We (NPO openEHR Japan) have started translating a set of archetypes into Japanese, first the ones that make up a referral letter (COMPOSITION.request, INSTRUCTION.service_request, EVALUATION.reason_for_encounter, SECTION.referral_details), with about 35 more planned (problem/diagnosis, adverse reaction, medication order, lab results, vital signs). Before we go further I would like to check with the CKM editors and other translation communities that the way we are working does not disturb the established translation process, and to ask what the appropriate way to contribute translations is when they are produced this way.
What we are doing
Instead of translating item by item in the CKM translation editor, we:
- download the current published ADL of the archetype from CKM;
- extract all translatable strings (purpose/use/misuse/keywords and text/description/comment per at-code) into a table;
- produce a draft with an LLM that is constrained by a project glossary and style guide, and must flag every term it is unsure about;
- have a clinician review the flagged terms and the glossary changes;
- inject the reviewed strings back into the ADL as a
["ja"]block — thedefinitionsection, the other languages and the metadata (uid,build_uid,revision) are left byte-identical, and a script checks that the ja code set equals the en code set with no placeholders left; - upload the resulting ADL to CKM through the translation upload, and then proceed with the normal review.
The work is kept in a public Git repository ( GitHub - skoba/openehr-japanese-translation: openEHR Japanese Translation · GitHub ) with one issue and one pull request per archetype, so every translation decision is traceable.
What I would like to ask
- Is uploading a translated ADL an intended and acceptable way to submit a translation, or is it considered a side path that bypasses something the translation editor does (translator roles, review assignment, change tracking, attribution)? We found that the upload works and that the file must keep the original file name, but we could not find this documented and do not want to rely on behaviour that is not meant to be used this way.
- Attribution. We put the reviewing clinician (and NPO openEHR Japan) as the translator in the
translationssection, not the tool. Is there a convention or expectation for AI-assisted translations — for example, a note inother_detailsor in the translation’s accreditation field — so that reviewers know how the draft was produced? - Archetypes under reassessment. For an archetype whose published revision is, say, 1.1.2 while the trunk is 1.1.3-alpha, should the translation be submitted against the published revision and then updated for the changed nodes after publication, or against the development revision? We chose the published one, but I would like to follow whatever the editors prefer.
- Reviewing. Does a translation submitted this way go through the same review as one made in the editor, and is there anything we can do to make the reviewer’s job easier — for example, attaching the list of flagged terms with our reasoning?
- If this workflow is acceptable, would it be useful to write it up (and make the tooling language-agnostic) for other translation communities, or would the editors rather keep translation inside the CKM editor for consistency?
We are happy to change any part of this to fit the CKM process. The goal is only to get consistent, reviewed Japanese translations into CKM faster than we could by hand, not to work around the process.
Shinji Kobayashi NPO openEHR Japan / Gifu University