Knowledge Graph Consolidation by Unifying Synonymous Relationships

Entity-centric information resources in the form of huge RDF knowledge graphs have become an important part of today's information systems. But while the integration of independent sources promises rich information, their inherent heterogeneity also poses threats to the overall usefulness. To some degree challenges of heterogeneity have been addressed by creating underlying ontological structures. Yet, our analysis shows that synonymous relationships are still prevalent in current knowledge graphs. In this paper we compare state-of-the-art relational learning techniques to analyze the semantics of relationships for unifying synonymous relationships. By embedding relationships into latent feature models, we are able to identify relationships showing the same semantics in a data-driven fashion. The resulting relationship synonyms can be used for knowledge graph consolidation. We evaluate our technique on Wikidata, Freebase and DBpedia: we identify hundreds of existing relationship duplicates with very high precision, outperforming the current state-of-the-art method.