Should you use a vote module for sentiment classification of online social text?

Abstract

In this work, we conduct a study where we compare the usage of a single classifier and the usage of a majority vote system composed of multiple classifiers. Each classifier was created using machine learning techniques and trained with real data. For the domain, we considered textual expressions present on online social networks, which can be volatile in characters count. This work seeks to prove two hypothesis: (1) the usage of a vote module that considers the output of an odd number of classifiers, will address the advantages characteristics of each classifier while mitigating their disadvantages; (2) the usage of a vote system will enable classifiers to correctly classify new labels that were not defined in the training process (like classifying neutral sentiment in addition to positive and negative). Our vote module is composed by a Naïve Bayes, a Logistic Regression, and a Support Vector Machine classifier. The tests that we conducted consider the online social textual content that varies in the character counting and our results suggests that there is no need for a vote system when considering the online social content, like comments, that is typically informal and do not surpass the count of 500 characters.