Modelagem de tópicos em redes sociais de políticos brasileiros

Abstract

With the emergence of social networks, it has become possible to reduce the distance between people and even establish relationships between unknown people who share a common ideal. Taking advantage of this idea, the Brazilian political scene has been adapting to this technology with the acting of candidates in social media since mid-2010. Social networks have become a fundamental tool for the dissemination of election campaigns and definitions of strategies to please the your electorate. The popularity of using social networks results in the generation of a multitude of complex data that can be analyzed and exploited. With the large volume of data, there is a need to apply methods and tools in order to process and synthesize information, facilitating its interpretation. Among the methods used, topic modeling aims to find latent topics contained in a document taking into account the occurrence of words. Thus, the present work aims to verify the efficiency of topic modeling algorithm known as LDA. Therefore, it was defined as target audience the Senators of the Republic in exercise from 2015, aiming to collect information from their publications on public pages on the social network Facebook. After data collection and application of the LDA algorithm, the results obtained were verified through a manual classification, trying to evaluate their performance in terms of hits and errors. In this process, there were a total of 7.694 publications with appropriate classification (≥ 0.7) from the 112.584 publications. This rate represents 6.8% of the publications analyzed. From this amount of publications, a manual verification was performed to determine the number of hits and errors, which showed that the algorithm obtained 5,239 hits (4.7%) and 2.455 errors (2.2%). Thus, in the present work the application of the gls LDA algorithm was unable to precisely synthesize the relevant subjects in a given data set, taking into account their average standard efficiency for the classification of topics.