The explosive growth of the user-generated content on the Web has offered a rich data source for mining opinions. However, the large number of diverse review sources challenges the individual users and organizations on how to use the opinion information effectively. Therefore, automated opinion mining and summarization techniques have become increasingly important. Different from previous approaches that have mostly treated product feature and opinion extraction as two independent tasks, we merge them together in a unified process by using probabilistic models. Specifically, we treat the problem of product feature and opinion extraction as a sequence labeling task and adopt Conditional Random Fields models to accomplish it. As part of our work, we develop a computational approach to construct domain specific sentiment lexicon by combining semi-structured reviews with general sentiment lexicon, which helps to identify the sentiment orientations of opinions. Experimental results on two real world datasets show that the proposed method is effective.