An angular shrinkage BERT model for few-shot relation extraction with none-of-the-above detection

Abstract

Few-shot relation extraction aims to solve the problem of insufficient annotated data in relation extraction tasks. Through the comparison between samples, few-shot relation extraction achieves lower-cost relation classification. However, most existing methods only do classification within the scope of enumerated relations. For one of the main challenges faced by the application of few-shot relation extraction-the recognition of the none-of-the-above instances, there has been few works on it. In this paper, we propose an angular shrinkage BERT model for the few-shot relation extraction task with none-of-the-above detection, which uses an additive angular loss to enlarge the margins of different classes in the feature space, and obtain highly discriminative features to improve the recognition ability for none-of-the-above instances. Meanwhile, we present a two-stage training strategy to enhance the stability of the performance. We evaluate our model on the most used few-shot relation extraction dataset FewRel. Experimental results show that our approach outperforms previous sentence-pair methods in scenarios containing none-of-the-above instances, and also achieves improvements on the traditional few-shot relation extraction task compared with our baseline model.