Abstract: Aiming at the method of knowledge map completion based on inductive relationship prediction, the existing knowledge graph completion methods based on inductive relationship prediction are limited to direct reasoning, and all entity sets must be known during training. This paper proposes a relationship prediction method based on graph neural network, extracts the local directed subgraphs of graph neural network for reasoning, introduces a node-edge bidirectional information transfer mechanism for inductive relationship reasoning, streng-thens the information communication between nodes and edges and effectively deals with the asymmetric relationship in the triplet. In view of the fact that different connection paths between entities reveal the nature of their relationship and help predictive reasoning, considering the relationship path between two entities, the relationship type suitable for inductive reasoning is used to represent the path, the edge embedding is defined attention formula, and the relationship between entities that have not been seen in the training set is predicted. Experimental results on common benchmark datasets suitable for inductive reasoning methods show that the method in this paper improves the prediction accuracy of triples compared with baseline models.

Key words: knowledge graph completion, graph neural network, inductive reasoning, link prediction

摘要： 针对基于归纳关系预测的知识图谱补全方法，现有的方法仅限于直推式推理，训练期间必须知道全部的实体集合。提出一种基于图神经网络的关系预测方法。首先提取图神经网络的局部有向子图进行推理，其次引入一个用于归纳关系推理的节点-边双向信息传递机制，以加强节点和边之间的信息交流并有效处理三元组中的非对称关系。鉴于实体之间不同的连接路径揭示了其关系的本质并有助于预测推理，因此考虑两个实体之间的关系路径，用适用于归纳式推理的关系类型表示其路径，定义了边嵌入的注意力公式，对在训练集中没有见过的实体进行关系预测。在适用于归纳推理方法的常用基准数据集上的实验结果表明，该方法相比基线模型提高了三元组的预测精度。

关键词: 知识图谱补全, 图神经网络, 归纳式推理, 链接预测