[Objective] This paper focused on Chinese text sentiment classification based on deep belief network, especially the parameter selection and performance analysis of the network. [Methods] Chinese e-commercial reviews are as the object of the study, the unigram, bigram, POS, simple dependency label, sentiment score and triple dependency features are extracted and used as the input of deep belief network by setting different layers and different input numbers to compute the accuracy of sentiment classification. [Results] The results demonstrate that the triple dependency features as the input got better classification performance than the other features, but the number of hidden layers doesn’t have an effect on the classification accuracy. [Limitations] The methods aren’t conducted and verified on other deep learning models. [Conclusions] Deep learning has a good performance for sentiment analysis, but how to set up parameters still need to be further considered.