A scalable rule learning approach fusing multiple sources of information

Logical rules enable knowledge graph reasoning in an explicit and interpretable way. Hybrid learning approaches that combine logical rules with representation learning has shown strong capabilities in knowledge graph reasoning tasks by integrating both aspects’ strengths and utilizing various knowledge elements of knowledge graph to assist in rule mining. Previous studies primarily focused on the structural information in knowledge graphs, ignoring other multi-source information. This situation limits the learning of high-quality rules. To address this, we introduce SiMi, a scalable rule learning approach fusing multiple sources of information.SiMi enhances the embedding representation of entities by utilizing external information, which provides semantic constraints for the rule learning process. This method can easily fuse multiple types of information, such as text, and can use different KGE models to aid rule learning. Extensive experiments on different datasets demonstrate the scalability of SiMi as well as the effectiveness in link prediction tasks.