Cross-Lingual Entity Alignment Model Based on Multi-entity Enhancement and Semantic Information

Changhe Li, Baohua Zhang, Chao Wang

Abstract: Traditional cross-lingual entity alignment models are primarily focus on developing effective methods for representing the structural information of knowledge graphs. They only consider the relationship of current entity and its neighbouring entities, overlooking the fact that multiple relations may exist between two given entities, which will decrease the accuracy. To solve this, we propose G-MESI (Global algorithm based on Multi-entity Enhancement and Semantic Information), which incorporates the representations of multiple entities and their corresponding relations to enhance the embedding of the current entity. In order to represent all relationships between entities, we use self-attention in this model, which calculates attention matrices for entity-to-relation and relation-to-entity, ultimately leading to an enhanced entity-to-entity embedding. Furthermore, we employ cross-lingual word embeddings to measure the similarity, and incorporate a global alignment algorithm for improving performance. We evaluate our model on the widely used DBP15K dataset and achieve state-of-the-art results, with a 4.1% increase in average Hit@1 accuracy and a 7.6% improvement in Hit@1 for the Chinese-English bilingual dataset using the local algorithm.