Corpus and Coverage

Although the individual member institutions differ in their collection profiles, documentation practices, and research priorities, the knowledge contained within their databases is fundamentally complementary: each archive engages, in its own way, with the same core entities—artworks, their photographic reproductions, and the scholarly contexts that surround them. To render this shared knowledge interoperable, data are ingested from a wide range of source formats, transformed into a shared data model, and—where necessary—reconciled against external authority vocabularies such as the Art & Architecture Thesaurus (AAT).

For a discussion of common usability challenges in entity reconciliation workflows, see Daquino / Kozlov / Mambelli (2026).

Table 1. Descriptive statistics of the data provided in artresearch.net by PHAROS member institutions.

System Architecture

artresearch.net is built on ResearchSpace, an open-source semantic research environment originally developed by the British Museum. It leverages Linked Open Data (LOD) technologies to move toward more dynamic and context-rich modes of information discovery (Oldman / Tanase / Santschi 2019). Within artresearch.net, data are published in accordance with the CIDOC Conceptual Reference Model (CRM) ontology. Collections are transformed into RDF and made queryable through SPARQL endpoints, with RDF stored in a Blazegraph triplestore; an IIIF-compliant image server ensures the delivery of high-resolution images.

For further technical details, see Klic (2022).

Multimodal Search

artresearch.net supports a wide range of scholarly practices—from locating specific artworks to uncovering previously unrecognized visual relationships across heterogeneous collections. To accommodate this diversity, the platform combines text-based and image-based search. Text Match identifies records in which query terms appear verbatim within metadata fields such as titles or artist names. Textual Similarity enables concept-driven and multilingual exploration by embedding textual metadata in a shared vector space. Extending this approach across modalities, Visual Similarity translates textual queries into visual embeddings, allowing users to retrieve artworks corresponding to described compositional, stylistic, or aesthetic features. Image-based search likewise distinguishes between exact and similarity-based modes. Visual Match detects near-identical reproductions by comparing digital fingerprints of images, supporting tasks such as cross-archive reconciliation. Visual Similarity retrieves artworks related to a given image in terms of style, palette, composition, or subject matter, enabling the identification of broader visual affinities rather than direct copies. Textual metadata are embedded using the Qwen3 model with 8 billion parameters (Yang et al. 2025). Cross-modal and image-based search rely on the NaFlex variant of the SigLIP 2 model with 400 million parameters (Tschannen et al. 2025). For near-exact image retrieval, similarity rankings are refined using RootSIFT feature matching (Arandjelovic / Zisserman 2012) with geometric verification via MAGSAC (Barath / Matas / Noskova 2019). To support interpretability, search results are labeled with indicators denoting the retrieval strategies employed (i.e., Text Match, Textual Similarity, Visual Match, or Visual Similarity), with records satisfying multiple criteria being prioritized in the ranking (Figure 1a).

See also https://artresearch.net/resource/search-help (last accessed 26 April 2026).