Análise da evolução do discurso com uso de técnicas de modelagem de tópicos

Abstract

With the growing use of social networks, the impact that content on these networks has on today’s society also grows, influencing billions of people around the world. Since in this virtual world the diversity of subjects discussed and ideas expressed is very large, the analysis of the content displayed on social networks becomes of great importance for understanding the influence that these networks have on the behavior of society as a whole. Thus, the present work aims to analyze the evolution of discourses in posts on social networks in order to identify patterns that allow us to assess changes in discourse, unusual events or other relevant factors in their content. In the current scenario of the pandemic generated by the Sars-cov-2 virus, the study of the speeches of Brazilian municipal governments in posts on social networks, more precisely on Facebook, becomes relevant for the aforementioned analysis of the behavior of the population during the pandemic. This work sought to evaluate the application of two machine learning techniques for the analysis of speech evolution: (i) LDA (Latent Dirichlet Allocation) with temporal discretization in time intervals, and (ii) ) TOT (Topics Over Time), without temporal discretization. The models are compared quantitatively and qualitatively. The main contribution of this work is the comparison of the advantages and disadvantages of each model. It was not possible to obtain a clear result of which of the two is the best model, because although the LDA model with discretization better represents the evolution by topics, the TOT model presents better metrics in the linear regression that was performed as part of the quantitative analysis.