OMANISHA: A benchmark dataset for identifying and categorizing Bengali misogynistic text

Misogynistic discourse in online spaces has profound psychological, social, and institutional consequences for women, contributing to persistent gender inequalities, the normalization of gender-based violence, and the suppression of female voices in digital communities. The unchecked proliferation of such hostile content across social media platforms has made automated misogyny detection a pressing social imperative. Despite Bengali's global prominence, computational resources for detecting gender-based online abuse remain limited. To address this gap, we introduce OMANISHA (Online Misogynistic Annotated Natural-language Instances for Sentiment and Hate Analysis), a Bengali dataset designed to advance misogynistic text detection. The dataset comprises 7,017 annotated samples independently developed by the authors based on linguistic patterns observed across diverse public online platforms. Each sample was assigned to one of four predefined labels: non-misogynistic (2,420), stereotype (1,744), derogation (1,527), and sexual harassment (1,326). Moreover, English translations are provided to enhance cross-lingual accessibility. Every sample was independently labeled by two native Bengali-speaking annotators, chosen from a four-member annotation team representing diverse gender, religious, ethnic, and geographical backgrounds. Annotation disagreements were resolved through structured consultation with a third annotator. Annotation quality was assessed using Cohen's Kappa (κ = 0.76) and Krippendorff's Alpha (α = 0.75), indicating substantial inter-annotator agreement and strong annotation reliability. Pairwise Jaccard Similarity scores among the classes range from 0.12 to 0.21, suggesting taxonomic distinctiveness across the defined categories. The preprocessing stages include duplicate removal, text normalization, and coherence checking to ensure overall data integrity. OMANISHA serves as a publicly accessible dataset for supporting explainable AI-driven content moderation, advancing low-resource NLP research, and promoting women-centric online safety.