Text sentiment classification is a popular subject of natural language processing and the crucial problem in product evaluation. Based on semantic relationship of word vector and sentence vector and the impact of user information, product information to text sentiment classification, Cosine Similarity Long-Short Term (CSLSTM) network is proposed. CSLSTM considers attention mechanisms of user information and product information in various semantic levels. And it involves a effective initialization method in hidden level weights of word-level attention matrix according to similarity of word vector and sentence vector. The competitive results are derived from three sentiment classification datasets, Yelp13, Yelp 14, and IMDB.