How Metadata Works | Metadata Crosswalks — What Gets Lost When One Schema Is Translated Into Another

Database A has a field called author.

Database B has a field called creator.

Easy.

Map one to the other.

Then someone asks whether a photographer, editor, software generator, corporate body and translator all count as “author.”

A metadata crosswalk is a mapping between description systems. It can preserve useful meaning across schemas, but it cannot create semantic equivalence where the source and target never meant the same thing.

This is the third pillar beneath How Metadata Works. The master introduces schema translation. This article owns the lossy boundary: exact mappings, broader and narrower mappings, conditional translation, cardinality mismatch, controlled-vocabulary mismatch and the cases where a responsible crosswalk must say “no faithful target exists.”

Quick Read

A metadata crosswalk maps properties, values or structures in one metadata schema into another. Some mappings are close to exact. Others are broader, narrower, one-to-many, many-to-one or conditional on context. Translation can lose granularity, vocabulary identity, ordering, cardinality, uncertainty, provenance or relationship direction. Interoperability therefore does not mean every source record can be converted perfectly into every target schema. A high-quality crosswalk names its source and target versions, classifies mapping strength, preserves unmapped information where possible, records transformation provenance and tests round trips so downstream users know which distinctions did not survive.

source schema + record → mapping rules → target schema → validation → semantic-loss receipt → downstream use → optional reverse mapping → compare with source

Syntactic Similarity Is the Weakest Evidence

Two fields are both named date.

One means publication date.

The other means date the catalogue record was last edited.

A name match produces the wrong crosswalk.

The Metadata master already states the governing warning:

same field label ≠ same semantics.

Definition Comes Before Mapping

For each source and target field, ask:

• What resource does this property describe?
• What relationship does it assert?
• What value type is allowed?
• Can it repeat?
• Is order meaningful?
• Which vocabulary constrains it?
• What uncertainty or qualification can it carry?

Only then does a mapping statement become meaningful.

Exact Mapping Is the Ideal Case

Source property and target property express the same relationship over the same kind of resource, with compatible value semantics and constraints.

Even here, check:

• cardinality;
• value datatype;
• language tags;
• identifier versus literal use;
• required versus optional status.

Semantic equivalence can fail at the constraint layer even when the property definitions are close.

Broader Mapping Loses Precision

Source schema distinguishes:

• dateCreated;
• dateIssued;
• dateModified.

Target schema has only:

date

Mapping all three into date may be valid for discovery.

The target can no longer tell which event each date described unless extra qualification survives elsewhere.

many precise meanings → one broad target = semantic compression.

Narrower Mapping Requires a Decision the Source Did Not Make

Source:

date = 2026-08-31

Target requires either:

• created;
• issued;
• modified.

The source does not contain enough information to choose safely.

A translator that guesses has invented metadata.

The correct mapping may be “unresolved” until source context supplies the missing distinction.

One-to-Many Mapping Can Preserve Structure

A source field contains:

Jane Tan (editor)

A richer target may split that into:

• agent = Jane Tan;
• role = editor.

This is not loss if the source semantics support the decomposition.

But the extraction rule should be explicit and provenance-aware.

Many-to-One Mapping Is Where Loss Hides

Source schema has:

• primary creator;
• contributor;
• translator;
• illustrator.

Target has one:

creator

If every person becomes creator, role distinctions disappear.

The target may still support discovery.

It no longer supports every source-side question.

Cardinality Mismatch Can Destroy Information Quietly

Source allows ten subjects.

Target field permits one.

Which subject survives?

First?

Most specific?

Most popular?

Any choice is an information policy, not a neutral technical step.

Ordering Mismatch Matters Too

Source records creators in credited order.

Target stores an unordered set.

Every name survives.

Credit order does not.

Completeness by field count can therefore hide relational loss.

Literal-to-Identifier Mapping Can Gain Structure

Source:

creator = "Jane Tan"

Target:

creator = ORCID/ROR/authority URI

The target gains stable identity if the entity resolution is correct.

But the mapping activity has made an inference: this string refers to this identified entity.

The first pillar, Metadata Provenance, owns the evidence trail for that reconciliation.

Identifier-to-Literal Mapping Usually Loses More

Target cannot store the URI.

It keeps only the display label.

“Jane Tan” survives.

The stable identity link does not.

When the record later moves to another system, entity reconciliation must happen again.

Controlled Vocabularies Create a Second Crosswalk Problem

Schema fields match perfectly.

Source subject values use Vocabulary A.

Target expects Vocabulary B.

Now property mapping is solved but concept mapping is not.

Concept Mappings Have Strength Too

One concept may correspond closely to another.

Or be broader.

Or narrower.

Or overlap only under certain contexts.

Flattening all of these to “same concept” produces false semantic interoperability.

Free Text Is Harder Still

Source value:

artificial intelligence and society

Target vocabulary offers:

• Artificial Intelligence;
• Technology and Society;
• AI Ethics.

Choosing one or several controlled concepts requires interpretation.

The crosswalk has moved from syntax into classification.

Language Tags Can Be Lost

Source stores:

title@en
标题@zh

Target has one untagged title string.

Which language wins?

If both are concatenated, can downstream machines still distinguish them?

Multilingual interoperability depends on preserving language metadata, not merely Unicode text.

Datatype Loss Changes Machine Meaning

Source stores an ISO date typed as a date.

Target stores free text.

The visible value survives.

Machine-level date operations may not.

The target has preserved presentation while weakening computability.

Relationship Direction Can Flip or Disappear

Source:

A isVersionOf B

Target only has:

relatedTo

The relationship survives as a link.

Version direction disappears.

That loss matters if the receiver needs to determine which resource supersedes which.

Nested Structure Can Collapse Into a String

Source has a structured address:

• street;
• city;
• postal code;
• country.

Target has:

addressText

Human readability survives.

Field-level geographic filtering may not.

Flattening Is Often Useful and Still Lossy

A simple discovery portal may not need the full museum-object data model.

Flattening a rich record into title, creator, date, thumbnail and access URL can be exactly the right exchange product.

The mistake is not flattening.

The mistake is forgetting that the flattened record cannot answer every question the source could answer.

DCMI Application Profiles Exist Because Communities Need Different Constraints

The current DCMI Metadata Terms are designed to be combined with compatible vocabularies in application profiles.

This is a powerful interoperability principle:

shared vocabulary does not require identical local metadata models.

Crosswalks exist partly because legitimate local specialisation remains.

DCAT Shows Interoperability Without One Database

W3C DCAT 3 defines a shared vocabulary for data catalogues so datasets and services from separate systems can be consumed, aggregated and discovered together.

The source catalogues do not need to become one internal database.

This is exactly the environment in which explicit semantic mappings and profiles matter.

Validation Only Proves the Target Record Fits the Target Rules

A transformed record passes schema validation.

Good.

That does not prove the crosswalk preserved the source meaning.

W3C SHACL is useful for validating RDF graphs against declared shapes and constraints.

But structural conformance and semantic fidelity remain different tests.

Round-Trip Testing Exposes Irreversible Loss

Map A → B.

Then map B → A.

Compare the reconstructed A with the original.

Lost:

• creator roles;
• language tags;
• date semantics;
• identifier links;
• ordering.

The round trip makes the semantic loss visible.

A crosswalk need not be reversible to be useful.

Its irreversibility should be known.

A Loss Ledger Is Better Than Silent Success

For each mapping rule, record:

• source property/version;
• target property/version;
• mapping strength;
• condition;
• transformation function;
• known lost distinctions;
• fallback;
• provenance of the mapping decision.

The crosswalk becomes auditable rather than magical.

Unmappable Is a Valid Output State

Source has conservation-condition history.

Target has no condition concept.

Do not force the information into description merely to claim 100% field coverage if the target semantics no longer support machine use.

Options include:

• preserve source extension;
• attach source record;
• emit an unmapped-field receipt;
• use a richer profile;
• accept intentional loss for the target job.

Extensions Can Protect Local Meaning

A shared exchange profile contains common fields.

A domain extension carries museum-specific or education-specific details.

This often produces better interoperability than forcing every local concept into the common core.

But Extensions Can Become Private Dialects

Every organisation invents hundreds of local fields.

No definitions.

No identifiers.

No mapping documentation.

Interoperability collapses again.

Extensions need governance, documentation and stable semantics.

Crosswalks Must Be Versioned

Schema A v1 → Schema B v2.

Then B v3 changes creator semantics.

The original crosswalk may no longer be safe.

The second pillar, Metadata Currentness, owns whether the mapping is still applicable today.

The fourth pillar, Schema Evolution, owns how the schemas themselves change without invalidating every older record.

Provenance Should Follow the Mapped Value

The transformed target says:

creator = Jane Tan

Its provenance should be able to say:

• derived from source field author;
• source record ID;
• crosswalk version;
• mapping rule;
• transformation timestamp.

Then downstream systems can distinguish source assertion from transformed representation.

Crosswalk Chains Compound Loss

Schema A → B → C → D.

Each step seems reasonable.

By D, creator roles, language, uncertainty and identifiers may all have disappeared.

When possible, transform from the richest canonical source rather than repeatedly translating a translation.

crosswalk of crosswalks is semantic generation loss.

AI Can Propose Crosswalks—but Should Not Hide Ambiguity

A model can compare field names, definitions and examples and suggest candidate mappings.

Useful.

But a plausible mapping should remain:

• machine-proposed;
• scored or qualified;
• reviewable;
• version-bound;
• testable on real records.

A model should not convert semantic uncertainty into a silent exact map merely because the column names look alike.

A Better Metadata-Crosswalk Model

source schema/version → property/value semantics → target schema/version → classify mapping → transform with provenance → validate target syntax → record semantic loss/unmapped values → round-trip test → receiver-specific use

A 30-Lens Metadata Crosswalk Audit

1. Source schema: what vocabulary/profile?
2. Source version: which release?
3. Target schema: what vocabulary/profile?
4. Target version: which release?
5. Resource type: are both schemas describing the same entity level?
6. Source definition: what does the property mean?
7. Target definition: what does the property mean?
8. Mapping strength: exact, broader, narrower, related?
9. Condition: when does the mapping apply?
10. Cardinality: one or many values?
11. Ordering: is sequence significant?
12. Datatype: literal, date, number, identifier?
13. Language: are language tags preserved?
14. Identifier: does stable entity identity survive?
15. Vocabulary: which controlled concepts are used?
16. Concept mapping: exact or approximate?
17. Nested structure: is hierarchy flattened?
18. Role: are agent roles preserved?
19. Direction: are inverse relationships preserved?
20. Uncertainty: can qualifiers survive?
21. Provenance: can mapped values trace back to source?
22. Transformation: what computation changes the value?
23. Unmapped: what has no target home?
24. Extension: can local meaning be retained safely?
25. Validation: does target syntax conform?
26. Semantic test: does the target answer the same question?
27. Round trip: what is lost A→B→A?
28. Chain: is this mapping based on another transformed copy?
29. Currentness: are schema and crosswalk versions still applicable?
30. Receiver: is the remaining information sufficient for the target job?

Laboratory 1: Date Collapse

Map created, issued and modified dates into a target with one generic date field. Then attempt the reverse mapping. List which distinctions cannot be recovered.

Laboratory 2: Multilingual Title

Transform a record with English, Chinese and Malay language-tagged titles into a system that permits one untagged title. Design three strategies and state the loss created by each.

Laboratory 3: Crosswalk Receipt

For five source properties, classify the mapping as exact, broader, narrower, conditional or unmappable. Attach source/target versions and a one-line loss statement to every rule.

For Primary Readers

Imagine moving toys from a box with separate spaces for cars, buses and bicycles into a box labelled only “vehicles.” All toys fit, but the old categories disappear. That is a many-to-one crosswalk.

For Secondary Readers

For each mapped field, state whether the target is exact, broader or narrower and identify one question the target can no longer answer.

For Advanced Readers

Model a metadata crosswalk as a typed semantic transformation between versioned application profiles. Fidelity should be evaluated over property semantics, value domains, cardinality, relation direction and controlled-vocabulary mappings; target conformance is necessary but insufficient for semantic preservation.

Common Misconceptions

• “Matching field names mean the fields are equivalent.” Definitions and constraints may differ completely.
• “If the target validates, the translation is correct.” Structural conformance does not prove semantic fidelity.
• “A good crosswalk should map 100% of fields.” Honest unmappable states are safer than invented semantics.
• “If every value survives visibly, no information was lost.” roles, order, datatype, language and identifiers can disappear while strings remain.
• “Crosswalks are timeless.” They are mappings between specific evolving schema/profile versions.

Research Corridor

• DCMI Metadata Terms — interoperable terms designed for use with other compatible vocabularies and application profiles.
• W3C DCAT 3 — catalogue interoperability across separate systems.
• W3C SHACL — RDF graph validation against declared shapes and constraints.
• eduKateSG — How Interoperability Works.

Frequently Asked Questions

What is a metadata crosswalk?

It is a set of mappings that translates properties, values or structures from one metadata schema or application profile into another.

Why are metadata crosswalks lossy?

Source and target schemas often differ in semantic precision, cardinality, value types, relationship structure and controlled vocabularies, so some distinctions have no exact target representation.

How can crosswalk loss be measured?

Classify mapping strength, record unmapped structures and run round-trip transformations to see which source distinctions cannot be reconstructed from the target representation.

Final Thought: Translation Works Better When It Admits Its Accent

Interoperability does not require pretending every schema speaks the same language.

A trustworthy crosswalk carries meaning as far as the target can hold it—and leaves a visible receipt for every distinction that could not make the journey.