Datasets:

SamiaT
/

NorSumm

license: cc0-1.0
task_categories:
  - summarization
language:
  - nb
  - nn
  - 'no'
size_categories:
  - n<1K
dataset_info:
  - config_name: nb
    features:
      - name: id
        dtype: string
      - name: article
        dtype: string
      - name: summaries
        sequence: string
    splits:
      - name: validation
        num_bytes: 165830
        num_examples: 30
      - name: test
        num_bytes: 217351
        num_examples: 33
    download_size: 247097
    dataset_size: 383181
  - config_name: nn
    features:
      - name: id
        dtype: string
      - name: article
        dtype: string
      - name: summaries
        sequence: string
    splits:
      - name: validation
        num_bytes: 165879
        num_examples: 30
      - name: test
        num_bytes: 219189
        num_examples: 33
    download_size: 249301
    dataset_size: 385068
configs:
  - config_name: nb
    data_files:
      - split: validation
        path: nb/validation-*
      - split: test
        path: nb/test-*
  - config_name: nn
    data_files:
      - split: validation
        path: nn/validation-*
      - split: test
        path: nn/test-*

configs: - config_name: default data_files: - split: dev path: "NorSumm_dev.json" - split: test path: "NorSumm_test.json"

Dataset Card for Dataset Name

This repository contains the Norwegian Summarisation Benchmark Dataset, which consists of 378 human-authored summaries from prominent Norwegian news sources across various domains. The dataset is designed to benchmark the abstractive summarisation capabilities of generative language models. NorSumm is the first manually created Norwegian news summarization datasat, created from scratch in Norwegian. The dataset consists of 378 manually generated summaries for 63 new articles. Each news article has been summarised by three native Norwegian speakers, and each generated summary has been translated to either Norwegian written form (Bokmål or Nynorsk) depending on which language the original summary has been written in.

The accompanying paper by Touileb et al. (2025) provides a comprehensive account of the data creation process and evaluates the performance of existing open large language models for Norwegian on this dataset. Additionally, it offers insights from a manual human evaluation, comparing human-authored summaries to those generated by models. The findings suggest that the dataset presents a challenging benchmark for assessing the summarisation capabilities of LLMs in Norwegian.

Source data

We used a subset of news articles from the Norwegian event extraction dataset EDEN (Touileb et al., 2024), as the source for summarisation. Due to the time-intensive nature of creating summaries, we only created summaries for the development and test splits of EDEN, comprising 30 and 33 news articles, respectively.

Dataset Description

The dataset contains pairs of articles and six generated summaries (generetaed by three individual people, in both Bokmål and Nynorsk). The news articles represen the Bokmål portion of the Norwegian Dependency treebank data NDT.

Curated by: Human annotators.
Funded by: MediaFutures: Research Centre for Responsible Media Technology & Innovation, University of Bergen.
Shared by: WP5 -- Norwegian language technology, at MediaFutures
Language(s) (NLP): Norwegian Bokmål and Nynorsk.
License: CC0-1.0

Dataset Sources [optional]

Repository: https://github.com/SamiaTouileb/NorSumm
Paper [optional]: Touileb et al. (2025). Benchmarking Abstractive Summarisation: A Dataset of Human-authored Summaries of Norwegian News Articles. Accepted at NoDaLiDa2025.

Uses

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

Terms of use

The license is the same as the underlying Norwegian Dependency Treebank and is Creative Commons (CC) Licence Name: Creative_Commons-ZERO (CC-ZERO).

Who are the annotators?

We hired three annotators with strong academic backgrounds related to journalism, all Norwegian native speakers. They were fairly compensated following an hourly contract, and were hired for a period of 6 months. All annotators have a background in media science or journalism.

Citation [optional]

BibTeX:

@article{touileb2025benchmarking, title={Benchmarking Abstractive Summarisation: A Dataset of Human-authored Summaries of Norwegian News Articles}, author={Touileb, Samia and Mikhailov, Vladislav and Kroka, Marie and {\O}vrelid, Lilja and Velldal, Erik}, journal={arXiv preprint arXiv:2501.07718}, year={2025} }

APA:

Touileb, S., Mikhailov, V., Kroka, M., Øvrelid, L., & Velldal, E. (2025). Benchmarking Abstractive Summarisation: A Dataset of Human-authored Summaries of Norwegian News Articles. arXiv preprint arXiv:2501.07718.

Dataset Card Authors

Samia Touileb

Dataset Card Contact

Samia Touileb