Crosswalks: SSSOM, Mapping Predicates and Ontology Alignment
No standard wins. That is not pessimism; it is the empirical record of the previous lesson and the entire domain tour to come. Every real organisation lives with several standards at once: the regulator mandates one vocabulary, the flagship supplier ships another, the legacy estate speaks a third, and the new AI platform extracts entities that match none of them. The scarce, valuable artifact is therefore not another standard but the bridge between standards: the crosswalk. This lesson teaches bridges as first-class engineering: the mapping predicates that actually mean something, SSSOM as the interchange format, the state of the art in automated matching, and the certification discipline that separates a bridge you can build on from a spreadsheet of hopeful guesses. We have shipped a public, tested crosswalk between two government-grade ontologies, so this is written from practice.
Why bridges beat monoliths
The monolith strategy, adopt one standard everywhere and translate at the borders, fails for a structural reason: the borders are where the information lives. Each standard encodes its community’s distinctions, and translation without an explicit bridge silently discards whichever distinctions the target cannot express. Do that through two or three hops and a large fraction of the original detail is gone, with nobody able to say where. A crosswalk makes the loss explicit and auditable: every correspondence is a recorded claim with a stated strength, every deliberate non-correspondence is recorded too, and the whole set is versioned and testable. In economic terms, a good crosswalk converts an O(n squared) integration problem, every system pairwise against every other, into O(n) mappings against a hub, which is exactly the architecture the National Digital Twin Programme pursues with IES and the reason lesson 26’s IES-HQDM bridge matters beyond its two endpoints.
Mapping predicates that mean something
The unit of a crosswalk is a single mapping: source term, predicate, target term, plus metadata. Everything hinges on the predicate, and the field’s workhorse vocabulary comes from SKOS, whose mapping properties were designed for exactly this cross-scheme role. The discipline is to use the weakest predicate the evidence supports, never the strongest one convenience suggests.
| Predicate | Claim being made | Use when | Risk if wrong |
|---|---|---|---|
| owl:equivalentClass | Logically identical extensions; full inference flows both ways | Rarely; only when both definitions provably coincide | Highest: reasoners propagate every consequence of the error through both ontologies |
| skos:exactMatch | Interchangeable for information retrieval across schemes | Confident correspondence without full logical commitment | Moderate: transitive by design, so chains of exactMatch can drift far from exact |
| skos:closeMatch | Similar enough for many applications, not interchangeable | Real overlap with known edge-case divergence | Lower: explicitly non-transitive, contains its own blast radius |
| skos:broadMatch / skos:narrowMatch | Target is more general / more specific than source | The common case between standards at different granularities | Direction errors invert hierarchies; always record orientation convention |
| skos:relatedMatch | Associated, neither equivalent nor hierarchical | Flagging relevance where no subsumption holds | Low, but overuse turns a crosswalk into a suggestion box |
| Asserted non-mapping | We looked, and there is deliberately no counterpart | Concepts one side cannot or will not express | Omitting these is the silent killer: absence of a mapping is otherwise indistinguishable from absence of effort |
Two habits from delivery work. First, the asserted non-mapping deserves equal ceremony with the mappings: when we built the IES-HQDM bridge, recording why certain concepts have no counterpart, with the reasoning attached, saved every subsequent integrator from re-litigating the gap, and it is the difference between a complete review of the source ontology and a partial one. Second, resist equivalence inflation. Under lesson 28’s false-friends analysis, most cross-standard matches that look exact are closeMatch or broadMatch on inspection; a crosswalk whose predicate distribution is dominated by exactMatch is usually advertising its own superficiality.
What breaks · owl:sameAs across registries
Finance’s fixed-asset register carries A-2201, “Pump P-101, Aberdeen-2”. The obvious join:
fin:A-2201 owl:sameAs ex:P101 .
# after the merge, both of these are entailed:
ex:P101 fin:depreciationSchedule fin:Sched7 . # the pump depreciates
fin:A-2201 ex:hasImpeller ex:IMP77 . # the ledger entry has an impellerowl:sameAs merges every property of both individuals, in both directions, forever, and the two registries have different identity criteria, which is exactly when the merge produces absurdity. Finance’s asset is the thing lesson 3 met: it survives physical replacement, so when P-101 is one day swapped out under the same asset line, the sameAs chain welds two physical pumps together through the ledger, and lesson 5’s impossible-machine graph returns wearing a finance badge. The repair is to say what is actually true: the asset entry and the pump are different individuals related by a modelled relationship, ex:recordedAs, or in lesson 24’s terms a relator binding a state of the pump to the ledger line. Identity is the strongest claim in the stack; across registries with different identity criteria, it is almost never the true one.
SSSOM: the interchange layer
For years every mapping project invented its own spreadsheet columns, which meant mappings could not be merged, compared or tooled. The Simple Standard for Sharing Ontological Mappings fixed this. SSSOM (Matentzoglu and colleagues, arXiv 2112.07051, published in the journal Database in 2022) defines a TSV-based format in which each row is one mapping carrying, alongside subject, predicate and object, the metadata that makes a mapping evaluable: justification (how the mapping was derived, from lexical matching through logical reasoning to human curation), author, date, confidence, and licence. A machine-readable header binds prefixes and set-level provenance, and a defined translation renders the whole thing as OWL or RDF when you need it in a triple store.
SSSOM has become the lingua franca of the mapping world with striking speed: the sssom-py toolkit handles conversion and validation, the OBO ecosystem publishes curated mapping sets in it, and EMBL-EBI rebuilt its cross-ontology mapping browser as OxO2, SSSOM-native and aimed explicitly at logically sound crosswalks (arXiv 2506.04286). The professional rule is simple: whatever your internal tooling, your crosswalk should be able to emit SSSOM, because that is the format reviewers, aggregators and successor teams can consume without a meeting. It also, quietly, enforces honesty: an SSSOM row without justification or confidence is visibly naked in a way a spreadsheet cell never was.
Worked example · the Aberdeen-2 asset register
Where we left it: IFC read cold, false friends flagged. Now the bridge gets built, as SSSOM rows.
subject_id predicate_id object_id conf justification
ex:PumpModel skos:closeMatch ifc:IfcPumpType 0.90 semapv:ManualMappingCuration
ex:PumpPosition skos:closeMatch ifc:IfcPump 0.85 semapv:ManualMappingCuration
ex:Pump sssom:NoTermFound - 1.00 physical unit has no IFC
counterpart before handoverEvery decision in those rows was earned earlier. The model level maps at the model level, because lesson 25 made HX-300’s stratum explicit. The register’s positions, not its physical pumps, match IFC occurrences, because lesson 26 split the two and lesson 28 caught the false friend. And the most valuable row is the last one, the asserted non-mapping with its reason attached: the physical unit’s absence from design-stage IFC is a fact future integrators would otherwise rediscover expensively. Note the predicate distribution: zero exactMatch. That is not timidity; it is what an honest audit of two standards with different commitments almost always yields, and a wall of exactMatch would have been the superficiality smell the lesson describes. Certification comes next: merge, reason, round-trip a fixture pump, and the declared losses must match the measured ones.
State of the register: a small, certified, honest bridge to its first neighbouring standard, and a template for every domain in the tour ahead.
Matchers: from LogMap to LLM agents
Can the mapping table be generated automatically? Partially, and the frontier has moved fast. The Ontology Alignment Evaluation Initiative, OAEI, has benchmarked matchers annually for two decades, and its recent editions tell the story. The classical generation, led by systems such as LogMap, combines lexical indexing with logical repair, ensuring the proposed alignment does not create unsatisfiable classes, and held the state of the art for roughly a decade. The 2024 campaign marked the turn: LLM integration lifted the CANARD system’s precision and F-measure by as much as 45 percent, and by the 2025 campaign LLM-augmented systems were mainstream across tracks, with LogMap-LLM, BERTMap and BioGITOM leading biomedical tasks. The emblem of the new generation is Agent-OM (arXiv 2312.00326, published at PVLDB 2025), which reframes matching as agent tool-use: paired retrieval and matching agents wielding ontology-matching primitives, achieving the best precision, recall and F1 on OAEI 2022 and 2023 tracks that dedicated matchers had owned for years. Libraries such as OntoAligner now package retrieval-based, LLM-based and hybrid alignment behind one interface, so a practitioner can assemble a candidate pipeline in an afternoon.
The practical reading of all this: machines now produce excellent candidate mappings and increasingly good similarity evidence, but every automated matcher, classical or agentic, still emits hypotheses, not commitments. The OAEI’s own methodology assumes a reference alignment created by humans. Use matchers to generate and rank candidates, capture their scores in the SSSOM confidence and justification fields, and route everything above your threshold to expert review. What the LLM wave has genuinely changed is the economics of the first 80 percent, which shifts the human budget to where it always belonged: the hard 20 percent where the false friends live.
Certifying a bridge
A crosswalk you merely published is an opinion; a crosswalk you certified is infrastructure. Certification, as we practise it on the IES-HQDM crosswalk and teach in our crosswalk course, means the bridge has survived four specific ordeals, automated wherever possible. Logical safety: merge both ontologies with the bridge axioms and run a reasoner; the result must be consistent, every class satisfiable, and free of accidental collapses in which two deliberately distinct classes become entailed as equivalent. This catches the subtle disasters, such as an innocent-looking pair of broadMatch-turned-subclass axioms whose composition quietly equates a person with an organisation. Data round-trip: take fixture instances, translate them across the bridge and assess what survives; the measured loss must match the loss the predicates declare, because a bridge that claims exactMatch and drops properties in transit is lying in a testable way. Metadata completeness: every row carries predicate, justification, confidence, author and date, enforceable by validating the SSSOM file itself, shapes over the mapping set, in exactly the CI spirit of lesson 27. Review and versioning: named reviewers, a changelog, and versioned releases, because both endpoint ontologies will evolve and an unversioned bridge silently rots. None of these ordeals is exotic; all four are scriptable with the tooling this module has already given you. The falsifiability doctrine applies with full force: a bridge that could not have failed any test certifies nothing.
Step back and the module’s arc closes. Lesson 28 taught you to read any standard; this lesson taught you to connect any two; and the connective discipline, explicit predicates, interchange formats, machine-checked safety, human-owned judgement, is the same falsifiability engineering as lesson 27, pointed outward. As you enter the domain tour, watch for the bridges as much as the standards: in every domain, the crosswalk is where the integration value concentrates, and the people who can certify one are the scarcest resource in the room.
Check yourself
The register’s ex:Impeller must be mapped to a target standard’s RotatingComponent, whose definition covers impellers, rotors and shafts. Which predicate belongs in the SSSOM row?
- skos:broadMatch
- skos:exactMatch
- owl:sameAs
- skos:relatedMatch
Show the answer, and why each wrong option is wrong
A is correct. Every impeller is a rotating component and not conversely, so the target term has the broader meaning: skos:broadMatch, read as “has broader match”, records exactly that directional subsumption, and the direction is information the round-trip test will rely on when it measures what detail dies in translation.
B is equivalence inflation: the extensions demonstrably differ (shafts are not impellers), and a reviewer who finds exactMatch here should downgrade their trust in every other row of the file. C is the wrong layer twice over: sameAs asserts identity of individuals, these are classes, and the previous block just showed what identity assertions do across artefacts with different commitments. D errs in the safe-looking direction, and is still wrong: the evidence supports a specific subsumption, and recording only “related” discards it. The discipline is the weakest predicate the evidence supports, not the weakest predicate available; both inflation and deflation are ways of lying about what you know.
