Exhibition Catalogues as Data: Sources of Structural Complexity and Steps Towards AI-Assisted Extraction

The paper examines how structural data complexity in historical exhibition catalogues shapes the possibilities and limits of artificial intelligence assisted extraction and structuring. Focusing on the Exhibitions of Living Masters corpus from the RKD – Netherlands Institute for Art History, it argues that what is often framed as “messy data” reflects layered document-level, cross-catalogue, and contextual complexities embedded in historical and institutional practices. The paper first develops a taxonomy that distinguishes among document-level issues (legibility, implicit content, annotations, multilinguality), cross-catalogue variation (fonts, layouts, metadata conventions), and contextual factors (historical shifts, legacy structuring bias), clarifying how each category produces different kinds of extraction failure and intervention. It then presents an initial qualitative comparison of two workflows: a sequential pipeline using open character recognition (OCR) and a large language model (LLM) and a vision-language model (VLM) operating directly on page images. The VLM approach more consistently preserves entry boundaries, interprets visual cues and repetition markers, and manages multilingual content, while both workflows remain fragile in fields requiring contextual knowledge, such as addresses, pricing, and historically shifting entities. The paper concludes by outlining implications for digital art history and GLAM institutions, showing how such workflows can extend digitization beyond page images to interoperable, reusable datasets, while stressing the continuing need for domain-informed modelling, expert validation, and attention to the labor that underpins AI-assisted data creation.