Few-shot relation classification using clustering-based prototype modification

Abstract

Few-shot relation classification is a natural language processing task that aims to enable models to recognize new relational categories of query instances by training on base classes with few labeled support instances. Many recently proposed prototypical-network-based approaches have exhibited excellent performance, some of which include additional information to further improve the model performance further. However, existing methods ignore the category information hidden in the query set and the hierarchical information that exists in relational hyponymy and fail to address the fundamental problem whereby the obtained prototype is always unreliable owing to the limited instances in the support set. In this study, we propose a novel prototypical-network-based model for the few-shot relation classification task to generate more precise relational prototypes by taking advantage of the category information hidden in the query set. In the model, we design a novel relational-similarity-based instance adaptive local loss that leverages relational hierarchical information to distinguish the hard instances belonging to similar relations effectively. Experiments on the FewRel dataset demonstrate that our approach achieves state-of-the-art performance.