PSSOP: A Prompt-style Approach for Few-Shot Text Classification

Pretrained language models have achieved significant performance in various NLP downstream tasks. However, their effectiveness is heavily dependent on having a sufficient amount of labeled data for supervised training and large parameter scales, which hinder efficient inference and fine-tuning during training. Conversely, smaller-scale models are more convenient to tune but often underperform, especially in scenarios where only a few examples are available for training (few-shot scenarios). This limitation poses a significant challenge in practical applications where obtaining extensive labeled data or tuning complex models may not be feasible. To address this serious deficiency, prompt learning methods have emerged as an effective solution in the NLP field. These methods aim to enhance pretrained models’ few-shot learning capabilities without requiring extensive model tuning. In this paper, we utilized the idea of building prompts and propose a novel solution called Prompt-Style Sentence Order Prediction (PSSOP) method that reconstructs downstream tasks as sentence order prediction (SOP) tasks. We introduce an additional learnable module along with a two-stage data augmentation and contrastive learning strategy to effectively addressing NLP few-shot classification tasks while constraining the model’s parameter scale. Experimental results on 6 classification task datasets including news classification, sentiment analysis, and aspect-based sentiment analysis demonstrate the efficacy of our proposed method.