Mintaka is a complex, natural, and multilingual question answering (QA) dataset composed of 20,000 question-answer pairs elicited from MTurk workers and annotated with Wikidata question and answer entities. Full details on the Mintaka dataset can be found in our paper: https://aclanthology.org/2022.coling-1.138/
To build Mintaka, we explicitly collected questions in 8 complexity types, as well as generic questions:
- Count (e.g., Q: How many astronauts have been elected to Congress? A: 4)
- Comparative (e.g., Q: Is Mont Blanc taller than Mount Rainier? A: Yes)
- Superlative (e.g., Q: Who was the youngest tribute in the Hunger Games? A: Rue)
- Ordinal (e.g., Q: Who was the last Ptolemaic ruler of Egypt? A: Cleopatra)
- Multi-hop (e.g., Q: Who was the quarterback of the team that won Super Bowl 50? A: Peyton Manning)
- Intersection (e.g., Q: Which movie was directed by Denis Villeneuve and stars Timothee Chalamet? A: Dune)
- Difference (e.g., Q: Which Mario Kart game did Yoshi not appear in? A: Mario Kart Live: Home Circuit)
- Yes/No (e.g., Q: Has Lady Gaga ever made a song with Ariana Grande? A: Yes.)
- Generic (e.g., Q: Where was Michael Phelps born? A: Baltimore, Maryland)
We collected questions about 8 categories: Movies, Music, Sports, Books, Geography, Politics, Video Games, and History
All questions were written in English and translated into 8 additional languages: Arabic, French, German, Hindi, Italian, Japanese, Portuguese, and Spanish
Mintaka is one of the first large-scale complex, natural, and multilingual datasets that can be used for end-to-end question-answering models.
- October 25, 2022:
- Evaluation script is now available! See the Evaluation section below.
- Answer entities now come with Wikidata labels in all languages when available.
- Superlative and Count questions are reformatted with a
supportingNumorsupportingEntfield when a supporting number or entity was provided by the annotator.
In this repo, we provide our randomly split train (14,000 samples), dev (2,000 samples), and test (4,000 samples) sets for Mintaka.
An example sample is shown below:
{
"id": "a9011ddf",
"question": "What is the seventh tallest mountain in North America?",
"translations":
{
"ar": "ما سابع أعلى جبل في أمريكا الشمالية؟",
"de": "Wie heißt der siebthöchste Berg Nordamerikas?",
"ja": "北アメリカで七番目に高い山は何ですか？",
"hi": "उत्तर अमेरिका में सातवां सबसे लंबा पर्वत कौन सा है?",
"pt": "Qual é a sétima montanha mais alta da América do Norte?",
"es": "¿Cuál es la séptima montaña más alta de Norteamérica?",
"it": "Qual è la settima montagna più alta del Nord America?",
"fr": "Quelle est la septième plus haute montagne d’Amérique du Nord ?"
},
"questionEntity":
[
{
"name": "Q49",
"entityType": "entity",
"label": "North America",
"mention": "North America",
"span":
[
40,
53
]
},
{
"name": 7,
"entityType": "ordinal",
"mention": "seventh",
"span":
[
12,
19
]
}
],
"answer":
{
"answerType": "entity",
"answer":
[
{
"name": "Q1153188",
"label":
{
"en": "Mount Lucania",
"ar": null,
"de": "Mount Lucania",
"es": "Monte Lucania",
"fr": "mont Lucania",
"hi": null,
"it": "Monte Lucania",
"ja": "ルカニア山",
"pt": "Monte Lucania"
}
}
],
"mention": "Mount Lucania"
},
"category": "geography",
"complexityType": "ordinal"
}
A description of the fields is given below:
id: a unique ID for the given samplequestion: the original question elicited in Englishtranslations: the translations of the English question into the following languages:ar: Arabic (Saudi Arabia)de: German (Germany)ja: Japanese (Japan)hi: Hindi (India)pt: Portuguese (Brazil)es: Spanish (Mexico)it: Italian (Italy)fr: French (France)
questionEntity: a list of annotated question entities identified by crowd workers.
{
"name": The Wikidata Q-code or numerical value of the entity
"entityType": The type of the entity. Options are:
entity, cardinal, ordinal, date, time, percent, quantity, or money
"label": [optional] The label of the Wikidata Q-code
"mention": The entity as it appears in the English question text
"span": The start and end characters of the mention in the English question text
}
answer: the answer as annotated by crowd workers
{
"answerType": The type of the answer. Options are:
entity, boolean, number, date, or string
"answer": A list of annotated answers. For entities, this will include:
{"name": Wikidata Q-code, "label": {multilingual labels of the entity from Wikidata}}
"mention": The original answer text elicited in English
"supportingEnt": [optional] For count and superlative questions, crowd workers
could provide additional entities to support the answer
"supportingNum": [optional] For superlative questions, crowd workers
could provide additional numeric values to support the answer
}
category: the category of the question. Options are:geography,movies,history,books,politics,music,videogames, orsportscomplexityType: the complexity type of the question. Options are:ordinal,intersection,count,superlative,yesnocomparative,multihop,difference, orgeneric
We have created an evaluation script to help score predictions on the Mintaka test set. This script expects results in a JSON file with the question ID as the key and the answer as the value.
Predictions for text evaluation should be strings, such as:
{
"fae46b21": "Mark Twain",
"bc8713cc": "1",
"d2a03f72": "Drake",
"9a296167": "5",
"e343ad26": "False",
"b41ae115": "Elvis Presto",
"4367c74a": "X-Men X2 X-Men: The Last Stand",
...
}
and predictions for KG evaluation can be strings, integers, floats, booleans, or lists, such as:
{
"fae46b21": "Q7245",
"bc8713cc": 1,
"d2a03f72": "Q33240",
"9a296167": 5,
"e343ad26": false,
"b41ae115": null,
"4367c74a":
[
"Q106182",
"Q219776",
"Q221168"
],
...
}
You can run the evaluation script as:
python evaluate/evaluate.py \
--mode [text or kg] \
--path_to_test_file [default is data/mintaka_test.json]
--path_to_predictions [path to prediction file] \
--lang [language of evaluation]
Below is an example for KG evaluation:
python evaluate/evaluate.py \
--mode kg \
--path_to_predictions evaluate/examples/example_kg_predictions.json
and for text evaluation:
python evaluate/evaluate.py \
--mode text \
--path_to_predictions evaluate/examples/example_text_predictions.json \
--lang en
The evaluation script returns:
- Exact Match: The percent of questions where the prediction exactly matches the labeled answer. For text answers, this is based on the answer scoring in DPR.
- F1: An F1 score calculated as the shared words or entities between the prediction and labeled answer (based on the SQuAD 2.0 eval script).
- Hits@1: A hits@1 score checking if at least one Top 1 prediction matches a labeled answer (based on the hits@K scoring in EmbedKGQA)
This project is licensed under the CC-BY-4.0 License.
If you use this dataset, please cite the following paper:
@inproceedings{sen-etal-2022-mintaka,
title = "Mintaka: A Complex, Natural, and Multilingual Dataset for End-to-End Question Answering",
author = "Sen, Priyanka and
Aji, Alham Fikri and
Saffari, Amir",
booktitle = "Proceedings of the 29th International Conference on Computational Linguistics",
month = oct,
year = "2022",
address = "Gyeongju, Republic of Korea",
publisher = "International Committee on Computational Linguistics",
url = "https://aclanthology.org/2022.coling-1.138",
pages = "1604--1619"
}