Master Art of Translation | The Terminology System — How Concepts, Terms, Glossaries and Termbases Keep Meaning Stable Across Languages

EDKSG-TRANS-MASTER-WORLD-050

Translation terminology management is the part of translation that decides whether the same concept keeps the same identity when it moves across languages, teams, documents, products and years. People searching for a translation glossary, bilingual glossary, terminology database, termbase, technical translation terminology, approved translations, terminology consistency, term extraction, CAT-tool terminology, AI translation glossary or machine-translation glossary are usually trying to solve one deeper problem: how do you stop an important concept from changing simply because different people choose different words?

A strong terminology system does more than collect vocabulary. It separates concepts from words, distinguishes preferred terms from merely possible synonyms, records definitions and domains, stores approved equivalents by language and locale, identifies forbidden or deprecated forms, supplies evidence, and pushes those decisions into translation, review, machine translation, AI workflows and quality assurance. That is why a professional termbase is not just a longer dictionary and why a bilingual glossary is most useful when it is built around meaning rather than around alphabetical convenience.

This guide explains how terminology works inside translation, how to build and govern a termbase, how glossaries differ from dictionaries and translation memories, how concept-oriented terminology prevents false equivalence, how terminology interacts with vocabulary and English, how approved terms should be researched and reviewed, and how human translators, CAT tools, machine translation and AI systems can all use the same terminology architecture without letting consistency become mechanical or inaccurate.

What this article owns in the Master Art of Translation architecture

This article owns the terminology layer of the translation spine. The earlier architecture nodes establish source-text analysis, equivalence, context and the translation unit. Terminology begins after those foundations are visible. It asks a narrower question: when a source expression represents a stable concept that matters across a project, what should that concept be called in each language, and how do we keep the choice controlled?

That boundary matters because terminology overlaps with vocabulary without being the same thing. Vocabulary teaches what words mean, how they collocate, how they change form and how they behave in English. Translation terminology decides which lexical form is authorised for a concept inside a specific multilingual system. A learner may know five accurate synonyms. A translator may still need to use only one because the product, law, scientific discipline or organisation has standardised it.

This article therefore does not replace the protected Vocabulary Learning Hub or How English Works. It routes outward when the underlying problem is vocabulary depth, morphology, grammar, syntax or English usage. The terminology node owns the governance decision that sits between understanding language and deploying one controlled multilingual label.

The hidden problem: the right word can still be the wrong term

A translator opens a dictionary, finds an accurate equivalent and uses it. Another translator later finds a different accurate equivalent. A reviewer prefers a third. A machine-translation system produces a fourth. Every sentence sounds acceptable in isolation. The document set becomes unstable.

This is not primarily a vocabulary problem. It is a governance problem. The organisation has never decided whether the source expression is an ordinary word or a controlled term, what concept it designates, what target equivalent should be preferred, which alternatives remain acceptable, and which alternatives should be rejected because they imply a different concept.

Terminology management exists because accuracy is sometimes distributed across documents rather than contained inside one sentence. A software product may use the same feature name in a button, a manual, a help centre, release notes and legal terms. A medical device may use one component name in engineering drawings, safety labels, service documentation and training. A school system may use an official assessment term across policy, curriculum and examinations. Consistency becomes meaning infrastructure.

Words, terms and concepts are different objects

A word is a linguistic form. A term is a linguistic designation used for a concept in a particular subject field or controlled environment. A concept is the underlying unit of knowledge: the thing, process, state, relation or category being designated.

This distinction is foundational. The same word can designate different concepts. “Cell” can refer to a biological unit, a spreadsheet location, a prison room or a telecommunications coverage area. The written form is the same. The concept changes. A terminology system should not store one giant entry called “cell” and attach every possible translation to it as if all meanings were interchangeable.

Instead, it creates separate concept records. Each record answers: what concept is this, in which domain, with which definition, and which designations name it in each language? Once the concept is stable, translation choices become easier to govern.

The concept-first rule

The strongest terminology systems are concept-oriented. They do not begin with “What is the translation of this word?” They begin with “Which concept is this source expression naming here?”

That changes the workflow. The terminologist identifies the domain and context, researches the concept, writes or validates a definition, distinguishes neighbouring concepts, then assigns one or more language-specific terms to the concept. This order reduces the chance that a familiar-looking bilingual equivalent is accepted merely because the strings resemble each other.

Concept-first work is especially valuable in technical, legal, scientific, medical and institutional translation, but the principle also helps in ordinary projects. Product names, user-interface labels, programme names, internal process names and brand phrases can all behave like controlled terminology when inconsistent naming creates confusion.

One concept can have many terms

A concept may have a preferred term, admitted synonyms, abbreviations, short forms, legacy terms, colloquial forms and deprecated labels. The terminology record should not pretend these forms do not exist. It should tell the translator what status each form has.

For example, a product may have an official interface term and a longer documentation term. A scientific concept may have an accepted full name and an acronym. A government programme may have an official legal title and a public-facing short name. The target language may need its own preferred form rather than a mechanical translation of every source variant.

This is why a good termbase stores status information. “Preferred” means use this form by default in the defined context. “Admitted” means acceptable under stated conditions. “Deprecated” means historically encountered but generally avoid. “Forbidden” means do not use because it is incorrect, misleading, legally risky or inconsistent with current policy.

One term can represent many concepts

The reverse problem is polysemy. A term string can represent different concepts across domains. “Port” can be a harbour, a computer connection, the left side of a vessel or a fortified wine. “Model” can be a representation, a product variant, a mathematical structure, a person displaying clothing or an AI system.

Terminology systems must therefore resist string-only matching. If a tool sees the source string and inserts the same target term everywhere, it may create consistent error. Context, domain and concept identification remain necessary.

This is one reason terminology QA should flag candidate terms rather than assume that every exact string match is automatically correct. Human review, metadata and context-aware automation all exist to prevent a termbase from becoming a very efficient error amplifier.

Termbase, glossary, dictionary and translation memory

These resources are related but serve different jobs.

A dictionary describes language broadly. It may provide definitions, senses, pronunciation, examples, grammatical information and possible equivalents. It is usually descriptive rather than project-governing.

A glossary is often a simpler list of terms and explanations or source-target equivalents. It may be perfectly sufficient for a small project.

A termbase is a structured terminology database. It can store concept-oriented records, multiple languages, definitions, term status, domains, notes, sources, grammatical data and workflow metadata.

A translation memory stores source-target segments that were translated previously. It remembers how sentences or phrases were translated in context. A termbase controls terminology inside those sentences.

The distinction matters operationally. A translation memory may suggest a complete sentence containing an outdated term. The termbase may tell the translator that the term has changed. Terminology should therefore remain independently governable.

What a professional terminology record needs

A minimal record can contain a source term and target term. A stronger record contains enough evidence and metadata for another translator to understand why the decision exists.

Useful fields include: concept identifier, definition, domain, source-language term, target-language term, locale, term status, part of speech, grammatical gender where relevant, number behaviour, approved abbreviation, prohibited variants, context example, source of evidence, date of approval, owner, reviewer, project applicability and notes on exceptions.

Not every project needs every field. The principle is proportionality. A three-page brochure may need a simple bilingual table. A global medical-device programme may need richly governed multilingual records with traceable sources and review history.

Definitions are the stabiliser

When terminology disagreements become circular, the missing element is often the definition. Two people argue about which word is better without first agreeing on the concept.

A good terminology definition is precise enough to distinguish the concept from its neighbours. It should not merely repeat the term. “A widget is a type of widget used for widgeting” is structurally useless. The definition should identify the broader category and the characteristic that separates this concept from related ones.

Definitions also expose false equivalence across languages. If two candidate target terms seem similar, compare their definitions and domain usage. The closer-looking word may designate a narrower, broader or culturally different concept.

Domain is not optional metadata

Domain tells the system where the concept belongs. It can be broad, such as medicine, finance or education, or narrow, such as radiology, retail banking or primary-school assessment.

Without domain information, ambiguous strings become difficult to control. The term “consideration” behaves differently in ordinary conversation and contract law. “Force” behaves differently in physics and everyday English. “Class” behaves differently in education and object-oriented programming.

Domain metadata also helps tools rank terminology suggestions. A termbase serving many business units can contain multiple valid translations for the same source string if each belongs to a different concept and domain.

Locale is part of terminology

A language code alone may be too coarse. Different locales can prefer different technical vocabulary, spelling, regulatory labels and public terminology.

English itself demonstrates the problem. “Elevator” and “lift” name the same broad concept but belong to different dominant usage patterns. “Cell phone” and “mobile phone” vary similarly. In professional translation, locale differences can be more consequential than ordinary vocabulary preferences because regulators, standards bodies and industry communities may use local official terms.

A mature termbase therefore records language and locale when the distinction matters. It does not assume that one target term is universal simply because readers can understand it.

Evidence hierarchy: where approved terms should come from

Terminology decisions should be supported by evidence appropriate to the domain. A useful hierarchy begins with binding or authoritative sources where they exist: legislation, regulators, standards, official product nomenclature, client-approved resources, scientific bodies, professional associations and recognised institutional databases.

Next come high-quality domain publications, specialist textbooks, peer-reviewed literature, manufacturer documentation and well-established professional usage. General dictionaries and web frequency are useful evidence but should not automatically outrank domain authority.

Search results are especially dangerous when low-quality machine-translated pages repeat the same error. Frequency is not proof. Terminology research asks not only “How often does this phrase occur?” but “Who uses it, in what domain, in which locale, with what definition and with what authority?”

Current professional reference points

Terminology practice has formal and institutional reference points. ISO 12616-1:2021 addresses fundamentals of translation-oriented terminography for producing bilingual and multilingual terminology collections. ISO 30042:2019 specifies TermBase eXchange, or TBX, for terminology-resource exchange; a third edition is under development in 2026. ISO/TR 24633-1:2026 provides a new companion overview to the TBX framework. ISO 17100:2015 remains the published translation-services standard while a new edition is being developed.

The European Commission describes terminology as essential for legal certainty, coherence and multilingual communication and maintains IATE, a concept-oriented terminology system containing roughly 600,000 records and more than six million terms. Commercial translation platforms such as Phrase and memoQ likewise distinguish termbases from translation memories and use terminology resources to guide translators, reviewers and machine-translation workflows.

These references do not create one universal terminology policy for every project. They show the recurring professional pattern: identify concepts, document terms, control status, preserve metadata and make terminology reusable across workflows.

The terminology lifecycle

Terminology is not a one-time spreadsheet exercise. It has a lifecycle.

First, a candidate term is identified. Then the concept is researched. A definition is confirmed or created. Existing source-language variants are collected. Target-language candidates are researched. One or more forms are approved. The record is deployed into translation tools. Translators and reviewers use it. Quality checks detect violations. New evidence may trigger review. A term may later be revised, deprecated or retired.

This lifecycle is what turns terminology into infrastructure. The record remains alive because the product, law, science, organisation and language can change.

Step 1: identify candidate terminology

Not every noun deserves a termbase record. Candidate terminology usually has one or more signals: it is domain-specific, repeated, safety-critical, legally significant, branded, likely to be translated inconsistently, difficult to interpret, required by a standard, or central to how the system is organised.

Automatic term extraction can help by identifying frequent or statistically distinctive expressions. Human review remains necessary because frequency alone does not distinguish important terms from common words.

A practical project can begin with a seed list of twenty to fifty high-leverage terms rather than attempting to capture every specialist word before translation starts.

Step 2: collect source context

A candidate term without context is only a string. Collect representative source sentences, section headings, diagrams, interface screenshots and definitions. Note where the term appears and whether the author uses it consistently.

Context helps distinguish homonyms and reveals grammatical behaviour. It also exposes whether the source itself is terminologically unstable. If two source labels refer to the same concept, the terminology project may need source-language harmonisation before target-language standardisation.

This is one reason terminology management can improve authoring as well as translation.

Step 3: define the concept

Write or verify a definition that distinguishes the concept. If a reliable authoritative definition exists, record its source. If the project needs an internal operational definition, mark it as such rather than presenting it as an external standard.

Then identify neighbouring concepts. What is broader? What is narrower? What is commonly confused with it? Which feature makes this concept different?

This conceptual neighbourhood is often more useful than a long list of synonyms because it tells translators what the term must not become.

Step 4: research target-language candidates

Generate candidates from dictionaries, official bilingual resources, parallel texts, specialist publications, corpora and authoritative websites. Search for the concept, not only for the source term.

Check whether the candidate is actually used by target-language professionals. Inspect collocations and grammatical behaviour. Verify whether it belongs to the right locale. Look for competing terms and determine whether they are true synonyms, regional variants, outdated forms or names for related concepts.

Record evidence as you research so the decision remains reviewable.

Step 5: assign status

Choose the preferred target term and classify alternatives. A useful status model can include preferred, admitted, deprecated and forbidden.

The reason for status should be documented when it is not obvious. A term may be forbidden because it names a different concept, conflicts with regulation, is associated with an old product version, belongs to the wrong locale or creates ambiguity.

Status makes terminology actionable. Without status, the termbase becomes a collection of possibilities rather than a decision system.

Step 6: record grammatical and lexical behaviour

A term is used inside sentences, so its grammatical behaviour matters. Record part of speech, countability, plural form, grammatical gender where relevant, required prepositions, common compounds and approved abbreviations.

This information is especially useful for inflected languages and for terminology QA. A checker that expects only one surface form may incorrectly flag valid inflections. The system needs enough linguistic information to distinguish variation from inconsistency.

This is one of the points where terminology connects directly to vocabulary and grammar knowledge.

Step 7: validate with the right owner

Who approves a term depends on the project. It may be a subject-matter expert, legal team, product owner, terminologist, senior translator, regulator-facing specialist or brand team.

Approval should not become committee theatre. The goal is clear authority. If everyone can change terminology informally, no one actually governs it.

For high-risk terminology, record reviewer and approval date. Traceability becomes useful when a future team asks why an unusual form was chosen.

Step 8: deploy into the workflow

A terminology resource has little value if translators never see it. Integrate the termbase into the CAT environment, writing system, review checklist, machine-translation glossary or AI prompt architecture as appropriate.

The translator should receive a suggestion at the moment the source term appears. Reviewers should be able to check terminology use. Automated QA should flag missing or forbidden target forms where technically possible.

Deployment converts terminology from documentation into behaviour.

Step 9: monitor real usage

Once translation begins, collect evidence from actual use. Does the preferred term fit naturally into sentences? Does it inflect correctly? Does it create unwanted ambiguity? Do target readers recognise it? Are translators repeatedly querying the same record?

A termbase should be stable enough to create consistency and flexible enough to correct bad decisions. Stability is not stubbornness.

Recurring translator objections are signals. Sometimes the translators need clarification. Sometimes the terminology record is incomplete. Sometimes the approved term is wrong.

Step 10: revise without losing history

When terminology changes, preserve history. Record the old form, new form, effective date, reason and affected products or document versions.

This prevents future teams from “rediscovering” an old term and assuming it is still current. It also supports migration: translation memories, documentation and interfaces may need controlled updates.

Terminology versioning is especially important in regulated, legal, scientific and software environments where different historical versions may need to remain interpretable.

The minimum viable termbase

A small project does not need enterprise terminology software to benefit from terminology discipline. A spreadsheet can work if each row is treated as a concept record rather than merely a bilingual word pair.

At minimum include: concept ID, domain, source term, definition, preferred target term, status, context example and evidence source. Add notes for exceptions.

The important shift is mental: the resource exists to control a concept, not to impress anyone with complexity.

When a simple glossary is enough

If the project is short, low risk and handled by one translator, a two-column or three-column glossary may be sufficient. The cost of managing rich metadata can exceed the benefit.

But even a simple glossary should distinguish “preferred translation” from “one possible translation.” If the list has been approved for the project, say so. If it is only a research aid, do not present it as binding.

Terminology architecture scales down as well as up.

Term extraction: finding the vocabulary that deserves control

Manual extraction begins with reading. Highlight repeated specialist nouns, compounds, product labels, process names, measurement terms, legal defined terms, abbreviations and unusual phrases.

Automatic extraction can rank candidates by frequency, statistical distinctiveness or linguistic patterns. This accelerates large projects but cannot decide conceptual importance by itself.

A frequent word such as “system” may not need control. A rare safety term may be critical. Selection requires domain judgement.

Terminology from corpora and comparable texts

Comparable target-language texts show how professionals already talk about the concept. They are often more valuable than isolated bilingual lookup because they reveal natural collocation, register and genre.

A translator working on laboratory equipment should read target-language manuals, standards and specialist documentation. A translator working on educational assessment should examine official target-language curriculum and testing terminology. A legal translator should inspect relevant official bilingual legislation where available.

Comparable texts answer a practical question: what would a competent target-language professional call this concept when not translating?

Parallel texts are powerful and dangerous

A parallel text is a source text paired with a translation. Official parallel corpora can provide valuable evidence, especially when the same institution controls terminology across languages.

But an existing translation can also contain errors, old terminology or context-specific choices. Repetition does not guarantee current correctness.

Use parallel texts as evidence, then verify the concept, date, domain and authority.

Terminology and collocation

Approving a noun is not enough if translators do not know how it behaves. Technical terms often have characteristic verbs, adjectives and prepositions.

A term record can therefore include common collocations. This supports natural target writing and prevents literal source-language combinations from leaking into translation.

Collocation information is especially useful when the preferred target term is correct but unfamiliar to general translators. It shows the term living inside real sentences.

Terminology and morphology

Languages change word forms through number, case, gender, tense, derivation and compounding. Terminology control should therefore distinguish concept identity from surface form.

A target term may appear in plural, inflected case forms or compounds without becoming a different term. Automated QA needs language-aware matching where possible.

Conversely, a related derivative may designate a different concept. “Authenticate,” “authentication” and “authenticator” belong to one conceptual family but do not perform the same grammatical or terminological job.

Abbreviations and acronyms

Acronyms create a special terminology problem because the source and target may share, translate or replace them differently.

Record the full form, abbreviation, language, status and first-use policy. Do not invent a target acronym merely because a translated full form could theoretically produce one.

In many professional communities the source-language acronym remains standard even when the full term is translated. Evidence from target-domain usage should decide.

Names are adjacent to terminology but not identical

People, organisations, products, laws, programmes and places need naming control, but they are entities rather than ordinary concepts.

A project should maintain a name list alongside the termbase. Record official translated names, transliterations, abbreviations, legal names and do-not-translate instructions.

This separation keeps the terminology model clean while still preventing identity drift.

Do-not-translate terminology

Some terms must remain unchanged: product names, command strings, trademarks, code identifiers, protocol names or brand expressions.

Do-not-translate status should be explicit. Otherwise a well-meaning translator or AI system may localise the string and damage recognisability or functionality.

When unchanged source-language terms appear inside target grammar, provide guidance on articles, pluralisation and surrounding syntax so the result remains natural.

Brand terminology

Brands often create proprietary names for features, programmes, subscription levels and processes. These terms may be semantically ordinary but operationally controlled.

Terminology governance prevents marketing, support, product and legal teams from naming the same feature differently across languages. It also prevents translators from “improving” deliberate naming.

The termbase should distinguish official brand terms from descriptive prose so translators know where variation is allowed.

Legal terminology

Legal terminology requires special caution because superficially similar institutions and concepts can differ across jurisdictions.

A target term should not imply legal equivalence merely because it sounds familiar. Research the concept, legal system and authoritative bilingual usage. Defined terms in contracts should be recorded and reused consistently.

Terminology management supports legal translation, but it does not turn linguistic work into legal advice. Where legal effect depends on jurisdictional interpretation, qualified legal review may be necessary.

Medical terminology

Medical terminology combines scientific precision, regulatory constraints and audience needs. A clinician-facing term may differ from a patient-facing explanation without changing the underlying concept.

The termbase should therefore record audience or usage notes when one concept has both technical and plain-language designations. It should also distinguish symptoms, diagnoses, procedures, devices and anatomical terms carefully.

Near-synonyms can be clinically meaningful. Terminology review should prioritise patient safety and authoritative medical usage over stylistic variety.

Scientific terminology

Science often appears internationally standardised, but translation still encounters competing nomenclature, abbreviations, legacy terms and discipline-specific meanings.

Check standards, scholarly usage and field conventions. Preserve distinctions between process, quantity, unit, model and measurement.

Do not replace repeated scientific terms with elegant synonyms merely to improve style. Repetition can signal conceptual identity.

Technical terminology

Technical documentation is one of the strongest use cases for terminology management because components, procedures and states recur across manuals, interfaces, service documents and training.

Termbases reduce variation such as one component receiving three target names across different manuals. They also support new translators who do not yet know the product.

For complex systems, connect terminology to diagrams, part numbers or screenshots. Visual reference can resolve ambiguity that text alone cannot.

Software and interface terminology

Software terms appear in buttons, menus, settings, notifications, documentation and support content. The same feature name must often remain recognisable across all of them.

Record whether a term is an interface label, conceptual feature name or descriptive action. A button may require a short form while documentation uses a longer noun phrase.

Context screenshots and string identifiers can belong in the terminology workflow because short UI strings are often semantically underdetermined.

Business terminology

Organisations create internal terminology around roles, processes, metrics and products. Translators may be tempted to normalise these expressions into generic business language.

First determine whether the expression is truly generic or whether it names an internal concept. If it is controlled, document it. If it is ordinary prose, allow natural variation.

Terminology control should protect identity without turning every corporate phrase into a rigid term.

Academic terminology

Academic translation depends on disciplinary vocabulary. The same ordinary word may become technical inside linguistics, economics, sociology, mathematics or philosophy.

Comparable target-language scholarship is essential. Use the terminology the target discipline recognises, not merely the closest bilingual-dictionary equivalent.

When a source author defines a term idiosyncratically, the translation may need to preserve that local definition even if broader literature uses the expression differently.

Educational terminology

Education systems have official names for subjects, stages, assessments, qualification levels and pedagogical concepts. These labels can look ordinary while carrying institution-specific meaning.

Do not casually replace one system’s concept with another country’s nearest familiar label if the institutions differ. A short explanatory gloss may be more accurate than false equivalence.

For learning materials, terminology also affects teachability. Students need stable labels for concepts so explanations, exercises and feedback connect.

Marketing terminology and transcreation

Marketing translation permits more creative adaptation than technical documentation, but brand terminology remains controlled. Product names, feature names, programme names and legal claims may need exact consistency even while surrounding copy changes freely.

A terminology system therefore creates the fixed points inside transcreation. It tells the creative translator which elements are negotiable and which are not.

This balance prevents brand consistency from becoming robotic and creativity from becoming identity drift.

Style guide versus termbase

A termbase answers “What do we call this concept?” A style guide answers broader questions about how the target language should be written.

Style guidance can cover tone, punctuation, numbers, dates, capitalisation, sentence length, inclusive language, address forms and spelling conventions. Terminology can link to those rules but should not become overloaded with every writing preference.

Keeping the resources separate but connected makes each easier to maintain.

Translation memory versus termbase

Translation memory reuses previous segments. Termbases reuse conceptual naming decisions. The two resources solve different kinds of repetition.

A translation memory match can be contextually wrong even when it is textually similar. A termbase match can be conceptually wrong when the same source string names another concept. Both resources require judgement.

Used together, they accelerate translation while keeping important lexical decisions controlled.

Terminology and machine translation

Machine-translation systems can use glossaries or terminology constraints to encourage preferred equivalents. This is useful for branded, technical and repetitive content.

But glossary injection does not guarantee perfect grammar or correct sense selection. The engine still has to integrate the term into the target sentence, and the source string may be ambiguous.

Terminology therefore improves machine translation most when the source terminology is stable, the glossary is conceptually clean and post-editing still verifies context.

Terminology and AI translation

Large language models can follow terminology instructions remarkably well when the instructions are explicit, but they may also substitute synonyms because natural variation is part of ordinary language generation.

For controlled translation, provide a terminology table with source term, approved target, status, domain and exceptions. Tell the model that preferred terms are mandatory where the specified concept is present and that it must not apply a term merely because the source string matches in another sense.

Then run a separate terminology QA pass. Generation and verification should remain distinct operations.

A practical AI terminology prompt pattern

A robust instruction can say: translate for the specified audience and locale; use the supplied preferred terms when they match the listed concepts; preserve do-not-translate items; do not invent target equivalents for unresolved terms; flag ambiguous source uses; keep terminology consistent across headings, body text and interface labels; and return a list of any terms that could not be applied confidently.

The important detail is concept matching. A blind instruction such as “always translate X as Y” can create errors when X is polysemous.

The prompt should therefore contain enough domain or definition context to tell the model what the term record means.

Terminology quality assurance

Terminology QA asks whether required terms were used, forbidden terms avoided and source concepts preserved.

Automated checks can detect exact term matches, missing target forms, prohibited variants, inconsistent capitalisation and sometimes morphology. Human review checks whether the concept actually matches the record, whether the term fits grammar and whether controlled consistency has become unnatural.

Quality assurance is strongest when automated precision and human interpretation are combined.

False positives in terminology QA

A checker may flag a translation even though the preferred term appears in a valid inflected form. It may detect a forbidden string that occurs inside another word. It may insist on a term when the source string represents a different concept.

These false positives matter because noisy QA trains reviewers to ignore warnings. Good terminology engineering improves matching rules, morphology handling, domain filters and exception logic.

A warning system is valuable only when users trust that warnings deserve attention.

False negatives in terminology QA

The opposite problem occurs when an incorrect term is not detected. The source may use an inflected variant, synonym or abbreviation not included in the record. The target may paraphrase the concept rather than using a forbidden term string.

Improve recall by adding known source variants, approved abbreviations and common error forms. For high-risk projects, supplement string checks with reviewer attention and context-aware tools.

No automated terminology check should be mistaken for full semantic review.

Terminology governance

Governance answers who can propose, approve, change and retire terms. It also defines where decisions are stored and how users learn that a term has changed.

Small teams can govern terminology informally with one owner and a shared glossary. Large organisations may need proposal workflows, reviewers by domain, version history and release processes.

The governance mechanism should be as light as possible while still preventing uncontrolled drift.

Roles in terminology work

Common roles include translator, terminologist, subject-matter expert, reviewer, product owner, legal reviewer and localisation manager. One person may perform several roles in a small project.

The critical point is not job title but responsibility. Someone must own conceptual accuracy. Someone must own target-language quality. Someone must have authority to approve the final status when disagreement remains.

Undefined authority creates endless terminology debate.

Terminology conflict resolution

When two target terms compete, compare them against explicit criteria: conceptual accuracy, authoritative usage, target-domain frequency, locale fit, register, legal or regulatory requirement, user comprehension, brand policy and grammatical usability.

Do not settle the conflict by personal taste alone. Record the evidence and decision.

If both terms are valid in different contexts, encode the contexts instead of forcing one universal winner.

Terminology change management

Changing an approved term can affect translation memories, software strings, documentation, training, legal text, search indexes and user habits.

Before change, identify impact. Decide whether legacy content should be updated or preserved. Communicate an effective date. Add the old term as deprecated so QA can detect it in new work.

Terminology is a network dependency. Rename one node carefully.

Terminology versioning

A useful record can include created date, modified date, status history and applicable product version. This allows translators to answer questions such as: Which term was valid in version 4? When did the regulatory label change? Is this legacy manual supposed to retain the old name?

Version-aware terminology prevents anachronistic “corrections” in historical content.

It also supports auditability without requiring every translator to remember the history personally.

Terminology migration and mergers

When organisations merge systems, terminology resources can contain duplicates, conflicting IDs, different definitions and different status models.

Do not simply concatenate two glossaries. Reconcile concepts. Determine whether similar records are identical, overlapping or distinct. Preserve provenance so old translations remain interpretable.

Terminology migration is concept mapping, not spreadsheet merging.

TBX and interoperability

TermBase eXchange, or TBX, exists because terminology needs to move between systems. A termbase trapped inside one application loses long-term value.

TBX provides a structured framework for exchanging terminology data. ISO 30042:2019 is the current published standard, while a new edition is advancing through ISO development in 2026. ISO/TR 24633-1:2026 adds a current companion overview to the framework.

You do not need to hand-write TBX files to benefit from the principle. Design terminology data so concepts, languages, terms and metadata remain separable and exportable.

IATE as a model of concept-oriented multilingual terminology

IATE demonstrates what large-scale concept-oriented terminology looks like. The European Union uses one system to document multilingual terms around shared concepts, with definitions, sources, domains and language-specific designations.

The lesson for smaller projects is not to imitate institutional complexity. It is to notice the architecture: concept record first, language-specific terms beneath it, evidence around the decision.

That same pattern can exist in a twenty-row school glossary or a million-term institutional database.

Failure mode 1: the alphabetic bilingual list

A source term appears in column A and a target term in column B. Nothing identifies the concept, domain or context.

This works until one source string has multiple meanings. Then translators cannot know which target equivalent belongs to which concept.

Repair by adding concept IDs, definitions or at least domain and context fields.

Failure mode 2: collecting synonyms without status

The glossary contains five target synonyms but does not say which one to use. Translators choose freely and the project remains inconsistent.

A terminology resource should reduce decision load. If every record merely reproduces the dictionary’s list of possibilities, it has not completed the governance job.

Assign preferred and admitted status with usage notes.

Failure mode 3: over-controlling ordinary vocabulary

Some teams put every common word into a glossary and demand one fixed translation. Target language becomes repetitive and unnatural.

Control only where conceptual identity, brand, safety, law, user comprehension or workflow consistency requires it. Ordinary vocabulary should remain responsive to context.

Terminology is selective rigidity.

Failure mode 4: translating the source label instead of the concept

A familiar source word triggers a familiar target word. No one checks what the source author means in this domain.

This creates false friends and false equivalence. The repair is concept research: definition, context, neighbouring concepts and target-domain usage.

The term record begins with knowledge, not typography.

Failure mode 5: termbase without deployment

The organisation creates a beautiful glossary PDF that translators never open.

Terminology should appear inside the workflow where possible: CAT-tool suggestions, linked resources, machine-translation glossaries, review checklists or AI instructions.

A terminology decision that cannot reach the person making the translation decision is not operational.

Failure mode 6: legacy terminology contaminates new content

Old translation memories continue serving a deprecated term. New translators trust the memory match because it is a 100% textual match.

Repair by maintaining terminology independently, marking old terms as deprecated and running QA against new work. Where necessary, batch-update translation memories under controlled conditions.

Reuse must remain subordinate to current terminology.

Failure mode 7: machine translation obeys the glossary in the wrong sense

A glossary forces every occurrence of “charge” to one target term because the system cannot distinguish price from accusation or electrical property.

The fix is sense-aware terminology: separate concept records, use domain context, reduce ambiguous source strings where possible and review machine output.

Consistency without concept control is systematic error.

Failure mode 8: the glossary becomes stale

Products change, regulations change, language changes and the glossary remains untouched for years.

Assign ownership and review triggers. A terminology resource should be revised when the domain changes, not only on an arbitrary calendar.

Staleness is especially dangerous because an old termbase looks authoritative.

Failure mode 9: no evidence

A term is marked approved because “someone decided it years ago.” No source, definition or rationale exists.

The record becomes difficult to challenge even when wrong and difficult to defend even when right.

Record enough evidence that a future reviewer can understand the basis of the decision.

Failure mode 10: terminology as personal preference

Reviewers replace accurate approved terms with synonyms they happen to prefer. Translators reverse the changes. Review becomes conflict rather than quality control.

Terminology governance separates preference from policy. If the approved term is wrong, change the term record through the proper process. Do not create silent exceptions sentence by sentence.

This makes review faster and less personal.

Worked example 1: “cell” in biology and spreadsheets

Suppose a multilingual education platform contains both science lessons and spreadsheet tutorials. The source word “cell” appears in both.

A string-oriented glossary has one entry and one target translation. It will eventually fail. A concept-oriented termbase has at least two records: biological cell and spreadsheet cell, each with domain, definition, target term and context example.

The same source string now routes through two different conceptual records. Terminology becomes accurate rather than merely consistent.

Worked example 2: “appeal” in ordinary and legal language

In ordinary English, “appeal” can mean attractiveness or a request. In law it can name a formal process of asking a higher authority or court to review a decision.

The legal concept should receive its own record, definition, jurisdiction notes and official target equivalent. A marketing translation of “appeal” should not inherit that legal term merely because the source string matches.

Domain metadata prevents absurd consistency.

Worked example 3: a renamed software feature

Version 3 calls a feature “Workspace.” Version 4 renames it “Project Space.” Old translation memories still contain the previous target name.

Create or update the terminology record with effective version, new preferred term and deprecated legacy form. QA flags the old term in new content. Legacy manuals remain interpretable because the history is preserved.

Terminology versioning turns a rebrand into a controlled migration instead of a search-and-replace accident.

Worked example 4: a medical plain-language pair

The clinical concept has a technical designation used by physicians and a plain-language phrase used in patient education. Both are correct. They are not interchangeable in every document.

The termbase records both designations and marks audience usage. The medical report uses the technical term. The patient leaflet uses the approved plain-language equivalent while preserving the concept.

Controlled terminology can include controlled variation.

Worked example 5: a legal defined term

A contract defines “Confidential Information” with a capitalised label. The translator uses a different synonym in each clause to avoid repetition.

The translation becomes legally harder to track because readers may wonder whether the variants refer to the defined concept or to ordinary confidential material.

The termbase marks the defined term as mandatory in the contract. Repetition is correct because identity matters more than stylistic variety.

Worked example 6: a technical component and a common noun

A machine manual uses “guard” as the official name of a protective component. Elsewhere the source says a worker should “guard against contamination.”

The termbase record applies only to the component concept. The verb remains ordinary language.

Tools and translators need context so the controlled noun does not colonise every matching string.

Worked example 7: an educational qualification

A qualification name has no exact institutional equivalent in the target education system. Translating it as the nearest local qualification would help readers recognise the level but falsely imply equivalence.

The terminology decision may retain the official source name, use an established translated name, or add a short explanatory description depending on publication purpose.

The concept record stores the institutional identity so future translators do not keep reinventing the mapping.

Worked example 8: an AI product name

A company calls a feature “Reasoning Mode.” Marketing wants a creative local name. Product wants the English name unchanged. Support has already used a literal translation.

This is a governance conflict. Define the feature concept, decide naming policy by locale, approve one form and record legacy variants as deprecated. Then deploy the decision across product, documentation and support.

The termbase becomes the shared contract between departments.

Worked example 9: terminology inside an AI prompt

A translation model receives a table saying “account” must always become one target term. In a finance document, “account” alternates between bank account, accounting record and narrative account.

The instruction is too crude. Replace it with concept-specific records: bank account, ledger account and narrative description, each with definition and target term. Tell the model to apply a record only when the concept matches.

Better terminology architecture produces better prompting.

Worked example 10: a multilingual website search term

The organisation has an official target term for a service, but users search predominantly for a common alternative. Which term belongs on the page?

Terminology policy and SEO can coexist. Keep the official term as the controlled designation while using the common search phrase naturally in explanatory copy where accurate. The termbase can record the search synonym as admitted for discovery but not as the official name.

Governance does not require ignoring real user language.

Diagnostic: should this expression become a termbase entry?

Ask six questions. Does inconsistency create confusion? Does the expression name a domain concept? Is it repeated? Is it legally, medically, technically or commercially important? Does the organisation already have a preferred form? Would a new translator benefit from guidance?

If several answers are yes, a record is probably worthwhile. If the expression is ordinary language whose best translation changes freely with context, keep it in the translator’s vocabulary rather than the termbase.

The goal is not maximum term count. It is maximum decision value.

Diagnostic: is this one concept or two?

Compare definitions, broader categories, distinguishing features and typical contexts. If the source string changes meaning across domains or sentences, separate the records.

Do not use one concept record merely because two expressions look alike. Concept identity is semantic, not orthographic.

This diagnostic prevents the most common structural error in simple bilingual glossaries.

Diagnostic: are two target forms true synonyms?

Check whether both forms designate the same concept in the same domain and locale, with the same register and legal or technical effect.

If one is broader, narrower, older, regional, colloquial or institution-specific, they are not fully interchangeable even if dictionaries list them together.

Term status should encode the distinction.

Diagnostic: is the preferred term actually usable?

Insert the term into representative sentences. Test plural forms, compounds, headings and interface labels. Ask a target-language specialist to read it in context.

A theoretically accurate equivalent can fail operationally if it is grammatically awkward, unknown to the domain community or misleading in common collocations.

Terminology validation should include usage, not only conceptual matching.

Terminology metrics that matter

Large programmes may track number of approved records, unresolved candidates, terminology violations, repeated translator queries, time to approve new terms and percentage of high-risk concepts covered.

Raw termbase size is a weak success metric. A huge database can be unusable. Better measures ask whether terminology reduces inconsistency, rework and uncertainty.

Quality should be measured by decision reliability, not collection size.

Terminology debt

Organisations accumulate terminology debt when products, teams and translations grow faster than naming governance. The symptoms include duplicate feature names, conflicting glossaries, old terms in documentation, reviewer arguments and machine translation that alternates between synonyms.

Terminology debt behaves like technical debt. It makes every future change more expensive because dependencies are unclear.

Repair begins with inventory, concept reconciliation, ownership and staged cleanup rather than mass replacement without context.

Terminology and source-language authoring

The best multilingual terminology systems often begin before translation. Authors use preferred source terms consistently, avoid accidental synonyms in controlled content and define new concepts before publication.

This improves machine translation, human translation, search and product comprehension simultaneously.

Controlled source terminology is not dull writing when applied selectively. It is a coordination device for concepts that must remain stable.

Terminology and controlled language

Controlled language extends beyond terminology by restricting grammar, sentence patterns or style to improve clarity and consistency. Terminology is one component of that broader approach.

Technical authoring may benefit from controlled wording because repeated structures improve translation reuse and reduce ambiguity. Literary or persuasive writing usually needs more freedom.

The translation brief decides how much control is appropriate.

Terminology and search

Users search with their own vocabulary, which may differ from official terminology. Search systems, help centres and SEO pages may therefore need aliases.

A termbase can record official and common forms without confusing their status. Search can index both while published product labels remain controlled.

This is an example of one concept supporting multiple designations for different functions.

Terminology and accessibility

Highly technical terminology can create access barriers. Patient information, public services and educational content may require approved plain-language equivalents or definitions.

The solution is not to erase specialist concepts. It is to manage layers of naming: technical designation, plain-language explanation and perhaps acronym.

Terminology architecture can support both precision and comprehension.

Terminology and inclusive language

Some terms change because professional or social communities adopt more accurate or respectful language. Terminology governance should be able to record current preferred forms and deprecate older ones while preserving historical traceability.

Decisions should be evidence-based and appropriate to domain and locale rather than driven by ad hoc personal preference.

A clear status history helps teams update responsibly.

Terminology and knowledge graphs

A rich terminology system begins to resemble a lightweight knowledge graph. Concepts relate to broader concepts, narrower concepts, parts, processes, products and organisations.

You do not need graph software to benefit from this thinking. Recording relationships can help translators understand why two near-synonyms are not interchangeable.

The conceptual network gives the term meaning beyond its spelling.

Terminology and retrieval-augmented AI

AI translation systems can retrieve project terminology before generating text. The quality of retrieval depends on the quality of the records.

A concept record with definition, domain and preferred target form gives the model better evidence than a bare source-target pair. Retrieval can also surface context notes and forbidden variants.

Structured terminology therefore becomes a high-value knowledge source for AI rather than a legacy resource from pre-AI translation workflows.

Terminology hallucination

AI systems may invent plausible-looking technical terms when they lack evidence. The output can sound professional while no target-domain expert actually uses the expression.

Terminology resources reduce this risk by constraining high-value concepts. For unresolved terms, instruct the system to flag uncertainty rather than coin a confident equivalent.

Fluency should not be allowed to masquerade as terminological authority.

Terminology review for AI output

After AI translation, extract all occurrences of controlled source terms and compare target forms against the termbase. Then inspect context to ensure the record was applicable.

Review forbidden terms separately. Check whether the model paraphrased around the preferred term and whether that paraphrase is acceptable.

This two-stage review catches both under-use and over-application of terminology.

Terminology for multilingual teams

When many translators work in parallel, terminology is shared memory. It prevents each translator from solving the same concept independently and creating several local solutions.

A query channel should connect translators to terminology owners. New decisions should flow back into the shared resource quickly enough that other translators benefit.

Terminology becomes a coordination protocol.

Terminology for students and language learners

Students can use a simplified concept-first method. Instead of memorising bilingual pairs, they write the concept in their own words, note the domain, record the target term and add one example sentence.

This prevents the common belief that every source word owns one permanent translation. Learners begin to see semantic range and context.

Translation terminology becomes an advanced form of vocabulary learning.

Primary-level terminology learning

Young learners do not need the word “terminography.” They can learn the concept through subjects. In science, one word names one classroom concept. In mathematics, a term such as “factor” has a specific meaning. In computing, “file” is not simply any piece of paper.

Ask children to draw the concept, define it simply and compare it with a nearby concept. Then add the equivalent in another language if appropriate.

The foundation is precision, not jargon.

Secondary-level terminology learning

Secondary students can build mini-termbases by subject: physics, chemistry, biology, mathematics, geography, literature or computing.

Each record can include definition, word class, related term, common confusion and bilingual equivalent. Students quickly discover that specialised vocabulary is organised by concept systems rather than alphabetically.

This improves both subject learning and translation judgement.

Advanced learner terminology projects

Advanced learners can research one domain and produce a multilingual terminology set with sources, preferred terms, definitions, contexts and status labels.

They should justify disputed choices using target-language evidence rather than intuition. They can compare dictionary entries, corpora, official terminology and AI suggestions.

This turns vocabulary study into evidence-based language work.

A classroom exercise: three strings, three concepts

Give students the word “field” in three sentences: a farmer walked across the field; the researcher entered the field of linguistics; the database field stores a date.

Ask them to identify three concepts before translating. Then compare target equivalents.

The exercise teaches concept-first translation with almost no technical terminology.

A classroom exercise: approved term versus natural synonym

Give students a technical passage and an approved glossary. Include several places where a natural synonym sounds attractive.

Ask them to decide when the approved term is mandatory and when ordinary language can vary. Require explanation.

The lesson is that good translation balances control and naturalness.

A classroom exercise: termbase audit

Provide a deliberately weak glossary containing polysemous strings, missing definitions, duplicate records and conflicting translations.

Students must redesign it around concepts. They learn that terminology quality is structural, not cosmetic.

This is excellent preparation for AI-era language work because it trains students to build the evidence that machines need.

Vocabulary depth and terminology

Terminology work depends on vocabulary depth: semantic range, collocation, morphology, register, word families and usage constraints.

A translator with shallow vocabulary may accept a target term because its dictionary definition looks close. A translator with deeper lexical knowledge notices that the candidate collocates differently or belongs to another register.

This is where the protected Vocabulary Learning Hub becomes the correct next route when the bottleneck is lexical knowledge rather than terminology governance.

How English Works and terminology

When English is the source or target, grammar and syntax determine how terms behave. Articles, countability, compounds, modifiers and word order can change whether a preferred term fits naturally.

English also compresses specialist concepts into noun compounds that may need unpacking before translation. Understanding the sentence remains a prerequisite for applying the termbase correctly.

When the underlying problem is English structure, route to How English Works rather than trying to turn the termbase into a grammar manual.

The terminology decision matrix

For a disputed term, compare candidates across eight dimensions: conceptual fit, domain authority, locale fit, grammatical usability, target-language naturalness, regulatory or brand requirement, user comprehension and consistency with existing approved resources.

No single dimension always wins. In a contract, legal authority may dominate. In a consumer interface, user comprehension may matter more. In a scientific paper, disciplinary convention may be decisive.

The matrix makes trade-offs explicit.

The terminology confidence scale

Not every term decision has equal certainty. Use a confidence label if helpful: validated, approved with strong evidence, provisional, candidate or unresolved.

This prevents provisional research from masquerading as binding policy. It also tells translators which terms can be trusted automatically and which deserve attention.

Confidence metadata is particularly valuable in fast-moving technical fields.

The terminology query template

When translators need a decision, the query should contain source term, source context, domain, current candidate translations, evidence, the uncertainty and the consequence of choosing incorrectly.

A good query is easier to answer because it narrows the conceptual question. “What does X mean?” is weak. “Does X here name the physical component defined in section 3 or the software function with the same label?” is strong.

Terminology research improves when questions become discriminating.

The terminology review checklist

Before approving a record, ask: Is the concept clearly defined? Is the domain correct? Are competing concepts distinguished? Is the preferred term attested in authoritative target usage? Is the locale correct? Are synonyms classified? Are deprecated and forbidden forms recorded? Is grammar usable? Is the evidence traceable? Does the term create confusion with another existing record?

Then test the term inside real sentences. Approval should survive actual use.

This checklist turns terminology from taste into evidence-based decision.

The project-start terminology sprint

At project start, spend a short focused period identifying the highest-risk terms. Review the source, existing client glossaries, translation memories, product documentation and authoritative external resources.

Approve a small seed termbase before full translation begins. Add new terms during work through a controlled query process.

This is usually faster than attempting to predefine every possible term or postponing all terminology decisions until review.

The continuous terminology loop

During translation, new concepts appear. Translators propose candidates. Experts validate them. Records are updated. QA uses the new decisions. Review feedback produces refinements. The cycle repeats.

The terminology system therefore learns from the project while remaining governed.

This loop is more realistic than treating the glossary as complete before work starts.

Terminology in agile product development

Software teams release quickly. New features may be named days before localisation. If terminology approval lags behind development, translators invent temporary labels that later leak into production.

Integrate terminology decisions into product naming. When a new feature is created, open a terminology record early. Record English source naming, definition, screenshots and target-locale decisions before strings multiply.

Terminology becomes part of product design rather than post-release cleanup.

Terminology in high-risk translation

Legal, medical, safety, financial and regulated content should use stricter terminology controls because the cost of semantic drift is higher.

Require stronger evidence, qualified review, traceability and targeted QA for high-risk concepts. Do not assume one process fits every term.

Risk-based terminology management directs effort where inconsistency could cause real harm.

Terminology in low-risk translation

Internal messages, exploratory content and low-stakes drafts may not justify heavy terminology governance.

Use a lightweight glossary or no glossary at all if the cost exceeds the benefit. The terminology architecture is a tool, not a ritual.

Professional judgement includes knowing when not to build infrastructure.

Terminology and literary translation

Literature usually resists rigid terminology control, but recurring fictional institutions, magical objects, place names, invented technologies and character epithets can behave like terms.

A series bible or name glossary helps maintain continuity across chapters and volumes. At the same time, ordinary repeated words may carry literary variation that should not be standardised mechanically.

Terminology in literature protects world identity, not bureaucratic consistency.

Terminology and subtitle translation

Subtitles face space and timing constraints. Long official terms may need approved short forms. Series, games and franchises also contain recurring names that audiences expect to remain stable.

The terminology resource can store short subtitle variants separately from full documentation terms while keeping both attached to one concept.

Controlled variation again solves the tension between consistency and format.

Terminology and interpreting

Interpreters cannot consult a large termbase while every sentence is unfolding. Preparation matters.

Build event-specific glossaries of names, acronyms, technical terms and likely concepts. Include pronunciation and short definitions. Prioritise retrieval speed over rich database fields.

The same concept-oriented architecture can produce a lighter runtime resource for live work.

Terminology and voice

Controlled terms should not flatten every sentence into institutional prose. Voice lives around the fixed points.

A literary narrator, friendly help-centre writer and formal legal drafter can all use the same product term while surrounding it with very different language.

Terminology preserves identity. Style preserves voice. Good translation needs both.

Terminology and semantic drift

Semantic drift occurs when repeated translation decisions slowly change the concept. One translator chooses a broader term, another paraphrases, a reviewer simplifies, and after several versions the target means something else.

A well-defined termbase acts as an anchor. The definition remains stable even if surface forms evolve.

This is one reason concept records are more durable than bilingual word pairs.

Terminology and semantic compression

Specialist terms compress large definitions into short labels. That compression is efficient only when the reader shares the concept.

When target readers do not know the term, translation may require a first-use explanation, parenthetical gloss or plain-language equivalent.

Terminology management should therefore consider reader knowledge, not only translator preference.

Terminology and ambiguity

Ambiguity can exist in the source term, the target term or both. A good record documents known ambiguities and gives disambiguating context.

If one target term could be misread dangerously, choose a more explicit form even when it is slightly longer. In interface constraints, consider redesign rather than forcing ambiguity into a tiny label.

Terminological elegance is secondary to correct concept recovery.

Terminology and neologisms

Fast-moving fields create new concepts faster than authoritative target terminology appears. AI, biotechnology, climate technology and digital culture frequently produce this problem.

Before coining a translation, search target-domain usage, professional communities and official sources. If no stable term exists, record the proposed form as provisional and monitor usage.

Coinage is sometimes necessary. The terminology system should make that uncertainty visible.

Terminology and borrowed words

Target communities may borrow the source-language term directly, adapt its spelling or create a native equivalent. None of these strategies is automatically superior.

Research actual domain usage and reader expectations. A native-looking translation may sound artificial if professionals overwhelmingly use the borrowing. A borrowed term may be inappropriate in public-facing material if readers do not know it.

Terminology decisions are sociolinguistic as well as semantic.

Terminology and transliteration

Names and technical terms from another script may require transliteration rather than translation. Record the transliteration system and official preferred forms.

Multiple romanisation systems can create duplicate-looking entities and inconsistent search results. A controlled name and term resource reduces this fragmentation.

Where an established translated term exists, distinguish it from transliteration so users know which operation is being performed.

Terminology and compound terms

Complex terms may contain smaller terms but should not always be translated compositionally. The established target expression can differ from the sum of its parts.

Record multiword terms as units when they designate one concept. Add component relationships only when useful.

This prevents translators from building a technically plausible but nonstandard phrase from individually approved words.

Terminology and phraseology

Some professional language depends on recurring phrases rather than isolated terms: warnings, procedural formulas, legal expressions and institutional wording.

A termbase may not be the only resource. Translation memory, phrasebase or style guide can store larger units. The architecture should place each recurring language pattern in the resource best suited to control it.

Do not force every language problem into terminology simply because terminology tools are available.

Terminology and definitions for AI

Definitions become especially valuable when AI systems must choose between homonymous records. A model can compare source context with concept definitions far better than with bare bilingual pairs.

Include concise, discriminating definitions rather than encyclopaedic essays. The goal is decision support.

Structured terminology designed for humans can become structured grounding for machines.

Terminology and retrieval precision

If retrieval returns too many records for a common source string, filter by domain, product, locale or version. If it returns too few, add synonyms and source variants.

The same precision-recall trade-off that appears in search appears in terminology retrieval. Too broad and translators see noise. Too narrow and they miss guidance.

Good metadata makes the balance controllable.

Terminology and reviewer behaviour

Reviewers should treat approved terminology as shared project policy, not as one translator’s preference. If they disagree, they should propose a terminology change with evidence.

This separates sentence-level editing from system-level decisions. It also prevents the same term from being changed back and forth across files.

Terminology governance reduces review churn.

Terminology and client review

Clients often provide terminology feedback late because they recognise product language better than translators do. Capture those decisions immediately.

Ask whether the change is a preference, a mandatory brand decision, a conceptual correction or a locale requirement. Update the record accordingly.

A good termbase converts client feedback into reusable knowledge.

Terminology and search-and-replace risk

When a preferred term changes, global search-and-replace can damage unrelated senses, code, quotations and historical content.

Use concept-aware or context-filtered migration where possible. Review high-risk replacements. Keep backups and version control.

String equality is not concept identity.

Terminology and quality thresholds

Not every terminology violation has equal severity. A stylistic synonym in a low-risk brochure may be minor. A wrong medical device component, legal defined term or dosage-related concept may be critical.

Review systems should prioritise by consequence, not merely count errors.

Risk classification helps teams spend attention intelligently.

Terminology and multilingual equity

Large organisations sometimes develop terminology deeply for one major language and leave smaller languages with ad hoc equivalents. This creates unequal quality and extra reviewer burden.

A multilingual concept model allows all languages to attach to the same conceptual spine while preserving locale-specific terms.

The concept record becomes shared infrastructure rather than a source-language hierarchy.

Terminology and pivot-language risk

Some multilingual workflows translate through an intermediate language. Terminology can drift when the pivot language collapses distinctions present in the original.

Where possible, link target terms to the underlying concept rather than only to the pivot string. Definitions and concept IDs help preserve distinctions across the chain.

Concept orientation protects multilingual architecture from source-language accidents.

Terminology and back-translation

Back-translation can reveal whether a controlled target term still points toward the intended concept, but it should not be treated as a perfect test. Languages divide conceptual space differently.

Use definitions and target-domain evidence alongside back-translation. The question is conceptual fidelity, not whether the back-translated string matches the original exactly.

Terminology evaluation remains concept-centred.

Terminology and human memory

Experienced translators remember many project terms, but human memory is not a governance system. People forget, teams change and versions diverge.

Externalising terminology decisions reduces cognitive load. Translators can spend attention on new semantic problems rather than re-deciding old ones.

The termbase is organisational memory for concepts.

Terminology and translation speed

Terminology work can look slow because research happens before drafting. Across a project, it saves time by preventing repeated lookup, reviewer disagreement and rework.

The speed gain is largest when terms recur frequently or when many translators share the same content.

Fast translation is often the result of decisions made once and reused responsibly.

Terminology and translation quality

Terminology quality is one dimension of translation quality, not the whole of it. A translation can use every approved term and still fail through wrong logic, tone, grammar, omission or context.

Conversely, a fluent accurate translation can still create serious inconsistency if it changes controlled terms.

The terminology system belongs inside a wider quality architecture.

Terminology and the translation unit

The previous architecture node asks how much text must move together. Terminology adds a cross-segment constraint. A term may recur across hundreds of units and require one conceptual identity.

This means local translation decisions are not always local. The translator may need document-level or project-level terminology memory while working on one sentence.

Translation units manage context boundaries. Termbases manage concept continuity across those boundaries.

Terminology and equivalence

The equivalence node explains why languages do not match one-to-one. Terminology turns that reality into controlled project decisions.

A source concept may require a target phrase rather than one word. A target community may use a borrowing. A legal concept may require explanation because no institutional equivalent exists.

Terminology management does not eliminate non-equivalence. It documents how the project handles it.

Terminology and context

The context node explains why meaning changes with co-text, document, situation, culture and interface. Terminology records use that context to decide whether a source string activates a controlled concept.

A term match without context can be wrong. A context model without terminology can still drift across repeated concepts.

The two architecture layers reinforce each other.

Terminology and source analysis

Source analysis identifies domain, audience, repeated terms, ambiguity and document purpose. That work supplies the raw material for terminology extraction.

If the translator skips source analysis, terminology decisions are more likely to be string-based rather than concept-based.

The architecture therefore remains cumulative: read first, establish equivalence boundaries, model context, choose translation units, then stabilise recurring concepts.

A twenty-question terminology audit

1. Does each record represent one concept? 2. Is the concept defined? 3. Is the domain explicit? 4. Is the target locale explicit where needed? 5. Is there one preferred term? 6. Are admitted synonyms classified? 7. Are deprecated forms recorded? 8. Are forbidden forms recorded? 9. Is the evidence traceable? 10. Are names separated from concepts? 11. Are abbreviations controlled? 12. Are grammatical forms usable? 13. Are homonymous source strings separated? 14. Are cross-project conflicts resolved? 15. Can translators access the termbase during work? 16. Can QA detect important violations? 17. Is there an owner? 18. Is version history preserved? 19. Are high-risk terms reviewed by qualified experts? 20. Can the data be exported or migrated?

If several answers are no, the terminology resource may look complete while remaining operationally weak.

A ten-question translator check before accepting a term

1. Which concept is active here? 2. Is the termbase record for that concept? 3. Does the project require the preferred term? 4. Does the target form fit grammar? 5. Is the locale correct? 6. Is this a defined legal or technical term? 7. Is the source using a different sense? 8. Does an exception note apply? 9. Has the term recently changed? 10. If I disagree, should I query the record rather than silently override it?

This check takes seconds once it becomes habit.

A ten-question reviewer check

1. Are all high-risk terms correct? 2. Are preferred terms consistently used? 3. Are forbidden forms absent? 4. Are different concepts being collapsed? 5. Are identical concepts being renamed unnecessarily? 6. Are terms grammatically integrated? 7. Has machine translation over-applied a glossary entry? 8. Has a translator paraphrased away a required term? 9. Are legacy terms appearing in new content? 10. Do any repeated queries indicate a weak termbase record?

Review feedback should improve the terminology system as well as the current document.

The terminology architecture for a small project

Use one shared spreadsheet. One row per concept. Keep source term, target term, definition, context, status and source. Assign one person to approve changes. Review the twenty most important terms before translation starts. Add terms when genuine problems arise.

Run a final manual terminology search before delivery. This lightweight process captures most of the benefit without enterprise overhead.

Discipline matters more than software.

The terminology architecture for a large programme

Use a managed termbase with concept IDs, multilingual terms, definitions, domains, status, sources, ownership, versioning and access control. Integrate it with CAT tools, machine translation, AI systems and QA. Establish proposal and approval workflows. Track unresolved terms and terminology debt.

Provide role-based views: translators need fast suggestions; terminologists need full metadata; product teams may need naming dashboards; AI systems need machine-readable records.

Complexity is justified when scale makes inconsistency expensive.

The terminology architecture for a school or learner

Use a concept notebook. One page or card per important concept. Write a simple definition, source term, target term, example sentence, related term and common confusion.

Organise by subject rather than only alphabetically. Review by concept families. Translate short passages using the resource and note when one source word needs different target words.

This makes terminology a bridge between subject knowledge, vocabulary and translation.

Frequently asked question: is a glossary the same as a termbase?

Not necessarily. “Glossary” is often used broadly for a list of terms and meanings or translations. A termbase is usually more structured and can hold concept-level multilingual data, status and metadata.

A simple glossary can still be excellent if it serves the project. The important distinction is whether the resource merely lists words or actually supports controlled terminology decisions.

Frequently asked question: should every repeated word become a term?

No. Frequency is only one signal. Ordinary words often need flexible translation. Add a term when conceptual consistency has value.

Overloaded termbases create noise and can make target language unnatural.

Frequently asked question: should one term always have one translation?

If one concept has one preferred target designation in the same domain and locale, usually yes for controlled content. But the same source string can represent different concepts, and one concept can have different approved designations by audience or format.

The rule is concept consistency, not string rigidity.

Frequently asked question: can AI build a termbase?

AI can extract candidate terms, suggest definitions, generate possible equivalents and organise evidence. It should not be the sole authority for high-stakes approval because it can invent plausible terminology or miss domain distinctions.

Use AI to accelerate research and structuring. Use authoritative evidence and qualified review to approve terms.

Frequently asked question: can a termbase improve machine translation?

Yes. Many translation platforms can pass glossary constraints to machine-translation engines. This often improves consistency for product, technical and branded terminology.

It does not remove the need for review. Engines can integrate a required term awkwardly or apply it to the wrong sense.

Frequently asked question: what if experts disagree?

Return to the concept and evidence. Ask whether the disagreement concerns meaning, locale, register, institution, product policy or personal preference.

Sometimes both forms are valid under different conditions. Encode the distinction. If one decision must be made, use the project’s defined authority and document the rationale.

Frequently asked question: when should a term be deprecated?

Deprecate when the form remains historically recognisable but should not be used in new content. Reasons can include rebranding, regulatory change, conceptual correction or modernised professional usage.

Keep the old form searchable so legacy translations can be interpreted and QA can detect it.

Frequently asked question: what belongs in the definition?

Enough information to distinguish the concept from nearby concepts. State what kind of thing it is and the characteristic that makes it different. Avoid circular wording.

For complex domains, link to authoritative fuller definitions rather than turning the termbase into an encyclopedia.

Frequently asked question: should translators be allowed to override terms?

They should be able to query or challenge a term when evidence shows the record is wrong or inapplicable. Silent override creates inconsistency. Blind obedience preserves errors.

The best workflow makes disagreement visible and resolvable.

The deeper idea: terminology is a memory of distinctions

A termbase is often described as a database of words. That description is too small.

The real asset is the distinction the organisation decided to preserve. This concept is not that concept. This label is official. This synonym is acceptable only in patient-facing text. This old name belongs to version 2. This target form is legally established. This abbreviation remains untranslated. This source string has two senses and must not be collapsed.

Terminology stores these distinctions outside individual memory so they survive across people, tools and time.

The governing rule

Do not ask first, “What word should we use?”

Ask, “Which concept are we naming, which distinctions must survive, who will use the term, and what evidence shows that this target designation belongs to that concept in this domain and locale?”

Then record the decision in a form that translators, reviewers and machines can reuse.

Where this node leads next

The terminology system sits at the point where repeated concepts become controlled multilingual decisions. From here, the Master Art of Translation architecture can move into lexical choice and collocation, grammar and syntax, reference and cohesion, tone and register, idioms and figurative language, translation memory, human translation workflow, machine translation, AI translation, post-editing, quality assurance, localisation, transcreation, interpreting and specialist domains.

The sequence remains cumulative. Source analysis tells us what the text is doing. Equivalence tells us why forms do not map one-to-one. Context tells us which meaning is active. Translation units tell us what must move together. Terminology tells us which recurring concepts must keep a controlled identity across the entire system.

When terminology is weak, the same concept acquires different names and different concepts collapse into the same name. When terminology is strong, language can vary where variation is useful while the important distinctions stay fixed.

---

Continue through the Master Art of Translation spine

Architecture root: Master Art of Translation | The Complete System for Moving Meaning Between Languages

Source analysis: Master Art of Translation | How to Read a Source Text Before You Translate It

Equivalence: Master Art of Translation | How Equivalence Works When Languages Do Not Match One-to-One

Context: Master Art of Translation | The Context Stack

Translation units: Master Art of Translation | The Translation Unit

Adjacent language systems

Vocabulary: Vocabulary Learning Hub — use this route when the underlying problem is vocabulary depth, collocation, morphology, word families or lexical retrieval rather than terminology governance.

English: How English Works — use this route when the underlying problem is English grammar, syntax, discourse, register or meaning construction rather than multilingual terminology control.

Professional references and further reading

European Commission — Terminology: an overview of concept-oriented terminology work and IATE.

ISO 12616-1:2021: terminology work in support of multilingual communication and fundamentals of translation-oriented terminography.

ISO 30042:2019: Management of terminology resources — TermBase eXchange (TBX).

ISO/TR 24633-1:2026: companion overview to the TBX framework.

Phrase — Term Bases Overview: practical terminology management inside a modern translation platform.

memoQ Qterm: terminology management and organisational termbase workflow.