Zainab-Binte-Khalid/code-switching-codesaviours-si26-zainab

Roman Urdu–English Code-Switching Dataset

Dataset Description

This dataset contains 1,901 sentences and 21,370 word-level language labels, built to capture how Roman Urdu and English are naturally mixed together in everyday Pakistani online communication.

Code-switching — blending two languages within a single sentence — is how the vast majority of Pakistanis actually write and speak online, on platforms like Twitter/X, Facebook, YouTube, Reddit, and WhatsApp. A sentence like "Aaj mera mood nahi hai for anything" is completely normal in this context, but most existing NLP models and datasets are built around monolingual text and don't handle this kind of mixing well. This dataset was created to help close that gap by providing real, word-level labelled examples that a model could learn from.

Motivation

Standard language identification and NLP tools typically assume a sentence is written in a single language. Roman Urdu–English code-switching breaks that assumption at the word level, not just the sentence level — a single sentence can contain both languages, sometimes even within back-to-back words. This dataset supports building and evaluating models (e.g. token classifiers, chatbots, sentiment analysis systems, or content moderation tools) that need to correctly interpret this kind of everyday, informal, mixed-language text.

How the Data Was Collected

Sentences were gathered from real, naturally occurring Roman Urdu–English text found in Pakistani online spaces, combined with existing public Roman Urdu datasets. Specifically, the dataset combines:

• Personally collected and hand-labelled sentences, sourced from real online conversations and posts
• Sentences drawn from existing public Roman Urdu text sources

All sentences were deduplicated across sources to avoid repetition, then labelled at the word level by hand.

Labels

Each individual word in a sentence is tagged with one of three labels describing its language:

Label	Meaning	Example
ENG	The word is English	busy, meetings, feeling
URD	The word is Urdu, written in Roman script	Aaj, ka, bohot, nahi
MIX	The word blends both languages within itself	e.g. compound/hybrid coinages

Dataset Structure

The dataset is provided as a single flat CSV file (dataset_labels.csv) with three columns:

• sentence — the full original sentence, unmodified
• word — one individual word taken from that sentence
• label — the language label for that specific word (ENG, URD, or MIX)

Because labelling happens at the word level, each sentence appears across multiple rows — one row per word it contains. To reconstruct a sentence's full word-by-word labelling, filter the CSV by matching sentence values.

Example

For the sentence "Aaj ka din bohot busy tha":

sentence	word	label
Aaj ka din bohot busy tha	Aaj	URD
Aaj ka din bohot busy tha	ka	URD
Aaj ka din bohot busy tha	din	URD
Aaj ka din bohot busy tha	bohot	URD
Aaj ka din bohot busy tha	busy	ENG
Aaj ka din bohot busy tha	tha	URD

Dataset Statistics

• Sentences: 1,901
• Total word-level entries: 21,370
• Label distribution: URD-majority, with a substantial share of English tokens and a small number of blended (MIX) tokens, reflecting realistic code-switching patterns rather than an artificially balanced split.

Intended Uses

This dataset is intended for:

• Training or evaluating token-level language identification models on Roman Urdu–English text
• Supporting downstream NLP tasks on code-switched South Asian social media text, such as sentiment analysis, chatbots, and content moderation
• Educational and research use in low-resource and code-switching NLP

Limitations

• The dataset reflects informal, conversational online text and may not generalize to formal writing.
• Label decisions for ambiguous or hybrid words were made manually and may reflect some subjectivity, as is common in code-switching annotation.
• The dataset size, while sufficient for many use cases, is modest compared to large monolingual corpora — it is best suited for fine-tuning or evaluation rather than training large models from scratch.

License

This dataset is released under the CC BY 4.0 license.

Author

Zainab Binte Khalid — Code Saviours SI-26, Project 2