Abstract:All along, KG completion relied on link prediction has always been the focus of researchers. However, overwhelming majority of them can only serve 2-ary KGs. While in practice, knowledge hypergraphs (KH) covering facts beyond binary relations are far more ubiquitous. When confronted with them, massive studies for KGs show inadaptability. The several work towards N-ary KHs generally simply extend KG methods. And they usually transform N-ary knowledge into role-value pairs or triples, largely simplifying inherent association within each piece of knowledge. Furthermore, previous models study each N-ary knowledge independently, resulting in structural correlations among them being completely neglected. Motivated by these, avoiding breaking knowledge structure in KHs like previous studies do, we propose the first KH reasoning model based on original knowledge formats, RHKH. Challenged by complicated compositions indicated by the original format of N-ary tuples, association within and among each tuple is discovered through an innovative relational hypergraph neural network, RHNN. It considers complex interactions between relation and entities involved in the same knowledge as well. To refine such interactions, semantic components at each arity-position of relations are distinguished, along with introducing position-specific shift. Extensive experiments demonstrate the effectiveness of RHKH.

RHKH: Relational Hypergraph Neural Network for Link Prediction on N-ary Knowledge Hypergraph