Skip to content

Semantic Bridge

The draft Semantic Mapping Publication Profile describes publication requirements for mapping artefacts beyond this Product Holon application.

7. Phase 3 — Align through a governed semantic bridge

7.1 Objective

Phase 3 creates and governs semantic correspondences between source-derived GDSN/GPC semantics and target publication semantics such as GS1 Web Vocabulary and schema.org. It is a design-time control-plane process. It does not run lexical mapping separately for each product.

7.2 Governed semantic bridge package

The bridge is a package of independently versioned artefacts, not a single ontology containing every mechanism:

Artefact Role
SSSOM mapping set Source of truth for candidate, reviewed, approved and rejected mappings.
OWL bridge module Compiled equivalence, subclass, subproperty or property-chain axioms where structurally sound.
DL-safe rule module Conditional semantic derivations evaluated with the ontology over named Product Holon individuals.
SPARQL projection/export package Extracts or serialises approved target graphs; may implement explicitly governed transforms that are not placed in the ontology.
Publication allow-list Determines which facts may be emitted to each audience profile.
Regression and conformance tests Demonstrate that released mappings still produce expected results after source or target changes.
Release manifest Links every compiled artefact to the reviewed SSSOM records and source vocabulary releases.

SPARQL queries are not part of the OWL ontology. They belong to the bridge package and have their own lifecycle.

7.3 Four publication routes

Every candidate product fact is assigned one of four routes.

Route Condition Treatment
Route 1 — direct semantic correspondence Source and target have compatible meaning, value space, cardinality and graph shape. Compile an approved OWL mapping or deterministic property alias.
Route 2 — structural transformation Meaning is compatible but source and target graph structures differ or the output property depends on a code/value condition. Apply an approved DL-safe rule or versioned transform, then export the result.
Route 3 — generic publication The fact is publishable but no approved first-class target term exists. Emit schema:additionalProperty / schema:PropertyValue with the source property identifier and label.
Route 4 — excluded The fact is B2B-only, commercially sensitive, inapplicable, unsupported or outside the publication profile. Retain it in the authorised internal or B2B representation; do not emit it publicly.

Route assignment is a mapping-governance decision, not a runtime guess by an LLM.

7.4 Candidate generation

Candidate mappings may be generated through:

  • exact or normalised label comparison;
  • synonym and annotation comparison;
  • structural similarity;
  • code-list and value-space comparison;
  • source lineage and documentation;
  • existing schema.org or Web Vocabulary alignments;
  • domain-expert proposals.

The demonstrated LOOM/BioPortal run extracted 28,101 unique terms and produced 237 unique bidirectional candidate pairs, serialised as 474 directional mapping records:

  • GDSN ↔ GPC: 38 unique pairs;
  • GDSN ↔ GS1 Web Vocabulary: 184 unique pairs;
  • GPC ↔ GS1 Web Vocabulary: 15 unique pairs.

All records were expressed as skos:closeMatch, which is appropriate for lexical candidates. The output is evidence for review, not authority for equivalence.

7.5 Limits of lexical matching

Labels are not semantic keys. Similar labels may hide different scope, range, cardinality or structure; dissimilar labels may express the same concept. Opaque generated GDSN CURIEs make labels necessary for candidate discovery, but labels alone remain insufficient for approval.

The source evidence found several examples:

  • allergenStatement in Web Vocabulary is a language-tagged string, while the structured GDSN allergen path is a coded multi-hop record;
  • a separate flat GDSN allergen-statement attribute is a more plausible direct candidate;
  • claimDetail has no single lexical match, while components of its surrounding domain align to several distinct Web Vocabulary record types;
  • multiple GDSN GTIN attributes map lexically to one gs1:gtin or schema:gtin concept, creating an n:1 cardinality issue that requires packaging-level review.

7.6 SSSOM mapping-set governance

Every mapping shall be represented as a first-class SSSOM record with, where applicable:

  • subject_id and subject_label;
  • predicate_id;
  • object_id and object_label;
  • mapping justification;
  • mapping tool and tool version;
  • source and target vocabulary releases;
  • confidence;
  • mapping cardinality;
  • author and reviewer identifiers;
  • review date and decision;
  • comments describing structural assumptions;
  • status: candidate, approved, rejected, superseded or deprecated.

An automated candidate should normally use semapv:LexicalMatching or the appropriate machine-matching justification. A reviewed decision records manual or composite curation. The exact predicate may remain skos:closeMatch, become skos:exactMatch, or compile into an OWL axiom only after the review establishes the required semantics.

7.7 Human review gate

The review process shall test:

  1. identifier resolution and term status;
  2. labels, definitions and domain context;
  3. source and target value spaces;
  4. datatype and language handling;
  5. cardinality and mapping cardinality;
  6. graph shape and nesting;
  7. target publication behaviour;
  8. loss, aggregation or transformation risk;
  9. authority and intended audience;
  10. regression examples and rejection cases.

No candidate shall be compiled into a reasoner-active equivalence solely because a lexical tool generated it.

7.8 Direct OWL mappings

An OWL equivalentProperty or equivalent-class axiom is appropriate only where the two terms denote the same relation or class over compatible graph structures. Domain aliases within Web Vocabulary are examples of straightforward equivalence. A source-derived GDSN-to-Web Vocabulary equivalence must be explicitly approved because Web Vocabulary does not publish links to the generated local GDSN namespace.

Where a two-hop object-property path reaches the same target individual expected by the target property, an OWL property chain may support a derived property. It shall not be used to convert a structured individual into a literal string or to branch on a data value.

7.9 DL-safe rule mappings

Conditional mappings may use DL-safe rules in an independently versioned module imported by the core bridge ontology. Every variable binds to named Product Holon individuals already present in the ABox. Representative cases include:

  • mapping nutrient code FAT and its quantity to schema:fatContent;
  • mapping ENERC, PRO- and NA to the corresponding schema.org nutrition properties;
  • deriving an approved textual allergen statement from a structured coded chain where no authoritative flat statement is available;
  • constructing an isomorphic publication assertion from a named source record.

The rule module shall be tested with the same semantic releases and mapping-set version as the Product Holon publication profile.

7.10 SPARQL projection and export

SPARQL may be used to:

  • select the entailed target graph;
  • construct the final JSON-LD publication graph;
  • filter Route 4 properties;
  • create generic PropertyValue nodes for Route 3;
  • serialise audience-specific views;
  • implement a transform that governance intentionally keeps outside the reasoner.

Where substantive transformation remains in SPARQL, the query is a governed compiled artefact with tests and provenance. Where SWRL has already entailed the target properties, the export query should remain thin and stable.

7.11 Generic PropertyValue projection

Route 3 preserves a publishable fact without claiming that it is a native Web Vocabulary or schema.org property. The representation shall include:

  • schema:additionalProperty;
  • a schema:PropertyValue node;
  • the source property identifier in schema:propertyID;
  • a human-readable name;
  • the value and unit where applicable;
  • source or provenance linkage where supported by the profile.

Downstream parsers may give generic properties less weight than first-class properties. The architecture treats Route 3 as loss-minimising publication, not semantic equivalence.

7.12 Current evidence-based mapping findings

The evidence pack records the full findings. The consolidated position is:

  • GTIN has strong schema.org/Web Vocabulary support, but multiple source GTIN attributes require packaging-level cardinality review;
  • gdsn:a812659001 and gs1:ingredientStatement are a tool-corroborated candidate pair pending formal human approval;
  • gpc:20002867 and gs1:sharpnessOfCheese are independently corroborated candidate terms, with value-level equivalence still requiring review;
  • gdsn:a-1463136308 is the plausible direct candidate for gs1:allergenStatement; the structured allergen chain is a separate Route 2 case;
  • nutrient codes ENERC, FAT, PRO- and NA have defined schema.org projection targets; other nutrients, such as calcium in the worked example, require Route 3 unless a governed target is added;
  • every residual facet reproduced in the four-Brick mapping tables had no LOOM candidate to Web Vocabulary and currently uses Route 3; the source baseline elsewhere reports a 19-of-20 count, but its reproduced per-Brick lists do not reconcile to that total, so no consolidated numeric claim is made here;
  • packagingRecycledContentRatio and claimDetail were present in the candidate term set but produced no direct Web Vocabulary match; they require structural review, decomposition, Route 3 or Route 4 treatment.

These are architectural inputs, not an approved GS1 crosswalk.

7.13 Bridge package versioning

A bridge release shall identify:

  • source GDSN and GPC releases;
  • target Web Vocabulary and schema.org releases;
  • SSSOM mapping-set version and checksum;
  • approved mapping count by predicate and route;
  • rule-module version;
  • export-query version;
  • allow-list version;
  • regression-test results;
  • known rejected and deprecated mappings.

7.14 Drift detection and regression testing

Regression tests shall detect:

  • removed or deprecated source and target terms;
  • changed domains, ranges, cardinalities or code values;
  • mappings whose graph shapes are no longer compatible;
  • unexpected inference closure;
  • lost or duplicated publication facts;
  • invalid JSON-LD or SHACL output;
  • changed Route 3 fallback counts;
  • changed public/private classification.

7.15 Phase outputs

Phase 3 produces a released governed semantic bridge package, including its SSSOM source of truth, compiled semantic artefacts, export/projection rules, publication allow-lists, tests and release manifest.

Part IV — Publication and consumption

The next phase applies the governed bridge to publication and consumption of Product Holon representations.

Continue to Publication and Consumption.