Automatic Classification of Portuguese Municipalities’ Posts On Social Media: A Bertimbau-Based Approach

Communication between municipalities and citizens through social media, particularly Facebook, has become a fundamental pillar of e-participation in Portugal. However, manually analyzing the vast volume of interactions generated for research purposes is a very laborious process and difficult to sustain in the long term. This article presents the development and validation of an Artificial Intelligence system for the automatic classification of municipal publications, using the pre-trained language model BERTimbau. The corpus used includes 6,897 publications from 84 Portuguese municipalities, classified in a standardized taxonomy of 11 strategic themes. To mitigate the challenges inherent in short texts and the strong class imbalance, the pipeline included lemmatization techniques and the use of Macro F1 as the central evaluation metric. The model achieved an Accuracy of 0.7446 and a Macro F1 of 0.7124, demonstrating effectiveness in identifying driving themes of participation, such as municipal works and political activity. The results validate the transition to automated systems, allowing researchers to obtain classified research data in a timely manner and eliminate the subjectivity of human analysis.