Skip to content
Faisal Mahmud TusherDigital marketing, SEO & analytics
with Faisal Tusher

International Marketing › Localisation

Building a Termbase and Translation Memory

How to build a termbase and keep a translation memory: extracting terms, approving them per language, do-not-translate names, ownership and corrections.
Tutorialintermediate9 min readUpdated

Why terms drift and what the two resources prevent

Terms drift because nobody decided them. Each translator meets a product or feature name, picks a reasonable equivalent and moves on. The next translator, or the same one a month later, picks a different one. Support articles, product pages and ads end up using several words for one thing, and customers and support teams cannot tell whether those words mean the same thing. The translation and transcreation guide describes the termbase in a paragraph: the approved translation per language, the terms not to translate, a definition and context, and the rejected alternatives, with reviewers correcting the termbase rather than only the asset. This tutorial builds it.

The translation guide mentions translation memory only as something professional translators often use. Here it is treated as a practice the business runs: a record, kept by the business, of every source sentence or segment and the translation a reviewer approved for it. Repeated wording such as feature descriptions, delivery lines, button labels and boilerplate then comes back the same way every time, whoever the translator is.

The two resources sit beside the localisation style guide, which records everything that is not a single term: address, register, voice and formats. The style guide links to the termbase rather than listing terms. It is also worth being clear about what the termbase is not. A reader-facing glossary explains words to customers on the site; the termbase is an internal resource for the people producing copy, and many of its entries would never appear in a public glossary.

Before you start

  • Have the source content to hand. The product catalogue, product pages, help articles, emails, app text and recent campaigns in the source language, because the candidate terms come from there.
  • Name the in-country reviewer for each language. A native speaker who knows the category approves every term in their language, as the translation guide advises for review generally.
  • Have the style guide started. The localisation style guide tutorial covers the language decisions the termbase does not.
  • Collect the keyword research. The approved keyword list from keyword research for a new market supplies the searched-for words.
  • Know the house naming rules. The source-language style guide and any Naming Convention for products and features tell you which names are fixed.
  • Agree one place to keep the files. A location the business controls, which every supplier and reviewer can read and only named people can edit.

The steps

  1. Pull candidate terms from the source content

    Go through the source content and list every word or phrase that names something the business sells or does: product and product line names, feature names, components, materials, service names, category words, units that recur, and the fixed phrases used in buttons, navigation and checkout. Include terms that look ordinary but carry a specific meaning in your catalogue. Note where each term appears, because context decides the translation. Start with the content in the first localisation wave rather than the whole estate; the list grows with each new job.

  2. Write a definition and context for each

    For every candidate, write a one-line definition in the source language and an example sentence from real copy. Translators choose correctly when they know what the thing is and where it appears, and a definition often reveals that two source terms mean the same thing, or that one term is used for two things. Settle those in the source first; a termbase cannot be more consistent than the source it is built from. The Editorial Style Guide is the place to record a source-language decision of that kind.

  3. Mark the names that are never translated

    Mark the brand name, product line names and any feature names the brand has decided to keep in the source language as do-not-translate. The translation guide lists these as part of the termbase, and the tutorial on naming a product or brand checks a new name with native speakers before it is chosen, precisely because it will not be translated. Record how each name behaves in a sentence, such as whether it takes local grammar, and keep the list identical to the style guide's list of source-language names.

  4. Agree each approved term with the in-country reviewer

    Send the list to the in-country reviewer for each language with the definitions and contexts. The reviewer proposes the approved term, checks it against how the category is actually described in that market and flags terms where no good equivalent exists. Where the source term is a piece of wordplay or a campaign coinage, mark it for transcreation rather than forcing a single equivalent. The reviewer's approval, with their name and the date, goes into the entry.

  5. Add the searched-for keywords from local research

    For terms that people search for, the approved term should normally be the word the market uses, not a translation of the home keyword. The keyword research tutorial finds those words with a native-speaking researcher; copy the approved keyword into the entry, with a note of its Search Volume range and intent where the research recorded them. Where the reviewer's preferred term and the searched-for term differ, the two agree which one leads in headings and which one appears in body copy, and the entry says so.

  6. Record rejected alternatives

    Every term that was considered and turned down goes into the entry with the reason: too formal, used by a competitor, means something else in this category, not what people search for. This is what stops a term drifting back when a new translator arrives with the same reasonable idea. The translation guide names rejected alternatives as one of the termbase's fields for that reason.

  7. Decide who owns and edits the files

    Write down who owns the termbase and the translation memory, who approves changes per language and who may only read them. A RACI matrix makes it explicit: the localisation owner is accountable for the files, the in-country reviewers are responsible for the terms in their language, suppliers are consulted and informed. The files belong to the business and live where the business keeps them, the same principle the operations hub applies to ad accounts and data, so that a change of supplier changes nothing about the approved wording.

  8. Start the translation memory from approved copy

    Seed the memory with source and target pairs that a reviewer has already approved: published pages, emails and help articles that passed in-country review. Leave out anything that never had language sign-off, because a memory full of unreviewed sentences repeats their errors faithfully. Each pair records the language, the date, the content type and who approved it. From then on, every approved job adds its pairs to the memory.

  9. Send every job with both attached

    Every localisation request carries the current termbase and memory, alongside the style guide and the brief. The localisation brief template has a termbase row for that purpose. Where machine translation is used, the translation guide states its role for product data, help articles and long informational content as a first draft constrained by the termbase, followed by a full post-edit by a qualified linguist; the AI for content guide adds that a brief should say which passages, such as product names, must stay exact. Without the termbase, a machine draft picks its own terms.

  10. Correct the source, not just the asset

    When a reviewer changes a term in an asset, the change goes into the termbase, with the old term moved to the rejected alternatives. When a reviewer rewrites a remembered sentence, the memory is updated so the corrected version is the one reused. The translation guide puts it plainly: fix the source of the error, or it recurs on every future job. Review both files at each market gate, and whenever the source catalogue or naming changes.

Termbase entry fields: a table

FieldWhat it holdsWho fills it
Source termThe word or phrase as used in the source languageLocalisation owner
DefinitionOne line saying what the thing isLocalisation owner with the product team
ContextAn example sentence from real copy and where it appearsLocalisation owner
StatusTranslate, do not translate, or transcreateBrand owner
Approved term per languageThe term to use in that languageIn-country reviewer
Searched-for termThe keyword the market uses, from local researchNative-speaking researcher
Rejected alternativesTerms considered and turned down, with the reasonIn-country reviewer
Grammar notesHow the term behaves in a sentence, plurals, genderIn-country reviewer
ApprovalReviewer name and dateIn-country reviewer
Change historyWhat changed, when and whyLocalisation owner

A remembered sentence is not an approved claim. A memory pair that contains a performance, safety, environmental or price claim still goes through the claims library and the market's compliance review each time it is used, because a claim's approval rests on its evidence, conditions and markets, not on the fact that it was translated before. Mark claim-bearing pairs in the memory so translators and reviewers know to check them rather than reuse them.

A laboratory glassware supplier builds its first termbase (illustrative)

Illustrative example — not a real result

Kolvrent Labware builds a termbase for Market A and Market B

Kolvrent Labware, an invented supplier of laboratory glassware, prepares editions for Market A and Market B. Its marketing lead pulls 214 candidate terms from the catalogue and the 60 product pages in the first wave, writes a definition and a context sentence for each, and finds that the source itself used two names for the same flask neck finish. That is settled first.

The 5 product line names are marked do-not-translate. Two in-country reviewers, a laboratory technician in Market A and a chemistry teacher in Market B, approve terms for their language over three sessions each. The Market A reviewer finds that a feature described as heat-resistant had been given three different names across earlier translated datasheets; one is approved, two go to the rejected list with the reason. Keyword research in Market B shows that buyers search for one product category using a borrowed English term, so that term leads in headings there, with the reviewer's term in body copy.

The translation memory starts with 1,150 approved sentence pairs from datasheets that had already passed review; 300 unreviewed pairs from an old supplier are left out. Both files sit in Kolvrent's own shared drive, with the localisation owner as the only editor. Six months later, when Kolvrent changes its Market B supplier, the new translator starts from the same files on the first day.

Common mistakes

  • Letting the supplier hold the files. The termbase and memory are the business's record of approved wording. Keep the master copies where the business controls them.
  • Building from an inconsistent source. If the source uses two words for one thing, the termbase inherits the confusion. Settle the source terms first.
  • Terms without definitions or context. A bare list of words forces translators to guess. Every entry needs a definition and an example sentence.
  • Translating keywords. The searched-for term comes from local research, not from translating the home keyword.
  • Seeding the memory with unreviewed copy. A memory repeats whatever it holds. Only reviewed pairs go in.
  • Reusing remembered claims. A remembered sentence with a claim is checked again through the claims library and local compliance.
  • Corrections that stay in the asset. Every reviewer change updates the termbase or the memory, or the same correction is needed on the next job.

The marketing localisation hub shows where the termbase and memory sit in the language workflow, and the naming convention entry in the glossary covers the source-side rules that keep names fixed before any of this begins.

Frequently asked questions

What is the difference between a termbase and a translation memory?

A termbase works at the level of single terms: each product, feature or category word with its approved translation per language, a definition, context, the rejected alternatives and whether it is translated at all. A translation memory works at the level of whole sentences or segments: it pairs each approved source sentence with its approved translation so the same sentence is not translated differently next time. The termbase keeps the vocabulary consistent; the memory keeps repeated wording consistent.

Who should own the termbase and translation memory?

The business, not the supplier. A named localisation owner holds the master files, decides who may edit them and makes sure every supplier works from the current version. In-country reviewers approve the terms in their language. If the files live only inside a supplier's systems, changing supplier means starting again, and the approved wording the business paid for walks out with them.

Does a translation memory mean approved sentences can be reused without review?

No. A remembered translation was right for the context it was approved in, not necessarily for the next one, and the reviewer still reads the new asset in context. Claims are the clearest case: a sentence carrying a claim goes through the claims library and the market's compliance review each time it is used, because a claim's approval depends on its evidence, conditions and markets, not on the fact that it was translated before.

Terms used on this page

Frameworks

Where this page is used

Also related

Related