• List of Backups
• View the diff.
• View the diff current.
• View the source.
• Go to Driving domain QA datasets.
  • 1 (2019-10-28 (Mon) 06:39:48)
  • 2 (2019-10-29 (Tue) 08:34:03)
  • 3 (2019-10-29 (Tue) 13:13:03)
  • 4 (2019-10-31 (Thu) 14:00:21)

Driving domain QA datasets †

The driving domain QA (Question Answering) datasets were constructed based on the driving domain blog posts published on the web. They consisted of a Predicate-Argument Structure QA (PAS-QA) dataset and a Reading Comprehension QA (RC-QA) dataset. We constructed a PAS-QA dataset in which a question asks an omitted argument for a predicate. We made 12,468 questions for the ga case (nominative), 3,151 questions for the wo case (accusative), 1,069 questions for the ni case (dative). We also constructed an RC-QA dataset which was a problem to extract the answer to the question from text, and made 20,007 questions. We constructed PAS-QA and RC-QA datasets with crowdsourcing because it enabled to create large-scale datasets in a short time. The data format of these QA datasets is the same as SQuAD 2.0. As for the PAS-QA nominative dataset and the RC-QA dataset, all problems have answers in documents. However, as for the PAS-QA accusative and dative dataset, some problems cannot be answered because there are no answers in documents. Please refer to the references for how to construct these datasets and how to make problems with no answers in documents. The following are examples of driving domain QA datasets.

• PAS-QA dataset
  • Document :
    • 私は右車線に移動した。 (I moved to the right lane.)
    • (Φが) バックミラーを見た。 ((Φ-NOM) saw the rearview mirror.)
  • Question :
    • “見た”の主語は何か？ (What is the subject of "saw"?)
  • Answer :　
    • 私 (I)

• RC-QA dataset
  • Document :　
    • 私の車の前をバイクにまたがった警察官が走っていた。 (A police officer straddling his bike was running in front of my car.)
  • Question : 　
    • 警察官は何に乗っていた？ (What was the police officer riding?)
  • Answer :
    • バイク (his bike)

Download †

Driving domain QA datasets Version 1.0 (tar.gz compression; 4,292,552 bytes): Download Form
cf.) To download datasets, you need to enter your name and email address, and agree to the download conditions.

Update history †

• Version 1.0 - Released on 10/31/2019

References †

• Norio Takahashi, Tomohide Shibata, Daisuke Kawahara and Sadao Kurohashi.
  Predicate-argument structure analysis based on a machine comprehension model in a specific domain,
  In Proceedings of the 25th Annual Meeting of Natural Language Processing (in Japanese), 2019.
  　https://www.anlp.jp/proceedings/annual_meeting/2019/pdf_dir/B1-4.pdf
  　　cf.) Describes how to construct these datasets.
• Norio Takahashi, Tomohide Shibata, Daisuke Kawahara and Sadao Kurohashi.
  Machine Comprehension Improves Domain-Specific Japanese Predicate-Argument Structure Analysis,
  In Proceedings of 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint Conference on Natural Language Processing, Workshop MRQA: Machine Reading for Question Answering, 2019.
  　https://mrqa.github.io/assets/papers/42_Paper.pdf
  　　cf.) Describes how to construct these datasets and how to make problems with no answers in documents.

Contact †

If you have any questions or problems about these datasets, please send an email to nl-resource at nlp.ist.i.kyoto-u.ac.jp. If you have a request to add source information or to delete a document in the datasets, please send an email to this mail address.