OpenWHO: A Document-Level Parallel Corpus for Health Translation

Dataset Description

OpenWHO is a document-level parallel corpus designed for MT evaluation in the health domain, with a special focus on low-resource languages. Sourced from the World Health Organization's (WHO) multilingual e-learning platform, OpenWHO.org, the dataset contains expert-authored, professionally translated educational materials for health workers.

• Domain: Public health education (e.g., COVID-19, Ebola, vaccination protocols, infection prevention).
• Source: Expert-authored and professionally translated content from the WHO.
• Format: Provided at both the document and sentence level.
• Size: 2,978 documents and 26,824 sentences across languages (ignoring the EN source).
• Languages: English paired with 20+ languages, including 9 low-resource ones: Georgian, Armenian, Azerbaijani, Kazakh, Tetun, Albanian, Somali, Sinhala, and Tamil.
• Clean: Shielded from web-crawling, making it a reliable test set for large models.

Languages

The full list of languages, their ISO 639-1 codes, and sentence counts are provided below. Low-resource languages are in bold.

Language	ISO 639-1	Script	Documents	Sentences
Macedonian	mk	Cyrillic	254	3695
Arabic	ar	Arabic	293	2623
French	fr	Latin	301	2385
Russian	ru	Cyrillic	315	2194
Georgian	ka	Georgian	131	2151
Armenian	hy	Armenian	125	1982
Azerbaijani	az	Latin	98	1677
Ukrainian	uk	Cyrillic	204	1632
Turkish	tr	Latin	81	1093
Tetun	tet	Latin	67	1086
Dutch	nl	Latin	77	1082
Albanian	sq	Latin	74	1029
Kazakh	kk	Cyrillic	103	936
Chinese	zh	Chinese	203	871
Spanish	es	Latin	149	596
Indonesian	id	Latin	80	329
Portuguese	pt	Latin	62	279
Somali	so	Latin	20	224
Sinhala	si	Sinhala	25	214
Tamil	ta	Tamil	26	207
Swahili	sw	Latin	14	107
... and 20+ other languages with fewer sentences.

Dataset Structure

The dataset has two main levels of granularity: document-level and sentence-level. This is reflected in the dataset configurations.

Data Configurations

You can load either documents or sentences. Each language has its own configuration.

• Documents: Use doc__{lang_iso3} to load monolingual documents. For example, doc__eng for English, doc__kat for Georgian.
• Sentences: Use sent__{lang_iso3} to load parallel sentences. For example, sent__sqi for English-Albanian pairs.

from datasets import load_dataset

# Load English-Albanian parallel sentences
en_sqi_sentences = load_dataset("raphaelmerx/openwho", "sent__sqi", split="train")

# Load all English documents
en_docs = load_dataset("raphaelmerx/openwho", "doc__eng", split="train")

To create a parallel document dataset, you can load the source and target documents and align them using their unique identifiers (course_id, section_index, subsection_index).

Data Fields

Sentence Configuration (sent__{lang_iso3})

A row in the sentence-level data represents a single translated sentence pair.

• src_lang: The source language code (always 'English').
• tgt_lang: The ISO 639-3 code of the target language.
• src_text: The sentence in the source language (English).
• tgt_text: The sentence in the target language.
• course_id: Identifier for the course the text is from.
• section_index: The index of the section within the course.
• subsection_index: The index of the subsection within the section.

Example (sent__sqi):

{
  "src_lang": "English",
  "tgt_lang": "sqi",
  "src_text": "The exposed individual is HIV positive.",
  "tgt_text": "Individi i ekspozuar është HIV pozitiv.",
  "course_id": "IPC-HH",
  "section_index": 1,
  "subsection_index": 2
}

Document Configuration (doc__{lang_iso3})

A row in the document-level data represents a full-text page from a course.

• text: The full text of the document/page.
• lang_iso3: The ISO 639-3 code of the language.
• course_id: Identifier for the course the text is from.
• section_index: The index of the section within the course.
• subsection_index: The index of the subsection within the section.

Example (doc__eng):

{
  "text": "Introduction to Hand Hygiene\n\nWelcome to this course on Infection Prevention and Control through Hand Hygiene. In this first module, we will discuss the importance of hand hygiene in preventing healthcare-associated infections.",
  "lang_iso3": "eng",
  "course_id": "IPC-HH",
  "section_index": 1,
  "subsection_index": 1
}

Citation Information

If you use this dataset in your research, please cite our paper:

@inproceedings{merx-etal-2025-openwho,
    title = "{O}pen{WHO}: A Document-Level Parallel Corpus for Health Translation in Low-Resource Languages",
    author = "Merx, Raphael  and
      Suominen, Hanna  and
      Cohn, Trevor  and
      Vylomova, Ekaterina",
    editor = "Haddow, Barry  and
      Kocmi, Tom  and
      Koehn, Philipp  and
      Monz, Christof",
    booktitle = "Proceedings of the Tenth Conference on Machine Translation",
    month = nov,
    year = "2025",
    address = "Suzhou, China",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.wmt-1.8/",
    doi = "10.18653/v1/2025.wmt-1.8",
    pages = "142--160",
    ISBN = "979-8-89176-341-8",
}