一种融合实体图上下文的三维旋转知识图谱表示学习

Three-dimensional Rotation Knowledge Graph Representation Learning with Fusing Entity Graph Context

知识图谱表示学习旨在将实体和关系投影到低维、连续的向量空间中,学习实体和关系语义信息的向量表示.然而,大多数现有的模型难以有效建模知识图谱的基本特征,即对称/反对称、逆、组合关系模式.此外,许多模型将知识图谱简化成不相关三元组构成的集合,忽略三元组中的实体在图中的邻域信息.针对以上问题,本文提出一种融合实体图上下文的三维旋转知识图谱表示学习模型.该模型首先在四元数数学框架的基础上,将实体表示为三维空间中的一组向量,将关系解释为实体间的三维旋转变换,以更好建模各种关系模式.然后,利用注意力机制从相邻节点和边中学习实体图上下文表示,并将其引入到三元组打分函数,以将实体的图上下文信息融合到表示学习模型中.在两个公开数据FB15k-237与WN18RR上的实验结果表明了本文所提模型的有效性.

Knowledge graph representation learning aims to project entities and relationships into a low-dimensional, continuous vector space, and learns vector representations of the semantic information of entities and relationships.However, most of the existing models are difficult to effectively model the basic characteristics of the knowledge graph, that is, the symmetric/antisymmetric, inverse, and combination relationship modes.In addition, many models simplify the knowledge graph to a collection of unrelated triples, ignoring the neighborhood information of entities in the triples in the graph.To address the above issues, a three-dimensional rotation knowledge graph representation learning model with fusing entity graph context is proposed.Based on the quaternion mathematical framework, the model first represents entities as a set of vectors in three-dimensional space, and interprets relations as three-dimensional rotation transformations between entities, which to better model various relation patterns.Then, the attention mechanism is used to learn the entity graph context representations from the neighboring nodes and edges, so that the representations can be introduced into the triple scoring function to integrate the context information into the model.The experimental results on two dataset FB15 k-237 and WN18 RR show the effectiveness of the proposed model.