Turkish Cyberbullying Detection with Fine-Tuned Pre-Trained Language Models

Abstract

With the rapid increase in internet usage and its pervasive presence in all aspects of life, social media platforms have seen a rise in negative behaviors alongside their positive contributions. One such negative behavior is cyberbullying, which refers to the misuse of information and communication technologies to harm others. Cyberbullying is becoming a significant social problem. This study aims to detect and classify Turkish sentences containing cyberbullying using deep learning models. To achieve this, the BERT model, known for its ability to understand the context of language, was chosen. Specifically, the BERTurk, DistilBERTurk, and ConvBERTurk models—designed for the Turkish language—were fine-tuned and retrained using a dataset of 3,388 tweets labeled as racist, sexist, offensive language, or neutral. The primary goal of this study is to perform a comprehensive comparison of multi-class Turkish cyberbullying detection models and to develop an Artifical Intelligence (AI) model that delivers highly accurate results on real-world data. According to the results, BERTurk achieved the highest F1 score of 0.88, while the DistilBERTurk model showed the lowest performance.