Dataset Card for XGLUE

Dataset Summary

XGLUE is a new benchmark dataset to evaluate the performance of cross-lingual pre-trained models with respect to cross-lingual natural language understanding and generation.

The training data of each task is in English while the validation and test data is present in multiple different languages. The following table shows which languages are present as validation and test data for each config.

Available Languages for Test and Validation Data

Therefore, for each config, a cross-lingual pre-trained model should be fine-tuned on the English training data, and evaluated on for all languages.

Leaderboards

The XGLUE leaderboard can be found on the homepage and consits of a XGLUE-Understanding Score (the average of the tasks ner, pos, mlqa, nc, xnli, paws-x, qadsm, wpr, qam) and a XGLUE-Generation Score (the average of the tasks qg, ntg).

Dataset Structure

Data Instances

ner

An example of 'test.nl' looks as follows.

{
  "ner": [
    "O",
    "O",
    "O",
    "B-LOC",
    "O",
    "B-LOC",
    "O",
    "B-LOC",
    "O",
    "O",
    "O",
    "O",
    "O",
    "O",
    "O",
    "B-PER",
    "I-PER",
    "O",
    "O",
    "B-LOC",
    "O",
    "O"
  ],
  "words": [
    "Dat",
    "is",
    "in",
    "Itali\u00eb",
    ",",
    "Spanje",
    "of",
    "Engeland",
    "misschien",
    "geen",
    "probleem",
    ",",
    "maar",
    "volgens",
    "'",
    "Der",
    "Kaiser",
    "'",
    "in",
    "Duitsland",
    "wel",
    "."
  ]
}

pos

An example of 'test.fr' looks as follows.

{
  "pos": [
    "PRON",
    "VERB",
    "SCONJ",
    "ADP",
    "PRON",
    "CCONJ",
    "DET",
    "NOUN",
    "ADP",
    "NOUN",
    "CCONJ",
    "NOUN",
    "ADJ",
    "PRON",
    "PRON",
    "AUX",
    "ADV",
    "VERB",
    "PUNCT",
    "PRON",
    "VERB",
    "VERB",
    "DET",
    "ADJ",
    "NOUN",
    "ADP",
    "DET",
    "NOUN",
    "PUNCT"
  ],
  "words": [
    "Je",
    "sens",
    "qu'",
    "entre",
    "\u00e7a",
    "et",
    "les",
    "films",
    "de",
    "m\u00e9decins",
    "et",
    "scientifiques",
    "fous",
    "que",
    "nous",
    "avons",
    "d\u00e9j\u00e0",
    "vus",
    ",",
    "nous",
    "pourrions",
    "emprunter",
    "un",
    "autre",
    "chemin",
    "pour",
    "l'",
    "origine",
    "."
  ]
}

mlqa

An example of 'test.hi' looks as follows.

{
  "answers": {
    "answer_start": [
      378
    ],
    "text": [
      "\u0909\u0924\u094d\u0924\u0930 \u092a\u0942\u0930\u094d\u0935"
    ]
  },
  "context": "\u0909\u0938\u0940 \"\u090f\u0930\u093f\u092f\u093e XX \" \u0928\u093e\u092e\u0915\u0930\u0923 \u092a\u094d\u0930\u0923\u093e\u0932\u0940 \u0915\u093e \u092a\u094d\u0930\u092f\u094b\u0917 \u0928\u0947\u0935\u093e\u0926\u093e \u092a\u0930\u0940\u0915\u094d\u0937\u0923 \u0938\u094d\u0925\u0932 \u0915\u0947 \u0905\u0928\u094d\u092f \u092d\u093e\u0917\u094b\u0902 \u0915\u0947 \u0932\u093f\u090f \u0915\u093f\u092f\u093e \u0917\u092f\u093e \u0939\u0948\u0964\u092e\u0942\u0932 \u0930\u0942\u092a \u092e\u0947\u0902 6 \u092c\u091f\u0947 10 \u092e\u0940\u0932 \u0915\u093e \u092f\u0939 \u0906\u092f\u0924\u093e\u0915\u093e\u0930 \u0905\u0921\u094d\u0921\u093e \u0905\u092c \u0924\u0925\u093e\u0915\u0925\u093f\u0924 '\u0917\u094d\u0930\u0942\u092e \u092c\u0949\u0915\u094d\u0938 \" \u0915\u093e \u090f\u0915 \u092d\u093e\u0917 \u0939\u0948, \u091c\u094b \u0915\u093f 23 \u092c\u091f\u0947 25.3 \u092e\u0940\u0932 \u0915\u093e \u090f\u0915 \u092a\u094d\u0930\u0924\u093f\u092c\u0902\u0927\u093f\u0924 \u0939\u0935\u093e\u0908 \u0915\u094d\u0937\u0947\u0924\u094d\u0930 \u0939\u0948\u0964 \u092f\u0939 \u0915\u094d\u0937\u0947\u0924\u094d\u0930 NTS \u0915\u0947 \u0906\u0902\u0924\u0930\u093f\u0915 \u0938\u0921\u093c\u0915 \u092a\u094d\u0930\u092c\u0902\u0927\u0928 \u0938\u0947 \u091c\u0941\u0921\u093c\u093e \u0939\u0948, \u091c\u093f\u0938\u0915\u0940 \u092a\u0915\u094d\u0915\u0940 \u0938\u0921\u093c\u0915\u0947\u0902 \u0926\u0915\u094d\u0937\u093f\u0923 \u092e\u0947\u0902 \u092e\u0930\u0915\u0930\u0940 \u0915\u0940 \u0913\u0930 \u0914\u0930 \u092a\u0936\u094d\u091a\u093f\u092e \u092e\u0947\u0902 \u092f\u0941\u0915\u094d\u0915\u093e \u092b\u094d\u0932\u0948\u091f \u0915\u0940 \u0913\u0930 \u091c\u093e\u0924\u0940 \u0939\u0948\u0902\u0964 \u091d\u0940\u0932 \u0938\u0947 \u0909\u0924\u094d\u0924\u0930 \u092a\u0942\u0930\u094d\u0935 \u0915\u0940 \u0913\u0930 \u092c\u0922\u093c\u0924\u0947 \u0939\u0941\u090f \u0935\u094d\u092f\u093e\u092a\u0915 \u0914\u0930 \u0914\u0930 \u0938\u0941\u0935\u094d\u092f\u0935\u0938\u094d\u0925\u093f\u0924 \u0917\u094d\u0930\u0942\u092e \u091d\u0940\u0932 \u0915\u0940 \u0938\u0921\u093c\u0915\u0947\u0902 \u090f\u0915 \u0926\u0930\u094d\u0930\u0947 \u0915\u0947 \u091c\u0930\u093f\u092f\u0947 \u092a\u0947\u091a\u0940\u0926\u093e \u092a\u0939\u093e\u0921\u093c\u093f\u092f\u094b\u0902 \u0938\u0947 \u0939\u094b\u0915\u0930 \u0917\u0941\u091c\u0930\u0924\u0940 \u0939\u0948\u0902\u0964 \u092a\u0939\u0932\u0947 \u0938\u0921\u093c\u0915\u0947\u0902 \u0917\u094d\u0930\u0942\u092e \u0918\u093e\u091f\u0940",
  "question": "\u091d\u0940\u0932 \u0915\u0947 \u0938\u093e\u092a\u0947\u0915\u094d\u0937 \u0917\u094d\u0930\u0942\u092e \u0932\u0947\u0915 \u0930\u094b\u0921 \u0915\u0939\u093e\u0901 \u091c\u093e\u0924\u0940 \u0925\u0940?"
}

nc

An example of 'test.es' looks as follows.

{
  "news_body": "El bizcocho es seguramente el producto m\u00e1s b\u00e1sico y sencillo de toda la reposter\u00eda : consiste en poco m\u00e1s que mezclar unos cuantos ingredientes, meterlos al horno y esperar a que se hagan. Por obra y gracia del impulsor qu\u00edmico, tambi\u00e9n conocido como \"levadura de tipo Royal\", despu\u00e9s de un rato de calorcito esta combinaci\u00f3n de harina, az\u00facar, huevo, grasa -aceite o mantequilla- y l\u00e1cteo se transforma en uno de los productos m\u00e1s deliciosos que existen para desayunar o merendar . Por muy manazas que seas, es m\u00e1s que probable que tu bizcocho casero supere en calidad a cualquier infamia industrial envasada. Para lograr un bizcocho digno de admiraci\u00f3n s\u00f3lo tienes que respetar unas pocas normas que afectan a los ingredientes, proporciones, mezclado, horneado y desmoldado. Todas las tienes resumidas en unos dos minutos el v\u00eddeo de arriba, en el que adem \u00e1s aprender\u00e1s alg\u00fan truquillo para que tu bizcochaco quede m\u00e1s fino, jugoso, esponjoso y amoroso. M\u00e1s en MSN:",
  "news_category": "foodanddrink",
  "news_title": "Cocina para lerdos: las leyes del bizcocho"
}

xnli

An example of 'validation.th' looks as follows.

{
  "hypothesis": "\u0e40\u0e02\u0e32\u0e42\u0e17\u0e23\u0e2b\u0e32\u0e40\u0e40\u0e21\u0e48\u0e02\u0e2d\u0e07\u0e40\u0e02\u0e32\u0e2d\u0e22\u0e48\u0e32\u0e07\u0e23\u0e27\u0e14\u0e40\u0e23\u0e47\u0e27\u0e2b\u0e25\u0e31\u0e07\u0e08\u0e32\u0e01\u0e17\u0e35\u0e48\u0e23\u0e16\u0e42\u0e23\u0e07\u0e40\u0e23\u0e35\u0e22\u0e19\u0e2a\u0e48\u0e07\u0e40\u0e02\u0e32\u0e40\u0e40\u0e25\u0e49\u0e27",
  "label": 1,
  "premise": "\u0e41\u0e25\u0e30\u0e40\u0e02\u0e32\u0e1e\u0e39\u0e14\u0e27\u0e48\u0e32, \u0e21\u0e48\u0e32\u0e21\u0e4a\u0e32 \u0e1c\u0e21\u0e2d\u0e22\u0e39\u0e48\u0e1a\u0e49\u0e32\u0e19"
}

paws-x

An example of 'test.es' looks as follows.

{
  "label": 1,
  "sentence1": "La excepci\u00f3n fue entre fines de 2005 y 2009 cuando jug\u00f3 en Suecia con Carlstad United BK, Serbia con FK Borac \u010ca\u010dak y el FC Terek Grozny de Rusia.",
  "sentence2": "La excepci\u00f3n se dio entre fines del 2005 y 2009, cuando jug\u00f3 con Suecia en el Carlstad United BK, Serbia con el FK Borac \u010ca\u010dak y el FC Terek Grozny de Rusia."
}

qadsm

An example of 'train' looks as follows.

{
  "ad_description": "Your New England Cruise Awaits! Holland America Line Official Site.",
  "ad_title": "New England Cruises",
  "query": "cruise portland maine",
  "relevance_label": 1
}

wpr

An example of 'test.zh' looks as follows.

{
  "query": "maxpro\u5b98\u7f51",
  "relavance_label": 0,
  "web_page_snippet": "\u5728\u7ebf\u8d2d\u4e70\uff0c\u552e\u540e\u670d\u52a1\u3002vivo\u667a\u80fd\u624b\u673a\u5f53\u5b63\u660e\u661f\u673a\u578b\u6709NEX\uff0cvivo X21\uff0cvivo X20\uff0c\uff0cvivo X23\u7b49\uff0c\u5728vivo\u5b98\u7f51\u8d2d\u4e70\u624b\u673a\u53ef\u4ee5\u4eab\u53d712 \u671f\u514d\u606f\u4ed8\u6b3e\u3002 \u54c1\u724c Funtouch OS \u4f53\u9a8c\u5e97 | ...",
  "wed_page_title": "vivo\u667a\u80fd\u624b\u673a\u5b98\u65b9\u7f51\u7ad9-AI\u975e\u51e1\u6444\u5f71X23"
}

qam

An example of 'validation.en' looks as follows.

{
  "annswer": "Erikson has stated that after the last novel of the Malazan Book of the Fallen was finished, he and Esslemont would write a comprehensive guide tentatively named The Encyclopaedia Malazica.",
  "label": 0,
  "question": "main character of malazan book of the fallen"
}

qg

An example of 'test.de' looks as follows.

{
  "answer_passage": "Medien bei WhatsApp automatisch speichern. Tippen Sie oben rechts unter WhatsApp auf die drei Punkte oder auf die Men\u00fc-Taste Ihres Smartphones. Dort wechseln Sie in die \"Einstellungen\" und von hier aus weiter zu den \"Chat-Einstellungen\". Unter dem Punkt \"Medien Auto-Download\" k\u00f6nnen Sie festlegen, wann die WhatsApp-Bilder heruntergeladen werden sollen.",
  "question": "speichenn von whats app bilder unterbinden"
}

ntg

An example of 'test.en' looks as follows.

{
  "news_body": "Check out this vintage Willys Pickup! As they say, the devil is in the details, and it's not every day you see such attention paid to every last area of a restoration like with this 1961 Willys Pickup . Already the Pickup has a unique look that shares some styling with the Jeep, plus some original touches you don't get anywhere else. It's a classy way to show up to any event, all thanks to Hollywood Motors . A burgundy paint job contrasts with white lower panels and the roof. Plenty of tasteful chrome details grace the exterior, including the bumpers, headlight bezels, crossmembers on the grille, hood latches, taillight bezels, exhaust finisher, tailgate hinges, etc. Steel wheels painted white and chrome hubs are a tasteful addition. Beautiful oak side steps and bed strips add a touch of craftsmanship to this ride. This truck is of real showroom quality, thanks to the astoundingly detailed restoration work performed on it, making this Willys Pickup a fierce contender for best of show. Under that beautiful hood is a 225 Buick V6 engine mated to a three-speed manual transmission, so you enjoy an ideal level of control. Four wheel drive is functional, making it that much more utilitarian and downright cool. The tires are new, so you can enjoy a lot of life out of them, while the wheels and hubs are in great condition. Just in case, a fifth wheel with a tire and a side mount are included. Just as important, this Pickup runs smoothly, so you can go cruising or even hit the open road if you're interested in participating in some classic rallies. You might associate Willys with the famous Jeep CJ, but the automaker did produce a fair amount of trucks. The Pickup is quite the unique example, thanks to distinct styling that really turns heads, making it a favorite at quite a few shows. Source: Hollywood Motors Check These Rides Out Too: Fear No Trails With These Off-Roaders 1965 Pontiac GTO: American Icon For Sale In Canada Low-Mileage 1955 Chevy 3100 Represents Turn In Pickup Market",
  "news_title": "This 1961 Willys Pickup Will Let You Cruise In Style"
}

Data Fields

ner

In the following each data field in ner is explained. The data fields are the same among all splits.

  • words: a list of words composing the sentence.
  • ner: a list of entitity classes corresponding to each word respectively.

pos

In the following each data field in pos is explained. The data fields are the same among all splits.

  • words: a list of words composing the sentence.
  • pos: a list of "part-of-speech" classes corresponding to each word respectively.

mlqa

In the following each data field in mlqa is explained. The data fields are the same among all splits.

  • context: a string, the context containing the answer.
  • question: a string, the question to be answered.
  • answers: a string, the answer to question.

nc

In the following each data field in nc is explained. The data fields are the same among all splits.

  • news_title: a string, to the title of the news report.
  • news_body: a string, to the actual news report.
  • news_category: a string, the category of the news report, e.g. foodanddrink

xnli

In the following each data field in xnli is explained. The data fields are the same among all splits.

  • premise: a string, the context/premise, i.e. the first sentence for natural language inference.
  • hypothesis: a string, a sentence whereas its relation to premise is to be classified, i.e. the second sentence for natural language inference.
  • label: a class catory (int), natural language inference relation class between hypothesis and premise. One of 0: entailment, 1: contradiction, 2: neutral.

paws-x

In the following each data field in paws-x is explained. The data fields are the same among all splits.

  • sentence1: a string, a sentence.
  • sentence2: a string, a sentence whereas the sentence is either a paraphrase of sentence1 or not.
  • label: a class label (int), whether sentence2 is a paraphrase of sentence1 One of 0: different, 1: same.

qadsm

In the following each data field in qadsm is explained. The data fields are the same among all splits.

  • query: a string, the search query one would insert into a search engine.
  • ad_title: a string, the title of the advertisement.
  • ad_description: a string, the content of the advertisement, i.e. the main body.
  • relevance_label: a class label (int), how relevant the advertisement ad_title + ad_description is to the search query query. One of 0: Bad, 1: Good.

wpr

In the following each data field in wpr is explained. The data fields are the same among all splits.

  • query: a string, the search query one would insert into a search engine.
  • web_page_title: a string, the title of a web page.
  • web_page_snippet: a string, the content of a web page, i.e. the main body.
  • relavance_label: a class label (int), how relevant the web page web_page_snippet + web_page_snippet is to the search query query. One of 0: Bad, 1: Fair, 2: Good, 3: Excellent, 4: Perfect.

qam

In the following each data field in qam is explained. The data fields are the same among all splits.

  • question: a string, a question.
  • answer: a string, a possible answer to question.
  • label: a class label (int), whether the answer is relevant to the question. One of 0: False, 1: True.

qg

In the following each data field in qg is explained. The data fields are the same among all splits.

  • answer_passage: a string, a detailed answer to the question.
  • question: a string, a question.

ntg

In the following each data field in ntg is explained. The data fields are the same among all splits.

  • news_body: a string, the content of a news article.
  • news_title: a string, the title corresponding to the news article news_body.

Data Splits

ner

The following table shows the number of data samples/number of rows for each split in ner.

train validation.en validation.de validation.es validation.nl test.en test.de test.es test.nl
ner 14042 3252 2874 1923 2895 3454 3007 1523 5202

pos

The following table shows the number of data samples/number of rows for each split in pos.

train validation.en validation.de validation.es validation.nl validation.bg validation.el validation.fr validation.pl validation.tr validation.vi validation.zh validation.ur validation.hi validation.it validation.ar validation.ru validation.th test.en test.de test.es test.nl test.bg test.el test.fr test.pl test.tr test.vi test.zh test.ur test.hi test.it test.ar test.ru test.th
pos 25376 2001 798 1399 717 1114 402 1475 2214 987 799 499 551 1658 563 908 578 497 2076 976 425 595 1115 455 415 2214 982 799 499 534 1683 481 679 600 497

mlqa

The following table shows the number of data samples/number of rows for each split in mlqa.

train validation.en validation.de validation.ar validation.es validation.hi validation.vi validation.zh test.en test.de test.ar test.es test.hi test.vi test.zh
mlqa 87599 1148 512 517 500 507 511 504 11590 4517 5335 5253 4918 5495 5137

nc

The following table shows the number of data samples/number of rows for each split in nc.

train validation.en validation.de validation.es validation.fr validation.ru test.en test.de test.es test.fr test.ru
nc 100000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000

xnli

The following table shows the number of data samples/number of rows for each split in xnli.

train validation.en validation.ar validation.bg validation.de validation.el validation.es validation.fr validation.hi validation.ru validation.sw validation.th validation.tr validation.ur validation.vi validation.zh test.en test.ar test.bg test.de test.el test.es test.fr test.hi test.ru test.sw test.th test.tr test.ur test.vi test.zh
xnli 392702 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010

The following table shows the number of data samples/number of rows for each split in mlqa.

| |train|validation.en|validation.de|validation.ar|validation.es|validation.hi|validation.vi|validation.zh|test.en|test.de|test.ar|test.es|test.hi|test.vi|test.zh|
|----|----:|------------:|------------:|------------:|------------:|------------:|------------:|------------:|------:|------:|------:|------:|------:|------:|------:|
|mlqa|87599| 1148| 512| 517| 500| 507| 511| 504| 11590| 4517| 5335| 5253| 4918| 5495| 5137|

nc

The following table shows the number of data samples/number of rows for each split in nc.

train validation.en validation.de validation.es validation.fr validation.ru test.en test.de test.es test.fr test.ru
nc 100000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000

xnli

The following table shows the number of data samples/number of rows for each split in xnli.

train validation.en validation.ar validation.bg validation.de validation.el validation.es validation.fr validation.hi validation.ru validation.sw validation.th validation.tr validation.ur validation.vi validation.zh test.en test.ar test.bg test.de test.el test.es test.fr test.hi test.ru test.sw test.th test.tr test.ur test.vi test.zh
xnli 392702 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 2490 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010 5010

paws-x

The following table shows the number of data samples/number of rows for each split in paws-x.

train validation.en validation.de validation.es validation.fr test.en test.de test.es test.fr
paws-x 49401 2000 2000 2000 2000 2000 2000 2000 2000

qadsm

The following table shows the number of data samples/number of rows for each split in qadsm.

train validation.en validation.de validation.fr test.en test.de test.fr
qadsm 100000 10000 10000 10000 10000 10000 10000

wpr

The following table shows the number of data samples/number of rows for each split in wpr.

train validation.en validation.de validation.es validation.fr validation.it validation.pt validation.zh test.en test.de test.es test.fr test.it test.pt test.zh
wpr 99997 10008 10004 10004 10005 10003 10001 10002 10004 9997 10006 10020 10001 10015 9999

qam

The following table shows the number of data samples/number of rows for each split in qam.

train validation.en validation.de validation.fr test.en test.de test.fr
qam 100000 10000 10000 10000 10000 10000 10000

qg

The following table shows the number of data samples/number of rows for each split in qg.

train validation.en validation.de validation.es validation.fr validation.it validation.pt test.en test.de test.es test.fr test.it test.pt
qg 100000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000

ntg

The following table shows the number of data samples/number of rows for each split in ntg.

train validation.en validation.de validation.es validation.fr validation.ru test.en test.de test.es test.fr test.ru
ntg 300000 10000 10000 10000 10000 10000 10000 10000 10000 10000 10000

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

[More Information Needed]

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

The dataset is maintained mainly by Yaobo Liang, Yeyun Gong, Nan Duan, Ming Gong, Linjun Shou, and Daniel Campos from Microsoft Research.

Licensing Information

The licensing status of the dataset hinges on the legal status of XGLUE hich is unclear.

Citation Information

@article{Liang2020XGLUEAN,
  title={XGLUE: A New Benchmark Dataset for Cross-lingual Pre-training, Understanding and Generation},
  author={Yaobo Liang and Nan Duan and Yeyun Gong and Ning Wu and Fenfei Guo and Weizhen Qi and Ming Gong and Linjun Shou and Daxin Jiang and Guihong Cao and Xiaodong Fan and Ruofei Zhang and Rahul Agrawal and Edward Cui and Sining Wei and Taroon Bharti and Ying Qiao and Jiun-Hung Chen and Winnie Wu and Shuguang Liu and Fan Yang and Daniel Campos and Rangan Majumder and Ming Zhou},
  journal={arXiv},
  year={2020},
  volume={abs/2004.01401}
}

Contributions

Thanks to @patrickvonplaten for adding this dataset.

Models trained or fine-tuned on xglue

None yet