Abstract

In the field of few-shot learning, the increasing emergence of novel relations poses a significant challenge to existing models that leverage knowledge from past relation classification tasks. This issue becomes more pronounced when the volume of novel relations overtakes that of existing ones, underscoring the need for models that are less dependent on base relations and more adept at learning from the novel ones. This paper addresses the challenging Class-Incremental Fewshot Relation Classification (CIFRC) problem by a model named ReliDispProto: Reliability and Dispersion-infused Prototypical network. ReliDispProto employs a sophisticated similarity metric that accounts for the reliability of query-to-prototype matches and the distribution dispersion of support classes. Based on the metric, unlabeled instances are efficiently classified into support classes across a series of sessions. While training on novel relations, ReliDispProto is fine-tuned by classification reliability assessed in each iteration. To further enhance its performance, we integrate teacher-student knowledge distillation and label smoothing techniques. These additions effectively alleviate issues such as catastrophic forgetting and overfitting. Experimental results on three public datasets reveal that ReliDispProto significantly outperforms existing state-of-the-art methods, achieving up to 15.63% improvements in accuracy. The datasets and source code for ReliDispProto are accessible from http://github.com/nickhcx/ReliDispProto.