Multi-class Emotions Classification by Sentic Levels as Features in Sentiment Analysis

Sentiment Analysis has become a critical research area in recent days and pervasive in real life. Considering the identification of Emotions from textual content, we propose the Hourglass of Emotions as the feature that comes from the intensity of affective dimensions and combination thereof. Thus, based on a news dataset labeled with six primary Emotions, we intend to solve the Multi-class Classification Problem comparing decomposition methods - One against All and One Against One - and several aggregation methods. As base classifiers algorithms, we adopted Support Vector Machine, Naive Bayes, Decision Tree and Random Forests. Anchored on the results, we found that it is feasible to use this new set of features. The combination of Support Vector Machine and WENG pairwise coupling method was the best one, producing an accuracy of 55.91%.