Advancing cyberbullying detection in low-resource languages: a transformer- stacking framework for Bengali

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Computing

Background:

• Cyberbullying is a global issue, with limited research in low-resource languages like Bengali due to data and methodology scarcity.
• Existing methods often lack language-specific preprocessing and advanced transformer models, hindering effective cyberbullying detection.

Purpose of the Study:

• To develop and evaluate an effective hybrid framework for Bengali cyberbullying detection.
• To address limitations in current approaches by incorporating Bengali-specific preprocessing and advanced transformer models.

Main Methods:

• Introduced three Bengali-specific preprocessing strategies to improve feature representation.
• Proposed Transformer-stacking, a hybrid framework combining XLM-R-base, multilingual BERT, and Bangla-Bert-Base with a multi-layer perceptron classifier.
• Evaluated the framework on a 44,001-sample Bengali cyberbullying dataset for binary and multiclass classification.

Main Results:

• Transformer-stacking achieved 93.61% F1-score and 93.62% accuracy for binary classification (Sub-task A).
• Achieved 89.23% F1-score and accuracy for multiclass classification (Sub-task B).
• Outperformed baseline models, ensemble techniques, and state-of-the-art methods, with results statistically validated by McNemar's test.

Conclusions:

• Transformer-stacking provides an effective and generalizable solution for Bengali cyberbullying detection.
• The study establishes a new benchmark for cyberbullying detection in underexplored low-resource languages.
• Demonstrated model scalability and adaptability through experiments on external datasets for hate speech and abusive language detection.