Identificação e caracterização de conteúdo tóxico de usuários em comunidades brasileiras no Reddit

Abstract

The proliferation of online social interactions in recent years, with the consequent growth in user-generated content, has brought the escalating issue of toxic language. While automatic machine learning models have been effective in moderating the vast amount of data on online social networks, low-resource languages, such as Brazilian Portuguese, still lack efficient automated moderation tools. We address this gap by creating a novel dataset collected from some of the most popular Brazilian Reddit communities. Using manually labeled data, we propose the use of both open and closed machine learning models trained for the task of toxicity classification. We also explore the use of Large Language Models (LLMs) to assist in generating high-quality synthetic data. Our main findings show that toxic comments in Brazilian communities exhibit distinct linguistic patterns, and that pre-trained language models (such as LLMs and Transformers) are essential for scalable and effective automated toxicity detection. These results highlight the importance of leveraging external knowledge learned by pre-trained models for application in specific tasks. With this research, we aim to contribute to the critical task of automated moderation on online social media, promoting a safer and more inclusive online environment for all.