Sentiment analysis determines people’s perceptions of various events by employing natural language processing techniques. Advances in machine learning and deep learning have made it possible to determine emotional tendencies more accurately. Nevertheless, RNN cannot capture long-term dependencies in text, LSTM can only unidirectionally extract contextual information in text, and the training of Transformer requires a lot of computational resources. Also in specific domains will be confronted with specific vocabularies, the processing of these vocabularies directly affects the final classification performance. In this paper, we propose a new GloVe-based Mamba model that combines the advantages of GloVe word embeddings and Mamba networks. First, the model adds specific words to Jieba for segmentation in the data preprocessing stage to make the segmentation results more accurate. Second, the model combines GloVe word embeddings to extract textual contextual information and solve the problem of lack of semantic information. In addition, the Mamba network is utilized to selectively process the input information to improve the accuracy and computational efficiency of Chinese text sentiment polarity classification.