Boosting Meta-Learning for Few-Shot Text Classification via Label-guided Distance Scaling

Researchers propose a Label-guided Distance Scaling (LDS) strategy to improve few-shot text classification by leveraging label semantics during both training and testing phases. The method addresses misclassification issues when randomly selected labeled samples don't provide effective supervision signals, demonstrating significant performance improvements over state-of-the-art models.

• →LDS strategy exploits label semantics as supervision signals in both training and testing stages to improve few-shot text classification accuracy.
• →The method includes a label-guided loss function that pulls sample representations closer to corresponding label representations during training.
• →A Label-guided Scaler is introduced to adjust sample representations with label semantics during testing, even when labeled samples are far from class centers.
• →Experimental results show the approach significantly outperforms existing state-of-the-art meta-learning models for text classification.
• →All datasets and code are made publicly available for research reproducibility.