In this paper we proposed a hierarchical generative model based on Nai¨ve Bayes and LDA for unsupervised sentiment analysis at sentence level and document level of granularity simultaneously. In particular, our model called NB-LDA assumes that each sentence instead of word has a latent sentiment label, and then the sentiment label generates a series of features for the sentence independently in the Nai¨ve Bayes manner. The idea of NB assumption at sentence level makes it possible that we can use advanced NLP technologies such as dependency parsing to improve the performance for unsuper...
