Abstract
This paper presents a novel hybrid deep learning architecture, the CustomBERTCNNAttentionModel, designed for the automated detection of cyberbullying in social media text. The proposed model integrates the contextual language understanding capabilities of Bidirectional Encoder Representations from Transformers (BERT) with the local feature extraction strengths of Convolutional Neural Networks (CNNs) and the dynamic relevance weighting of multihead self-attention mechanisms. Evaluated on the Kaggle Cyberbullying Dataset, which includes both binary and multiclass labels, the model demonstrates superior performance compared to traditional classifiers and ensemble methods. The architecture effectively handles imbalanced and noisy text data, achieving an accuracy of 0.9853 in binary classification tasks. A comprehensive evaluation using standard metrics and visual analysis through confusion matrices confirms the model's robustness and its capacity to generalize across diverse types of cyberbullying. These results highlight the effectiveness of combining transformer-based embeddings with attention-enhanced convolutional structures for detecting harmful online behavior and contribute to the advancement of intelligent moderation systems.