About dpv#
This page says how the schema reads DPV, where each part comes from, and what the model leaves out. It is for someone who will use the data or change the converter.
Terms are records#
DPV states each concept three ways at once: dpv:Marketing a rdfs:Class, skos:Concept, dpv:Purpose. A LinkML class cannot also be a value, so a schema with one class per term could not say that a process’s hasPurpose is dpv:Marketing.
So the schema gives a class to each concept that terms are instances of, a kind, such as DpvPurpose for dpv:Purpose. Every term of the vocabularies is then a record. DpvConcept holds a concept’s record, DpvProperty a property’s, and DpvResource a thing a use case describes, such as a process.
| What DPV states | What the schema holds |
|---|---|
A type a concept holds beyond rdfs:Class and skos:Concept | A class under its own IRI, 149 of them, each under the nearest kind above it along skos:broader |
| A property | A slot under its own IRI, with DPV’s local name as its key, 307 of them |
| A term | A record in src/dpv/data/, its rdf_type naming each kind it is an instance of |
Where two namespaces define one property name, the core keeps the name and the other takes its namespace first: hasRiskLevel and eu_aiact_hasRiskLevel. A kind’s class takes the namespace the same way.
A kind’s record and its class are the same thing seen two ways. dpv:Purpose is a record, and DpvPurpose is the class a host names as a concept’s type. So project/owl/ describes this schema, and upstream/*.ttl is DPV’s ontology.
The parts of a release#
src/dpv/data/ holds one file per namespace family: the core, personal data, technology, artificial intelligence, risk, justifications, locations, laws and authorities, sectors and standards. A further file holds the contributors, and another the examples.
scripts/dpv_to_linkml.py writes each file from upstream/, along with dpv_terms.yaml and the kinds’ modules. It stops on a predicate no table names, so a new edition cannot lose a fact unnoticed. tests/test_coverage.py walks the Turtle both ways, and tests/test_catalogue.py compares the catalogue’s JSON-LD projection with it triple by triple.
What the examples add#
DPV’s 88 example files are its one published account of how the terms join up in data. dex:E0026, for one, describes a process with a purpose and two nested processes. Each file’s statements are records in dpv-2.3-examples.yaml. The file declares no prefixes, so the converter supplies them: ex: as DPV’s page says, and exA:, exB: and a bare colon, which DPV leaves unstated, under http://example.com.
A value in a record must look like a reference, and tests/dpv_rules.py holds it to a term that exists. An unknown key fails the validator, and that is the gain over a schema of classes with no slots.
What stays open#
DPV uses dcam:domainIncludes and dcam:rangeIncludes, not rdfs:domain and rdfs:range. They say what a property usually joins, not what it must. So DpvResource holds any property, and the validator refuses no value for leaving those classes. tests/test_examples.py reads them anyway, and tests/example_departures.yaml lists each place DPV’s own examples leave them: 58 values, 65 subjects and 30 literals where a range names a class.
A statement the slots cannot hold, such as a literal under hasLocation or a predicate with no slot, goes in the resource’s statements.
What the records leave out#
skos:narrower, which only the locations state, 68,032 times. 4,959 are the inverse ofskos:broader. The other 63,073 list what eachloc:non-Xholds: every country but X, and but X’s member countries when X is a union.tests/test_coverage.pyregenerates every pair.schema:domainIncludesandschema:rangeIncludes, which repeat thedcam:pair on every property.- The statements of how DPV’s site publishes a file:
prof:hasResourceand the 337 resource descriptors. dpv-27560, DPV’s schema for consent records. Its Turtle file names four classes and states nothing about them, so there is nothing to render.
What the data shows#
These are in DPV’s own files. tests/test_catalogue.py and tests/dpv_rules.py pin them, so a later edition that fixes one fails a test.
- 18 names under DPV’s namespaces are used and defined nowhere, such as
dpv:Systemandtech:AI. One is an IRI with a line break in it. - 60 statements use
skos:sameAs, and each vocabulary headerrdfs:Label. Neither is a term of its vocabulary. - Three example files do not parse:
E0030,E0059andE0061. dcat:Resourcestates two creation dates, and five terms name two vocabularies underrdfs:isDefinedBy.- 153 terms have no definition, nearly all laws and authorities.
Mappings#
src/dpv/mappings/ renders the two mapping files DPV publishes in machine-readable form, for ISO/IEC 29100:2024 terms and for ODRL. A row reads with DPV’s term as its subject, as DPV’s own statements do, and skos:broadMatch stands for DPV’s skos:broader. tests/test_mappings.py holds each row to both the CSV and the Turtle.
DPV’s other mapping folders, for DCAT, Dublin Core, gist, PROV, schema.org and SEMIC, hold a web page only, so nothing renders them. dpv-odrl.ttl does not parse, so the test reads its comment lines, and one row names dpv:hasTechnology, which 2.3 does not state.