CON-GAN-BERT: combining Contrastive Learning with Generative Adversarial Nets for Few-Shot Sentiment Classification

Abstract

Sentiment classification is a classical and important task of natural language processing (NLP), with the development of the Internet, there are multifarious reviews, comments and news produced everyday which need high cost to annotate, so it has become a challenge to develop a more effective sentiment classification model which requires less training samples. In this paper, we propose a sentence level sentiment classification model based on Contrastive Learning, Generative Adversarial Network and BERT (CON-GAN-BERT). Experiments on several public Chinese sentiment classification datasets show that CON-GAN-BERT significantly outperforms strong pre-training baseline, and still obtaining good performances for Few-Shot Learning without any data augmentation or unlabeled data.