Repository navigation
External import strategy discussion (MIREOT vs SLME) #71
Replies: 2 comments
Clarification: this is the answer to RES-59, not a reopening of the MIREOT-vs-SLME decision
Thanks @Zaalouk — this is a genuinely useful piece of work, and the DAnO half of it is verified-correct (see below). Two framing points and three technical corrections, so the record is unambiguous. 1. The decision was already taken, and it standsThe MIREOT-vs-SLME question was decided on 2026-08-03 in KB RES-A-12. It has now been ported to GitHub and made public as #74 (per MAN-8: GitHub is the source of truth for ADIRO decisions). The YouTrack article is frozen and points there. #74 is closed as resolved, it records a decision. The implementation is tracked in RES-68: SLME That core rule was already normative in the repo when this comment was posted — This discussion is the deliverable for RES-59, the review action attached to that note — i.e. it answers the four open questions the decision deliberately left open. The title ("External import strategy discussion (MIREOT vs SLME)") reads like the decision is being re-litigated, which is what prompted this clarification. The answers from here and #70 are now folded into #74's "Open questions — now answered" table. 2. Correction:
|
| Method | Contains | Size |
|---|---|---|
BOT (⊥) |
seed + all their super-classes — "a view from the BOTTOM of the class-hierarchy upwards" | medium |
TOP (⊤) |
seed + all their sub-classes | large |
STAR (⊥⊤*) |
seed + inter-relations, "not necessarily sub- and super-classes" | very small |
So BOT is precisely the method that would pull BFO/PROV/CCO in, and STAR is the one that would not. The recommendation in §3 is inverted: STAR is the mitigation here, BOT is the risk. Nothing in this discussion motivates moving off the STAR default.
Mechanically, on ⊥⊤*-locality: geo:SpatialObject ⊑ bfo:BFO_0000001 is ⊤-local (replacing BFO_0000001 with ⊤ yields a tautology), so it is dropped from the ⊤-module and therefore from STAR. The same holds for geo:hasGeometry ⊑ locn:geometry. But owl:equivalentClass/owl:equivalentProperty axioms are neither ⊥- nor ⊤-local, so geo:Feature ≡ (BFO union ⊔ as:Place) would survive STAR — your instinct that the equivalence alignments are the sharp edge is right, even though the method conclusion isn't.
3. Correction: SLME isn't non-deterministic — the input graph is the variable
"SLME's automatic transitive closure becomes unpredictable/unstable… would non-deterministically drag in whichever external upper-ontology parent happens to co-reside in the loaded graph at the time"
SLME is deterministic for a fixed input ontology. What varies is which graph you hand it, and that is entirely under our control. The fix is input scoping, not method switching — and RES-A-12 already mandates it via "version-pin every external source".
I verified this: 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. They live in a separate file that is never pulled in unless you deliberately merge it. So the entire re-homing hazard in §4 is conditional on a choice we simply don't make. I've made this explicit in #74: pin the core vocabulary file, not a merged alignment graph.
Terminology: pulling in a source's own ancestors is over-inclusion/bloat, not re-homing. Re-homing means deliberately asserting an ADIRO-local parent for an external term.
4. The DAnO findings are correct — but MIREOT doesn't fix them
I checked dano.ttl directly and both observations hold:
dano:hasConfidence→rdfs:domain dano:DrawingElement, whiledano:DrawingElementMeta'srdfs:commentsays it holds "inference confidence and methods used". Genuine inconsistency in DAnO.dano:depicts→rdfs:domain dano:DisplayElement, no range asserted.
Worth noting this also closes a stated method caveat in our own title-block review, which read DAnO from its generated specification rather than its raw TTL and flagged "confirm domains/ranges before writing any alignment axiom". Now confirmed.
However — MIREOT cannot rewrite an external term's rdfs:domain. Its third URI asserts a local superclass only. If we want confidence hung off a different class, or depicts to apply to metadata:DrawingElement rather than dano:DisplayElement, that is a mint-our-own-term + rdfs:subPropertyOf/skos:closeMatch case. These are alignment cases, not MIREOT cases.
5. The bigger open question this surfaces: import vs. align for DAnO
This discussion assumes we will extract from DAnO. Two things in the repo point the other way:
- The title-block vocabulary review already concluded "align to
dano:, do not import it" — importing drags 16 drawing-element classes into a suite with a blocking DL gate for the sake of seven datatype properties. - The WIP
aec_provenancemodule (see PR Add UC-01 extraction concepts: title-block field scheme + provenance & geometry modules (and suite tidy-up) #76) already implements a third pattern: minimal local typing stubs for PROV-O terms +rdfs:isDefinedBy, explicitly notowl:imports, so the module stays self-contained and the reasoning CI resolves offline viacatalog-v001.xml.
That third pattern — local stub + alignment — isn't in the original MIREOT/SLME framing at all, and it's arguably the right answer for DAnO. It is now option 1 in the published rule, ahead of SLME and MIREOT, and #77 tracks settling DAnO specifically.
Two caveats on it, for the record:
- "Local stub + alignment" covers two distinct moves. Reusing the external IRI as a stub (what
aec_provenancedoes for PROV-O) is not the same as minting an ADIRO term and relating it withrdfs:subPropertyOf/skos:closeMatch(what the DAnO domain mismatches in §4 require). Both avoid importing; only the first keeps the external IRI as the subject. - It is a deliberate divergence from ODK, which treats an SLME extract as the normal way to reuse an external term, and where a term that is a name without a definition is discouraged.
versioning-vs-odk.mdnow records it as such rather than claiming parity. The trade is accepted because every ADIRO PR is reasoned over the merged suite behind a blocking DL gate. aec_provenanceitself is still on Add UC-01 extraction concepts: title-block field scheme + provenance & geometry modules (and suite tidy-up) #76, so until that merges the worked example is WIP rather than part of the published suite.
Net effect
- The
STARdefault stands; noBOT/TOPcase has been established. - No confirmed re-homing case, so no MIREOT case yet.
- The reasoner gate (RES-36) is done and blocking, and RES-59 is resolved — so RES-68 has no remaining blockers.
- The rule is now published in-repo via docs: make the external-ontology import rule normative in-repo #78; External ontology imports — strategy (MIREOT vs SLME) [ported from KB RES-A-12] #74 remains the decision record behind it.
- GeoSPARQL adoption remains deferred (UC-03: optional GeoSPARQL geometric-layer extension for reference symbols (deferred) #36); DAnO import-vs-align is now tracked in Decide DAnO import-vs-align (and whether aec_provenance supersedes its provenance terms) #77.
|
Closing this as resolved. This thread was the RES-59 review deliverable rather than an open question, and RES-59 is now resolved. The strategy it reviewed is settled and normative in the repo, so leaving the thread open risks reading as though the decision were still being re-litigated. Where things ended up:
Thanks @Zaalouk122576 - the review is what surfaced all three. The point about equivalence axioms and locality was right, and it is now part of the published rule: an upward Live follow-on work has moved to:
Anything that needs revisiting is best raised there, or as a new thread. |
Uh oh!
There was an error while loading. Please reload this page.
This discussion surveys the different import strategies, specifically MIREOT vs SLME, according to the below points.
Which geometry vocabulary (GeoSPARQL vs OMG+FOG vs other) fits ADIRO's needs, and which specific terms are required.
GeoSPARQL covers what ADIRO needs for now:
geo:Geometry/geo:hasGeometry/geo:asWKTfor attaching serialized geometry to a DrawingElement (CQ 1.3), the Simple Features relations (geo:sfContains,geo:sfWithin,geo:sfDisjoint,geo:sfOverlaps) for containment and spatial-inconsistency checks (CQ 2.3), and optionally the RCC8/RCC9 relations for reachability-style adjacency (CQ 4.1/4.3). Thegeof:*query functions also give us a way to compute these relations directly on geometry literals at extraction/validation time, which fits the SWRL/DL/SHACL reasoning stack without requiring pre-asserted topology. See the full breakdown in the discuission @70.DAnO — reference/URL and which terms/scope we'd take.
DAnO's scope (DAnO's canonical namespace/IRI is https://w3id.org/dano) for ADIRO breaks into three groups: provenance (
dano:inferredBy,dano:inferredWith,dano:inferredFrom,dano:inferredAt) for traceability of AI-generated models,dano:hasConfidencefor the human-in-the-loop flagging workflow, and the description/depicts relations (dano:DescriptionElement,dano:Composite,dano:depicts/dano:isDepictedBy) for relating extracted labels to geometry. See the full breakdown in the discuission @70.STAR as default vs needing BOT/TOP for specific cases.
From DAnO's perspective:
DAnO's core classes (
dano:DrawingElement,dano:DrawingElementMeta) are declared as direct subclasses ofowl:Thing, not of any upper ontology. This matters for the STAR-vs DAnO -BOT/TOP question, as since DAnO itself doesn't commit to an upper-ontology backbone, a STAR module extraction (SLME default) ofdano:DrawingElement/DisplayElement/depicts would pull in onlyowl:Thingas the superclass chain, there's no BOT/TOP scaffolding to preserve or lose. In other words, this gives us nothing to "need BOT/TOP for" on its own.From GeoSPARQL's perspective:
For
geo:sfContains,geo:sfWithin,geo:sfDisjoint,geo:sfOverlapseach only declarerdfs:domain geo:SpatialObject,rdfs:range geo:SpatialObject, and oneowl:equivalentPropertypointing to their Egenhofer counterpart (e.g.sfDisjoint owl:equivalentProperty ehDisjoint) ( seegeo.ttlin GeoSPARQL's codebase) . That's a flat, near-leaf axiom set. Therefore, a STAR (top+bottom) module seeded on these terms should stay small and self-contained, since the only outward edge is one paired equivalent property.The
geof:*query functions (geof:relate,geof:sfEquals,geof:sfWithin, RCC8 functions) are listed as simpleskos:Concept/skos:memberentries in a flat collection, not deep class-restriction hierarchies functions. These too are STAR-friendly, there is no need for BOT/TOP asymmetry since there's essentially no logical hierarchy to preserve one-sidedly.geo:Geometry/geo:SpatialObjectin the core ontology only (geo.ttl) also stay simple:Geometry rdfs:subClassOf SpatialObject.Where we would need BOT/TOP instead of default STAR: As soon as we extract against the merged
alignments.ttlgraph (rather than justgeo.ttl), these same terms pick up many more super/sub-class axioms simultaneously:geo:SpatialObjectgainsrdfs:subClassOf obo:BFO_0000001,prov:Entityandowl:equivalentClass pos:SpatialThing.geo:Geometrygainsrdfs:subClassOf obo:BFO_0000031,locn:Geometry,schema:GeoShape,cco:ont00000853.geo:hasGeometrygains four externalrdfs:subPropertyOfparents (locn:geometry,dcterms:spatial,cco:ont00001917, etc.).If ADIRO's extraction seed is run against this merged alignment graph, a plain STAR module would pull in the whole upward closure (BFO/PROV/CCO/schema.org axioms) unintentionally. On the other hand, A BOT (bottom) module, which only keeps subclass/subproperty axioms going down from the seed, discarding the upward superclass chain , is the safer choice there if we only want
geo:Geometry/geo:SpatialObject/geo:hasGeometrywithout dragging in BFO/PROV/CCO.Any genuine re-homing cases that justify MIREOT over SLME.
From DAnO's perspective:
The one place where DAnO's native placement of a term could plausibly conflict with how ADIRO wants to model it is
dano:DrawingElementMeta, which the class comment ( in DAnO's repodano.ttl) says holds "inference confidence and methods used", yet the actualhasConfidenceaxiom is asserted with domaindano:DrawingElement, not DrawingElementMeta . If ADIRO's provenance model (CQG 3: confidence scores, inference records) wants to consolidate all of these under a single "prediction/annotation" superclass distinct from DAnO's split, that's a genuine re-homing case, we would want to use MIREOT rather than SLME/STAR.Similarly,
dano:depicts/dano:isDepictedByis asserted with domaindano:DisplayElementonly , if ADIRO's wants depicts to apply more broadly (e.g. directly offmetadata:DrawingElementrather than DAnO's narrower DisplayElement subclass), that's another re-homing case that favours MIREOT's selective/asserted-axiom approach over a STAR-module extraction that would otherwise import DAnO's whole DisplayElement branch just to keep the domain typing intact.From GeoSPARQL's perspective:
Several of our candidate terms are asserted with simultaneous superclass/superproperty links to multiple external upper ontologies in the same alignment graph:
geo:SpatialObject→obo:BFO_0000001andprov:Entityandowl:equivalentClass pos:SpatialThingall at once.geo:Feature→dcterms:Location,prov:Location,gn:Class, plus anowl:equivalentClassunion of three BFO classes andas:Place.geo:hasGeometry→ four different subPropertyOf parents from four separate vocabularies.This is exactly the scenario where SLME's automatic transitive closure becomes unpredictable/unstable: since these upper alignments are all declared in one shared
alignments.ttlgraph, extracting a syntactic-locality module seeded ongeo:hasGeometryorgeo:SpatialObjectwould non-deterministically drag in whichever external upper-ontology parent happens to co-reside in the loaded graph at the time "a form of unwanted "re-homing" " (the term becoming logically subordinate to BFO/PROV/CCO/GeoNames axioms we never asked for). On the other hand, MIREOT, where we manually restate just the term (and its single, intendedrdfs:subClassOf, e.g.Geometry ⊑ SpatialObjectfromgeo.ttlalone) without pulling the alignment file at all, is the safer, deterministic choice for these specific terms.By contrast, the simpler leaf terms (
geo:sfContains,geo:sfWithin,geo:sfDisjoint,geo:sfOverlaps, thegeof:*functions) don't show this multi-parent pattern ingeo.ttlitself, for those, SLME/STAR extraction fromgeo.ttlalone is safe, since there's no re-homing risk as long as we don't also loadalignments.ttl.All reactions