Detecção de discurso tóxico em redes sociais brasileiras: um estudo comparativo entre modelos clássicos de aprendizado de máquina e modelos de linguagem pré-treinados

Abstract

The dissemination of toxic comments on social media platforms represents a significant challenge for digital content moderation, particularly in Brazilian Portuguese, a language characterized by the expressive use of profanity and slang that creates ambiguity in the automatic distinction between genuine offense and communicative informality. This paper presents a comparative study of four automatic toxicity classification approaches, including TF-IDF with Logistic Regression, TF-IDF with Support Vector Machine (SVM), BERTimbau and XLM-RoBERTa, the latter two fine-tuned on the ToLD-BR dataset, comprising 21,000 Twitter comments annotated for toxicity. The models were evaluated using Accuracy, Macro F1-score, Precision and Recall. Results showed that Transformer models outperformed classical methods across all evaluated metrics, with BERTimbau achieving Macro F1-score of 0.77 and Accuracy of 0.77, representing a gain of 3.37 percentage points over the best baseline. BERTimbau, a monolingual model trained exclusively on Brazilian Portuguese, outperformed the multilingual XLM-RoBERTa, suggesting that language specialization is a relevant factor for this task. Qualitative error analysis revealed that the main challenge lies in the ambiguity between offensive and expressive colloquial language use in Brazilian Portuguese. These findings reinforce the importance of language-specific models for automated content moderation.