With the rapid development of the Internet, the microblog platform, BBS, e-Commerce etc. gathered a lot of short messages/text, which contained subjective sentences. These sentences often had obvious inclination which reflected the sentiment of the author. By mining the author’s sentiment, such as like, angry, indignation, averseness, etc., we can analyze people’s opinion for some policy, people’s preferences for some commodities. So, in this paper, we proposed a new method for sentiment classification by combination of several machine learning algorithms, which included feature extraction and ontologies. The further optimization was also given in this paper. We evaluated our method on several datasets and achieved good results.