Language Intelligence
The Scale of the Written Record
The scholarly literature on fine art spans five centuries of catalogues raisonnés, auction records, exhibition catalogues, letters, inventories, and critical essays — written in dozens of languages, scattered across thousands of institutions. No single scholar, however dedicated, can read more than a fraction of this material in a lifetime. Connections that would be obvious if all the evidence were visible remain hidden because the relevant passages sit in different volumes, different archives, different languages.
Language models change this calculus. They can process text at a volume and speed that is simply impossible for a human reader, surfacing patterns, contradictions, and corroborations across the full breadth of the written record.
Source-Anchored Intelligence
The critical discipline in our approach is citation fidelity. Every claim extracted by our language systems is traced back to its source document, page, and passage. We do not permit our models to synthesise conclusions that cannot be independently verified by returning to the original text.
This means our systems function as research assistants, not oracles. They identify relevant passages, flag contradictions between sources, and surface previously unnoticed connections — but the scholarly judgment of what those connections mean remains with the human researcher.
Multilingual Processing
Art-historical scholarship is inherently multilingual. A single provenance chain might pass through Italian inventories, French auction records, German exhibition catalogues, and English dealer correspondence. Our language systems are designed to operate across these linguistic boundaries, extracting structured information from documents regardless of their source language and normalising it into a unified evidence graph.
We pay particular attention to the challenges of historical language — archaic spellings, abbreviated Latin, period-specific terminology, and the variant transliterations of proper names that make traditional keyword search inadequate for serious archival research.
Current Applications
Our active language intelligence programmes include automated provenance extraction from auction catalogue text; cross-referencing of exhibition histories across multilingual sources; sentiment and certainty analysis of attribution language in scholarly literature; and entity resolution — the painstaking work of determining when two different names in two different documents refer to the same person, place, or artwork.