结合谓词感知与图注意力机制的链接预测方法

知识图谱补全旨在预测三元组中缺失的部分使知识图谱趋于完整.针对基于神经网络等模型的链接预测方法忽略了实体间的关联信息,导致模型不能覆盖三元组周围局部邻域中固有的隐藏信息,提出图注意力机制与谓词感知结合的方法.首先,利用图注意力机制定义了一个关系嵌入矩阵,描述任意给定实体邻域内实体间的关系;其次,引入谓词增强实体间语义理解程度,构造了基于谓词嵌入向量的注意力值计算公式,以便有效地度量实体间语义联系的强度;此外,利用实体邻居间的边关系预测多跳实体间的直接关系以补全知识图谱.在数据集WN18RR、Kinship、FB15K的实验结果表明了该方法能有效提高三元组的预测精度.

Link prediction based on predicate awareness and graph attention mechanism

The purpose of knowledge graph completion is to predict the missing parts in the triplet and make the knowledge graph complete. In view of the fact that the link prediction methods based on neural network and other models ignore the association information between entities, resulting in the model can''t cover the inherent hidden information in the local neighborhood around the triple. Aiming at this problem, this paper proposed a method combining graph attention mechanism with predicate perception. Firstly, it defined a relational embedding matrix to describe the relationship between entities in the neighborhood of any given entity by using graph attention mechanism. Secondly, it introduced predicate to enhance the semantic understanding between entities, and constructed the attention value calculation formula based on predicate embedding vector to effectively measure the strength of semantic connection between entities. In addition, it used the edge relationship between entity neighbors to predict the direct relationship between multi-hop entities to complete the knowledge graph. Experimental results on WN18RR, Kinship and FB15K datasets show that the proposed method can effectively improve the prediction accuracy of triplets.