APPLICATION OF THE TRANSFORMER ARCHITECTURE FOR FAKE NEWS DETECTION IN PORTUGUESE NEWS ARTICLES

Abstract

The dissemination of false content in digital environments represents a challenge for society and for the development of automated text analysis tools. This study aimed to experimentally evaluate the performance of the pretrained BERTimbau model, based on the Transformer architecture, in the binary classification of true and false news articles written in Brazilian Portuguese. The Fake.br-Corpus was used, comprising 7,200 news articles divided in a stratified manner into training, validation, and test sets, with 5,760, 720, and 720 records, respectively. The texts were tokenized using the tokenizer associated with the neuralmind/bert-base-portuguese-cased model, and fine-tuning was performed with the Transformers library in the Google Colaboratory environment. The best checkpoint was selected based on the F1-score obtained on the validation set and was evaluated exclusively on the test set. The model achieved an accuracy of 99.86%, precision of 100.00%, recall of 99.72%, and F1-score of 99.86%. The confusion matrix showed 360 real news articles correctly classified, 359 false news articles correctly classified, no false positives, and one false negative. The results indicate high performance on the evaluated corpus, although they do not guarantee generalization to other sources, periods, or production contexts. It is concluded that the study demonstrates the experimental feasibility of a BERTimbau-based pipeline for text classification in Portuguese, without replacing human fact-checking processes.