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 model for bidirectional graph attentionmechanism with hierarchical type information

知识图谱嵌入旨在将实体与关系映射到低维且稠密的向量空间中。目前已有的嵌入模型仍存在以下两个方面的缺陷:现有的模型大多只关注知识图谱的语义信息,而忽略了大量三元组的隐藏信息;现有的模型仅关注了实体的单向信息,而忽略了双向的潜在信息。针对以上问题,提出了一种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Bi-HTGAT,该模型设计了层次类型注意力机制,考虑不同关系下每种类型的不同实体对中心实体的贡献。同时引入了关系的方向注意力机制,通过融合不同方向的邻居信息来更新实体和关系嵌入,最终聚合两部分信息以得到实体的最终嵌入。在基准数据集上的实验证明,Bi-HTGAT在链接预测任务上性能明显优于其他基线模型,充分证明了Bi-HTGAT能够进一步提高嵌入结果的精准度。

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.