Skip to content

Latest commit

 

History

History
58 lines (45 loc) · 2.69 KB

README.md

File metadata and controls

58 lines (45 loc) · 2.69 KB

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.


Language Num. parallel
English 141,457
Spanish 113,215
Portuguese 81,695
French 63,692
German 58,375
Russian 52,928
Italian 50,968
Indonesian 47,511
Dutch 31,270
Arabic 29,229
Vietnamese 19,600
Chinese 18,887
Thai 14,770
Japanese 12,669
Korean 12,189
Hindi 9,929
Czech 7,200
Turkish 4,503

License