Datasets:
license: cc0-1.0
configs:
- config_name: default
data_files:
- split: test
path: udhr-lid.csv
multilinguality:
- multilingual
language:
- tir
- rmn
- arb
- mxv
- mal
- fij
- som
- cot
- fur
- vie
- zlm
- bam
- chr
- maz
- yad
- ztu
- ykg
- ccp
- alt
- ayr
- njo
- bci
- gyr
- run
- haw
- rgn
- cak
- kwi
- fra
- agr
- duu
- ilo
- nhn
- kdh
- cnh
- bod
- mya
- ady
- pol
- ydd
- cos
- lot
- arl
- glv
- gag
- bfa
- afr
- lij
- zlm
- ibb
- toi
- tzm
- ron
- ojb
- san
- eng
- bum
- pam
- kqs
- dje
- auc
- smo
- por
- fry
- lad
- pov
- tyv
- guc
- huu
- ese
- kbp
- eve
- yrk
- lin
- tdt
- qvc
- top
- nav
- twi
- oss
- lia
- ame
- hun
- lit
- que
- qug
- nku
- csa
- lao
- knc
- kjh
- jav
- mam
- ita
- ppl
- aar
- tbz
- ssw
- bug
- srp
- kaz
- min
- mad
- orh
- tgk
- kat
- uig
- tzo
- hat
- shn
- kbd
- niv
- idu
- krl
- abk
- mto
- gla
- ijs
- cri
- uzn
- tah
- tob
- kir
- quy
- hnj
- srr
- lvs
- nan
- hns
- snk
- swh
- ekk
- guu
- div
- dzo
- spa
- hms
- ell
- ace
- war
- ind
- cjy
- cfm
- nds
- ewe
- tel
- src
- fuf
- vmw
- zro
- men
- kqn
- nzi
- taj
- khk
- ddn
- nso
- mxi
- pon
- fvr
- hau
- ktu
- tem
- yor
- pnb
- ltz
- evn
- cjs
- nba
- niu
- dan
- acu
- zgh
- chj
- heb
- lua
- quz
- uig
- cbi
- jav
- cpu
- wuu
- mah
- kmb
- mcd
- ben
- lus
- ajg
- azj
- tha
- dga
- isl
- sus
- fuf
- fkv
- jiv
- mor
- nio
- als
- buc
- kde
- nbl
- ceb
- ven
- sun
- cbt
- swb
- tur
- dyo
- sin
- pbu
- ada
- pap
- qvh
- loz
- pan
- qva
- sme
- bax
- tuk
- hsb
- hus
- qvn
- ban
- cha
- zyb
- hin
- tat
- uzn
- qxu
- gej
- quc
- mnw
- bho
- udu
- kha
- kbr
- tsz
- pau
- mkd
- shp
- ike
- lue
- tgl
- yap
- yua
- koi
- hrv
- emk
- tet
- ndo
- cbu
- vep
- cmn
- sag
- nym
- rus
- gjn
- guk
- kri
- ote
- lun
- vai
- bis
- arn
- tsn
- gle
- hak
- gkp
- ura
- tca
- xho
- wln
- amc
- mos
- lld
- bul
- qxn
- bcl
- ctd
- dip
- dag
- kek
- bre
- mri
- fin
- sah
- cym
- kan
- fao
- gsw
- sey
- bem
- bos
- bin
- chv
- tpi
- ami
- oaa
- lob
- ast
- nno
- sco
- tuk
- khm
- pes
- pbb
- tam
- ibo
- san
- sid
- plt
- guj
- hsn
- kin
- lug
- slr
- koo
- xsm
- jpn
- oki
- deu
- rar
- pcm
- hni
- vec
- gld
- sot
- crs
- fuv
- srp
- npi
- nya
- kea
- blt
- roh
- cbr
- chk
- kal
- mfq
- quh
- kor
- slv
- cof
- shk
- zul
- qwh
- fon
- mic
- prs
- mag
- bel
- iii
- mar
- dyu
- boa
- swe
- pis
- mlt
- amh
- umb
- cnr
- mai
- toj
- csw
- ina
- bba
- cbs
- kng
- oci
- pcd
- miq
- lat
- qvm
- wwa
- bos
- urd
- kmr
- ido
- gaa
- epo
- gaz
- cat
- hye
- cni
- suk
- gug
- gan
- cjk
- tzh
- zam
- ces
- cic
- mcf
- not
- kaa
- tso
- piu
- fat
- mzi
- snn
- tly
- eus
- nld
- nob
- wol
- hlt
- sna
- tiv
- ton
- hea
- skr
- lns
- rup
- cab
- glg
- tgl
- yao
- nyn
- aii
- tzm
- slk
- ukr
- kkh
- zdj
- amr
- yue
- crh
- hil
tags:
- UDHR
- udhr
- language identification
- LID
- glot
- GlotLID
pretty_name: UDHR-LID
UDHR-LID
Why UDHR-LID?
You can access UDHR (Universal Declaration of Human Rights) here, but when a verse is missing, they have texts such as "missing" or "?". Also, about 1/3 of the sentences consist only of "articles 1-30" in different languages. We cleaned the entire dataset from XML files and selected only the paragraphs. We cleared any unrelated language texts from the data and also removed the cases that were incorrect.
Incorrect? Look at the ckb and kmr files in the UDHR. Both are the same! ckb is known for the Arabic script, although it can also be written in Latin. Clearly, a unique file cannot belong to two different languages (at least in the text language identification task). We also deleted files that we believe those scripts are no longer in use.
The deleted files include:
- ckb_Latn (Arabic is in use.)
- azb_Latn (Arabic is in use.)
- khk_Mong (Cyrillic is in use, but the other varieties of Mongolian language in Inner Mongolia use the Mongolian writing system.),
- vie_Hani (Latin is in use.)
For dealing with scripts in other languages, if you are interested, check Glotscript code and paper. We have prepared a tool for detecting the script of a text, as well as metadata to determine the correct script for each language.
We believe UDHR should remain a test corpus in NLP, not a training corpus. Of course, we are not opposed to great works such as Franc built on top of UDHR. However, if your work scale is much bigger than UDHR, do not put UDHR in your data. Use it as test/validation, or find out what is wrong with your training data with help of UDHR. Be aware that a part of UDHR may be hosted on other websites such as Wikipedia, news websites like BBC, collaborative translation communities like Tatoeba. Before using UDHR as a test, exclude any sentence where UDHR is a part of your training.
We created this corpus for language identification evaluation task in our GlotLID paper, but feel free to use it for your own task. The texts here are not in order, and they're not parallel. However, each row of data belongs to the determined language, long, cleaned, and has rich linguistic content!
Usage (HF Loader)
from datasets import load_dataset
dataset = load_dataset('cis-lmu/udhr-lid', split='test')
print(dataset[0]) # First row of udhr-lid
Download
If you are not a fan of the HF dataloader, download each language directly:
! wget https://huggingface.co/datasets/cis-lmu/udhr-lid/resolve/main/udhr-lid.csv
or clone the whole repository:
! git clone https://huggingface.co/datasets/cis-lmu/udhr-lid
License
UDHR is the most translated copyright-free document in the world. We license the actual packaging, the metadata and the annotations of these data under the cc0-1.0 (waiving all of the rights under copyright law).
Citation
If you use any part of this data in your research, please cite it (along with http://www.unicode.org/udhr/d/) using the following BibTeX entry.
@inproceedings{
kargaran2023glotlid,
title={{GlotLID}: Language Identification for Low-Resource Languages},
author={Kargaran, Amir Hossein and Imani, Ayyoob and Yvon, Fran{\c{c}}ois and Sch{\"u}tze, Hinrich},
booktitle={The 2023 Conference on Empirical Methods in Natural Language Processing},
year={2023},
url={https://openreview.net/forum?id=dl4e3EBz5j}
}