A termbase is the difference between a translation that reads well once and a product that says the same thing in every market for years. Terminology management is the discipline of deciding which words a brand uses, recording them with enough context that another person will use them the same way, and keeping that record alive as the product and the market change. This guide covers what a termbase is, how it differs from a glossary, how to build one step by step, which tools will not cost you anything, how to maintain it, and the mistakes that quietly ruin most of the termbases built in a hurry.


1. What a Termbase Is

A termbase is a structured, searchable database of terms and their approved equivalents across languages, where each term carries the information a translator needs to use it correctly. The word comes from "terminology database", and the key word is structured. A termbase is not a list. It is a set of entries, and each entry is a record with defined fields: the term itself, its language, its part of speech, its definition, its approved translation, and the metadata that tells you when the term is preferred, when it is allowed, and when it must never be used.

That structure is what makes a termbase useful to a machine as well as a person. A CAT tool or a translation management system can read a termbase, highlight the terms inside a segment, and warn a translator when the approved equivalent is not present. A plain list in a spreadsheet cannot do that reliably, because the tool has no way to know which column is the term, which is the translation, and which rows are notes rather than entries.

The standard format for exchanging termbases between tools is TBX, the TermBase eXchange format standardised as ISO 30042. You do not need to memorise that to start, but you will meet it the moment you try to move a termbase from one system into another, which is why it is worth knowing the name.


2. Glossary Versus Termbase

The two words are used loosely in the industry, and the confusion causes real problems. A glossary is a simple two-column or three-column list: a source term and its target equivalent, sometimes with a note. A termbase is a database of entries with many fields per term. Both have their place, and the right choice depends on how much the term needs to be explained and how many people will use it.

DimensionGlossaryTermbase
StructureFlat list of term pairsEntries with defined fields per term
Typical fieldsSource, target, sometimes a noteTerm, language, part of speech, definition, context, status, domain, source, author, date
Handles synonyms and variantsPoorly, usually one row per variantYes, with admitted and deprecated terms linked to a preferred one
Works inside a CAT toolSometimes, as a simple matchYes, with term recognition and enforcement
Best forA quick one-off job or a small projectAny product or brand used across many translators and years

A practical rule: if the term needs a definition, a context sentence or a status, it belongs in a termbase. If it is a brand name with one obvious equivalent and no nuance, a glossary row will do. Most real projects start with a glossary and grow into a termbase once the first translator asks "why did we choose this word?".


3. Why Terminology Management Pays

Terminology is not housekeeping. It is one of the few activities in localization that improves quality, speed and cost at the same time.

  • Consistency. The same concept carries the same word in every file, every release and every market. Consistency is the quality dimension users notice first and complain about loudest.
  • Speed and reuse. When terminology matches, translation memories match too, which means more reuse and less rewriting. Terminology errors also block clean reuse, because the tool cannot confirm a segment whose terms changed.
  • Compliance and risk. In legal, medical and regulated domains the wrong term is not a style issue, it is a liability. A termbase that records a forbidden term and its approved replacement is a control, not a convenience.
  • Onboarding. A new translator, or an AI system, can produce usable output on day one if the terminology is written down. Without it, every new contributor repeats the same decisions and the same mistakes.
  • Brand. Product names, feature names and marketing claims are brand assets. A termbase is where "do not translate", "translate as", and "never use this variant" are recorded.

There is a difference between doing terminology work and running a terminology programme. The programme has a scope, a stated goal, a workflow and an owner, and it is justified in the language of the business: fewer review cycles, fewer corrections after publication, less rework on every release. That is the frame we use in Strategic Terminology Management for Global Success, where the starting point is not the tool but the questions a termbase has to answer for the organisation.

What the course works through. Defining scope and goals, aligning terminology with business strategy, designing the workflow, securing stakeholder buy-in and measuring return. The closing session looks forward: how terminology feeds machine translation glossaries, LLM prompts, metadata structures and knowledge graphs. The line the session opens with is a fair summary of the whole discipline: if content is king, terminology is queen.

4. Terminology in an AI Localization Workflow

An AI-assisted localization pipeline has four stages, and terminology runs through all of them. Understanding where the termbase sits explains why it stops being a nice-to-have the moment a machine enters the workflow.

  1. Terminology and assets. The glossary and termbase are agreed first, before any text is translated. This is the stage that defines the words the rest of the pipeline must use.
  2. Pre-translation. A machine translation engine or a large language model produces a draft. When the termbase is available to the engine, whether through constrained decoding in MT or through a glossary injected into the prompt, the draft arrives using the approved terms. When it is not, the model picks its own vocabulary and the terminology drifts from the first sentence.
  3. Post-editing. The human editor corrects meaning, fluency and style. Post-editing is fastest when terminology is already right, because term fixes are the ones that ripple through a file and force the editor to re-read segments already passed.
  4. Quality assurance. Automated checks and human LQA verify that the approved terms are used and the forbidden terms are absent. A terminology check is one of the cheapest automated checks there is, and it only works if a termbase exists to check against.

Two properties of AI make this structure matter more than it did before. First, models are confident: they will produce a plausible but wrong term without any signal that they guessed, so errors do not announce themselves. Second, models are controllable: the biggest quality gains come from changing the instructions and the assets you give them, not from swapping the model. That is the control lever. A well-built glossary and a clear style guide are the two levers a linguist pulls to move AI output from acceptable to reliable.

For the step-change version of this workflow, our Creating AI Localization Pipelines and AI-Driven Localization Pipelines courses build the pipeline end to end, and the glossary-rewriting method at the heart of it is the one taught in From Post-Editor to AI Quality Specialist.


This is where the course material is blunt about the limits of the machine. Large language models do not understand meaning, they predict language, which is why they produce fluent output that is sometimes wrong. Independent benchmarks cited in Terminology Management and Generative AI put hallucination rates above 15 to 20 per cent depending on task and domain. Without authoritative linguistic data underneath, no model holds consistency, factual accuracy or brand voice at scale.

That is the practical argument for treating terminology as AI infrastructure rather than as a glossary attached to a project. The same session traces the path from flat glossaries to structured termbases to knowledge graphs, and explains where retrieval and grounding enter the pipeline, which is the difference between a model that guesses your product names and one that is told them.

5. How to Prepare Terminology for AI

A termbase written for a human translator and a termbase written for a machine are not the same artefact. The human tolerates a loose note, a synonym, an ambiguous status. The machine does not. Preparing terminology for AI means making each entry explicit enough to be followed without interpretation.

5.1. Make it machine-readable, not just human-readable

The distinction is practical. A human-readable resource assumes a person will resolve ambiguity with judgement. A machine-readable resource removes the ambiguity: it says which term is approved, which is forbidden, and what the term means in this product. When a linguist rewrites a glossary entry this way, the change in AI output can be immediate, because the model now has an instruction instead of a hint.

5.2. Use a structured format and keep it clean

  • TBX (ISO 30042) is the structured exchange format. Use it when the termbase has to move between a CAT tool, a TMS and a prompt-building script, so the fields survive the trip.
  • CSV or XLSX is portable and readable by every script and platform, but flat. It is fine for a glossary and risky for a termbase with many fields, because nothing stops a column being rearranged.
  • One term, one entry. Keep each entry atomic so a tool can import it without guessing. Do not merge several concepts into one row.

5.3. Fill the fields that a model and a reviewer both need

  • Part of speech, so the model inflects the term correctly in the target language rather than using it in the wrong grammatical role.
  • Definition, so the term is disambiguated when the same word means different things in different products.
  • Status, preferred, admitted, deprecated or forbidden, so the instruction is unambiguous.
  • Do not translate and forbidden term flags, because the fastest quality win is often telling the model what it must leave alone.
  • Context, a real sentence that shows how the term behaves.

5.4. Keep it consistent and constrain it

Consistency beats richness. One approved term per concept, no competing synonyms in the preferred list, and the same status vocabulary everywhere. Then constrain the model with the termbase: inject the relevant approved terms into the prompt, tell it the forbidden ones, and repeat the instruction for the terms that matter most. The same termbase then serves the pre-translation prompt, the CAT tool term recognition and the final automated check, which is what closes the loop.

Working through this properly is the core of Terminology Management and AI and Strategic Terminology Management, and it connects back to the research method covered in Translating Terminology and Resources.


5.5. Ground the model, do not hope it remembers

Preparation for AI is mostly about turning terminology into something a system can be given, not something it is expected to know. In practice that means clean structured entries, a stable identifier per concept, and a format the tool can ingest without interpretation. The course session on terminology as AI infrastructure covers the two mechanisms that matter here: retrieval-augmented generation, where the right entries are fetched into the prompt, and term-augmented generation, where the approved terms constrain the output directly.

There is also a human step that no pipeline removes. Automated term extraction, definition drafting and clustering of candidate terms speed the work up considerably, but every entry that reaches a termbase still needs a reviewer who can say whether the concept is the same concept. The course calls this the human in the loop, and it is the control that keeps a fast workflow honest.

6. The Anatomy of a Good Entry

An entry is only as useful as the fields you fill in. The terms below are the ones that pay for themselves; you do not need all of them, but the more of these you populate, the fewer questions come back.

FieldWhat it recordsWhy it matters
TermThe word or phrase, in its canonical formThis is the unit being managed
Language and localeWhich language, and which market variantEnglish is not one target; en-GB and en-US differ
Term type / statusPreferred, admitted, deprecated or forbiddenTells the reader what to do, not just what exists
Part of speechNoun, verb, adjective, and so onPrevents using a noun where the grammar needs a verb
DefinitionWhat the concept meansDisambiguates a term that has several meanings
ContextA real sentence where the term appearsShows how the term behaves in the wild
Domain / subjectThe field the term belongs to"Platform" means something different in software and in transport
Source and clientWhere the decision came from and whom it servesLets a later reader trace authority
Cross-referencesSynonyms, abbreviations, related termsKeeps the entry connected to the rest of the set
MetadataAuthor, date, project, versionMakes change tracking and review possible

7. How to Build a Termbase, Step by Step

Building a termbase is a process, not an afternoon. The order below is the one that avoids rework.

7.1. Step 1. Define the scope and the owner

Scope is a business decision before it is a linguistic one. The course framework asks three questions at this point: which content actually carries risk if the term is wrong, who owns the decision for each domain, and what result the programme is expected to produce in the first year. A termbase scoped to a product line, with a named owner and a measurable goal, survives. A termbase scoped to everything, with no owner, does not.

Decide what the termbase covers: one product, one client, one domain, or the whole brand. Then name a single owner. A termbase with two owners has none, and entries will drift. Write down who approves a term and who can only propose one.

7.2. Step 2. Gather what already exists

Most terminology already exists somewhere. Collect the client glossary, the style guide, previous translations, approved term lists, regulatory texts, and any terminology the client considers fixed. If you work in a regulated domain, international reference bases such as IATE for the European Union and UNTERM for the United Nations, and vendor terminology portals from major software companies, are useful starting points for established terms. Note their source, because a term borrowed from a reference base still needs client approval.

7.3. Step 3. Extract candidate terms

You are looking for the terms that matter: repeated concepts, product names, technical nouns, and anything the client has opinions about. You can find candidates by reading, by using a concordancer to surface repeated words and phrases, or by letting a language model propose candidates. Extraction only proposes; a human decides. A model that invents a term is a liability, and the entry it produces is worse than no entry.

7.4. Step 4. Draft the entries

For each candidate, write the source term and the proposed target, with the fields from the anatomy above that you need. Keep definitions short and factual. If you cannot write a definition, you may not understand the concept well enough to fix the term yet, which is itself useful information.

7.5. Step 5. Validate with the subject matter expert or the client

Terminology is a decision, and the decision belongs to the person who owns the content. A translator can propose "preferred" with evidence, but a product owner, a legal reviewer or a medical affairs team confirms it. Record who validated and when. This step is where a termbase becomes authoritative instead of aspirational.

7.6. Step 6. Set status and metadata

Mark each term as preferred, admitted, deprecated or forbidden. Add the part of speech, the domain, the source and the validation date. Status is the field that turns a list into instructions: it tells a translator which variant to use and which to avoid.

7.7. Step 7. Import it into the tools people actually use

A termbase that lives only in a spreadsheet will not be used. Import it into the CAT tool or TMS so it appears in the editor and can enforce or flag terms. This is where the TBX format earns its keep, because it moves between systems with the structure intact.

7.8. Step 8. Freeze a version and test it

Before the termbase goes live, freeze a version, note the date, and run a test project through it to check the terms surface where they should. Fix what breaks, then release. Versioning is what allows you to say later exactly which terminology a given delivery used.

7.9. Step 9. Train and publish the workflow

Tell the team how to query a term and how to propose a change. An unused termbase and a missing one cost the same. A short guide, a named contact and a simple proposal form are enough to keep it alive.

Index before you build. Start with the terms the client has already fixed and the ones that keep coming back in review comments. A small, accurate termbase of twenty authoritative entries beats a large one of two hundred unverified rows, because translators learn to trust the first and ignore the second.


8. The Advantages of Terminology Extraction

Terminology extraction is the step that turns a pile of source text into a candidate list. Done well, it is what makes the termbase complete; done badly, it is what makes it noisy. There are two ways to do it, and the best results come from combining them.

8.1. Automatic extraction

Software proposes candidates. A concordancer surfaces the words and multiword phrases that repeat across a corpus, a terminology tool applies statistical and linguistic filters, and a language model can suggest candidate pairs and draft definitions. The advantages are speed and coverage: a machine reads the whole corpus, not a sample, and it finds the repeated multiword terms that a translator scanning by eye will miss. The limits are equally clear. Automatic extraction proposes; it does not decide. It cannot tell an approved term from a passing one, it will happily surface noise, and an unverified term it invents is worse than no entry, because a machine will apply it confidently.

8.2. Manual extraction

A specialist reads the source with intent and records the terms that need a decision, especially the ones that carry legal, medical or brand weight. Manual extraction is slower, but it brings the judgement that no filter supplies: which term is a concept of the product and which is just a word, and which variant the client will accept. It is the stage where domain knowledge pays for itself.

8.3. The hybrid that works

Use the machine for the first pass and the human for the decision. Let a tool and a model produce a wide candidate list, then have a linguist or terminologist validate, merge duplicates, set status and write the short definitions. This keeps the coverage of extraction and the accuracy of review, and it is the only way to build a termbase at scale without drowning the reviewer.

8.4. Why the effort pays

  • Consistency at the source. You fix a term once and it propagates through translation memories, prompts and every downstream file.
  • Faster post-editing and review. Fewer terminology defects means fewer segments reopened, and fewer review comments that stop the same error repeating.
  • Better AI output. A complete, machine-readable glossary is what turns a generic model into one that follows your vocabulary.
  • Lower cost. Every term decided in advance is a query that never reaches the translator, the reviewer or the vendor manager.
  • Compliance and onboarding. In regulated domains the extraction record is evidence, and for a new contributor, human or machine, it is the shortest path to usable output.

The extraction workflow, from corpus to validated entry, is covered in depth in the terminology track collected at TranslaStars courses, alongside the AI workflow courses above.


The course material separates three tasks that are often mixed together. Extracting candidate terms is pattern work and machines are good at it. Drafting a definition is editorial work and machines are useful but unreliable. Deciding whether two candidates are the same concept is judgement, and it stays with the terminologist. Running the three in that order, with a reviewer at the end, is what produces a candidate list worth importing rather than a list that has to be cleaned by hand.

9. Free and Low-Cost Tools

You do not need to buy a terminology suite to run proper terminology management. The free layer covers most of the work.

ToolWhat it does for terminologyCost layer
OmegaTOpen-source CAT tool with glossary and termbase support inside the editorFree
WeblateWeb-based translation platform with a shared glossary per projectFree to self-host
CrowdinCloud localization platform with glossaries and term enforcementFree tier for smaller projects
PoeditCAT tool for software strings with glossary supportFree core, paid Pro features
XbenchQuality and terminology checking, including term consistencyFree version
AntConcFree concordancer to surface repeated terms in your own corporaFree
IATE, UNTERM, vendor portalsReference terminology you can consult and adaptFree to consult

The pattern is consistent: the free tools handle extraction, storage and checking, and the paid layer adds scale, connectors and enforcement across large teams. If your reason for buying a platform is terminology alone, you can probably start free. If your reason is automation across hundreds of files, that is a different purchase with a different justification.


Tools only help once you know what you are looking for. The course Translating Terminology: Best Resources is built around that gap: where to look for the accurate translation of a complex term, how to judge between several plausible equivalents, and which resources to trust for documents from international organisations such as the EU and the UN. A termbase is only as good as the sources behind its entries.

10. Formats and Exchange

Terminology changes hands, so the format matters almost as much as the content.

  • TBX (ISO 30042). The structured exchange format for termbases. Use it when moving between CAT tools and TMS so fields survive.
  • CSV and XLSX. Portable and familiar, but flat, and easy to damage. Good for a simple glossary, risky for a termbase with many fields.
  • Native formats. Every CAT tool and TMS has its own, which is fine inside that tool and a trap when you leave it.

Whichever format you use, agree one convention for column names and for how status is written, and keep a plain-text or CSV export as a backup. The termbase you cannot read after a tool licence lapses is not an asset, it is a hostage.


11. How to Maintain It

Building a termbase is a project. Maintaining it is an operation, and operations need a rhythm.

11.1. Give it an owner and a change process

One owner approves entries in the active language pairs. Anyone can propose a change, but a proposal is not an entry until it is reviewed. Record the decision, the reason and the date, so a term changed in March can be explained in September.

Review on a cadence

Review the termbase at defined points: before a new release, before a new language is added, and at least once per quarter for an active product. A termbase reviewed only when something breaks is already behind.

Handle the lifecycle of a term

Terms are born, they change, and they die. When a product renames a feature, mark the old term deprecated rather than deleting it, so existing translations and memories still resolve. Deletion loses the history that explains why the new term exists.

Keep metrics

Track how many terms are covered, how many queries were resolved, and how many entries changed status. Coverage and query volume tell you whether the termbase is being used; a termbase nobody queries is either perfect or invisible, and it is rarely perfect.

Back up and version

Keep dated exports. If a translation memory or a termbase is corrupted, the release history is the only way to rebuild it, and stored versions are cheaper than reconstruction.


12.The Mistakes That Ruin Termbases

  1. No definitions or context. A term pair with no explanation cannot be applied to a new sentence or a new translator. It becomes a guess.
  2. No part of speech. Without it, the same term gets used as a noun and a verb, and one of them is wrong.
  3. No status field. If everything is "a term", nothing is approved. The reader cannot tell preferred from tolerated from forbidden.
  4. No owner. Shared ownership becomes no ownership, and entries contradict each other within months.
  5. Dumping everything in. A termbase that includes every noun is noise. Include what needs a decision.
  6. No versioning. Without versions you cannot answer "which terminology did this delivery use?", which is exactly the question a client asks when something is disputed.
  7. Ignoring case, number and gender. A target term that only works in one grammatical form will break in the next sentence.
  8. Confusing source and target columns. Obvious in the abstract, common in a rushed spreadsheet, and fatal when the tool imports the wrong column as the term.
  9. No link to the style guide. Terminology and style overlap. If the two documents disagree, both lose authority.
  10. Never revisiting it. A termbase that has not changed since launch is describing a product that no longer exists.

13.Governance at Scale

When several vendors, several languages and a client all touch the same terminology, the termbase stops being a file and becomes a governance problem. Three rules keep it under control.

One authoritative source. There is a single master termbase, and everything else is a derived export. If two teams each maintain their own master, they will diverge and the divergence will surface in front of the client.

An escalation path. Translators propose, reviewers confirm, the owner arbitrates. When a vendor and a client disagree on a term, the escalation path should already exist, so the disagreement is resolved by decision rather than by whoever delivered last.

Confidentiality. Termbases often contain unreleased product names and client strategy. Treat a termbase as confidential client material, store it where client material belongs, and never paste it into a public tool. This is the same obligation that applies to the source text, and it does not relax because a termbase feels like metadata.

Terminology is one of the disciplines the Translating Terminology and Resources course covers in depth, and it sits inside the broader process work of the Localization Management Program.

Translating Terminology and Resources course cover
The terminology workflow end to end · Translating Terminology and Resources

At scale, governance is also a staffing question. As terminology becomes the layer that grounds AI output, the role moves from maintaining lists to owning the data that constrains generation. The course closes on that shift: fewer people doing term entry by hand, more people deciding which concepts are authoritative, how they are structured and when a model is allowed to use them.

Frequently Asked Questions

What is the difference between a glossary and a termbase?

A glossary is a flat list of term pairs, usually source and target with an occasional note. A termbase is a structured database of entries, with fields such as part of speech, definition, context, status and domain. A glossary answers "what is the equivalent?"; a termbase also answers "when, why and how should I use it?".

Can I build a termbase with free tools?

Yes. OmegaT, Weblate, Crowdin's free tier, Poedit and Xbench cover storage, term recognition and checking, and a free concordancer such as AntConc helps with extraction. You only need a paid platform when terminology has to be enforced across large teams and automated workflows.

How big should a termbase be?

There is no target size, and bigger is not better. A termbase should contain the terms that need a consistent decision, with enough context to be applied correctly. A small, accurate, trusted termbase outperforms a large, unverified one because people actually use it.

What is TBX?

TBX, the TermBase eXchange format, is the standardised XML format for termbases, published as ISO 30042. It is the format to use when moving a termbase between CAT tools and TMS because it carries the structure with the terms.

Should I use a language model to build terminology?

For proposing candidates and drafting definitions, yes, and it saves real time. For deciding which term is approved, no. Terminology is a decision owned by the client or the subject matter expert, and a model cannot authorise a term. Use AI for extraction, keep the approval human, and record who decided.

Who should own the termbase?

One person, usually a terminologist or localization lead, with the authority to approve entries in the active languages. Everyone else proposes. A termbase with several owners and no single arbiter will contradict itself within a few months.

Related reading on this blog: the interview question about terminology in multi-translator projects, and how to choose the TMS that will store your termbase.

Terminology is a skill employers test. It is one of the most repeated requirements in language job postings, and it is one of the first things an interviewer asks about. The terminology and management tracks are collected at TranslaStars courses, and the openings that value it most are on the jobs board.