Most of the previous approaches in opinion mining focus on the classifications of opinion polarities, positive or negative, expressed in customer reviews. In this paper, we present the problem of extracting contextual opposite sentiments in classified free format text reviews. We adapt the sequence data model to text mining with Part-of-Speech tags, and then we propose a belief-driven approach for extracting contextual opposite sentiments as unexpected sequences with respect to the opinion polarity of reviews. We conclude by detailing our experimental results on free format text movie review data.