mqa / README.md
maximedb's picture
Update README.md (#2)
5218028
metadata
annotations_creators:
  - no-annotation
language_creators:
  - other
language:
  - ca
  - en
  - de
  - es
  - fr
  - ru
  - ja
  - it
  - zh
  - pt
  - nl
  - tr
  - pl
  - vi
  - ar
  - id
  - uk
  - ro
  - 'no'
  - th
  - sv
  - el
  - fi
  - he
  - da
  - cs
  - ko
  - fa
  - hi
  - hu
  - sk
  - lt
  - et
  - hr
  - is
  - lv
  - ms
  - bg
  - sr
  - ca
license:
  - cc0-1.0
multilinguality:
  - multilingual
pretty_name: MQA - a Multilingual FAQ and CQA Dataset
size_categories:
  - unknown
source_datasets:
  - original
task_categories:
  - question-answering
task_ids:
  - multiple-choice-qa

MQA

MQA is a Multilingual corpus of Questions and Answers (MQA) parsed from the Common Crawl. Questions are divided in two types: Frequently Asked Questions (FAQ) and Community Question Answering (CQA).

from datasets import load_dataset
all_data = load_dataset("clips/mqa", language="en")
{
  "name": "the title of the question (if any)",
  "text": "the body of the question (if any)",
  "answers": [{
    "text": "the text of the answer",
    "is_accepted": "true|false"
  }]
}
faq_data = load_dataset("clips/mqa", scope="faq", language="en")
cqa_data = load_dataset("clips/mqa", scope="cqa", language="en")

Languages

We collected around 234M pairs of questions and answers in 39 languages. To download a language specific subset you need to specify the language key as configuration. See below for an example.

load_dataset("clips/mqa", language="en") # replace "en" by any language listed below
Language FAQ CQA
en 174,696,414 14,082,180
de 17,796,992 1,094,606
es 14,967,582 845,836
fr 13,096,727 1,299,359
ru 12,435,022 1,715,131
it 6,850,573 455,027
ja 6,369,706 2,089,952
zh 5,940,796 579,596
pt 5,851,286 373,982
nl 4,882,511 503,376
tr 3,893,964 370,975
pl 3,766,531 70,559
vi 2,795,227 96,528
id 2,253,070 200,441
ar 2,211,795 805,661
uk 2,090,611 27,260
el 1,758,618 17,167
no 1,752,820 11,786
sv 1,733,582 20,024
fi 1,717,221 41,371
ro 1,689,471 93,222
th 1,685,463 73,204
da 1,554,581 16,398
he 1,422,449 88,435
ko 1,361,901 49,061
cs 1,224,312 143,863
hu 878,385 27,639
fa 787,420 118,805
sk 785,101 4,615
lt 672,105 301
et 547,208 441
hi 516,342 205,645
hr 458,958 11,677
is 437,748 37
lv 428,002 88
ms 230,568 7,460
bg 198,671 5,320
sr 110,270 3,980
ca 100,201 1,914

FAQ vs. CQA

You can download the Frequently Asked Questions (FAQ) or the Community Question Answering (CQA) part of the dataset.

faq = load_dataset("clips/mqa", scope="faq")
cqa = load_dataset("clips/mqa", scope="cqa")
all = load_dataset("clips/mqa", scope="all")

Although FAQ and CQA questions share the same structure, CQA questions can have multiple answers for a given questions, while FAQ questions have a single answer. FAQ questions typically only have a title (name key), while CQA have a title and a body (name and text).

Nesting and Data Fields

You can specify three different nesting level: question, page and domain.

Question

load_dataset("clips/mqa", level="question") # default

The default level is the question object:

  • name: the title of the question(if any) in markdown format
  • text: the body of the question (if any) in markdown format
  • answers: a list of answers
    • text: the title of the answer (if any) in markdown format
    • name: the body of the answer in markdown format
    • is_accepted: true if the answer is selected.

Page

This level returns a list of questions present on the same page. This is mostly useful for FAQs since CQAs already have one question per page.

load_dataset("clips/mqa", level="page")

Domain

This level returns a list of pages present on the web domain. This is a good way to cope with FAQs duplication by sampling one page per domain at each epoch.

load_dataset("clips/mqa", level="domain")

Source Data

This section was adapted from the source data description of OSCAR

Common Crawl is a non-profit foundation which produces and maintains an open repository of web crawled data that is both accessible and analysable. Common Crawl's complete web archive consists of petabytes of data collected over 8 years of web crawling. The repository contains raw web page HTML data (WARC files), metdata extracts (WAT files) and plain text extracts (WET files). The organisation's crawlers has always respected nofollow and robots.txt policies.

To construct MQA, we used the WARC files of Common Crawl.

People

This model was developed by Maxime De Bruyn, Ehsan Lotfi, Jeska Buhmann and Walter Daelemans.

Licensing Information

These data are released under this licensing scheme.
We do not own any of the text from which these data has been extracted.
We license the actual packaging of these data under the Creative Commons CC0 license ("no rights reserved") http://creativecommons.org/publicdomain/zero/1.0/

Should you consider that our data contains material that is owned by you and should therefore not be reproduced here, please:
* Clearly identify yourself, with detailed contact data such as an address, telephone number or email address at which you can be contacted.
* Clearly identify the copyrighted work claimed to be infringed.
* Clearly identify the material that is claimed to be infringing and information reasonably sufficient to allow us to locate the material.

We will comply to legitimate requests by removing the affected sources from the next release of the corpus.

Citation information

@inproceedings{de-bruyn-etal-2021-mfaq,
    title = "{MFAQ}: a Multilingual {FAQ} Dataset",
    author = "De Bruyn, Maxime  and
      Lotfi, Ehsan  and
      Buhmann, Jeska  and
      Daelemans, Walter",
    booktitle = "Proceedings of the 3rd Workshop on Machine Reading for Question Answering",
    month = nov,
    year = "2021",
    address = "Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.mrqa-1.1",
    pages = "1--13",
}