Bangla Racism and Body-Shaming Detection Dataset
Description
This dataset contains 5,000 manually annotated Bangla-language text instances collected and curated for the detection of racism and body-shaming content on Bangla social media. It accompanies the paper "An Optimized Multi-Transformer Framework with Explainable AI for Racism and Body-Shaming Detection in Bangla Social Media." Files Bangla_Racism_BodyShaming_Dataset.csv — the full labeled dataset (UTF-8 encoded) Format CSV, UTF-8 (with BOM for Excel compatibility) Fields Column Description Sentence The target Bangla sentence being classified Sentiment Label: "Body Shaming" or "Racism" Story with Violence A short contextual narrative embedding the sentence, framed with a violent/harsh undertone Story without Violence A short contextual narrative embedding the same sentence, framed neutrally Class distribution Body Shaming: 2,573 samples Racism: 2,427 samples Total: 5,000 samples Language Bangla (Bengali script) Intended use Training and evaluation of transformer-based models (e.g., BanglaBERT, mBERT, MuRIL, XLM-RoBERTa) for hate speech / abusive language detection in low-resource Bangla NLP, and for explainable AI (XAI) research using techniques such as SHAP. Notes Samples were manually verified prior to inclusion. One fully empty row present in the original raw export was removed during cleaning. Please cite the accompanying paper if you use this dataset.
Files
Steps to reproduce
Collected Bangla-language text from public social media platforms (Facebook, YouTube, Instagram) and comment sections of online news portals, targeting content related to racism and body-shaming. Expanded coverage and lexical diversity using assisted text generation to capture a wider range of expression (explicit, implicit/euphemistic, sarcastic, and context-dependent instances). Manually reviewed, edited, and verified every generated/collected instance by hand before inclusion. Annotated each instance into exactly one of two classes — Racism or Body Shaming — based on the contextual meaning of the full sentence (not on the presence of individual offensive keywords), following predefined annotation guidelines. Removed duplicate, ambiguous, and internally inconsistent records during verification. Compiled the final labeled dataset of 5,000 instances (2,573 Body Shaming, 2,427 Racism) in CSV format, with each row containing the target sentence plus two contextual narrative variants (with/without violence framing).
Institutions
- Bangladesh UniversityDhaka Division, Dhaka