Abstract: In order to improve the accuracy of sentiment classification of short text, a T-CLSTM model was proposed to according to its characteristic. The model generates word topic vectors with LDA model, and constructs sliding window word topic context and hierarchical word topic context to extend the short text information. The composition of word topic, word topic context and the effect of the sliding window size on the topic context were discussed. The word vector and word topic context vectors are used as input features to train models for sentiment classification. Experimental results on the COAE2014 corpus show that the proposed model can obtain 92.3% accuracy, which is 2% and 4% higher than that of baseline algorithms SVM and LSTM.