Given the swift progress in natural language processing(NLP) technology, text classification and sentence-level classification have gained increasing significance in areas such as retrieval of information, sentiment analysis, and question answering systems. Traditional sentence classification methods often depend on large-scale labeled training data and face difficulties in managing semantic similarity. To confront these issues, this research introduces an innovative methodology using dual contrastive learning integrated with label extended semantic features for sentence classification. Specifically, we propose an effective representation method that combines word embeddings and tag-extended semantic embeddings, which is capable of extracting profound semantic attributes of labels. Secondly, based on a dual contrastive learning framework, we promote the model to enhance the acquisition of semantic similarity between sentences by formulating matching pairs containing positive and negative examples. During the training process, label expansion semantic features are integrated into the contrastive loss function, which augments the model’s comprehension and modeling capabilities of label expansion semantics. Our empirical evaluations conducted on publicly accessible datasets in both Chinese and English reveal enhancements in accuracy attributable to our proposed methodology.