Datasets:

drelhaj
/

Multilingual-Summarization

license: cc-by-4.0
language:
  - ar
  - zh
  - cs
  - en
  - fr
  - el
  - he
  - hi
  - ro
  - es
task_categories:
  - summarization
tags:
  - multi-document
  - single-document
  - abstractive
  - multilingual
configs:
  - config_name: single
    description: Single-document summarisation pairs (article + summary).
  - config_name: multi
    description: Multi-document summarisation clusters with 3 human summaries.

MultiLing Multilingual Summarisation Corpus

The MultiLing Multilingual Summarisation Corpus is a comprehensive multilingual benchmark for single-document and multi-document abstractive summarisation, originally created for the MultiLing 2011 and MultiLing 2013 shared tasks held under ACL.
This release consolidates, cleans, and reformats the original resources into a standard, machine-readable dataset suitable for modern sequence-to-sequence and large language model research.

The corpus covers ten languages:

Arabic, Chinese, Czech, English, French, Greek, Hebrew, Hindi, Romanian, Spanish

and contains both:

Single-document summarisation pairs (source article, 1 gold summary)
Multi-document summarisation clusters (10 related articles, 3 human-written abstractive summaries)

All texts originate from WikiNews, following the Creative Commons BY licence. See http://multiling.iit.demokritos.gr/

✨ Key Features

Single-Document Summarisation

40 languages in the original collection; the present release includes the main ten used for MultiLing 2013.
For each language, every document has:
- One source article
- One human abstractive summary
Fully parallel across languages.

Multi-Document Summarisation

Each topic consists of 10 articles describing an event sequence.
Topics appear consistently across all languages that contributed to that year’s task.
Each cluster includes three human-written abstractive summaries, produced independently.
Human summaries were constrained to 240–250 words (or equivalent byte limits for Chinese).

Parallel & Comparable Structure

The corpus was originally designed to allow:

Cross-lingual and multilingual summarisation
Comparative analyses of summarisation difficulty across languages
Multilingual evaluation of automatic summarisation metrics (ROUGE, AutoSummENG-MeMoG, NPowER)

📘 Source and Citation

This dataset is derived from the corpus described in:

Li L, Forăscu C, El-Haj M, Giannakopoulos G. (2013)
Multi-Document Multilingual Summarization Corpus Preparation, Part 1: Arabic, English, Greek, Chinese, Romanian.
In Proceedings of the MultiLing 2013 Workshop on Multilingual Multi-Document Summarization, pp. 1–12.
ACL 2013, Sofia, Bulgaria.
PDF: https://aclanthology.org/W13-3101.pdf

Please cite the paper above when using this dataset.

🗂 Dataset Structure

Single-Document Format

Each sample includes:

language — ISO folder name (ar, en, fr, etc.)
doc_id
document_text
summary

Multi-Document Format

Each sample includes:

cluster_id — e.g. M000, M014, M103
language
doc_ids — list of the ten document identifiers
documents_text — concatenated with <DOC id=…> wrappers
summary_1, summary_2, summary_3 — three reference human summaries

All files are provided in CSV and JSONL, with train/dev/test splits.

🔧 Recommended Use Cases

Multilingual abstractive summarisation
Cross-lingual evaluation of LLMs
Multi-document summarisation research
Training summarisation models on parallel news texts
Research on multilingual evaluation metrics
Cross-lingual transfer learning
Low-resource summarisation investigations

📊 Splits

The dataset is released with deterministic:

train
validation
test

splits for single-document and multi-document subsets.
For multi-document summarisation, splits are cluster-based to prevent data leakage.

📥 Loading the Dataset

from datasets import load_dataset

ds = load_dataset("YOUR_DATASET_NAME", "multi")
# or
ds = load_dataset("YOUR_DATASET_NAME", "single")

🔒 Licence

All texts originate from WikiNews under Creative Commons BY 2.5/3.0 licences. This consolidated dataset is released under CC-BY-4.0.

🙏 Acknowledgements

MultiLing is the result of a large international community effort involving contributors from more than ten universities and research centres. This cleaned and repackaged release builds on that original work to make the corpus more accessible for modern NLP research.