BanglaCAT: A Context-Aware Bangla Social Media Toxicity Dataset with Hierarchical Annotation

Authors:
Preprint
Publisher website
Google Scholar
Abstract
Toxic and abusive language on social media is an increasing concern, particularly for low-resource languages such as Bangla, where reliable resources for context-aware and fine-grained toxicity detection remain limited. This data article presents BanglaCAT, a context-aware Bangla social media toxicity dataset that records not only whether a reply is toxic but also the type of toxicity it expresses. Unlike datasets that label comments in isolation, each BanglaCAT record holds a target reply together with the conversational context it appeared under, so that short, ambiguous, sarcastic or context-dependent replies are labeled as they were actually read.The dataset contains 19,453 training context–reply pairs and 2,329 real-world test pairs. The training data were generated with Claude Sonnet 5 and then checked by three human reviewers working independently, each of whom could accept the suggested label or replace it. The final label was decided by majority vote. The test data were collected separately from public Bangla discussions on Facebook, YouTube, and news portals. Each test item was labeled independently by three human annotators. Items that did not receive a 2-of-3 majority on the joint Level-1/Level-2 label were passed to a second, independent group of three annotators, whose own majority decided the label; the two items still unresolved after that round were settled from the pooled votes of both groups.BanglaCAT follows a hierarchical annotation scheme. At Level-1, each target reply is classified as Toxic or Non-toxic. Toxic replies are further assigned one of 11 Level-2 categories: Insult, Threat, Hate Speech, Profanity, Harassment, Identity Attack, Toxic Endorsement / Agreement, Incitement / Encouragement of Harm, Mockery / Ridicule, Sexual Harassment / Objectification, and Dehumanization. The structure supports both binary toxicity detection and finer-grained classification of harmful language in context. To our knowledge, BanglaCAT is the first Bangla resource we have been able to identify that combines paired conversational context, a binary toxicity level, 11 mutually exclusive toxicity subtypes, and a separately collected human-annotated real-world evaluation partition. A Bangla version and a row-aligned English version are released for work on context-aware toxicity detection, hierarchical classification, content moderation, and cross-lingual NLP.