The authors present a hierarchical neural network H-RNN-CNN as a general model to represent text in sentiment analysis. Firstly, since information may lose in long text, the authors divide the text by sentence and use them as middle layer. Secondly, recurrent neural network is used to process sequence and relationship across sentences is captured by convolutional neural network. Moreover, the effectiveness of the variants of recurrent neural network and the pre-trained embedding are discussed. Experiment results demonstrate that the approach works well on several datasets.