Representation-Based Completion of Knowledge Graph with Open-World Data

Abstract

The method of knowledge graph completion (KGC) by adding external knowledge with new entities was discussed in this paper. Adopting the TransE-based representation of relations and triples in Knowledge Graph, we extract triples from open-world data and evaluate their correctness to fulfill KGC, where vectors are used for similarity evaluation. From the “structural” point of view, triples were first built from open-world data according to the similarity between TransE-based representation of pairs of entities and that of relations in KG. From the “semantic” point of view, the correctness of each external triple was evaluated by measuring the distance in the triple locally and ranking in the entire KG globally. ON the FreeBase and DBPedia KGs by different KG representation models and KGC methods, experimental results show that our proposal outperforms some state-of-the-art methods.