Build a community glossary to keep your translations honest
When your agency serves clients in many languages, the same term gets translated a dozen different ways, and small errors slip in that change what a client understands. A shared glossary that your bilingual staff and community help build is the anchor that keeps every translated notice, form, and message accurate and consistent.
A small translation error is an access barrier
Translation quality is not a cosmetic concern for the people you serve. A consent form that softens a risk, a notice that flips a deadline, or an interpreter card that names the wrong right can change what a client agrees to or whether they seek care. These errors are hardest to catch in exactly the languages that have the fewest reviewers, which are often the languages your clients speak.
Consistency is its own kind of accuracy. When one flyer calls a program by one name and the intake form calls it another, a client cannot tell they are the same thing. When a term like "eligibility rule" is translated four different ways across your materials, the materials read as if they came from four different agencies. A glossary fixes the term once so every document speaks with one voice.
Building the glossary with your community, and not just handing it to a vendor, is what makes it right. Your bilingual staff and the people you serve know which word carries the intended meaning in their variety of the language, which formal register fits a health or legal setting, and which literal translation would read as cold or confusing. That knowledge lives in the community, so that is where the glossary is built.
Build the glossary
Start small, with the terms that carry the most weight, and grow it as you go.
- List your load-bearing termsPull the words that recur across your materials and where an error would hurt: program and benefit names, "eligibility," "interpreter," "consent," "deadline," rights and legal terms, and any word that changes what a client does.
- Capture your agency’s own names and acronymsProgram, office, and agency names and abbreviations, such as SNAP, WIC, your agency’s own initials, and local program names, are exactly where translation breaks down, so pin down their agreed rendering in each language, deciding for each one whether to keep it, transliterate it, or translate it.
- Gather candidate translationsFor each term, collect how it is currently rendered across your existing documents in each language. You will usually find the same term translated more than one way. Bring the variants together so the disagreement is visible.
- Decide the canonical term with bilingual staffFor each language, have bilingual staff and, where you can, community members agree on one translation to use everywhere. Choose the one that is accurate, natural in your clients’ variety of the language, and right for a formal health or legal register.
- Note the register and any real variantsRecord the tone you want, such as a formal form of address, and note the terms where communities genuinely differ, so a reviewer knows a choice was deliberate and not a slip.
- Keep it livingAdd terms as new materials and programs appear, date the entries, and revisit the glossary when rules or programs change. A glossary that is never updated quietly goes out of date.
Use it to quality-check translations
The glossary is the reference a reviewer checks every translated document against.
- Check every recurring term against the glossaryRead the translated document with the glossary open. Where a key term does not match the agreed translation, fix it. This is the fastest way to catch the consistency problems that make materials look untrustworthy.
- Read for meaning, not just wordsCompare the translation against the original meaning, not word by word. Watch for false friends, a word that looks right but carries a different sense, and for a translation that reads smoothly while saying something the original did not.
- Guard the negations and the numbersThe errors that do the most harm are a dropped "not" in a safety rule, a flipped deadline, a wrong dollar amount, or a wrong phone number. Check these against the source every time, because they read as fluent when they are wrong.
- Confirm the register and toneMake sure the formal level matches the setting and stays consistent, so a health or legal notice does not slip into a casual voice partway through.
- Send unresolved calls back to the communityWhen staff disagree on a term, do not guess. Take it back to bilingual staff or community members, decide, and record the decision in the glossary so it is settled for next time.
The errors a glossary QA catches
These are the recurring failure modes worth naming for your reviewers.
A word that looks like a fair translation but carries a different feeling or sense. "Shame" and "embarrassment" are close in English but land very differently, and the wrong choice changes the tone of a whole message.
A "never enter client data" that becomes "rarely," or a lost "not," turns a safety rule into permission. In prohibitions and consent language, the negation is the most important word on the page.
"Caseworker," "social worker," and "navigator" are different roles, and a "qualified interpreter" is not a "translator." A swapped title can misdirect a client or misstate a right.
The same program or rule named differently across a flyer, a form, and a webpage reads as if the materials are unrelated. Consistency is what tells a client these documents belong together.
A notice that opens formally and slips into a casual voice, or the reverse, reads as careless in a setting where clients are deciding whether to trust you. The glossary fixes the intended register so it holds throughout.
The most dangerous error reads perfectly and says the wrong thing. Smooth wording is never evidence that a detail is right, so the number, the date, and the rule get checked against the source regardless.
Machine translation and LLMs frequently expand an abbreviation to the wrong thing, or invent an expansion outright, and clinical and agency acronyms are highly ambiguous with many possible expansions (Moon et al., 2014; Ji et al., 2023).
A proper name gets translated literally, word by word, instead of kept or transliterated, so your agency’s or a program’s name turns into a generic phrase a client will not recognize (Sharma et al., 2023; Yan et al., 2024).
AI can flag, people decide
An AI tool can speed up the mechanical parts of this work. It can scan a set of documents and surface where a term was translated inconsistently, draft a first list of candidate translations, and flag passages that drift from the source meaning for a person to look at. Used this way, it turns a slow manual comparison into a shorter review.
What it cannot do is decide. AI does not know which word your community actually uses, which register fits a legal setting in that language, or whether a subtle choice reads as warm or cold to the people you serve. It also makes confident mistakes, so anything it flags or drafts is a candidate a bilingual person confirms, never a final answer. The glossary stays a community document that AI only helps maintain.
This is how LINC keeps its own multilingual content honest. We build a canonical glossary of our key terms, use AI to check every translated page against it and to surface likely errors, and then have people confirm each call. The method on this page is the one we run on ourselves.
A quick QA checklist
- Every key term matches the agreed glossary translation.
- Negations and prohibitions are intact, with no dropped or softened "not."
- Numbers, dates, deadlines, amounts, addresses, and phone numbers match the source.
- Job titles and rights are named correctly, and interpreter is not confused with translator.
- The register is formal where the setting calls for it, and consistent throughout.
- Nothing that reads fluently was left unchecked against the original meaning.
- Unresolved term choices went back to the community and were recorded.
Sources
- Dinu, G., Mathur, P., Federico, M., & Al-Onaizan, Y. (2019). Training Neural Machine Translation to Apply Terminology Constraints. Proceedings of ACL 2019 (short paper). link
- Haque, R., Hasanuzzaman, M., & Way, A. (2020). Analysing terminology translation errors in statistical and neural machine translation. Machine Translation, 34(2-3), 149-195. doi:10.1007/s10590-020-09251-z
- Sharma, R., Katyayan, P., & Joshi, N. (2023). Improving the Quality of Neural Machine Translation Through Proper Translation of Name Entities. arXiv preprint. link
- Yan, J., Yan, P., Chen, Y., Li, J., Zhu, X., & Zhang, Y. (2024). Benchmarking GPT-4 against Human Translators: A Comprehensive Evaluation Across Languages, Domains, and Expertise Levels. arXiv preprint. link
- Moon, S., Pakhomov, S., Liu, N., Ryan, J. O., & Melton, G. B. (2014). A sense inventory for clinical abbreviations and acronyms created using clinical notes and medical dictionary resources. Journal of the American Medical Informatics Association, 21(2), 299-307. doi:10.1136/amiajnl-2012-001506
- Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12), 1-38. doi:10.1145/3571730
This guide describes a practice any agency can run, with or without AI. A glossary is most useful when it is short, shared, and actually used, so start with the terms that matter most and keep it where your whole team can reach it.