Learning and Deduction of Rules for Knowledge Graph Completion

Abstract

The amount of training data or knowledge determines the precision of Knowledge Graph Completion (KGC), since traditional methods require as much training data as possible for model training or embedding to ensure the accurate expression of the knowledge contained in KG. It is necessary to extend the rules into KGC to improve the precision of KGC with respect to the small amount of training data. In this paper, we propose a framework RuleB for KGC including rule learning and rule deduction. First, we adopt six types of basic rules in the framework for searching the triples in KG. Then, we propose an optimization algorithm RWK based on the random walk strategy and K-sized traverse to reduce the execution time of triple search. Second, we give the corresponding deduction strategy for the different basic rules to obtain the new rule knowledge (NRK). Further, we give the corresponding strategy for combining the NRK and KG to obtain new entity knowledge (NEK), as well as that for NEK-based KGC, where NEK contains new triples that have not been presented in the KG previously. Experimental results of rule learning show the efficiency and accuracy of our methods.