Dynamic Prototype Selection by Fusing Attention Mechanism for Few-Shot Relation Classification

Abstract

In a relation classification task, few-shot learning is an effective method when the number of training instances decreases. The prototypical network is a few-shot classification model that generates a point to represent each class, and this point is called a prototype. The mean is used to select prototypes for each class from a support set in a prototypical network. This method is fixed and static, and will lose some information at the sentence level. Therefore, we treat the mean selection as a special attention mechanism, then we expand the mean selection to dynamic prototype selection by fusing a self-attention mechanism. We also propose a query-attention mechanism to more accurately select prototypes. Experimental results on the FewRel dataset show that our model achieves significant and consistent improvements to baselines on few-shot relation classification.