Indonesian Cyberbullying Detection Using IndoBERTweet-BiGRU Model on Class-Imbalanced X (Twitter) Data

Abstract

Cyberbullying on social media platforms, particularly X (formerly Twitter), has become a serious issue that negatively affects users' mental health and well-being. Automatic cyberbullying detection in Indonesian remains challenging due to the widespread use of informal language, slang, abbreviations, and highly imbalanced class distributions. This study proposes a hybrid deep learning model that integrates IndoBERTweet with a Bidirectional Gated Recurrent Unit (BiGRU) to improve cyberbullying detection performance on Indonesian tweets. A dataset of Indonesian tweets was collected from X and annotated using a multi-stage dual large language model (LLM) labeling strategy to reduce the time and effort required for manual annotation while maintaining label consistency. To address class imbalance, this study investigates the effectiveness of Focal Loss and label distribution modification through multiple experimental scenarios. The proposed approach was evaluated using accuracy, precision, recall, and F1-score. The best performance was achieved by combining Focal Loss with a modified four-class label configuration consisting of Rude and Vulgar Words, Sexual Harassment, Body Shaming and Hate Speech, and Non-Cyberbullying. This configuration obtained an accuracy of 0.93, precision of 0.90, recall of 0.90, and F1-score of 0.90. These findings demonstrate that integrating contextual language representations with sequential modeling, supported by an efficient LLM-assisted labeling strategy and class imbalance handling, provides an effective approach for Indonesian cyberbullying detection and offers a practical solution for large-scale social media content moderation.