abdullaharoon/Urdu-Multi-Domain-Benchmark

Urdu Multi-Domain Benchmark (v1.1)

A benchmark for domain robustness of Urdu text classifiers: 26 datasets in two scripts (Nastaliq and Roman Urdu), grouped into 7 transfer matrices over four task types (sentiment and polarity, abusive language, fake news, question-answer pair validation). All domains of a matrix share one label space, so every source-to-target pair is a valid domain shift.

• Paper: Domain Robustness of Multilingual NLP Models Across Urdu and Roman Urdu Scripts (Haroon and Bashir, FAST-NUCES).
• Code, audits and model predictions: https://github.com/AbdullahPatti/Urdu-Multi-Domain-Script-Research
• DOI: https://doi.org/10.5281/zenodo.23053495

Content warning. The abusive-language and hate-speech domains contain offensive, hateful and sectarian text. Classifiers trained on these data should not be deployed without human review.

What changed from v1.0

v1.1 supersedes v1.0 (DOI 10.5281/zenodo.22195610). Domains were selected by explicit inclusion criteria, and every split is free of exact and near-duplicate train-test overlap:

• Excluded candidates (duplicated sources, labels predictable from the origin of an item, fewer than 100 test items):
• Nastaliq/Hate Speech/Domain_E_Inter_Faith_Sectarian_Ethnic: same tweets and labels as Hate Speech A (ISE Level-1 re-uses the 22K Offensive pool)
• Nastaliq/Sentiment Analysis/Domain_D_Governance_Political: same texts and labels as Sentiment A
• Roman/Sentiment Analysis/Domain_D_Utilities_Apps_Courses: same 11,000-review corpus as Roman Sentiment B
• Roman/Twitter or Social Media Opinions/Domain_A_Twitter_Political_Social: positive/negative items are the RUSAD corpus (= Roman Sentiment B); all neutral items come from a different source, so the label is predictable from the source
• Nastaliq/Fake News Detection/Domain_C_General_Pakistani_News: fake and real items come from two separate files; fake items are largely fact-check verdicts, so the label is predictable from the source
• Nastaliq/Fake News Detection/Domain_D_Fact_Checking_Platform: real items carry an outlet dateline ("92 News", 90% of real vs 3.5% of fake items), so the label is predictable from the source
• Nastaliq/Sentiment Analysis/Domain_C_Civic_Social_Topics: too small to evaluate (162 items, 33 in test; minimum is 100 test items)
• Near-duplicate-aware splits. LLM-generated domains were re-split so that clusters of template near-duplicates (character 5-gram Jaccard >= 0.8) do not straddle train and test.
• Question-grouped QA splits. No question (or near-duplicate question) occurs in both splits.
• Cross-domain overlap removed inside each matrix.

Domains

Matrix	ID	Domain	Origin	Train	Test	Licence	Distribution
N-SA	SA-A	Twitter Social Media	organic	781	194	none stated (citation requested)	rebuild
N-SA	SA-B	Movie Reviews	translated	6,695	1,638	ODbL 1.0	file
N-SA	SA-E	Socioeconomic Agricultural	generated	1,629	407	xAI terms (attribute Grok)	file
N-HS	HS-A	Social Media Offensive	organic	17,152	4,133	none stated	rebuild
N-HS	HS-B	Social Media Hate Speech	organic	18,486	3,540	MIT	file
N-HS	HS-C	Politics Sports Health Family	generated	1,360	340	xAI terms (attribute Grok)	file
N-HS	HS-D	Sports Labor Arts Education	generated	1,356	339	xAI terms (attribute Grok)	file
N-FN	FN-A	Multi Topic News	organic	634	260	none stated	rebuild
N-FN	FN-B	Ax to Grind	organic	7,217	1,645	none stated	rebuild
N-QA	QA-A	Religious Hadith	organic	992	250	MIT	file
N-QA	QA-B	General Knowledge	organic	7,564	1,889	MIT	file
N-QA	QA-C	Long Form QnA	generated	896	224	xAI terms (attribute Grok)	file
N-QA	QA-D	Short Form QnA	generated	2,764	691	xAI terms (attribute Grok)	file
R-SA	SA-A	Mixed Public Discourse	organic	19,262	2,140	MIT	file
R-SA	SA-B	Social Media Politics Drama	organic	8,781	2,163	CC BY 4.0	file
R-SA	SA-C	YouTube Entertainment	organic	15,587	3,759	unknown	file
R-CA	CA-A	YouTube Comments	uncertain	3,469	866	unknown	file
R-CA	CA-B	Content Creator Comments	generated	1,040	260	xAI terms (attribute Grok)	file
R-CA	CA-C	Social Media Diverse Vocab	generated	1,040	260	xAI terms (attribute Grok)	file
R-CA	CA-D	General Online Video	generated	960	240	xAI terms (attribute Grok)	file
R-P3	PR-A	Daraz Ecommerce	organic	13,489	3,294	CC BY 4.0	file
R-P3	PR-B	Restaurant Food	organic	600	148	unknown	file
R-P3	PR-D	General Social Comments	organic	19,349	4,026	CC BY 4.0	file
R-P3	TW-B	Consumer Electronics Gaming	generated	1,160	290	xAI terms (attribute Grok)	file
R-P3	TW-C	Public Civic Services	generated	1,160	290	xAI terms (attribute Grok)	file
R-P3	TW-D	Mixed Social Commentary	generated	1,280	320	xAI terms (attribute Grok)	file

Origin: organic (public corpus of native text), translated (machine-translated from English), generated (produced by the first author with xAI Grok and checked by hand), uncertain. Full provenance and URLs: domains.csv.

Getting the data

Domains marked file are included directly (data/), each as a config of this dataset:

from datasets import load_dataset
ds = load_dataset("abdullaharoon/Urdu-Multi-Domain-Benchmark", "R-SA_SA-B")

Domains marked rebuild come from public corpora whose authors state no licence, so their text is not redistributed here. manifests/ holds the row order, a SHA-1 of each normalised text and our label; rebuild.py downloads the authors' GitHub repositories and writes the train/test files:

pip install pandas openpyxl
python rebuild.py            # writes data/<domain>/{train,test}.csv for the four rebuild domains

The rebuild recovers every item except 2 of 21,285 ISE-Hate tweets, which differ in the authors' current file; rebuild.py reports them.

Licences

Each domain keeps the licence of its source (table above). Traced sources with a permissive licence are released under MIT, CC BY 4.0 or ODbL 1.0. Four published corpora (Urdu Sentiment Corpus, ISE-Hate, Bend the Truth, Ax-to-Grind Urdu) state no licence and are distributed through rebuild.py only; please cite their authors. Generated domains were produced with xAI Grok; its terms ask that such output be attributed to Grok and prohibit using it to develop models that compete with xAI. 3 organic domains were obtained from public uploads whose authors we could not identify; they are released for research use and will be removed on request from a rights holder. Our own contributions (splits, manifests, audits, generated-data curation) are released under CC BY-NC 4.0.

Citation

Please cite the paper and the original corpora listed in domains.csv.