Ranking blog posts that express opinions regarding a given topic should serve a critical function in helping users. We explored a couple of methods for opinion retrieval in the framework of probabilistic language models. The first method combines topic-relevance model and opinion-relevance model, at document level, that captures topic dependence of the opinion expressions. The second method combines the aforementioned topic-opinion relevance models at sentence level, and accumulates the negative cross entropy between the combined relevance models and each sentence model to obtain a document-level score. This paper reports the overview of our methods and the evaluation results on the Opinion Retrieval Task at the TREC 2007 Blog Track.