Glossary

Goal

The glossary fixes, for each term of a business domain, the agreed definition, so that two departments stop using the same word for two things or two words for one. Nothing in the artifacts they exchange reveals the collision: the requirement is written, reviewed, approved and built, then the divergence surfaces in acceptance testing or in production, as a number that will not reconcile or a rule that fires on the wrong population. At a health insurer, the sales department says "assuré" meaning the person who signs the contract and pays the premium, the benefits department says "assuré" meaning the person whose risk is covered. The two dashboards count a different population, both are correct, and the gap passes for an arithmetic error for months.

The glossary attacks that, by making a domain's vocabulary decidable: each term's definition is dated, carried by a named owner and reachable from wherever the term is used. BABOK gives the reason in one sentence: a term can mean different things to different people, and a list of established terms and definitions supplies the common language in which ideas are exchanged. HERMES, the Swiss Confederation's project management method, says the same for a public administration: its terminology guarantees a common language and a common understanding between the core organisation and the project organisation, between the project and the programme, between project management, application management and portfolio management.

The decision the glossary supports is the decision to write anything at all in the words of the domain: a requirement, a business rule, an interface contract, the definition of an indicator, a communication to the regulator. Each of those artifacts silently re-runs the vocabulary arbitration when no glossary exists, and each of them arbitrates it differently. The third benefit BABOK attributes to the technique is the one that carries it: the glossary simplifies the writing and the maintenance of all the business analysis information produced on top of it. It is rarely read for itself, it is the substrate the other artifacts stand on.

The deliverable is the set of entries: one preferred term per concept, its definition, its non-preferred synonyms, its label in every language in use, its owner, its status, the date of its last review and its source of authority. To those are added the rejected-term entries, which record the colliding labels that were split and route the reader to the disambiguated terms. They are the highest-value entries in a glossary and the ones most glossaries leave out. The glossary HERMES publishes applies that per-language label to a whole method: its tables carry every phase, scenario, module, outcome, task and role in German, French, Italian and English, with the number of the method element.

Usage

When to use it

• Start of an initiative: BABOK places the creation of the glossary in the early stages of the project, where it still eases the transfer of knowledge.
• A term used differently by two departments: the collision stays invisible in the artifacts they exchange.
• A contested consolidated indicator: two numbers diverge because two departments count two populations.
• Regulated domain: the term has a legal definition, and the entry carries the reference to the article.
• Bilingual or trilingual organisation: every concept must carry the label actually used in each language.
• Large requirements corpus: the vocabulary exceeds what a reviewer keeps in mind, and terminological consistency is checked against a list.
• Taking over a legacy domain: the vocabulary of the incumbent system has displaced the vocabulary of the business.
• Ahead of a concept model or a data dictionary: both consume a vocabulary that has already been settled.
• Newcomers arriving in the domain: onboarding an employee, a supplier, an external team.

When not to use it

• A business glossary already carried by a data catalogue: govern and enrich the tool in place, never open a parallel glossary.
• Ambiguity about a field's format or permitted values: that is the data dictionary (BABOK 10.12).

Description

The admission criterion: the discipline of refusal

BABOK gives four inclusion criteria, and a term enters the glossary as soon as any one of them holds: the term is specific to the domain, it carries several definitions, the definition given to it departs from its everyday sense or there is a reasonable risk of misunderstanding. Those four criteria are a filter, and their real function is exclusion. The commonest failure of a glossary is the four-hundred-entry glossary that defines words everybody knows and in which the eight entries that carry the meaning of the domain are buried and therefore never read. The missing term gets noticed and gets fixed; the superfluous entry gets read, tires the reader and closes the document. The admission criterion is published before the harvest, because the discipline that makes a glossary readable is the discipline of refusal.

BABOK adds four development rules, and the fourth is the one practice drops. Definitions are clear, concise and brief. Acronyms used in a definition are spelled out. Stakeholders can access the glossary easily and reliably. Editing the glossary is restricted to designated stakeholders. A wiki page open to everyone drifts: the content of a glossary is decided, and a decided artifact has a door. BABOK further recommends that the glossary be created early in the initiative, and it describes the designation of a point of contact responsible for its upkeep and its distribution.

The entry: what a glossary line carries

BABOK names three fields: the term, its definition, its common synonyms. Practice adds four, and each of them is defended by a standard or by a limitation BABOK itself states.

• Preferred term: the elected label, one per concept and per language. ISO 25964-1 calls it the preferred term, the W3C's SKOS vocabulary calls it skos:prefLabel. It is the word the initiative's artifacts use from now on.
• Definition: the statement that delimits the concept. Its form follows ISO 704 and ISO/IEC 11179-4.
• Non-preferred synonyms: the words people actually say, recorded as entry points that route back to the preferred term (skos:altLabel, the USE / UF relation of ISO 25964-1). Nothing is deleted: a synonym erased from the glossary goes on circulating in the organisation, now with no way in.
• Label per language: in a Swiss organisation, a concept carries a French label, a German one and often an Italian one. Each is a term in its own right, elected in its own language. The relation between them is equivalence, an inter-language relation that terminology keeps apart from synonymy.
• Owner: the person with the authority to say what the term means and to settle disputes about it. BABOK puts the absence of an owner at the head of its limitations, and a term with no owner is a term that will be neither updated nor arbitrated.
• Status and date of last review: proposed, approved or deprecated. The status tells the reader how much credit to give the entry, and the date makes staleness visible.
• Source of authority: the legal reference, the standard or the decision that fixes the definition. A definition backed by an article of law is not reopened. A definition with no source will be reopened at the first disagreement, then at every disagreement after it. The reference is the instrument that fixes the rule, not the page that publishes it: the amount of the Swiss health-insurance deductible is fixed by OAMal art. 103, and the federal office that circulates it is not its source of authority.

Two optional fields earn their place as soon as the domain has depth: related terms (skos:related) and the broader / narrower pair (skos:broader / skos:narrower), where a hierarchy exists between the concepts.

The relations to resolve

A glossary is built by resolving the relations among the candidate terms. This is where the difficulty of the technique sits, and it is the part a reader will not find in BABOK. The underlying model is ISO 1087's: a concept is a unit of knowledge abstracted from the objects of the world, a designation is a sign that represents it and a definition is the statement that delimits it. Term and concept are therefore two things of a different nature.

Synonymy is the case of several terms for one concept within one language. The restriction to a single language is constitutive, and that is how ISO 1087 defines the word. Synonymy resolves by election: one preferred term, the others recorded as non-preferred synonyms that point at it. None is lost, and the organisation's real vocabulary stays a way into the decided vocabulary.

Equivalence is the relation between the labels of one concept across languages, and it is a different relation from synonymy. ISO 704 does not conflate them: its clause 7.7, "Relations between designations and concepts", gives synonymy its own subclause (7.7.2) and equivalence its own (7.7.3). The distinction is most of what multilingualism does to a Swiss glossary. The German label of a concept is a term in its own right, elected in its own language, from the real usage of that language. The German column of the glossary is where that difference shows, in both directions: it carries German words where German has them and the French word itself where German-speaking Switzerland uses the French word.

A collision is the converse case of synonymy: one term for several concepts. Terminology distinguishes two species of it, and ISO 1087 defines both. Homonymy links identical forms whose concepts are unrelated, of different meaning and different origin: the bank you deposit money in and the bank of a river. Polysemy links one designation to several concepts whose senses are related, grown from a shared root that has diverged. ISO 704 handles the two in a single clause (7.7.5, "Polysemy and homonymy"), and that grouping already says what matters to a glossary: the resolution is the same in both cases.

In a business domain the dominant case is polysemy. "assuré" denotes the preneur d'assurance and the personne assurée: two senses of the same word, in the same domain, grown from the same root. "caisse" denotes the health insurer, the pension fund and the compensation fund: three institutions, one base sense declined three times. Neither is a homonym, and it is the kinship of the senses that makes the danger. A true homonym fools nobody: the two concepts have no business together and the context separates them unaided, so nobody at an insurer reads "bank" and pictures a river. A polyseme, by contrast, sails through a review without a reviewer blinking, because the two departments read the same sentence, each in their own sense, and the sentence stays true for both. The collision that costs money is polysemy, because it passes review.

Two operational consequences follow. The first is that the compromise definition is the temptation specific to polysemy: when two senses are neighbours, there is always a sentence broad enough to cover both, and it is that availability which makes it dangerous. Facing a true homonym, nobody would attempt the sentence that covers the money and the river. The second is for detection: a homonym is spotted on distribution alone, its contexts having nothing in common, whereas the contexts of a polyseme overlap heavily. That is why collision detection is run on a corpus segmented by department rather than on the text taken in bulk, even when a model does the detecting.

A collision resolves by splitting: two disambiguated preferred terms and the ambiguous label withdrawn from use. The point that practice almost always misses deserves its own statement. A colliding label cannot be carried as a synonym. A synonym entry asserts that this label denotes this concept; a colliding label denotes two, so recording it as an alias of one of them is a false assertion, which the glossary then propagates with the authority of an approved entry. The label is recorded as a rejected term, in an entry that points at the disambiguated terms and says why it is rejected.

Hierarchy, finally, is the case of a concept that subsumes another. It is recorded for two operational reasons. It lets a definition stand on its superordinate concept, the mechanism that makes the rules for writing a definition something other than a matter of style. And it makes a rule's scope checkable: does the rule bear on the superordinate concept or on only one of its subordinates, a question that has an answer only if the hierarchy is written down.

Three of these relations resolve within a single language: synonymy, collision and hierarchy. The graph carries all three. Equivalence links the languages to each other, and its place is the per-language label column.

Writing a definition worth having

BABOK says "clear, concise and brief", which is an intention. ISO 704 gives the method. The standard holds that the intensional definition is preferable to all others and is used whenever it is possible, because it alone exposes the characteristics that delimit the concept inside its system. Its form comes down to two parts: the immediate superordinate concept, then the delimiting characteristics. The 2022 edition replaced "generic concept" with "superordinate concept", and the replacement carries meaning: the superordinate covers the generic relation and the partitive one, where the generic covers only one. The word immediate is operational too: the superordinate named is the nearest one, since a distant ancestor would delimit too widely and teach nothing.

Applied to Swiss health insurance, the form gives: quote-part (the coinsurance), a share of the participation aux coûts (the immediate superordinate concept) equal to 10 % of the costs above the deductible, capped at CHF 700 per calendar year for an adult and CHF 350 for a child (the delimiting characteristics). The definition names a concept, and it is the preferred term of that concept, here participation aux coûts, that must itself be a glossary entry. The distinction between the two is the one between term and concept, and a definition puts it to work on every line. That mechanism is what makes the hierarchy operational and what turns the glossary into a system of definitions: each definition rests on another entry, and the whole holds together through those supports, where an alphabetical list holds together through nothing.

ISO 704 separates essential characteristics, needed to understand the concept, from delimiting characteristics, needed to separate it from its neighbours in the same system. The practical consequence is strong: a definition is written against its siblings. Franchise (the deductible) and quote-part (the coinsurance) are two subordinates of the same superordinate, and each of the two definitions has to be written so that no reader could mistake it for the other.

The standard also names the deficient definitions, and its clause 6.5 distinguishes three: the circular definition (6.5.2), the inaccurate definition (6.5.3), which covers the too-broad as well as the too-narrow, and the negative definition (6.5.4). Definition by negation is a defect of terminology before it is a fault of writing: where a positive definition is possible, defining a concept by what it is not delimits nothing.

A fourth defect is met as often as the other three, and its authority lies elsewhere. An entry whose definition collapses into a synonym or a translation delimits nothing either, and it is ISO/IEC 11179-4 and ISO 1087 that exclude it. ISO/IEC 11179-4 requires the definition to be a descriptive phrase or sentence, which a bare word or an abbreviation is not, and it carries the other formulation rules that apply word for word to the glossary's "definition" column: the definition is singular, it stands alone (it is understood without opening another entry, apart from that of its superordinate concept), it is free of circular reasoning and it states what the concept is. ISO 1087 supplies the word that says why the defect is one: a synonym is a designation, and a designation represents the concept where a definition delimits it. The two objects have two functions, and the object, concept, designation, definition model is enough to keep them apart.

Two tests can be run on any definition in under a minute. The substitution test replaces the term with its definition in a real sentence taken from a requirement: if the sentence becomes false, absurd or circular, the definition is bad. ISO 704 sets it out in clause 6.4.4 under the name substitution principle, which makes it a citable requirement in review. The sibling test, by contrast, is a practitioner's habit: it reads the definition next to that of its nearest neighbour, and if a reader could swap them without noticing, the delimiting characteristics are missing.

Running the exercise

1. Fix the scope and publish the admission criterion
   One initiative, one domain or the whole enterprise. BABOK's four criteria are announced as the rule of entry before the harvest, failing which the glossary will absorb whatever is put in front of it.
2. Harvest the candidate terms where they already live
   General terms and conditions, legal and regulatory texts, existing lexicons, requirements, interview notes, screen labels, report column headers, error messages, ticket titles. Never a blank page. For each candidate, record where it was found and who uses it: that provenance is what makes the next step possible.
3. Detect the collisions before writing a single definition
   Cross the label against the department or the source. Two labels appearing in the same contexts signal a candidate synonymy. One label whose contexts contradict each other from department to department signals a candidate collision, and in a business domain that is almost always a polysemy, which is to say a collision whose contexts resemble one another: what shows it is the provenance of the document. The step most often skipped is the one that carries the value.
4. Resolve the relations
   Elect the preferred term of each synonym cluster, split each colliding label into disambiguated terms and record the ambiguous label as a rejected term, write down the superordination relations wherever they exist.
5. Write the definition
   Per ISO 704: immediate superordinate concept plus delimiting characteristics, positive, non-circular, self-sufficient, singular, acronyms spelled out. Run the substitution test of clause 6.4.4 and the sibling test.
6. Attach the source of authority
   For a regulated term the reference is the enactment: LAMal art. 42 para. 1 for Tiers garant and OAMal art. 103 for the amount of the deductible, which the Act delegates to the Federal Council. A federal office that circulates an amount is a source of distribution, and an entry that rests on it is an entry that will have to be reopened at the first challenge. For an internal term, the decision that fixed it and its date.
7. Name a person per term or per domain
   The owner settles the disputes and is the person the update trigger is addressed to. A team decides nothing and answers for nothing.
8. Arbitrate what cannot be agreed
   BABOK gives as its second limitation the difficulty different stakeholders have in agreeing on a single definition. Two outcomes are honest and a third is a trap. Honest: the sponsor decides, and the losing department's word is recorded as a non-preferred synonym; or the two readings really are two concepts, and the term is split. The trap is the compromise definition, a sentence broad enough for both departments to sign and for neither to use. It hides the collision behind an approved entry, which is strictly worse than open disagreement.
9. Put the glossary where the term is met
   The test is operational: does a reader who trips over an unknown term get its definition without leaving the artifact they are reading? The glossary is called from the requirements tool, injected as a tooltip in the wiki, echoed in the footer of the ticket template, at the head of the interface specification. That is the checkable translation of BABOK's access recommendation, and the glossary lodged in its own separate document is read by the person who wrote it.
10. Restrict editing
    Propose widely, approve narrowly. The status field, from "proposed" to "approved" and then "deprecated", is what makes that circuit visible.
11. Review on trigger
    The triggers are real: a new term enters a requirement, a regulation changes, a product is launched, a merger imports a second vocabulary, a system is replaced. A standing task called "review the glossary" is a task that is deferred indefinitely. A light periodic pass, annual or before a release train, remains useful for spotting the entries to deprecate, and everything else runs on events.
12. Treat each language's label as a term in its own right
    In a multilingual organisation the relation is equivalence, from one language to the other: the label is elected in its own language, from real usage, and it is verified against a source in that language. Where German-speaking Switzerland uses the French word, translating would produce a word nobody uses.

From a project glossary to an enterprise glossary

A glossary bounded to a single initiative resolves ambiguity within its perimeter and leaves it intact outside. The word two departments on a project have finally agreed on goes on denoting two things in the rest of the organisation, and the next project re-runs the arbitration without knowing it. The scope that makes a collision disappear for good is that of the whole enterprise: one preferred term per concept, valid for every department, keeps the divergence from being reborn elsewhere, and it is the only level at which the cross-domain collision shows. "Contract" does not denote the same thing in human resources, where it binds an employee, and in procurement, where it binds a supplier: two concepts, one label, which no review internal to either department confronts. That is what the Federal Chancellery's TERMDAT bank puts on display by presenting each term within its domain, where the reader sees that one and the same word does not overlap from one branch of law to another. The enterprise glossary takes up the discipline: the entry carries the domain in which the concept holds, and two domains that use one label for two concepts are split apart like any other collision.

Swiss multilingualism adds a trap that only this scope exposes. In German, Prämie denotes the risk premium, the consideration for the cover, whereas Beitrag, which one would happily render as "premium", denotes in the second pillar the savings share of the contribution: rendering "premium" as Beitrag shifts the concept from risk to savings, an error no review of the French text catches. French has its own trade turns of phrase, such as the verb policer, to issue a policy, which everyday language ignores and which a definition must nonetheless use to speak the language of the domain. These are cross-language equivalences in the sense of ISO 704, and they are verified against the real usage of each language, never against a plausible translation.

The same logic extends outside the enterprise. A partner, a reinsurer or a provider uses its own vocabulary for the shared concepts, and the mapping between its term and the internal preferred term is one more equivalence, which is worth carrying in the glossary as an access label. Recorded, it spares every exchange the re-translation each side runs on its own without checking it; omitted, it lets two organisations believe they mean the same concept when they share only the word.

What makes the exercise fail

The glossary nobody maintains

This is BABOK's first limitation and the central trap of the technique. The abandoned glossary becomes false while remaining consulted, first of all by newcomers, who have no other source to check it against. A domain with no glossary would have sent the same person to ask a colleague.

The circular definition

"Insured: a person who is insured." It is detectable mechanically, which explains why it survives so long: nobody looks.

The two-sense term carried as a single entry

The compromise definition is its direct consequence, and it produces a text both departments approve and neither applies.

The colliding label recorded as a synonym

A subtler version of the same defect and a more dangerous one, because the glossary then asserts an identity that does not exist, with the authority of an approved entry. The counter is the split and the rejected-term entry.

The glossary saturated with everyday vocabulary

BABOK's four admission criteria exist to prevent it. An entry defining "invoice" in the sense of an invoice drowns the eight entries that carried the meaning of the domain.

A language's label manufactured by translation

A plausible compound nobody uses enters the glossary, then the screens and the correspondence, and the organisation starts speaking a language that belongs to none of its customers.

Two glossaries

The project's and the enterprise's, contradicting each other. The younger one loses, and the effort invested in it is a dead loss.

Spilling over into the neighbouring techniques

A definition that carries a field's format has become a data dictionary entry. A definition that carries a fact type, "the insured holds a policy", has become a fragment of concept model. Both drifts duplicate an artifact that will be maintained elsewhere and will diverge.

Relation to the neighbouring techniques

Three techniques work the same material and are constantly confused. They form a continuum: the glossary, then the concept model (BABOK 10.11), then the data dictionary (BABOK 10.12). The glossary records terms and their agreed definition. The concept model records the meaning of, and the relations between, the concepts, adding to the vocabulary the layer of fact types, categorisations and roles. The data dictionary records the data elements, their format and their permitted values and the way they compose.

The decision rule fits in three lines. The ambiguity is in the word: that is the glossary. The ambiguity is in what is true of the thing: that is the concept model. The ambiguity is in the field: that is the data dictionary.

The glossary is the cheapest of the three and the way into the other two. Where the arbitration is missing, where the abstract-modelling skill is absent or the upkeep budget is not there, the fallback to the glossary is a legitimate choice: a glossary that is kept up beats a concept model nobody will maintain.

The fact type given as an example in the "relations recorded" row, "the insured holds a policy", contains its own demonstration of the continuum. Which insured: the preneur d'assurance, who signs the contract, or the personne assurée, whose risk is covered? The insurer's glossary rejects that label, because it denotes two. The concept model therefore needs the glossary to have elected before it: failing that, its fact type links a concept to a label that covers two, and the diagram inherits the ambiguity it was supposed to dispel. The continuum is a dependency: the downstream technique consumes the vocabulary the upstream technique has settled.

AI considerations

Four uses hold up, in increasing order of value, and all of them sit upstream of the human decision.

The harvest of candidate terms comes first, because the corpus already exists and nobody has the patience to comb it by hand. A model run over the general terms and conditions, the regulation, the requirements, the ticket titles, the screen labels and the interview notes returns the candidates with their frequency and their occurrences. The output is a list of candidates, and BABOK's admission criterion is applied afterwards, by a human.

The detection of collisions is the highest-value use, because it attacks the step that is skipped the most. On a corpus segmented by department, a model surfaces two things no reader can hold in their head across hundreds of documents: this label is used in two incompatible senses depending on the source, which marks a candidate collision; these two labels appear in the same contexts and probably denote a single concept, which marks a candidate synonymy. The machine produces the candidates, and arbitrating between two incompatible senses stays a decision of the organisation.

The mechanical check for defects comes next. The defects of a definition are for the most part text checks a machine runs without tiring where a reviewer runs them once and then stops looking: circularity, including one hop away where the eye never catches it (ISO 704 6.5.2), the negative definition (6.5.4) or the too-broad one (6.5.3), the superordinate concept used without being an entry itself, the acronym not spelled out, the definition reduced to a synonym where ISO/IEC 11179-4 requires a descriptive phrase. That check is worth installing in the review chain.

The detection of drift, finally, compares the approved definition with the term's real usage in the tickets, the specifications and the correspondence of the past year. Where usage has moved away from the definition, the glossary is going stale in the only way that matters, and that is a comparison a machine sustains over time and human attention does not. It attacks BABOK's first limitation.

A definition is a decision of the organisation. A model reading the documents of two departments correctly establishes that "assuré" is used there in two incompatible senses. It cannot establish which of the two the organisation will adopt, because until the organisation has decided there is nothing to extract: the fact does not yet exist. A definition produced by a machine and never arbitrated is a plausible sentence with no authority behind it, and it is worse than a missing entry, because an entry that looks approved ends up cited in a requirement and then built.

Three corollaries follow, and all of them bear on the fields a machine fills most readily with the plausible. For a regulated term, the authority is the legal text: a reference produced by a model is checked against the text, and the article cited has to be the one that fixes the rule, the ordinance where the Act delegates the amount. A language's label is verified the same way against a source in that language, since a morphologically correct compound costs the machine no more than an attested word. Finally, submitting internal contracts, tickets and correspondence to an external service commits the organisation under the revised Federal Act on Data Protection, in force since 1 September 2023, and that decision is taken before the corpus leaves the company.

Examples

The example is the glossary of a Swiss health insurer active in French-speaking and German-speaking Switzerland, restricted to the terms of cost sharing and of the contractual relation. It is the ideal demonstration domain for the technique: every one of its terms is defined by law, disputed between two departments or carries a French label in German.

The labels are reproduced as the insurer uses them, in French and in German. That is the technique: a designation is elected in its language and is not translated, so rendering these labels in English would fabricate terms that no Swiss insurer uses and would destroy the very lesson the table exists to teach. The two rejected-term entries are the point of the table.

An ordinary glossary would have kept "assuré" as a single term with a compromise definition and "caisse" as a synonym of "caisse maladie", because that is the institution most often meant at an insurer. Both temptations are temptations of polysemy: it is because the senses are related that a broad enough sentence looks possible, and that an alias looks defensible. Both choices produce a false assertion carried by an approved entry. The table treats them as rejected terms, each pointing back at the concepts it covered, and the Status column names the cause of the rejection.

Visualisations

The graph of the relations shows the work the table of entries does not show. A glossary entry is a result: it carries a term, a definition and synonyms, and it has erased the question that produced it. The graph puts that question back where it arises, between the candidate labels and the concepts they denote, and it makes visible the two distinctions everything depends on. The first is that labels and concepts are objects of a different nature, which the table cannot say, since its rows carry terms and its concepts are implicit. The second is that the relations do not call for the same resolution: synonymy resolves by election, losing nothing, whereas a collision forces a split and the withdrawal of the label. That asymmetry is what a reader should carry away.

The graph keeps to the relations that resolve within a single language, which makes it a model of the method. Inter-language equivalence is demonstrated where the glossary table puts it, in the German label column, where the reader sees the elected label in its place in the deliverable.

Cost

Tooling

A spreadsheet, a document or a wiki page is enough for a bounded project glossary, and their licence cost is nil. Two weaknesses come with them, and BABOK names both: there is no door on the editing, and the glossary is not reachable from where the term is met. That choice holds up when the scope is one initiative and the readership one team, and it becomes expensive as soon as either widens.

The wiki with a glossary macro (Confluence, SharePoint) clears the step that counts: the term becomes a tooltip on every page that uses it. That is the direct answer to BABOK's recommendation of easy and reliable access, and it is the cheapest way to get a glossary read.

Requirements management tools (Jama, Polarion, IBM DOORS Next or Jira with a glossary app) lodge the glossary next to the requirements that use the terms. Changing a definition then surfaces the requirements it touches, and the terminological consistency of a requirements corpus becomes a query. It is the option that pays for itself as soon as the corpus is large.

The business glossary modules of the data catalogues (Collibra, Alation, DataHub, OpenMetadata) are the same platforms as those of the data dictionary, and the module is a different one. What it brings the glossary is the stewardship circuit, the lifecycle of the status and above all the link from the business term to the data elements that realise it. That link is the bridge between the glossary and the dictionary, and it is why an enterprise buys one platform rather than two tools. The underlying stewardship model is DAMA-DMBOK's, which files the business glossary under metadata management.

Terminology tools (SDL MultiTerm, Coreon and their equivalents) are the right family as soon as the glossary is multilingual, which in Switzerland is the ordinary case. They implement the ISO 704 and ISO 1087 model natively: one record per concept, carrying its labels in each language, which a spreadsheet with one row per term cannot express without lying. When the glossary has to be published, exchanged between systems or consumed by a machine, SKOS (W3C) is the interchange format: it carries exactly the glossary's set of relations, prefLabel, altLabel, broader, narrower, related.

Two references, one Swiss and one international, are worth consulting before writing an entry. TERMDAT, the terminology database of the federal administration, kept by the terminology section of the Federal Chancellery, holds several hundred thousand multilingual records (German, French, Italian, Romansh, English) of Swiss legal and administrative terminology, freely consultable. It is at once the source to query for a regulated term and the example of a professional multilingual glossary built on the concept rather than on the word. SEVOCAB, the public database of ISO/IEC/IEEE 24765, plays the same role for the vocabulary of systems and software engineering: a governed, public glossary in which every entry carries its source.

Sources

• IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.23 Glossary.
• IIBA, A Guide to the Business Analysis Body of Knowledge (BABOK Guide) v3, §10.11 Concept Modelling and §10.12 Data Dictionary: the two techniques neighbouring the glossary.
• ISO 704:2022, Terminology work - Principles and methods: the intensional definition (immediate superordinate concept and delimiting characteristics), the substitution principle (6.4.4), the three deficient definitions of clause 6.5 (circular 6.5.2, inaccurate 6.5.3, negative 6.5.4) and clause 7.7 on the relations between designations and concepts, which gives synonymy (7.7.2) and equivalence (7.7.3) two separate clauses and treats polysemy and homonymy together (7.7.5).
• ISO 1087:2019, Terminology work and terminology science - Vocabulary: the object, concept, designation, definition model on which the distinction between term and concept rests and the definitions of synonymy (within one language), polysemy (related senses) and homonymy (unrelated concepts).
• ISO 25964-1:2011, Information and documentation - Thesauri and interoperability with other vocabularies - Part 1: Thesauri for information retrieval: preferred and non-preferred terms, as well as the equivalence, hierarchical and associative relations. Its equivalence relationship is the preferred / non-preferred relation within a language, not to be confused with the inter-language equivalence of ISO 704.
• W3C, SKOS Simple Knowledge Organization System Reference: the machine-readable form of the same set of relations, prefLabel, altLabel, broader, narrower, related.
• ISO/IEC 11179-4, Information technology - Metadata registries (MDR) - Part 4: Formulation of data definitions: the rules for formulating a definition, singular, a descriptive phrase or sentence, self-sufficient, free of circularity. It is the descriptive-phrase rule that excludes the definition collapsed into a synonym.
• ISO/IEC/IEEE 24765:2017, Systems and software engineering - Vocabulary, and the public SEVOCAB database: a governed, publicly maintained professional glossary, taken as the model of upkeep.
• Swiss Federal Chancellery, TERMDAT: the terminology database of the federal administration, several hundred thousand multilingual records of Swiss legal and administrative terminology, freely consultable.
• LAMal / KVG, SR 832.10, Federal Act of 18 March 1994 on Health Insurance: art. 42 (tiers garant and tiers payant), art. 64 (cost sharing), art. 65 (premium reduction), the sources of authority of the worked example.
• OAMal / KVV, SR 832.102, Ordinance of 27 June 1995 on Health Insurance: art. 103 (the amount of the deductible and the annual cap on the coinsurance) and art. 104 (the contribution to hospital-stay costs and its exemptions), the articles that fix the amounts the Act delegates to the Federal Council.
• Federal Office of Public Health, Health insurance: premiums and cost sharing: the distribution of the amounts and the usage source against which the German labels of the example were verified, Franchise, Selbstbehalt and Spitalbeitrag.
• DAMA International, DAMA-DMBOK: Data Management Body of Knowledge, metadata management chapter: the business glossary and its stewardship model.
• HERMES 2022, Project Management Reference Manual, Swiss Confederation, Federal Chancellery, §A.5 The interfaces of HERMES project management and the glossary: the method's terminology as the guarantee of a common language between the core organisation, the project organisation and project management, application management and portfolio management; the six four-language tables that give phases, scenarios, modules, outcomes, tasks and roles with the number of the method element.