Um Modelo de Classificação de Polaridade em Cinco Níveis

Abstract

Sentiment analysis is the area of study that observes people’s opinions, feelings, assessments, attitudes and emotions around entities such as products, services, organizations and events. In the real world, companies and organizations are often interested in knowing public opinions about their products and services. Consumers are also interested in knowing the opinion of those who bought a product before buying it. Other people are interested in knowing the opinions of others about certain candidates in an election process before making a decision about who will vote. The objective of this work is to develop a supervised learning method capable of classifying tweets in five levels of polarity (complete approval, punctual positive opinion, neutral opinion, punctual negative opinion and complete rejection) using tweets of the political context as a case study. In order to do this, we investigated whether three-level polarity detection techniques are capable of providing good evidence for training and classifier testing in the context of the five polarity levels in tweets. Based on this idea, we propose learning strategies with the Decision Trees, Naive Bayes and SVM classifiers using as features models: bag-of-words, evidence from results of methods that classify polarity in three levels and a meta-level extraction method. The results showed that there is a gain in the accuracy of the classifiers when combining the different models of features.