With the popularity of the Internet, microblogging as a representative of the social network has generated a lot of data. Exploring useful information from these data has become an important direction for today's research. According to the characteristics of microblogging text, this paper presents a method based on joint classifier to filter out noise microblogging, and then uses LDA model for subject discovery. The joint classifier model is composed of naive Bayesian, support vector machine and decision tree. The accuracy of the combined classifier is 87%, which can clearly show that this classification method is feasible and effective.