Simple and effective meta relational learning for few-shot knowledge graph completion

Abstract

Conventional knowledge graph completion methods are effective for completing knowledge graphs (KGs), but they face significant challenges when dealing with relations with only a limited number of associative triples. To address the issue of incompleteness and long-tail distribution of relations in KGs, few-shot knowledge graph completion emerges as a promising solution. This approach predicts new triplets about a relation by leveraging only a handful of associated triples. Previous methods have focused on aggregating neighbor information and imposing sequential dependency assumptions. However, these methods can be counterproductive when they involve unrelated neighbors and rely on unrealistic assumptions, which hinders the learning of meta-representations. This paper proposes a simple and effective meta relational learning model (SMetaR) for few-shot knowledge graph completion that maintains the complete feature information of few-shot relations through a linear model. This approach effectively learns the meta-representation of few-shot relations and enhances meta-relational learning capabilities. Extensive experiments on two public datasets reveal that the model outperforms existing few-shot knowledge graph completion methods, demonstrating its effectiveness.