关键词: 实体对齐, 无监督, 多模态, 渐进式策略, 分配问题

Abstract: Although the current entity alignment method utilizes the structural information between entities in the knowledge graph and achieves a good alignment effect, a large amount of side information contained between entities is ignored. This information has unique characteristics and can significantly enhance the alignment effect. The availability of entity profile information in entity alignment is analyzed, and an unsupervised entity alignment method fusing graphic and text information is proposed. This method enhances the feature representation of the entity by fusing the literal information and visual information of the entity; uses the two-way threshold nearest neighbor algorithm to filter out the entity pairs whose distance measurement is too high; uses a progressive strategy to dynamically increase the similarity threshold for Control the quality of the generation of aligned entity pairs and control their generation speed; define the results obtained by the allocation algorithm to optimize the progressive strategy. In order to verify the method proposed in this paper, experiments were carried out on ZH_EN, JA_EN, and FR_EN of the DBP15K data set, and the results of 10 methods such as PSR, EVA, and DATTI were compared and analyzed. Hits@1 on FR_EN is 95.7% and 97.4%, and Hits@10 on FR_EN is 99.0%, which shows better performance.

Key words: Entity alignment, Unsupervised, Multimodal, Progressive Strategy, Assignment problem