Identificação de discurso de ódio nas redes sociais por meio de redes neurais artificiais LSTM e RoBERTa

Abstract

This work addresses the detection of hate speech on social networks using RNA, focusing on the comparison between the LSTM architectures and the BERT variant, RoBERTa. The alarming growth of speeches on these platforms makes it essential to use Artificial Intelligence (AI) for the automatic identification of this type of content. The study seeks to analyze the performance of these architectures in the classification of hate speech, evaluating their effecti veness metrics and comparing the results with other models that used the same database. For this, textual data were collected and preprocessed, followed by training and tuning the models, testing different configurations of batch size and number of epochs. The results indicated that the RoBERTa model outperformed LSTM, presenting greater accuracy in the classification of content. However, a multilingual BERT model demonstrated even better performance than RoBERTa, possibly due to its ability to handle linguistic variations and generalize across different contexts. Despite the effectiveness of transformers, high computational cost and training time have been challenges for their large-scale implementation. We conclude that IA, especially transformer-based models, represent a promising tool in combating online hate speech. However, optimizations are straightforward to balance performance and computational efficiency, making these models more viable for practical applications.