README.md

metadata

language:
  - aa
  - ab
  - ace
  - ach
  - ady
  - aeb
  - aii
  - ak
  - alz
  - apc
  - apd
  - ar
  - arn
  - arz
  - as
  - av
  - awa
  - ay
  - ayl
  - ba
  - bal
  - ban
  - bbc
  - bci
  - bem
  - ber
  - bew
  - bho
  - bik
  - bjn
  - bo
  - br
  - brx
  - bts
  - btx
  - bua
  - bug
  - ce
  - cgg
  - ch
  - chk
  - chm
  - ckb
  - cnh
  - cpf
  - crh
  - crs
  - ctg
  - cv
  - din
  - doi
  - dov
  - dv
  - dyu
  - dz
  - ee
  - efi
  - fa
  - ff
  - fj
  - fo
  - fon
  - fr
  - fur
  - gaa
  - gn
  - gom
  - gv
  - hi
  - hil
  - hne
  - hrx
  - iba
  - ilo
  - iso
  - iu
  - kaa
  - kac
  - kbd
  - kek
  - kg
  - kha
  - ki
  - kl
  - kr
  - kri
  - ks
  - ktu
  - kv
  - lg
  - li
  - lij
  - lmo
  - ln
  - ltg
  - lu
  - luo
  - lus
  - mad
  - mag
  - mai
  - mak
  - mam
  - meo
  - mfe
  - mh
  - min
  - mos
  - ms
  - mwr
  - nd
  - ndc
  - new
  - nhe
  - nr
  - nso
  - nus
  - nv
  - oc
  - om
  - os
  - pa
  - pag
  - pam
  - pap
  - pcm
  - qu
  - quc
  - rhg
  - rn
  - rom
  - sa
  - sah
  - sat
  - scn
  - se
  - sg
  - shn
  - skr
  - ss
  - sus
  - syl
  - szl
  - tcy
  - tet
  - ti
  - tiv
  - tn
  - to
  - tpi
  - trp
  - ts
  - tum
  - ty
  - tyv
  - udm
  - ve
  - vec
  - war
  - wo
  - yua
  - yue
  - zap
  - zza
multilinguality:
  - multili
viewer: false
license: cc-by-4.0
task_categories:
  - translation

Dataset origin: https://github.com/google-research/url-nlp/tree/main/gatitos

GATITOS Multilingual Lexicon

The GATITOS (Google's Additional Translations Into Tail-languages: Often Short) dataset is a high-quality, multi-way parallel dataset of tokens and short phrases, intended for training and improving machine translation models. Experiments on this dataset and Panlex focusing on unsupervised translation in a 208-language model can be found in BiLex Rx: Lexical Data Augmentation for Massively Multilingual Machine Translation.

About the Data and Data Collection

This dataset consists in 4,000 English segments (4,500 tokens) that have been translated into each of 173 languages, 170 of which are low-resource, and three of which are mid-high resource (es, fr, hi). All translations were made directly from English, with the exception of Aymara, which was translated from the Spanish.

This dataset contains primarily short segments: 93% single tokens, and only 23 sentences (0.6%) have over 5 tokens. As such it is best thought of as a multilingual lexicon, rather than a parallel training corpus. The source text is frequent words in the English Language, along with some common phrases and short sentences. Care has been taken to ensure that they include good coverage of numbers, months, days of the week, swadesh words, and names of the languages themselves (including the endonym).

Single tokens are remarkably fickle to translate because of the common issue of polysemy. Some languages in this dataset only provide one translation per source token; others provide multiple translations. When multiple translations are given, they are broken up into multiple lines. Therefore, you can get the average number of translations per token by looking at the number of lines in the file versus the number of deduplicated English lines in the file. The three languages with the most average translations per English token are Betawi, Kachin, and Occitan, each with over 1.5 on average.

Sometimes translators have left annotations like "pike (verb)", "او (مونث)", "фамили(я)", "أسوأ(repeated)" etc.. Wolof is particularly noticeable in this regard, with notes like "xont (verb) which means to feed an animal". Several languages also give definitions instead of translations for some words, e.g. Tok Pisin translation of mole" to "wanpela liklik animal we i save stap long hol long graun".

BCP-47 code	Language Name	Endonym	Alternate names	script
aa	Afar	Qafar		Latn
ab	Abkhaz	Аҧсуа		Cyrl
ace	Achenese	Aceh	Achinese	Latn
ach	Acholi	Acholi		Latn
ady	Adyghe	Черкес	West Circassian	Cyrl
aeb	Tunisian Arabic	اللغة العربيّة التونسيّة		Arab
aii	Assyrian	ܣܘܪܝܬ		Syrc
ak	Twi	Twi	Akan (See note below)	Latn
alz	Alur	Lur		Latn
apc	North Levantine Arabic	شامي‎		Arab
apd-SD	Sudanese Arabic	عربي سوداني		Arab
ar-MA	Darija	الدارجة		Arab
arn	Mapudungun	Mapudungun	Mapuche	Latn
arz	Egyptian Arabic	اللهجة المصرية	Masri	Arab
as	Assamese	অসমীয়া	Asamiya, Asomiya	Beng
av	Avar	магӏарул мацӏ	Avaric	Cyrl
awa	Awadhi	अवधी		Deva
ay	Aymara	Aymar aru		Latn
ayl	Libyan Arabic	العربية الليبية		Arab
ba	Bashkir	Башҡорт		Cyrl
bal	Baluchi	بلوچی	Balochi	Arab
ban	Balinese	Basa Bali		Latn
bbc	Batak Toba	Batak Toba	Toba Batak	Latn
bci	Baoulé	wawle		Latn
bem	Bemba (Zambia)	Ichibemba		Latn
ber	Tamazight (Tifinagh script)	ⵜⴰⵎⴰⵣⵉⵖⵜ	Berber Languages	Tfng
ber-Latn	Tamazight (Latin Script)	Tamaziɣt	the Berber languages	Latn
bew	Betawi	Betawi	Betawi Malay, Jakartan Malay, Batavian Malay	Latn
bho	Bhojpuri	भोजपुरी		Deva
bik	Central Bikol	Sentral Bikol	Bikol Naga; Bikol	Latn
bjn	Banjar	Banjar		Latn
bjn-Arab	Banjar	بنجر; ڤندير بنجر		Arab
bm	Bambara	Bámánánkán		Latn
bm-Nkoo	Bambara (N'Ko Script)	ߓߡߊߣߊ߲ߞߊ߲		Nkoo
bo	Tibetan	བོད་མི།; བོད་ཀྱི།; བོད་སྐད།; བོད་ཡིག།	Lhasa Tibetan, Standard Tibetan	Tibt
br	Breton	brezhoneg		Latn
brx	Bodo	बोडो	Boro	Deva
bts	Batak Simalungun	Bahasa Simalungun	Simalungun	Latn
btx	Batak Karo	Batak Karo		Latn
bua	Buryat	Буряад		Cyrl
bug	Buginese	basa Ugi	Bugis	Latn
ce	Chechen	Нохчийн		Cyrl
cgg	Chiga	Kiga	Kiga	Latn
ch	Chamorro	Finuʼ Chamoru		Latn
chk	Chuukese	Trukese	Trukese	Latn
chm	Meadow Mari	олык марий	Eastern Mari	Cyrl
ckb	Kurdish (Sorani)	سۆرانی	Central Kurdish	Arab
cnh	Hakha Chin	Laica	Laiholh	Latn
cpf	Unknown French Creole			Latn
crh	Crimean Tatar	Крымскотатарский		Cyrl
crh-Latn	Crimean Tatar	qırımtatar tili		Latn
crs	Seselwa Creole French	kreol seselwa	Seychellois Creole, kreol	Latn
ctg	Chittagonian	Chittagonian		Beng
cv	Chuvash	Чăваш		Cyrl
din	Dinka	Thuɔŋjäŋ		Latn
doi	Dogri	डोगरी		Deva
dov	Dombe	ChiDombe	Zambezi, Tonga	Latn
dv	Dhivehi	ދިވެހި	Maldivian, Divehi	Thaa
dyu	Dyula	Julakan	Dioula, Jula	Latn
dz	Dzongkha	རྫོང་ཁ		Tibt
ee	Ewe	Eʋegbe	Eve, Anlo, Anwona	Latn
efi	Efik	Usem mbom Efik; Ikɔ-efik		Latn
es	Spanish	español, castellano	Castilian	Latn
fa-AF	Dari	دری		Arab
ff	Fulfulde	[many]	Fula, Fulah, Fulbe, Fulani, Pular, Pulaar	Latn
fj	Fijian	Fiji; Fijian		Latn
fo	Faroese	Føroysk		Latn
fon	Fon	fongbē		Latn
fr	French	français		Latn
fur	Friulian	Furlan		Latn
gaa	Ga	Gã		Latn
gn	Guarani	avañeʼẽ	Guaraní	Latn
gom	Konkani	कोंकणी		Deva
gv	Manx	Gaelg		Latn
hi	Hindi	हिन्दी		Deva
hil	Hiligaynon	Hiligaynon		Latn
hne	Chhattisgarhi	छत्तीसगढ़ी		Deva
hrx	Hunsrik	Hunsrik		Latn
iba	Iban	Iban		Latn
ilo	Ilocano	Iloko	Iloko, Iluko, Ilokano	Latn
iso	Isoko	Isoko		Latn
iu	Inuktitut	ᐃᓄᒃᑎᑐᑦ	Eastern Canadian Inuktitut	Cans
kaa	Kara-Kalpak	Қарақалпақ; Қарақалпақша	Qaraqalpaq	Cyrl
kac	Kachin	Jinghpaw	Jingpho	Latn
kbd	Kabardian	Къэбэрдей; Адыгэ	East Circassian	Cyrl
kek	Kekchí	Qʼeqchiʼ	Q'eqchi'	Latn
kg	Kongo	Kikongo		Latn
kha	Khasi	khasi		Latn
ki	Kikuyu	Gĩgĩkũyũ	Gikuyu	Latn
kl	Kalaallisut	Kalaallisut	Greenlandic	Latn
kr	Kanuri	Kànùrí		Latn
kri	Krio	Krio	Sierra Leonean Creole	Latn
ks	Kashmiri	كٲشُر	Koshur	Arab
ks-Deva	Kashmiri	कॉशुर		Deva
ktu	Kituba	Kituba		Latn
kv	Komi	Коми		Cyrl
lg	Luganda	Oluganda	Ganda	Latn
li	Limburgan	Limburgs		Latn
lij	Ligurian	Ligure		Latn
lmo	Lombard	lombard		Latn
ln	Lingala	Lingála	Ngala	Latn
ltg	Latgalian	latgalīšu		Latn
lu	Luba-Katanga	Tshiluba		Latn
luo	Luo	Luo	Dholuo	Latn
lus	Mizo	Mizo ṭawng	Lushai, Duhlian	Latn
mad	Madurese	Madurâ		Latn
mag	Magahi	मगही		Deva
mai	Maithili	मैथिली		Deva
mak	Makasar	Mangkasara	Makassarese, Makassar, Macassar	Latn
mam	Mam	Ta yol Mam		Latn
meo	Kedah Malay	Siti		Latn
mfe	Morisien	Morisien	Mauritian Creole	Latn
mh	Marshallese	Majol	Ebon	Latn
min	Minangkabau	Minangkabau		Latn
mni-Mtei	Meiteilon (Manipuri)	ꯃꯤꯇꯩꯂꯣꯟ	Meitei, Meetei, Meitheilon, Meeteilon	Mtei
mos	Mooré	Moaaga; Mòoré	Mossé, Mossi	Latn
ms-Arab	Malay (Jawi script)	ملايو		Arab
mwr	Marwari	मारवाड़ी	Marwadi; Merwari, Mewari (usually ref. dialects)	Deva
nd	North Ndebele	IsiNdebele	Ndebele, isiNdebele saseNyakatho, Zimbabwean Ndebele	Latn
ndc-ZW	Ndau	Cindawu		Latn
new	Newari	नेवाः	Nepal Bhasa, Newar	Deva
nhe	Eastern Huasteca Nahuatl	Nahuatlajtoli Kuextekan ikisayo		Latn
nr	South Ndebele	Isindebele	Transvaal Ndebele	Latn
nso	Sepedi	Sepedi	Pedi, Northern Sotho	Latn
nus	Nuer	Thok Naath		Latn
nv	Navajo	Diné; Naabeehó	Navaho	Latn
oc	Occitan	Occitan	Provençal	Latn
om	Oromo	Afaan Oromoo	Oromiffa, Oromigna, Afaan Oromoo	Latn
os	Ossetian	Ирон	Iron, Ossetic	Cyrl
pa-Arab	Lahnda Punjabi (Pakistan)	لہندا پنجابی		Arab
pag	Pangasinan	Pangasinan		Latn
pam	Pampanga	Kapampangan	Kapampangan	Latn
pap	Papiamento	Papiamentu		Latn
pcm	Nigerian Pidgin	Naijá, Pidgin for Nigerian	Naijá, Pijin, Broken	Latn
qu	Quechua	Runa Simi		Latn
quc	K'iche'	K'iche'	Quiché	Latn
rhg-Latn	Rohingya	Ruháingga		Latn
rn	Rundi	Ikirundi		Latn
rom	Romani	Rromani ćhib	Romany, Romanes, Roma	Latn
sa	Sanskrit	संस्कृतम्		Deva
sah	Yakut	Саха	Yakutian, Sakha, Saqa, Saxa	Cyrl
sat-Latn	Santali, Santhali	Santali	Santhali	Latn
scn	Sicilian	Sicilianu		Latn
se	Northern Sami	davvisámegiella		Latn
sg	Sango	Sango		Latn
shn	Shan	တႆး; လိၵ်ႈတႆး		Mymr
skr	Saraiki	سرائیکی	Multani	Arab
ss	Swati	Siswati		Latn
sus	Susu	Sosokhui		Latn
syl	Sylheti	সিলেটি		Beng
szl	Silesian	Ślůnski		Latn
tcy	Tulu	ತುಳು		Knda
tet	Tetum	Tetun		Latn
ti	Tigrinya	ትግርኛ	Tigrigna	Ethi
tiv	Tiv	Tiv		Latn
tn	Tswana	Setswana	Setswana	Latn
to	Tongan	Tonga		Latn
tpi	Tok Pisin	Tok Pisin	New Guinea Pidgin	Latn
trp	Kok Borok	Kok Borok		Latn
ts	Tsonga	Xitsonga	Xitsonga	Latn
tum	Tumbuka	Chitumbuka		Latn
ty	Tahitian	Reo Tahiti; Tahiti		Latn
tyv	Tuvinian	тувинский		Cyrl
udm	Udmurt	Удмурт кыл		Cyrl
ve	Venda	Tshivenḓa	Tshivenda, Setswetla	Latn
vec	Venetian	Veneto		Latn
war	Waray	Waray	Waray-Waray	Latn
wo	Wolof	Wolof		Latn
yua	Yucateco	Maayat'aan	Yucatec Maya	Latn
yue	Cantonese	廣東話	Yue Chinese (technically a superset)	Hant
zap	Zapotec	Diidxazá		Latn
zza	Zaza	Zazakî	Dimli	Latn

Please note that the language name can be a point of considerable tension, and there is rarely a "right" name for a language! This said, here are some non-exhaustive notes that went into the names in the second column above, based on conversations with speakers. We welcome more comments from the community:

• ak: Calling this language "Twi" instead of "Akan" here is reflective of the corresponding translation models active on Google Translate. These are likely to mix different Akan dialects, including Twi proper, Fante, Akuapem Twi, etc.. Therefore, from a linguistic perspective, one would call this "Akan" and use the corresponding code "ak". However, this is a more academic word for the language; according to native speaker consultants, "Akan" is usually used to describe the ethnic group rather than the language, so they suggested simply calling this "Twi", which is most likely accurate for the large majority of model outputs.
• ckb: Central Kurdish is a more academic name that may not be as widely known to speakers of the language, so Sorani is preferred -- though they are both acceptable.
• dv: There is not a consistent romanization of this language, and both Divehi and Dhivehi are common. The first consonant in this language is a dental, non-aspirated d. When transliterating Indian language names there is no standard about how to represent this sound. In the south, "dh" usually refers to the dental consonant; in the North, it just as often refers to the aspirated one. However, in both cases, it is common that "d" alone would represent the retroflex consonant, so it is possible that Divehi would be mispronounced more often.
• gom: Technically, gom is the code for Goan Konkani only. Other varieties of Konkani, like Mangalore Konkani (which oddly enough does not have an ISO code), are mutually unintelligible! There is certainly dispute about whether to treat Konkani as one language or not.
• ilo: Native speakers give a variety of different answers around the correct name. Some say that Ilocano is more formal, whereas Ilokano is more informal; some say that both of these actually refer to the people, and that Iloko is the only correct term for the language.
• lg: Speakers usually refer to this language as "Luganda"; linguists sometimes call it "Ganda", but like Sepedi, this is not usually used by people familiar with the language.
• lus: There is agreement that Lushai is an obsolete colonial name and should not be used. Mizo is the preferred name.
• mni-Mtei: The conventional and most common name for this language is Manipuri. However, in reality, this is only the language of one of the many ethnicities in Manipur, so that name can be seen as offensive to some. Therefore, we use the endonym Meiteilon, with the conventional name (Manipuri) in parentheses.
• nso: Although a linguist might say that Sepedi is a dialect of Northern Sotho, speakers often say that they call the whole language Sepedi, and that Northern Sotho has offensive colonial overtones. Others, however, may disagree with this. The endonym Sesotho sa Leboa is not uncommon. There is agreement that no-one uses Pedi.
• qu: These translations are technically in Southern Quechua. However, this is a point of some contention. Some will opine that "Quechua" is not a language but a family of languages. Others, however, say that they all feel like one language, and that the regional differences are slight, and it's easy for Peruvians, Bolivians, and even Ecuadorians to understand each other. The majority opinion among the naive speakers we talked to seemed to be closer to the latter, so we have opted for the name "Quechua".
• ti: In Tigray Tigrigna is the most common Romanization, and in Eritrea Tigrinya is.
• zap: Linguists usually talk about the Zapotec languages, rather than considering Zapotec a single language. However, just as with Quechua above, conversations with a native speaker indicated that they perceived Zapotec as one language, with a variety of local variants that are all easy enough to understand (excepting one or two). A broader survey would help resolve this question.
• Those languages in this dataset that are not on Google Translate have yet to undergo a more careful name review; if you see a name in the list above that you want to correct, please contact us!

Note on language codes:

• we use chm for Meadow Mari, although chm is properly the superset, and mhr is the individual code for Meadow Mari.
• Similarly, we use rom for all Romani dialects, ak for Twi (see note above), and ff for all Fulfulde dialects.
• We use fa-AF and ar-MA for Dari and Darija respectively. The specific codes prs and ary are probably more precise.
• apd-SD and ndc-ZW are probably redundant; apd and ndc alone refer to Sudanese Arabic and Ndau resp.

Special note on Bantu languages: Bantu language names have a special noun-class prefix for languages, like xiTsonga, sePedi, kiSwahili, isiZulu, luGanda, siSwati, isiNdebele, tshiVenda, and so on. However, there is no pattern as to whether this prefix is used in the conventional English form of the name. For instance, Xhosa usually leaves off the isi- in English, whereas Lingala almost never leaves off the li-. Others, like (xi)Tsonga, seem to be frequently used in both forms. Linguistic resources occasionally strip the prefixes, yielding names like "Ganda", "Pedi" and so on, perhaps out of a desire for consistent handling of the noun class prefix. (Even the ba in Bantu is such a prefix, so these languages are sometimes called the Ntu languages!)

Disclaimer: This is not an officially supported Google product.

Contact

We welcome contributions, corrections, and suggestions, especially from speakers of these languages. Please file bugs on this repo or contact icaswell꩜google.com if you have suggestions or corrections. This repository may be updated sporadically with corrections or additional translations.

License

This data is released with the CC-BY-4.0 license.

Citation

@misc{jones2023bilexrxlexicaldata,
      title={Bilex Rx: Lexical Data Augmentation for Massively Multilingual Machine Translation}, 
      author={Alex Jones and Isaac Caswell and Ishank Saxena and Orhan Firat},
      year={2023},
      eprint={2303.15265},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2303.15265}, 
}