THIN-Hate: A Multi-Domain Corpus of Tamil and Tanglish Hate, Implicit and explicit in Nature

THIN (Tamil and Tanglish Hate, Implicit and explicit in Nature) is a corpus of 17,498 YouTube comments across five domains (Caste, Political, Religious, Women, Migration), written in Tamil script, Romanized Tamil (Tanglish), Tamil-English code-mixed text and English. Each comment is annotated on four levels: hate class (Hate, Hate with Offensive, Offensive, Non-Hate, Cannot decide), explicit versus implicit form, target, and language variety. The corpus keeps Hate and Hate with Offensive as separate classes from purely Offensive language and labels both explicit and implicit hate (1,726 implicit comments in the full corpus). Gold labels are the 2-of-3 annotator majority or an adjudicated decision. Splits are grouped by video and stratified by domain and class. This record contains the train (12,343 comments) and dev (1,676 comments) splits with gold labels. The test split (3,479 comments) is withheld for a future shared task and will be released after it ends. Author handles and original IDs are removed, and mentions, URLs and phone numbers are masked. See README.md for the full description. Content warning: the data contains hateful, abusive and offensive language, quoted for research on detecting it.