Awesome

WikiLingua: A Multilingual Abstractive Summarization Dataset

UPDATE:
We have created new Train/Test splits for all 17 languages that can be downloaded here. These splits were created to ensure that there is no (document, summary) pair overlap across any of the 17 languages so that they can be safely used for multilingual evaluations.

This repo contains dataset introduced in the following paper:

WikiLingua: A New Benchmark Dataset for Multilingual Abstractive Summarization

Download the dataset using this link.

Please refer to this Collab notebook to see how to align articles in other languages with the parallel English articles.

Reference

Please cite the following paper:

@inproceedings{ladhak-wiki-2020,
    title={WikiLingua: A New Benchmark Dataset for Multilingual Abstractive Summarization},
    author={Faisal Ladhak, Esin Durmus, Claire Cardie and Kathleen McKeown},
    booktitle={Findings of EMNLP, 2020},
    year={2020}
}

Description

The dataset includes ~770k article and summary pairs in 18 languages from WikiHow. We extracted gold-standard article-summary alignments across languages by aligning the images that are used to describe each how-to step in an article.

The table below shows number of article-summary pairs with a parallel article-summary pair in English.

License

• Article provided by wikiHow https://www.wikihow.com/Main-Page, a wiki building the world's largest, highest quality how-to manual. Please edit this article and find author credits at wikiHow.com. Content on wikiHow can be shared under a Creative Commons license.

• Refer to this webpage for the specific attribution guidelines.