2024/12/13 by Guanghua Hou, Shuhui Cao, Hou, Guanghua +5
Computer Science · #Text and Document Classification Technologies
paper · pdf · doi:10.48550/arxiv.2412.10110
As an algorithmic framework for learning to learn, meta-learning provides a\npromising solution for few-shot text classification. However, most existing\nresearch fail to give enough attention to class labels. Traditional basic\nframework building meta-learner based on prototype networks heavily relies on\ninter-class variance, and it is easily influenced by noise. To address these\nlimitations, we proposes a simple and effective few-shot text classification\nframework. In particular, the corresponding label templates are embed into\ninput sentences to fully utilize the potential value of class labels, guiding\nthe pre-trained model to generate more discriminative text representations\nthrough the semantic information conveyed by labels. With the continuous\ninfluence of label semantics, supervised contrastive learning is utilized to\nmodel the interaction information between support samples and query samples.\nFurthermore, the averaging mechanism is replaced with an attention mechanism to\nhighlight vital semantic information. To verify the proposed scheme, four\ntypical datasets are employed to assess the performance of different methods.\nExperimental results demonstrate that our method achieves substantial\nperformance enhancements and outperforms existing state-of-the-art models on\nfew-shot text classification tasks.\n