A Hierarchical Two Stage BERT Model for Cyberbullying Detection

Abstract

Cyberbullying has emerged as a persistent social issue in tandem with the rapid proliferation of social media usage, particularly in developing nations such as Indonesia. Online bullying extends beyond explicit insults to encompass more nuanced forms, including denigration and social exclusion, which are often subtle and context-dependent. This diversity presents significant challenges for automated detection systems, especially when cyberbullying is treated as a single, homogeneous category. To address this challenge, this study proposes a two-stage classification framework for detecting cyberbullying in the Indonesian language using the IndoBERT model. In the first stage, a binary classification model is employed to differentiate between bullying and non-bullying content. In the second stage, cyberbullying tweets are further categorized into three specific types: harassment, denigration, and exclusion. The model is trained using a curated dataset of Indonesian tweets collected from publicly available social media content. The results indicate that the proposed framework achieves strong and consistent performance, particularly in fine-grained cyberbullying classification, with an overall accuracy of approximately 92% in the second stage. These findings suggest that hierarchical classification can yield robust performance and effectively aid in the detection of cyberbullying.