Abstract: A Rough Set and Information Gain based on sentiment feature selection method is proposed for building a solid foundation in sentiment analysis. The novel method firstly uses Information Gain to select a feature subset which has high relativity with the class attribute. Secondly, the features which have high redundancy will be eliminated by Rough Set. Experimental results on several datasets reveal the method makes accuracy increase 4-9 percentages than other methods. It is an outstanding feature selection method and has significance in sentiment analysis.