Topic-related Sentiment Analysis for Discovering Contradicting Opinions in Weblogs

This work addresses the problem of analyzing the evolution of community opinions across time. First, a two-step approach is introduced to determine a continuous sentiment value for each topic discussed in a text based on SentiWordNet as lexical resource. Sentences are clustered according to their topic using Latent Dirichlet Allocation. Both steps are extensively evaluated and tested. The output is then exploited for studying contradictions among weblog posts and comments. We introduce a novel measure for contradictions based on a mean value and the variance of opinions among different posts. In addition, a method is proposed, which identifies posts with contradicting opinions on certain topics on a basis of such a measure. It can be used to analyze and track opinion evolution over time and to identify interesting trends and patterns. The developed algorithm is applied to a dataset of medical blogs and comments on political news with promising performance and accuracy.