Detecção de Discurso de Ódio em Comentários Relacionados à Política

Abstract

In an era where people are increasingly connected, the spread of hate speech on social networks has become more frequent. Consequently, computational technology has emerged as a valuable tool to identify and mitigate hate speech on these platforms. Given the available computational power, we used Natural Language Processing to detect hate speech in texts from social networks. Besides addressing detection, another goal was to investigate the impact of lexical distance between the languages of the corpora used in model training, exploring encoders and decoders based on Transformer architecture. Therefore, we investigated the inclusion of Cross-Lingual Learning (CLL) to enhance hate speech detection in different languages, employing various CLL techniques and the application of multiple languages as training sources for the model. The results revealed that applying CLL, especially with multiple source languages, significantly improved the effectiveness of the models in classifying hate speech. Moreover, encoder-based models were more efficient when the lexical distance between languages was closer, achieving 96.92% in the F1-score metric. In contrast, decoder models were more efficient when the lexical distance between languages was farther, achieving 96.58% in the F1-score metric. Thus, this work highlights that linguistic diversity and the lexical distance used in Transformer-based models are crucial for developing effective systems to detect hate speech. Finally, the findings of this research reinforce the feasibility of using CLL and multiple languages to improve hate speech detection, offering new directions and insights for future research in this area.