FuAlign: Cross-lingual entity alignment via multi-view representation learning of fused knowledge graphs

Abstract

Listen

Translate article

Cross-lingual entity alignment is essential in knowledge graph completion. Recently, knowledge representation learning (KRL) has gained remarkable achievements in cross-lingual entity alignment. Most existing methods first learn entity representations of different knowledge graphs and transform them into a unified space based on a seed set of pre-aligned entity pairs. However, the transformation is error-prone due to the distribution differences between the seed set and the whole entity set. This paper presents FuAlign, a novel cross-lingual entity alignment framework based on multi-view KRL of a pre-fused knowledge graph. FuAlign first fuses two matching knowledge graphs based on the given seed set. Then, it exploits multi-view representation learning to map the fused knowledge graph into a unified space. FuAlign improves the performance from two aspects. First, it represents entities in a unified embedding space, thus avoiding the error-prone transformation between different embedding spaces. Second, the proposed multi-view representation learning model captures different kinds of information such as semantics, entity context, and long-term entity dependency in knowledge graphs. We conduct extensive experiments to evaluate the performance of the proposed method. Experimental results show that FuAlign outperforms most baseline methods and achieves comparable performance with the Bert-based model.