With the increasing popularity of opinion-rich resources, opinion mining and sentiment analysis has received increasing attention. Sentiment analysis is one of the most effective ways to find the opinion of authors. By mining what people think, sentiment analysis can provide the basis for decision making. Most of the objects of analysis are text data, such as Facebook status and movie reviews. Despite many sentiment classification models having good performance on English corpora, they are not good at Chinese or other languages. Traditional sentiment approaches impose many restrictions on the raw data, and they don't have enough capacity to deal with long-distance sequential dependencies. So, we propose a model based on recurrent neural network model using a context vector space model. Chinese information entropy is typically higher than English, we therefore hypothesise that context vector space model can be used to improve the accuracy of sentiment analysis. Our algorithm represents each complex input by a dense vector trained to translate sequence data to another sequence, like the translation of English and French. Then we build a recurrent neural network with the Long-Short-Term Memory model to deal the long-distance dependencies in input data, such as movie review. The results show that our approach has promise but still has a lot of room for improvement.