Sentiment analysis has been a hot area in the research field of language understanding, but complex deep neural network used in it is still lacked. In this study, we combine convolutional neural networks (CNNs) and BLSTM (bidirectional Long Short-Term Memory) as a complex model to analyze the sentiment orientation of text. First, we design an appropriate structure to combine CNN and BLSTM to find out the most optimal one layer, and then conduct six experiments, including single CNN and single LSTM, for the test and accuracy comparison. Specially, we pre-process the data to transform the words into word vectors to improve the accuracy of the classification result. The classification accuracy of 89.7% resulted from CNN-BLSTM is much better than single CNN or single LSTM. Moreover, CNN with one convolution layer and one pooling layer also performs better than CNN with more layers.