Supervised Contrastive Learning with Term Weighting for Improving Chinese Text Classification

With the rapid growth of information retrieval technology, Chinese text classification, which is basis content security, has become a widely discussed topic. In view huge difference compared with English, task more complex in semantic representations. However, most existing classification approaches typically regard feature representation and selection as key points, but fail to take into account learning strategy that adapts task. Besides, these compress word vector, without considering distribution term among categories interest. order improve effect unified method, called Supervised Contrastive Learning Term Weighting (SCL-TW), proposed this paper. contrastive makes full use large amount unlabeled data model stability. SCL-TW, we calculate score weighting optimize process augmentation text. Subsequently, transformed features are fed temporal convolution network conduct representation. Experimental verifications conducted on two benchmark datasets. The results demonstrate SCL-TW outperforms other advanced by an amazing margin.