POSE: A Positional Embedding Model for Knowledge Hypergraph Link Prediction

Abstract

AbstractLink prediction in knowledge hypergraphs has been widely recognized as crucial for various downstream tasks of knowledge-enabled applications, from question answering to recommender systems. However, most current approaches are directly extended from binary relation of the knowledge graph to n-ary relation, thus cannot capture entities’ role and positional information in each n-ary tuple. To accommodate the transformation of relations from binary to n-ary in the knowledge hypergraph, in this work, we propose POSE, which exploits the semantic properties of tuples at both role and position levels. POSE explores an embedding space with basis vectors and represents the role and positional information of entities through a linear combination, which promotes similar representations for entities with related roles and the same positions. Then, a relation matrix is further employed to capture the compatibility of both information with all associated entities, and a scoring function is used to measure the plausibility of tuples composed of entities with specific roles and positions. Meanwhile, POSE achieves full theoretical expressiveness and predictive efficiency. Experimental results show that POSE achieves an average improvement of 4.1% on MRR compared to state-of-the-art knowledge hypergraph embedding methods. Our code is available at https://github.com/zirui-chen/POSE.KeywordsLink predictionKnowledge hypergraphPostional embedding