BADD: A Large-Scale Bengali Dataset for Arrogance Detection

Abstract

BADD is a human-annotated Bangla social-media dataset for arrogance detection and fine-grained arrogance analysis. The current release contains 22,427 comments based on three independent human annotation files. No AI system or model-derived signal was used as an annotator or for human consensus. The original annotation scheme includes three labels: Arrogant, Non-Arrogant-Toxic, and Non-Arrogant. For the public release, Non-Arrogant-Toxic and Non-Arrogant are mapped to Non-arrogant, resulting in 6,126 Arrogant and 16,301 Non-arrogant instances. Annotation rules are documented in `ANNOTATION_GUIDELINE.md`. Arrogant comments are further categorized into five types: Superiority & Self-Importance, Knowledge & Moral Superiority, Contempt & Condescension, Dominance & Entitlement, and Mixed / Other Arrogance. The dataset includes four source/provenance categories: YouTube, Facebook, News portal, and AI. Here, “AI” denotes a data-source category only and does not indicate AI-based annotation. The canonical Bengali dataset is provided in `BADD_final_dataset_bengali.csv`. Reproducible splits are available as `BADD_train_bengali.csv` (17,941), `BADD_validation_bengali.csv` (2,243), and `BADD_test_bengali.csv` (2,243). An English machine-translated companion is provided in `BADD_final_dataset_english.csv`, and an aligned Bengali-English version in `BADD_final_dataset_bilingual.csv`. Translation was performed using `facebook/nllb-200-distilled-600M` from `ben_Beng` to `eng_Latn` with deterministic beam search (`num_beams=2`, `do_sample=False`). Only comment text is translated; source, label, and category remain unchanged. Column definitions are provided in `DATA_DICTIONARY.csv`. Translation settings are recorded in `TRANSLATION_METADATA.json`, while `TRANSLATION_QA_SUMMARY.md` summarizes quality screening. The current QA procedure flags 197 rows for manual inspection; these are screening indicators, not confirmed translation errors. `README.md` provides the package overview and dataset statistics, while `SHA256SUMS.txt` supports file-integrity verification. The Bengali and English files use `comment`, `source`, `arrogance_label`, and `category`; the bilingual file uses `comment_bn`, `comment_en`, `source`, `arrogance_label`, and `category`. The Bengali dataset is the canonical research version, while the English dataset is a machine-translated companion. BADD supports research in Bangla NLP, text classification, computational social science, toxicity-aware language analysis, and arrogance detection.