CONTACT US

Glossary Building: From Project List to Career Termbase

How translators build the terminology their quality depends on, and keep it.

Introduction: Why Glossaries Matter

Fifteen years ago, a reviewer for an Indian NGO flagged a term in a 40,000-word file: she wanted “redevabilité”, not “responsabilité”. The English term was “accountability to affected populations” (AAP). In another context, she might well have been right. But for AAP, “responsabilité à l’égard des populations touchées” was the established UN terminology the report had to follow, and I had recorded it in my ECOSOC glossary. Without that glossary, I might have accepted a perfectly plausible correction that was wrong for this context.

Terminology is often where specialized translation is won or lost, and it is certainly where much of the time goes. Research cited by the Conference of Translation Services of European States (COTSOES) suggests that experienced translators spend 20 to 25 percent of their working time on terminology research. For inexperienced translators, the figure climbs to 40 to 60 percent. Every hour of research either disappears with the project or gets recorded once and keeps paying off for years. The difference is a glossary.

From Word List to Concept Entry: What a Glossary Actually Is

A two-column word list is not a glossary. It answers one question (“what did I use last time?”); it says nothing about questions that actually cost you time, like which of two competing renderings the client prefers, and whether a term changes meaning across contexts.

A real glossary is built on concept orientation: each entry represents one concept, one unit of meaning, and everything you know about that concept lives together in the entry. If English monitoring and surveillance name two different activities in your domain (in conservation and public health they usually do, and French keeps them apart as “suivi” and “surveillance”), they get two entries. If a single French word covers several concepts (“réserve” as a nature reserve, a legal reservation or a financial buffer), each meaning gets its own entry with its own domain label. Synonyms travel together: preferred, accepted and rejected variants of the same concept belong in one entry, each carrying a status, a decision rather than just a meaning.

It’s the difference between a list that describes words and a resource that answers “which one do I use, for this client, in this context?”

Terminology as a field began with Eugen Wüster, an Austrian engineer who dreamed of perfect univocity: one concept, one term, ambiguity engineered out of technical language. Standardization bodies still run on that dream, and for controlled vocabularies it works. But translators live in texts, and texts refused to cooperate.

From the 1990s, researchers rebuilt the field around how terms actually behave:

  • Rita Temmerman showed that variation and metaphor are normal features of scientific language, not defects.
  • Marie-Claude L’Homme’s team put the focus where translators feel it daily: on the verbs and collocations that surround a term.
  • Pamela Faber’s frame-based school built EcoLexicon, the University of Granada’s environmental knowledge base, a resource worth bookmarking if you translate in that sector.

Term variation is not your failure to find “the right answer”; it is a documented property of specialized language. And a term without its surrounding phraseology is half a term, so capture a context sentence while you have one in front of you.

Working translators, it turns out, agree. In a survey of professional translators reported by Stella Tagnin of the University of São Paulo, whose students spent years building corpus-driven English-Portuguese glossaries, only 8 percent were satisfied with a bare equivalent. The most popular single answer, at 23 percent, was “all of the above”: equivalents plus definitions, authentic examples in both languages and cross-references.

Translator’s Tip: Usually, when you hesitate between two French renderings, you are not missing “the” term; you are navigating real variation. Put both in the entry, mark one preferred and note what triggers the other. You will need the reasoning, not just the winner.

The Project Glossary: Preparing Terminology Before You Translate

An NGO sends you a 30,000-word report to translate into French. Terminology preparation starts before segment one, but it does not mean compiling an exhaustive termbase. On a one-off project, that is effort in the wrong place. TerminOrgs’ starter guide says the same thing: start small and risk-based. In practice, a critical list of roughly 30 to 50 high-risk terms that grows as the project reveals more.

Which terms make the list

Selection is risk-based first, frequency-based second. A term earns its place on the list when getting it wrong, or rendering it inconsistently, would be expensive:

  • Client-critical names: the organization’s programs, initiatives, units and job titles. No public term bank knows these; only the client’s own usage does.
  • Regulatory and legal designations: treaty names, protected-area categories, notifiable diseases. These have official renderings you are not free to improvise.
  • Terms with several accepted equivalents: English stakeholder can become “partie prenante”, “acteur” or “intervenant” depending on house style; community-based wavers between “communautaire” and “à assise communautaire”. The risk here is inconsistency, not ignorance.
  • Terms of art disguised as ordinary words: control in disease control, surveillance, evidence, capacity building (“renforcement des capacités”). These are where a fluent but unprepared translation quietly goes wrong.
  • High-frequency repeats: even a low-risk term used eighty times must be locked to one rendering.

Everything generic (ex. report, framework, meeting) stays off the list. A useful filter question for each candidate: if two translators rendered this differently, would the client notice or care? If not, drop it.

Harvesting the candidates in the first place takes less time than it sounds. For a report-length text, one targeted read does it: the title, the executive summary, headings, captions and whatever phrasing the organization visibly repeats. For anything longer, run a quick extraction pass over the source file and filter the output (how those tools think is covered below).

Where do the French equivalents come from?

In a strict order. The client’s own published French corpus comes first: previous reports, the French pages of their website, their partners’ usage. Consistency with what the organization has already published is usually worth more than abstract correctness, and program names simply cannot be sourced anywhere else. And wherever an official designation applies: the instruments themselves, in their authentic French. EUR-Lex for EU law, the conventions and resolutions in their own French versions. Only then do the public term banks enter, and choosing among them is a skill in itself. The main ones for institutional English-French work:

  • IATE, the EU’s inter-institutional termbase: authoritative for EU policy and legislation; check each entry’s reliability rating and domain before trusting it.
  • UNTERM, the United Nations terminology database: the reference for UN body names, resolution language and SDG vocabulary.
  • TERMIUM Plus, the Canadian government’s term bank: enormous and excellent for scientific and technical domains, but it documents Canadian usage.
  • FranceTerme, France’s official terminology database: terms published in the Journal officiel, binding for French public administration and the reference for France-targeted texts.
  • Le Grand dictionnaire terminologique, from the Office québécois de la langue française: rich definitions and usage notes, oriented to Quebec.
  • Domain portals such as the FAO terminology portal for food security or fisheries, and WHO’s glossaries for public health.

Client usage beats every term bank, except where an official designation binds the client too. After that, match the bank to the text’s institutional ecosystem: a UN-adjacent report calls for UNTERM and the relevant agency portal, an EU-funded project for IATE, a Canadian audience for TERMIUM. Variety of French is messier, and it cuts across both: a Quebec-preferred term can be exactly wrong for a readership in Dakar or Geneva. When the banks disagree, and they will, decide by ecosystem and audience, then record what you chose and why.

What an entry needs

Full terminological records are overkill for a single assignment; a bare word pair is too little. The fields that earn their keep: the English term, the French term, the source of the French term (“UNTERM”, “IATE, 4-star”, “client annual report 2024, p. 12”), a status (candidate, self-validated, client-approved), a short context or note where the term is ambiguous and any forbidden variants the client has rejected. Definitions can usually stay as links. Keep the format simple (a clean spreadsheet is fine) but keep the columns disciplined, so the file imports into a termbase without rework.

Translator’s Tip: The source column is the thirty seconds that save an argument later. I lost an afternoon to a missing one once.

Validation and Enforcement: The Client Loop and the CAT Tool

A glossary nobody agreed to is an opinion. The simplest thing you can do is get the contested terms approved before you translate, not after, a principle the American Translators Association’s guidance on term validation states plainly: ideally, terms should be validated before a project starts.

The mechanics are simple. From your 30-to-50-term list, pull out only the genuinely contested or client-critical items, usually 10 to 30 rows. Send them as one batched sheet with your proposed French term, the authority backing it and the credible alternative where one exists. Ask closed, decision-ready questions: “We propose ‘partie prenante’, used in UNTERM and your 2023 report; ‘acteur’ also appears in IATE. Confirm ‘partie prenante’?” is answerable in five seconds. “How would you like us to handle stakeholder?” is not.

Ask at kickoff who on the client side owns terminology decisions; no owner means no sign-off, and no sign-off means the debate happens at review stage, at maximum cost. Record every answer in the glossary with its status set to client-approved and the email date as its source. Approved terms are then locked: they do not change mid-project without a new client decision.

And because real clients sometimes go quiet, tie the query sheet to a response deadline your schedule can absorb. If no answer comes, proceed with your best-attested rendering at self-validated status, flag the pending items in your delivery note and never promote them to client-approved on your own authority.

Import the glossary into your CAT tool as a project termbase (MultiTerm for Trados Studio, the native termbase in memoQ, CSV or TBX import almost everywhere; TBX is the standard exchange format for terminology data, and there is more on it below) and terminology appears in every segment as you work.

Then make the tool police it: Trados Studio’s Terminology Verifier and memoQ’s QA settings both check that termbase terms were actually used in the target, and both can flag forbidden terms, so the variants your client rejected trigger warnings if they sneak back in. Enable fuzzy or stemmed matching; French morphology (plurals, elision, agreement) will otherwise defeat the checker in both directions. Before delivery, run the terminology check across the full file set and treat every hit as either a fix or a documented false positive.

Two habits, while you translate. New terms discovered mid-flight go straight in with candidate status (capturing a term at the moment of discovery costs seconds; reconstructing it later costs hours), and contested ones join the next query batch, sent early enough for answers to arrive while applying them is still cheap. And if a locked term must change because the client changed their mind, change the termbase entry first, then use the QA check to retrofit every earlier segment. Never rely on memory.

When the project ends, spend twenty minutes on consolidation: promote validated candidates, delete noise and merge the glossary into a master termbase for that client. Client-specific renderings must not leak into other clients’ work, so keep client masters separate from your reusable domain glossaries. That twenty minutes is why the next project starts preloaded, which is where the compounding begins. (For what happens when the tools drive the translator instead of the other way around, see How the CAT Tool Is Costing You Credibility.)

Translator’s Tip: Log every client decision with its date, its decider and the verbatim wording (“approved by C. Martin, email of May 12”). Approvals outlive projects, but they also outlive the people who made them.

Building Your Own Corpus

The corpus is not a last resort. It is the reference, built before you need it: the client’s own material, and the documents their sector runs on. Translators borrow the method from corpus linguistics. In domains where the term banks thin out, like carbon-credit methodologies or a new EU due-diligence regime, its absence just costs more.

Two kinds of corpus matter here. Parallel corpora pair source texts with their human translations, aligned sentence by sentence, and are mines for verified equivalents. The classics are public: Europarl aligns European Parliament proceedings across 21 languages; the OPUS collection gathers over a thousand parallel corpora, from software documentation to legislation; JRC-Acquis covers the body of EU law in 22 languages and is a staple of legal terminology work. (Every parallel lookup tool you already use, from concordancers to Linguee, is built on this kind of data.) Comparable corpora are the workaround when no translations exist: independent texts in each language, collected on the same subject, which let you observe how each language community talks about the domain natively.

The fastest way to build one is the seed-word method, introduced by Marco Baroni and Silvia Bernardini’s BootCaT tool in 2004 and now built into Sketch Engine as its corpus-from-the-web feature. You supply 10 to 20 seed terms pulled from the client’s documents. It queries the web, downloads the pages and strips the boilerplate. You get a tagged corpus, usually in under ten minutes. Run a keyword comparison against a general reference corpus and the domain’s vocabulary rises to the top; open a concordance and you can watch any candidate term behave in dozens of real sentences.

I do my own corpus work in LogiTerm Pro, which searches extensive document collections I have assembled myself, full text and aligned bitexts side by side. Sketch Engine is the better choice if you need to build a corpus from the Web. Pick for the shape of your sources.

Do it twice: once in English, once in French. The French corpus is the underrated half, because it answers the questions no term bank can. Which verb does the French professional literature actually use with “aire protégée privée”? Is your candidate equivalent attested at all outside translated texts? Half an hour of building beats an afternoon of guessing.

Translator’s Tip: Before a corpus can answer questions, it just has to exist. Build the target-language corpus at project start, not when you hit the first problem.

How Term Extraction Tools Think

Once you have a corpus (or just a long source text), extraction tools can propose the term candidates for you. Understanding roughly how they think makes you a better user of them, because you learn what they miss.

Why frequency fails

Ranking words by frequency fails twice. Common grammar words dominate any frequency list. And specialized terms are usually multiword units that nest inside one another: in a medical corpus, the string soft contact occurs frequently, but only because it lives inside soft contact lens; frequency alone cannot tell a real term from a fragment of one.

The C-value

Modern extractors therefore combine two stages. A linguistic filter keeps only plausible term shapes: sequences like noun + noun or adjective + noun, sometimes with a preposition (which is how patterns like coloboma of retina survive), minus a stop-list of general words. A statistical score then ranks the survivors by termhood, a measure of how term-like each candidate is. The classic score, still the reference point for the whole field, is the C-value, published by Katerina Frantzi, Sophia Ananiadou, and Hideki Mima in 2000:

C-value(a) = log2|a| × f(a)                                   if a is not nested

C-value(a) = log2|a| × ( f(a) − (1/P(Ta)) × Σ f(b) )          if a occurs inside longer candidates

Here a is the candidate string, |a| its length in words, f(a) its frequency in the corpus, Ta the set of longer candidates containing a, P(Ta) how many there are and b each of those longer strings. Longer candidates get a boost (multiword noun phrases are more likely to be genuine terms), and a candidate that mostly appears inside longer terms gets discounted by the average frequency of its parents.

In the authors’ corpus of eye-pathology records, basal cell carcinoma occurs nearly a thousand times and also nests inside longer terms like adenoid cystic basal cell carcinoma; the formula weighs both facts and still ranks it near the top, while relegating fragments like soft contact.

A refinement called the NC-value re-ranks the list using context words, on the observation that real terms keep company with characteristic verbs, nouns and adjectives. You can try the method free through TerMine, the University of Manchester text-mining center’s demo service. One quirk worth knowing: because log2 of 1 is 0, the original formula ignores single-word terms, which is why later adaptations adjust the length factor.

Tools you can actually use

For daily practice, the algorithm arrives pre-packaged. OneClick Terms (from the Sketch Engine team, whose freelancer plans start around 12 euros per month) extracts monolingual or bilingual term candidates from uploaded documents and exports TBX, CSV or Excel. memoQ’s translator edition has term extraction built in, feeding candidates straight into a termbase with concordance context. Trados Studio users have the free projectTermExtract plugin, plus the indispensable free Glossary Converter for moving data between MultiTerm, Excel and TBX. On a zero budget, AntConc‘s keyword and n-gram tools produce a serviceable candidate list from any text collection.

And your richest source may already be on your disk: extraction run over your own translation memories harvests term pairs from work you have already validated, one client or domain at a time.

There is always noise (false candidates) and silence (real terms missed), and you should tune toward noise, because scanning and rejecting candidates is fast while discovering silence mid-project is expensive. And extraction produces candidates, never terms. The human validation pass is the work. Budget for it.

Translator’s Tip: Set a frequency threshold of 2 or 3 when extracting from your own TMs. It filters typos and one-off improvisations, and what remains is very close to a validated bilingual glossary of past decisions.

The Master Termbase: Turning Projects into an Asset

Project glossaries are sprints. The career asset is the master termbase: one concept-oriented database you filter per project. The freelance-realistic architecture is a single master with three disciplined picklist fields (client, domain, status), so any project view is one filter away, plus separate termbases only for genuinely confidential client material and for domain glossaries you might one day share.

My own split, since the question comes up: corpus in LogiTerm, termbase in MultiTerm. The reason is enforcement. A termbase is worth what your QA can check, so it has to live where the terminology verifier runs. Mine runs in Trados Studio, so MultiTerm wins by default. In memoQ, the native termbase wins for exactly the same reason. The criterion is not which tool is better, it is which one your CAT tool can police.

The field set worth maintaining is short. Per concept: a definition or short gloss, a domain from a controlled list and the client tag. Per term: the term itself, part of speech, a status (preferred, admitted, deprecated, forbidden), one authentic context sentence, the source and a free note. Two more fields, entry ID and modification date, your tools maintain for you. Everything beyond that tends to die unfilled; every field you add must have a defined purpose and a controlled format, or it becomes noise that makes maintenance harder.

Maintenance itself follows a rhythm: capture continuously (context and source are unrecoverable later), then a quarterly duplicate-and-gaps pass, then an annual audit to retire obsolete terms and apply client renamings globally. When two entries turn out to describe the same concept, merge them and keep the better-documented one. The rule that prevents most future cleanup is the same concept orientation from the beginning of this article: one concept per entry, synonyms inside the entry with statuses, polysemes split. It is precisely the concept-shaped data that survives tool migrations and feeds AI systems without poisoning them.

Standards exist so that this asset outlives any one tool, and they are worth knowing by name. ISO 704:2022 codifies the principles of terminology work. ISO 12620 (now split into parts 1 and 2, 2022) governs data categories, the standardized field names cataloged at DatCatInfo. ISO 26162 specifies termbase design, software and content in three parts.

ISO 30042:2019 defines TBX, TermBase eXchange, the XML interchange format, whose practical profile for translators, TBX-Basic, is maintained by the industry consortium TerminOrgs (Version 4 appeared in November 2025).

A TBX file mirrors concept orientation structurally: each entry has a concept level, a language level and a term level. A minimal valid TBX-Basic entry looks like this:

The definition belongs to the concept, not to any one language. The two French renderings sit inside the same entry, and the status decides which one you use. The context sentence attaches to the term it illustrates, not to the entry as a whole. In the file itself those three levels are called conceptEntry, langSec and termSec.

In day-to-day freelance practice, spreadsheets remain the lingua franca: every tool imports CSV and Excel, and clients can actually open them. TBX earns its place for migration and archival, because it preserves the concept structure that flat tables strain to represent.

The pragmatic pattern: keep the master in your termbase (or a disciplined spreadsheet), export XLSX for clients and bulk edits, export TBX for tool changes and archives, and test any migration with a ten-entry round trip before trusting it with ten thousand.

Translator’s Tip: Even in a spreadsheet, keep a concept ID column and let synonyms share it. It costs nothing today and it makes your data mergeable and migratable, whatever you plug it into later.

Glossaries in the Age of AI

Large language models don’t make glossaries obsolete. They punish sloppy ones.

Start with extraction

LLMs are genuinely useful as recall boosters: given a text, they propose candidate terms quickly, including some that statistical methods rank too low to surface. But the research record through 2025 (surveyed in Antonio San Martín’s work on AI-assisted terminology) finds performance inconsistent across languages, domains and even runs of the same model, with documented cases of hallucinated terms and confused regional variants. The structural weakness is auditability: a C-value list is attested by construction (every candidate exists in your corpus), while an LLM list is not, unless you check. Use both, and check everything the model gives you.

Equivalents

Equivalents show the same problem. Testing a custom GPT-4 against the Croatian national termbase Struna, Bruno Nahod found that 60 percent of its proposed term equivalents matched the expert version exactly. Counting everything a terminologist could work from rather than accept outright, he put the usable share near 90 percent. It is a small study, around twenty terms, so read it as a shape rather than a measurement. The shape is this: usable is not correct, and the distance between them is fluent enough to pass.

Machine Output

Terminology is also what keeps machine output in line. The 2025 edition of the WMT terminology shared task, the field’s benchmark, found that supplying correct terminology improves both term accuracy and overall quality, and that feeding a noisy termbase measurably degrades it. Whatever automation sits in a workflow inherits the discipline of the terminology behind it, or the lack of it. Adherence is probabilistic either way, which is why the terminology QA pass in your CAT tool is not optional.eeding models a noisy termbase measurably degrades output, which quietly converts termbase hygiene from good practice into a hard technical requirement.

The International Federation of Translators’ 2025 position paper names lack of terminological consistency among machine translation’s persistent weaknesses and insists on the human “at the core” of the workflow.

Translator’s Tip: Let AI propose; never let it validate.

How the Pieces Fit Together

Build the corpus first: the client’s own material and the reference documents their sector runs on. Scope the risky terms, mine the corpus before any term bank, and let extraction surface what you would have missed. Choose among the banks by ecosystem and audience. Shape what you find into concept entries with statuses and sources. Validate with the client before translating, enforce mechanically in the CAT tool and consolidate afterward into a master that makes the next project faster. Then let the same validated data steer whatever automation the workflow includes.

None of this is improvised. It is a discipline with a literature and a curriculum, taught today in translation faculties from Geneva to São Paulo, increasingly with AI on the syllabus. For the freelancer, the practical canon is freely available: the TerminOrgs starter guides, the COTSOES recommendations, and Lynne Bowker and Jennifer Pearson’s Working with Specialized Language.

Conclusion

A glossary looks like a list of words. It is actually a record of decisions: which term, for which client, in which context, on whose authority. Built that way, entry by entry, it becomes the rare professional asset that appreciates with use: it makes each project faster than the last, turns quality from an aspiration into a checkable property and now doubles as the grounding data that keeps AI-assisted workflows honest. The tools will keep changing. The principle underneath them does not: in terminology, as in everything else we do, context is everything, and the translator who records context owns the asset that no machine can fabricate.

If you work in the environmental or humanitarian sector, you can see these principles applied on our resources page, in the Climate & Biodiversity EN-FR Glossary and the Humanitarian Coordination EN-FR Mini-Glossary: both were built with the workflow described here and are offered, as always, as companions and starting points rather than prescriptions.

References

  • Bowker, L. & Pearson, J. (2002). Working with Specialized Language: A Practical Guide to Using Corpora. Routledge. https://www.routledge.com/Working-with-Specialized-Language-A-Practical-Guide-to-Using-Corpora/Bowker-Pearson/p/book/9780415236997
  • Cabré, M. T. (2003). “Theories of terminology: Their description, prescription and explanation.” Terminology 9(2), 163-199. https://benjamins.com/catalog/term.9.2.03cab
  • COTSOES (2003). Recommendations for Terminology Work. https://cotsoes.org/sites/default/files/public_files/COTSOES_Recommendations_for_Terminology_Work.pdf
  • Faber, P. (ed.) (2012). A Cognitive Linguistics View of Terminology and Specialized Language. De Gruyter Mouton. EcoLexicon: https://ecolexicon.ugr.es/
  • FIT (2025). Position Paper on Machine Translation in the Age of AI. https://en.fit-ift.org/position-and-discussion-papers/
  • Frantzi, K., Ananiadou, S. & Mima, H. (2000). “Automatic recognition of multi-word terms: the C-value/NC-value method.” International Journal on Digital Libraries 3(2), 115-130. https://link.springer.com/article/10.1007/s007999900023
  • ISO 704:2022, Terminology work: Principles and methods. https://www.iso.org/standard/79077.html
  • ISO 30042:2019, Management of terminology resources: TermBase eXchange (TBX). https://www.iso.org/standard/62510.html
  • Kilgarriff, A. (2013). “Terminology finding, parallel corpora and bilingual word sketches in the Sketch Engine.” ASLIB Translating and the Computer 35. https://www.sketchengine.eu/wp-content/uploads/2015/05/Terminology_finding_2013.pdf
  • L’Homme, M.-C. (2020). Lexical Semantics for Terminology: An Introduction. John Benjamins. https://benjamins.com/catalog/tlrp.20
  • Nahod, B. (2024). “Can We Substitute Field Experts with Customized Large Language Model in Processing Specialized Languages? – A Case Study.” Proceedings of the XXI EURALEX International Congress, 725-739. http://euralex.org/wp-content/themes/euralex/proceedings/Euralex%202024/EURALEX2024_Pr_p725-739_Nahod.pdf.pdf
  • San Martín, A. (2025). “Toward Human-Centered AI-Assisted Terminology Work.” https://arxiv.org/abs/2512.18859
  • Tagnin, S. E. O. (2015). “Corpus-driven glossaries in translator training courses.” Oslo Studies in Language 7(1), 359-377. https://journals.uio.no/osla/article/view/1447
  • TerminOrgs. Terminology Starter Guide and TBX-Basic Version 4 (2025). https://www.terminorgs.net/
  • WMT 2025 Terminology Shared Task findings. https://www2.statmt.org/wmt25/terminology.html

Sources were verified at the time of writing; standards versions and tool features evolve, so check the linked originals for current details.