Multilingual Detection of Cyberbullying on Social Networks Using a Fine-Tuned GPT-3.5 Model

Cyberbullying on social networks has emerged as a global problem with serious consequences on the mental health of victims, mainly children, and adolescents. Although there are AI-based solutions to address this issue, they face limitations such as a lack of multilingual datasets, detecting sarcasm, and detecting idioms. Research presents an innovative approach to effective cyberbullying detection using a fine-tuned GPT-3.5 model. Our main contribution is the creation of an extensive multi-label dataset of approximately 60,000 data in English, and Spanish, spanning diverse dialects. This data set was obtained by combining and processing multiple datasets from reliable sources. In addition, we developed a fine-tuned model based on GPT-3.5, capable of identifying hate speech, and offensive language in textual content on social networks. We conducted a thorough evaluation comparing our model to specialized solutions such as Perspective API, Moderation, Content Safety, Toxic Bert, and Gemini. The results demonstrate that our approach outperforms existing models in metrics such as precision, f1-score, and accuracy, making it the most suitable choice for effective cyberbullying detection. This research lays the groundwork for a future app where users can be alerted to harmful content online.