Viewsari · Lives of the artists

Viewsari

A provenance-aware knowledge graph built from Giorgio Vasari's Le Vite de' piu eccellenti pittori, scultori, e architettori (1568), developed at KIT / FIZ Karlsruhe (ISE group) as part of a doctoral dissertation.

Browse annotations Ontology introduction

Ground truth

Biographies

270

Paragraphs

Volumes

ObliquER

852

Mentions

443

Entities

Knowledge extracted from interpretive texts is not discovered but constructed; therefore, knowledge graphs must model the construction process itself.

Cultural heritage materials, once digitized, remain largely unstructured: human-readable, but not systematically queryable, linkable, or analyzable. Large language models have improved access to such content, yet they cannot guarantee epistemic traceability. Viewsari addresses this gap by treating knowledge extraction from historical text as an interpretive activity, recording agents, prompts, software versions, and source contexts for each extracted statement.

The approach is neuro-symbolic: the Viewsari ontology (symbolic component) provides a formal schema that structures and constrains extraction tasks assigned to a large language model (neural component). Vasari's Lives (1568) serves as testbed — a founding document of art history characterized by implicit entity references, attributions presented as facts, and long-tail entities absent from standard knowledge bases.

Research Questions

RQ1 What representational commitments must a knowledge graph make to remain accountable to the interpretive act that produced it, and how can these commitments be operationalized in a provenance-aware ontology?

The Viewsari ontology formalizes a three-layer model covering bibliographic structure, document components, and extracted entities with full PROV-O provenance. Every mention, entity, and extraction run is traceable back to its source paragraph and the agent that produced it.

Ontology overview WIDOCO documentation Knowledge base SPARQL endpoint

RQ2 Under what conditions can large language models, treated as interpretive agents, recognize and link implicit and out-of-knowledge-base entity mentions that lie beyond span-level supervised baselines?

The ObliquER pipeline uses LLM-based named entity recognition with few-shot and ontology-guided prompting strategies. Extraction runs are recorded as per-paragraph PROV activities with full prompt text, timestamps, and model provenance. Entity linking from the ground truth maps mentions to Wikidata or OOKB identifiers.

ObliquER pipeline Knowledge graph explorer Interactive demo Annotation corpus

RQ3 Which design patterns recur across provenance-aware knowledge graphs in the digital humanities, and to what extent does the Viewsari methodology transfer to structurally distinct interpretive corpora?

Vasari's text features implicit references, long-tail entities absent from standard knowledge bases, and attributions presented as facts. The Viewsari approach addresses these through OOKB entity handling, coreference resolution across biographies, and a provenance-first KG design that separates extraction from interpretation — principles applicable to other cultural heritage corpora.

Knowledge base (OOKB entities) Ground truth KG explorer Publications

Knowledge Graph Preview