Conference proceeding
BRAC: A Transformer-Based Model for Detecting Religion-Oriented Aggression in Bengali Social Media Comments
Computer and information technology, pp 311-316
19 Dec 2025
Abstract
Religious aggression on social media poses a critical threat to digital peace and social stability. In Bangladesh, where religion is deeply interwoven with culture and society, social platforms often amplify intolerant and violent narratives that escalate into offline conflict. Detecting such aggression is therefore both a computational and social imperative. While significant advances have been made in English hate speech detection, Bangla-spoken by over 230 million people-remains under-resourced. The lack of annotated corpora, linguistic complexity, dialectal variation, and frequent code-mixing continue to hinder Bangla NLP research. To address this gap, we present the Bangla Religious Aggression Comments (BRAC), comprising 20,000 annotated social media comments labeled for both aggression and targeted religion. We further provide benchmark evaluations across machine learning, deep learning, and Transformerbased models. Random Forest achieved 90.8% accuracy for aggression detection, while SVM attained 80.18 % for target religion classification. CNN models improved performance, reaching 94.5 % for aggression and 88.2 % for target religion. The best results were obtained with BanglaBERT, achieving 95.7 % accuracy for aggression detection and 90.4 % for target religion classification. You can access the dataset from https://github.com/Riad071/BRAC-Bengali-Religious-Aggressive-Comments.
Metrics
1 Record Views
Details
- Title
- BRAC: A Transformer-Based Model for Detecting Religion-Oriented Aggression in Bengali Social Media Comments
- Creators
- Riad Hossain - Chittagong University of Engineering & TechnologyKobra Elahi - East Delta UniversitySharjina Afrin - East Delta UniversityAyesha Banu - Chittagong University of Engineering & TechnologyMohammad Sayed Talukder - Drexel University
- Publication Details
- Computer and information technology, pp 311-316
- Publisher
- IEEE
- Number of pages
- 6
- Resource Type
- Conference proceeding
- Language
- English
- Academic Unit
- Computer Science
- Scopus ID
- 2-s2.0-105041691354
- Other Identifier
- 991022198378904721