In view of the problem that traditional long-short memory network (LSTM) can't automatically select the most important latent semantic factors in text categorization, this paper proposes an improved LSTM model. First, the traditional LSTM operation relationship is extended to the bidirectional mode, so that the network fully remembers the context of the input feature words; then the pooling layer is added in front of the output layer to better select the most important latent semantic factors. The experimental results of the Internet Movie Database review data show that the model is superior to the traditional long-short-time memory neural network and other similar models, thus revealing that the improved scheme proposed in this paper is effective for improving the accuracy of text classification.