计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (9): 183-189.DOI: 10.3778/j.issn.1002-8331.1904-0264

• 模式识别与人工智能 • 上一篇    下一篇

基于Att_GCN模型的知识图谱推理算法

王红,林海舟,卢林燕   

  1. 中国民航大学 计算机科学与技术学院,天津 300300
  • 出版日期:2020-05-01 发布日期:2020-04-29

Knowledge Graph Inference Algorithm Based on Att_GCN Model

WANG Hong, LIN Haizhou, LU Linyan   

  1. College of Computer Science and Technology, Civil Aviation University of China, Tianjin 300300, China
  • Online:2020-05-01 Published:2020-04-29

摘要:

针对目前知识图谱推理方法中,传统神经网络方法不能有效考虑实体之间的相关性问题,提出了一种将图卷积神经网络(GCN)与注意力机制(Attention)相结合的知识图谱推理方法。该方法利用注意力机制对知识图谱中的实体与其邻域实体进行相关性计算,得到实体特征向量。通过图卷积神经网络的参数共享技术学习实体的所有邻域实体特征。将实体特征和关系特征进行特征融合,得到每个实体的隐性特征向量。通过实体分类实验和链接预测实验进行分析,结果表明,该方法的准确率较传统神经网络方法有进一步提高,为搜索、问答、推荐以及数据集成等领域提供了方法支持。

关键词: 知识图谱推理, 实体分类, 链接预测, 图卷积神经网络(GCN), Attention机制

Abstract:

In view of the fact that the traditional neural network method can not effectively consider the correlation between entities in current Knowledge Graph Reasoning methods, a Knowledge Graph Reasoning method combining Graph Convolution Neural network(GCN) and Attention Mechanism(Attention) is proposed. This method uses Attention Mechanism to calculate the correlation between entities and their neighborhood entities in knowledge graph, and obtains the entity feature vector. All the neighborhood entity features of entities are learned by parameter sharing technology of graph convolution neural network. The entity features and relationship features are fused to get the implicit feature vectors of each entity. The results of entity classification experiment and link prediction experiment show that the accuracy of this method is further improved than that of traditional neural network method, which provides methodological support for search, question and answer, recommendation and data integration.

Key words: knowledge graph reasoning, entity classification, link prediction, Graph Convolution Neural(GCN)network, Attention mechanism