Comparação entre abordagem neurais e não neurais para modelagem de tópicos em tweets sobre Institutos Federais Brasileiros.

Abstract

The analysis of public perception of educational institutions on social media can reveal topics of community interest and support improvements in communication and academic management. Topic Modeling (TM) techniques allow for the automation of this identification, but studies evaluating aspects such as coherence, diversity, stability, and computational time of TM-generated topics in the educational context remain scarce. This work investigates these dimensions by applying TM to a novel corpus of tweets about Brazilian Federal Institutes, segmented by five macro-regions. The results show that, although traditional methods such as LDA and NMF perform better in terms of diversity and stability, respectively, the BERTopic approach with the BERTweet-BR model produced topics with higher coherence compared to LDA.