themohal/persoarabic-lid

PersoArabicLID

Language identification data for 19 languages written in Perso-Arabic scripts, including Saraiki (skr), Urdu, Western Punjabi, Sindhi, Pashto and several Kurdish varieties. The main use is telling closely related languages that share one script apart (for example Saraiki vs Urdu vs Punjabi).

This is a re-packaged copy of the PALI dataset by Ahmadi, Agarwal and Anastasopoulos. Original repository: https://github.com/sinaahmadi/PersoArabicLID. All credit for the data belongs to the original authors; please cite their paper (below).

Loading

from datasets import load_dataset

ds = load_dataset("themohal/persoarabic-lid", "merged")     # upstream's recommended training mix
print(ds["train"][0])

Configs and splits

config	upstream folder	contents
clean	datasets/0	clean text, no synthetic noise
noisy_20 ... noisy_100	datasets/20 ... datasets/100	text with 20% to 100% synthetic character noise
noisy_all	datasets/all	the five noisy levels merged
merged	datasets/merged	clean + noisy data, extra clean rows added for Urdu, Persian, Arabic and Uyghur

config	split	rows
clean	train	118,973
clean	validation	12,518
noisy_20	train	100,200
noisy_40	train	100,200
noisy_60	train	100,200
noisy_80	train	100,200
noisy_100	train	100,200
noisy_all	train	103,201
merged	train	267,399

Schema

field	type	description
text	string	one sentence or short text
label	ClassLabel (19 classes)	language code used by the upstream repository
language	string	English language name
iso639_3	string	ISO 639-3 code of the language

Labels

label	language	iso639_3	clean/train rows	clean/validation rows
ar	Arabic	arb	7,200	800
azb	Azeri Turkish (South Azerbaijani)	azb	7,199	801
bal	Balochi	bal	1,797	114
brh	Brahui	brh	180	5
ckb	Central Kurdish (Sorani)	ckb	7,200	800
fa	Persian	fas	7,200	800
glk	Gilaki	glk	7,198	802
hac	Gorani (Hawrami)	hac	7,193	674
kas	Kashmiri	kas	7,211	434
ku	Northern Kurdish (Kurmanji)	kmr	7,198	800
mzn	Mazanderani	mzn	7,199	801
pa	Western Punjabi	pnb	7,200	800
ps	Pashto	pus	7,200	800
sd	Sindhi	snd	7,200	800
sdh	Southern Kurdish	sdh	7,200	800
skr	Saraiki	skr	7,200	800
trw	Torwali	trw	1,798	87
ud	Urdu	urd	7,200	800
ug	Uyghur	uig	7,200	800

Notes on the label codes, which are inherited from upstream and are not all standard:

• ud is Urdu (ISO 639-1 would be ur), pa is Western Punjabi (pnb), ku is Northern Kurdish (Kurmanji) in Arabic script.
• The upstream README lists both sd and sdh as Sindhi. sdh corresponds to the Southern Kurdish corpus (SouthernKurdish_sdh.txt upstream); Sindhi is sd.
• Use the iso639_3 column when you need an unambiguous code.

Things to know before using it

• No official test split. The upstream repository describes test sets but ships only train.txt files. In clean, a validation split was carved out by this repackaging (10% of the unique sentences per language, seed 42, with all copies removed from train). It is not the paper's test set, so scores are not directly comparable with the paper.
• Noisy configs cover 15 languages. Arabic (ar), Persian (fa), Urdu (ud) and Uyghur (ug) have no synthetic noise upstream, so they appear only in clean and merged.
• Small validation sets for low-resource languages. Brahui, Torwali and Balochi have very little text, so their validation counts are tiny (see the table above); do not read much into per-language scores there.
• Noisy configs have no eval split. The same sentence appears at several noise levels in noisy_all and merged, so a random split would leak. Evaluate on clean/validation or on your own data.
• Upsampling. Languages with little text (for example Brahui, Torwali, Balochi) are upsampled upstream, so train contains repeated rows.
• Synthetic noise. The noisy configs are generated with character-mapping scripts to imitate orthographic variation between related languages. They are not natural text.
• Domain. The sentences come from corpora such as Wikipedia dumps, which are clean and formal. Expect lower accuracy on informal, social-media or code-mixed text.
• Licensing. The upstream repository is MIT licensed. Some corpora are derived from Wikipedia dumps (the upstream file names carry dump dates), whose text is CC BY-SA; check the paper and upstream corpora before any commercial reuse.

Citation

@inproceedings{ahmadi2023pali,
    title = "{PALI}: A Language Identification Benchmark for {Perso-Arabic} Scripts",
    author = "Ahmadi, Sina and Agarwal, Milind and Anastasopoulos, Antonios",
    booktitle = "Proceedings of the 10th Workshop on NLP for Similar Languages, Varieties and Dialects (VarDial)",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "The 17th Conference of the European Chapter of the Association for Computational Linguistics"
}