This paper studies sentiment analysis in Weibo. The study focuses on three types of tasks:emotion sentence identification and classification, emotion tendency classification, and emotion expression extraction. An unsupervised topic sentiment model, UTSM, is proposed based on the LDA Collocation model to facilitate automatic hashtag labeling. A Gibbs sampling implementation is presented for deriving an algorithm that can be used to automatically categorize emotion tendency with computer. To address the issue of lower recall ratio for emotion expression extraction in Weibo, dependency parsing is used to divide dependency model into two categories with subject and object. Six dependency models are also constructed from evaluation objects and emotion words, and a merging algorithm is proposed to accurately extract emotion expression. Result of experiments indicates that the presented method has a strong innovative and practical value.