基于交互式连接图注意力网络的知识图谱补全方法

Knowledge Graph Completion Method Based on Interactively Connected Graph Attention Network

知识图谱为许多智能信息服务应用提供了底层支持,包括智能搜索、公共安全、金融、医疗等领域,但现有的知识图谱通常是不完整的,知识图谱补全已经成为亟需解决的问题。现有的知识图谱补全方法忽略了邻居节点以及关系所富含的重要信息,往往只是简单地将邻居节点和关系拼接起来,忽略了不同关系和邻居节点对于节点有着不同的重要性。为此,文中提出了一种基于交互式连接图注意力网络的知识图谱补全方法(ICGAT)。该方法首先通过寻找两跳邻居节点,挖掘出潜在可能的关系,扩充每个节点的三元组;然后将每个三元组中的关系与节点的特征融合,并且采用节点与邻居节点交互式连接的方法,用4个空间向量来表示交互式连接的关系;最后将交互式连接的向量输入图注意力网络,得到关系和邻居节点对该节点的权重,以此说明其重要性。为了有效地表示一对多、多对多等复杂关系的三元组,该方法使用RotatE模型作为预训练模型。在链接预测任务中的实验结果表明,ICGAT方法在WN18RR和FB15k-237数据集中的平均排名(MR)和排名前10命中率(HR@10)均有一定的提升,说明ICGAT能够提高链接预测任务的准确性。

Knowledge graph provides underlying support for many intelligent information service applications,including intelligent search,public safety,finance,medical care and other fields.However,the existing knowledge graph is usually incomplete,and knowledge graph completion has become an urgent problem to be solved.The existing knowledge graph completion method models ignore the important information rich in neighbor nodes and relationships,and often simply splice neighbor nodes and relationships together,ignoring the different importance of different relationships and neighbor nodes to nodes.To solve this problem,this paper proposed a knowledge graph completion method(ICGAT)based on the interactive connection graph attention network.The method firstly finds out the potential relationship by finding two-hop neighbor nodes,and expands the triples of each node.Then it fuses the relationship in each triplet with the features of the node,and adopts the method of interactive connection between nodes and neighbor nodes,using 4 space vectors to represent the interactively connected relationship.Finally,the vector of interactive connection was input into the graph attention network to obtain the weight of relations and neighbor nodes to the node,so as to illustrate its importance.In order to effectively represent triples of complex relationships such as one-to-many,many-to-many,etc.,this method used the RotatE model as a pre-training model.The experimental results in the link prediction task show that the performance of the mean rank and HR@10indicators of the ICGAT method in the WN18RR and FB15k-237 datasets have been improved to a certain extent,indicating that ICGAT can improve the accuracy of the link prediction task.