LaMuN (Large-scale Multilingual News Image Captioning Benchmark) is a large-scale multilingual dataset for news image captioning. It contains 836,436 news image-caption instances across 59 languages, including a large number of mid- and low-resource languages.

The dataset was collected from three international news organisations:

BBC News
Voice of America (VOA)
Deutsche Welle (DW)

The articles span the period from 2001 to December 2025.

Unlike generic image captioning, news image captioning requires models to combine visual information with the accompanying news article. Captions frequently contain information that cannot be inferred from the image alone, including names of people, organisations, locations, and events.

LaMuN is intended to support research on multilingual and low-resource multimodal language understanding, particularly the generation of contextually grounded captions for news images.

Supported Tasks

The primary task supported by LaMuN is:

News image captioning / Image-Text-to-Text: given a news image and its accompanying article, generate a concise journalistic caption describing the image in the target language.

The dataset may also be useful for related research including:

multilingual vision-language modelling;
multimodal representation learning;
entity-aware image captioning;
multimodal entity grounding;
image-article retrieval and alignment;
cross-lingual multimodal transfer; and
low-resource multimodal NLP.

Languages

LaMuN contains data in 59 languages:

Albanian, Amharic, Arabic, Armenian, Bengali, Bosnian, Bulgarian, Burmese, Central Kurdish, Chinese (Mandarin), Croatian, Dari, English, French, Georgian, German, Gujarati, Haitian Creole, Hausa, Hindi, Igbo, Indonesian, Japanese, Khmer, Kinyarwanda, Korean, Kurmanji Kurdish, Kyrgyz, Lao, Lingala, Macedonian, Marathi, Nepali, Nigerian Pidgin, North Ndebele, Oromo, Pashto, Persian, Polish, Punjabi, Romanian, Russian, Scottish Gaelic, Serbian, Shona, Sinhala, Somali, Swahili, Telugu, Thai, Tibetan, Tigrinya, Turkish, Ukrainian, Urdu, Uzbek, Vietnamese, Welsh, and Yoruba.

The collection spans multiple language families and includes high-, mid-, and low-resource languages.

Dataset Structure

Data Instances

Each instance represents an image associated with a news article. The dataset provides the information required for news image captioning, including:

the news image;
the associated news article/content;
the original image caption;
additional article metadata, such as the article title; and
the URL of the original news article.

Multiple images can originate from the same news article.

Data Splits

LaMuN contains training and test splits:

Split	Instances
Train	778,556
Test	57,880
Total	836,436

The dataset contains 383,395 unique news articles.

Most languages contain 1,000 test instances. Scottish Gaelic contains 585 test instances and Yoruba contains 295. Scottish Gaelic and Yoruba are test-only languages in the current release.

No separate development split is provided.

Dataset Creation

Curation Rationale

Existing news image captioning resources have predominantly focused on English. LaMuN was created to provide substantially broader linguistic coverage and to facilitate research on multilingual news image captioning, particularly for languages with comparatively limited multimodal resources.

Source Data

Data Sources

The dataset was constructed using publicly available news articles and associated images from:

BBC News;
Voice of America (VOA); and
Deutsche Welle (DW).

The sources were selected based on their multilingual coverage, journalistic standards, accessibility, and licensing considerations.

The collected material covers news published between 2001 and December 2025.

Data Collection and Processing

The original news records were collected without manually editing their content.

The following filtering was applied during dataset construction:

Images with either a height or width of less than 180 pixels were removed.
Instances with captions containing three words or fewer were removed.

The resulting dataset contains more than 800,000 image-caption instances across 59 languages.

Annotations

The captions in LaMuN are naturally occurring news captions obtained from the original news sources. They were not newly written or crowdsourced specifically for this dataset.

As a result, the captions reflect real-world journalistic captioning practices across the included news organisations and languages.

Personal and Sensitive Information

LaMuN is based on real-world news reporting and may therefore contain information about identifiable individuals, organisations, locations, and events.

Named entities are an important part of news image captioning and have not been systematically anonymised or removed.

Users should treat the dataset as real-world journalistic material and take appropriate care when processing information concerning identifiable individuals or sensitive events.

Considerations for Using the Data

Biases

LaMuN reflects the content and editorial practices of BBC News, Voice of America, and Deutsche Welle. Consequently, the dataset may reproduce biases associated with:

editorial and topic selection;
geographic coverage;
political and social reporting;
the representation of individuals and communities;
language-specific reporting practices; and
differences in coverage between news organisations.

The dataset should not be interpreted as a representative sample of all journalism or language use in any of the included languages.

Language and Source Imbalance

The number of examples varies substantially across languages. English and several other widely resourced languages contain considerably more instances than many low-resource languages.

News-source coverage also differs by language. Some languages contain material from all three news organisations, while others are represented by only one or two sources.

Researchers should account for these imbalances when comparing performance across languages or constructing multilingual training sets.

Other Known Limitations

LaMuN consists of naturally occurring news data rather than manually standardised annotations. Caption length, writing style, article length, topic distribution, and the number of images associated with an article therefore vary across languages and news organisations.

The dataset also represents the news coverage available from the three selected sources during the collection period and does not provide comprehensive coverage of all news outlets, regions, communities, or viewpoints.

Additional Information

Licensing and Usage

LaMuN was collected from publicly available material published by BBC News, Voice of America, and Deutsche Welle.

The collection process adhered to the terms of service of the respective sources and respected access restrictions specified through their robots.txt directives.

For each instance, the dataset provides the URL of the original news article to preserve attribution and allow verification of the source material.

LaMuN is released for non-commercial research purposes. Users are responsible for ensuring that their use of the underlying content complies with the terms and conditions of the respective news providers.

The dataset should not be interpreted as transferring ownership or licensing rights over the original images or news content from their respective copyright holders.

Dataset Curators

LaMuN was created by:

Yuji Chen — Lancaster University
Purushoth Velayuthan — University of Nevada, Reno
Alistair Plum — University of Luxembourg
Hansi Hettiarachchi — Lancaster University
Saroj Basnet — George Mason University
Menan Velayuthan — Utrecht University
Marcos Zampieri — George Mason University
Tharindu Ranasinghe — Lancaster University

Citation

If you use LaMuN in your research, please cite:

@inproceedings{chen2026lamun,
  title     = {Large-scale Multilingual News Image Captioning with LLMs},
  author    = {Chen, Yuji and
               Velayuthan, Purushoth and
               Plum, Alistair and
               Hettiarachchi, Hansi and
               Basnet, Saroj and
               Velayuthan, Menan and
               Zampieri, Marcos and
               Ranasinghe, Tharindu},
  year      = {2026}
}

Downloads last month: 34

Total file size:

14.3 GB