Impacto do pré-treinamento via MLM no desempenho e na interpretabilidade do BERTimbau para a classificação de notícias falsas

Abstract

The surge of misinformation on social media and digital platforms constitutes one of the major challenges of the information age, impacting public opinion formation and threatening trust in democratic institutions and processes. Fake news has spread with unprecedented speed and reach, driven by algorithmic dynamics and accelerated online content consumption. Given this scenario, this work investigated the impact of Masked Language Modelpre-training on the performance and interpretability of the BERTimbau model for fake news classification in Portuguese. Using the FakeRecogna 2.0 corpus and a methodology based on StratifiedGroupKFold cross-validation with 10 folds, a reference model was compared to a version specialized in the Brazilian journalistic domain. The application of the eXplainable Artificial Intelligence technique, LIME, allowed auditing the classifier’s behavior regarding the concept drift phenomenon across three temporal windows, landmarks previously identified in the literature as significant change points for this dataset. Quantitative results indicated a performance gain using MLM pre-training, with the average F1-Score rising from 0.8936 to 0.9708 and average recall jumping from 0.8250 to 0.9479, suggesting that domain adaptation may have enhanced the model’s sensitivity to textual nuances that previously resulted in false negatives. From a qualitative perspective, the analysis raises the hypothesis that the classifier correlates syntactic cohesion and stopword frequency with journalistic veracity, whereas it appears to use structural viralization patterns as indicators of deceptive content. Furthermore, a trend of sanitary thematic consistency was observed in true news throughout the three periods, while fake news appeared to migrate more rapidly towards political polarization. Despite high metrics, limitations were noted in interpreting metaphors and potential noise stemming from fact-checking agency metadata, reinforcing that domain specialization and explainability are relevant tools in the pursuit of robustness and transparency in fake news classification in Portuguese.