ltg/norquad

Dataset Card for Dataset Name

NorQuAD is the first Norwegian question answering dataset for machine reading comprehension, created from scratch in Norwegian. The dataset consists of 4,752 manually created question-answer pairs.

Dataset Details

Dataset Description

The dataset provides Norwegian question-answer pairs taken from two data sources: Wikipedia and news.

• Curated by: Human annotators.
• Funded by: The UiO Teksthub initiative
• Shared by: The Language Technology Group, University of Oslo
• Language(s) (NLP): Norwegian Bokmål
• License: CC0-1.0

Dataset Sources

• Repository: https://github.com/ltgoslo/NorQuAD
• Paper: Ivanova et. al., 2023

Uses

The dataset is intended to be used for NLP model development and benchmarking.

Dataset Structure

Data Instances

{
    "id": "1",
    "context": "This is a test context",
    "question": "This is a question",
    "answers": {
        "answer_start": [1],
        "text": ["This is an answer"]
    },
}

Data Fields

  id: a string feature.
  context: a string feature.
  question: a string feature.
  answers: a dictionary feature containing:
    text: a string feature.
    answer_start: a int32 feature.

Dataset Splits

NorQuAD consists of training (3808 examples), validation (472), and public test (472) sets.

Dataset Creation

Curation Rationale

Machine reading comprehension is one of the key problems in natural language understanding. The question answering (QA) task requires a machine to read and comprehend a given text passage, and then answer questions about the passage. There is progress in reading comprehension and question answering for English and a few other languages. We would like to fill in the lack of annotated data for question answering for Norwegian. This project aims at compiling human-created training, validation, and test sets for the task for Norwegian.

Source Data

Wikipedia: 872 articles were sampled from Norwegian Bokmal Wikipedia.

News: For the news category, articles were sampled from Norsk Aviskorpus, an openly available dataset of Norwegian news.

Data Collection and Processing

Wikipedia:In order to include high-quality articles, 130 articles from the ‘Recommended‘ section and 139 from the ‘Featured‘ section were sampled. The remaining 603 articles were randomly sampled from the remaining Wikipedia corpus. From the sampled articles, we chose only the “Introduction“ sections to be selected as passages for annotation.

News: 1000 articles were sampled from the Norsk Aviskorpus (NAK)—a collection of Norwegian news texts for the year 2019. As was the case with Wikipedia articles, we chose only news articles which consisted of at least 300 words.

Who are the source data producers?

The data is sourced from Norwegian Wikipedia dumps as well as the openly available Norwegian News Corpus, available from the Språkbanken repository.

Annotations

In total, the annotators processed 353 passages from Wikipedia and 403 passages from news, creating a total of 4,752 question-answer pairs.

Annotation process

The dataset was created in three stages: (i) selecting text passages, (ii) collecting question-answer pairs for those passages, and (iii) human validation of (a subset of) created question-answer pairs.

Text selection

Data was selected from openly available sources from Wikipedia and News data, as described above.

Question-Answer Pairs

The annotators were provided with a set of initial instructions, largely based on those for similar datasets, in particular, the English SQuAD dataset (Rajpurkar et al., 2016) and the GermanQuAD data (Moller et al., 2021). These instructions were subsequently refined following regular meetings with the annotation team. The annotation guidelines provided to the annotators are available (here)[https://github.com/ltgoslo/NorQuAD/blob/main/guidelines.md]. For annotation, we used the Haystack annotation tool, which was designed for QA collection.

Human validation

In a separate stage, the annotators validated a subset of the NorQuAD dataset. In this phase, each annotator replied to the questions created by the other annotator. We chose the question-answer pairs for validation at random. In total, 1378 questions from the set of question-answer pairs were answered by validators.

Who are the annotators?

Two students of the Master’s program in Natural Language Processing at the University of Oslo, both native Norwegian speakers, created question-answer pairs from the collected passages. Each student received a separate set of passages for annotation. The students received financial remuneration for their efforts and are co-authors of the paper describing the resource.

Citation

BibTeX:

@inproceedings{
ivanova2023norquad,
title={NorQu{AD}: Norwegian Question Answering Dataset},
author={Sardana Ivanova and Fredrik Aas Andreassen and Matias Jentoft and Sondre Wold and Lilja {\O}vrelid},
booktitle={The 24th Nordic Conference on Computational Linguistics},
year={2023},
url={https://aclanthology.org/2023.nodalida-1.17.pdf}
}

APA:

[More Information Needed]

Dataset Card Authors

Vladislav Mikhailov and Lilja Øvrelid

Dataset Card Contact

vladism@ifi.uio.no and liljao@ifi.uio.no