Joint Few-Shot Text Classification Aided by Label Semantic and Sentence-Aware Interaction

Meta-learning has achieved remarkable performance to address few-shot text classification. However, previous studies fail to attach importance to the label semantic completely and overlook the implicit interaction between lexical features of sentences. In this paper, we explore three enhancement components of the meta-learner aided by the label semantic and sentence-aware interaction, e.g., the label-augmented encoder, the interaction extractor, and the label semantic discriminator. Significantly, these modules are agnostic to the choice of few-shot text classification methods and can be easily incorporated into various existing meta-learning frameworks to improve the classification performance and adaptive ability of the model. We conduct extensive experiments on five benchmark datasets. The results demonstrate that meta-learning approaches upgraded by the above three enhancement components obtain considerable superiority over state-of-the-art models in all datasets. In particular, the average accuracy of 1-shot classification and 5-shot classification is increased by 3.60% and 2.28%, respectively.