HyperspherE: An Embedding Method for Knowledge Graph Completion Based on Hypersphere

Knowledge graph completion (KGC) aims to predict missing facts by mining information already present in a knowledge graph (KG). A general solution for KGC task is embedding facts in KG into a low-dimensional vector space. Recently, several embedding models focus on modeling isA relations (i.e., instanceOf and subclassOf), and produce some state-of-the-art performance. However, most of them encode instances as vectors for simplification, which neglects the uncertainty of instances. In this paper, we present a new knowledge graph completion model called HyperspherE to alleviate this problem. Specifically, HyperspherE encodes both instances and concepts as hyperspheres. Relations between instances are encoded as vectors in the same vector space. Afterwards, HyperspherE formulates isA relations by the relative positions between hyperspheres. Experimental results on dataset YAGO39K empirically show that HyperspherE outperforms some existing state-of-the-art baselines, and demonstrate the effectiveness of the penalty term in score function.