“The research is happening in the text fields” – Are Linked Open Data and Art History a good match? Giacomo Nanni Linda Freyberg Sabine de Günther Marian Dörk

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

 “The research is happening in the text fields” – Are Linked Open Data and Art History a good match? 2023 Digital Humanities 2023 (DH2023), Graz 10.5281/zenodo.8107731 lod2023-art-history lod2023-art-history.txt, the project's staged copy 73b837e3f6705fdd5fcda7372e2d3041aa5f546b cc-by art-history

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

 44 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them Where the boundary falls between what an art-historical research project can express as Linked Open Data and what it must leave as prose. Structured vocabularies buy accuracy and queryability - so what is lost when the reflexive, contextual part of humanities research is forced into them? Argues that only the quantitative, queryable part of art-historical research is worth modelling as triples, and bridges the rest to it through visual interfaces rather than forcing it into ontologies. critique linked open data modelling in art history practice-reportpractice-report mechanismsan ontology can never close: no cultural artefact is finally contextualisable, and the vocabularies are uneven across subtopics mechanismsthe target moves: meanings shift over the life of the project doing the modelling mechanismsthe cost of manual tagging is what forces institutions to publish shallow metadata, which then constrains what anyone can ask of the data mechanismsdeep modelling is not durable either - the British Museum modelled its whole collection in CIDOC CRM and then could not keep it accessible One project, one discipline: the argument is illustrated by Restaging Fashion rather than tested across cases.what this limits: the generality of the argument: it is illustrated by the Restaging Fashion project rather than tested across cases The boundary it proposes - quantitative to triples, reflexive to interfaces - is asserted rather than operationalised; nothing says where a given sentence falls.what this limits: the proposed boundary itself: QUANTITATIVE INFORMATION ... WHICH CAN BE QUERIED AND EXPRESSED STATISTICALLY, ARE WORTH TO BE MODELED IN TRIPLE STRUCTURES - asserted, not operationalised Linked Data, reference ontologies and authority files, applied to art-historical research and found to fit only part of it linked open datalinked open data CIDOC CRMCIDOC CRM free-form textfree-form text information visualizationinformation visualization ontologiesontologies metadatametadata triple structurestriple structures representation - as machine-readable structure, whose point is that a formal query can be run over it representation - as something another researcher can read and understand, which the paper argues needs no formalisation at all the unrepresentable - material the paper holds cannot be captured by either, quoting Galloway claim_type: critiquea second reading reaching the same value. Recorded: critique — note: a second reading reaching the same value from a different sentence. Not applied - a person decides. method: conceptual-analysisa second reading DIFFERING from the recorded value. Recorded: practice-report — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. phenomenon: the trade-off between semantic formalisation and the expressiveness of humanities research writinga second reading DIFFERING from the recorded value. Recorded: linked open data modelling in art history — note: a second reading, verbatim-verified, disagreeing with the value on the record. Not applied - a person decides. builds-on → Galloway (2011: 91) on the unrepresentable builds-on → Daquino et al. (2022) on institutional uptake of LOD-native systems builds-on → Berners-Lee (2006) on Linked Data contradicts → the programme of modelling cultural heritage research exhaustively in formal ontologiesdiffers in: THE AXIS IS WHICH PART OF THE WORK IS BEING MODELLED. 'IN TERMS OF REPRESENTING THE RESEARCH PROCESS IT IS NOT WORTH MAKING THIS PART MACHINE-READABLE as a basis for formal queries.' The exhaustive-formalisation programme is contradicted for ONE COMPONENT and left standing for the rest - which is exactly the qualification a bare 'contradicts' deletes. Linked Open Data —contrasted-with→ free-form textstrength of the relation: 1 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span poses the choice between LOD and free-form text as a question; contrasted-with is loose.Linked Open Datafree-form text CIDOC CRM —is-a→ ontologiesstrength of the relation: 1CIDOC CRMontologies The central question of the paper is a boundary question: what becomes structured data and what stays prose. Standards buy accuracy and unambiguity, but research also needs room for observations that are not yet settled. Modelling a collection early forces a trade-off between standardisation and semantic accuracy. Even major museums publishing LOD share only shallow models of their objects. The British Museum structured its whole collection in CIDOC CRM but could not keep the database consistent and accessible. Manual tagging is precise but does not scale, and automatic tagging fails on inconsistent data - so neither route reaches existing collections. More elaborated, reflexive or contextualized information may need other forms of representation, such as visual interfaces. Free text is treated as the starting point for semantic contextualisation rather than as a residue to be tidied away. reference modelsreference models CIDOC CRMCIDOC CRM LODLOD unrepresentableunrepresentable

Introduction

The scientific discourse in object-focused and historical disciplines follows certain methods and builds a narration with objects, events, evidence and conclusions.

In Digital Humanities (DH) research this process and its specific structure has to be represented on a data level. Linked Open Data, Semantic Web ontologies and, in particular, reference models like CIDOC CRM (2022) offer a precise structural vocabulary to describe cultural heritage objects and their relations. These frameworks can support GLAM institutions to make the data on their objects findable, accessible and interoperable.

However, in research projects the course of developing arguments and describing findings is crucial. We use Art History research, in particular, the project “Restaging Fashion - Digital contextualization of vestimentary sources” (Refa), as one example to show how emerging discourses within the research process can be mapped in a knowledge organization system and which aspects may not be formally represented. A recurring question is: What should be modeled as linked open data (LOD) and what should be entered as free-form text? Despite the proven utility of data standards, vocabularies, and authority files providing for accuracy, context, and unambiguity, there is also a need for an open space to formulate and present tentative observations. With this contribution we present how we bridge LOD and full text descriptions through information visualization.

Semantic data modeling in Museums

Semantic data modeling of a collection during an early stage of the research prompts a trade-off to be made between data standardization and semantic accuracy.Sophisticated ontologies like CIDOC CRM encourage the detailed description of an object through an event-driven structure, which makes it a very powerful tool for the cultural heritage sector by providing for the possibility of adding meaning to the original data.

Since 2006, when the term Linked Data was coined (see Berners-Lee 2006), according to Daquinoet al."only a few pioneers have abandoned legacy cataloguing and archiving systems to fully embrace the Semantic Web paradigm and manage their catalogues through LOD-native management systems" (2022).

When considering the LOD datasets published by The British Museum and the Rijksmuseum, who also create tools such as ResearchSpace (Oldman / Tanase 2018) and dedicated platforms to engage with their collections, it appears that these museums only share relatively shallow data models for their artworks on the web. The Rijksmuseum provides basic metadata on objects structured in DC , EDM and LIDO schemas. Only one example is modeled deeply: Linked Art van Gogh object metadata download . But the creators clearly state that: “both the implementation of the model as well as the use of identifiers will be subject to change” (RijksData 2022). The British Museum, on the other hand, structured its entire collection using CIDOC CRM, but was unable to preserve its database in a consistent and accessible manner (Lincoln, 2015). Data that should be linked together is no longer accessible, making the research of scholars no longer available ( see Programminghistorian 2017).

The potential of these applications is obvious: LOD allows data to be reused and redistributed, revealing relationships with other data on the Web (Berners-Lee 2006). The resources involved, however, are somewhat excessive both from a technological perspective (see De Decker 2015)and regarding the cataloging of information. Manual tagging can be performed by domain experts in a very precise manner, but it cannot be performed manually to large volumes of already existing content due to the lack of resources (Simou et al. 2017) or may not be conducted automatically due to inconsistencies and/or the lack of standardization in the datasets. This leads institutions to publish only a subset of the available information, exposing minimal metadata (Doerr et al. 2010), which might prevent researchers from using the data and constrains them to conventional research methods.

Data-driven art historical discourse

Art historical research on and with collections is deeply concerned with the history of objects and their discourse such as provenance history, history of interpretation, dating or attribution based on comparable works. This can be expressed with the event structure of CIDOC CRM including descriptive information on the objects like dimensions, inscriptions, and alternative titles also linked to relevant vocabularies like ICONCLASS and AAT . But a crucial part of humanities research consists of annotating that basic formal information, by elaborating and contextualizing the history of the artworks and their interpretation, often concluding in the form of a paper or monograph. Undertaking this part of research in the form of LOD is very intricate and laborious. Complex ontologies will never be completely exhaustive since no cultural artifact or experience could be fully and finally contextualized and vocabularies are not equally accurate and detailed for every subtopic. In addition, contexts and meanings shift over time, even during the timeframe of a research project. According to Galloway "there are some things that are unrepresentable. And the computer is our guide into that realm" (2011: 91). In terms of representing the research process it is not worth making this part machine -readable as a basis for formal queries. It is more important to present it in a way that is accessible and understandable to other researchers, which does not require this high level of formalization, because it would end up in overcomplicated queries.

Visualizing context

One way to equally accommodate formal LOD structures as well as full text descriptions is through custom interfaces. In particular, information visualization can be used for the contextualization of images, sources, texts and their relations.

Fig. 1: Draft of Refa interface

As part of the Restaging Fashion project, we devised different experimental prototypes, linked specific parts of the textual description to artworks, to classification entries, to archival sources as well as to other cultural heritage objects (see Fig. 1.) with the aim to establish an overview of the context.

Instead of hiding the full text in unstructured note fields, we approach this as a starting point for the semantic contextualization of the research process.

In conclusion, certain aspects of humanities research can be properly expressed by semantic web structures. In particular quantitative information, for instance, on objects and their relations, which can be queried and expressed statistically, are worth to be modeled in triple structures. Other, more elaborated, reflexive, or contextualized information may require different forms of representations like visual interfaces.

CIDOC Conceptual Reference Model

Galleries, Libraries, Archives and Museums

Dublin Core

Europeana Data Model

Lightweight Information Describing Objects

https://iconclass.org/

https://www.getty.edu/research/tools/vocabularies/aat/