Khubaib01/roman-urdu-sentiment-embeddings

Roman Urdu Sentiment Embeddings Dataset

Overview

This repository contains a research-grade Roman Urdu Sentiment Analysis dataset released as anonymized sentence embeddings. Roman Urdu is a low-resource language with high linguistic variability, heavy slang usage, and frequent code-mixing with English. This dataset is curated to support robust sentiment analysis research while ensuring strict privacy preservation.

Raw text data has not been released. Instead, all messages are transformed into dense vector representations, making the dataset safe for open-source, academic, and industrial use.

---

Dataset Summary

• Language: Roman Urdu (Urdu–English code-mixed included)
• Task: Sentiment Analysis
• Labels: Positive, Negative, Neutral
• Total samples: 99,000
• Samples per class: 33,000 (fully balanced)
• Embedding dimension: 768
• Embedding dtype: float32
• Collection period: November 2025
• License: CC BY-NC 4.0

---

Data Sources

The original corpus was collected from real-world Roman Urdu usage, including:

• WhatsApp chat conversations
• YouTube comments
• Social media posts and discussions
• Slang-heavy informal text
• Formal and semi-formal Roman Urdu
• Urdu–English code-mixed sentences

This diversity ensures the dataset reflects authentic language usage across multiple domains.

---

Privacy and Ethics

To ensure ethical data release and user privacy:

• No raw text is included
• All samples are converted to vector embeddings
• No personally identifiable information (PII) is present
• Original messages cannot be reconstructed

This approach allows safe reuse without compromising privacy.

---

Preprocessing and Quality Control

The dataset underwent extensive preprocessing:

• Data cleaning and deduplication
• Sentiment label validation
• Balanced class enforcement
• Noise reduction
• Embedding integrity checks

Embedding Statistics

• Min: -0.8144
• Max: 1.0553
• Mean: 0.00188

No NaN or infinite values are present.

---

Embedding Generation

Sentence embeddings were generated using a transformer-based multilingual model with a 768-dimensional hidden representation. The embeddings are optimized for Roman Urdu and code-mixed Urdu–English text.

They are suitable for:

• Linear and neural classifiers
• Transfer learning
• Clustering and similarity search
• Benchmarking low-resource NLP systems

---

Benchmarking Results

A sentiment classifier trained on this dataset (Khubaib01/roman-urdu-sentiment-xlm-r) was benchmarked against publicly available Roman Urdu sentiment models.

Model	Accuracy (Slang)	Macro F1 (Slang)	Accuracy (Formal)	Macro F1 (Formal)
Khubaib01/roman-urdu-sentiment-xlm-r	0.8444	0.8351	0.7508	0.7318
ayeshasameer/xlm-roberta-roman-urdu-sentiment	0.7833	0.7803	0.6246	0.6246
tahamueed23/urdu-roman-urdu-sentiment	0.4333	0.3374	0.4850	0.3872
Aimlab/xlm-roberta-roman-urdu-finetuned	0.2833	0.2084	0.2558	0.2171

The proposed model achieves state-of-the-art performance, particularly on slang-heavy and informal Roman Urdu text.

---

Files

• embeddings.npz
  NumPy compressed file containing embeddings of shape (99000, 768).

• labels.csv
  CSV file containing sentiment labels aligned by index with the embeddings.

---

Usage Example

import numpy as np
import pandas as pd

embeddings = np.load("embeddings.npz")["arr_0"]
labels = pd.read_csv("labels.csv")

print(embeddings.shape)
print(labels.head())

---

Intended Use

This dataset is intended for:

• Low-resource language research
• Roman Urdu sentiment benchmarking
• Embedding-based NLP experiments
• Academic research and education

---

Limitations

• Raw text is unavailable due to privacy constraints
• Not suitable for generative language modeling
• Best used for classification and representation learning

---

Citation

If you use this dataset, please cite:

Roman Urdu Sentiment Embeddings Dataset  
Khubaib01, 2025

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0) license.