融合层次类型信息的双向图注意力机制的 知识图谱嵌入模型.

Knowledge graph embedding designs to map the entities and relations into low-dimensional and dense vector spaces. The existing embedded models still have the following two defects: most of the existing models only focus on the semantic information of knowledge graph, but ignore plentiful hidden information of triples; the existing models only concern the unidirectional information of the entity, but ignore the bidirectional potential information. This paper proposed the knowledge graph embedding model Bi-HTGAT to solve these problems. The model designed a hierarchical type attention mechanism, considered the contribution of different entities of each type to the central entity under different relations. At the same time, this paper introduced the directional attention mechanism of the relations, fused the neighbor information in different directions to update the entities and relations embedding, and finally aggregated the two parts of information to obtain the final embedding of the entity. The experimental results show that Bi-HTGAT performs better than other baseline models in link prediction, which fully proves that Bi-HTGAT can further improve the accuracy of embedding results. [ABSTRACT FROM AUTHOR]