Khubaib01/RomanUrdu-NLP-Sentiment-Corpus

---
license: apache-2.0
task_categories:
- text-classification
language:
- ur
tags:
- code
size_categories:
- 100K<n<1M
---

# RomanUrdu-NLP-Sentiment-Corpus
**Largest Open-Source Roman Urdu Sentiment Dataset with Slang Robustness**

---

## Overview

This repository presents the largest publicly available Roman Urdu sentiment analysis dataset, containing **134,052 labeled text samples** collected from chats and social media platforms. The dataset is designed to be:

- Robust to slang and informal Roman Urdu
- High-quality through LLM-assisted labeling and human validation
- Balanced across sentiment classes
- Suitable for research and real-world NLP applications

This dataset supports research in:

- Sentiment Analysis
- Low-resource language NLP
- Code-mixed and slang-aware text modeling
- Social media opinion mining

---

## Dataset Design Goals

The dataset was created with the following objectives:

1. Robustness to slang, abbreviations, and spelling variations
2. Large-scale corpus for deep learning models
3. High annotation quality through hybrid labeling
4. Open-source accessibility under Apache 2.0
5. Future extensibility with emotion labels

---

## Dataset Structure

Each row contains two attributes:

| Column | Description |
|--------|-------------|
| `message` | Roman Urdu text |
| `label` | Sentiment class (`Positive`, `Neutral`, `Negative`) |

---

## Dataset Statistics

### General Statistics

- Total samples: **134,052**
- Unique messages: **109,409**
- Most frequent message: `"Good"` (24 occurrences)
- Labels: **3** (Positive, Neutral, Negative)

---

### Class Distribution

| Label | Percentage |
|-------|------------|
| Positive | 28% |
| Neutral | 32% |
| Negative | 40% |

This distribution reflects real-world social media sentiment skew.

---

## Message Length Statistics

### Word Length (per message)

```python
count 134052
mean 13.55 words
std 19.46
min 0
25% 5
50% 9
75% 16
max 3212
```

### Character Length (per message)

```python
count 134052
mean 66.62 chars
std 102.15
min 1
25% 22
50% 41
75% 81
max 19074
```

### Average Word Length by Label

| Label | Avg Words |
| -------- | --------- |
| Negative | 18.05 |
| Positive | 13.68 |
| Neutral | 7.87 |

Negative samples tend to be longer and more expressive, while neutral messages are shorter and concise.

## Annotation Methodology

The dataset was created in two major phases:

### Phase 1: Initial Dataset (99K Samples)

- Labeled using LLM-assisted annotation

- Verified by human annotators and validators

- Released previously in the form of embeddings

- Used to train the baseline model:
`Khubaib01/roman-urdu-sentiment-xlm-r`

> - Read the paper here: [Paper](https://doi.org/10.5281/zenodo.18080524)

### Phase 2: Extended Dataset (134K Samples)

- Additional samples labeled using the trained model

- All newly labeled samples validated by human reviewers

Focused on including:

- Slang

- Informal expressions

- Local dialect usage

- Social media language patterns

This hybrid annotation pipeline ensures:

- Scalability

- Consistency

- High label reliability

## Benchmark Model

A sentiment classification model trained on the initial 99k dataset:

**Model Name:**
`Khubaib01/roman-urdu-sentiment-xlm-r`

**Performance:**

- Achieved 84% accuracy

- Ranked highest among available Roman Urdu sentiment models on HuggingFace at time of evaluation

- Benchmarked against multiple multilingual and Roman Urdu models

This model was also used to assist labeling for the extended dataset.

## Slang & Robustness Focus

Unlike many clean benchmark datasets, this dataset includes:

- Local slang

- Abbreviations (e.g., "bkl", "yr", "bhai", "scene off")

- Misspellings

- Mixed English + Roman Urdu

- Informal sentence structures

This makes the dataset suitable for:

- Real-world deployment

- Chatbots

- Social media analysis

- Low-resource language research

## Future Work

Planned extensions include:

- Emotion labels (anger, joy, sadness, fear, etc.)

- Multi-label emotion classification

- Offensive and toxicity detection

- Language normalization benchmarks

## Core Author

**Muhammad Khubaib Ahmad**
Core Engineer & Researcher
Creator of:

- Roman Urdu Sentiment Dataset (134k)

- 99k Roman Urdu embeddings dataset

`Khubaib01/roman-urdu-sentiment-xlm-r` model

## Contributors (Human Validation & Annotation)

The following contributors reviewed labels and worked as data validators and annotators:

- **Ayesha Khalid**

- **Faiez Ahmad**

- **Khadija Faysal**

Their role ensured quality control and reduced noise and labeling errors.

## License

This dataset is released under the **Apache License 2.0**.

You are free to:

- Use

- Modify

- Distribute

- Train models

- Use commercially

With proper attribution.

## Citation

If you use this dataset in your research, please cite:

```bibtex
@misc{muhammad_khubaib_ahmad_2026,
author = { Muhammad Khubaib Ahmad },
title = { RomanUrdu-NLP-Sentiment-Corpus (Revision 98d0169) },
year = 2026,
url = { https://huggingface.co/datasets/Khubaib01/RomanUrdu-NLP-Sentiment-Corpus },
doi = { 10.57967/hf/7931 },
publisher = { Hugging Face }
}
```

## Ethical Considerations

- All data has been anonymized.

- No personal identifiers are included.

- Data collected from public sources and chat-style corpora.

- Dataset intended for research and educational purposes only.

## Author Contact

**Email:** [e-mail]

**LinkedIn:** [Muhammad Khubaib Ahmad](https://www.linkedin.com/in/muhammad-khubaib-ahmad-)