KorQuAD: Korean Extractive Question Answering
KorQuAD 1.0 is the Korean machine reading comprehension dataset built on the SQuAD 1.0 protocol: crowdworkers read a Korean Wikipedia passage, write a question, and mark the minimal answer span. 70,079 answerable QA pairs. Includes a runnable Potato config to reproduce or extend the annotation.
About this dataset
KorQuAD 1.0 is a Korean machine reading comprehension dataset introduced by Lim, Kim, and Lee (LG CNS) in 2019. It follows the SQuAD 1.0 protocol so that Korean QA systems can be trained and compared on the same extractive task English models use.
The data was collected from Korean Wikipedia. The authors selected 1,647 articles — including Wikipedia's featured ('알찬 글') and good ('좋은 글') articles plus a random sample — and extracted paragraphs after removing images, tables, and very short passages. Crowdworkers then read each passage and wrote questions in natural Korean.
For every question a worker selected the minimal answer span in the passage, at the morpheme level. Workers were told to write questions in their own words, to vary the register (both casual and respectful Korean forms), to favor questions requiring reasoning, and to avoid questions answerable by simple translation. KorQuAD 1.0 is fully answerable — it has no SQuAD-2.0-style unanswerable questions. The release contains 70,079 QA pairs (train 60,407 / dev 5,774 / test 3,898).
The Potato config below reproduces this data-collection task: it shows a Korean Wikipedia passage, collects a written question, and captures the minimal answer span. Use it to grow KorQuAD, build a domain-specific Korean QA set, or run the same SQuAD-style protocol in another language.
- Released
- 2019 (arXiv, LG CNS)
- Language
- Korean
- Source corpus
- Korean Wikipedia (1,647 articles)
- QA pairs
- 70,079 (train 60,407 / dev 5,774 / test 3,898)
- Answer format
- Minimal extractive span (morpheme-level)
- Unanswerable questions
- None (SQuAD 1.0 style)
Configuration Fileconfig.yaml
This Potato config reproduces the annotation task. Save it as config.yaml and run potato start config.yaml to try it.
# KorQuAD 1.0 - Korean Extractive Question Answering
# Based on Lim et al., arXiv 2019 (KorQuAD1.0)
# Paper: https://arxiv.org/abs/1909.07005
# Dataset: https://korquad.github.io
#
# KorQuAD 1.0 is the Korean counterpart of SQuAD 1.0: crowdworkers read a
# passage from Korean Wikipedia, write a question in natural language, and
# select the minimal answer span in the passage. All questions are answerable
# (there are no SQuAD-2.0-style unanswerable questions in KorQuAD 1.0).
#
# Annotation workflow (as described in the paper, sections 3.1-3.2):
# 1. Read the given Korean Wikipedia passage.
# 2. Write a question in your own words that is answerable from the passage.
# Vary the question form (both casual and respectful register are welcome).
# 3. Highlight the minimal span in the passage that answers the question.
annotation_task_name: "KorQuAD - Korean Extractive QA"
task_dir: "."
data_files:
- sample-data.json
item_properties:
id_key: "id"
text_key: "text"
output_annotation_dir: "annotation_output/"
output_annotation_format: "json"
port: 8000
server_name: localhost
annotation_schemes:
# Step 1: Write a question in Korean answerable from the passage
- annotation_type: text
name: question
description: "이 문단을 읽고 답을 찾을 수 있는 질문을 한국어로 작성하세요 (Write a Korean question answerable from the passage)"
# Step 2: Highlight the minimal answer span in the passage
- annotation_type: span
name: answer_span
description: "질문의 답이 되는 최소한의 구간을 표시하세요 (Highlight the minimal span that answers the question)"
labels:
- "정답 (Answer)"
label_colors:
"정답 (Answer)": "#3b82f6"
annotation_instructions: |
You will be shown a passage from Korean Wikipedia. Following the KorQuAD 1.0
data-collection protocol (Lim et al., 2019):
1. Read the passage carefully.
2. Write a question in Korean, in your own words, that can be answered from the
passage. You are encouraged to vary the question form — Korean allows both
casual and respectful (존댓말) phrasings — and to write questions that require
some reasoning rather than simple word matching. Avoid questions that are
answerable only by translation.
3. Highlight the minimal span in the passage that answers your question. Keep the
span as short as possible (KorQuAD answers are minimal spans at the morpheme
level).
Every question must be answerable from the passage — KorQuAD 1.0 contains no
unanswerable questions.
html_layout: |
<div style="padding: 15px; max-width: 800px; margin: auto;">
<div style="background: #f0f9ff; border: 1px solid #bae6fd; border-radius: 8px; padding: 16px; margin-bottom: 16px;">
<strong style="color: #0369a1;">문단 (Passage):</strong>
<p style="font-size: 16px; line-height: 1.8; margin: 8px 0 0 0;">{{text}}</p>
</div>
</div>
allow_all_users: true
instances_per_annotator: 50
annotation_per_instance: 3
allow_skip: true
skip_reason_required: false
Sample Datasample-data.json
[
{
"id": "korquad_001",
"text": "세종대왕은 조선의 제4대 국왕으로, 1443년에 훈민정음을 창제하였다. 훈민정음은 오늘날 한글이라고 불리며, 자음과 모음으로 이루어진 표음 문자이다. 세종은 백성이 쉽게 글을 배우고 쓸 수 있도록 하기 위해 새로운 문자를 만들었다."
},
{
"id": "korquad_002",
"text": "서울특별시는 대한민국의 수도이자 최대 도시이다. 한강이 도시의 중심을 가로질러 흐르며, 조선 시대부터 오늘날까지 수도의 역할을 이어오고 있다. 서울에는 경복궁을 비롯한 여러 궁궐이 남아 있다."
}
]
// ... and 2 more itemsGet This Design
Clone or download from the repository
Quick start:
git clone https://github.com/davidjurgens/potato-showcase.git cd potato-showcase/text/reading-comprehension/korquad-korean-qa potato start config.yaml
Dataset & paper
Lim et al., arXiv 2019
Citation (BibTeX)
@article{lim2019korquad,
title = "{K}or{Q}u{AD}1.0: {K}orean {QA} Dataset for Machine Reading Comprehension",
author = "Lim, Seungyoung and Kim, Myungji and Lee, Jooyoul",
journal = "arXiv preprint arXiv:1909.07005",
year = "2019",
url = "https://arxiv.org/abs/1909.07005"
}Details
Annotation Types
Domain
Use Cases
Tags
Found an issue or want to improve this design?
Open an IssueRelated Designs
FQuAD: French Question Answering Dataset
FQuAD is the French extractive QA dataset from Illuin Technology, built on the SQuAD 1.1 protocol. University students read a French Wikipedia paragraph, write a question, and select the smallest answer span. Over 60,000 answerable QA pairs, with a runnable Potato config to reproduce or extend it.
GermanQuAD: German Extractive Question Answering
GermanQuAD is deepset's German machine reading comprehension dataset (MRQA 2021). Trained annotators read a German Wikipedia passage, write a self-sufficient question reformulated to reduce lexical overlap, and mark the answer span. 13,722 answerable QA pairs, with a runnable Potato config to reproduce or extend it.
Natural Questions - Open-Domain Question Answering
Open-domain question answering over Wikipedia passages, based on Google's Natural Questions dataset (Kwiatkowski et al., TACL 2019). Annotators identify both short and long answer spans and determine answerability.