Prompt and Contrastive Learning for Few-shot Sentiment Classification

Abstract

Sentiment classification is a hot topic in the field of natural language processing. Currently, state-of-the-art classification models follow two steps: pre-training a large language model on upstream tasks, and then using human-labeled data to fine-tune a task-related model. However, there is a large gap between the upstream tasks of the pre-trained model and the downstream tasks being performed, resulting in the need for more labeled data to achieve excellent performance. Manually annotating data is expensive. In this paper, we propose a few-shot sentiment classification method based on Prompt and Contrastive Learning (PCL), which can significantly improve the performance of large-scale pre-trained language models in low-data and high-data regimes. Prompt learning aims to alleviate the gap between upstream and downstream tasks, and the contrastive learning is designed to capture the inter-class and intra-class distribution patterns of labeled data. Thanks to the integration of the two strategies, PCL markedly exceeds baselines with low resources. Extensive experiments on three datasets show that our method has outstanding performance in the few-shot settings.