You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
External ontology imports — strategy (MIREOT vs SLME) [ported from KB RES-A-12]
#74
Ported from the internal YouTrack KB (RES-A-12, written 2026-08-03) to GitHub Discussions on 2026-09-18.
Per MAN-8 / MAN-A-3, GitHub is the source of truth for ADIRO design decisions. The YouTrack article is now frozen and points here. This discussion is closed as resolved — it records a decision, not an open question. New questions about it belong in a new Discussion or issue.
Status: DECIDED. The review action (RES-59) is resolved; implementation is tracked in RES-68. The policy below is already normative in the repo — see docs/contribute/versioning/ §Imports.
Context for ADIRO's ontology versioning epic (RES-27).
ADIRO will definitely import terms from external ontologies — near-term for geometrical concepts (GeoSPARQL, or the W3C Linked-Building-Data OMG/FOG geometry ontologies), and likely DAnO, among others. This note records how we should do that and why, so we choose deliberately rather than by default.
A note on ROBOT (it's the toolkit, not a method). The decision below is MIREOT vs SLME — two extraction methods. ROBOT is the tool that can perform either of them via robot extract --method MIREOT|STAR|BOT|TOP. Separately, robot report (quality checks) and robot reason (consistency / unsatisfiable-class detection) are what we use for the QC / reasoner gate (RES-36). So ROBOT appears in two distinct roles: (1) extracting the import module, and (2) the consistency gate that catches a bad import. It is not itself an alternative to MIREOT.
The question that started this
When you reference an external term with MIREOT you must supply three things: the source ontology URI, the source term URI, and a target (local) superclass URI. The intuition: the first two already identify the term, and its parent hierarchy is retrievable from the source — so why hand-specify a parent? If the tool can find the parent, it can find the whole chain; "stop at the parent" feels arbitrary.
This intuition is correct about a real weakness — but it conflates two different things.
What MIREOT actually does (identity vs. placement)
MIREOT = Minimum Information to Reference an External Ontology Term (Courtot et al., 2011). Its third URI is not there because the parent is unknowable; it's there to force you to commit, in your ontology, to where the term sits:
URIs 1–2 = identity (the term's "ID card").
URI 3 = the integration axiom ("job assignment") — without it the imported term is an isolated node with no logical link to ADIRO.
MIREOT intentionally ignores the source hierarchy, to (a) avoid bloat (pulling in hundreds of irrelevant intermediate classes) and (b) let you re-home the term under a local parent that fits our domain, which may differ from the source's. Hardcoding the local superclass also decouples us from later changes in the source's structure.
So the manual parent is a feature for the deliberate re-homing case — and pure downside when our intent is instead to faithfully reuse the source's structure (exactly the case the intuition targets).
Is there consensus? No — MIREOT is a pragmatic compromise
Its own authors describe it as a "balanced compromise" that trades logical completeness for usability, admitting that importing only selected information means "incomplete or incorrect inferences could conceivably be made."
Documented failure at scale — the key evidence:"To MIREOT or not to MIREOT? A case study of the impact of using MIREOT in the Experimental Factor Ontology (EFO)" (Slater et al., KAUST BORG, ~2018). Importing EFO's MIREOT-referenced ontologies in full showed EFO logically inconsistent with its references, and combining EFO with many of them individually produced large numbers of unsatisfiable classes — "potential problems within a major ontological ecosystem." A follow-on ("…finding and repairing hidden contradictions in biomedical ontologies," BMC Med Inform Decis Mak, 2020) develops repair methods. This is directly analogous to ADIRO importing from geometry + DAnO + others.
The better default for us: SLME (via ROBOT extract)
SLME = Syntactic Locality-based Module Extraction (Grau, Horrocks, Kazakov, Sattler — "Modular Reuse of Ontologies," 2008). It automatically computes exactly the axioms/ancestors needed so the extracted module preserves every entailment over the terms you selected — no manual parent, logically safe by construction. It is one of the methods offered by robot extract.
Methods: STAR (smallest such module — usual default for reference reuse), BOT, TOP.
Trade-off vs MIREOT: SLME modules can be larger (they include the logically-necessary context) — but that's the point; that context is what keeps reasoning sound.
Limitation: SLME faithfully slices the source — it will not re-home a term under a different local parent for you.
seed + all their super-classes (view from the bottom of the hierarchy upwards)
medium
TOP (⊤)
seed + all their sub-classes (view from the top downwards)
large
STAR (⊥⊤*)
seed + the inter-relations between them — not necessarily sub- or super-classes
very small
So BOT is the method that pulls in the upward ancestor chain, and STAR is the one that does not. If the worry is "don't drag an upper ontology (BFO / PROV / CCO / schema.org) in behind a geometry term," then STAR is the mitigation and BOT is the risk — not the other way round.
Two useful consequences of ⊥⊤*-locality worth knowing when reasoning about what STAR will actually pull:
An upward rdfs:subClassOf / rdfs:subPropertyOf alignment (e.g. geo:SpatialObject ⊑ bfo:BFO_0000001) is ⊤-local and is therefore dropped by STAR.
An owl:equivalentClass / owl:equivalentProperty alignment is neither ⊥- nor ⊤-local and is retained by STAR. So e.g. geo:Feature ≡ (BFO union ⊔ as:Place)would come through even on STAR — if such an axiom were in the input graph at all.
That last point is the real control: what matters most is which graph you feed the extractor, not which method you pick. SLME is fully deterministic for a fixed input.
Recommendation for ADIRO
Default extraction method = SLME (robot extract --method STAR) for reusing external terms — automatic, entailment-preserving, no arbitrary manual placement.
Version-pin every external source (reproducible extracts; also satisfies our versioning rule that pinned imports be documented + justified — see RES-27). Pin the core vocabulary file, not a merged/alignment graph — this is what makes the extract deterministic and keeps upper-ontology alignments out by construction.
Reserve MIREOT (--method MIREOT) only for the deliberate case where we intentionally re-home an external term under ADIRO's own taxonomy — and reasoner-check it.
Mandatory safeguard regardless of method: the QC / reasoner gate in CI — robot reason (consistency + unsatisfiable-class detection) and robot report across the merged ontology — so any import that introduces unsatisfiability is caught before release. ✅ This is now DONE and blocking (RES-36): .github/workflows/ontology-reasoning.yml runs HermiT + ROBOT report over the merged suite at latest on every src/**.ttl PR, via scripts/run_reasoning.sh, resolving imports offline through src/catalog-v001.xml. The reasoner is a blocking gate; report is advisory.
Third option, added 2026-09-18: local stub + alignment (don't import at all)
RES-A-12 as originally written framed this as a binary. In practice ADIRO has a third, lighter pattern, and it is already in use in the WIP aec_provenance module: declare a minimal local typing stub for the external term (rdf:type owl:Class/owl:ObjectProperty + rdfs:label + rdfs:isDefinedBy <source>) and align to it, without owl:imports and without any extraction.
What it buys: the module stays self-contained; the reasoning CI resolves everything offline; zero risk of importing foreign axioms; the external IRI is still used, so downstream consumers get the interoperability signal.
What it costs: you inherit none of the source's entailments. The stub is a name, not a definition.
When it's right: when we want the external IRI for interoperability signalling and our own axioms carry all the logic we need — which is the common case for lightweight alignment targets (PROV-O, DAnO).
When it's wrong: when we genuinely want to reason with the source's axioms. Then extract with STAR.
The repo's existing title-block review reaches the same conclusion for DAnO independently — "align to dano:, do not import it", because importing drags 16 drawing-element classes into a suite with a blocking DL gate for the sake of seven datatype properties.
Decision rule, in order of preference:
Only need the IRI + our own axioms → local stub + rdfs:isDefinedBy alignment.
Need the source's entailments → SLME STAR from the pinned core file.
Need the term re-homed under an ADIRO parent → MIREOT, reasoner-checked.
Never → owl:imports of a whole external ontology.
Note that MIREOT does not let you rewrite an external term's rdfs:domain/rdfs:range. Its third URI asserts a local superclass only. If an external property's domain is too narrow for us, that is a mint-our-own-term + rdfs:subPropertyOf/skos:closeMatch case (option 1), not a MIREOT case.
Open questions — now answered
These were the review questions carried by RES-59, answered by @Zaalouk in #70 and #71:
Question
Answer
Which geometry vocabulary, and which terms?
GeoSPARQL.geo:Geometry / geo:hasGeometry / geo:asWKT (CQ 1.3); Simple Features geo:sfContains/sfWithin/sfDisjoint/sfOverlaps (CQ 2.3); optionally RCC8/9 and the geof:* query functions. See #70, #71. Adoption itself is still deferred (see #36).
DAnO — reference and scope?
https://w3id.org/dano, CC BY 4.0, RUB. Scope: provenance (inferredBy/inferredWith/inferredFrom/inferredAt), hasConfidence, and depicts/isDepictedBy. Open: align-vs-extract is unresolved and now tracked as its own issue.
STAR default, or do we need BOT/TOP?
STAR stands as the default. No case has been found that needs BOT/TOP. The concern raised in #71 — avoiding BFO/PROV/CCO closure — is served by STAR, not BOT (see the table above), and is better handled by pinning the core file.
Genuine re-homing cases justifying MIREOT?
None confirmed yet. The GeoSPARQL cases are over-inclusion, not re-homing, and are avoided by input scoping. The DAnO cases (hasConfidence asserted with rdfs:domain dano:DrawingElement while DrawingElementMeta is documented as holding confidence; depicts with rdfs:domain dano:DisplayElement) are domain mismatches, which MIREOT cannot fix — they are mint-and-align cases.
Verified 2026-09-18: GeoSPARQL core (http://www.opengis.net/ont/geosparql, v1.1.1) has no owl:imports and contains none of the BFO / PROV / CCO / locn: / schema: alignment axioms — those live in a separate, non-imported alignments file. DAnO's dano.ttl confirms hasConfidence rdfs:domain dano:DrawingElement and depicts rdfs:domain dano:DisplayElement (no range).
References
Courtot et al., MIREOT: The minimum information to reference an external ontology term, Applied Ontology, 2011.
Towards semantic interoperability: finding and repairing hidden contradictions in biomedical ontologies, BMC Med Inform Decis Mak, 2020 — https://doi.org/10.1186/s12911-020-01336-2
Grau, Horrocks, Kazakov, Sattler, Modular Reuse of Ontologies: Theory and Practice, JAIR, 2008.
reacted with thumbs up emoji reacted with thumbs down emoji reacted with laugh emoji reacted with hooray emoji reacted with confused emoji reacted with heart emoji reacted with rocket emoji reacted with eyes emoji
Uh oh!
There was an error while loading. Please reload this page.
Context for ADIRO's ontology versioning epic (RES-27).
ADIRO will definitely import terms from external ontologies — near-term for geometrical concepts (GeoSPARQL, or the W3C Linked-Building-Data OMG/FOG geometry ontologies), and likely DAnO, among others. This note records how we should do that and why, so we choose deliberately rather than by default.
The question that started this
When you reference an external term with MIREOT you must supply three things: the source ontology URI, the source term URI, and a target (local) superclass URI. The intuition: the first two already identify the term, and its parent hierarchy is retrievable from the source — so why hand-specify a parent? If the tool can find the parent, it can find the whole chain; "stop at the parent" feels arbitrary.
This intuition is correct about a real weakness — but it conflates two different things.
What MIREOT actually does (identity vs. placement)
MIREOT = Minimum Information to Reference an External Ontology Term (Courtot et al., 2011). Its third URI is not there because the parent is unknowable; it's there to force you to commit, in your ontology, to where the term sits:
So the manual parent is a feature for the deliberate re-homing case — and pure downside when our intent is instead to faithfully reuse the source's structure (exactly the case the intuition targets).
Is there consensus? No — MIREOT is a pragmatic compromise
Its own authors describe it as a "balanced compromise" that trades logical completeness for usability, admitting that importing only selected information means "incomplete or incorrect inferences could conceivably be made."
Documented failure at scale — the key evidence: "To MIREOT or not to MIREOT? A case study of the impact of using MIREOT in the Experimental Factor Ontology (EFO)" (Slater et al., KAUST BORG, ~2018). Importing EFO's MIREOT-referenced ontologies in full showed EFO logically inconsistent with its references, and combining EFO with many of them individually produced large numbers of unsatisfiable classes — "potential problems within a major ontological ecosystem." A follow-on ("…finding and repairing hidden contradictions in biomedical ontologies," BMC Med Inform Decis Mak, 2020) develops repair methods. This is directly analogous to ADIRO importing from geometry + DAnO + others.
The better default for us: SLME (via ROBOT
extract)SLME = Syntactic Locality-based Module Extraction (Grau, Horrocks, Kazakov, Sattler — "Modular Reuse of Ontologies," 2008). It automatically computes exactly the axioms/ancestors needed so the extracted module preserves every entailment over the terms you selected — no manual parent, logically safe by construction. It is one of the methods offered by
robot extract.STAR(smallest such module — usual default for reference reuse),BOT,TOP.What the three SLME methods actually do
Per the ROBOT
extractdocumentation — worth stating explicitly, because it is easy to get backwards:BOT(⊥)TOP(⊤)STAR(⊥⊤*)So
BOTis the method that pulls in the upward ancestor chain, andSTARis the one that does not. If the worry is "don't drag an upper ontology (BFO / PROV / CCO / schema.org) in behind a geometry term," thenSTARis the mitigation andBOTis the risk — not the other way round.Two useful consequences of ⊥⊤*-locality worth knowing when reasoning about what
STARwill actually pull:rdfs:subClassOf/rdfs:subPropertyOfalignment (e.g.geo:SpatialObject ⊑ bfo:BFO_0000001) is ⊤-local and is therefore dropped bySTAR.owl:equivalentClass/owl:equivalentPropertyalignment is neither ⊥- nor ⊤-local and is retained bySTAR. So e.g.geo:Feature ≡ (BFO union ⊔ as:Place)would come through even onSTAR— if such an axiom were in the input graph at all.That last point is the real control: what matters most is which graph you feed the extractor, not which method you pick. SLME is fully deterministic for a fixed input.
Recommendation for ADIRO
robot extract --method STAR) for reusing external terms — automatic, entailment-preserving, no arbitrary manual placement.--method MIREOT) only for the deliberate case where we intentionally re-home an external term under ADIRO's own taxonomy — and reasoner-check it.robot reason(consistency + unsatisfiable-class detection) androbot reportacross the merged ontology — so any import that introduces unsatisfiability is caught before release.✅ This is now DONE and blocking (RES-36):
.github/workflows/ontology-reasoning.ymlruns HermiT + ROBOTreportover the merged suite at latest on everysrc/**.ttlPR, viascripts/run_reasoning.sh, resolving imports offline throughsrc/catalog-v001.xml. The reasoner is a blocking gate;reportis advisory.Third option, added 2026-09-18: local stub + alignment (don't import at all)
RES-A-12 as originally written framed this as a binary. In practice ADIRO has a third, lighter pattern, and it is already in use in the WIP
aec_provenancemodule: declare a minimal local typing stub for the external term (rdf:type owl:Class/owl:ObjectProperty+rdfs:label+rdfs:isDefinedBy <source>) and align to it, withoutowl:importsand without any extraction.STAR.The repo's existing title-block review reaches the same conclusion for DAnO independently — "align to
dano:, do not import it", because importing drags 16 drawing-element classes into a suite with a blocking DL gate for the sake of seven datatype properties.Decision rule, in order of preference:
rdfs:isDefinedByalignment.STARfrom the pinned core file.owl:importsof a whole external ontology.Open questions — now answered
These were the review questions carried by RES-59, answered by @Zaalouk in #70 and #71:
geo:Geometry/geo:hasGeometry/geo:asWKT(CQ 1.3); Simple Featuresgeo:sfContains/sfWithin/sfDisjoint/sfOverlaps(CQ 2.3); optionally RCC8/9 and thegeof:*query functions. See #70, #71. Adoption itself is still deferred (see #36).https://w3id.org/dano, CC BY 4.0, RUB. Scope: provenance (inferredBy/inferredWith/inferredFrom/inferredAt),hasConfidence, anddepicts/isDepictedBy. Open: align-vs-extract is unresolved and now tracked as its own issue.STARdefault, or do we needBOT/TOP?STARstands as the default. No case has been found that needsBOT/TOP. The concern raised in #71 — avoiding BFO/PROV/CCO closure — is served bySTAR, notBOT(see the table above), and is better handled by pinning the core file.hasConfidenceasserted withrdfs:domain dano:DrawingElementwhileDrawingElementMetais documented as holding confidence;depictswithrdfs:domain dano:DisplayElement) are domain mismatches, which MIREOT cannot fix — they are mint-and-align cases.Verified 2026-09-18: GeoSPARQL core (
http://www.opengis.net/ont/geosparql, v1.1.1) has noowl:importsand contains none of the BFO / PROV / CCO /locn:/schema:alignment axioms — those live in a separate, non-imported alignments file. DAnO'sdano.ttlconfirmshasConfidence rdfs:domain dano:DrawingElementanddepicts rdfs:domain dano:DisplayElement(no range).References
extract— http://robot.obolibrary.org/extract ;report— http://robot.obolibrary.org/report ;reason— http://robot.obolibrary.org/reasonRelated
All reactions