Dataset Card for JDocQA

Dataset Summary

Japanese Document Question Answering (JDocQA), a large-scale document-based QA dataset, essentially requiring both visual and textual information to answer questions, which comprises 5,504 documents in PDF format and annotated 11,600 question-and-answer instances in Japanese.

Supported Tasks and Leaderboards

From JDocQA's paper:

We consider generative question answering where a model generates a textual answer following the document context and textual question. For realistic applications of a wide range of user questions for documents, we prepare four categories of questions: (1) yes/no, (2) factoid, (3) numerical, and (4) open-ended.

In yes/no questions, answers are “yes” or “no.”

In factoid questions, answers are some facts, such as named entities, that typically appear in the given documents.

In numerical questions, answers are numeric values, often including some numerals (some units, e.g., km or Japanese numerals such as “8個 (objects)” and “8人 (persons)”). These numeric values are written in the documents or are calculated from other numbers in the documents.

In open-ended questions, free-form responses are required. For such questions, we aim to assess complex comprehension abilities, such as the ability to form opinions or brief explanations based on the provided contexts and questions.

Figure 1 presents samples of these four categories of questions. All examples include diverse images and question types related to some Japanese documents collected. We also include unanswerable questions for each question category.

Languages

The language data in JDocQA is in Japanese (BCP-47 ja-JP).

Dataset Structure

Data Instances

import datasets as ds

dataset = ds.load_dataset(
  path="shunk031/JDocQA", 
  # Rename to the same wording as in the paper: Document -> Report / Kouhou -> Pamphlet
  rename_pdf_category=True,
  # Set to True to use loading script for huggingface datasets
  trust_remote_code=True,
)

print(dataset)
# DatasetDict({
#     train: Dataset({
#         features: ['answer', 'answer_type', 'context', 'multiple_select_answer', 'multiple_select_question', 'no_reason', 'normalized_answer', 'original_answer', 'original_context', 'original_question', 'pdf_category', 'pdf_name', 'question', 'question_number', 'question_page_number', 'reason_of_answer_bbox', 'text_from_ocr_pdf', 'text_from_pdf', 'type_of_image', 'pdf_filepath'],
#         num_rows: 9290
#     })
#     validation: Dataset({
#         features: ['answer', 'answer_type', 'context', 'multiple_select_answer', 'multiple_select_question', 'no_reason', 'normalized_answer', 'original_answer', 'original_context', 'original_question', 'pdf_category', 'pdf_name', 'question', 'question_number', 'question_page_number', 'reason_of_answer_bbox', 'text_from_ocr_pdf', 'text_from_pdf', 'type_of_image', 'pdf_filepath'],
#         num_rows: 1134
#     })
#     test: Dataset({
#         features: ['answer', 'answer_type', 'context', 'multiple_select_answer', 'multiple_select_question', 'no_reason', 'normalized_answer', 'original_answer', 'original_context', 'original_question', 'pdf_category', 'pdf_name', 'question', 'question_number', 'question_page_number', 'reason_of_answer_bbox', 'text_from_ocr_pdf', 'text_from_pdf', 'type_of_image', 'pdf_filepath'],
#         num_rows: 1176
#     })
# })

An example of the JDocQA dataset (training set) looks as follows:

{
    "answer": "本文中に記載がありません",
    "answer_type": 3,
    "context": "_II.調査内容(2.虹本マニュアルの策定)(3)基本マニュアルの記載項目前述の方針等を踏まえ、基本マニュアルの具体的な記載項目(目次だて)は以下のとおりとする。小項目・内容Iはじめにマニュアルの目的、立会義務_(消防法第13条第3項)安全対策の基本事項(SS立会い者とローリー乗務員による相互確認・相互協力の重要性)ローリー荷卸しの手順の基本的流れ ※詳細版のみIIローリー荷邊し時の作業内容1ローリー到着時(荷爺し前)1.ローリー停車位置の確認:計導2.納品書の相互確認3.アースの接続4.消火器の配置5.積荷の相互確認6.地下タンク和在庫及び和荷卸し数量の確認7・詳細版には、各項目ごとに、_-SS立会い者、ローリー乗務2荷邊し時(ホースの結合)03-.注油口の確認、ホースの結合を記載3.ベーパー回収ホース接続ee4荷卸し作業中の安全馬視特に重要な基本事3荷卸し終了時1.配管内、ホース内の残油の確認2.注油口の確認ハッチ内残油確認3.在庫確認4.5.後片付け6.ローリーの退出自事故・災害時の対処(初動対応)1コンタミ(混油)事故発見時(緊急処置)、連絡2オーバーフロー(漏油)事故発見時(緊急処置)、連絡3火災発見時(緊急処置)、初期消火IV通報・緊急連絡緊急時連絡先、通報内容参考チェックリスト例",
    "multiple_select_answer": 3,
    "multiple_select_question": ["はい", "いいえ", "わからない", "本文中に記載が見つけられませんでした"],
    "no_reason": 0,
    "normalized_answer": "本文中に記載がありません",
    "original_answer": "本文中に記載が見つけられませんでした",
    "original_context": "_II.調査内容(2.虹本マニュアルの策定)(3)基本マニュアルの記載項目前述の方針等を踏まえ、基本マニュアルの具体的な記載項目(目次だて)は以下のとおりとする。小項目・内容Iはじめにマニュアルの目的、立会義務_(消防法第13条第3項)安全対策の基本事項(SS立会い者とローリー乗務員による相互確認・相互協力の重要性)ローリー荷卸しの手順の基本的流れ ※詳細版のみIIローリー荷邊し時の作業内容1ローリー到着時(荷爺し前)1.ローリー停車位置の確認:計導2.納品書の相互確認3.アースの接続4.消火器の配置5.積荷の相互確認6.地下タンク和在庫及び和荷卸し数量の確認7・詳細版には、各項目ごとに、_-SS立会い者、ローリー乗務2荷邊し時(ホースの結合)03-.注油口の確認、ホースの結合を記載3.ベーパー回収ホース接続ee4荷卸し作業中の安全馬視特に重要な基本事3 荷卸し終了時1.配管内、ホース内の残油の確認2.注油口の確認ハッチ内残油確認3.在庫確認4.5.後片付け6.ローリーの退出自事故・災害時の対処(初動対応)1コンタミ(混油)事故発見時(緊急処置)、連絡2オーバーフロー(漏油)事故発見時(緊急処置)、連絡3火災発見時(緊急処置)、初期消火IV通報・緊急連絡緊急時連絡先、通報内容参考チェックリスト例",
    "original_question": "基本マニュアルの具体的な記載項目としている事故・災害時の対処の中で、オーバーフロー（漏油）事故が起こった場合は発見時にどのような処置が求められますか？",
    "pdf_category": 2,
    "pdf_name": "public_document00152.pdf",
    "question": "基本マニュアルの具体的な記載項目としている事故・災害時の対処の中で、オーバーフロー（漏油）事故が起こった場合は発見時にどのような処置が求められますか？\n解答は自由に記述してください。",
    "question_number": [4, 656, 1, 4],
    "question_page_number": "9",
    "reason_of_answer_bbox": [""],
    "text_from_ocr_pdf": "_II.調査内容(2.虹本マニュアルの策定)(3)基本マニュアルの記載項目前述の方針等を踏まえ、基本マニュアルの具体的な記載項目(目次だて)は以下のとおりとする。小項目・内容Iはじめにマニュアルの目的、立会義務_(消防法第13条第3項)安全対策の基本事項(SS立会い者とローリー乗務員による相互確認・相互協力の重要性)ローリー荷卸しの手順の基本的流れ ※詳細版のみIIローリー荷邊し時の作業内容1ローリー到着時(荷爺し前)1.ローリー停車位置の確認:計導2.納品書の相互確認3.アースの接続4.消火器の配置5.積荷の相互確認6.地下タンク和在庫及び和荷卸し数量の確認7・詳細版には、各項目ごとに、_-SS立会い者、ローリー乗務2荷邊し時(ホースの結合)03-.注油口の確認、ホースの結合を記載3.ベーパー回収ホース接続ee4荷卸し作業中の安全馬視特に重要な基本事3荷卸し終了時1.配管内、ホース内の残油の確認2.注油口の確認ハッチ内残油確認3.在庫確認4.5.後片付け6.ローリーの退出自事故・災害時の対処(初動対応)1コンタミ(混油)事故発見時(緊急処置)、連絡2オーバーフロー(漏油)事故発見時(緊急処置)、連絡3火災発見時(緊急処置)、初期消火IV通報・緊急連絡緊急時連絡先、通報内容参考チェックリスト例",
    "text_from_pdf": "",
    "type_of_image": [0],
    "pdf_filepath": "/home/shunk031/.cache/huggingface/datasets/downloads/extracted/f3481b9f65c75efec1e5398f76bd8347e64661573961b69423568699f1d7083a/pdf_files/public_document00152.pdf"
}

Data Fields

From JDocQA's README.md and the paper:

answer:
answer_type: (1) Yes/No questions, (2) Factoid questions, (3) Numerical questions, (4) Open-ended questions.
context: Removed noises from 'original_context'.
multiple_select_answer:
multiple_select_question:
no_reason: Unanswerable question -> 0, Answerable question -> 1, Multi page question -> 2. They can be jointly flagged such as 1,2.
normalized_answer:
original_answer: Annotated answers.
original_context: Extracted texts from PDF.
original_question: Annotated questions.
pdf_category: Document category.
pdf_name: PDF name.
question: Question query for models.
question_number:
question_page_number: Where annotators found answer of the questions.
reason_of_answer_bbox:
text_from_ocr_pdf:
text_from_pdf:
type_of_image: (1) Table, (2) Bar chart, (3) Line chart, (4) Pie chart, (5) Map, (6) Other figures, (7) Mixtured writing style from left to the right and from upside to the downside, (8) Drawings, (9) Others. Note that this enrty is for statistical purpose in our paper, and some labels are missing, which are represented as null.
pdf_filepath: full file path to the corresponding PDF file.

pdf_category

We renamed the several category names upon the paper for the interpretability.

Document category in the PDF set as Report in the paper.

Kouhou category in the PDF set as Pamphlet in the paper.

Data Splits

From JDocQA's paper:

学習，検定，テストセットにそれぞれ 9,290 件，1,134 件，1,176 件の質問応答が含まれるようにデータセット全体を分割した．同一 PDF ファイルは必ず同一の分割に出現する．

Dataset Creation

Curation Rationale

From JDocQA's paper:

To address the demand for a large-scale and fully annotated Japanese document question answering dataset, we introduce a JDocQA dataset by collecting Japanese documents in PDF styles from open-access sources including multiple formats of documents: slides, reports, websites and pamphlets and manually annotating question-answer pairs on them.

Source Data

From JDocQA's paper:

We gather public documents, such as, municipality pamphlets and websites, that are created by Japanese governmental agencies or local governments.

Initial Data Collection and Normalization

From JDocQA's paper:

We manually collected PDF documents from open-access resources such as Japanese National Diet Library (NDL)’s digital collection, web archive projects (WARP) and websites of Japanese government ministries. We manually gathered documents such as reports, pamphlets or websites that are published by public or quasi-public sectors, such as local governments or public universities through WARP. We also gather Japanese ministry documents such as slides and reports from their websites following the government agencies’ policies. Those documents cover a wide range of topics, for instance, economic policies, education policies, labor issues, health and hygiene, agriculture, forestry, fisheries, culture and arts, history, related to governmental policy or policy guidelines, as well as the everyday affairs of local governments. These documents also include visual elements such as figures, tables, charts, pictures, or mandala charts, complex figures with a combination of texts and objects typically seen in the Japanese public administrative sector’s official document. We classify these documents into four categories, namely, pamphlet, slide, report, and website considering the form of the documents.

We extracted texts from PDF documents with PyPDF2. We also notice that some PDF documents are probably created from paper scans, and we cannot extract embedded texts from such documents. Therefore, we extracted texts from the document page images by OCR (Optical Character Recognition) as an alternative source. After the text extraction or OCR, we removed mistakenly recognized symbols and emojis, or duplicated characters from texts when the same character continuously and repeatedly appeared more than five times.

Who are the source language producers?

From JDocQA's paper:

JDocQA dataset comprises 5,504 files and 11,600 question-and-answer pairs in Japanese.

Annotations

Annotation process

From JDocQA's paper:

As documents include rich textual and visual elements (e.g., graphs, charts, maps, illustrations, and a mix of vertical and horizontal written text), we made question answer pairs that are related to both textual and visual information. We ask annotators to write up two to four question-answer annotations in each document. We also ask not to use any AI-tools such as OpenAI ChatGPT during the annotation process. Each question is accompanied with the supporting facts as marked in red in Figure 1 and Figure 3. We classify a subset of questions that have multiple supporting facts in multiple pages as multi-page questions. Multi-page questions are considerably difficult from their single-page counterparts. For unanswerable questions, we ask annotators to write questions that lack supporting facts in the documents, making them impossible to answer based on the given documents.

We prepared three types of images for visual inputs for multimodal models. The first type of images are those of the whole page of the documents including the annotated question answering pairs. The second type of images are those cropped by bounding boxes on which annotators based their answers such as tables or figures of the pages. When multiple bounding boxes are annotated to a single question-answer pair, multiple cropped images are combined together into a single image here. The third type of images are blank (white) images that are used for ablation studies.

Who are the annotators?

From JDocQA's paper:

We ask 43 annotators in total for the question-answering pairs annotation on documents.

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

From JDocQA's paper:

We assume our datasets are useful for both research and development of generative language models and their applications for Japanese document question answering.

Discussion of Biases

From JDocQA's paper:

We carefully avoid private documents and choose considerably public documents published by public or quasi-public sectors for the publicity of our dataset usage. All of the documents and webpages are publicly available online and we follow our institutional rules to gather them. We follow our institutional rules and also consult external advisors for data collection processes.

Other Known Limitations

From JDocQA's paper:

We also consider our dataset with unanswerable questions can contribute to harnessing the hallucination problem of large language models. However, this doesn’t mean that the fintuned models with unanswerable questions do not perform hallucinations at all.

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

From JDocQA's README.md:

JDocQA dataset annotations are distributed under CC BY-SA 4.0. We are delighted to see many derivations from JDocQA! When you create any derivations, e.g., datasets, papers, etc, from JDocQA, please cite our paper accordingly. If your derivations are web-based projects, please cite our paper and include the link to this github page.

Citation Information

@inproceedings{onami2024jdocqa,
  title={JDocQA: Japanese Document Question Answering Dataset for Generative Language Models},
  author={Onami, Eri and Kurita, Shuhei and Miyanishi, Taiki and Watanabe, Taro},
  booktitle={Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  pages={9503--9514},
  year={2024}
}

Contributions

Thanks to @mizuumi for creating this dataset.