Leveraging multimodal features for knowledge graph entity alignment based on dynamic self-attention networks

Abstract

Entity alignment establishes correspondence between entities across different knowledge graphs, which plays an important role in knowledge fusion. The key of entity alignment is to represent and fuse the features of entities across graphs. To improve efficiency, recent methods introduce graph embedding techniques that map different knowledge graphs into common latent spaces to encode structure and attribute features of entities. However, the structure embedding is generated by aggregating coarse connection information between entities without capturing semantics of the relations, which is incapable of dealing with entities of similar structures. Besides, multiple features are integrated by static arithmetic weighting method, which has difficulty in balancing the relative importance of structure and attribute features. To address these problems, we propose a dynamic self-attention based entity alignment model (DSEA), which leverages multimodal features for entity alignment through a dynamic self-attention network. To learn informative representations, we build a multimodal feature embedding component (MFE) to leverage structure, name, and position importance of entities. In MFE, semantics of adjacent entities and relations are exploited to enhance the structure embedding through a fine-granularity semantic augmentation network. Then a dynamic self-attention network is designed to adaptively evaluate the relative importance of structure and attribute features by learning the weights according to their contribution in entity alignment. Based on the feature weights, a similarity function of entities is formulated and the entity correspondence is established according to the similarities of entities. Extensive experiments conducted on five real-world datasets show that DSEA significantly outperforms state-of-the-art methods.