Abstract: Aiming at the problem of manual annotation in the text sentiment analysis, a new method based on five tuple of appraisal expression is proposed. This  method just needs appropriate sentiment dictionary. The sentiment tendencies of comments are analyzed without lots of markup work. Through the combination of unsupervised and supervised learning methods to construct the evaluation thesaurus and evaluation object list; the extraction of appraisal expression is based on these lists, using linear chain conditional random fields model, which is in the chain of sentiment words. Finally, evaluation objects are divided into four categories and emotional words are divided into five types according to the relationship between semantic collocation, combined with the influence of sentence pattern, negative word and degree word on the sentiment analysis, a method of calculating the sentiment tendency of the text is put forward. Compared with other methods, this method based on the appraisal expression has obtained better F value, and it has a certain cross domain.

Key words: sentiment classification, information extract, opinion mining, sentiment analysis