Bengali cyberbullying detection: A comprehensive dataset for advanced analysis

Area of Science:

• Computational Linguistics
• Social Computing
• Cyberpsychology

Background:

• Cyberbullying is a significant digital concern, amplified by social media.
• Research on cyberbullying is extensive, yet studies on Bengali cyberbullying are limited.
• This paper addresses the scarcity of research on Bengali cyberbullying.

Purpose of the Study:

• To analyze Bengali cyberbullying on social media platforms.
• To identify prevalent themes and features of negative Bengali comments.
• To evaluate and compare the performance of various machine learning models for cyberbullying detection.

Main Methods:

• Analysis of over 70,000 Bengali social media comments.
• Sentiment analysis to classify comments as positive or negative.
• Latent Dirichlet Allocation (LDA) for topic modeling of negative comments.
• Application and comparison of machine learning models: Support Vector Machine, XGBoost, CNN+BiLSTM+GRU, mBERT, XLM-R.
• Integration of BERT embeddings with CNN and Artificial Neural Network (ANN) models.
• Utilizing Local Interpretable Model-agnostic Explanations (LIME) for model interpretability.

Main Results:

• mBERT achieved the highest accuracy at 92%.
• The CNN+BiLSTM+GRU hybrid model reached 91% accuracy.
• Incorporating BERT embeddings into CNN and ANN models improved performance to 93% accuracy.
• LDA successfully extracted features related to age, gender, ethnicity, religion, and miscellaneous categories from negative comments.

Conclusions:

• Advanced deep learning models, particularly mBERT and hybrid architectures, are effective for detecting Bengali cyberbullying.
• BERT embeddings significantly enhance the performance of cyberbullying detection models.
• LIME provides valuable insights into model predictions, increasing transparency and trust in cyberbullying detection systems.
• This research contributes a robust framework for understanding and combating Bengali cyberbullying.