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

Description

BanglaCAT is a context-aware toxicity dataset for Bangla (Bengali) social-media text. Each instance is a conversational pair consisting of a preceding comment (context) and a target comment whose toxicity is judged in light of that context. Unlike most existing Bangla resources that label isolated comments, this dataset is designed so that conversational context can both amplify and mitigate perceived toxicity (e.g., sarcasm, veiled threats, toxic endorsement, consensual banter, quoted or reported speech). Partitions: Train (19,453 pairs): LLM-generated candidates, cross-validated by independent LLMs, and human spot-checked. Released in original Bangla and English. Test (2,329 pairs): Real comments collected from public Facebook, YouTube, and Bangladeshi news-portal threads. Triple-annotated by human annotators; gold labels obtained by exact 2-of-3 joint hierarchical majority vote. Released in Bangla and English (row-aligned). Label schema (hierarchical): Level-1: Toxic / Non-toxic Level-2 (toxicity_type): None (when Non-toxic), or one of 11 fine-grained types when Toxic — Insult, Threat, Hate Speech, Profanity, Harassment, Identity Attack, Toxic Endorsement / Agreement, Incitement / Encouragement of Harm, Mockery / Ridicule, Sexual Harassment / Objectification, Dehumanization. The gold label is always the complete pair (level_1_class, toxicity_type). Level-1 and Level-2 are never voted independently, preventing invalid combinations. ## Dataset structure ``` BanglaCAT_Mendeley/ ├── train/ │ ├── bangla_context_toxicity_train.csv # Bangla (19,453 rows) │ └── bangla_context_toxicity_train_ENGLISH.csv # English translation, row-aligned ├── test/ │ ├── bangla_toxicity_test_set_HIERARCHICAL_FINAL.csv # Bangla (2,329 rows) │ └── bangla_toxicity_test_set_HIERARCHICAL_FINAL_ENGLISH.csv # Bangla + English columns ├── docs/ │ └── DATA_DICTIONARY.md ├── CITATION.cff ├── LICENSE.md ├── CHECKSUMS.txt └── README.md Train and test are strictly disjoint. Intended use is research on content moderation, hate-speech detection, and conversational AI safety.

Steps to reproduce

1. Download and extract the dataset archive from this Mendeley Data deposit (DOI: 10.17632/8vr3w68c7x.1). 2. Load the training and test CSVs with UTF-8 encoding (recommended: `pandas.read_csv(..., encoding="utf-8-sig")`). 3. Use the hierarchical label pair `(level_1_class, toxicity_type)` as the gold target. - Level-1: `Toxic` / `Non-toxic` - Level-2: `None` when Non-toxic; one of 11 fine-grained types when Toxic (Insult, Threat, Hate Speech, Profanity, Harassment, Identity Attack, Toxic Endorsement / Agreement, Incitement / Encouragement of Harm, Mockery / Ridicule, Sexual Harassment / Objectification, Dehumanization). 4. For context-aware models, feed both the preceding turn (`previous_reply` / `context`) and the target turn (`target_reply` / `reply`) as input. For context-free baselines, use only the target turn. 5. Keep the train (19,453 pairs) and test (2,329 pairs) partitions strictly separate. Do not mix rows across partitions. 6. Optional: use the English versions of the same files for non-Bangla-speaking experiments; they are row-aligned with the Bangla originals. 7. Evaluate Level-1 (binary) and/or the full 12-class joint hierarchical label, depending on the experimental setup. Full column definitions and taxonomy are given in `README.md` and `docs/DATA_DICTIONARY.md` inside the archive.

Institutions

• East Delta University

  Chittagong, Chittagong
• Chittagong University of Engineering & Technology

  Chittagong, Chittagong

Categories

Licence