Gains:
- Ability to ensure terminology consistency by establishing a termbase and clearly defining approved and prohibited equivalents
- Using artificial intelligence to quickly extract term candidates and make the final term decision with the expert and the customer
- Ability to connect termbase to the tool/prompt during translation and detect incompatibilities with a post-translation term QA tour
In a technical translation, if you translate the word "user account" as "user account" in one place, "member account" in another place, "user profile" in another place, the user will be lost, the brand will appear inconsistent, and the search/documentation will be broken, even if the text is technically "correct". The silent but most decisive quality element of translation is terminological consistency. In this unit, you will learn terminology management, creating a termbase (term database), term extraction with AI, and how to maintain terminology discipline throughout a translation. The aim is to be an expert who meets the same concept in every project in the same way everywhere and is loyal to the corporate language.
Basic concepts
Terminology is words and expressions that correspond to specific concepts of a field or institution ("checkout", "return policy", "stem cell"). Termbase (TB — term database) is a structured dictionary that keeps the source-target equivalents, definition, usage note and "prohibited equivalents" of these terms; It connects to CAT tools and automatically suggests translation. Term extraction is the process of automatically collecting terms that need to be translated/standardized from a text. Glossary is a simpler, two-column list of terms; It is used instead of termbase in small jobs.
Why is it so important? Because machine translation and LLM by themselves do not guarantee terminology consistency. NMT can translate each sentence independently; LLM may change its decision in long text. What provides term discipline is the termbase that you have prepared and connected to the vehicle.
Tip: Spend 30 minutes setting up termbase on your first job with a client. This investment increases both speed and quality in all subsequent work and gives the reputation of "brand consistent in your translations".
Term extraction and termbase construction with AI
AI speeds up the most tedious part of setting up a termbase: finding candidates. You can give a source text to LLM and say, "list the domain-specific terms in this text, those that are repetitive, and those that require consistency in translation." You then approve this list of candidates, determine the correct target response and confirm with the customer when necessary. Critical point: AI suggests term candidates, the expert and the customer make the term decision. Whether a brand calls "appointment" or "booking" is the customer's choice, not the AI's.
For each term, put the following into Termbase: source term, approved target counterpart, short description/context, "do not use" (prohibited) equivalents, and domain/project tag if applicable. Writing forbidden equivalents is the most effective way to avoid inconsistency (for "account", DO NOT use just "account"; "profile", "membership").
Maintaining terminology discipline throughout translation
Installing Termbase is not enough; It needs to be implemented. Three layers:
- During translation: When translating to LLM, add termbase to the prompt ("MUST use these terms with these equivalents"). If you are using a CAT tool, connect termbase; The tool shows the match instantly.
- Post-translation QA: Scan the finished text against termbase. I asked the LLM "are there places in this text that do not comply with termbase?" or run the term QA function of the CAT tool.
- End-of-project update: Add new terms and customer feedback to termbase when work is finished; termbase is a living document.
Caution: When you give LLM a long termbase it may "forget" some terms or change them depending on the context. So never skip the post-translation QA round; Prompting is not a substitute for checking.
three mini cases
Case 1 — Termbase deleted the discrepancy. In a 60,000-word documentation made by a software company with 5 different freelance translators, the same 40 terms were translated in dozens of different ways. Once a common termbase was established and all translations were aligned to it, complaints such as "I can't find this screen" to user support decreased significantly.
Case 2 — AI reduced inference from 2 days to 2 hours. Instead of manually extracting a list of terms from a 200-page device document, a medical translation team had the LLM draft it: 380 candidate terms were listed in 2 hours. The expert approved these in 1 day and clarified their responses; It would have taken weeks if done manually. Decisions belonged to people again.
Case 3 — Prohibition rule averted crisis. In a brand's termbase, "customer" was approved for "customer" and "user" was prohibited (due to brand positioning). When a new translator typed "user", term QA caught it and the brand language distortion was fixed before it went live.
Four copyable templates
1) Term candidate extraction:
Your role: terminologist. From the following [domain] text extract terms that require consistency in translation: domain-specific concepts, repetitive phrases, product/feature names, abbreviations. List each term in the source language, add a brief context note. The decision is mine; You just suggest a candidate. Adding common words (and, but).Text: [...]
2) Creating Termbase line:
Prepare termbase row for the following terms: source term | suggested target response | short description | wrong answers to avoid. If you are not sure, mark the answer with [?]; I will make the decision.Alan: [...] | Terms: [...]
3) Termbase compatible translation:
Translate this text [source]→[target]. MUST apply the termbase below; DO NOT USE prohibited equivalents:- [source1] → [target1] (use: [forbidden1])- [source2] → [target2]If you encounter a term that is not in Termbase, mark it with [?].Text: [...]
4) Post-translation term QA:
Scan the translation below against the termbase I provided. Just list the INCOMPATIBILITIES: which term was used, where, with what wrong meaning, what should be the correct one. Reporting compatible locations.Termbase: [...] | Translation: [...]
Weak prompt / Strong prompt
Weak: "Be careful with the terms." (Which term, which equivalent, which prohibition — is unclear; the machine chooses itself.)
Strong: "When translating this text, apply the following termbase: 'dashboard'→'dashboard', 'log in'→'log in'. If you come across a non-termbase term, put [?] and make it up yourself."
Difference: strong prompt specifies the approved counterpart, the forbidden counterpart, and the unknown term behavior; consistency is guaranteed.
Table of terminology tools/approaches
Approach
What does it do?
limit
Glossary (two columns)
Small business, fast start
No definition/prohibition
Termbase (connected to CAT)
Automatic recommendation, QA
Installation requires labor
Term extraction with LLM
Speeds up the candidate list
People decide
Term QA scan
Catches dissonance
Termbase must be up to date
List of prohibited provisions
Maintains brand language
Requires constant maintenance
Common mistakes
- Starting a big job without installing Termbase. The inconsistency subsequently grows exponentially.
- Accepting the response suggested by AI without question. Terminology decision is up to the client/expert.
- Not writing prohibited responses. Just writing the "right" does not prevent the "wrong".
- The post-translation term is skipping QA. Putting termbase in the prompt does not guarantee its implementation.
- Not updating Termbase. A dead termbase will mislead over time.
Multilingual termbase and customer approval
In large projects, the termbase is often multilingual: a concept has a counterpart in more than one target language (source → English, German, Turkish, Arabic). This ensures consistent translation of the same concept across all languages and speeds up work when a new language is added. The point to be noted in the multilingual termbase is that each language has its own cultural and grammatical equivalent; It is wrong to blindly copy the equivalent in one language into another.
Terminology also has a chain of approval dimension. In critical terms, the final decision should often be made by the client's subject matter expert or brand team, not the translator; Because they know best the official/legal/brand equivalent of a term. That's why on big jobs it's good practice to tag the termbase as "approved", "candidate" and "disputed": you ask the customer for candidate and controversial terms altogether, get them approved, then lock them. This discipline prevents mid-task “this term should actually be” surprise and mass retranslation.
In summary
Terminological consistency is the invisible but most decisive quality element of translation. The master translator establishes a termbase on the job; It uses AI to quickly extract term candidates, but leaves the decision with the customer and the expert; clearly writes approved and prohibited responses; It connects termbase to the prompt/tool during translation and never skips the post-translation term QA round. Termbase is a living document; grows in every business and protects the brand.
Application task
Select a field text (technical, medical or marketing). Extract at least 15 term candidates with the "term candidate extraction" template, confirm them and write the equivalents, definitions and prohibited equivalents with the "termbase line" template. Then translate a short text with this termbase, scan for incompatibilities with the "term QA" template and fix what you find.
checklist
- [ ] I have set up a termbase/glossary at work.
- [ ] I approved AI's term candidates and made the decision.
- [ ] In addition to the approved responses, I also wrote the prohibited responses.
- [ ] I connected termbase to the prompt/tool during translation.
- [ ] I did a post-translation term QA scan and fixed the incompatibilities.