BANT: Bangla Arrogance and Non-Arrogant Toxicity Dataset
Description
BANT (Bangla Arrogance and Non-Arrogant Toxicity Dataset) is a human-annotated Bangla social-media dataset developed for arrogance detection, non-arrogant toxicity analysis, fine-grained arrogance classification, and multilingual NLP research. Version 1 contains 22,427 comments collected from publicly accessible YouTube, Facebook, and news-portal content, together with a separately identified subset of 918 synthetic Bangla comments. The dataset was independently annotated by three human annotators using three original labels: Arrogant, Non-Arrogant-Toxic, and Non-Arrogant. Final labels were derived by majority voting. Comments labelled Arrogant additionally include one of five fine-grained categories: Superiority & Self-Importance, Knowledge & Moral Superiority, Contempt & Condescension, Dominance & Entitlement, and Mixed / Other Arrogance. This release supports both a binary task (Arrogant vs. Non-arrogant) and a three-class task (Arrogant, Non-Arrogant-Toxic, Non-Arrogant). It includes fixed train, validation, and test splits of 17,941, 2,243, and 2,243 instances, respectively. Bengali, machine-translated English, and aligned bilingual versions are provided. The package also includes annotation guidelines, data dictionaries, translation metadata and QA information, release manifests, SHA-256 checksums, and reproducibility scripts. No AI-generated prediction was used as an annotator or consensus signal. The dataset is released under the CC BY 4.0 license.
Files
Steps to reproduce
The BANT dataset was constructed through a multi-stage collection, annotation, curation, and validation pipeline. 1. Bangla textual content was collected intermittently between January 2024 and March 2025 from public YouTube, Facebook, and news-portal sources. Public comments were collected from manually selected posts using Python and Selenium WebDriver 4.28.1 with a headless Chromium browser. A separate subset of 918 synthetic Bangla comments was generated using ChatGPT and retained as an AI provenance category. 2. Collected text was cleaned and normalized using Unicode NFKC normalization, removal of zero-width characters, whitespace normalization, duplicate screening, and removal of irrelevant interface text. Direct identifiers matching URLs, email addresses, user handles, and Bangladeshi phone-number patterns were replaced with generic placeholders. 3. Exactly three human annotators independently reviewed separate copies of the annotation material. The original label space consisted of Arrogant, Non-Arrogant-Toxic, and Non-Arrogant. For Arrogant instances, annotators additionally selected one of five fine-grained arrogance categories. 4. Final three-class labels were obtained by row-wise majority voting across the three submitted human annotations. The resulting distribution is 6,126 Arrogant, 9,183 Non-Arrogant-Toxic, and 7,118 Non-Arrogant instances. 5. A binary task view was derived by retaining Arrogant as Arrogant and mapping both Non-Arrogant-Toxic and Non-Arrogant to Non-arrogant. This produces 6,126 Arrogant and 16,301 Non-arrogant instances. 6. The final 22,427 instances were divided into fixed train, validation, and test partitions containing 17,941, 2,243, and 2,243 rows, respectively. The same row membership is preserved across the binary and three-class releases. 7. An English companion was generated from the finalized Bengali text using facebook/nllb-200-distilled-600M with ben_Beng as the source language and eng_Latn as the target language. Deterministic beam decoding was used with num_beams=2 and do_sample=False. Labels, source metadata, and categories were copied unchanged. 8. Release-level quality checks verify row counts, label distributions, absence of exact and normalized-text duplicates, split disjointness and completeness, alignment between Bengali, English, and bilingual files, and consistency between the binary and three-class task views. The deposited package includes the fixed datasets, data dictionaries, annotation guideline, release manifest, SHA-256 checksums, translation metadata, and reproducibility scripts. The supplied validation script can be executed from the package root using: python -m pip install -r reproducibility/requirements.txt python reproducibility/validate_release.py --root . The historical online source content may change or be removed over time; therefore, the deposited Version 1 files should be treated as the stable research snapshot for reproduction and benchmarking.
Institutions
- Chittagong University of Engineering & TechnologyChittagong, Chittagong