JMCL: Joint Multi-Contrastive Learning for Text Classification

Contrastive learning has been widely applied in representation learning, demonstrating superior performance across various fields such as vision, text, and speech. This paper proposes a method called Joint Multi-contrastive Learning (JMCL) that combines supervised contrastive learning with self-supervised contrastive learning to enhance performance in text classification tasks. By employing label-aware data augmentation in supervised learning, the method encodes input data using different labels, facilitating the tight clustering of similar samples and the effective separation of dissimilar samples. In self-supervised learning, random perturbations of text embeddings are utilized for data augmentation, increasing sample diversity and enhancing model transferability. Experimental results demonstrate that the JMCL method achieves improvements across four out of five benchmark datasets based on BertGCN, with a notable increase of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{0. 7 \%}$</tex> on the Ohsumed dataset compared to BertGCN. These findings validate the effectiveness of Joint Multi-contrastive Learning in text classification tasks and offer new insights and directions for future research.