关键词: 知识图谱, 实体对齐, 图卷积网络, 邻域匹配, 实体名称

Abstract: To address the structural heterogeneity among cross-lingual knowledge graphs in entity alignment tasks, a fusing entity name information and neighborhood matching in entity alignment network is proposed. Firstly, a weighted adjacency degree matrix is constructed based on the specific adjacency degree information of entities, which is utilized in Highway-GCNs to obtain entity embeddings and compute distances between entity pairs. Secondly, the matching scores between the neighborhoods of entity pairs are calculated based on graph structural information, which are used to iteratively update the distances between entity pairs. Subsequently, the Jaccard distance between same-language entity pair names is computed to update the distances between entity pairs. Finally, the updated distances between entity pairs are used for alignment evaluation. Experiments conducted on the public dataset DBP15k, the proposed model achieves Hit@1 scores of 85.6%, 88.9% and 95.2% on three cross-lingual datasets respectively, surpassing all baseline models based on glove word vector models. This indicates that the proposed model effectively improves the accuracy of entity alignment results.

Key words: knowledge graph, entity alignment, graph convolutional network, neighborhood matching, entity name