Transformer-Based Approach to Cyberbullying Detection Using Fine-Tuned DistilBERT
Аннотация
This study offers a deep learning framework for classifying cyberbullying in order to ensure fairness and robustness, a balanced dataset of 20,000 labeled tweets was used, divided into four categories: religion, gender/sexual, ethnicity/race, and non-cyberbullying. Preprocessing included thorough text cleaning, which included deleting special characters, links, and numbers. The AdamW optimizer was implemented to train the model with a batch size of 256 over 15 epochs at a learning rate of 5e-5. Additional layers of batch, dropout, and dense normalization that decreases the overfitting and enhanced generalization. The experimental results showed a loss of 0.07 and a sparse categorical accuracy of 99.38%, which was a significant improvement over current technique. The suggested method demonstrates how effective transformer-based models are at detecting multi-class cyberbullying.