How can a museum translate CIDOC CRM 7.1.3 records without altering the historical relationships in a cultural heritage knowledge graph? The CIDOC Conceptual Reference Model is an ontology for connecting objects, people, places, events, periods and evidence across museums, archives and research collections. Its classes, properties and identifiers are not just technical decoration: they describe who did something, what happened, where it happened, and which historical claim is supported. A multilingual catalogue succeeds when a visitor can read the record naturally while those relationships remain intact.
Searches for CIDOC CRM localization, multilingual museum ontologies, RDF heritage metadata translation, CRM event modelling, ontology label translation and ISO 21127 cultural heritage interoperability often lead to the same underlying problem. A translator encounters human-readable labels next to stable machine identifiers, concept definitions and evidence-bearing relations. Changing the wrong layer can quietly reclassify an object, assign an action to the wrong participant or turn an uncertain date into a confirmed event. Translating only the words that should change is therefore an exercise in understanding the model.
This practical guide explains CIDOC CRM 7.1.3 for curators, collection managers, data engineers, translators, digital humanities scholars and heritage portals worldwide. It uses the February 2024 official release aligned with ISO 21127:2023 rather than treating later working drafts as final specifications. It follows a record from source modelling through language selection, terminology review, validation and public display. The goal is not to make one language sound like another; it is to make the same defensible account available to more readers without rewriting the historical evidence.
Why cultural heritage translation needs an evidence-preserving model
Museums do more than attach words to pictures. They maintain research claims about objects and their journeys through time. A visitor wants a readable explanation, a scholar wants traceable evidence, and a shared portal needs stable data links. CIDOC CRM connects these needs through entities and relationships. The principle is straightforward: language is an access path into the same evidence, rather than a license to alter the evidence itself.
A dependable workflow begins with a small record sample, not with replacing every string. Find the entity identifiers, event roles, property definitions, authority references, historical qualifiers, rights statements and language-bearing literals. Decide which are translatable and which must remain fixed. Test what the end user sees after indexing and display. The hardest errors are often those that preserve correct syntax while changing who did what.
Thirty-two practical decisions in multilingual CIDOC CRM work
1. An ontology is a model of relationships
A collection may hold a bronze vessel, a photograph of it and a scholarly claim about its excavation. CIDOC CRM describes distinct entities and relationships rather than merely storing a paragraph of catalogue prose. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
If every record is flattened into translated sentences, evidence links and event roles can disappear without an obvious error. Identify the relevant object, event, actor and document nodes before changing their display text. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Ask whether a cross-language query still locates the same object and the same claim. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
2. Official releases and draft vocabulary
The CIDOC CRM site lists official 7.1.3 and more recent draft documents with ongoing proposals. An official implementation should declare its actual version and use the corresponding class and property definitions. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
Treating a 2026 draft as a mandatory production update can create incompatible property meanings or unexpected validation failures. Pin the ontology version and store migration work separately from ordinary translation changes. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Compare serialized class names against the cited official release, not a draft used for discussion. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
3. Stable URIs versus translated names
An RDF resource can have one persistent URI and several human-readable labels in different languages. The identifier provides machine identity; each label gives a reader a linguistic way to recognize the entity. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
Replacing the URI with a translated title creates an apparently new person or object in linked-data searches. Preserve resource identifiers verbatim and attach reviewed language-tagged label variants where the model permits. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Search using both language labels and confirm that the graph returns exactly one underlying resource. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
4. Classes describe kinds of things
An E21 Person and an E74 Group can participate in the same exhibition without becoming the same kind of entity. Class membership answers what a node represents, while a label explains it to a human reader. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
A translated word such as association may tempt an editor to change a person’s type or treat a workshop as an individual. Keep ontology class references fixed and translate presentation labels according to established terminology. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Inspect inferred types and counts before and after localization to catch unexpected reclassification. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
5. Properties carry factual assertions
A relationship in the graph can connect an event with a participant, time-span or location. Property identifiers express particular meanings; a translated display label does not authorize a new relationship. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
Substituting a broad everyday verb for a specific property may turn participation into ownership or production into discovery. Leave property URIs intact and create a glossary that explains their meaning accurately in each supported language. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Review sample subject-predicate-object statements beside their public-language renderings. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
6. People, groups and institutions
A named conservator, a museum department and a historical workshop may all appear in the same object history. Preserve the distinction between individual persons, collective actors and institutions, including changing memberships over time. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
A simplified translation that names the museum as the artist can erase the actual maker and confuse attribution. Retain actor node identity and translate relationship explanations without conflating agent types. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Check that the attribution panel names the right actor in each role after language switching. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
7. Events differ from enduring objects
A sculpture persists over time, while its manufacture, exhibition, restoration and transfer occur as different events. The model’s event orientation lets institutions explain an object’s history without merging the object with any single episode. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
A target-language catalogue may wrongly make an exhibition date look like the date of creation. Separate object descriptions from event descriptions and translate each event type consistently. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Trace one object’s production and exhibition events to confirm their dates and actors remain separate. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
8. E5 Event and E7 Activity
A storm that damages a site differs from a planned conservation intervention performed by a qualified team. Both can be historically significant, but an intentional activity should not be translated as an accidental happening. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
Calling restoration natural deterioration alters the account of deliberate action and human responsibility. Use verified event-category terminology and preserve any recorded participants, causes and intentions. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Compare the same two records in both languages and ask whether the agency of action remains clear. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
9. E12 Production and E65 Creation
An archaeological pot and a digital artwork can have distinct histories of coming into existence. Different CRM concepts may model making physical things and creating conceptual or informational entities. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
A translator who treats every instance of creation as physical manufacture can misdescribe what came into being. Follow the specific implemented CRM class and approved domain vocabulary rather than using one generic verb everywhere. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Verify that the resulting translated record refers to the same product or created work. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
10. Place is not merely a place-name
A record of excavation at one location may also identify a workshop, a later storage facility and a current display room. Place nodes and event-place relationships tell the reader why a location matters. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
Translating a place name without its role can make an excavation site appear to be the site of manufacture. Preserve the linked E53 Place resources and the predicates that give them meaning in context. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Check maps, timelines and object captions together; their locations must not contradict one another. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
11. Time-spans and uncertainty
A curator may record a period such as the late nineteenth century rather than an exact calendar day. E52 Time-Span modelling can retain temporal information without inventing precision absent from the evidence. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
A target phrase such as in 1890 may falsely narrow a broad historical estimate into a verified date. Preserve structured temporal bounds and translate approximate expressions with equivalent epistemic strength. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Compare the public timeline with source temporal fields, including earliest and latest possible limits. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
12. Appellations and alternative names
One historic city may be known by several language forms, spellings and names used in different periods. Names should connect to a place or actor without being mistaken for several distinct identities. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
A careless import can create duplicate place nodes when a historic spelling is translated as a new resource. Maintain name variants with language, source and contextual status rather than replacing the underlying authority identity. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Confirm that both the old and current name resolve to the same appropriately qualified record. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
13. Identifiers and catalogue accessions
An object can carry an institutional accession number alongside public titles and alternative identifiers. An identifier is evidence of record linkage, not prose to be made more elegant. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
Localizing punctuation inside accession numbers can break the join between a catalogued object and conservation files. Freeze identifier values and translate only surrounding captions such as accession number or record ID. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Run exact-match lookups using the source identifier after target-language publication. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
14. Types and controlled concepts
Two collections may describe a vessel using different languages while referring to the same controlled object category. E55 Type and controlled vocabularies can connect a stable concept to language-specific expressions. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
Substituting a familiar but broader noun can merge a ritual vessel with an everyday drinking cup. Use approved labels tied to concept identifiers, and note when a cultural term has no exact equivalent. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Test classification filters in multiple languages and check that membership does not change. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
15. Materials and techniques
A historic photograph may involve albumen printing, while another image uses a modern digital process. Translate technique terminology at the correct technical granularity, retaining documented material distinctions. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
Calling both simply photographs may be acceptable in a public overview but not as a replacement for research metadata. Preserve detailed technique concepts and add accessible explanatory language in separate descriptive fields. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Review specialist search results to ensure a filter for one technique does not retrieve all technologies indiscriminately. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
16. Dimensions and units
An object record might report height, depth, mass and measurement method with different levels of reliability. The measured quantity and its method are evidence; a display language may adapt the unit presentation only deliberately. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
Swapping centimetres and millimetres or silently rounding can make a conservation measurement materially false. Keep numerical source values and unit references, use controlled conversion rules where authorized, and document any transformed display. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Compare multiple measured dimensions and their labels with the primary measurement record. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
17. Ownership versus custody
A borrowed painting may be owned by one institution, conserved by another and displayed at a third venue. Actor relations distinguish legal ownership, current custody, temporary responsibility and the location of exhibition. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
A translation that says owned by where the evidence says in the care of misstates rights and responsibility. Choose relationship wording according to the actual predicate and any date-qualified documentation. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Review the provenance and exhibition panels against loan records where access is authorized. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
18. Production versus modification
A clock made in one century may have been repaired, fitted with new components and photographed much later. The making event must not absorb every subsequent intervention into an apparently single manufacturing episode. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
A simplified statement such as made in 1950 can wrongly date an eighteenth-century object that was restored in 1950. Keep separate events and make their chronology visible without inventing undocumented steps. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Ask whether a historian can reconstruct the order of recorded interventions from the localized display. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
19. Attributions are evidence-bearing claims
A painting may be labelled workshop of, attributed to or formerly believed to be by a particular artist. Such phrases communicate degrees and histories of scholarly confidence rather than stylistic preference. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
Removing the qualification may transform a contested attribution into an authoritative declaration. Translate uncertainty markers precisely and preserve the source and date of each attribution statement. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Compare source and target answers to the question who is claimed to have made the object. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
20. E13 Attribute Assignment
An institution may record when an attribution was made and which scholar or committee made it. An assignment activity is distinct from the underlying fact it discusses; both can matter to later research. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
Presenting an assignment as if it were unqualified objective truth conceals the provenance of the assertion. Retain the attributing actor, date and source where the chosen model represents them. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Inspect whether a reader can see who made the claim and whether it has been revised. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
21. Documents and evidence trails
A letter, inventory book, excavation photograph and database note may support different parts of a heritage account. A document reference points to evidence, not necessarily to an event in which the document itself participated. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
Flattening every source into a single translated narrative can remove the ability to check competing accounts. Preserve document identifiers, page references, quotations and provenance before rewriting explanatory text. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Follow a claim back to its cited source in both languages and verify that the reference still opens. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
22. Provenance gaps
An old collection record might identify an owner in 1920 and a museum accession in 1970 but leave the intervening years unknown. Responsible translation must maintain the gap instead of supplying a plausible ownership story. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
A smooth paragraph can conceal missing evidence even while individual translated sentences sound reasonable. Use explicit wording such as recorded from or later documented at, with careful qualification. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Review the rendered timeline and confirm that no connection is shown as proven without supporting records. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
23. Rights and restrictions
A digital reproduction can have different reuse conditions from the physical object and its archival description. Rights labels must remain attached to the correct resource and should not be inferred from the object’s age. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
A translation that treats public display permission as unrestricted commercial reuse may mislead visitors and partners. Keep rights URIs, rights holders and qualified license text associated with their designated entities. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Open the public image page and compare the permission statement with the authoritative rights record. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
24. Language tags and fallback
A linked-data endpoint might return English and French labels but have no approved term in another requested language. A language fallback is a display decision, not evidence that an absent translation exists. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
If the interface silently labels a source term as a translation, readers may mistake it for a reviewed equivalent. Show approved labels when present and describe the fallback convention honestly where needed. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Test missing-label cases, mixed-script names and embedded quotations rather than only fully translated records. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
25. Broader terms versus exact matches
A cultural concept may be related to another vocabulary’s class without sharing exactly the same scope. Mapping terms across ontologies requires semantic comparison, not merely matching their English headings. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
An unjustified exact-match link can propagate incorrect classifications into every participating catalogue. Use carefully reviewed mapping relations and document why the chosen strength of equivalence is defensible. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Run example queries at both ends of each mapping and inspect false matches. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
26. RDF triples and graph shape
A CRM statement can be expressed as a subject, predicate and object with separately identified resources. Translating a display string should not change the underlying triple or graph connectivity. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
Moving a relation from one event node to another may be syntactically valid while rewriting history. Validate graph identifiers, statement counts and relevant relation paths across localized representations. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Ask whether a machine query returns the same participating actors and places before and after translation. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
27. SPARQL query equivalence
Two readers may ask in different languages which objects were produced by a particular workshop. The language filter may change which label is returned; it should not alter the set of identified works. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
A localized query that accidentally filters on literal display text can drop relevant records or duplicate results. Query by stable identifiers and use language-aware selection for descriptions and titles. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Compare entity counts and sample result identifiers across locales, not just visible headings. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
28. Serialization and escaping
A graph may be delivered as RDF/XML, Turtle or JSON-LD with language-tagged literal strings. Markup punctuation and syntax markers have technical jobs that translation cannot override. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
Changing namespace prefixes, braces or encoded quotation marks as ordinary punctuation can make a valid model unreadable. Restrict language editing to literal content explicitly designated for localization and validate each exported serialization. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Parse the target output with a standards-compatible parser, then inspect representative literal values. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
29. Authority updates and corrections
A museum may later replace an obsolete place-name label or revise an attribution after further research. A multilingual record needs a revision trail that separates new scholarship from new language wording. A translation can change what a visitor reads, but it must not change the answer a researcher could recover from the underlying evidence. Treat the technical relationship as part of the meaning, not as expendable background.
Overwriting an earlier claim without a date can make an archived citation appear to support a different statement. Keep change dates, sources and correction reasons; update dependent translations only after authoritative review. A useful working comparison keeps the source wording beside the target expression and shows which record identifiers or linked facts justify the interpretation. The best wording is clear without claiming more than the record supports.
Reopen older citations and compare whether their historical context remains discoverable. Test the public display as well as the stored data. Sometimes the mapping is sound but a search snippet, interface label or aggregated timeline tells a different story.
30. Community perspectives and contested names
A culturally significant artifact may have an institutional name and additional names supplied by source communities. Respectful language work distinguishes documented voices, authority decisions and contested or historical designations. This is a difference between a linguistic form and a factual assertion. The reader can receive a natural explanation while the model keeps the evidence in its original relationships.
Erasing all older terms may hide evidence, while presenting a harmful historical term as neutral current usage can distort context. Record preferred public labels with provenance and note when a term appears only in a historical quotation. Solve that problem at the correct level: do not change an identity-bearing field merely to improve a sentence, and do not preserve a misleading sentence simply because the underlying data validates.
Review the public wording with relevant communities or curatorial expertise when appropriate. Try the same question in both languages and compare the real objects, dates and people returned. A satisfactory translation gives equivalent answers, not merely equivalent-looking words.
31. Interoperability with LIDO records
LIDO exchange records and CRM event graphs may describe the same museum object for different purposes. The interchange schema and conceptual ontology can complement each other without being identical representations. Professional localization becomes safer once the purpose of each element is visible. A title, type, event and source citation do different jobs even when a website displays them together.
A simplistic one-to-one field conversion may collapse event distinctions and create misleading mappings. Trace object IDs, events, people, places and rights through a small sample rather than assuming every field corresponds directly. If a term has more than one possible rendering, record the reason for the chosen expression and the domain specialist who resolved it. This avoids a second translator silently reversing the decision.
Compare the exported LIDO portal page with the CRM-based institutional knowledge graph. Carry the check through export, indexing and presentation. A technically correct file can still mislead someone who only reads the search result or a shortened caption.
32. A complete release test
A multilingual heritage platform can pass its XML and RDF checks while displaying a wrong maker or date. Structural validity and historical correctness require separate tests and accountable specialist review. The practical standard is not word-for-word sameness; it is preservation of the same documented relationship across languages. A concise expression may be appropriate if its qualification survives.
Shipping without checking the rendered interface can preserve flawless machine data beside a misleading visitor-facing summary. Use a release sample containing uncertain attribution, multiple events, older names, missing translations and restricted imagery. Compare a representative real record before applying a rule to thousands of objects. If the case exposes uncertainty or competing meanings, separate the editorial question from the file-format question.
Record pass or fail against source evidence, then verify the repaired page after it is republished. Then verify the change with the people who own the evidence, not only with those who manage the interface. A polished catalogue is trustworthy only when its claims remain traceable.
Worked situations: diagnose the failure before translating again
An excavation site becomes a production place
A curator records a bowl’s production in one region and excavation in a different country. A translator shortens both relations to the same word for place. Retain the two event nodes and the roles connecting each to its location. The translated public text should distinguish where the bowl was produced from where archaeologists found it. Reopen the two map views and make sure their dates and captions refer to different activities. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A workshop is rewritten as a master artist
A catalogue says a sculpture was produced by a named workshop, with no documented individual maker. A localized label names the celebrated workshop founder as the sculptor. Preserve the collective actor and the uncertainty about individual authorship. Explain a workshop relationship without silently claiming a particular person’s hand. Check both the actor authority and the provenance statement before approving the label. The repair should be tested in the exported graph and in the display that a visitor actually uses.
Two names create duplicate people
An artist’s name appears in several scripts and one export replaces its persistent URI with a transliterated label. Restore the shared identifier and store the name variants with documented language or script information. Do not solve the issue by deleting historically useful names. Query the graph using each displayed name and confirm the results identify one actor. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A restoration date becomes a creation date
A painting underwent conservation in 1984, but its historical production date is much earlier. A merged translated timeline makes 1984 appear beside made. Restore distinct event labels, dates and actors. The conservator’s role remains important, but it must not displace the maker or period of origin. Compare each localized timeline entry against the corresponding event resource. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A claim becomes stronger than the evidence
A source database says probably attributed to a named artist. A fluent target-language caption drops both probably and attributed. Reinstate the uncertainty and clarify whether the statement comes from a cataloguer, a published study or a later museum review. Read the claim alongside its citation; the level of certainty must match. The repair should be tested in the exported graph and in the display that a visitor actually uses.
An image inherits the object’s legal status
An old artifact may be out of copyright, but a newly produced photograph is associated with separate licensing conditions. Keep the physical object, digital image and each applicable rights statement distinct. Translate the permission explanation instead of inferring freedom of reuse. Verify links and rights text on the actual download and display pages. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A material label merges incompatible processes
A search facet groups photographic prints made by two different historical processes because both were translated using one convenient general term. Keep the distinct controlled concepts and, if useful, add explanatory descriptions rather than merging their IDs. Compare facet results in both languages and investigate any unexpected change in record counts. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A graph parser passes a false account
Every RDF triple in a translated release is syntactically valid, but one actor relation now points to the collector instead of the maker. Compare the source and target graph paths for the affected object, restore the correct actor link, and add a semantic regression test. A parsing success message is not proof that a historical statement remains true. The repair should be tested in the exported graph and in the display that a visitor actually uses.
A practical hand-off and release checklist
Collect the source records and record the implemented ontology version before a translator changes wording. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Locate the identifiers, relation predicates, dates, event roles, controlled types, rights and source citations. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Mark translatable literals separately from immutable resource identifiers and structural syntax. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Document authority decisions for persons, groups, places and culturally contested names. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Build a compact test set covering missing labels, uncertain attributions, multiple events and overlapping places. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Use language-tagged labels when supported, and specify how a requested language falls back to another language. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Validate output files as RDF or the applicable serialization without treating parser success as semantic proof. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Compare graph queries and public results across language versions using stable identifiers. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Ask a curatorial or domain expert to review historically consequential attributions and rights statements. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Publish with a correction route so that new evidence can update labels without concealing past claims. When the result is unclear, preserve the source relation and escalate the interpretation rather than guessing.
Frequently asked questions
Is CIDOC CRM a translation file format?
No. It is a conceptual ontology for describing cultural heritage information. Actual data may be expressed in graph serializations and linked to multilingual labels.
Which release should an implementation cite?
Document the release actually used. As checked in October 2026, the official release is 7.1.3; newer 7.3 and 7.4 documents are listed as drafts.
May ontology class identifiers be translated?
Do not translate the machine identifiers in a stored graph. You can provide reviewed explanations of their concepts in different languages.
Can two names refer to the same person?
Yes. Alternative spellings and language variants may point to one stable actor identity while retaining their sources and historical context.
Does a valid RDF file guarantee factual accuracy?
No. Parsing and schema checks do not prove the event roles, dates, rights or attribution in a heritage statement are correct.
Should every historical place name be modernized?
No. Keep source-era context and authority identity; add a modern explanatory name only when appropriate and clearly labelled.
How is CIDOC CRM different from LIDO?
CIDOC CRM provides a conceptual model for relationships and evidence. LIDO is an exchange schema for museum object metadata. They can be related without becoming interchangeable.
How do we handle missing language labels?
Use a transparent fallback policy, preserve the source form and never imply that a fallback string is a reviewed translation.
What is the most important review question?
Would a researcher comparing the translated page to the underlying evidence find the same actors, events, places, dates and qualifications?
Who should approve an uncertain historical term?
A language specialist should work with the relevant curator, researcher or community expert rather than making an unsupported editorial decision.