A Chinese Few-Shot Text Classification Method Utilizing Improved Prompt Learning and Unlabeled Data

Insufficiently labeled samples and low-generalization performance have become significant natural language processing problems, drawing concern for few-shot text classification (FSTC). Advances in prompt learning significantly improved the of FSTC. However, methods typically require pre-trained model tokens vocabulary list training, while different models token coding structures, making it impractical to build effective Chinese from previous approaches related English. In addition, a majority current do not make use existing unlabeled data, thus often leading unsatisfactory real-world applications. To address above limitations, we propose novel FSTC method called CIPLUD that combines an which are used small amount data. We two modules: Multiple Masks Optimization-based Prompt Learning (MMOPL) module One-Class Support Vector Machine-based Unlabeled Data Leveraging (OCSVM-UDL) module. The former generates prefixes with multiple masks constructs suitable templates labels. It optimizes random combination problem during label prediction joint probability length constraints. latter, by establishing OCSVM trained vector space, selects reasonable pseudo-label data each category large After selecting mixed them annotated obtain brand new training then repeated steps modules as iterative semi-supervised optimization process. experimental results on four benchmark datasets demonstrate our proposed solution outperformed other average accuracy improvement 2.3%.