Knowledge Graph Completion Based on Logical Rules and Graph Neural Network

Abstract

Knowledge Graph Completion(KGC) aims to utilize the existing knowledge in a Knowledge Graph(KG) to derive new facts. This process is pivotal in several tasks and domains and has garnered increasing attention from researchers. However, most existing KGC methods focus on modeling fact triples in a KG without fully considering deep semantics or associations between entities and relationships in the KG. To address this issue, logical rules can be used to reflect the implicit associations between relationships in a KG. The semantic nature of the KG implies that higher-order neighborhoods around fact triples contain deep semantic information. Therefore, to fully explore the inherent semantics and correlations of entities and relationships in KGs, this study proposes a novel model for a KGC based on logical rules and Graph Neural Networks(GNNs). The framework first employs an automatic rule learning process based on the efficient Expectation-Maximization(EM) iterative optimization algorithm. It then performs joint embedding training on the obtained high-quality logical rules, entities, and relationships in the KG to model the complex relationship patterns in the KG and improve the generalization of the embedding representation. Subsequently, attention embedding propagation is performed by simultaneously considering the importance of logic rules and triples to aggregate higher-order neighborhood information, and entity and relation embedding representations incorporating deep semantics and associations are obtained for the KGC. In this study, extensive experiments are conducted on four public datasets for link prediction, and the results demonstrate the effectiveness of the proposed model.