Detecção de discurso de ódio: análise de modelos clássicos e redes neurais com estratégias de balanceamento

Abstract

With the growth of internet access and, consequently, the expansion of social media, new opportunities for communication, interaction, and information exchange have emerged on a global scale. However, this virtual environment has also enabled the spread of negative content, such as hate speech. Hate speech can be understood as any expression that promotes violence or discrimination against individuals or groups based on characteristics such as race, religion, gender, sexual orientation, among others. In this context, the present study focused on evaluating machine learning models by comparing traditional algorithms and neural networks applied to textual data extracted from social media containing hate speech. Balancing techniques such as class weight, oversampling, and undersampling were employed as a way to handle class imbalance. The results showed that traditional classifiers demonstrated greater consistency in the weighted F-Score and AUC metrics, even in the original scenario without balancing treatment, reaching values above 0.93. On the other hand, neural network-based models, such as MLP, CNN, and LSTM, proved to be more sensitive to the type of balancing, suggesting the need for adjustments to improve their generalization capacity.