Solving Data Imbalance in Text Classification with Constructing Contrastive Samples

Contrastive learning (CL) has been successfully applied in Natural Language Processing (NLP) as a powerful representation method and shown promising results various downstream tasks. Recent research highlighted the importance of constructing effective contrastive samples through data augmentation. However, current augmentation methods primarily rely on random word deletion, substitution, cropping, which may introduce noisy hinder learning. In this article, we propose novel approach to address imbalance text classification by samples. Our involves use Label-indicative Component generate high-quality positive for minority class, along with introduction Hard Negative Mixing strategy synthesize challenging negative at feature level. By applying supervised these samples, are able obtain superior representations, significantly benefit tasks imbalanced data. effectively mitigates distributional biases promotes noise-resistant To validate effectiveness our method, conducted experiments benchmark datasets (THUCNews, AG’s News, 20NG) well FDCNews dataset. The code is publicly available following GitHub repository: https://github.com/hanggun/CLDMTC.