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)

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

Character Length (per message)

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

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:

@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: muhammadkhubaibahmad854@gmail.com

LinkedIn: Muhammad Khubaib Ahmad