Differentiable Learning of Rules with Constants in Knowledge Graph

Anonymous

Abstract: Knowledge reasoning, helping overcome the incompleteness issue of knowledge graph(KG), significantly contributes to the development of large KG, which consists of relations and constants. Rule mining studies the problem of capturing interpretable patterns over KG, which is one of the key tasks of knowledge reasoning. However, previous works mainly focus on the combination of different relations, and are limited for ignoring the importance of constants. In this paper, we propose that constants should be considered in rule mining process, and introduce an Elegant Differentiable rUle learning with Constant mEthod (EduCe). Based on soft constant operator and dynamic weight, the model we proposed can mine more diverse and accurate logical rules while controlling the number of parameters, which is also a great challenge to this problem. Experiment results on several benchmark datasets demonstrate the effectiveness and accuracy of rule with constants.