Análise de sentimentos no Twitter sobre a Reforma da Previdência no ano de 2019

Abstract

Social networks are prevalent virtual environments, being present in the daily lives of millions of people around the world. There are several networks, each with their specializations, such as sharing videos, images, or text. Twitter is one of the most popular, focused on sharing short texts, the so-called tweets. This course conclusion paper aims to perform a sentiment analysis regarding the Brazilian Social Security Reform, approved in 2019. The sentiment analysis combines machine learning concepts, natural language processing, linguistics, and textual analysis. It aims to extract opinions from texts through a classification traditionally related to a “positive” or “negative” feeling. In this context, we conducted a set of experiments to determine the impact of different pre-processing techniques, an important step that aims to improve the classification’s performance. We perform a comparison of four machine learning algorithms used to classify the texts: Naïve Bayes, Logistic Regression, Support Vector Machines (SVM), and Random Forest. The results of the pre-processing revealed that different techniques could affect each algorithm differently. The algorithms also obtained performances considered adequate, with Naïve Bayes having the lowest accuracy, 84% with the application of text pre-processing techniques. Analyzing the classification results, Naïve Bayes was the one that classified more as positive, while the others as neutral. Comparisons were also made between the classifications’ results, showing that the most similar was Logistic Regression and SVM, with 91% of equal classifications. The most distinctive was Naïve Bayes and Random Forest, at only 46% equal ratings.