waroodzkhan/code-switching-codesaviours-si26-warood

Code-Switching Codesaviours SI-26 Dataset

Dataset Description

This dataset contains 150 naturally occurring Roman Urdu–English code-switched sentences, commonly spoken by Pakistani speakers in casual, everyday communication. Each sentence has been broken down into individual words, and every word is labeled by language.

Roman Urdu–English code-switching is extremely common in Pakistan (spoken/written by an estimated 230+ million people) but is poorly handled by existing NLP models, which are typically trained on monolingual data. This dataset aims to help build and evaluate models that can understand this kind of mixed-language text.

Collection Method

Sentences were collected to reflect real, natural code-switching patterns across everyday contexts, including:

• Casual conversation
• University and academic life
• Work and professional communication
• Plans with friends
• Family conversations
• Relationships and feelings
• Shopping, food, and lifestyle
• Social media and internet culture

Each sentence reflects genuine Roman Urdu–English mixing patterns as used in real conversation (e.g. WhatsApp messages, social media comments, and everyday speech), rather than formal or translated text.

Dataset Structure

The dataset is a flat CSV where each row represents a single word from a sentence, along with its language label.

Column	Description
sentence	The full original sentence the word belongs to
word	An individual word/token from the sentence
label	Language label for that word

Total sentences: 150 Total word entries: 1207

Label Meanings

Label	Meaning
URD	Roman Urdu word (Urdu written in Latin/English script, e.g. "yaar", "bohot", "hai")
ENG	English word
MIX	A word that blends both languages (hybrid/coined term)

Label Distribution

• URD: 681 words
• ENG: 526 words

Intended Use

This dataset can be used for:

• Training or fine-tuning language identification models for code-switched text
• Roman Urdu NLP research
• Code-switching pattern analysis
• Building better NLP tools for Pakistani/South Asian social media text

Limitations

• Dataset size is relatively small (150 sentences); larger datasets would improve model robustness
• Roman Urdu spelling is inherently inconsistent (no standardized orthography), which may introduce variation
• Sentences primarily reflect informal, conversational register

Authors

Created as part of the Code Saviours SI-26 program, Week 6 project — Warood Zahra Khan