[Objective] To solve the problem that the traditional static word vector embedding method can't effectively deal with polysemy in Chinese text, and it is difficult to mine the contextual emotional features and internal semantic association structure. [Methods] Firstly, in one channel, the emotional elements related to the text are integrated into Word2Vec and FastText word vectors by using rough data reasoning, and CNN extracts the local features of the text; Secondly, the other channel uses BERT for word embedding supplement, and BiLSTM obtains the global features of the text. Finally, the attention calculation module is added for the deep interaction of dual channel features. [Results] The highest accuracy of the experiment on three Chinese data sets is 92.43%, which is 0.81% higher than the highest value of the benchmark model.[Limitations] The selected data set is only for coarse-grained emotion classification modeling, and experiments in fine-grained field have not been considered. [Conclusions] Judging the emotion category according to its output results effectively improves the performance of the model in Chinese text emotion classification.