VDCL : A supervised text classification method based on virtual adversarial and contrast learning

In recent years, Dual Contrastive Learning (DualCL) has achieved better results in the field of text classification by making full use of label and inter-sample information in a supervised environment. But how to improve the generalization performance and robustness of neural language models is still a question that researchers have been exploring. In this paper, we propose a framework for introducing virtual adversarial training into Dual Contrastive Learning, The framework first pre-trains a large neuro-linguistic model and then fine-tunes the model using Dual Contrastive Learning to learn both the features of the input samples and the parameters of the classifier. In summary, this means that virtual adversarial training is used to improve the robustness and generalization performance of the model, and then Dual Contrastive Learning is used to learn better feature representations using supervised signals between training samples. Our experiments on three benchmark text classification datasets demonstrate the improved feature extraction and classification accuracy of VDCL and confirm the capability of VDCL in the area of text classification.