Análise comparativa de métodos de detecção automática de mensagens ofensivas em textos curtos e ruidosos

Abstract

With the increasing use of social media and the ease of access to digital content—especially among children and adolescents—there has been a significant rise in cases of cyberbullying and online harassment in recent years. In response, several content moderation tools have been developed, such as comment filters, reporting systems, and user profiles dedicated to moderation. However, due to the vast amount of information constantly generated on social media platforms, manual moderation has become impractical, highlighting the importance of automated moderation in reducing the incidence of digital crimes. This work addresses the automatic identification of aggressive behavior in offensive messages found in short and noisy texts using machine learning and deep learning algorithms. A public dataset extracted from platform X was used, containing 20,001 sentences labeled as aggressive 39.1% or non-aggressive 60.9%. Supervised learning models were trained using stratified cross-validation, employing text preprocessing techniques and various algorithms, including BERT, FastText, and ensemble methods, with the goal of assessing the effectiveness of these approaches in the automatic detection of textual aggressiveness. The results showed that the BERT and FastText models achieved excellent recall scores, reaching 96.5% and 95.8%, respectively, significantly outperforming the baseline model in detecting offensive messages.