Improving Meta-learning for Few-Shot Text Classification via Label Propagation

Meta-learning has shown remarkable success in few-shot learning, and a popular metric-based meta-learning method known as prototypical network has gained widespread adoption for addressing few-shot text classification tasks. However, its effectiveness is hampered by the reliance on limited labeled samples to define class prototypes, which may not accurately reflect the true class distribution, especially given the sparsity of textual data. This misalignment can consequently reduce the performance of few-shot text classification. To address this problem, we propose an optimization method for the prototypical network named LP-PN by leveraging a semi-supervised learning technique known as label propagation. LP-PN utilizes unlabeled samples from query set to optimize the representation of corresponding class prototypes, thus aligning prototypes more closely with the actual class distribution. Furthermore, to overcome the limitations of static distance metrics that fail to capture class differences, we incorporate a dynamic distance metric based on the attention mechanism in LP-PN. We evaluate our method across four benchmark datasets, and the results show that LP-PN demonstrates competitive performance compared with recent few-shot text classification methods.