Choose the right representation
Distinguish a thing, a value, a class and a relationship; then choose an encoding that preserves the distinction when the question changes.
Technical modelling guide Modelling judgement Chapter 02
Method basisUpdated
A graph can give almost anything a node. The modelling question is whether that node identifies something the domain needs to distinguish. The useful model is neither the one with the most classes nor the one with the fewest triples: it is the one whose commitments explain the evidence and survive reuse.
Begin with what the words refer to
In the fictional Harbour Court dossier, “property” appears on a listing, an inspection report and a title record. One building contains two dwellings. The address helps readers find the dossier, but it cannot tell them whether an inspection concerned Flat 1, Flat 2 or the whole building. Start by separating those possible subjects before creating a general Property class.
A particular building is an individual. Building is a proposed class of individuals. An identifier refers to the particular building; a label gives readers a name. A description makes statements about it. A report is an information artefact carrying such statements. Changing a label, correcting a statement and replacing a building are therefore different changes, even when a screen presents them together.
This separation prevents a common error: turning a useful join into identity. The title record and building may be related, but owl:sameAs would assert that they are the same individual. A shared address is not a reason for that assertion. Nor do two different IRIs, on their own, establish that their referents must differ; the intended identities still need to be explained.
- A direct property can express a domain relation between two instances: for example, Flat 1 partOf Building B. Its domain and range carry formal semantics.
- A material relator is an independently justified dependent domain entity: a candidate Agreement has its own attributes and mediates at least two bearers. The scoped method requires both.
- An assertion record is about the claim “Flat 1 partOf Building B” and records information such as its author and assertion time. This statement description is not thereby a material relator.
A class, a controlled value or an ordinary value?
A class says what its instances are. A controlled concept gives a governed meaning to a classification value. A literal supplies a value such as text, a number or a date. These are different jobs. “Recorded condition: poor” does not require a PoorBuilding class simply because a form offers a “poor” option.
Suppose the group needs a stable set of condition categories, each with a definition and scope note. A SKOS concept scheme can govern that vocabulary. The assessed thing is related to an assessment or classification using a concept from the scheme. The concept is not the assessed building, the assessment act or a particular measurement. Its label can change without turning every classified building into a new kind.
By contrast, a report version has a distinct production history and can be cited or revised independently. That provides reasons to identify it as a resource. A literal report number may remain useful, but a string cannot itself carry the report’s relationships. “Resource or literal?” is therefore a question about needed identity and meaning, not whether the source cell happens to contain text.
Use the smallest representation that preserves the required distinctions. A recorded numerical value does not need a new object merely because RDF permits one. If two observations have different methods, times or subjects, however, flattening their values onto one bearer loses which value belongs to which observation. Additional structure then has a concrete consumer: the question that distinguishes those observations.
Name the relation before drawing the arrow
Write the proposed statement as a sentence: “this report describes this inspection”; “this inspection has this subject”; “this dwelling is part of this building”. Examine its direction and its endpoints. A generic “has” property may conceal distinctions that the practitioners already understand.
Part-of, instance-of and subclass-of are particularly easy to confuse. Flat 1 being part of Harbour Court does not make Dwelling a subclass of Building. Likewise, an inspection being about a building does not make Inspection a type of Building. rdfs:subClassOf means that every instance of the subclass is an instance of the superclass; it is not a folder, topic link or part/whole relation. RDFS §3.4 defines that commitment.
A domain verb also exposes placement errors. “Was inspected on” is about an inspection event; “report generated on” is about the report’s production. Attaching a single “date” to the building forces readers to infer which event it means. ODR-0042 places a property at the most specific class that introduces its meaning, rather than hoisting everything onto the page’s main subject.
Two reports need not mean two inspections
This small RDF illustration deliberately identifies the inspection separately from its report versions. The prefixes identify teaching terms, not proposed OPDA vocabulary. Definitions, applicability declarations, provenance and constraints are omitted so that the referent distinction is visible.
@prefix ex: <https://example.org/harbour-court/> .
@prefix model: <https://example.org/teaching-model/> .
ex:inspection-12-aug a model:Inspection ;
model:hasSubject ex:flat-1 .
ex:report-v1 a model:ReportVersion ;
model:reportsInspection ex:inspection-12-aug .
ex:report-v2 a model:ReportVersion ;
model:reportsInspection ex:inspection-12-aug . The two report resources point to the same inspection resource. A consumer can now ask for the inspected subject without selecting the latest document date. The graph still does not say why the report was revised, whether either conclusion is accurate or which time applies to each assertion. Those are additional modelling questions, not defects that a longer identifier resolves.
Four structures that must not be collapsed
Once relationships need dates, sources or qualifications, their diagrams can begin to look alike. Similar shapes do not establish similar meanings. Keep these four cases explicit:
| Structure | Question it answers | What does not follow |
|---|---|---|
| Qualified association | In what context, period or capacity did these participants relate? | The association is not automatically a material relator. |
| Material relator | What dependent domain relationship mediates bearers and has characteristics of its own? | Adding two endpoints and a date is not a philosophical justification. |
| Statement reification | Which particular assertion has this source, attribution or confidence? | Addressing the statement neither asserts its truth nor supplies a domain lifecycle. |
| Term mapping | How do these governed meanings correspond, and why? | A correspondence does not make their instances identical. |
For Maya’s participation as seller, a direct relation may answer a narrow question. When the participation’s period and transaction must be distinguished from another simultaneous participation, a qualified resource gives those facts an unambiguous home. This is the motivation behind n-ary relation patterns; it is not an instruction to reify every join. The W3C n-ary relations Note illustrates several reasons to represent additional relation information.
A material-relator interpretation needs the further analysis in the dependence section. A statement reifier follows the selected RDF 1.2 carrier profile. A reviewed mapping follows SKOS, SSSOM and SEMAPV requirements. Do not interchange these carriers merely because each can be drawn as a box with incoming and outgoing arrows.
Say where a property applies without changing its semantics
Suppose reportsInspection is intended only for report versions. A truthful RDFS domain would commit every subject using that property to the selected report class. It does not demand that a particular input explicitly contains that type statement, and it does not check the presence of an inspection reference. Those validation questions belong to SHACL.
If a property is intended for either reports or another distinct information-resource type, repeating RDFS domains does not express the alternatives: it entails both memberships. Switching off local inference cannot make that published meaning disappear. RDFS domain semantics remain in force.
ODR-0037 therefore requires one complete mechanism on each property side: one truthful RDFS IRI, useful Schema.org inclusion hints, or a governed waiver. Subject and value sides are decided separately; do not mix mechanism families on one side. SHACL then expresses required types, counts, datatypes and conditional obligations for the chosen targets. A shape is not a substitute for documenting the property’s applicability.
Try to break the proposed representation
Give reviewers two reports from one inspection, two inspections of different dwellings at one address, and a corrected label on an unchanged resource. Ask them to identify the subject, event and evidence without looking at the source layout. If they cannot, revise the representation before adding more fields.
Then remove an optional qualifier. Does the remaining direct statement still have an intelligible meaning? Introduce a second participation. Can its dates be assigned unambiguously? These tests reveal whether a relation needs qualification and whether the proposed identity is doing useful work. They do not require domain participants to read Turtle.
Decision basis and further reading
ODR-0047 sets the domain-structure concern; ODR-0042 governs property placement and qualification; ODR-0037 governs exclusive applicability. ODR-0043 defines the bounded OWL surface, while ODR-0058 defines the RDF 1.2 statement carrier.
The selected method and its standards profiles are normative within their recorded scope. The examples above are explanatory models, not declarations of current-candidate implementation or approval.
Comments
Loading comments…
Sign in to post a comment