Visual Analysis of Conflicting Opinions.

Résumé En

Understanding the nature and dynamics of conflicting opinions in texts is a challenging issue concerning sense making and decision making in visual analytics. In this paper we address this issue with a visual analysis of how positive and negative reviews of the controversial bestseller The Da Vinci Code differ. An integrative approach is proposed to capture the dynamics of opinions at both macroscopic and microscopic levels. More than 3,000 customer reviews of the book retrieved from Amazon.com, including 1,738 positive and 918 negative reviews, are analyzed in this approach to address questions such as: what are the differentiating features of positive and negative reviews? How did positive and negative reviews evolve through time? To what extent can discriminating features be automatically extracted without prior knowledge of the subject? To what extent can these features accurately predict the general opinion of a given review? In order to identify differentiating features, we focus on underlying terminology variations and use TermWatch, a tool for terminology variation analysis, to construct multi-layered networks of terms based on syntactic, semantic, and statistic associations. Time series of terms with strong variations are constructed on a monthly basis. We have utilized a number of visualization and modeling tools to conduct a visual analysis of these identified features. The results show not only what positive and negative reviews have in common, but also what differentiate them persistently over time. Furthermore, how earlier reviews may have influenced subsequent reviews are also investigated. Challenges for future work are identified.