Artificial Intelligence and Art Market Discourse: Language Modelling and Knowledge Retrieval from the Getty Provenance Index – German Sales, 1900–1945

This thesis examines the Getty Provenance Index (GPI) German Sales as a large language resource of approximately 830,000 German-language auction entries from Germany, Austria, and Switzerland (1900–1945). The project treats the corpus as a discursive language system and evaluates the application of artificial intelligence for language analysis at scale. The thesis contributes a practical methodology for preparing, analysing, and searching large art-market archives across multiple model configurations. It also delivers a new dataset of named expert attestations, enabling research on expert knowledge and the interaction between academic and commercial market spheres.

The thesis is structured using 4 case studies and develops a methodology for AI-based language analysis of this art-market archive. The second chapter investigates text embeddings and clustering methods to group auction entries into stylistic subsets. The third explores BERT and attention-weight analysis as an exploratory tool for examining descriptive and technical language in painting descriptions. The fourth applies semantic search techniques to support knowledge retrieval from the challenging dataset. Retrieval-augmented generation is evaluated across multiple model configurations as a prototype for AI-supported search in large art-market archives. The final case study demonstrates a text-retrieval pipeline combining SQL pre-filtering with a large language model to extract named experts from auction entries.