Knowledge Graph Embedding with Relation Rotation and Entity Adjustment by Quaternions

Wen Sun, Qingqiang Wu, Xiaoli Wang, Junfeng Yao, Zhifeng Bao

Abstract: Recent knowledge graph embedding models have shown promising results on link prediction, by employing operations on quaternion space to capture correlations between entities. However, they only used three quaternion embeddings for rotation calculation that fails to capture the interaction between entities and relations. The single relation quaternion to rotate the head entity also makes the connection between the head and tail entities weak. To address the problem, this paper proposes a novel knowledge graph embedding model denoted as QuatPE, to utilize paired relations to simultaneously rotate the head and tail entities in a triple. We employ the adjustment vectors to adjust the position of the same entity in a quaternion space when facing different triples. With paired relations, QuatPE can strengthen the connection between the head and tail entities which enhances the representation capabilities. By using the adjustment vectors, QuatPE also helps to better handle complex relation patterns, such as 1-to-N, N-to-1, and N-to-N. Experimental results show that QuatPE can achieve significant performance on well-known datasets for link prediction. Researchers can reproduce our results by following the source codes at https://github.com/galaxysunwen/QuatPE-master.