Domain-Specific AI Translation: Glossaries & Prompts
The word “protocol” means something very different in medicine than in computer networking. Domain-specific translation is where general AI translation skills meet the precision demands of specialized fields — and where a single wrong term can have serious consequences.
Why Domains Are Hard
General-purpose AI models know a little about everything, but specialized fields have vocabulary that:
- Carries precise meanings — a legal “party” isn’t a celebration
- Varies by jurisdiction or region — medical terminology differs between US and UK English, let alone across languages
- Must be consistent — if you translate “compliance” as “conformité” on page 1, it can’t become “respect des règles” on page 5
Without guidance, AI models default to the most common meaning of a word — which is often the wrong one in a specialized context.
Glossary Priming
The most effective technique is including a glossary in your prompt — a list of terms with their required translations:
Translate the following medical discharge summary from English
to Spanish (Latin America).
Audience: Patients and their families (use accessible language)
Domain: Cardiology
Glossary:
- "myocardial infarction" = "infarto de miocardio"
- "stent placement" = "colocación de stent"
- "blood thinner" = "anticoagulante"
- "follow-up appointment" = "cita de seguimiento"
Text to translate:
[discharge summary]A glossary of even 10-15 key terms dramatically improves consistency and accuracy for domain content.
Context Priming
Beyond terminology, tell the model about the domain:
You are translating a software user interface from English to German.
Context:
- This is a B2B analytics dashboard
- Users are data analysts familiar with technical terms
- UI labels should be concise (max 3 words where possible)
- Keep established English terms that German users expect
(e.g., "Dashboard", "Login", "Streaming")Context priming helps the model understand not just what words mean, but how they’re used in practice — the register, the audience’s expectations, and the conventions of the field.
Domain-Specific Challenges
| Domain | Key challenge | Prompt strategy |
|---|---|---|
| Legal | Jurisdiction-specific terms, untranslatable concepts | Specify legal system and target jurisdiction |
| Medical | Patient-facing vs. clinician-facing language | Specify audience and reading level |
| Technical | Product names, code terms, UI strings | Define which terms to keep in English |
| Financial | Number formats, currency conventions, regulatory terms | Specify locale and regulatory framework |
A Practical Tip
Build your glossary incrementally. Start with the 10 most critical terms, translate a section, review the output, then add terms the model got wrong. After a few iterations, you’ll have a reliable glossary you can reuse across documents in that domain.
Of course, even the best prompts produce output that needs human review. Next, we’ll look at how to efficiently review and refine AI translations — the post-editing workflow.