TP-RotatE: A knowledge graph representation learning method combining path information and rules to capture complex relational patterns

Abstract

Representation learning on a knowledge graph (KG) aims to map entities and relationships into a low-dimensional vector space. Traditional methods for representation learning have predominantly focused on the structural aspects of triples within the KG. While existing approaches have endeavored to integrate path information and rules to enhance the structural richness of KGs, these efforts have been constrained by the lack of consideration for complex relational representations and contextual information. In this study, we introduce TP-RotatE, an innovative method that leverages the semantic context of triples to effectively capture more intricate relational patterns. Specifically, our model harnesses contextual information surrounding the head entity and distills relevant rules. These rules are then integrated with path information to offer a more holistic perspective on the relationships embedded within complex vector spaces. Furthermore, the synergy between rules and paths empowers the knowledge-embedded model to handle the intricacies of complex relationships. Experimental results on a benchmark dataset confirm that TP-RotatE surpasses current baseline methods in KG inference tasks, achieving state-of-the-art performance.