DSKRL: A dissimilarity-support-aware knowledge representation learning framework on noisy knowledge graph

Abstract

Knowledge graphs are widely used in a variety of knowledge-driven applications. Due to the inefficiency of manual construction of knowledge graphs, automatic construction mechanism has become the dominant method, which introduces noise. However, most knowledge representation learning approaches assume that there is no noise in the knowledge graph and hence ignore noise detection. In this paper, we propose a dissimilarity-support-aware knowledge representation learning framework, which accomplishes knowledge representation learning and noise detection simultaneously. Specifically, we introduce triple dissimilarity and triple support to construct the model energy function which is based on translation-based methods. The triple dissimilarity measures the matching extent of entities and relations in triples and the triple support measures the credibility of the matching extent. In order to make triple dissimilarity and triple support estimation effective and comprehensive, we synthesize structural information and auxiliary information (entity hierarchical type and relation path information) in triple dissimilarity and triple support. We conduct experiments on three datasets for the knowledge graph noise detection task and the knowledge graph completion task. The experimental results show that our model achieves significant and consistent improvements compared to all baseline methods.