MLSUM: The Multilingual Summarization Corpus

• Document summarization combines long-document understanding, relevance discrimination, and short synthesis.Source: §1 · p. 1
• Most available summarization datasets are in English, concentrating research efforts on English-language data.Source: §1 · p. 1
• Multilingual pretraining and transfer learning enable models to apply task knowledge from pivot languages to other languages.Source: §1 · p. 1
• Pivot-language training remains limited because it prevents testing whether models fit one language as well as another, especially for low-resource languages.Source: §1 · p. 1
• MLSUM contributes the first large-scale multilingual summarization dataset, strong multilingual abstractive baselines, and comparative cross-lingual analysis.Source: §1 · pp. 1–2

• Earlier multilingual summarization work included MEAD, genetic-algorithm approaches, and the MultiLing benchmark, which covered 40 languages but offered relatively few examples.Source: §2.1 · p. 2
• Translation-based multilingual summarization for low-resource languages showed only slight baseline improvements and remained far from human performance.Source: §2.1 · p. 2
• Multilingual transfer datasets such as XNLI and MLQA evaluate classification or question answering, but rely on human translations and provide evaluation data.Source: §2.2 · p. 2
• XTREME covers 40 languages and nine tasks, but excludes summarization.Source: §2.2 · p. 2
• XGLUE includes multilingual news-title generation for five languages, but not document summarization.Source: §2.2 · p. 2
• Existing summarization corpora span English news, patents, social media, and other domains, with differing scale, references, and supervision.DUC provides multiple reference summaries but little training data; Gigaword uses headlines generated from article openings rather than human summaries.Source: §2.3 · p. 3

• MLSUM is introduced as the first large-scale multilingual summarization corpus, with more than 1.5 million articles in French, German, Spanish, Turkish, and Russian.Source: §3 · p. 3
• The corpus is designed as a multilingual extension of CNN/Daily Mail, using similarly constructed news data and comparable training-sample amounts except for Russian.Source: §3 · p. 3, §3.1 · p. 4
• The paper reports corpus methodology, language statistics, and baseline or state-of-the-art model performance, with article URLs and construction code supporting replication.Source: §3 · p. 4, §3.1 · p. 4
• Newspapers were selected for broad topic coverage, large public archives, and human-written highlights or summaries extractable from webpages.Source: §3.1 · p. 4
• Articles were crawled from 2010 to 2019, discarding articles under 50 words or summaries under 10 words.Source: §3.1 · p. 4
• Russian has ten times fewer training samples than the other MLSUM languages, with the shortest and most abstractive summaries.Source: §3.2 · pp. 4–5

• Experiments train supervised and unsupervised, extractive and abstractive models separately for each language.Source: §4 · p. 5
• The mixed-language model produced no significant performance difference from per-language training.Source: §4 · p. 5
• Oracle selects input sentences maximizing ROUGE-L against the reference summary, indicating the extractive upper bound.Source: §4.1 · p. 5
• Random extracts N words randomly, with N fixed to the average summary length, providing an unbiased reference point.Source: §4.1 · p. 5
• Lead-3 selects the first three input sentences, while TextRank extracts the most central sentences using sentence co-similarities.Source: §4.1 · p. 5
• Abstractive systems use encoder-decoder generation, including Pointer-Generator copying out-of-vocabulary tokens and M-BERT-based multilingual generation.Source: §4.2 · pp. 5–6

Metrics

• ROUGE measures n-gram overlap between an evaluated summary and a human reference summary.Source: §5 · p. 6
• METEOR evaluates machine-generated text using a harmonic mean of unigram precision and recall, weighting recall more heavily.Source: §5 · p. 6
• Novelty reports the percentage of novel n-grams in generated summaries to assess reliance on extraction and copying.Source: §5 · p. 6

Multilingual evaluation

• Existing neural metrics based on question answering, factualness discrimination, or human annotations were trained only on English datasets.Source: §5 · pp. 6–7

Analysis factors

• The analysis attributes cross-language performance differences to dataset properties and language-specific metric or model biases.Dataset factors include article structure, summary abstractiveness, and data quantity.Source: §6 · p. 7
• Multilingual datasets are needed to study language-specific effects because pivot-language evaluations cannot isolate them.Source: §6 · p. 7

Overall results

• The overall system ranking is preserved across ROUGE-L and METEOR, apart from close swaps between Lead 3 and Pointer Generator.Source: §6 · p. 7

Russian

• Russian has comparatively low performance, reflecting its highly abstractive corpus and roughly tenfold smaller training set.Pretraining increases Russian Pointer Generator performance by +66% in ROUGE-L when using M-BERT.Source: §6 · p. 7

Abstractiveness

• Pointer-generator networks are poorly abstractive, whereas M-BERT is consistently more abstractive, particularly for Russian.The paper links this pattern to copying behavior and exposure to additional texts during pretraining.Source: §6.1 · p. 7

Dataset abstractiveness

• Oracle performance is lower for French and Spanish than for English and German despite similar novelty rates across German, French, and Spanish.Lead-3 ROUGE-L is 35.20 for German and 33.09 for English, versus 19.69 for French and 13.70 for Spanish.Source: §6.2 · p. 8

Model biases

• TextRank performs close to Pointer Generator on English corpora but substantially worse on other languages, suggesting an English bias.The reported TextRank-to-Pointer Generator ratio is 0.85 to 1.21 on English corpora and 0.37 to 0.65 elsewhere.Source: §6.2 · pp. 8–9
• M-BERT consistently outperforms Pointer Generator, with larger gains for German than French, potentially reflecting language-specific effects of self-attention.The paper hypothesizes that self-attention may better preserve information in languages with freer word order.Source: §6.2 · p. 9

• MLSUM contains over 1.5M article-summary pairs in French, German, Russian, Spanish, and Turkish.Source: §7 · p. 10
• The dataset complements CNN/Daily Mail English data and supports comparative analysis of state-of-the-art summarization approaches across languages.Source: §7 · p. 10
• The experiments highlight biases in existing summarization models and differences in model performance across languages.Source: §7 · p. 10
• Future work includes adding Arabic and Hindi and adapting neural evaluation metrics to multilingual summarization.Source: §7 · p. 10

• The Karzai clan’s regional influence is rooted in Kandahar and shapes national political networks.Source: – FRENCH – · p. 14
• Kandahar is presented as a historical center of Afghan state formation and a source of political legitimacy distinct from Kabul.Source: – FRENCH – · p. 14
• Kandahar and the surrounding south are described as the historical heartland of the Taliban insurgency.Source: – FRENCH – · p. 14
• Taliban influence is associated with growing insecurity, limited employment, intimidation, and the departure of educated residents toward Kabul.Source: – FRENCH – · p. 14
• The United States increasingly directed operations in Helmand and Kandahar after sending successive reinforcements.Source: – FRENCH – · p. 14

• Die Unabhängigkeit hatte hohe menschliche und wirtschaftliche Kosten: Ein Drittel der Bevölkerung starb in den Kämpfen, und viele Plantagen wurden zerstört.Source: – GERMAN – · p. 15
• Frankreich und andere Kolonialmächte verweigerten Haiti zunächst die Anerkennung und unterstützten ein Embargo sowie Forderungen nach Reparationszahlungen.Source: – GERMAN – · p. 15
• Frankreich verlangte 150 Millionen Francs für die Anerkennung Haitis und drohte andernfalls mit einer Invasion und erneuter Versklavung.Source: – GERMAN – · p. 15
• Haiti nahm Schulden auf und zahlte; die Schuldenlast lähmte bis 1947 die Wirtschaft und legte laut Text den Grundstein für Armut und Korruption.Source: – GERMAN – · p. 15

• Bis zu 60 Minuten Verspätung mit bereits an Bord befindlichen Passagieren meldeten mehrere Fluggesellschaften.Source: – SPANISH – · p. 16
• Obwohl der Flughafen ab 15.00 Uhr wieder alle vier Start- und Landebahnen betrieb, verzeichneten die Flüge weiterhin erhebliche Verspätungen.Source: – SPANISH – · p. 16
• 30 Minuten betrugen die durchschnittlichen Verspätungen, nachdem die Abwesenheit von 5 der 18 eingeplanten Fluglotsen zwei Start- und Landebahnen geschlossen hatte.Source: – SPANISH – · p. 16

• Lehramtsbewerber ohne Anstellung demonstrierten in Ankara wegen fehlender Stellen und erhielten Unterstützung von TEKEL-Arbeitern sowie Muharrem İnce.Source: – TURKISH – · p. 16
• Muharrem İnce kritisierte, dass Schulen ohne Lehrkräfte blieben, während viele Lehramtsabsolventen arbeitslos seien, und verwies auf unterschiedliche Einstellungszahlen nach Fächern.Source: – TURKISH – · p. 16
• Die Plattform erklärte, jährlich würden Absolventen wegen unzureichender Bildungspolitik arbeitslos; bei ausbleibenden Einstellungen kündigte sie Streiks und einen Hungerstreik an.Source: – TURKISH – · p. 16
• Die Demonstrierenden forderten reguläre Einstellungen, eine Begrenzung der Lehrerausbildung am tatsächlichen Bedarf und ein transparenteres Verfahren statt KPSS.Source: – TURKISH – · p. 16