Hard-Negative-Aware Contrastive Learning for Knowledge Inference on N-Ary Facts

N-ary relational Knowledge Bases (KBs) allow knowledge to be represented beyond binary relational facts. With a surge in interest in modeling n-ary relational facts, recent approaches have achieved considerable progress on the knowledge inference task. However, these models have generally ignored the hard negative entities. The hard negative entities share the same entity-relation query couple in the primary triple with the ground truth entity, but do not satisfy the n-ary relational query. The existence of hard negative entities is a very important and unique characteristic of n-ary relational KBs in contrast to binary relational KBs. We propose a hard-negative-aware contrastive learning based method for knowledge inference on n-ary relational facts. Our framework is simple and effective, and can be directly applied to any recent approaches on n-ary relational facts without further modification. The empirical findings across three real-world datasets demonstrate the superiority of this framework.