Adversarial perturbation and bidirectional attention mechanism for few-shot English text classification

In few-shot text classification, how well the query and support sets are encoded largely decides the final accuracy. Yet, most prior methods overlook the pairwise correspondences between them and treat all features as equally important, neglecting the varying informativeness within each set. Therefore, we propose a novel adversarial perturbation and bidirectional attention mechanism for few-shot English text classification. The model combines the global information extraction ability of GRU and the local detail learning ability of attention mechanism to model text features. The adversarial training strategy is introduced to improve the stability and accuracy of the model when facing fuzzy expression or perturbation input. To capture nuanced interactions, we employ a bidirectional attention module that aligns support and query instances, highlighting within-class distinctions and forging category prototypes with enhanced discriminative power. The proposed model is tested on the public few-shot classification data sets, and results show that the classification accuracy rates with proposed method are better than other advanced models.