A Novel Feature Reduction Method in Sentiment Analysis

Abstract

With the genesis of the internet and the world wide web, we have seen an enormous growth of data and information on the web, as well as an increase in digital or textual opinions, sentiments and attitudes that have been remarked upon in reviews. More reviews in document-level have expressed a high-dimensional in feature space. The main task of feature selection and feature reduction is a reduction dimension in feature space while, at the same time, ensuring that is no loss in the minimum of accuracy. There are several factors to consider in reduction dimension of a term - document matrix of feature space. It can lead to removal of irrelevant and useless features; including as a result, more efficient categories, easier analysis more accurately of sentiment after reduction. For this aim, we have proposed a novel feature reduction method using standard deviation based on more variation or dispersion of features in feature space. We used three popular classifiers, namely: Naive Bayes, Maximum Entropy and Support Vector Machine for sentiment classification and ensemble of these classifiers. We then compared our proposed method with other feature reduction methods used on book and music reviews. Results show that classification by using the novel method improved the accuracy of sentiment classification.