Prompt-based Data Augmentation and Dual-enhanced Semantic Contrastive Learning for Few-shot Classification

This study addresses overfitting and generalization issues in few-shot learning by introducing the PAD framework, which leverages data augmentation, multi-template prompt learning, and dual semantic contrastive learning. Utilizing the back-translation and paraphrasing with ChatGPT, this framework first enhances sample generation. It incorporates demonstrations in input text and multi-template tasks to activate latent knowledge in pre-trained models. Furthermore, the Dual-Enhanced Semantic Contrastive (DESC) strategy improves differentiation between positive and negative samples through dual loss functions. Experimental results across multiple datasets confirm that the proposed method significantly boosts model generalization.