Abstract. The Robert Gordon University (RGU) participated in the Opinion Retrieval Task of the Trec 2007 Blog Track. At the core of the system we developed is a set of training documents labeled with respect to opinion. These documents are used to train a classifier in order to classify the documents that are relevant to the given Trec topics. However, a major limitation with these training documents is that they are not always generic enough to serve as good training examples for all the 50 Trec topics. We therefore propose a solution that generalizes their content by exploiting the context of opinion related language constructs such as adjectives, verbs, and adverbs. Our system significantly improves over RGU’s previous Blog Track systems.