NFDI4DS Ontology for DS & AI
- NFDI4DS Ontology is a domain-specific semantic framework for DS/AI research, offering modular, BFO-aligned integration for FAIR data exchange.
- It systematically represents research artifacts, organizational roles, and digital resources, enabling efficient graph-based querying and cross-referential analysis.
- The ontology underpins the NFDI4DS Knowledge Graph by simplifying complex relationships with SWRL rules and shortcut properties for effective knowledge retrieval.
The NFDI4DS Ontology (NFDI4DSO) is a domain-specific ontology for Data Science (DS) and AI developed within the NFDI4DataScience (NFDI4DS) consortium. It is presented as a modular extension of NFDICore, explicitly mapped to the Basic Formal Ontology (BFO), and intended to serve as the semantic foundation of the emerging NFDI4DS Knowledge Graph (NFDI4DS-KG). Its stated purpose is to provide a semantically precise, interoperable representation of DS/AI resources, actors, structures, and digital artifacts so that they can be linked, described consistently, and made available according to the FAIR principles (Gesese et al., 2024, Silva et al., 4 Aug 2025).
1. Position within NFDI and declared objectives
NFDI4DSO was developed in the context of the German National Research Data Infrastructure (NFDI), described as a national initiative comprising 26 consortia. Within that landscape, NFDI4DataScience focuses on Data Science and AI, with the goal of enhancing the accessibility and interoperability of research data, connecting the digital artifacts involved in DS/AI research, and ensuring that these artifacts are available according to the FAIR principles (Gesese et al., 2024).
The ontology is therefore positioned neither as an isolated local schema nor as a foundational ontology. The 2025 overview characterizes it most precisely as a “domain-specific, modular extension of NFDICore Ontology, providing the conceptual basis for the NFDI4DS Knowledge Graph” (Silva et al., 4 Aug 2025). In that sense, NFDI4DSO occupies two roles simultaneously: it is a domain ontology for DS/AI research artifacts and an application-oriented ontology for the actual research knowledge graphs operated within the consortium.
Its FAIR support is described in functional terms. Findability is supported through annotation with shared classes and properties; Accessibility through exposure of metadata in the knowledge graph and portal; Interoperability through extension of NFDICore and alignment with BFO and other ontologies; and Reusability through semantically explicit descriptions of artifacts and organizational context (Gesese et al., 2024). The 2025 RKG paper adds that FAIR implementation also relies on persistent identifiers such as DOIs, ORCIDs, and ROR, standards-based services such as SPARQL, and provenance and licensing metadata (Silva et al., 4 Aug 2025).
2. Domain coverage and representational scope
NFDI4DSO is intended to describe both domain-relevant DS/AI resources and artifacts and the organizational structure of the NFDI4DS consortium (Gesese et al., 2024). The resources explicitly named in the ontology description include datasets, models, ontologies, code repositories, execution platforms, and repositories. The broader RKG framing extends the intended semantic space to publications, software, machine learning models, persons, organizations, research resources, licenses, provenance information, and repository-specific metadata such as downloads, stars, forks, and model cards (Gesese et al., 2024, Silva et al., 4 Aug 2025).
The ontology also covers consortium-internal structures and role-bearing participants. The paper explicitly mentions persons / personas, consortium members, spokespersons, task area leads, organizations and affiliations, and consortium-level structures and roles (Gesese et al., 2024). This means that NFDI4DSO is not limited to content metadata for datasets and software; it also models governance and organizational relations within the consortium.
A notable feature is the simultaneous treatment of artifact metadata and consortium metadata. The NFDIcore 2.0 paper describes NFDI4DSO as representing “metadata on resources in the data science domain and content-related index data, e.g., metadata for training datasets and machine learning models” (Bruns et al., 2024). This situates the ontology at the intersection of research-information management and domain-content indexing.
The ontology is also structurally richer than a flat metadata schema. The original paper states that it contains classes of several ontological kinds, including independent continuants, roles, and processes (Gesese et al., 2024). That distinction matters because it enables modeling of who participates in what, in which role, and through which process, rather than only direct binary catalog-style assertions.
3. Foundational commitments, modularity, and external alignments
The immediate semantic foundation of NFDI4DSO is NFDICore, described as a mid-level ontology for metadata on NFDI resources such as individuals, organizations, projects, and data portals (Gesese et al., 2024). NFDICore is already mapped to BFO and Schema.org, and follows a modular architecture that allows domain-specific consortia to build extensions. The NFDI4DS ontology follows the same pattern as other named extensions such as CTO for NFDI4Culture and MWO for NFDI-MatWerk (Gesese et al., 2024).
The quantitative extension pattern is explicitly documented:
| Ontology component | NFDICore | NFDI4DSO adds |
|---|---|---|
| Classes | 51 | 42 |
| Object properties | 55 | 38 |
| Data properties | 8 | 9 |
NFDICore also contains 18 annotation properties and 5 SWRL rules (Gesese et al., 2024).
A central commitment of NFDI4DSO is alignment to BFO. The ontology is described as being mapped to BFO in the same way as NFDICore, with BFO providing high-level categories such as independent continuant, role, and process (Gesese et al., 2024). The NFDIcore 2.0 paper clarifies the broader rationale: BFO is used to provide universal foundational distinctions, reduce ambiguity, and improve the treatment of roles, processes, and artifact semantics across heterogeneous research domains (Bruns et al., 2024).
Additional mappings are stated to schema.org, FaBiO (the FRBR-aligned Bibliographic Ontology), and the Conference Ontology (Gesese et al., 2024). The 2025 RKG paper places this within a broader interoperability program involving DDI Alliance, DataCite, OpenAIRE, and the European Open Science Cloud (EOSC), though it is explicit that the ontology itself is clearly aligned upward with NFDICore, BFO, and schema.org (Silva et al., 4 Aug 2025).
This layered design implies a vertical semantic stack: an upper-level alignment layer centered on BFO, a shared NFDI mid-level layer in NFDICore, and a DS/AI-specific extension layer in NFDI4DSO (Gesese et al., 2024). A plausible implication is that NFDI4DSO is intended to support both strict ontological integration and practical cross-consortium federation.
4. Ontology design patterns and reasoning model
The ontology is described as being developed in a modular fashion, using a bottom-up, iterative, and user-centered approach (Gesese et al., 2024). The 2025 overview complements this by locating ontology work within the NFDI4DS metadata working group, which was established to explore existing metadata standards and schemas, identify schema-related gaps, address these gaps for the community, and work with foundational ontologies to ensure interoperability within NFDI (Silva et al., 4 Aug 2025).
A distinctive modeling pattern in NFDI4DSO is the combination of BFO-compliant expressive modeling with shortcut properties that simplify data integration and querying. The paper explains this through the example of spokespersonhood. In the full pattern, a person participates in a leadership process, the consortium participates in that same process, the person has a spokesperson role, and the role is realized in the leadership process. From that richer path, a direct property nfdi4dso:spokesperson is inferred (Gesese et al., 2024).
The stated SWRL rule is:
This is described as emblematic of the ontology’s design philosophy: ontological rigor is preserved in the underlying representation, but user-facing relations are simplified through inference (Gesese et al., 2024). The NFDIcore 2.0 paper presents the same pattern for inferred properties such as publisher and contactPoint, reinforcing that this is a general NFDI design strategy rather than an isolated modeling choice (Bruns et al., 2024).
The same section also clarifies a frequent misconception. NFDI4DSO is not presented merely as a lightweight metadata vocabulary. Its use of roles, processes, role realization, and SWRL-based shortcut inference indicates a more expressive knowledge representation model (Gesese et al., 2024).
5. Role in the NFDI4DS knowledge-graph architecture
NFDI4DSO is explicitly designed to form the foundation of the NFDI4DS Knowledge Graph, which is under development (Gesese et al., 2024). The KG is described as having two main components.
The Research Information Graph (RIG) contains metadata about NFDI4DS consortium resources, persons, and organizations, and serves as the backend for the NFDI4DS web portal. The Research Data Graph (RDG) contains content-related index data integrated from heterogeneous data sources in the consortium (Gesese et al., 2024). The 2025 overview reiterates this division and places the ontology in the semantic layer that mediates between extraction and enrichment pipelines on the one hand and graph population on the other (Silva et al., 4 Aug 2025).
The broader RKG architecture in the 2025 paper is summarized as four components: community benchmarking, tools and pipelines for IE and FAIR assessment, vocabularies, schemas, and ontologies, and the RKGs themselves (Silva et al., 4 Aug 2025). In that description, the ontology is the semantic normalization layer through which heterogeneous metadata from repositories such as Hugging Face, GitHub, Zenodo, and arXiv can be represented consistently.
The ontology is already tied to concrete graph access patterns. The original paper gives a query scenario for retrieving the co-spokespersons of the NFDI4DS consortium, exposed through both SHMARQL for navigational access and SPARQL for direct querying (Gesese et al., 2024). The example query links the consortium entity to a shortcut property, then retrieves person metadata including ORCID, affiliation, and first and last names. This use case illustrates a central practical function of the ontology: it makes organizational, person, and artifact metadata retrievable through graph-native interfaces.
Both RIG and RDG are intended to be accessible, searchable, and available via the NFDI4DS Registry platform, and the paper notes planned collaboration with other NFDI consortia to integrate domain-specific knowledge into the RDG (Gesese et al., 2024). This suggests an architecture in which NFDI4DSO is not only a local modeling device but also a vehicle for cross-consortial semantic integration.
6. Tooling, publication, evaluation, and open issues
NFDI4DSO is implemented using an OWL-based formalization of terminological knowledge and the Protégé ontology editor. Documentation is generated with Widoco, which the paper says was used to create enriched and customized documentation automatically. In addition to OWL axioms, the ontology uses SWRL rules (Gesese et al., 2024).
The paper gives concrete publication and access points. The stable version v1.0.0 is located at https://github.com/ISE-FIZKarlsruhe/NFDI4DS-Ontology/tree/main, while the latest development version is at https://github.com/ISE-FIZKarlsruhe/NFDI4DS-Ontology/tree/develop-1.0.1. The first public version of the knowledge graph is accessible via the SPARQL endpoint https://nfdi.fiz-karlsruhe.de/4ds/sparql and the SHMARQL interface https://nfdi.fiz-karlsruhe.de/4ds/shmarql (Gesese et al., 2024, Silva et al., 4 Aug 2025).
The evaluation status is explicitly limited. The poster paper does not report a full formal evaluation or benchmarking study and gives no detailed discussion of logical consistency checks, competency-question coverage results, user studies, precision/recall metrics, or ontology quality metrics (Gesese et al., 2024). Instead, future work is said to include extensive ontology evaluation, using competency questions based on persona definitions from the NFDI4DS consortium (Gesese et al., 2024). The NFDIcore 2.0 paper shows that competency-question-driven validation is already a broader NFDI methodology, and notes that NFDI4DS contributed personas during requirements elicitation (Bruns et al., 2024).
A second limitation concerns the available documentation. The 2025 overview states that it does not provide a namespace IRI for the ontology, a full list of classes or properties, OWL/RDFS declarations, Turtle examples, SHACL shapes, class hierarchies, domain/range axioms, cardinality restrictions, logic formulas, or ontology versioning details (Silva et al., 4 Aug 2025). A recurrent misconception is therefore to treat the 2025 paper as a full formal specification; it documents the ontology’s architectural role and intended function more strongly than its complete formal schema.
An open modeling issue appears in the NFDIcore 2.0 analysis. Because BFO allows roles only on independent continuants, but NFDI4DataScience sometimes treats processes as resources, some competency questions become difficult to represent consistently—for example, questions about standards “for” a process such as sharing data (Bruns et al., 2024). This is a design tension rather than a resolved defect. It indicates that, in the NFDI4DS setting, careful separation may be required between a process, a resource describing a process, and a standard about that resource or process.
Taken together, these features define NFDI4DSO as a BFO-aligned, NFDICore-based ontology for Data Science and AI whose present significance lies in three coupled functions: representing DS/AI artifacts and consortium structures, providing the semantic backbone for the NFDI4DS Knowledge Graph, and mediating between expressive ontology engineering and practical graph querying through inferred shortcut relations (Gesese et al., 2024).
