mlsumSummarization

mlsum

MLSum is a multilingual summarization dataset crawled from different news websites. The GEM version supports the German and Spanish subset alongside specifically collected challenge sets for COVID-related articles to test out-of-domain generalization.

You can load the dataset via:

import datasets
data = datasets.load_dataset('GEM/mlsum')

The data loader can be found here.

website

N/A

paper

ACL Anthology

authors

Thomas Scialom, Paul-Alexis Dray, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano

Quick-Use

Contact Name

If known, provide the name of at least one person the reader can contact for questions about the dataset.

Thomas Scialom

Multilingual?

Is the dataset multilingual?

yes

Covered Languages

What languages/dialects are covered in the dataset?

German, Spanish, Castilian

License

What is the license of the dataset?

other: Other license

Communicative Goal

Provide a short description of the communicative goal of a model trained for this task on this dataset.

The speaker is required to produce a high quality summary of news articles in the same language as the input article.

Additional Annotations?

Does the dataset have additional annotations for each instance?

none

Contains PII?

Does the source language data likely contain Personal Identifying Information about the data creators or subjects?

yes/very likely

Dataset Overview

  • Where to find the Data and its Documentation

  • Languages and Intended Use

  • Credit

  • Dataset Structure

Where to find the Data and its Documentation

Download

What is the link to where the original dataset is hosted?

Gitlab

Paper

What is the link to the paper describing the dataset (open access preferred)?

ACL Anthology

BibTex

Provide the BibTex-formatted reference for the dataset. Please use the correct published version (ACL anthology, etc.) instead of google scholar created Bibtex.

@inproceedings{scialom-etal-2020-mlsum,
title = "{MLSUM}: The Multilingual Summarization Corpus",
author = "Scialom, Thomas  and
Dray, Paul-Alexis  and
Lamprier, Sylvain  and
Piwowarski, Benjamin  and
Staiano, Jacopo",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2020.emnlp-main.647",
doi = "10.18653/v1/2020.emnlp-main.647",
pages = "8051--8067",
abstract = "We present MLSUM, the first large-scale MultiLingual SUMmarization dataset. Obtained from online newspapers, it contains 1.5M+ article/summary pairs in five different languages {--} namely, French, German, Spanish, Russian, Turkish. Together with English news articles from the popular CNN/Daily mail dataset, the collected data form a large scale multilingual dataset which can enable new research directions for the text summarization community. We report cross-lingual comparative analyses based on state-of-the-art systems. These highlight existing biases which motivate the use of a multi-lingual dataset.",
}
Contact Name

If known, provide the name of at least one person the reader can contact for questions about the dataset.

Thomas Scialom

Contact Email

If known, provide the email of at least one person the reader can contact for questions about the dataset.

{thomas,paul-alexis,jacopo}@recital.ai, {sylvain.lamprier,benjamin.piwowarski}@lip6.fr

Has a Leaderboard?

Does the dataset have an active leaderboard?

no

Languages and Intended Use

Multilingual?

Is the dataset multilingual?

yes

Covered Dialects

What dialects are covered? Are there multiple dialects per language?

There is only one dialect per language, Hochdeutsch for German and Castilian Spanish for Spanish.

Covered Languages

What languages/dialects are covered in the dataset?

German, Spanish, Castilian

Whose Language?

Whose language is in the dataset?

The German articles are crawled from Süddeutsche Zeitung and the Spanish ones from El Pais.

License

What is the license of the dataset?

other: Other license

Intended Use

What is the intended use of the dataset?

The intended use of this dataset is to augment existing datasets for English news summarization with additional languages.

Add. License Info

What is the 'other' license of the dataset?

Restricted to non-commercial research purposes.

Primary Task

What primary task does the dataset support?

Summarization

Communicative Goal

Provide a short description of the communicative goal of a model trained for this task on this dataset.

The speaker is required to produce a high quality summary of news articles in the same language as the input article.

Credit

Curation Organization Type(s)

In what kind of organization did the dataset curation happen?

other

Curation Organization(s)

Name the organization(s).

CNRS, Sorbonne Université, reciTAL

Dataset Creators

Who created the original dataset? List the people involved in collecting the dataset and their affiliation(s).

Thomas Scialom, Paul-Alexis Dray, Sylvain Lamprier, Benjamin Piwowarski, Jacopo Staiano

Funding

Who funded the data creation?

Funding information is not specified.

Who added the Dataset to GEM?

Who contributed to the data card and adding the dataset to GEM? List the people+affiliations involved in creating this data card and who helped integrate this dataset into GEM.

The original data card was written by Pedro Henrique Martins (Instituto de Telecomunicações) and Sebastian Gehrmann (Google Research) extended and updated it to the v2 format. The COVID challenge set was created by Laura Perez-Beltrachini (University of Edinburgh). Data cleaning was done by Juan Diego Rodriguez (UT Austin).

Dataset Structure

Data Fields

List and describe the fields present in the dataset.

The data fields are:

  • text: the source article (string).
  • summary: the output summary (string).
  • topic: the topic of the article (string).
  • url: the article's url (string).
  • title: the article's title (string).
  • date: the article's date (string).
Reason for Structure

How was the dataset structure determined?

The structure follows previously released datasets. The topic and title fields were added to enable additional tasks like title generation and topic detection.

How were labels chosen?

How were the labels chosen?

They are human written highlights or summaries scraped from the same website.

Example Instance

Provide a JSON formatted example of a typical instance in the dataset.

{
'date': '00/01/2010',
'gem_id': 'mlsum_de-train-2',
'gem_parent_id': 'mlsum_de-train-2',
'references': [],
'target': 'Oskar Lafontaine gibt den Parteivorsitz der Linken ab - und seine Kollegen streiten, wer ihn beerben soll. sueddeutsche.de stellt die derzeit aussichtsreichsten Anwärter für Führungsaufgaben vor. Mit Vote.',
'text': 'Wenn an diesem Montag die Landesvorsitzenden der Linken über die Nachfolger der derzeitigen Chefs Lothar Bisky und Oskar Lafontaine sowie des Bundesgeschäftsführers Dietmar Bartsch beraten, geht es nicht nur darum, wer die Partei führen soll. Es geht auch um die künftige Ausrichtung und Stärke einer Partei, die vor allem von Lafontaine zusammengehalten worden war. Ihm war es schließlich vor fünf Jahren gelungen, aus der ostdeutschen PDS und der westedeutschen WASG eine Partei zu formen. Eine Partei allerdings, die zerrissen ist in Ost und West, in Regierungswillige und ewige Oppositionelle, in Realos und Ideologen, in gemäßigte und radikale Linke. Wir stellen mögliche Kandidaten vor. Stimmen Sie ab: Wen halten Sie für geeignet und wen für unfähig? Kampf um Lafontaines Erbe: Gregor Gysi Sollte überhaupt jemand die Partei alleine führen, wie es sich viele Ostdeutsche wünschen, käme dafür wohl nur der 62-jährige Gregor Gysi in Betracht. Er ist nach Lafontaine einer der bekanntesten Politiker der Linken und derzeit Fraktionsvorsitzender der Partei im Bundestag. Allerdings ist der ehemalige PDS-Vorsitzende und Rechtsanwalt nach drei Herzinfarkten gesundheitlich angeschlagen. Wahrscheinlich wäre deshalb, dass er die zerstrittene Partei nur übergangsweise führt. Doch noch ist nicht klar, ob eine Person allein die Partei führen soll oder eine Doppelspitze. Viele Linke wünschen sich ein Duo aus einem westdeutschen und einem ostdeutschen Politiker, Mann und Frau. Foto: Getty Images',
'title': 'Personaldebatte bei der Linken - Wer kommt nach Lafontaine?',
'topic': 'politik',
'url': 'https://www.sueddeutsche.de/politik/personaldebatte-bei-der-linken-wer-kommt-nach-lafontaine-1.70041'
}
Data Splits

Describe and name the splits in the dataset if there are more than one.

The statistics of the original dataset are:

Dataset Train Validation Test Mean article length Mean summary length
German 242,982 220,887 11,394 10,701 570.6 (words) 30.36 (words)
Spanish 290,645 266,367 10,358 13,920 800.5 (words) 20.71 (words)

The statistics of the cleaned version of the dataset are:

Dataset Train Validation Test
German 242,835 220,887 11,392 10,695
Spanish 283,228 259,886 9,977 13,365

The COVID challenge sets have 5058 (de) and 1938 (es) examples.

Splitting Criteria

Describe any criteria for splitting the data, if used. If there are differences between the splits (e.g., if the training annotations are machine-generated and the dev and test ones are created by humans, or if different numbers of annotators contributed to each example), describe them here.

The training set contains data from 2010 to 2018. Data from 2019 (~10% of the dataset) is used for validation (up to May) and testing(May-December 2019).

What does an outlier of the dataset in terms of length/perplexity/embedding look like?

Some topics are less represented within the dataset (e.g., Financial news in German and Television in Spanish).

Dataset in GEM

  • Rationale for Inclusion in GEM

  • GEM-Specific Curation

  • Getting Started with the Task

Rationale for Inclusion in GEM

Why is the Dataset in GEM?

What does this dataset contribute toward better generation evaluation and why is it part of GEM?

As the first large-scale multilingual summarization dataset, it enables evaluation of summarization models beyond English.

Similar Datasets

Do other datasets for the high level task exist?

yes

Unique Language Coverage

Does this dataset cover other languages than other datasets for the same task?

yes

Difference from other GEM datasets

What else sets this dataset apart from other similar datasets in GEM?

In our configuration, the dataset is fully non-English.

Ability that the Dataset measures

What aspect of model ability can be measured with this dataset?

Content Selection, Content Planning, Realization

GEM-Specific Curation

Modificatied for GEM?

Has the GEM version of the dataset been modified in any way (data, processing, splits) from the original curated data?

yes

GEM Modifications

What changes have been made to he original dataset?

data points removed, data points added

Modification Details

For each of these changes, described them in more details and provided the intended purpose of the modification

The modifications done to the original dataset are the following:

  • Selection of 2 languages (Spanish and German) out of the dataset 5 languages due to copyright restrictions.
  • Removal of duplicate articles.
  • Manually removal of article-summary pairs for which the summary is not related to the article.
  • Removal of article-summary pairs written in a different language (detected using the langdetect library).
Additional Splits?

Does GEM provide additional splits to the dataset?

yes

Split Information

Describe how the new splits were created

For both selected languages (German and Spanish), we compiled time-shifted test data in the form of new articles for the second semester of 2020 with Covid19-related keywords. We collected articles from the same German and Spanish outlets as the original MLSUM datasets (El Pais and Süddeutsche Zeitung). We used the scripts provided for the re-creation of the MLSUM datasets. The new challenge test set for German contains 5058 instances and the Spanish one contains 1938.

We additionally sample 500 training and validation points as additional challenge sets to measure overfitting.

Split Motivation

What aspects of the model's generation capacities were the splits created to test?

Generalization to unseen topics.

Getting Started with the Task

Previous Results

  • Previous Results

Previous Results

Measured Model Abilities

What aspect of model ability can be measured with this dataset?

Content Selection, Content Planning, Realization

Metrics

What metrics are typically used for this task?

METEOR, ROUGE, Other: Other Metrics

Other Metrics

Definitions of other metrics

Novelty: Number of generated n-grams not included in the source articles.

Proposed Evaluation

List and describe the purpose of the metrics and evaluation methodology (including human evaluation) that the dataset creators used when introducing this task.

ROUGE and METEOR both measure n-gram overlap with a focus on recall and are standard summarization metrics. Novelty is often reported alongside them to characterize how much a model diverges from its inputs.

Previous results available?

Are previous results available?

yes

Other Evaluation Approaches

What evaluation approaches have others used?

The GEM benchmark results (https://gem-benchmark.com/results) report a wide range of metrics include lexical overlap metrics but also semantic ones like BLEURT and BERT-Score.

Dataset Curation

  • Original Curation

  • Language Data

  • Structured Annotations

  • Consent

  • Private Identifying Information (PII)

  • Maintenance

Original Curation

Original Curation Rationale

Original curation rationale

The rationale was to create a multilingual news summarization dataset that mirrors the format of popular English datasets like XSum or CNN/DM.

Communicative Goal

What was the communicative goal?

The speaker is required to produce a high quality summary of news articles in the same language as the input article.

Sourced from Different Sources

Is the dataset aggregated from different data sources?

yes

Source Details

List the sources (one per line)

www.lemonde.fr www.sueddeutsche.de www.elpais.com www.mk.ru www.internethaber.com

Language Data

How was Language Data Obtained?

How was the language data obtained?

Found

Where was it found?

If found, where from?

Multiple websites

Language Producers

What further information do we have on the language producers?

The language producers are professional journalists.

Topics Covered

Does the language in the dataset focus on specific topics? How would you describe them?

4/5 of the original languages report their topics (except Turkish) and the distributions differ between sources. The dominant topics in German are Politik, Sport, Wirtschaft (economy). The dominant topics in Spanish are actualidad (current news) and opinion. French and Russian are different as well but we omit these languages in the GEM version.

Data Validation

Was the text validated by a different worker or a data curator?

not validated

Was Data Filtered?

Were text instances selected or filtered?

algorithmically

Filter Criteria

What were the selection criteria?

In the original dataset, only one filter was applied: all the articles shorter than 50 words or summaries shorter than 10 words are discarded.

The GEM version additionally applies langID filter to ensure that articles are in the correct language.

Structured Annotations

Additional Annotations?

Does the dataset have additional annotations for each instance?

none

Annotation Service?

Was an annotation service used?

no

Consent

Any Consent Policy?

Was there a consent policy involved when gathering the data?

no

Justification for Using the Data

If not, what is the justification for reusing the data?

The copyright remains with the original data creators and the usage permission is restricted to non-commercial uses.

Private Identifying Information (PII)

Contains PII?

Does the source language data likely contain Personal Identifying Information about the data creators or subjects?

yes/very likely

Categories of PII

What categories of PII are present or suspected in the data?

sensitive information, generic PII

Any PII Identification?

Did the curators use any automatic/manual method to identify PII in the dataset?

no identification

Maintenance

Any Maintenance Plan?

Does the original dataset have a maintenance plan?

no

Broader Social Context

  • Previous Work on the Social Impact of the Dataset

  • Impact on Under-Served Communities

  • Discussion of Biases

Previous Work on the Social Impact of the Dataset

Usage of Models based on the Data

Are you aware of cases where models trained on the task featured in this dataset ore related tasks have been used in automated systems?

no

Impact on Under-Served Communities

Addresses needs of underserved Communities?

Does this dataset address the needs of communities that are traditionally underserved in language technology, and particularly language generation technology? Communities may be underserved for exemple because their language, language variety, or social or geographical context is underepresented in NLP and NLG resources (datasets and models).

no

Discussion of Biases

Any Documented Social Biases?

Are there documented social biases in the dataset? Biases in this context are variations in the ways members of different social categories are represented that can have harmful downstream consequences for members of the more disadvantaged group.

no