Challenging Linguistic and Cultural Diversity: Evaluation of AI Models in the Detection of Hate Speech in Brazilian Social Networks

In Brazil, some cases of hate speech can be qualified as a crime. However, identifying and categorizing offensive comments among the vast number of interactions on social media is complex. Automatic detection of sensitive content is an expanding field, but it faces obstacles due to the subtleties of language and varied forms of expression. Brazil's rich cultural diversity, shaped by its experiences, culture, traditions, and history of colonization, introduces additional challenges. This linguistic and cultural diversity plays a crucial role in the ethical evaluation of models, raising questions about whether a model, even if trained specifically for the Portuguese of Brazil and contextualized to the country, can capture the wide variety of linguistic and cultural variations present. This study aims to investigate the effectiveness of artificial intelligence models in classifying hate speech in a ternary setting, focusing on their ability to generalize to new data and thus approach real conditions. Seven fine-tuning classifiers based on the BERT and BART models were used.