ML-SPD: Machine Learning based Sentiment Polarity Detection

Abstract

Internet revolution creates very important trends in people's life like news-portals, online-education, home-offices, online shopping, social media, etc. Without any controversy, social media is one of the most important outcomes of the Web. Today, social media is more than a communication channel where people have the opportunity to express their feelings, write their comments on microblogging sites, discussion groups, review sites, etc. These common habits have resulted in two important consequences: 1) Accumulation of very huge data on online platforms, 2) The requirement of automatic systems to classify these accumulated big data by subjective and sentimentally. In many cases, Sentiment Polarity Detection (SPD) in text may be an urgent requirement, rather than identifying the subject of the text. For instance, positively or negatively labeled product reviews may give sufficient summary information to readers about the review. In this study, to solve SPD problem we explore different text representation models in conjunction with state-of-the-art traditional Machine Learning techniques: Support Vector Machines (SVM), Neural Networks (NN), Nave Bayes (NB), and combination of NB and SVM classifier (NBSVM). We perform experiments on three publicly available benchmark movie review datasets in different languages: CornellPD in English, HUMIR in Turkish and SerbSPD-2C in Serbian. Experimental results confirm that the presented techniques achieve improvements over the previously published techniques applied to movie reviews datasets in Turkish and English. Developed software package “ML-SPD” is made publicly available to the research community so it can serve as a good baseline for future research.