 Social media has revolutionized the way people communicate, allowing for instantaneous global communication. However, this has led to the spread of toxic messages, leaving victims feeling powerless and vulnerable. To combat this, we have developed an automated detector of toxic messages on social media. Our approach combines traditional machine learning algorithms with deep learning methods, including a transformer architecture and topic modeling techniques. We found that all models performed well, achieving high accuracy scores. However, the best model, Transformer BR Tweet, only achieved a score of 91.4%. While this is a significant improvement over other models, it does not justify the additional computational costs associated with using a transformer model. This article was authored by Andrea Bonetti, Marcelino Martinez-Sober, Julio C. Torres and others. We are article.tv. Links in the description below.