Sentiment analysis for helpful reviews prediction

Abstract

Nowadays, every purchase we plan can be alleviated by the advice of those that tried in the past the given product. As more and more reviews are available, it would be practical to filter the relevant reviews not only to speed up the decision process but also to improve it. Gathering only the helpful reviews would reduce information processing time and save effort. To develop this functionality we need reliable prediction algorithms to classify and predict new reviews as helpful or not, even if the review has not been voted yet. In this paper, we propose a new approach which predicts reviews helpfulness based on sentiment analysis. Our approach focused on sentiment features such as the degree of positivity and the degree of negativity, in addition to the simplistic counts computed directly from reviews. It also extracts emotions dimension by means of emotion lexicon. We proposed a solution to internally construct an emotion lexicon in order to overcome challenges of invented terms, domain dependency, and spelling mistakes. We applied the proposed approach to Facebook pages of six medical products. We obtain a prediction accuracy of 97.95% through SVM algorithm. We found that sentiment degree and sadness emotion are the most decisive sentiment features to predict review helpfulness. The word count and frequencies are important as they reflect the richness and the seriousness of the review, but sentiment and emotions are more decisive as they engage and influence users.