Context-Aware Quaternion Embedding for Knowledge Graph Completion

In this paper, we study the learning representation of entities and relationships in the link prediction task of knowledge graph. The knowledge graph is a collection of factual triples, but most of them are incomplete. At present, some models use complex rotation to model triples, and obtain more effective results. However, such models generally use specific structures to learn the representation of entities or relationships, and do not make full use of the context information of entities and relations. In addition, in order to achieve high performance, models often need larger embedding dimensions and more epoches, which will cause large time and space cost. To systematically tackle these problems, we develop a novel knowledge graph embedding method, named CAQuatE. We propose two concepts to select valuable context information, then design a context information encoder to enhance the original embedding, and finally use quaternion multiplication to model triples. The experiment and results on two common benchmark datasets show that CAQuatE can significantly outperform the existing state-of-the-art model in the knowledge graph completion task by obtaining lower dimensional representation vectors with fewer epoches and no additional parameters.