Reforma da Previdência no Brasil: Uma análise a partir de dados do Twitter

Abstract

The social security reform voted and approved in 2019 was one of the most relevant topics in Brazilian politics in recent years. Given the relevance of the topic, it is essential to analyze the opinion of citizens about it. This analysis can be done using data from social networks that Brazilians increasingly used over the years with the advent of digital inclusion. The social network best known for quick sharing of news and opinions in the form of short texts is Twitter, where these short texts are called tweets. In this final course assignment, we propose a qualitative analysis of the dataset already assembled with tweets about the social security reform in Brazil in 2019. The classification of users from the dataset was performed, distinguishing ordinary users from automated accounts (known as bots); we performed this analysis using the Botometer tool, which classified 8,180 users as bots among the 237,894 unique users from the dataset. With the completed classification of users, we analyze the possibility of a message propagation “network” had been used to propagate a certain opinion on the topic. Everything led to believe that there was no use of this type of “network”, as the bots produced 40,151 tweets out of the 980,577 tweets. We also classify the tweets through sentiment analysis, where the possible outcomes could be: “positive”, “negative” or “neutral”. Four models were used to do these classifications: Naïve Bayes, Logistic Regression, Random Forest, and Support Vector Machines (SVM). These models allowed a comparison of results with the results obtained by the creator of the dataset, as he used them in his work to classify the tweets. Regarding the classification of tweets produced by the bots, most tweets received the neutral classification for all models used. Something that may have led to this result is that media generally use automated accounts to replicate news published on their websites; thus, many media accounts were considered bots in this analysis. In the pre-processing stage, we treated hashtags and compared the new results of this treatment with the initial results. It was possible to observe that both for the original classification and for the new classification, Naïve Bayes classified most of the tweets as positive. In contrast, the other models classified most tweets as neutral. We observed this behavior in other analyses performed in the dataset, such as the classification of the thousand tweets who received the most favorites.