Fuzzy entity alignment via knowledge embedding with awareness of uncertainty measure

Abstract Entity alignment refers to associate entities in different knowledge graphs if they are semantically identical. Embedding-based entity alignment approaches encode entities in a continuous embedding space where entities are aligned based on the similarity of learned embeddings. However, there exists ambiguity and uncertainty in entity alignment caused by single alignment metric. In this paper, a fuzzy entity alignment method FuzzyEA is proposed to model the uncertainty in alignment process based on intuitionistic fuzzy set (IFS). Iterative TransE model is designed to learn relational structure of knowledge graphs, where mutual selection and error correction mechanism is proposed to enhance the effect of iteration. The alignment results obtained by name/description embedding and structure embedding are fused based on Dempster’s combination rule. Experiments on three benchmark datasets demonstrate that the proposed FuzzyEA consistently outperforms other entity alignment methods and contributes to promising improvement in alignment accuracy and discrimination ability.