Datasets:

ArXiv:
License:
The Dataset Viewer has been disabled on this dataset.

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 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}, 
}
Downloads last month
91