Enhancing Cyberbullying Detection in Arabic Text through Ensemble Stacking Models
Abstract
Cyberbullying has become a major social concern in this modern era of digital communications. Cyberbullying can have detrimental effects on the individuals involved ranging from psychological to pathological. Hence, detecting any act of cyberbullying in an automated manner will help to prevent any unfortunate results. In this regard, data-driven approaches, such as Machine Learning (ML), particularly Deep Learning (DL), have shown promising results. DL approaches provide highly accurate predictive models for text classification. However, literature shows that ML approaches, particularly DL, have not been extensively studied for Arabic text classification of cyberbullying. The prevalence of non-diacritical writing, dialectal variability, and morphological complexity presents challenges in developing high-accuracy text classification systems for Arabic text. Subsequently, the application of DL to cyberbullying detection problems within Arabic text classification can be considered a novel approach due to the complexity of the problem and the tedious process involved, besides the scarcity of relevant research studies. Thus, this research aims to develop a highly advanced DL model that can automatically detect cyberbullying. We evaluated seven deep learning (DL) models for Arabic cyberbullying classification: Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Bidirectional Gated Recurrent Unit (Bi-GRU), hybrid Bi-LSTM-LSTM, CNN-Bi-GRU, CNN-Bi-LSTM, and Bi-LSTM-Bi-GRU. Subsequently, an ensemble stacking model was implemented, integrating the top-performing DL models in terms of accuracy. The stacking models were designed to optimize predictive accuracy by synergistically combining the strengths of the individual models. The ensemble stacking model consisted of DL models with a meta-learner layer of classifiers. In this research, the first model combined the two best-performing DL models: Bi-LSTM and Bi-LSTM-Bi-GRU. The second model combined the four best-performing models: CNN, Bi-LSTM, Bi-GRU, and Bi-LSTM-Bi-GRU. The final model combined all seven trained DL models. Our results indicate that the stacking DL model with the meta-learner layer of the Random Forest (RF) classifier achieved the highest accuracy of 94.73%, outperforming other models.