In order to alleviate the influence of word spelling mistakes, wrong order and mixed emotions on the final grasp of the overall sentiment score of reviews, these noise and mixed emotions can be dealt with in a targeted and orderly manner. The present invention designs a comprehensive deep capsule network classification model , simulating the logical steps of human reading, by capturing feature information at the word level, phrase level, and sentence level, respectively, to model reviews, specifically, modeling at the word level and phrase level with misspelling errors and word order errors Corresponding to the noise problem, the modeling at the sentence level corresponds to the emotional mixing problem, that is, each short sentence is regarded as a synonymy, and the impact of different synonymy groups on the final overall emotional attitude is dynamically considered. In terms of implementation, the BERT WordPiece vector and convolution can be used as word-level and phrase-level features, and then the capsule network can be used to obtain the final vector representation at the sentence level for classification.