BanTH: A Multi-label Hate Speech Detection Dataset for Transliterated Bangla

Dataset Overview

• Hate detection (binary) and target identification (multi-label) dataset on 37.3 k transliterated Bangla samples.
• ~38% hate samples and 7 target classes: Political, Religious, Gender, Personal Offense, Abusive/Violence, Origin, and Body Shaming.
• Sourced from 26 YouTube channels of 3 categories: News & Politics, People & Blogs, and Entertainment.
• Each sample is labeled by 3 annotators and verified by experts.

Data Format

Column Title	Description
text_id	Unique identifier for each text entry.
text_content	The actual transliterated Bangla text.
hate label	Indicates whether the text contains hate speech (Yes/No).
political	Flags political hate content (Yes/No).
religious	Flags religious hate content (Yes/No).
gender	Identifies gender-based hate content (Yes/No).
personal offense	Indicates if the text contains personal attacks (Yes/No).
abusive	Flags text with abusive language or can promote violence (Yes/No).
origin	Hate speech targeting someone's origin, e.g. race, nationality (Yes/No).
body_shaming	Identifies body-shaming content (Yes/No).
BN	Back-transliteration of the text content in Bangla.
EN	Translation of the text content in English.