README.md

---
license: cc-by-sa-4.0
task_categories:
- text-classification
- question-answering
- zero-shot-classification
language:
- bn
tags:
- communal
- violence
- dataset
- classification
- bengali
- low-resource
- annotator-disagreement
- multi-label
- bengali
- bangla
- social-media
- anonymyzed
- human-annotation
pretty_name: DANGA
size_categories:
- 10K<n<100K
---
> [!WARNING]
> **Content Warning:** This dataset contains **violent, hateful, and severely offensive language** in Bengali, including communal slurs, dehumanizing rhetoric, threats, and incitement to violence targeting religious, ethnic, and cultural communities. It is intended **solely for research purposes** (hate speech detection, content moderation, NLP). Do not use this dataset to generate, promote, or amplify harmful content.

<!-- > [!NOTE]
> This dataset has been released as part of the **Adaption Competition**, with support from **Adaptive Data by Adaption**, whose initiative motivated us to formalize, restructure, and substantially update this dataset prior to its open release.
 -->

# BanDANGA: A Bangla Dataset on Aggressive Narratives and Group-based Attacks

**দাঙ্গা** (DANGA) is an expert-annotated Bengali dataset of **12,720** social media texts classified for communal and sectarian violence. It captures violence across four identity dimensions: religion, ethnicity, socioculture, and nondenominational. Each dimensions were annotated with up to four expression types: derogation, antipathy, prejudication, and repression. The dataset includes full annotator disagreement metadata with individual votes, resolution strategies, and anonymized annotator pairs. This makes it suitable for multi-label classification, preference learning (DPO), and LLM fine-tuning on low-resource hate speech detection.


## Dataset Summary

## Authors & Attribution

DANGA was developed by **Istiak Shihab** and **Nazia Tasnim** as part of a broader effort to advance resources for the Bengali language.

The work was carried out in collaboration with **Bengali.AI**, a non-profit focused on building and promoting open technologies for Bengali.

---

## Dataset Summary

| | Count |
|---|---|
| Total samples | 12,720 |
| Violent | 4,459 (35.1%) |
| Non-violent | 8,261 (64.9%) |
| Multi-category violent | 146 |
| Annotator pairs | 4 (8 annotators) |
| Disputed samples | 4,963 (39.0%) |
| Expert-resolved disputes | 1,553 |

## Schema

Each record is a JSON object with the following structure:

```json
{
  "text": "হিন্দু মালুরা বাংলাদেশে বসবাস করে ...",
  "violent": true,
  "labels": {
    "religion": ["derogation", "antipathy", "prejudication"],
    "ethnicity": [],
    "socioculture": [],
    "nondenominational": ["prejudication"]
  },
  "annotation": {
    "disputed": true,
    "resolution": "third-party",
    "annotators": ["C", "D"],
    "votes": {
      "annot_1": {
        "religion": ["derogation", "antipathy", "prejudication"],
        "ethnicity": [],
        "socioculture": [],
        "nondenominational": ["derogation", "antipathy", "prejudication"]
      },
      "annot_2": {
        "religion": ["derogation", "antipathy", "prejudication"],
        "ethnicity": [],
        "socioculture": [],
        "nondenominational": []
      }
    }
  }
}
```

### Fields

| Field | Type | Description |
|---|---|---|
| `text` | string | Bengali social media text (YouTube comments) |
| `violent` | bool | Whether the text contains any violence expression |
| `labels` | object | Gold-standard labels across 4 identity dimensions |
| `annotation` | object | Full annotator disagreement metadata |

### Identity Dimensions

| Dimension | Column | Target Communities | Examples |
|---|---|---|---|
| **Religio-communal** | `religion` | Religious identity groups | Muslim, Hindu, Christian, Ahmadia, Shia, Atheist, Baul |
| **Ethno-communal** | `ethnicity` | Ethnic identity groups | Bihari, Rohingya, Chakma, Adibashi |
| **Sociocultural** | `socioculture` | Regional/geographic/cultural identity | Sylheti, Kashmiri, Brahmanbaria, Cultural Baul |
| **Nondenominational** | `nondenominational` | Individual, gender, political targets | Misogyny, homophobia, political entities, government |

### Expression Types (Degree of Violence)

Each identity dimension is annotated with zero or more expression types:

| Expression | Description | Count |
|---|---|---|
| **Derogation** | Communal slurs, incivility, dehumanization, bullying | 2,212 |
| **Prejudication** | False accusation, victim blaming, stereotyping, justifying mistreatment | 2,125 |
| **Antipathy** | Alienation, deportation, stripping rights, internalized hatred | 827 |
| **Repression** | Direct threats, incitement to harm, encouraging violence | 517 |

### Annotation Metadata

Each sample includes full disagreement provenance:

| Field | Values | Description |
|---|---|---|
| `disputed` | `true` / `false` | Whether annotators disagreed |
| `resolution` | `sided_with_X` / `third-party` / `null` | How the dispute was resolved |
| `annotators` | `["X", "Y"]` | Anonymized annotator pair (A–H) |
| `votes.annot_1` | labels object | First annotator's original labels |
| `votes.annot_2` | labels object | Second annotator's original labels |

**Resolution distribution (4,963 disputed samples):**

| Resolution | Count |
|---|---|
| Sided with first annotator | 2,205 |
| Sided with second annotator | 1,205 |
| Third-party expert label | 1,553 |

## Taxonomy

The dataset employs a **4×4 orthogonal taxonomy**:

- **4 Identity dimensions** (WHO is targeted): Religio-communal, Ethno-communal, Sociocultural, Nondenominational
- **4 Expression types** (HOW violence is expressed): Derogation, Antipathy, Prejudication, Repression
- Posts can have **multiple identity categories** and **multiple expression types** simultaneously (multi-label)

This produces a theoretical space of 16 fine-grained violence classes.

## Source

| Metric | Value |
|---|---|
| Language | Bengali (বাংলা) |
| Source | YouTube, Facebook, Newspaper comments |

## Intended Uses

- **Violence detection** in Bengali social media
- **Multi-label classification** research
- **Annotator disagreement modeling** and calibration
- **LLM fine-tuning** for hate speech and communal violence tasks
- **Preference learning (DPO/RLHF)** using annotator votes as chosen/rejected pairs
- **Cross-lingual transfer** for low-resource language hate speech detection

## License

This dataset is released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/). It is intended for research purposes only.

## Anonymization Policy

To protect the privacy of individuals whose content appears in this dataset, the following anonymization measures were applied:

- **Annotator identities** are fully anonymized. All annotators are referred to only by a randomly assigned letter (A–H). No names, institutional affiliations, or demographic information about annotators are disclosed.
- **Author/poster identities** from source platforms (YouTube, Facebook, newspaper comment sections) are not included in the dataset. Usernames and profile references have been removed or replaced.
- **Personal mentions** within text (e.g., tagged usernames, phone numbers, identifiable personal details) were removed or masked where detected during preprocessing.
- The raw source URLs or post IDs that could be used to re-identify individuals are not released as part of this dataset.

Researchers who discover re-identification risks are encouraged to contact the authors.

## Ethical Considerations

This content is preserved for research purposes, specifically to build systems that can detect and mitigate such violence. The following guidelines apply:

- The dataset **must not** be used to generate, promote, or amplify hate speech or communal violence
- The dataset is intended for **research use only** (NLP, content moderation, computational social science)
- All annotators have been **fully anonymized** (see Anonymization Policy above)
- The data was collected from **publicly available** social media comments; personal identifiers have been removed
- Users of this dataset are expected to adhere to responsible AI and research ethics guidelines