基于BERT和路径对比学习的归纳关系预测

在以往的知识图谱关系预测任务中，主要方法仅限于直推式推理，它们在新出现实体和关系情况下不能利用先验知识去处理归纳学习的问题。提出了基于BERT与路径对比学习的关系预测方法（BERT-based and path comparison learning，BPCL）。首先，利用卷积神经网络捕获子图目标三元组的上下文邻域信息，并将子图线性化为关系路径，利用BERT初始化边特征；其次，引入正、负关系路径；最后，联合对比学习和自监督学习训练对新出现实体之间的关系预测。在适用于归纳推理方法的常用基准数据集上，验证了该模型的预测精度有所提高。

Prediction of inductive relationship based on BERT and path contrast learning

In previous knowledge graph relationship prediction tasks, the main approach was straightforward reasoning, they couldn''t use a priori knowledge to deal with inductive learning in the case of unseen entities and relations. This paper proposed a relationship prediction method based on BERT and path contrast learning, called BPCL. Firstly, it used the convolutional neural network to capture the contextual neighborhood information of the target triplet of the subgraph, and linearized the subgraph into a relational path. And it used BERT to initialize edge features. Secondly, it introduced the comparative learning of positive and negative relational paths. Finally, it carried out the relationship prediction by combining contrast learning and supervised training. This paper verifies the improved prediction accuracy of the model on a common benchmark dataset applicable to inductive inference methods.