Emotion-Driven Cyberbullying Detection: A Hybrid Deep Learning Approach Integrating CNN-LSTM and BERT for Enhanced Accuracy
Ajanthaa Lakkshmanan, Rishikesh Reddy Kandadi, Satya Anila Manda · 2024
This study presents a novel hybrid deep learning framework for detecting cyberbullying on social media platforms by incorporating both offensive language and emotional sentiment analysis. The proposed model integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) for feature extraction and sequence modeling, along with Bidirectional Encoder Representations from Transformers (BERT) for contextual understanding. By analyzing both explicit harmful content and the underlying emotional triggers, the model effectively identifies subtle forms of cyberbullying that traditional methods often miss. Trained on a multilingual dataset, the framework leverages stacked word embeddings (GloVe + FastText) and achieves high accuracy in detecting emotionally charged cyberbullying. The experimental results demonstrate superior performance with an F1-score of 0.97, outperforming existing models. This approach highlights the importance of addressing emotional sentiment in cyberbullying detection, providing a more comprehensive and effective solution for online safety.