Learning three-way affinity embeddings for knowledge base completion

Abstract

Knowledge bases are an extremely important database for knowledge management, which is very useful for question answering, query expansion and other related tasks. However, it often suffers from incompleteness. In this paper, we propose a Three-Way Affinity Embeddings model (TWAE) to map both the entity and relationship into two vectors and consider any two of them direct interaction, and then predict the possible truth of additional facts. The basic idea is that the confidence of the additional predicted fact is determined by three-way affinities in the triplet using the latent representation of each item. Experiments show that our model performs excellent.