Efficient on-demand Opinion Mining

Every day, a multitude of people express their opinions regarding diverse entities, such as services, places and products, in blogs (e.g., The BBC “Have Your Say” Blog), online forums (e.g., slashdot.org) and review sites (e.g., www.amazon.com). This constantly growing availability of opinionated content has created massive amounts of extremely valuable information. Currently, search engines are unable to explore such information, because (1) it is difficult to distinguish opinionated content from factual content, and (2) opinionated content may present different connotations or polarities (i.e., positive or negative, interesting or boring etc.). Recently, some attention has been devoted to the first problem − opinion retrieval, which consists of distinguishing opinionated content from factual content. However, research on opinion mining, which consists in classifying opinionated content with regards to the opinion it expresses, is still lacking. The main challenge is that the search space is huge due to the sparseness typically associated with textual evidence, and thus, the classification model needs to be very complex in order to achieve accurate results. In this paper we present a novel strategy for opinion mining, based on a lazy, on-demand, associative classification approach which reduces the complexity of the model by adopting a highly specific bias during the inductive process. The proposed approach was evaluated using collections obtained from two actual application scenarios: an online forum and a large product review site. The results demonstrate that the proposed approach can provide gains up to 9%, when compared against the state-of-the-art general purpose classification approach.