How do you translate Linked Art 1.0 museum records without changing which object is being described, who made it, where it travelled or what the evidence actually supports? Linked Art is an interoperable way to publish cultural heritage information as linked data, using a profile of CIDOC CRM and JSON-LD. Its multilingual power comes from treating people, objects, concepts, events and statements as connected resources rather than as one large paragraph awaiting translation. That distinction matters when an artwork has several titles, disputed attribution, changing ownership or a history documented in several languages.
People looking for Linked Art JSON-LD localization, multilingual museum linked data, cultural heritage API translation, CRM entity identifiers, art provenance graphs or language-tagged names are usually trying to solve two problems at once. They want visitors to encounter understandable descriptions in their own language, while software continues to retrieve the same object, actor and event. If a translator replaces a persistent identifier with a translated word, merges two roles, or recasts an uncertain claim as a fact, fluent prose can silently damage the data model.
This practical guide explains how to translate the human-readable parts of Linked Art records, protect structural relationships and review published results. It addresses museum registrars, cataloguers, curators, software teams, translators, art historians and readers who may never have written JSON-LD before. The focus is a professional workflow for meaning preservation, not a substitute for the published Linked Art 1.0 model or an institution’s evidence, rights review and governance. We will move from ordinary descriptions to linked entities, work through failure cases and finish with release tests a multilingual publishing team can actually perform.
How Linked Art makes multilingual collection records more trustworthy
Think of a museum catalogue as a map of claims rather than a book of captions. A reader sees an artwork and a short description; the map behind that page connects the artwork to its maker, materials, history, images and cited evidence. Translation adds language access to that map. It must not redraw the roads. The following decisions show where professional localization can improve communication while protecting the relationships that make linked cultural data useful across institutions.
Forty decisions for translating Linked Art 1.0 accurately
1. Begin with the object, not the sentence
A conventional catalogue description can hide several kinds of assertion inside a single sentence: an object exists, a maker is proposed, an event is dated and a location is associated with it. Linked Art separates these claims so they can be connected to other records. Before translating any label, ask whether the data concerns a physical object, a visual or textual work, a digital representation, an event, a person or a concept. Those are different targets. A natural phrase that describes a digital photograph should not become a claim about the physical object itself. Record the node being translated before discussing the wording.
2. The job of a stable identifier
An identifier answers which resource a record means. A label answers how a person reads it. A museum may display a painting’s title in Spanish while retaining the same object URI that identifies it in English. Replacing the URI with translated words breaks references from loan records, exhibition sites, research citations and other institutions. Identifiers may be opaque and aesthetically unhelpful; they are not supposed to perform the job of a title. Keep identifiers and references fixed unless the institution intentionally migrates its identity system through an approved process. Translation alone gives no authority to create a replacement identity.
3. Linked Art is a profile, not every possible CRM statement
Linked Art chooses practical patterns from a broader conceptual framework. A catalogue team accustomed to the whole CIDOC CRM ontology may expect every theoretically possible relationship to be available through the Linked Art profile. A translator should not invent absent predicates to make a sentence fit. First check the specific pattern and version the publisher implements, then translate the human-language values that the pattern permits. If a source statement cannot be expressed faithfully, record an editorial question or request a model decision. Silent invention creates a record that looks interoperable while being semantically different from what the source system knows.
4. Why JSON-LD matters to language
JSON-LD looks similar to ordinary JSON, but its context maps compact properties to linked-data vocabulary. This means a field name or type can denote a specific relationship rather than simply describe a visual interface. Translate text-valued names and statements where the model expects natural language; do not translate JSON keys, control vocabulary, identifiers, booleans, structural punctuation or context definitions. A label may be readable in Japanese and French while a machine sees the same underlying class. Review a localized file as data and as rendered prose, because either side can appear correct while the other is broken.
5. Names are records in their own right
An object’s preferred title is not necessarily a literal attribute that can be overwritten for each language. Linked Art’s basic patterns represent names and identifiers as distinguishable data. A name may have a type, linguistic context or source. Retain original-language names when the material identity requires them, and attach translated labels as additional human-readable expressions where appropriate. A translated exhibition title should not erase an inscription recorded on the object. Museums may also have historical titles whose wording is troubling; present contextual editorial language without silently falsifying the documentary title by replacing it everywhere.
6. Distinguish names from identifiers
A registration number, accession code or externally assigned authority identifier is not a name merely because it appears next to a name in a visitor interface. If a data import translates an accession prefix, it may cease to match internal collection management records. Conversely, presenting only an opaque identifier as a visitor-facing heading harms comprehension. Maintain each in its proper role: identifiers for machine and administrative continuity; language-specific display names for reading; explanatory notes for difficult cases. Testing should compare the pre- and post-localization identity graph as well as the public labels.
7. The use of classified terms
Linked Art uses controlled concepts and classifications to distinguish kinds of things. For example, a display label might say drawing, while the underlying term points to a managed art vocabulary concept. Translating the English label does not authorize assigning a different concept URI. Examine the definition, scope note and broader or narrower relationships before selecting an equivalent term. If the destination language lacks a precise established term, document the near-equivalence or use an explanatory phrase. A beautiful translation that changes the class of an object can distort searches and statistics across a museum’s entire collection.
8. Concept labels are not object descriptions
A controlled concept may be shared by hundreds of objects. Its name should follow the vocabulary’s language practices; an individual object description may be more contextual or more tentative. Translators sometimes apply an artwork’s poetic wording as the vocabulary label, then unintentionally change the category that aggregates unrelated works. Keep concept-level terminology stable and object-level narrative flexible. If a proposed label would make sense only for one unusual item, it probably belongs in that item’s note rather than as the primary name of the common concept. This distinction supports both accessible prose and meaningful multilingual faceted search.
9. Human agents and their roles
The same person can be an artist, collector, seller, lender, restorer, photographer or donor at different moments. A short biography may compress these roles, but linked data needs the actual event or relationship to remain intact. When translating agent records, verify whether a role is asserted about a production, custody transfer, exhibition or another activity. An owner is not automatically a maker, and a seller is not always the legal owner. Role labels should communicate the capacity in which someone participated, not inflate it into an unrelated credential or responsibility.
10. Organisations and institutional names
Institutions often have legally registered names, local-language names, former names and widely recognised short forms. They are not necessarily interchangeable. Keep the resource identifier stable, record the institution’s preferred translation when available and preserve the relevant historical name where a transaction occurred under that name. A merger, renaming or relocation can be a historical fact rather than a simple language variant. If an archive says an organisation held an item in 1932, do not retrofit today’s corporate name into the event as though that were the actor at the time. Readers benefit from a modern explanatory label, but the recorded historical relationship must survive.
11. Production is an event, not just an adjective
A painting may be described as made by an artist, produced in a workshop, printed after an original design or assembled during a later restoration. Linked Art can represent production through events and participation. This matters because a translator tempted to condense a production statement may turn workshop of into created by, or after into by. Each small phrase changes the attribution. Retain the distinction among creator, attributed creator, workshop, school, later reproduction and uncertain association. When evidence is incomplete, preserve the source’s degree of certainty and link the statement to its authority where possible.
12. Time spans and uncertain dates
Art-historical records may contain circa dates, broad periods, reign names, revised datings or explicit ranges. A cultural calendar label is not a mathematical substitute for the dates encoded in a time span. If a heading says early seventeenth century, do not give it a precise Gregorian year without a documented conversion. A structured earliest and latest boundary may be more useful than a single display date, but such boundaries must come from evidence or documented editorial rules. Translate the readable date explanation while respecting any machine-level ranges and distinctions between event time and cataloguing time.
13. Places and geography over time
A current country name is not always the correct description of a historical place. A work made in a city before modern borders were established may need a historic place label, a present-day locator and notes explaining the relation. Keep the place resource and its historical associations consistent; a translation cannot move an object to today’s administrative unit merely for convenience. Distinguish where an object was made, found, purchased, kept, exhibited or photographed. The same place can have several language names, while several different places may share a translated name. Geographical clarity requires identifiers and historical context together.
14. Physical objects and their digital images
A physical sculpture, a conservation photograph and a web thumbnail are different resources with different creators, rights and technical properties. A museum might own the sculpture but license only a specific digital image, or permit educational display without permitting downloads. Translating ‘image available’ into ‘artwork free to use’ collapses these boundaries. Linked Art provides ways to describe digital content, and a publishing pipeline must distinguish content descriptions from object descriptions. Review each human-facing label against the actual resource it accompanies, especially in multi-asset records where one object has several digital representations.
15. Visual work, physical carrier and inscription
A manuscript’s physical pages may carry a written text that has a life beyond that particular copy. A painting may depict a textual inscription whose words are part of the artwork, not editorial commentary. Translation can provide a transcription, a translation of the inscription and a catalogue explanation, but they should not be presented as identical. Where a literal transcription is the evidence, retain it and provide an accessible translation alongside it. If an inscription is damaged or disputed, mark conjecture rather than manufacturing a clean text. These layered distinctions protect scholarship while allowing wider readership.
16. Subject versus depiction
A portrait can depict a named person, embody an allegory and be classified by artistic genre. These are not identical claims. A source phrase such as portrait traditionally identified as a ruler should not become verified depiction of that ruler. When translating aboutness records, ask whether the object depicts, refers to, symbolizes, is classified as or is interpreted to concern the target. Even a familiar artistic term may conceal a contested interpretation. Preserve the relation type and the source of the reading; a fluent description that overstates certainty can spread misinformation through linked searches.
17. Provenance as a sequence of events
The history of an object’s custody or ownership is often assembled from separate events rather than a single unbroken chain. Translate the recorded transfers and their supporting notes without filling documentary gaps. A catalogue may distinguish physical custody from ownership, as well as acquisition, loan, gift and bequest. Treat each term as a different kind of assertion unless evidence shows otherwise. If one link is unknown, write that clearly rather than making adjacent known events look continuous. Provenance translation is especially sensitive because it can intersect with restitution claims, legal rights and the evidentiary limits of old records.
18. A loan is not an acquisition
A museum can receive an object, catalogue it, insure it and display it while never acquiring title. In localizing Linked Art records, the user-facing words received, purchased, borrowed, lent and owned must remain connected to the correct activities and parties. Avoid a target-language verb that vaguely covers several transactions when the dataset needs to distinguish them. If the local language normally uses one broad term, add a concise qualifier about the legal or operational relationship. The point is not to impose English syntax, but to prevent the reader from inferring an ownership transfer that did not occur.
19. Custody and movement events
An object might move from one storage room to another without any change of ownership. A transport company can hold physical custody for a short interval without becoming the collecting institution. Translate movement notes with attention to origin, destination, dates, parties and status. ‘Scheduled to move’ is not equivalent to ‘moved’; ‘received in transit’ is not equivalent to ‘received into the collection’. Where the model records only a confirmed event, do not insert planned or cancelled movements as though they happened. A multilingual interface should make it easy to tell historical action from workflow intention.
20. Attribution and competing interpretations
An expert may attribute a painting to one artist while a later catalogue favours another. Linked data can preserve multiple statements, their sources and varying confidence, but a translation may hide this complexity by choosing the most familiar name. Keep separate assertions separate. Translate labels such as attributed to, probably, formerly attributed to and contested with their real epistemic strength. Do not replace a disputed person’s identifier with that of a more famous figure simply because a common translated name is ambiguous. A user should be able to see where knowledge ends and interpretation begins.
21. Statements and their evidentiary basis
An ordinary sentence often sounds authoritative because it is grammatical. A linked-data statement may be explicitly sourced to an inscription, sale catalogue, research paper or curatorial assessment. Translators need to know whether a claim is being quoted, summarized or endorsed. Retain bibliographic references, evidence relationships and statement types. A quote from an old auction advertisement may be inaccurate by today’s scholarship; translating it as the museum’s present position would mislead readers. If the source record offers an evidentiary distinction, surface that distinction in the translated language rather than flattening the record.
22. Historical quotations and harmful terminology
Cultural records can contain discriminatory, colonial or outdated terms that remain relevant as documentary evidence. A translation should neither amplify harmful wording without context nor silently erase the fact that the historical source used it. Separate literal transcription, faithful translation and explanatory editorial language. Label quoted historical terms as quotations and provide contextual notes when appropriate. A public-facing interpretation can choose inclusive current language while still preserving the evidence and provenance of the original phrase. This approach gives readers a safer, more accurate account than pretending the offensive wording never existed.
23. Named entities across scripts
Romanisation, transliteration and translation are different actions. A Japanese artist’s name may appear in Japanese script, in one established romanisation and in several alternative catalogues. The variations can point to one identified agent, but they should not be treated as distinct people simply because strings differ. Conversely, two different people can share the same romanised name. Keep identifiers, original scripts and documented variants together. A search index may use aliases broadly, while the scholarly display chooses a preferred form. The translator’s work is to improve recognition without changing entity identity.
24. Rights are not one universal permission
Object rights, photograph copyright, metadata licensing and website terms can differ. In Linked Art projects, a data consumer may receive a machine-readable rights statement for one asset and explanatory text for another. Translate the statement faithfully but retain its source, scope, URI and jurisdiction where relevant. Do not turn ‘may be viewed online’ into ‘public domain’, or ‘rights not evaluated’ into ‘free of restrictions’. If a platform must display language-specific explanations, they should clarify the rights resource rather than create a competing permission. Legal uncertainty belongs in a review queue, not in invented confident wording.
25. Conservation events and treatments
A condition note describes an observed state; a treatment record describes an action taken; a proposal describes an intended intervention. Those are three different categories of fact. An object may be described as stabilized after a past treatment without being judged suitable for new handling today. Translation must preserve who observed, who treated, when and under what limitations. Technical conservation terms often require discipline-specific review, because an everyday synonym can reverse the meaning of a material process. Avoid converting a condition observation into an instruction or treating an unapproved proposal as completed work.
26. Exhibitions and display histories
The exhibition in which an object appeared is an event, not an intrinsic property that lasts forever. A displayed title, gallery location and loan status may be valid for one exhibition period only. Translate exhibition names with an approved institutional form when one exists, and preserve historical programme titles where they function as cited evidence. If a label says an object was planned for an exhibition but withdrawn, do not publish it as exhibited. Readers researching object histories rely on the difference. Dataset releases should test that exhibit events, participants and dates still align after localization.
27. Collections and membership
Membership in a collection, authorship of an object and possession by an institution are distinct relationships. A work may be presented in a virtual exhibition collection without ever being owned or physically stored by the host institution. A translation that says ‘from the museum’s collection’ may imply ownership where the record only documents selection or association. Explain the relevant collection relationship in accessible language, but do not silently change the graph to express possession. When working across partner institutions, record whose description is being translated and which collection membership is asserted by the source.
28. Cataloguer voice versus source voice
Museum metadata can combine a curator’s prose, a donor’s description, a conservation expert’s note and a historical quotation. A multilingual page should not make all four sound like one omniscient narrator. Keep quotations, citations and attributed assessments recognisable. If style guidelines require consistent voice, apply that polish only within the editorial layer. The source’s degree of certainty, terminology and historical viewpoint remain relevant. This is especially important when one language edition is written for children and another for academic readers: readability can vary, but the claims and their provenance should not.
29. API endpoints and stable references
Linked Art’s API provides conventions for accessing linked resources. When a record has a canonical object endpoint, its URI should not change simply to reflect the language of the page. A site may choose language negotiation, additional display pages or separate translation resources, but the design must preserve entity identity. Test links from an artwork to an artist and from a provenance activity to its participants. Do not assume that replacing English words in URLs creates supported API paths. The API’s documented behaviour, not an editor’s guess, determines valid resource references.
30. Search and browse are not identical
A visitor might search for an artist in Arabic yet encounter a preferred French display title for an artwork. Search aliases, transliterations and translated labels can help retrieval without redefining the underlying entity. A browse hierarchy, however, depends on controlled classifications and relations. If someone translates a term into a broader category merely to improve recall, the faceted results become inaccurate. Evaluate search recall, precision and display comprehension separately. The best multilingual experience allows people to find relevant records through several linguistic paths and still arrive at the same correctly identified object.
31. Multilingual notes and readability
Notes are useful when names, dates or attribution cannot be rendered clearly through short labels. Write them as explanatory sentences for an international reader, not as private shorthand between specialists. Define uncommon technical terms the first time they appear. Separate what the object is from what researchers think about it, and mark conjecture openly. A translation should read naturally in the target language, yet never claim more certainty than the source supports. Editorial clarity is an additional layer of communication, not permission to edit the evidence underneath.
32. Language tags and directionality
A translated label should carry the correct language metadata when the publishing pattern supports it. A language tag that says English for Arabic text can cause screen readers, search indexing and font selection to behave incorrectly. Right-to-left scripts may require layout handling without changing the order of data properties. Names in mixed scripts need particular care because numerals, punctuation and embedded Latin identifiers can display awkwardly. Review actual rendered pages with native readers and assistive technology, rather than assuming that a valid JSON serialization guarantees a usable visitor experience.
33. Transliteration is not authority reconciliation
Two records with slightly different spellings are not automatically the same entity. Proper authority reconciliation examines source identifiers, dates, relationships, institutional records and other evidence. A translator can identify candidate equivalences and supply alternative names, but should not create same-as links on spelling alone. An erroneous equivalence can merge two artists’ works throughout a distributed graph and can be difficult to undo. Record uncertainty and request specialist approval where identity is unresolved. The translation team’s priority is readability while preserving the identity evidence that makes linked data reliable.
34. Version changes and model compatibility
A publisher might use a profile aligned with Linked Art 1.0 while an internal source system evolves independently. Translation should not silently force an upgrade or introduce fields unavailable to the current service. Keep a version record, test the supported properties and confirm JSON-LD context expectations. If documentation and the deployed API differ, describe the observed implementation as such and raise a technical issue. Release notes should identify material changes in labels or interpretation without rewriting historical events. Version discipline protects consumers who depend on stable machine-readable outputs.
35. Validation before publication
Mechanical validation checks whether JSON parses, whether required fields follow the supported profile and whether links resolve. Editorial validation asks a different question: does the translated label express the same claim? A successful parser cannot detect that ‘loan’ became ‘purchase’ or ‘possibly made by’ became ‘made by’. Use both classes of checks. Sample records across object types, actor roles, disputed attributions, languages and rights states. Compare important identifiers and relationships against the source, then review the localized display with subject specialists who can recognise nuanced art-history terms.
36. Handling incomplete and missing data
Unknown does not mean none, and an empty field does not authorize a translator to manufacture a likely value. A missing date, unspecified donor or unconfirmed place may be an honest limit of the repository. If a translation interface requires a fallback, label it as unspecified or not recorded according to the institution’s editorial policy. Do not infer from adjacent graph nodes; a collector known to own an object at one point does not prove who sold it later. Preserve missingness distinctly from a documented negative assertion, because researchers often investigate precisely these gaps.
37. Source language and relay translation
When original labels pass through an intermediate language, meanings can drift. A Japanese inscription translated into English and then into Italian is not equivalent to translating the primary evidence from Japanese. Identify the source language, the existence of earlier translations and which version is authoritative for each field. Where primary-language expertise is unavailable, say which secondary version was used and mark uncertainty appropriately. A relay translation can be valuable for access, but it must not masquerade as a direct scholarly transcription or hide conflicts between competing editions.
38. Accessibility and spoken labels
A multilingual database should serve people who use screen readers, speech interfaces or simplified text. An accessible description can explain an artwork without replacing the scholarly object record. Avoid reading opaque URIs aloud as substitutes for names, and ensure that abbreviations in collection captions are intelligible when spoken. Dates and ordinal identifiers can have unexpected pronunciation across languages. Test the title, artist name, rights notice, navigation and evidence notes together. Accessibility is not merely a front-end decoration; it is a test of whether public-language layers distinguish the things being described.
39. Interoperability with neighbouring standards
Linked Art can connect to CIDOC CRM, Getty vocabularies and IIIF experiences, but it does not replace every museum procedure, cataloguing schema or imaging protocol. An IIIF canvas is not automatically a physical artwork, and a Getty concept is not a free-text title. Preserve identifiers at interoperability boundaries instead of translating a whole record into a single flat string. When exporting to another model, document which distinctions survive, which become notes and which require additional data. The goal is faithful information movement, not pretending all cultural heritage standards describe the same kind of entity.
40. Planning editorial approval
Before translation begins, an institution should name the owners of terminology, art-historical attribution, rights language and technical publishing. A language expert may approve idiomatic phrasing but lack authority to rewrite provenance. A curator may assess interpretations but not legal permissions. A developer can validate schema structure but not cultural nuance. Identify which decisions need which reviewers. A short ownership table prevents final-day disagreements over whether a revised word is a harmless style change or a different assertion about an object. Clear decision rights speed publication without reducing rigour.
41. Release and rollback strategy
Treat a multilingual data release as a versioned publication. Retain a source snapshot, translation decisions, validation results and records of approved exceptions. A rollback should restore labels and related content without quietly changing stable IDs. If an error becomes public, correct it transparently with the institution’s normal editorial process; do not erase evidence that a prior interpretation existed when that history is relevant. Consumers may cache resource descriptions, so plan for propagation and explicit correction. Successful localization means the same graph remains trustworthy when users view it through different languages and applications.
Worked localization cases: diagnose the broken relationship
A workshop attribution loses its uncertainty
Suppose a source record reads attributed to the workshop of an identified painter. A destination editor wants to shorten the public title to painting by the painter because the longer phrase seems repetitive. That edit converts a proposed relationship into direct authorship. The correct repair is to preserve workshop attribution in the role statement and explain it in natural language: the work is associated with the painter’s workshop, rather than personally confirmed as painted by the artist. Check whether the linked production event assigns an agent, workshop or uncertainty statement before sign-off.
Two people share the same romanised name
A catalogue describes two different historical figures whose names become identical when transliterated. A search tool groups their records and a translator proposes one unified display entry. The error is not merely cosmetic: events involving one person will now be attributed to the other. Keep the distinct agent identifiers, original scripts, biographical context and alternate names. Search can show disambiguating dates or roles. Merge only after authority evidence supports identity. The localized interface must help readers distinguish the people while leaving the underlying actor graph unchanged.
The image licence is mistaken for ownership
A visitor sees a high-resolution photograph of a carved object and a caption indicating permission for online display. An enthusiastic translator renders the notice as the museum owns all rights to this artwork. That statement could be false about the object, the photograph and other derivatives. Find the specific right or licence statement, identify which resource it governs and translate its permitted uses accurately. If the rights language is ambiguous, escalate it rather than promising broad reuse. Test the notice beside the exact image viewers will download or share.
A city name is modernised into the wrong event
A historian records that an artefact was manufactured in a city known by a former political name. The target-language version replaces that name with a contemporary administrative label and omits the historical qualifier. A reader may now assume the current state existed at the time of manufacture. Retain the historic toponym as evidence where appropriate, provide the modern name as contextual orientation and maintain the same linked place resource if that identification is established. Explain uncertain geographic identifications instead of silently forcing them onto present-day maps.
A planned exhibition is recorded as completed
The catalogue says a painting was proposed for a travelling exhibition but withdrawn before opening. A compressed translation says exhibited internationally. This invents a public appearance and can contaminate provenance narratives or exhibition histories elsewhere. Check the event status and documentary evidence, and distinguish proposal, invitation, inclusion in a catalogue and actual display. When an exhibition resource exists, its relation to an artwork must express the observed historical event rather than a plan. A concise yet accurate label could say selected for the exhibition but not displayed.
A donor becomes a seller
An acquisition note mentions an object passing through a private collection and later entering a museum as a gift. One language version uses a verb associated with sale, confusing the donor with a commercial seller. The repair begins with the event type: gift, purchase, deposit and loan must not become interchangeable. Preserve identities of participants and whatever consideration, if any, was recorded. If no price exists, do not invent one. Later readers may use this record for provenance and legal research; the transaction’s nature matters far more than elegant sentence rhythm.
A historic title becomes a present-day catalogue claim
An old sales catalogue used a title that scholars now consider inaccurate. The museum’s source data stores that title as a historical name, alongside an updated display label. A translator merges the two into one definitive modern title and removes the historical attribution. This erases the reason researchers can match an old catalogue to the current object. Keep the earlier title with its documentary status and date where known, while presenting the current title as the museum’s chosen display form. Conflicting names are often evidence, not editorial mistakes.
An empty date becomes an approximate year
A data export lacks a production date but contains a later acquisition event. To avoid a blank card, the translation team uses the acquisition year as the artwork’s date. This is an inference from the wrong event and may be centuries off. Distinguish production, acquisition, publication and exhibition dates; each describes a different historical action. If production is unknown, display an honest unknown or not recorded label. Add a date only when an accepted source supports it and the correct event relation can be represented.
The label is correct but the link points elsewhere
A localized page shows the translated title of the right painting, yet the API object URI points to a different work after an automated string replacement. Human proofreading alone may miss the mismatch if both paintings share an artist. Validate identifiers, linked-agent references, image references and event targets against the source record. Use a sample that includes near-duplicate titles and multiple editions. The release should fail when core identity links change without explicit approved remapping. Machine consistency and reader-friendly language are parallel requirements.
A curatorial hypothesis sounds like a verified fact
A note proposes that a vase may have been used in ceremonies. In translation it becomes the vase was used in ceremonies. The change seems minor, but converts interpretation into evidence. Recover the modal language, its supporting research and any alternative readings. A good public description can still be clear: researchers have suggested a ceremonial use, but the function is not established. Where the hypothesis has a named author or publication, preserve that attribution. Visitors deserve to know which statements are observed and which remain scholarly arguments.
A practical approval checklist
For each release, select representative multilingual records and compare source and target identities, object-to-image links, production roles, acquisition and custody events, temporal qualifications, geographical context, classifications, provenance sources and rights scopes. Ask a language reviewer to flag unnatural prose, a domain reviewer to flag altered claims, and an engineer to validate the JSON-LD and API references. Archive the review decisions and the data version. Where one kind of review has not happened, record that limitation rather than implying comprehensive approval. The best release is not one without questions; it is one whose important questions are visible, owned and resolved before visitors rely on the information.
Frequently asked questions
Is Linked Art the same as CIDOC CRM?
No. Linked Art is a practical application profile drawing on a streamlined subset of CIDOC CRM, with established patterns for cultural heritage publication. Consult its profile rather than assuming the full ontology is implemented.
Should I translate JSON-LD field names?
No. Translate the intended human-readable content while retaining structural keys, identifiers, context mappings and machine-controlled values according to the model.
Can different languages share one object identifier?
Yes. That is a major benefit: a stable identity can have multiple appropriate language labels. A translated title does not require a new physical-object identity.
What if the source contains a historical error?
Preserve the original evidence record and raise a documented correction with the responsible curator. Editorial corrections are distinct from translation and may need to preserve the history of interpretation.
Do provenance records prove continuous ownership?
Not automatically. They may contain gaps, assumptions and partial evidence. Translate those limitations faithfully and do not infer undocumented transfers.
Can an artwork image have different rights from the object?
Yes. A digital photograph, a physical object and the metadata describing them can have different rights contexts. Make the resource scope explicit.
How should uncertain authorship be translated?
Keep qualifications such as attributed to, workshop of, possibly and formerly attributed to. Do not promote a hypothesis into direct authorship.
What is the most important technical test?
Confirm that object, actor, event, concept and rights links still resolve to the intended resources, alongside validation of JSON-LD structure and multilingual display.
Does a machine-valid record guarantee an accurate translation?
No. Structural validity cannot catch a change in legal status, time, attribution or meaning. Native-language editorial review and specialist checking remain essential.
Where should a new language version live?
Use the publishing architecture chosen by the institution, following Linked Art API and profile conventions. Add readable language layers without improvising new identifiers or unsupported URLs.
Official guidance and further reading
Start with the Linked Art Model 1.0, the basic patterns for names and identifiers, the Linked Art API 1.0, and the CIDOC CRM. These primary references define the relevant patterns; a translated article cannot replace their technical requirements.
For connected museum data problems, see the related eduKateSG guides to CIDOC CRM event graphs and multilingual labels, VRA Core visual resource records and Spectrum museum procedures. Each addresses a neighbouring responsibility; none should be confused with Linked Art resource identity and linked-data publication.