HYBRID FEATURE SELECTION FRAMEWORK FOR SENTIMENT ANALYSIS ON LARGE CORPORA

Abstract

Sentiment analysis has recently drawn considerable research attention in recent years owing to its applicabilityin determining users’ opinions, sentiments and emotions from large collections of textual data. The goal ofsentiment analysis centred on improving users’ experience by deploying robust techniques that mine opinionsand emotions from large corpora. There are several studies on sentiment analysis and opinion mining fromtextual information; however, the existence of domain-specific words, such as slang, abbreviations andgrammatical mistakes further posed serious challenges to existing sentiment analysis methods. In this paper, wefocus on the identification of an effective discriminative subset of features that can aid classification of users’opinions from large corpora. This study proposes a hybrid feature-selection framework that is based on thehybridization of filter- and wrapper-based feature selection methods. Correlation feature selection (CFS) ishybridized with Boruta and Recursive Feature Elimination (RFE) to identify the most discriminative featuresubsets for sentiment analysis. Four publicly available datasets for sentiment analysis: Amazon, Yelp, IMDB andKaggle are considered to evaluate the performance of the proposed hybrid feature selection framework. Thisstudy evaluates the performance of three classification algorithms: Support Vector Machine (SVM), Naïve Bayesand Random Forest to ascertain the superiority of the proposed approach. Experimental results across differentcontexts as depicted by the datasets considered in this study clearly show that CFS combined with Borutaproduced promising results, especially when the features selected are passed to Random Forest classifier.Indeed, the proposed hybrid framework provides an effective way of predicting users’ opinions and emotionswhile giving substantial consideration to predictive accuracy. The computing time of the resulting model isshorter as a result of the proposed hybrid feature selection framework.