A Contextualized Entity Representation for Knowledge Graph Completion

Abstract

Knowledge graphs (KGs) have achieved great success in many AI-related applications in the past decade. Although KGs contain billions of real facts, they are usually not complete. This problem arises to the task of missing link prediction whose purpose is to perform link prediction between entities. Knowledge graph embedding has proved to be a highly effective technology in many tasks such as knowledge reasoning, filling in the missing links, and semantic search. However, many existing embedding models focus on learning static embeddings of entities which pose several problems, most notably that all senses of a polysemous entity have to share the same representation. We, in this paper, propose a novel embedding method, which is named KG embedding with a contextualized entity representation (KGCR for short), to learn the contextual representations of entities for link prediction. KGCR encodes the contextual representations of an entity by considering the forward and backward contexts of relations which helps to capture the different senses of an entity when appearing at different positions of a relation or in different relations. Our approach is capable to model three major relational patterns, i.e., symmetry, antisymmetry, and inversion. Experimental results demonstrate that KGCR can capture the contextual semantics of entities in knowledge graphs and outperforms existing state-of-the-art (SOTA) baselines on benchmark datasets for filling in the missing link task.

Supported by the Key R&D Program Project of Zhejiang Province under Grant no. 2019C01004 and Zhejiang Education Department Project under Grant no. Y201839942.