A Few-Shot Knowledge Graph Completion Model With Neighbor Filter and Affine Attention

Abstract

In recent times, extensive scholarly focus has been directed towards the knowledge graph completion (KGC) due to the large number of triples that perform well in training tasks. However, the relations of realistic knowledge graphs (KGs) usually have long-tailed distributions, posing a great challenge in inferring new triples of task relationships from a limited number of triples. To tackle this challenge, methodologies for few-shot knowledge graph completion (FKGC) have been devised. These approaches employ a limited set of reference triples to forecast novel triples for various relations. However, existing FKGC approaches suffer from the drawbacks of not fully utilizing the structural information in KGs and ignoring the fine-grained information of interactions between entity pairs. In this paper, a FKGC model with neighbor filter and affine attention (NFAA) is proposed. The NFAA model filters 2-hop neighbors into a neighborhood scope for an entity aggregator and constructs a relation generator utilizing the affine attention mechanism to efficiently infer new triples for the few-shot relation task. Evaluations are performed using two publicly available benchmark datasets: NELL-one and Wiki-one. Experimental results validate the superiority of the NFAA model relative to several state-of-the-art approaches.