Translate | ISO 24613-1:2024 Lexical Markup Framework, Word Senses, Multilingual Lexicons and Termbase Mappings

How do you translate terminology consistently when one word has several senses and each language divides those meanings differently? A multilingual lexicon, termbase or translation memory must represent more than an English headword followed by a convenient foreign-language equivalent. ISO 24613-1:2024, the core model of the Lexical Markup Framework (LMF), provides a conceptual basis for modelling lexical resources used by computer applications. This guide explains how to organize word forms, senses, usage evidence, mappings and review decisions for translators, terminologists, dictionary editors, language-technology teams and multilingual educational publishers.

Consider the English word ‘bank’: a financial institution, the edge of a river, and a verb meaning to tilt a vehicle have different meanings. A bilingual table that assigns all three occurrences to one target term is not a vocabulary resource; it is a hidden error generator. The same problem appears with words such as ‘charge’, ‘issue’, ‘claim’, ‘due’, ‘appeal’ and ‘draft’. Correct translation begins by identifying the sense and communicative function in context, then choosing an appropriate target expression. Machine-readable labels and controlled identifiers help preserve that decision when thousands of texts are updated over time.

At a glance: define the purpose and users of the lexicon; decide whether a record represents a lemma, written or spoken form, sense, term, named entity or multiword expression; keep source evidence and usage constraints attached to the correct level; distinguish equivalence from relatedness; record regional, register and domain restrictions; prevent machine systems from guessing missing translations; check licensing and provenance; and review how vocabulary records behave in real sentences. ISO 24613-1:2024 is a published core modelling standard, not an automatic translation engine and not a claim that one word always has one universal counterpart.

Quick answer: how to create a translation-ready multilingual lexicon

Choose the domain, audience, languages, scripts and evidence sources before designing fields.
Represent distinct meanings separately even when the spelling is identical.
Attach each candidate translation to the specific meaning and usage condition it serves.
Record examples, provenance, domain limits and review status rather than storing unexplained equivalents.
Test full sentences and downstream tools to identify inappropriate cross-language substitutions.

Chapter 1. What the LMF core model contributes

The second edition, published in January 2024, establishes a core metamodel for representing lexical information in monolingual and multilingual resources used by computer applications. A metamodel describes conceptual types and relationships; it is not simply a spreadsheet template and does not itself supply accurate translations. The public ISO description does not authorize inventing a detailed field prescription on its behalf. This article’s practical field suggestions are educational design choices rather than quotations or a conformity checklist.

For each design decision, build a tiny contrastive test: two phrases that look similar but demand different linguistic treatment. If the data model cannot distinguish them, revise the model rather than hiding the distinction in a note nobody reads.

Chapter 2. Begin with lexical meaning, not a word list

A useful lexicon describes how a form behaves in actual communication. A written string is not identical to a grammatical category or a sense. Homographs may belong to different entries, one lemma may have several inflected forms and one term may denote a specialized concept. The unit of translation is the meaning expressed in a particular utterance, not the shape of its isolated letters.

The editor’s essential question is where the evidence belongs. A property about spelling should not be attached indiscriminately to every sense; a domain restriction should not be discarded when exporting a single target candidate.

Chapter 3. Separate forms, entries, senses and concepts

A team might store the headword as a lexical entry, its spellings and inflections as forms, and its contextual meanings as senses. A specialized concept record may point to several language-specific expressions. These distinctions prevent one update to a spelling from overwriting a domain-specific definition. They also allow a reviewer to tell whether a missing translation is an unresolved mapping or a deliberate gap.

A second reviewer should be able to reproduce the choice from the record alone: context, intended meaning, candidate expression, constraints, source of evidence and status. Otherwise the system contains an undocumented judgment.

Chapter 4. Build equivalence that is honest about its limits

Words rarely map one-to-one across every context. A candidate equivalent might work in accounting but not in conversation, or in one region but not another. Label the relation according to what evidence supports. Distinguish approximate translation, descriptive paraphrase, technical designation, obsolete label and forbidden false friend. Avoid asserting perfect equivalence merely to complete a dataset.

Chapter 5. Use evidence, not invented definitions

Record where a meaning and usage example came from, when it was checked and which reviewer approved the interpretation. Use your own original examples or properly licensed sources. If two documents conflict, preserve the disagreement until a domain expert resolves it. A polished glossary with weak evidence remains fragile, particularly when it feeds machine translation or automated content generation.

Chapter 6. Treat morphology and grammar as translation information

Number, tense, aspect, gender, case, classifier choice and agreement can determine the target construction. A termbase that keeps only the dictionary form may be unable to produce a grammatical sentence. Store the distinction between a lemma and its forms. When automatic inflection is used, require language-specific tests rather than assuming a universal rule applies.

Chapter 7. Preserve register, region and audience

A school reader, immigration officer, clinician, software user and literary audience can understand the same topic through different lexical choices. Geographic variety and formality are not decoration; they affect acceptability and interpretation. A good record states who uses a form and in what genre, not only which language it belongs to.

Chapter 8. Handle scripts, normalization and search

Unicode text may contain visually similar but technically distinct sequences; scripts and transliterations introduce further variation. Normalization can improve lookup but can also erase meaningful contrasts if applied without care. Use search keys separately from the original form, protect diacritics and audit right-to-left display. A search alias is not evidence that two senses are synonymous.

Chapter 9. Prepare revisions and interoperability

Language resources evolve. Record stable identifiers, versions, source references and the status of deprecated mappings. Export formats should retain relations and provenance, not flatten every sense into a two-column table. Test whether a consuming application can still distinguish preferred and rejected candidates after conversion. A migration that loses restrictions is a semantic data loss, not merely a technical inconvenience.

Chapter 10. Keep the user and task in view

A lexicon succeeds when a translator or application chooses the right expression in a real task. Test ambiguous sentences, inflected forms, false friends, idioms and restricted contexts. Review not only coverage, but also harmful confidence: a resource that returns an unqualified answer for an unresolved sense is more dangerous than one that asks for human attention.

Case laboratory: 40 translation and lexicon decisions

The examples below are original constructed teaching cases. They show how a language-resource editor can preserve meaning when words, terms, forms, senses, scripts and application constraints intersect. The ISO core model supplies a conceptual point of reference; these are not claims about mandatory proprietary fields.

Case 1. Bank as financial institution

Translation problem. A banking app contains ‘open a bank account’ beside a travel article describing the river bank. Both texts are imported into the same English-headword list.

Meaning that could be lost. The dangerous shortcut is one global target-language replacement for bank, which makes a geographical passage read like financial advice.

Better resource-design decision. Separate senses, associate each with examples and domain labels, and require the caller to supply or infer context before presenting an equivalent. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 2. Charge as price or accusation

Translation problem. Customer support writes ‘we will reverse the charge’, while a news translation mentions a criminal charge. The English surface form is identical.

Meaning that could be lost. A financial term inserted into a legal sentence can change both meaning and seriousness; a criminal term in a billing email can sound alarming.

Better resource-design decision. Create distinct sense records and review candidate equivalents with realistic billing and legal sentences rather than dictionary labels alone. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 3. Issue as problem or publication

Translation problem. An editor uses ‘the May issue’ while a technician reports ‘an issue with the sensor’. The source corpus lacks explicit domain tags.

Meaning that could be lost. A language model may treat an editorial issue as a defect or misread a defect report as a periodical item.

Better resource-design decision. Attach publication and problem senses with attestations, then use context-rich examples to classify new occurrences before accepting translation. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 4. Claim as assertion or insurance request

Translation problem. A supplier claims that its component passed a test, while an insurer records a customer claim for reimbursement.

Meaning that could be lost. A glossary that collapses assertion and formal reimbursement makes compliance notes and customer communications misleading.

Better resource-design decision. Keep separate usage patterns, including who makes the claim and which procedure receives it; mark uncertain cases for subject expert review. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 5. Draft as preliminary text or airflow

Translation problem. An office edits a draft report while an engineer studies a ventilation draft. The common spelling invites a naïve equivalence rule.

Meaning that could be lost. One lexical record cannot silently determine whether the text concerns document status or physical movement of air.

Better resource-design decision. Associate domain and collocation with each sense, and test both document-editing and engineering examples before exporting the mapping. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 6. Due as owed or expected

Translation problem. A classroom says ‘assignment due Friday’; accounts payable says ‘amount due’; a report says ‘the change was due to rainfall’.

Meaning that could be lost. The wrong target expression can turn an explanation into a payment demand or blur a deadline.

Better resource-design decision. Model due-date, amount-owed and causal-expression uses separately and flag phrase-level constructions requiring more than one translated word. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 7. Appeal as attractiveness or legal review

Translation problem. A museum advertises visual appeal while a court notice describes the deadline for an appeal.

Meaning that could be lost. A general notion of attractiveness is not equivalent to a formal legal review procedure; confusion can affect rights and deadlines.

Better resource-design decision. Record register and procedural context, and use specialized review for legal senses instead of treating the word as a style choice. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 8. Record noun versus verb

Translation problem. An archive contains a record, while a mobile app asks the user to record audio. The headword spelling is shared but grammar differs.

Meaning that could be lost. A two-column glossary may force the noun into a verb position or choose a phrase with the wrong action.

Better resource-design decision. Distinguish grammatical category, inflection and sense, then test the entry both as an object and as an instruction. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 9. Lead metal versus guide

Translation problem. A chemistry sheet refers to lead contamination; a project plan asks a mentor to lead a team.

Meaning that could be lost. Spelling alone cannot identify pronunciation or meaning, making isolated-word lookup unreliable.

Better resource-design decision. Separate entries or senses appropriately, associate pronunciation and subject evidence, and prevent unreviewed machine substitution. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 10. Present noun, adjective and verb

Translation problem. A teacher asks students to present findings; another sentence describes the present moment; a third mentions a birthday present.

Meaning that could be lost. Without part-of-speech and sense distinctions, a generated target sentence may be grammatical but refer to a gift rather than a presentation.

Better resource-design decision. Use contextual examples and category labels, and require full-sentence QA for all three uses. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 11. Set as mathematical collection

Translation problem. A mathematics lesson defines a set of objects, while a television article describes the film set and a recipe says ‘let it set’.

Meaning that could be lost. Treating these as synonyms because they share an English lemma destroys subject-specific meaning.

Better resource-design decision. Keep the mathematical concept, production location and physical-change senses separate, including examples from each kind of text. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 12. Current as flow or present

Translation problem. An electrical report mentions current through a wire while an HR newsletter describes the current policy.

Meaning that could be lost. One may need a technical quantity term, the other an adjective for present validity.

Better resource-design decision. Label electrical and temporal senses, record context triggers and assess any ambiguous document title manually. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 13. Cell in biology and spreadsheets

Translation problem. A scientist studies a cell, a finance team edits spreadsheet cell B12, and a company describes a prison cell in history material.

Meaning that could be lost. The shared label disguises different referents and different translation conventions.

Better resource-design decision. Preserve concept descriptions and domain links, and prevent the software from inferring the biology meaning from the first example in its database. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 14. Light as illumination or weight

Translation problem. A camera guide discusses soft light while a product listing says lightweight and a poem describes lightness of mood.

Meaning that could be lost. The same form may involve distinct physical, comparative and metaphorical meanings that cannot be substituted mechanically.

Better resource-design decision. Add usage examples and sense-specific translation candidates; a mood expression may require a full phrase rather than the base adjective. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 15. Fine as penalty or quality

Translation problem. A regulation lists a fine for late filing, while a quality review describes fine detail and a conversation says ‘I’m fine’.

Meaning that could be lost. A word-level mapping can turn a polite reassurance into a legal penalty, or alter a technical description.

Better resource-design decision. Classify penalty, detail and conversational uses separately, and explain why a translation requires a different construction in each. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 16. Close as near or shut

Translation problem. A location says ‘close to the station’, while a software button says ‘Close window’.

Meaning that could be lost. A grammatical environment and software action determine whether a target expression describes distance or commands closure.

Better resource-design decision. Store imperative UI context and descriptive adjective context, then check the rendered button action rather than only term equality. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 17. Borrow and lend direction

Translation problem. A learner’s sentence says ‘I borrowed the book from Mei’; another says ‘Mei lent the book to me’.

Meaning that could be lost. Languages may encode the perspective differently. Treating the verbs as interchangeable reverses who gave and received the item.

Better resource-design decision. Record event roles and example arguments, not only loose dictionary equivalence, and test the same exchange from both perspectives. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 18. Hire and rent distinctions

Translation problem. A travel website advertises renting a vehicle while an employment ad hires staff; some English varieties use ‘hire’ for both.

Meaning that could be lost. Region and domain affect the intended target term, and a superficial match can turn a rental service into recruitment.

Better resource-design decision. Separate employment and temporary-use senses, with regional labels, and examine actual purchase or employment actions. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 19. Teacher and tutor labels

Translation problem. An education site uses ‘tutor’ for a small-group teacher, while a university text uses ‘tutor’ for an academic adviser.

Meaning that could be lost. A target language may have distinct terms for private tuition and institutional support roles.

Better resource-design decision. Record the service and learner relationship, preserve the site’s own naming definitions and avoid a generic role label that changes expectations. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 20. False friend in formal notice

Translation problem. A bilingual editor encounters a cognate that resembles an English word but has a narrower meaning in the target language.

Meaning that could be lost. A visually familiar form can look credible while shifting a restriction into a permission or a provisional decision into a final one.

Better resource-design decision. Store rejected lookalikes as warnings, with counterexamples, so translators learn what not to select. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 21. Regional word for a dwelling

Translation problem. Housing guidance addresses apartment, flat and condominium, while translated examples target readers in different regions.

Meaning that could be lost. A single canonical expression may obscure tenure type, building form or local terminology.

Better resource-design decision. Record geographic usage and concept boundaries; if the source refers to a legal property class, seek locally authorized vocabulary. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 22. Plural morphology

Translation problem. A product interface shows ‘1 file’ and ‘2 files’ and assumes the same grammatical plural rule works in every target language.

Meaning that could be lost. Some languages need different plural categories, and some display nouns differently according to number.

Better resource-design decision. Model the content variation and test actual quantities rather than translating a static singular string in isolation. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 23. Grammatical gender

Translation problem. An automated notice inserts a translated profession term after a person’s name but supplies no grammatical context.

Meaning that could be lost. A forced masculine or feminine form may be inaccurate or insensitive in a language with agreement requirements.

Better resource-design decision. Design context-aware wording approved for the intended audience; do not claim the LMF core model automatically settles pronoun policy. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 24. Compound nouns

Translation problem. A technical glossary includes ‘pressure relief valve’, but a segmentation system treats each word as an independent term.

Meaning that could be lost. The compositional word-by-word translation can misname the device and its engineering role.

Better resource-design decision. Store the multiword expression with its own conceptual definition, usage evidence and preferred domain equivalent. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 25. Idiom versus literal phrase

Translation problem. A customer support guide contains ‘break the ice’ in a social context, while a maintenance manual literally instructs workers to break ice.

Meaning that could be lost. Idiom translation may require a different metaphor; maintenance instructions require literal action and safety accuracy.

Better resource-design decision. Record separate phrase-level meanings and test whether the target reader would perform or understand the intended action. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 26. Translation candidate with missing evidence

Translation problem. A terminology editor adds a plausible target word for an unfamiliar specialist concept without a source citation.

Meaning that could be lost. Other editors may assume the candidate is approved because the database displays it as a normal result.

Better resource-design decision. Use an explicit provisional status, owner and question; block its automatic release until domain evidence and review exist. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 27. Preferred versus permitted synonyms

Translation problem. A software product uses one term consistently in its navigation, but external technical documentation also uses a synonym.

Meaning that could be lost. Both may be linguistically acceptable, yet switching them in the interface can confuse returning users.

Better resource-design decision. Distinguish a project-preferred variant from an allowable alternative and preserve the context of that preference. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 28. Deprecated product vocabulary

Translation problem. A user guide refers to an old feature name while the new application has adopted a different name.

Meaning that could be lost. Replacing historical terms in archived documents can falsify the record, but reusing obsolete terms in new releases creates inconsistency.

Better resource-design decision. Store dates and version applicability, and map old forms to new ones without silently rewriting historical evidence. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 29. Brand versus ordinary noun

Translation problem. A proper product name contains a common English word that translation software wants to translate literally.

Meaning that could be lost. Replacing a protected name with a descriptive common noun can make the output refer to another product.

Better resource-design decision. Mark names with their correct status and approved localized variants; keep semantic descriptions as separate explanatory fields. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 30. Scientific abbreviations

Translation problem. A laboratory glossary includes an abbreviation shared with an unrelated computing term in the same organization.

Meaning that could be lost. Expansion without disciplinary context may misidentify a measurement, method or component.

Better resource-design decision. Store abbreviation senses with full-form evidence, domain restrictions and verified examples in which each acronym occurs. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 31. Language code versus country code

Translation problem. A multilingual database stores the target locale under a short string and assumes language is identical to country.

Meaning that could be lost. The same language may have multiple regional varieties and scripts; a country can host many languages.

Better resource-design decision. Represent language, script and region deliberately, and audit metadata rather than using country flags as translations. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 32. Script variants of the same expression

Translation problem. An editor has an expression in one script and an authorized transliterated display for a search index.

Meaning that could be lost. A display form, search alias and semantic translation are not the same kind of relation.

Better resource-design decision. Preserve the original lexical form and label the transformations explicitly, then test retrieval without collapsing distinct words. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 33. Unicode normalization

Translation problem. Two user-entered words look identical on screen but consist of different Unicode sequences in the source database.

Meaning that could be lost. A search system may return duplicates or fail to match, while aggressive cleanup may alter a genuinely significant form.

Better resource-design decision. Maintain reversible original data, apply documented normalization in an appropriate search layer and inspect exceptions. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 34. Collocation strength

Translation problem. An English phrase uses ‘make a decision’ while a target language expresses the same concept with a different customary verb.

Meaning that could be lost. A dictionary entry that translates make and decision independently can produce an awkward or misleading collocation.

Better resource-design decision. Store recurring phrase patterns, use representative examples and evaluate naturalness without losing the original decision sense. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 35. Metaphor in academic writing

Translation problem. A science lesson calls a process a ‘bottleneck’, meaning a constraint rather than a literal narrow bottle neck.

Meaning that could be lost. The target language may prefer a conventional technical metaphor or an explicit description depending on audience.

Better resource-design decision. Record the underlying concept and register, and let reviewers choose wording that preserves explanation rather than matching the image mechanically. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 36. Search synonym versus translation equivalence

Translation problem. A website expands queries using loosely related words so readers can find content even when they use informal vocabulary.

Meaning that could be lost. Search usefulness does not mean the two expressions can replace each other in formal writing.

Better resource-design decision. Keep retrieval relations separate from translation relations; test them under different success criteria. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Case 37. Legal term with jurisdictional limits

Translation problem. A bilingual contract glossary offers one equivalent for ‘trust’ even though the destination legal system organizes fiduciary relationships differently.

Meaning that could be lost. An apparent lexical match may hide a substantive institutional difference.

Better resource-design decision. Add a jurisdiction note and descriptive explanation; require competent legal review where exact equivalence is unavailable. Do not approve a target term until an editor can point to a usable sense, contextual example and appropriate evidence. The explicit absence of a match is often the most accurate data value.

Case 38. Medical everyday versus clinical term

Translation problem. A patient booklet says ‘heart attack’ while a specialized report uses a diagnostic term for a precisely defined condition.

Meaning that could be lost. The reader’s level and actual clinical scope matter. An overly broad translation may make two distinct medical concepts appear identical.

Better resource-design decision. Keep definitions and audience constraints; route clinical claims through qualified review instead of smoothing terminology. Re-test any proposed mapping in a sentence that contrasts it with its nearest competing sense. This makes an apparently convincing dictionary match answer to a practical interpretation challenge.

Case 39. Inflected search query

Translation problem. A student types a word in an inflected form, but the dictionary contains only the lemma.

Meaning that could be lost. The search might return no result or treat inflection as a separate meaning, reducing access without improving accuracy.

Better resource-design decision. Link forms to the correct lexical entry and test morphology-sensitive retrieval for realistic input examples. Preserve the decision’s provenance, its regional and domain applicability, and the reviewer who authorized it. These matter when terminology is reused months later by a different project.

Case 40. Historic senses in an archive

Translation problem. A nineteenth-century document uses a word in an obsolete meaning not found in a contemporary product glossary.

Meaning that could be lost. Replacing it with today’s most frequent sense distorts the historical claim.

Better resource-design decision. Add period and source context, preserve an explanatory note where necessary, and keep the archival evidence attached. Check both directions of use. A target expression may be a good translation in one context without yielding the same English source expression when translated back.

Terminology: a practical lexicon editor’s reference

Lexical entry: A record representing a word or other lexical unit, distinct from an uncontrolled translation string.
Lemma: A conventional citation form under which relevant inflected forms may be organized.
Word form: A particular spoken or written realization, which can vary with grammar and orthography.
Sense: A meaning associated with a lexical expression in particular contexts.
Polysemy: Multiple related senses associated with one lexical form, not proof of a universal target equivalent.
Homograph: A written form shared by words that may differ in meaning or other linguistic properties.
Part of speech: A grammatical classification used to distinguish noun, verb, adjective and other functions.
Termbase: A managed terminology resource typically organized around domain-specific concepts and designations.
Multilingual lexicon: A lexical resource containing language-specific entries and relations across languages.
Equivalent candidate: A proposed target expression for a specified sense or function; evidence and restrictions still matter.
False friend: A word resembling a form in another language but leading to an inappropriate meaning.
Register: A pattern of usage associated with formality, profession, audience or communicative setting.
Collocation: A recurrent combination of expressions whose natural target wording may not be word-for-word.
Multiword expression: A unit whose meaning or conventional use spans several words.
Provenance: The source, date and editorial history supporting a lexical assertion.
Deprecated: A status indicating that a formerly used expression should not necessarily be chosen for new content.
Transliteration: A representation of script form, not a semantic translation.
Normalization: A documented transformation used for consistent data handling or search, subject to reversibility checks.
Variant: An alternative realization conditioned by region, spelling, register or other factors.
Domain label: A constraint identifying where a sense or term is appropriate.
Sense disambiguation: Determining the intended meaning of an expression from its surrounding context.
Human review: Expert confirmation of semantic, linguistic and operational correctness beyond automatic lookup.
Interoperability: The ability for resources and software to exchange useful semantic distinctions, not just strings.
Core model: An underlying conceptual representation that can support a family of lexical resource structures.

Ten-stage multilingual lexicon workflow

Step 1. Define purpose and risk

Decide whether the lexicon supports education, legal translation, health communication, product localization or search. These workflows have different tolerance for ambiguity and different review requirements.

The practical deliverable is an auditable record, not an unsupported list of convenient equivalents. Make the source evidence and authorized approval easy to locate.

Step 2. Specify language and script coverage

Record language tags, writing systems and regional varieties with appropriate granularity. Do not substitute a country label for the user’s actual language.

One skilled reviewer should attempt to falsify the mapping using a deliberately contrasting sentence. A record that survives such a test has stronger value than a fluent isolated-word translation.

Step 3. Prepare a source and rights register

Track where every candidate example and definition originated, whether it can be used and whether it requires attribution, permission or restricted distribution.

A usable decision leaves its limitations visible. Clear constraints prevent later users from turning a helpful example into an unconditional rule.

Step 4. Identify lexical units

Separate lemmas, inflections, multiword expressions, acronyms, named entities and terms. One UI string can contain more than one lexical unit.

The practical deliverable is an auditable record, not an unsupported list of convenient equivalents. Make the source evidence and authorized approval easy to locate.

Step 5. Disambiguate meanings

Write contrastive definitions and contextual examples. Do not connect a target equivalent to a headword when it applies to only one of that headword’s senses.

One skilled reviewer should attempt to falsify the mapping using a deliberately contrasting sentence. A record that survives such a test has stronger value than a fluent isolated-word translation.

Step 6. Record target candidates and restrictions

State domain, register, geography and intended use. Label unverified proposals clearly and allow unresolved gaps to remain unresolved.

A usable decision leaves its limitations visible. Clear constraints prevent later users from turning a helpful example into an unconditional rule.

Step 7. Check grammar and forms

Test inflection, agreement, phrase structure and collocations; avoid assuming the target word inherits English morphology.

The practical deliverable is an auditable record, not an unsupported list of convenient equivalents. Make the source evidence and authorized approval easy to locate.

Step 8. Test in consuming applications

Export a trial dataset and examine the translation tool, dictionary front end, search index or educational exercise that consumes it. Check whether constraints survive the export.

One skilled reviewer should attempt to falsify the mapping using a deliberately contrasting sentence. A record that survives such a test has stronger value than a fluent isolated-word translation.

Step 9. Approve revisions with identifiers

Keep stable record identities, review notes, dated status and a reproducible mapping from older versions to newer ones.

A usable decision leaves its limitations visible. Clear constraints prevent later users from turning a helpful example into an unconditional rule.

Step 10. Monitor errors and learner feedback

Use observed failures to improve definitions and examples. Distinguish a software lookup defect from a faulty lexical judgment or a source-text ambiguity.

The practical deliverable is an auditable record, not an unsupported list of convenient equivalents. Make the source evidence and authorized approval easy to locate.

Frequently asked questions

Is LMF a translation engine?

No. It provides a core metamodel for lexical resources. A separate application and editorial process determine translations.

Does every English word have one equivalent?

No. Meaning, grammatical function, domain and culture can require several different target expressions or explanatory paraphrases.

How do I model an ambiguous word?

Separate its meanings, include contrasting usage examples and attach each translation candidate to the appropriate meaning.

Can a thesaurus synonym become a translation automatically?

No. Retrieval relatedness and contextual interchangeability are different tests.

Should I preserve examples from corpus texts?

Use examples only under lawful rights and appropriate privacy rules; record provenance and distinguish invented from attested examples.

What happens when a term has no accepted target?

Keep the gap explicit, ask a domain specialist and consider a clearly identified descriptive paraphrase rather than fabricating an authoritative term.

Do regional labels matter?

Yes. A term accepted in one regional variety or professional setting can be awkward or wrong elsewhere.

Can machine translation populate my lexicon?

It can generate candidates for controlled review, but raw machine outputs are not verified terminology.

How do I handle inflections?

Preserve the lemma/form distinction and test real grammatical contexts in each target language.

What about words in different scripts?

Maintain the original form and label transliterations and search aliases distinctly from semantic equivalents.

Is the 2024 edition the same as the 2019 edition?

No. The official ISO catalogue identifies 2024 as the current second edition of Part 1, replacing the 2019 first edition.

Does this guide reproduce ISO implementation clauses?

No. It draws on the public scope and provides original editorial and translation examples, not a normative implementation specification.

Four-week learning and quality programme

In week one, gather twenty terms from a carefully defined subject area. For each, identify lemma, part of speech, source definition, evidence and at least two context sentences. Week two builds a small multilingual resource: mark candidate equivalents, absent matches, domain restrictions and permitted variants. Week three tests difficult cases—polysemy, idioms, inflection, false friends and regional usage—in realistic sentences, not isolated lookup drills. In week four, exchange records with another editor who has no access to your private notes. Ask them to explain and defend each mapping from the stored evidence, check the data in a consuming application, and classify every uncertainty. The objective is fewer false confidently returned answers, not a larger but fragile glossary.

Sources and protected learning connections

The public scope and edition were checked on 5 November 2026 against ISO 24613-1:2024, Language resource management — Lexical markup framework (LMF) — Part 1: Core model. ISO lists the second edition as published January 2024, replacing Part 1’s 2019 edition. The discussion here is original educational interpretation and does not reproduce protected standard clauses or assert technical conformity.

For the overarching principles of meaning transfer, use the existing Master Art of Translation. For vocabulary acquisition, semantic depth and classroom learning, use the protected Vocabulary Learning Hub. For sentence meaning, context and grammar, use How English Works: Translation and Meaning Across Languages. This article addresses only the bounded task of building translation-ready lexical data; it is a child lesson, not another master hub.