Sentiment analysis is a very important technology in text mining.However

a number of systems require amounts of annotated training data in different fields.In order to solve these problems

an approach to polarity classification based on sentiment tags is proposed.Firstly

on the basis of all the documents

the sentiment-topic model is developed and the sentiment tags for each review are extracted.Then each review is divided into two sub-texts by these sentiment tags

and each sub-text is classified by exploiting the co-training algorithm.Finally

the category results of two sub-texts are combined to determine document-level polarity of each review.Experimental results show that compared with other algorithms

the method improves the classification precision without a large number of annotated samples.