This dissertation argues that machine extraction from interpretive sources is an act of source production itself and, therefore, must be subjected to the same level of source criticism. It develops this principle into a formal framework for constructing knowledge graphs, in which every claim that a generative model extracts is treated as a source to be criticized, rather than as an accepted fact.
The central scholarly contribution lies in this insight that the construction of knowledge graphs from interpretive historical sources requires the explicit representation of interpretations and their provenance as independent units of knowledge in order to ensure traceability, critical evaluation, and scholarly transparency.
When large language models extract knowledge from interpretive historical texts, the output often seems like the kind of conclusions that human readers would otherwise draw. For instance, they must determine whether "the minister of culture to the Medici court in Florence" refers to Giorgio Vasari or whether a "small devotional panel" denotes a particular artwork. Conventional knowledge graph construction pipelines presuppose that such mention-to-referent resolutions are resolved prior to representation. ... mehrThis dissertation challenges this pre-interpretive entity assumption.
We emphasize this as a defect, especially for interpretive sources such as Giorgio Vasari's The Lives of the Most Eminent Painters, Sculptors, and Architects (1568). Such cases blur the line between what is detected or extracted, what is reconstructed, and what is invented.
This dissertation develops and evaluates a framework that incorporates the interpretive process within the extraction pipeline as an essential element of knowledge representation. Existing provenance-aware methods typically record only the source text for each claim, omitting the specific conditions under which the claim is extracted. This omission, which we term the provenance gap, is fundamentally distinct from the opacity targeted by explainable artificial intelligence techniques, which answer how a model processes input rather than under which conditions – such as prompt templates, model versions, or reasoning steps – a specific claim is produced. What is lost in this gap is not detail but accountability. A claim whose conditions of production go unrecorded cannot be reconstructed or contested, and responsibility for them is lost within the graph.
As agentic frameworks become standard, the need for comprehensive provenance documentation intensifies. Reasoning traces generated by large language models constitute additional artifacts that require the same level of provenance documentation as the extractions they accompany. These traces supplement, but do not replace, the provenance record.
Viewsari addresses the provenance gap by explicitly tracking extraction conditions and maintaining a clear distinction between textual mentions and real-world referents, a separation often overlooked by standards such as the CIDOC Conceptual Reference Model (CRM).
We use Vasari's The Lives as our main case study. This important art-historical text blends historical biography with narrative storytelling, and artworks are rarely named by their official, canonical names (if they have one) and are instead described by who commissioned them, where they are, what they show, or their purpose. Because the text uses implicit references, includes works not found in standard databases, and blends description with interpretation, where no clear facts are to be recovered, it is a challenging test bed for conventional extraction pipelines, bringing their performance to a limit.
The dissertation constructs four engineering artifacts in sequence. At its foundation is the Viewsari ontology, a provenance-aware OWL 2 ontology that integrates PROV-O and the W3C Web Annotation Ontology within a three-layer structure spanning the bibliographic, structural, and content layers. This architecture is domain-neutral except for the content layer, which is tailored to the artworks in the case study. The ontology enforces three principal constraints on the pipeline, namely a target schema, a four-part mention typology (explicit, implicit, coreferent, generic) that separates resolution strategies, and provenance slots that model extraction reasoning as PROV-O activities. Central to this contribution is that historical descriptions should not be linked directly to world objects, but rather through explicitly modeled acts of interpretation. Viewsari implements this principle by separating the mention, the interpretation, and the referent.
ObliquER (Oblique Entity Recognition), built on this schema, is a generative extraction pipeline that employs a modular prompt system to separate domain-specific content from structural templates. This modularity enables systematic testing and facilitates adaptation to new domains.
We design ObliquER to recognize and link both explicit and implicit mentions and to include out-of-knowledge-base entities rather than discarding them. At the entity level, of the 15,804 artwork entities in the populated knowledge graph, 12,529 (79.3 %) are out-of-knowledge-base, meaning they cannot be linked to any Wikidata entity, while only 3,275 (20.7 %) are Wikidata-linked. At the mention level, the two types diverge sharply: explicit, named mentions resolve to Wikidata far more often than implicit ones (49.0 % versus 30.2 % linked in the populated graph), so implicit mentions account for the bulk of the out-of-knowledge-base population.
We find that implicit references and absence in contemporary knowledge bases track each other. Entities that Vasari expects readers to recognize without explicit naming are often those lacking canonical identifiers. Restricting a pipeline to explicit mentions would therefore exclude not only a portion of the artwork population but specifically those entities most integral to the canonical narratives that The Lives itself established.
ObliquER recovers implicit artwork mentions that surface-form-level zero-shot baselines such as Universal NER and GLiNER2 cannot represent. On the implicit subset, these baselines score a strict F1 of 0.00 and 0.02, respectively, by architectural construction, against ObliquER's 0.321, so the result reflects recognition capability for interpretive sources rather than an improvement over the prior best. On out-of-knowledge-base linking, ObliquER reports an F1 of 0.721, demonstrating that out-of-knowledge-base mentions are reliably detectable as such, an operational requirement for treating them as first-class entities in the resulting knowledge graph rather than as extraction failures.
Running ObliquER over this schema yields the Viewsari knowledge graph (2,455,976 triples, including 15,804 artwork entities, with complete provenance on every machine-extracted annotation), which, together with its web interface, serves two roles. As a digital humanities tool, the graph makes 12,529 previously uncataloged artworks from The Lives - entities absent from Wikidata and therefore invisible to conventional extraction pipelines - searchable and citable, and the interface lets non-technical art historians inspect and challenge the provenance of any claim. As a methodological example, a comparison with CIDOC CRM shows that Viewsari simultaneously adds a representation principle for implicit entities and extraction provenance, as the graph demonstrates.
To assess the application to other data, a transferability pilot adapts the Viewsari ontology and ObliquER to a 19th-century German corpus on the Illuminati (Kloß' Annalen), with the ground truth linking to FactGrid instead of Wikidata. The pilot delineates which components are general – the three-layer architecture, mention types, pipeline stages, and prompt structure – and which are domain-specific. Comparative analysis across digital humanities projects reveals a shared emphasis on provenance, but only the Viewsari ontology implements structures of computational provenance and mention/referent separation, abstracted and proposed as a reusable contribution to digital humanities ontology design (the Computational Provenance pattern), capturing the machine-generated dimension of generative extraction that other projects document only informally or omit.
A catalog of seventy competency questions, developed in collaboration with art historians and knowledge engineers, connects requirements to evaluation. All questions are answerable on the populated graph, and any remaining gaps precisely identify missing data rather than reflect unasked queries. This process clarifies the project's deliverables for each audience, namely a queryable, source-anchored response to art-historical questions and a provenance-aware, first-class representation of the extraction process for knowledge engineering.
The central methodological claim of this dissertation is that constructing a knowledge graph from interpretive historical sources requires modeling the interpretive work performed at each decision point, not merely extracting or detecting facts. The process must document the conditions under which each extraction is produced, recording them with the same rigor as the claims themselves. This documentation is essential because without it, the connection between a claim and its interpretive act is lost, and with it the epistemic responsibility for the claim. A graph that cannot say under which conditions it asserted something cannot be held accountable for the assertion. Such documentation is not a form of explainability in the technical sense; it does not expose the model's internal computations, but constitutes hermeneutic accountability, recording the textual basis, the interpretive convention, and the computational agent by which each claim was made. In this, it extends the historian's source criticism to machine extraction. Just as historiography never takes a source at face value but asks under what conditions, by whom, and to what end it was produced, Viewsari subjects each machine-extracted claim to the same scrutiny, treating each extraction as a source to be criticized rather than a fact to be accepted. Viewsari demonstrates that such modeling is feasible at scale, making interpretation integral to knowledge representation instead of something lost when moving from text to graph.
1