Modified Prototypical Networks for Few-Shot Text Classification Based on Class-Covariance Metric and Attention

This paper propose a modified Prototypical Networks based on the Class-covariance metric and Attention (PNCA) against the few-shot text classification issue. In this method, we use the Mahalanobis distance as the classification of class-specific feature vectors, furthermore, we add the word-level features and class-level features of attention mechanism, which highlight the standout knowledge of few-shot data and learn a more discriminative prototype representation. Our experiments based on the Chinese News Classification(CNC) Dataset show this model almost reach state-of-the-art, and the accuracy is 1%~2% higher than the existing models.