PHAROS Consortium: Art-Historical Linked Data

• PHAROS Consortium is an international coalition uniting major art-historical photographic archives by reconciling heterogeneous metadata through linked open data approaches.
• The project employs CIDOC-CRM protocols, rule-based matching, and embedding methods to address challenges like ambiguity and varied cataloguing practices.
• Its socio-technical framework, with curated interfaces and provenance tracking, advances dynamic scholarly dialogue and supports sustainable research infrastructures.

The PHAROS Consortium is an international coalition of leading art-historical photographic archives dedicated to reconciling and exposing heterogeneous collections as interoperable, semantically rich Linked Open Data. Emerging from a recognition that entrenched heterogeneity in cataloguing practices impedes the promise of cross-collection research, PHAROS coordinates infrastructure, workflows, and epistemic frameworks among thirteen partner institutions holding over thirty-one million photographic documents representing approximately one million distinct artworks. Its central aims are to enable discovery and research through linked data, develop reusable reconciliation workflows sensitive to art-historical nuance, and architect a socio-technical system that sustains and scales as partners, data, and research questions evolve (Daquino et al., 21 Jan 2026).

1. Consortium Structure and Holdings

PHAROS unites thirteen major archives across Europe and North America, including institutions such as Bibliotheca Hertziana (Max Planck Institute for Art History), Bildarchiv Foto Marburg, the Frick Art Reference Library, Fondazione Federico Zeri, the Kunsthistorisches Institut in Florenz, the Netherlands Institute for Art History (RKD), and the Warburg Institute. Collectively, these repositories comprise over thirty-one million photographic documents, capturing around one million unique artworks.

A subset of seven partners operates artresearch.net, a Linked-Data research platform established atop ResearchSpace. This platform has ingested 1.7 million photographs, utilizing a CIDOC-CRM application profile expressly designed to harmonize diverse metadata and facilitate cross-collection entity reconciliation. PHAROS’s mission encompasses (1) maximizing research and discovery by enabling interoperable linked data for art-historical photography, (2) formulating sustainable reconciliation policies and workflows attuned to domain-specific epistemologies, and (3) constructing an ontology-driven infrastructure capable of accommodating ongoing institutional and scholarly development (Daquino et al., 21 Jan 2026).

2. Metadata Heterogeneity and Its Typology

The integration of independent archives—each with decades of unique cataloguing traditions and a spectrum of descriptive schemas such as ICCD’s Scheda OA/F, MIDAS, MARC21, and bespoke systems—brings forth six salient forms of metadata heterogeneity:

• Reticence: Systematic omission or absence of fields due to incomplete policies or operational exigency.
• Flattening: Encapsulation of divergent information (e.g., photographer and studio) within a single metadata element.
• Coercion: Overspilling secondary data (e.g., provenance notes) into primary fields (e.g., artwork titles).
• Dumping: Allocation of uncategorizable or surfeit information to unrestricted text fields.
• Varying reliability: Lack of explicit markers for confidence or scholarly consensus levels within assertions.
• Assumed certainty: Presentation of values without uncertainty qualifiers, leading to the concealment of interpretive doubt.

These patterns undermine one-to-one URI mapping and demand more sophisticated approaches to entity reconciliation and metadata alignment (Daquino et al., 21 Jan 2026).

3. Domain-Specific Reconciliation Challenges

Several recurring reconciliation challenges typify the PHAROS corpus:

• Ambiguous or Qualified Attributions: Artist identities often include qualifiers, such as “circle of…”, family-level identifiers (“Bellini”), or institutionally divergent pseudonyms (“Pseudo Ambrogio di Baldese”). Authority sources like ULAN and Wikidata may offer incompatible mappings.
• Material and Technique Ambiguity: Terms such as “albumen” might simultaneously refer to raw material, photographic process, or print type—dimensions distinctly enumerated in thesauri like AAT.
• Provenance Granularity Collapse: Provenance can amalgamate agents, buildings, and collections in free text, obscuring relationships (e.g., between “Frick Art Research Library,” the “Frick Collection,” or “Hotel George V, Auction Tajan” as agent/building).
• Absence of Universal Artwork Authorities: With Getty CONA minimally populated in LOD, artwork reconciliation relies heavily on custom metadata alignments and image-similarity algorithms.
• Unstable Photographer Attributions: Studio and personal identities may shift, overlap, or be rendered ambiguous by marriage, collective operation, or unresolvable fragments (e.g., “Beyer, ?”).

Such complexities underscore the necessity of reconciliation mechanisms that directly model ambiguity and uncertainty (Daquino et al., 21 Jan 2026).

4. Reconciliation Strategies and Modelling Approaches

PHAROS frames ambiguity as intrinsic, eschewing forced one-to-one mappings in favor of a palette of modelling strategies:

• Many-to-many mappings: Local entities $e_\ell$ may be mapped to sets of external URIs $\{u_1, u_2, \ldots\}$ to accommodate homonymy, pseudonymy, or contested attributions.
• Uncertainty qualifiers: When links are plausible but unverified, rdfs:seeAlso is used in place of owl:sameAs, maintaining the potential for future revision.
• Anonymous or collective entities: New URIs denote collective or familial entities (“Bellini” as a workshop), avoiding presupposed equivalence to individual identities.
• Umbrella terms: Aggregative labels (e.g., “Böhm”) are instantiated as reified nodes (“Böhm (umbrella term)”) supporting hierarchical browsing and facet navigation.
• Associative relations: Non-equivalence is encoded for visually or functionally linked, but non-identical, works (e.g., related preparatory drawings and final paintings).
• Named-graph provenance: All statements are institutionally sourced, supporting traceability and future correction.

The explicit treatment of uncertainty is operationalized via a four-part framework:

Strategy Type	Description	Role in Workflow
Identification	Automated routines surface homonyms, pseudonyms, and incomplete data	Candidate generation for reconciliation pipeline
Modelling	Selection of predicates, reified nodes, alternate statements	Representing ambiguity in the data model
Workflow	Rule-based and embedding-based matching augmented by LLM pruning, vetting	Multiphase, human-in-the-loop pipeline
Interface	Hierarchical navigation, provenance badges, and attestation management	User interaction, surfacing and contextualizing data

No formal calculus is supplied, but the core function may be modeled as a reconciliation mapping $\mathcal{R}: L \rightarrow \mathcal{P}(A)$, where for entity $e$:

1. Rule-based matching: $$\mathcal{R}_1(e) = \left\{ a \in A ~\middle|~ \mathrm{sim}_s(e.\mathrm{label}, a.\mathrm{label}) \geq \tau_s ~\land~ \mathrm{overlap}(e.\mathrm{dates}, a.\mathrm{dates}) \right\}$$
2. Embedding-based expansion: $\mathcal{R}_2(e)$ is the set of top $k$ embedding-space neighbors for $e$, LLM-pruned
3. Final set: $$\mathcal{R}(e) = \mathcal{R}_1(e) \cup \mathrm{prune}(\mathcal{R}_2(e))$$, all subject to manual review

Cross-authority URI consistency is maintained through link expansion, deprecated/replacement URI updating, cycle/conflict filtration (prioritizing curated sources), and explicit provenance retention (Daquino et al., 21 Jan 2026).

5. Socio-Technical Infrastructure and Interface Design

artresearch.net, powered by ResearchSpace and a CIDOC-CRM application profile, provides the principal technical substrate. The infrastructure encompasses ontology-driven reconciliation, authority matching services, curated interfaces, and visualization tools supporting hierarchical facets and provenance annotation. The system is constructed to accommodate ongoing ingestion, authority scheme evolution, and expert curation. Each archive is responsible for validating reconciliation candidates concerning its holdings; vetted mappings are merged and made accessible within a centralized, provenance-annotated knowledge graph.

Interface strategies incorporate hierarchical facets that present summary or aggregate entities by default, permitting detailed exploration while maintaining user-oriented clarity. Provenance badges and alternate attestation displays foreground the interpretive plurality and evolution of scholarly consensus across sources (Daquino et al., 21 Jan 2026).

6. Key Lessons and Future Implications

The PHAROS experience yields several broadly relevant principles for large-scale cultural heritage infrastructures:

• Reconciliation is an iterative, socio-technical process, not a fixed technical task; governance structures, versioning, and robust provenance must be integral to any sustainable system.
• Full automation is illusory; LLMs and embeddings can propose, but only expert human adjudication can resolve the layered ambiguities endemic to historical cultural data.
• Compromises between philological precision and user experience are necessary; umbrella terms, hierarchical facets, and explicit modelling of disagreement support a spectrum of research needs.
• Semantic Web principles—Open World Assumption and “Anyone can say Anything about Any thing”—require modelling rather than erasure of disagreement; this is achieved via many-to-many relations, lightweight reification, and named graphs.
• Reconciliation is not aimed at enforcing a single truth, but at enabling dialogic, provenance-aware pluralism; interpretive diversity is preserved and surfaced as a resource rather than suppressed as a flaw (Daquino et al., 21 Jan 2026).

A plausible implication is that PHAROS’s reconciliation framework can be extrapolated to other domains where metadata heterogeneity and interpretive ambiguity are structural features, not idiosyncratic exceptions. Preservation of disagreement and explicit modelling of uncertainty are positioned as essential capabilities for any interoperable, sustainable, and semantically robust digital heritage infrastructure.