Repository navigation
ADIRO title-block properties — Data vs Object property design (research note) #61
Replies: 2 comments 1 reply
|
Thanks @AhmedElnagar1 for this amazing work! Please find below my recommendations: 1- We need to collect a sample size that is large as possible to record which fields actually appear. If a field appears on less than ~40% of sheets, it should be deferred to avoid bloating the ontology with empty properties. 2- Please avoid implementing fields inherited from manufacturing heritage, such as hasResponsibleDepartment, hasTechnicalReference, and hasProjectionMethod, because they are rarely a labelled cell on a building drawing. While these are valid fields in the ISO 7200 standard they reflect mechanical engineering traditions where first- or third-angle projection symbols are critical, however, in architectural and structural AEC sets, these symbols and administrative departments are almost never printed. If we implement them, this will add unnecessary complexity to the ontology and creates extraction targets for data that sampled evidence suggests the model will not find. 3- I suggest that we focus on the implementation of fields that are related to the standards that are only relevant to our use cases ( Targeting a Minimal Set of Standards): Moreover, we need to use ROBOT to avoid conflicting terms ( as the Redundant terms mentioned in the article), also reconciling these terms via ROBOT ensures we follow the DAnO principle of "reuse whenever reasonable" (NFR3) to maintain precise terminology and avoid "overly strict property assertions" that hinder adaptability (NFR2). In addition, using a semantic importing prevents the "duplication problem" where the title-block vocabulary might otherwise accidentally mint twins for existing sheet-level metadata. 4- I suggest creating a separate Titleblock ontology that links to aec_drawing_metadata as this permits a modular approach that is consistent with the finding that a title block is not the same thing as a document: the title block is a physical graphical region that asserts metadata, while the document is the conceptual deliverable those assertions describe. Also, for an ML extraction pipeline, this distinction is load-bearing because it allows the system to model cases where two different sheets carry conflicting values for the same document. By attaching extracted values to an adiro:Titleblock individual first, you can promote them to the Document level only once they are validated. 5- To ensure we are not overdoing it with external references, I suggest that we focus on what we immediately need by targeting the 29 core terms + any extra terms needed from the standards mentioned in point number 3 ( that are related to our use cases), and deferring the 20 marginal ones inherited from standards like ISO 7200’s manufacturing heritage. We can also treat standards like IEC 82045 as a superset vocabulary to verify if a field belongs, but do not feel obligated to implement every clause. Finally, DAnO the dano:depicts property to link drawing elements to external ontologies like ifcOWL, BEO, or BOT without requiring a full import of those frameworks :
Quoted from Ontology-based reasoning in automatic floor plan analysis ( DanO’s paper) Therefore, we can leverage equivalentTo or mapping assertions so extracted data can be written back into BIM. |
|
Thanks @Zaalouk122576
|
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
ADIRO title-block properties — Data vs Object property design
Research note for ontology development.
1. Standards surveyed
1.1 Core — defines the fields themselves
1:50 – m,cm.DocumentType/ Planart vocabulary and for scale parsing on German projects.SCALE: NONEas a legitimate value. Revision letters capped at 2 characters.1.2 Schema & metadata standards — tells us how to model the fields
https://standards.iso.org/iso/21597/-1/ed-1/en/Container.rdf(ct:prefix). Verified contents:Party/Person/Organisation;Documentwith internal / external / secured / encrypted / folder subclasses;ContainerDescription;Linkset; and file-level properties (filename,filetype,format,checksum,creationDate,modificationDate,priorVersion,nextVersion,createdBy,modifiedBy,publishedBy). Annex B addresses Dublin Core interoperability; Annex D covers SHACL validation. Scope, assessed honestly: this is container and file mechanics plus party provenance. It supplies almost none of the title-block content fields in §6–§8 — there is noidentifier/name/description/titleinct:itself (those arrive via Dublin Core), nocheckedBy/approvedBy, and it spellsOrganisation. Treat as an export contract, not a vocabulary source — see finding 1.dici:Information module, CC BY 4.0dici:hasContainerState→dici:InformationContainerStatewith named individualsInitial/WorkInProgress/Shared/Published/Archived. Originator is partial (dici:isCreatedBy, generic — it does not disambiguate originator vs legal owner vs client, which is the hard part). Volume/System and Level/Location have only conceptual analogues (ProductBreakdownStructure,LocationBreakdownStructure), not the printed code strings. Suitability S0–S7, Discipline, Type and sequential number are absent — and the Agents module (dica:) has role classes but explicitly no discipline vocabulary. Nothing for title blocks at all: noRevisionclass, no scale, paper size, projection method, plan key, confidentiality, or checked/approved-by. Costs: importing it pulls BFO/IAO into a suite governed by a blocking OWL 2 DL gate, and the source repo has been dormant since 2022-06-30, still at v0.5. Verdict: mint our own. Borrow the container-state modelling pattern; treat DiCon alignment as optional on-demand work, like ICDD.IfcDocumentInformation(ISO 16739)IfcDocumentStatusEnumandIfcDocumentConfidentialityEnum. Needed anyway for round-tripping into BIM.DocumentPurpose,DocumentContent,DocumentFormatand links actors to ICDD container parties. Reusable for the "why does this document exist" dimension.hasClassification.dcterms:for creator/title/identifier/date,prov:for the extraction provenance chain,skos:for every controlled vocabulary below.1.3 Note on two standards sometimes cited in error
ISO 9050 (glazing — light transmittance, solar factor) and ISO 9060 (classification of pyranometers) are unrelated to title blocks and are almost certainly a mis-transcription. Likely intended candidates are ISO 7200 (fields), ISO 5457 (sheet layout), or ISO 21597 (container ontology) — all covered above.
2. Six findings from the wider survey that change the design
Align to ICDD for interchange, not for coverage. (Revised 2026-08-11. This finding originally read "don't design from scratch — align to ICDD", called it "the single most important" standard for ADIRO, and said "do not reinvent these — subclass and align". That overstated ICDD's coverage, and the overstatement did real damage: it left the alignment work parked behind "obtain the standards" while implying ICDD would supply fields it does not contain.) What ICDD genuinely earns is an export contract. It is the ISO container format for linked document delivery, and a published ICDD ↔ DIN SPEC 91391-2 mapping exists — so it is the interchange vocabulary if a graph ever has to leave or enter the organisation as a deliverable. The useful alignment is roughly four axioms:
adiro:Person/adiro:Organization→ct:Person/ct:Organisation,adiro:Document→ct:Documentas the anchor forassertsMetadataFor, and file provenance where it maps. Everything else inct:is file mechanics ADIRO does not need, and the ISO 7200 role trio (createdBy/checkedBy/approvedBy) has nothing to align to and stays ADIRO-native regardless. Make the alignment conditional on a stated ICDD / openCDE deliverable requirement: it is purely additive, so it is cheap to add when a real requirement appears, and premature before one exists.A title block is not the same thing as a document. ISO 7200 fields describe the document; the title block is the graphical region that asserts them. For an ML extraction pipeline this distinction is load-bearing: two sheets can carry conflicting values for the same document. Recommendation: model
adiro:Titleblockas a distinct class with a bounding box and anassertsMetadataForlink to the document, and attach extracted values to the Titleblock, promoting them to the Document only once validated. This is what makes the "surface conflicts for human review" competency questions answerable.Several fields are deliberately outside the title block. ISO 7200 §4 holds the number of title-block fields to a minimum, with scale, projection symbol, general tolerance and surface-texture requirements handled dynamically and presented outside the title block only when used. So ADIRO needs these as document properties even though the detector may not find them inside the title-block region — and ASME puts the whole Revision History block outside the title block too.
Revision must be a class, and the standards agree. ISO 7200 offers revision index or date of issue; ISO 19650 adds suitability; ASME Y14.35 mandates a revision history with letter + date + description + approver per row. Flattening this into literals loses the history. See §5.
Status is three different things. ISO 7200 document status (lifecycle), ISO 19650 suitability (S0–S7, fitness-for-purpose of this revision), and ISO 19650 container state (a CDE location property) are routinely conflated in practice. Keep three separate object properties.
Confidentiality is a first-class field, not an afterthought. IFC has
IfcDocumentConfidentialityEnumon every document. Given that public-sector work is governed by national classification frameworks (in Germany, the VS grades), ADIRO should carryadiro:hasConfidentialityClassificationas an object property pointing to a controlled vocabulary that can hold both IFC's generic values and national grades. This also lets SHACL enforce "no restricted document may be linked to an unclassified container."3. The decision rule: Data Property vs Object Property
Use a Data Property (
xsd:string,xsd:date,xsd:integer,xsd:float) when the value:Use an Object Property when the value:
Organizationindividual, hundreds ofhasLegalOwnertriples)Practical tie-breaker: if a competency question would ever ask "find all documents where X is the same", X is an Object Property.
4. Supporting classes
adiro:Documentct:Document(ISO 21597-1) andIfcDocumentInformationadiro:TitleblockassertsMetadataFor→ Document. See finding 2adiro:Sheetadiro:Organization,adiro:Personct:Partysubtypes;IfcActorSelectequivalentadiro:Projectadiro:Revisionadiro:DocumentTypeadiro:Disciplineadiro:SuitabilityStatusadiro:DocumentLifecycleStatusIfcDocumentStatusEnumadiro:ContainerStateadiro:ConfidentialityClassificationadiro:ClassificationCodeadiro:Languageadiro:SourceFile5.
adiro:Revisionas its own classA document then has
adiro:hasRevision→ manyRevisionindividuals, plusadiro:hasCurrentRevision→ one. Benefits: full history, per-revision approver (which is what the audit trail actually needs), and the ability to answer "what changed between P02 and P03".6. ISO 7200 field mapping
Clause number, ISO field name and obligation status (M = mandatory, O = optional) are given so the mapping is checkable; see the standards notice at the top of this page. Definitions, recommended character lengths and layout are not reproduced — consult ISO 7200:2004 for those.
hasLegalOwnerOrganizationidentificationNumberxsd:stringhasRevisionRevision(§5)hasRevisionRevision(§5)sheetNumberxsd:stringnumberOfSheetsxsd:integerhasLanguageLanguage(ISO 639)titlerdf:langStringsupplementaryTitlerdf:langStringhasResponsibleDepartmentOrganizationhasTechnicalReferencePersonapprovedByPersoncreatedByPersonhasDocumentTypeDocumentTypehasClassification/keywordxsd:stringhasLifecycleStatusDocumentLifecycleStatuspageNumberxsd:integernumberOfPagesxsd:integerpaperSizexsd:string("A1")Note on 5.1.4 / 5.1.5: ISO 7200 permits using the date-of-issue field alone in place of a revision index. ADIRO should therefore not make
revisionIndexmandatory onRevision— a SHACL shape requiring at least one of index or date is the correct constraint.7. ISO 19650 additions
partOfProjectProjecthasOriginatorOrganizationvolumeSystemCodelevelLocationCodeA01,B03M)hasDocumentTypehasRolehasDisciplineDisciplinesequentialNumberhasRevision/hasSuitabilityStatushasContainerStateContainerState8. Fields present on real drawings but absent from ISO 7200
These are the ones a naïve ISO-7200-only ontology will silently drop. All are commonly inside AEC title blocks even though ISO 7200 pushes some of them outside.
scalexsd:stringNONE. Keep as string — ratios,NTS, and multi-scale sheets all occurdimensionUnits1:50 – m,cm— parse and store separately from scalehasProjectionMethodhasClientOrganizationcheckedByPersonplanKeypurpose,intendedUseIfcDocumentInformationvalidFrom,validUntilxsd:dateIfcDocumentInformationhasConfidentialityClassificationIfcDocumentConfidentialityEnum+ national grades. See finding 6organizationIdentifierpaperSizehasNorthPointOrientationxsd:floathasAnnotationBlock9. IFC round-trip alignment
Recommended
owl:equivalentProperty/ mapping assertions so extracted title-block data can be written back into BIM:IfcDocumentInformationIdentificationidentificationNumberNametitleDescriptionsupplementaryTitleLocationfileLocationPurpose/IntendedUse/Scopepurpose/intendedUse/scopeRevisionhasCurrentRevision→revisionIndexDocumentOwnerhasLegalOwnerEditorscreatedBy(multi-valued)CreationTime/LastRevisionTimeRevisiondateOfIssueElectronicFormatfileFormatConfidentialityhasConfidentialityClassificationStatushasLifecycleStatusNote the mismatch to flag: IFC collapses revision into a single label and two timestamps, so a full ADIRO revision history cannot round-trip losslessly into IFC. Export the current revision only, and keep the history in the graph.
10. Extraction-provenance properties (ADIRO-specific)
Not in any of the standards, but required by the human-in-the-loop and explainability competency questions:
extractedFromSourceFilehasBoundingBoxBoundingBoxextractionConfidencexsd:floatextractedByModelMLModelprov:wasGeneratedByhasValidationStatusconflictsWithModelling note: if per-value confidence and provenance are needed (they are), plain triples won't carry them. Either use RDF-star (
<< :doc adiro:title "..." >> adiro:extractionConfidence 0.94) or reify each extracted value as anadiro:ExtractedValueindividual. The second is uglier but has far better tool support in SPARQL-adjacent stacks and in SHACL.11. Summary
11.1 Consolidated inventory — source, meaning, and an assessment of usefulness
Compiled from §4–§10 so the whole vocabulary can be read in one place. "Source" is where the term comes from:
a standard clause, an external vocabulary, or ADIRO where we mint it because nothing existing covers it
(§1.4). "Denotes" is ADIRO's own plain-language meaning — not a standard's definition, which is not
reproduced here (see the standards notice at the top).
"Assessment" is a critical judgement, not a description. Each term is rated on whether it is worth minting
and maintaining:
Classes
Titleblockdm:Titleblock; reuse, do not re-mintOrganizationct:Partysubtype; IFCIfcActorSelectPersonct:Partysubtypedm:Person; reuseProjectdm:Project; reuseRevisiondm:DrawingRevisionfirst (q7)SourceFileBoundingBoxDocumentTypeDisciplineDocumentct:Document; IFCIfcDocumentInformationassertsMetadataForto mean anything, but nothing promotes to it until validation exists. Mint minimally; resist adding properties to itSuitabilityStatusConfidentialityClassificationDocumentLifecycleStatusIfcDocumentStatusEnumdm:StatusCodealready occupies this ground informally. Reconcile before mintingContainerStateClassificationCodeLanguageMLModelprov:extractedByModelpointing at a literal would serve every current need. Promote to a class when there is a second consumerSheetdm:DrawingSheetalready exists and is the class the CVAT pipeline labels. Do not minttb:Sheet; usedm:DrawingSheetObject properties
assertsMetadataForTitleblockto theDocumentit makes claims abouthasClientcreatedBycheckedBycreatedBy; no external equivalent existsapprovedByhasRevisionhasDisciplinehasDocumentTypepartOfProjectextractedFromhasBoundingBoxhasValidationStatushasLegalOwnerhasClient, and frequently the same string on a BH-authored sheet. Expect this pair to dominate the mis-binding matrixhasOriginatorhasSuitabilityStatusRevision, not the document; that placement is the valuable parthasLifecycleStatusdm:hasStatusCodehasConfidentialityClassificationhasCurrentRevisionhasRevision+ index/date ordering, but it is the only thing IFC can round-trip (§9) and the query you actually run most. Keep as a materialised convenience, and say so in the commentextractedByModelprov:MLModel: point it at a literal until there is a reason not toconflictsWithhasRolehasDiscipline— on most projects the Role and Discipline segments carry near-identical information, and the value is a one- or two-character code parsed out of the drawing number rather than a printed field. Pick one unless sampling shows they divergesupersedesdateOfIssue. An explicit edge is only worth it if revisions branch, which drawings essentially never dohasContainerStatehasClassificationClassificationCodehasLanguagehasResponsibleDepartmenthasTechnicalReferencehasProjectionMethodhasAnnotationBlockaec_drawing_metadataalready models these regions (:Legend,:Note,:TextualNote) with:contains. This is region modelling, which belongs in that module and already lives thereData properties
identificationNumberdm:drawingIdentifier(§1.1)titlerdf:langString, because bilingual title blocks are normal. Reconcile withdm:drawingTitlescaleNTSand multi-scale sheets all occur. Reconcile withdm:hasScalerevisionIndexsh:orshape. Reconcile withdm:revisionCodedateOfIssueDD.MMvsMM.DDambiguity makes it the most interesting field to measure. Reconcile withdm:issueDatesheetNumberplanKeyextractionHintsupplementaryTitletitlenumberOfSheetssheetNumberrevisionDescriptiondimensionUnitsscaleneeds parsing rather than storing rawvolumeSystemCodelevelLocationCodeextractionConfidencesequentialNumberidentificationNumberwith little independent query value. Store the whole identifier and parse on demandkeywordpurposeIfcDocumentInformationhasSuitabilityStatus— which is the purpose, expressed as a controlled code. Never a printed title-block cell. Included because IFC has it, not because we need itintendedUseIfcDocumentInformationpurposetoo. Two IFC attributes that collapse into suitability for our purposesvalidFromIfcDocumentInformationvalidUntilIfcDocumentInformationorganizationIdentifierhasNorthPointOrientationdm:hasOrientation→dm:OrientationValuealready exists. Check before minting; it belongs to the region ontology if anywherepageNumbersheetNumber— ISO 7200 separates sheet from page for multi-page documents in a manufacturing context. On a drawing they are the same thing. Mint onenumberOfPagesnumberOfSheets— same argumentpaperSizeA1)dm:sheetSizealready exists and holds exactly this. It is also listed twice in this note (§6 and §8), which is how the duplication went unnoticedWhat the assessment adds up to
Five rows should not be minted at all.
Sheet,hasAnnotationBlock,pageNumber,numberOfPagesandpaperSizeeach duplicate something that already exists — four of them inaec_drawing_metadata, andpageNumber/numberOfPagesduplicate their own sheet-level twins in this very table. That is on top of the~12 term overlaps with
aec_drawing_metadataalready identified separately, so the real duplication problem islarger than either count alone suggests.
Twenty more are marginal, and they cluster revealingly. Most come from one of three places:
hasResponsibleDepartment,hasTechnicalReference,hasProjectionMethod, the page/sheet split. Real fields in the standard; rare on building drawings.purpose,intendedUse,validFrom,validUntil. Added because IFC has anattribute, not because a title block prints one, and
purpose/intendedUselargely restate suitability.sequentialNumber,hasRole, and arguablyvolumeSystemCode/levelLocationCode. These are parsed out of the drawing number rather than read from labelled cells. Theyare outputs of a parser, not extraction targets, and treating them as fields for the model to find invites
confident nonsense.
Plus two that are simply out of scope for a title-block vocabulary:
ContainerState/hasContainerStatedescribe where a file sits in a CDE and are never printed on a sheet.So the honest shape of the work is closer to 48 terms than 73 — 29 core plus 19 useful — with 20 deferred
pending evidence and 5 dropped. That materially changes pass 4, which is where most of the marginal rows live:
it may be a much smaller pass than "the extension list" implies, or worth splitting into "extension properties
we have evidence for" and "the rest".
The one thing that would settle it is next step 3 — sample 20–30 real title blocks and record which fields
actually appear. Every 🟠 above is a testable claim, and running that sample before pass 4 turns this table from
opinion into a measurement. It is now the highest-value open task in this note for exactly that reason.
11.2 Revisions to the assessment — current position
Applied after two rounds of external review. This is the authoritative list; §11.1's emoji are the original
assessment.
Disciplinedcommon:Disciplineis an existing class hierarchy that UC-01 reuses. UC-01 deliberately dropped adisciplineCodestring so thatrdf:typechecks give automatic sub-discipline rollup; a flat SKOS scheme loses thatscaledm:hasScale(string) plus UC-07scaleRatio(decimal) already cover it. The field is still core to extraction; it binds todm:hasScalehasNorthPointOrientationnorthArrowAngleis the same measurement, better specified, and drawing-level rather than title-block-levelhasLegalOwnerDocumentType,hasDocumentTypeinferredAt; re-extraction comparison is impossible without oneMLModelSourceFiledano:inferredFromis a stringBoundingBoxdano:hasGeometryis more general and reusableextractedByModeldano:inferredBy(actor) anddano:inferredWith(algorithm) is a better decompositionhasResponsibleDepartment,hasTechnicalReference,hasProjectionMethodTwo rules agreed in review that apply to every remaining 🟠 row:
vocabulary with properties nothing populates. The field-frequency survey is therefore the gating input for the
whole 🟠 set — not a nice-to-have.
ISO 21597-1 (ICDD) as the anchor for document-to-document links needed by cross-sheet linking. Treat IEC 82045
as a completeness check, not an implementation obligation. Where an external ontology already models
something, link out to it (DANO's
depictspattern) rather than importing it wholesale.Recommended next steps
Obtain the normative— done, and it should never have been a blocker. The file is publicly resolvable (URL and verifiedContainer.rdffrom ISO 21597-1ct:local names in §1.2). Nothing in the title-block vocabulary is to be held on ICDD acquisition again.Evaluate DiCon's[closed 2026-08-11 — evaluated, negative.] See §1.2: it covers container state exactly and almost nothing else of §7, carries a BFO/IAO dependency, and has been dormant since 2022. Mint §7 as specified. Borrow DiCon's container-state pattern (named individuals rather than a closed enum) as independent corroboration of the modelling choice, and treat any DiCon alignment as optional on-demand work.dici:Information module before minting the §7 ISO 19650 properties.Standards to obtain
ISO 7200:2004 · ISO 19650-2:2018 + UK NA · ISO 21597-1:2020 (+ ontology files) · DIN 1356-1:2024-04 · DIN SPEC 91391-1:2019 · IEC 82045-2:2004 · IFC 4.3 schema docs (free online)
Related notes published here: Building the ADIRO title-block TBox — module design and build plan (the build spec derived from this research) · Evaluating a predicted knowledge graph against a ground-truth knowledge graph · Ground-truth labelling: ontology-bound keys, or verbatim captions?
All reactions