annotations_creators:
- expert-generated
language_creators:
- found
language:
- es
license:
- cc-by-sa-4.0
multilinguality:
- monolingual
pretty_name: Spanish Question Answering Corpus (SQAC)
size_categories:
- unknown
source_datasets:
- original
task_categories:
- question-answering
task_ids:
- extractive-qa
SQAC (Spanish Question-Answering Corpus): An extractive QA dataset for the Spanish language
BibTeX citation
@article{gutierrezfandino2022,
author = {Asier Gutiérrez-Fandiño and Jordi Armengol-Estapé and Marc Pàmies and Joan Llop-Palao and Joaquin Silveira-Ocampo and Casimiro Pio Carrino and Carme Armentano-Oller and Carlos Rodriguez-Penagos and Aitor Gonzalez-Agirre and Marta Villegas},
title = {MarIA: Spanish Language Models},
journal = {Procesamiento del Lenguaje Natural},
volume = {68},
number = {0},
year = {2022},
issn = {1989-7553},
url = {http://journal.sepln.org/sepln/ojs/ojs/index.php/pln/article/view/6405},
pages = {39--60}
}
See the pre-print version of our paper for further details: https://arxiv.org/abs/2107.07253
Introduction
This dataset contains 6,247 contexts and 18,817 questions with their answers, 1 to 5 for each fragment.
The sources of the contexts are:
- Encyclopedic articles from Wikipedia in Spanish, used under CC-by-sa licence.
- News from Wikinews in Spanish, used under CC-by licence.
- Text from the Spanish corpus AnCora, which is a mix from diferent newswire and literature sources, used under CC-by licence.
This dataset can be used to build extractive-QA.
Supported Tasks and Leaderboards
Extractive-QA
Languages
ES - Spanish
Directory structure
- README.md
- dev.json
- test.json
- train.json
- sqac.py
Dataset Structure
Data Instances
JSON files
Data Fields
Follows (Rajpurkar, Pranav et al., 2016) for squad v1 datasets. (see below for full reference). We added a field "source" with the source of the context.
Example
{
"data": [
{
"paragraphs": [
{
"context": "Al cogote, y fumando como una cafetera. Ah!, no era él, éramos todos nosotros. Luego llegó Billie Holiday. Bajo el epígrafe Arte, la noche temática, pasaron la vida de la única cantante del universo que no es su voz, sino su alma lo que se escucha cuando interpreta. Gata golpeada por el mundo, pateada, violada, enganchada a todos los paraísos artificiales del planeta, jamás encontró el Edén. El Edén lo encontramos nosotros cuando, al concluir la sesión de la tele, pusimos en la doméstica cadena de sonido el mítico Last Recording, su última grabación (marzo de 1959), con la orquesta de Ray Ellis y el piano de Hank Jones. Se estaba muriendo Lady Day, y no obstante, mientras moría, su alma cantaba, Baby, won't you please come home. O sea, niño, criatura, amor, vuelve, a casa por favor.",
"qas": [
{
"question": "¿Quién se incorporó a la reunión más adelante?",
"id": "c5429572-64b8-4c5d-9553-826f867b07be",
"answers": [
{
"answer_start": 91,
"text": "Billie Holiday"
}
]
},
...
]
}
],
"title": "P_129_20010702_&_P_154_20010102_&_P_108_20000301_c_&_P_108_20000601_d",
"source": "ancora"
},
...
]
}
Data Splits
- train
- development
- test
Content analysis
Number of articles, paragraphs and questions
- Number of articles: 3,834
- Number of contexts: 6,247
- Number of questions: 18,817
- Questions/context: 3.01
- Number of sentences: 48,026
- Sentences/context: 7.70
Number of tokens
- Total tokens in context: 1,561,616
- Tokens/context 250.30
- Total tokens in questions: 203,235
- Tokens in questions/questions: 10.80
- Tokens in questions/tokens in context: 0.13
- Total tokens in answers: 90,307
- Tokens in answers/answers: 4.80
- Tokens in answers/tokens in context: 0.06
Lexical variation
46.38 of the words in the Question can be found in the Context.
Question type
| Question | Count | % |
|---|---|---|
| qué | 6,381 | 33.91 % |
| quién/es | 2,952 | 15.69 % |
| cuál/es | 2,034 | 10.81 % |
| cómo | 1,949 | 10.36 % |
| dónde | 1,856 | 9.86 % |
| cuándo | 1,639 | 8.71 % |
| cuánto | 1,311 | 6.97 % |
| cuántos | 495 | 2.63 % |
| adónde | 100 | 0.53 % |
| cuánta | 49 | 0.26 % |
| no question mark | 43 | 0.23 % |
| cuántas | 19 | 0.10 % |
Dataset Creation
Methodology
6,247 contexts were randomly chosen from the three corpus described below. We commisioned the creation of between 1 and 5 questions for each context, following an adaptation of the guidelines from SQUAD 1.0 Rajpurkar, Pranav et al. “SQuAD: 100, 000+ Questions for Machine Comprehension of Text.” EMNLP (2016). In total, 18,817 pairs of a question and an extracted fragment that contains the answer were created.
Curation Rationale
For compatibility with similar datasets in other languages, we followed as close as possible existing curation guidelines. We also created another QA dataset with Wikipedia to ensure thematic and stylistic variety.
Source Data
- Spanish Wikipedia: https://es.wikipedia.org
- Spanish Wikinews: https://es.wikinews.org/
- AnCora corpus: http://clic.ub.edu/corpus/en
Initial Data Collection and Normalization
The source data are scraped articles from the Spanish Wikipedia site, Wikinews site and from AnCora corpus.
Who are the source language producers?
[More Information Needed]
Annotations
Annotation process
We commissioned the creation of 1 to 5 questions for each context, following an adaptation of the guidelines from SQUAD 1.0 Rajpurkar, Pranav et al. “SQuAD: 100, 000+ Questions for Machine Comprehension of Text.” EMNLP (2016).
Who are the annotators?
Native language speakers.
Dataset Curators
Carlos Rodríguez and Carme Armentano, from BSC-CNS.
Personal and Sensitive Information
No personal or sensitive information included.
Considerations for Using the Data
Social Impact of Dataset
[More Information Needed]
Discussion of Biases
[More Information Needed]
Other Known Limitations
[More Information Needed]
Contact
Carlos Rodríguez-Penagos or Carme Armentano-Oller (plantl-gob-es@bsc.es)
Funding
This work was funded by the Spanish State Secretariat for Digitalization and Artificial Intelligence (SEDIA) within the framework of the Plan-TL.
License
This work is licensed under a Attribution-ShareAlike 4.0 International License.