---
pretty_name: Urdu Hate Speech
tags:
- urdu
- hate-speech
- toxicity
- text-classification
- safety
- moderation
task_categories:
- text-classification
task_ids:
- hate-speech-detection
language:
- ur
license: mit
annotations_creators:
- expert-generated
language_creators:
- found
multilinguality:
- monolingual
size_categories:
- 10K<n<100K
source_datasets:
- original
---

# Urdu Hate Speech

A binary Urdu text classification dataset for hate-speech detection.

- Task: hate vs. not_hate (2 classes)
- Language: Urdu (ur)
- Primary columns: `text`, `label`  
  (In the uploaded files the columns are `Tweet` and `Tag`; see “How to Use” for renaming.)
- Labels:
  - 0 → not_hate
  - 1 → hate

## Dataset Summary

This dataset contains short-form Urdu text (e.g., social media posts) labeled for hate speech. It supports research, moderation assistance, and demos. It should not be used as the sole basis for punitive actions.

## Data Collection

- Source: Public tweets collected via the Twitter API/search endpoints using Urdu keywords, hate-speech lexicon terms, and topic-specific queries.
- Scope: Short-form posts; code-mixed Urdu/English may occur.
- Filtering:
  - Removed retweets/quote-tweets where applicable.
  - Deduplicated exact and near-duplicate messages.
  - Kept only text content (no images or external media).
  - Heuristic script filtering to retain primarily Urdu content.
- Privacy: User handles and URLs were normalized (see preprocessing). Only textual content is provided.

## Annotation

- Process: Manual annotation by domain-aware annotators following written guidelines.
- Label set:
  - `not_hate (0)`: Neutral/benign, counterspeech, or non-hateful opinions.
  - `hate (1)`: Dehumanizing, insulting, or advocating violence/discrimination against protected groups or individuals based on protected characteristics.
- Quality: Disagreements were resolved by discussion/adjudication. Use with human-in-the-loop for deployment scenarios.

## Preprocessing

Applied to produce the cleaned corpora (reflected in the uploaded files):

- Text normalization:
  - Unicode normalization (NFC).
  - Removed diacritics and tatweel (kashida) when present.
  - Collapsed multiple spaces and normalized punctuation spacing.
- Token replacement:
  - Replaced user mentions with `@user`.
  - Replaced URLs with `<url>`.
  - Stripped leading `RT` markers.
- Content filtering:
  - Removed non-textual artifacts (e.g., leftover HTML entities).
  - Deduplicated messages.
- Casing:
  - Urdu is caseless; original letter forms preserved.
- Files:
  - `Urdu_Hate_Speech.xlsx`: cleaned, manually labeled tweets (`Tweet`, `Tag`).
  - `preprocessed_combined_file (1).xlsx`: combined/cleaned set used in modeling; same column schema (`Tweet`, `Tag`).

Note: Additional balancing (e.g., TF–IDF + SMOTE) was applied only during model training, not as part of the released raw labels.

## Dataset Structure

### Data Files Included

- `Urdu_Hate_Speech.xlsx`
- `preprocessed_combined_file (1).xlsx`

### Data Fields

- `Tweet` (string): the input text in Urdu.
- `Tag` (int): class id in {0, 1}
  - 0: not_hate
  - 1: hate

If you prefer canonical names, rename to:
- `text` (string)
- `label` (ClassLabel: not_hate/hate)

### Data Splits

Splits are not fixed and can be defined at load time. Example mapping:

- train: `preprocessed_combined_file (1).xlsx`
- validation: `Urdu_Hate_Speech.xlsx`

Adjust as needed.

## How to Use

Install:
```bash
pip install -U datasets
# For Excel files:
pip install -U "datasets[excel]" openpyxl
```

Load and (optionally) rename columns:
```python
from datasets import load_dataset

DATASET_ID = "Adnan855570/urdu-hate-speech"

ds = load_dataset(
    DATASET_ID,
    data_files={
        "train": "preprocessed_combined_file (1).xlsx",
        "validation": "Urdu_Hate_Speech.xlsx",
    },
    # sheet="Sheet1",  # uncomment if you need a specific Excel sheet
)

# If your files use 'Tweet' and 'Tag', rename to canonical names:
ds = ds.rename_columns({"Tweet": "text", "Tag": "label"})

# Cast label to ClassLabel with names
from datasets import ClassLabel, Value, Features
features = Features({
    "text": Value("string"),
    "label": ClassLabel(names=["not_hate", "hate"]),
})
ds = ds.cast(features)
```

Quick usage with Transformers:
```python
from transformers import pipeline
clf = pipeline("text-classification", model="Adnan855570/urdu-roberta-hate", top_k=None)

ex = ds["train"][0]["text"]
print(clf(ex))
```

## Ethical Considerations and Limitations

- Context sensitivity: May misclassify satire, reclaimed slurs, or code-mixed content.
- Bias: Labels may reflect annotator and platform biases.
- Risk: False positives can suppress legitimate speech; false negatives can miss harmful content.
- Recommendation: Use with human-in-the-loop; audit for disparate impact; calibrate decision thresholds.

## Licensing

This dataset is released under the MIT License.

## Citation

Please cite if you use this dataset.
```bibtex
@misc{urdu_hate_speech_ds_2025,
  title  = {Urdu Hate Speech Dataset},
  author = {Adnan},
  year   = {2025},
  howpublished = {\url{https://huggingface.co/datasets/Adnan855570/urdu-hate-speech}}
}
```

## Acknowledgements

- Thanks to contributors and the Urdu NLP community.