temporal modeling methods to effectively represent and integrate temporal information into both entities and relations. Then we construct temporal enhanced graph attention to produce target temporal-relational entity embeddings with temporal entities and temporal relations. Thanks to the powerful design, TGA-EA achieves promising performances with sparse alignment seeds. Extensive experiments on five datasets demonstrate our approach’s obvious advantages over previous works.