Detection of Neutral Phrases and Polarity Shifting of Few Phrases for Effective Classification of Opinionated Texts

In this paper we present an approach to effectively classify opinions expressed by web users from the opinionated texts. Our earlier work used a threshold based approach for finding users opinion from opinionated texts collected from the first few results of web pages returned by different search engines. The Threshold based approach performed well for finding opinions from opinionated texts comprising mainly of positive opinionated texts. The same approach performed poorly for an even set of positive and negative opinionated texts. We aim to obtain a better classification of an opinion into positive and negative opinion for a balanced set of positive and negative opinionated texts. In our present approach, we extract phrases from opinionated texts and use these phrases to provide opinions of the opinionated texts. A better classification is accomplished by reducing neutral phrases as well as polarity shifting of few phrases. We create two lexicons namely general lexicon and domain lexicon and use these lexicons to achieve our objective. We obtain a better result to detect polar phrases using these lexicons compared to few well known lexicons. We also obtain a better result for classifying opinions from opinionated texts on products using our approach on different data sets.