AraBERT-based Approach to Arabic Cyberbullying Detection in Facebook comments رسالة ماجستير

Abstract

Cyberbullying is one of the significant issues with communication platforms like Facebook. It is especially alarming that, in contrast to traditional bullying, it can have serious emotional consequences and follow victims all the time, thus concern regarding cyberbullying is growing on websites like Facebook. Pre-trained language models have achieved significant success in a variety of natural language processing tasks. Although research has been conducted on optimizing BERT-based models for detecting cyberbullying and creation of various pre-trained Arabic models, limited attention has been paid to Arabic cyberbullying detection and low resource available in Arabic language. This thesis aims to investigate the effectiveness of using AraBERT, a pre-trained Arabic language model, in detecting Arabic cyberbullying comments. We create a balanced dataset out of 20,000 Facebook comments in Arabic language and manually label it as either bullying or non-bullying. We employed fine-tuning techniques to adapt AraBERTv2 to the cyberbullying detection task. Through experimentation with freezing layers’ technique and unfreezing different layers of the model, we explored the trade-off between leveraging pre-trained knowledge and adapting the model to the specific task. Our findings demonstrate that the tuning of all layers in AraBERTv2 achieved the highest performance, resulting in an accuracy of 91.9% and an F1 score of 93.0%