Visual analysis of conflicting opinions

Understanding the nature and dynamics of conflicting opinions isa profound and challenging issue. In this paper we address severalaspects of the issue through a study of more than 3,000 Amazoncustomer reviews of the controversial bestseller The Da VinciCode, including 1,738 positive and 918 negative reviews. Thestudy is motivated by critical questions such as: What are thedifferences between positive and negative reviews? What is theorigin of a particular opinion? How do these opinions change overtime? To what extent can differentiating features be identifiedfrom unstructured text? How accurately can these features predictthe category of a review? We first analyze terminology variationsin these reviews in terms of syntactic, semantic, and statisticassociations identified by TermWatch and use term variationpatterns to depict underlying topics. We then select the mostpredictive terms based on log likelihood tests and demonstratethat this small set of terms classifies over 70% of the conflictingreviews correctly. This feature selection process reduces thedimensionality of the feature space from more than 20,000dimensions to a couple of hundreds. We utilize automaticallygenerated decision trees to facilitate the understanding ofconflicting opinions in terms of these highly predictive terms.This study also uses a number of visualization and modeling toolsto identify not only what positive and negative reviews have incommon, but also they differ and evolve over time.