Hugging Face's logo

Datasets:
akariasai
/
xor_tydi_qa

Tasks:
Question Answering
Sub-tasks:
open-domain-qa
Languages:
Arabic
Bengali
Finnish
Size:
10K<n<100K
ArXiv:
License:
Dataset card Files Community
3
xor_tydi_qa / README.md
albertvillanova's picture
albertvillanova HF staff
Replace YAML keys from int to str (#2)
91e85a5
preview code
|
raw
history blame contribute delete
9.09 kB
metadata
annotations_creators:
  - crowdsourced
language_creators:
  - expert-generated
  - found
language:
  - ar
  - bn
  - fi
  - ja
  - ko
  - ru
  - te
license:
  - mit
multilinguality:
  - multilingual
size_categories:
  - 10K<n<100K
source_datasets:
  - original
  - extended|tydiqa
task_categories:
  - question-answering
task_ids:
  - open-domain-qa
paperswithcode_id: xor-tydi-qa
pretty_name: XOR QA
dataset_info:
  - config_name: xor-retrieve
    features:
      - name: question
        dtype: string
      - name: lang
        dtype:
          class_label:
            names:
              '0': ar
              '1': bn
              '2': fi
              '3': ja
              '4': ko
              '5': ru
              '6': te
      - name: answers
        dtype: string
    splits:
      - name: train
        num_bytes: 1698662
        num_examples: 15250
      - name: validation
        num_bytes: 259533
        num_examples: 2110
      - name: test
        num_bytes: 219046
        num_examples: 2499
    download_size: 3702288
    dataset_size: 2177241
  - config_name: xor-full
    features:
      - name: question
        dtype: string
      - name: lang
        dtype:
          class_label:
            names:
              '0': ar
              '1': bn
              '2': fi
              '3': ja
              '4': ko
              '5': ru
              '6': te
      - name: answers
        dtype: string
    splits:
      - name: train
        num_bytes: 7250913
        num_examples: 61360
      - name: validation
        num_bytes: 444672
        num_examples: 3473
      - name: test
        num_bytes: 706664
        num_examples: 8176
    download_size: 14018298
    dataset_size: 8402249

Dataset Card for XOR QA

Table of Contents

Dataset Description

Dataset Summary

XOR-TyDi QA brings together for the first time information-seeking questions, open-retrieval QA, and multilingual QA to create a multilingual open-retrieval QA dataset that enables cross-lingual answer retrieval. It consists of questions written by information-seeking native speakers in 7 typologically diverse languages and answer annotations that are retrieved from multilingual document collections.

Supported Tasks and Leaderboards

There are three sub-tasks: XOR-Retrieve, XOR-EnglishSpan, and XOR-Full.

  • XOR-retrieve: XOR-Retrieve is a cross-lingual retrieval task where a question is written in a target language (e.g., Japanese) and a system is required to retrieve English paragraphs that answer the question. The dataset can be used to train a model for cross-lingual retrieval. Success on this task is typically measured by R@5kt, R@2kt (the recall by computing the fraction of the questions for which the minimal answer is contained in the top 5,000 / 2,000 tokens selected). This task has an active leaderboard which can be found at leaderboard url

  • XOR-English Span: XOR-English Span is a cross-lingual retrieval task where a question is written in a target language (e.g., Japanese) and a system is required to output a short answer in English. The dataset can be used to train a model for cross-lingual retrieval. Success on this task is typically measured by F1, EM. This task has an active leaderboard which can be found at leaderboard url

  • XOR-Full: XOR-Full is a cross-lingual retrieval task where a question is written in the target language (e.g., Japanese) and a system is required to output a short answer in a target language. Success on this task is typically measured by F1, EM, BLEU This task has an active leaderboard which can be found at leaderboard url

Languages

The text in the dataset is available in 7 languages: Arabic ar, Bengali bn, Finnish fi, Japanese ja, Korean ko, Russian ru, Telugu te

Dataset Structure

Data Instances

A typical data point comprises a question, it's answer the language of the question text and the split to which it belongs.

{
    "id": "-3979399588609321314", 
    "question": "Сколько детей было у Наполео́на I Бонапа́рта?", 
    "answers": ["сын"], 
    "lang": "ru", 
    "split": "train"
}

Data Fields

  • id: An identifier for each example in the dataset
  • question: Open domain question
  • answers: The corresponding answer to the question posed
  • lang: BCP-47 language tag
  • split: identifier to differentiate train, validation and test splits

Data Splits

The data is split into a training, validation and test set for each of the two configurations.

train validation test
XOR Retrieve 15250 2113 2501
XOR Full 61360 3179 8177

Dataset Creation

Curation Rationale

This task framework reflects well real-world scenarios where a QA system uses multilingual document collections and answers questions asked by users with diverse linguistic and cultural backgrounds. Despite the common assumption that we can find answers in the target language, web re- sources in non-English languages are largely lim- ited compared to English (information scarcity), or the contents are biased towards their own cul- tures (information asymmetry). To solve these issues, XOR-TYDI QA (Asai et al., 2020) provides a benchmark for developing a multilingual QA system that finds answers in multiple languages.

Source Data

annotation pipeline consists of four steps: 1) collection of realistic questions that require cross-lingual ref- erences by annotating questions from TYDI QA without a same-language answer; 2) question translation from a target language to the pivot language of English where the missing informa- tion may exist; 3) answer span selection in the pivot language given a set of candidate documents; 4) answer verification and translation from the pivot language back to the original language.

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

The Dataset is created by extending TyDiQA dataset and translating the questions into other languages. The answers are obtained by crowdsourcing the questions to Mechanical Turk workders

Annotations

Annotation process

The English questions from TyDiQA are translated into other languages. The languages are chosen based on the availability of wikipedia data and the availability of tranlators.

Who are the annotators?

The translations are carried out using the professionla tranlation service (Gengo)[https://gengo.com] and the answers are annotated by MechanicalTurk workers

Personal and Sensitive Information

The dataset is created from wikipedia content and the QA task requires preserving the named entities, there by all the Wikipedia Named Entities are preserved in the data. Not much information has been provided about masking sensitive information.

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 people associated with the creation of the dataset are Akari Asai, Jungo Kasai, Jonathan H. Clark, Kenton Lee, Eunsol Choi, Hannaneh Hajishirzi

Licensing Information

XOR-TyDi QA is distributed under the CC BY-SA 4.0 license

Citation Information 

@article{xorqa,
    title   = {XOR QA: Cross-lingual Open-Retrieval Question Answering},
    author  = {Akari Asai and Jungo Kasai and Jonathan H. Clark and Kenton Lee and Eunsol Choi and Hannaneh Hajishirzi}
    year    = {2020}
}

Contributions

Thanks to @sumanthd17 for adding this dataset.