Ben-Misandry-5000: An Annotated Bengali Misandry Dataset for Six-Class Gender-Based Hate Speech Classification

Description

5,000 Bengali comments annotated for anti-male hate speech detection. Unlike existing Bengali gender-hate corpora (Ben-Misog, BOISHOMMO, BanTH, etc.), which label only anti-woman content, this is the first Bengali misandry-labelled corpus. Built via a taxonomy-driven synthetic pipeline using Azure OpenAI GPT-4.1 with few-shot grounding, normalized to pure Bengali script, deduplicated, and quality-checked. Mirrors the peer-reviewed Ben-Misog methodology (Roy et al., IEEE STI 2024) for cross-corpus experiments. Overview - 5,000 comments, six-class misandry taxonomy - Source: Azure GPT-4.1 synthetic generation - Two label layers: 6-class + derivable binary - Novel class: mockery_emotional_dismissal - 100% Bengali script, 0 duplicates, imbalance ratio 1.002 Taxonomy (6 Classes) - discredit — belittles men's competence/worth - stereotype_objectification — reduces men to provider/emotionless tropes - mockery_emotional_dismissal — mocks male vulnerability (novel) - threats_violence — threatens harm against men - dominance_derailing — dismisses men's issues/female-supremacist framing - non_misandric — neutral/control class Data Fields - text — Bengali comment, 21–117 chars, mean 10.2 words - class — one of six taxonomy labels Dataset Class Distribution : Discredit: 833 samples (16.66%) Stereotype & Objectification: 833 samples (16.66%) Mockery & Emotional Dismissal: 833 samples (16.66%) Threats & Violence: 833 samples (16.66%) Dominance & Derailing: 833 samples (16.66%) Non-Misandric: 835 samples (16.70%) Total Dataset Size: 5,000 samples (100%) Files Provided ben_misandry.csv — 5,000 rows × 2 columns (text, class), UTF-8, no BOM. Ready for classical ML and transformer classifiers (BanglaBERT, XLM-RoBERTa, mBERT, MuRIL, IndicBERT).

Institutions

• North South University

  Dhaka Division, Dhaka

Categories

Licence