Translation Memory and Glossary: What They Do and How to Use Them
Translation memory and a glossary solve two different problems. Translation memory saves effort and cost by reusing what you already translated. A glossary protects meaning by fixing how important words are handled. Together they make machine translation more consistent and make human review faster, because a correction is made once instead of on every page.
What is a translation memory?
A translation memory is a database of pairs: a source sentence and its approved translation in one language. When the same sentence appears again, the system offers the stored translation instead of translating it from scratch. Typical benefits:
- Less cost. Repeated text such as buttons, footers, shipping notes and product descriptions with shared sentences is translated once.
- More consistency. The same sentence gets the same translation everywhere.
- Reuse of human work. A sentence a reviewer corrected stays corrected.
Memories can also offer fuzzy matches: text that differs only slightly, such as in capital letters or punctuation. A fuzzy match is a suggestion that needs a glance, not an automatic yes.
The exchange format for memories is TMX (Translation Memory eXchange). Whether a tool exports it, and how complete the export is, varies, so test it. See avoiding translation vendor lock-in.
What is a glossary?
A glossary is a list of terms with rules. Two kinds cover most needs:
- Do not translate. The term stays exactly as written: brand names, product names, model numbers, technical words. In HTML you can also mark text on the page with the
translate="no"attribute. MDN documents this attribute, and notes that it is respected by automatic translation systems such as Google Translate. - Preferred translation. You choose how a term is translated into a given language, for example your category name or a feature name that must read the same everywhere.
A glossary answers questions a machine cannot guess: is "plan" a subscription or a schedule on your site? Is "Home" a page name or the word for a house?
How they differ
| Item | Translation memory | Glossary |
|---|---|---|
| Unit | Whole sentences or segments | Single terms or short phrases |
| Filled by | Your approved translations | You, on purpose |
| Effect | Reuse and cost saving | Correctness and consistency of key words |
| Failure it prevents | Paying twice, saying the same thing two ways | Brand name translated, wrong sense of a word |
| Grows | Automatically, as you translate | Only when you add entries |
A simple routine to build both
- List your protected terms first. Brand, product, feature and plan names, units and codes. Mark them do-not-translate.
- Add terms with a required translation. Start with ten to thirty that matter most, in each language you offer.
- Run the machine translation so the glossary applies from the beginning.
- Review the pages that matter (headlines, buttons, prices, legal text). Fix what is wrong by editing the translation.
- Turn recurring fixes into glossary entries. If you correct the same word three times, it belongs in the glossary.
- Re-run only what needs it. Re-translate the strings affected by the new entries rather than the whole site.
- Export both regularly so they are yours, in standard formats.
How LgTranslate handles them
LgTranslate keeps a translation memory that can reuse an earlier translation of identical text. Sharing memory between accounts is off by default. Its glossary has two entry kinds, do-not-translate and preferred translation, per language. In the design:
- Do-not-translate terms are protected during machine translation so the engine does not change them; preferred translations are passed to engines that support a glossary.
- Quality checks flag a translation where a do-not-translate term is missing or a preferred translation was not used. These are warnings, not blockers.
- Matching ignores capital letters by default. For most languages terms match whole words; for languages written without spaces (such as Chinese, Japanese, Korean and Thai) matching happens inside text, so a short term can match inside a longer word, and you should check short terms.
- A new glossary entry does not rewrite existing translations. Translations a person edited or reviewed are never replaced by a machine re-run.
- You can export your translations in several formats, including TMX, XLIFF, PO and CSV, and import them back.
Some details of the editor form are still limited (for example, it saves every term as not case-sensitive, and the API offers more control). We document those limits in the help centre instead of hiding them.
Our guide on reviewing machine translation shows where recurring fixes come from.
Common mistakes
- A huge glossary on day one. Hundreds of entries nobody reviewed cause more harm than good. Start small.
- Short ambiguous terms. A two-letter term will match in places you did not intend. Prefer full words or phrases.
- Forgetting regional variants. A preference for one regional form of a language might not apply to another; check how your tool treats codes such as
pt-PTandpt-BR. - Never exporting. A memory you cannot take with you is a hostage.
Next step
Write down your ten most important terms today. In LgTranslate, add them to the glossary before your first translation run, then compare the result with a run without them. You will see quickly which entries earn their place.
Key takeaways
- Translation memory reuses whole earlier translations; a glossary controls individual terms.
- Build the glossary small and let recurring fixes decide what to add.
- New glossary entries influence future runs and checks; re-run or edit existing translations.
- Keep both in formats you can export.
Sources
- translate global attribute – MDN Web Docs – marking content that should not be translated (existing guide source, see docs/ARTICLE_VERIFY.md)
- Managing multi-regional and multilingual sites – Google Search Central – consistency and avoiding mixed-language pages (checked 2026-11-06)
- Choosing a language tag – W3C Internationalization – language and region tags (checked 2026-11-06)
Frequently asked questions
What is the difference between translation memory and a glossary?
Translation memory stores whole earlier translations (usually sentences) so identical or similar text is not translated twice. A glossary stores individual terms, such as a product name that must stay as written or a word with a required translation.
Will adding a glossary entry fix my existing translations?
Not by itself. A new entry affects future translation and quality checks. Existing translations need to be re-run or edited. Translations a person wrote or reviewed should not be overwritten by a machine re-run.
Can I take my translation memory with me if I leave a vendor?
That depends on the vendor, so test it before you commit. TMX is the common exchange format for translation memory. Ask for an export and check that your text and terms come out intact.