Valorizing Past Art Historical Research with LLMs for European Cultural Heritage: A Case Study of the Corpus Rubenianum

The vast body of past art historical research, including encyclopaedias, monographs and art catalogues, represents a wealth of meticulous bibliographic and scholarly research. However, a significant portion of this valuable material exists only as scanned images or unstructured OCRed PDF documents, posing a challenge to contemporary researchers seeking to fully exploit these findings. This limitation hinders the seamless integration of established knowledge with newer discoveries and state-of-the-art digital analysis methods. This paper explores an approach to bridge this gap by demonstrating the potential of Large Language Models (LLMs) to extract structured information from these historical resources. To illustrate this potential, our research focuses on the Corpus Rubenianum, an impressive catalogue of over 40 books meticulously documenting the work of Peter Paul Rubens. This monumental work, based on the lifelong research of Ludwig Burchard, is universally recognised as the definitive resource on Rubens. Each volume is written by a leading scholar and aims to embody all that is currently known about the artist's oeuvre, with over 2,500 compositions and 10,000 works of art listed, based on Burchard's extensive documentation. Through the use of LLMs, we aim to extract key structured data from a selection of the volumes of the corpus available in PDF only, including but not limited to structured bibliographic references a structured list of provenances for each artwork a structured list of Greek mythological figures depicted in each painting (iconography). The second part of this paper will demonstrate how this extracted structured data could be effectively valorised through various digital methods, including but not limited to: enriching current bibliographic information with structured extracted bibliographic references from PDF sources; creating an index of characters depicted on the paintings with pointers to the respective artworks; visualising structured provenance information for specific artworks in an interactive timeline or network graph of individuals or institutions that owned an artwork during a particular period; illustrating the potential of feeding the structured iconographic descriptions into a vision transformer model to develop an augmented reality layer that displays iconographic information alongside the artwork, allowing a wider audience to 'read' the artwork. A prototype demonstrating these LLM extraction capabilities was recently presented and awarded second place at the ai4culture hackathon (in partnership with europeana.eu - February 2025). By demonstrating the extraction of structured information from a significant historical resource such as the Corpus Rubenianum, and its subsequent digital manipulation and visualisation, this paper will highlight the transformative potential of these technologies to unlock data from valuable past research that is only available in unstructured PDF-like formats. This approach can facilitate more comprehensive research and enhance public engagement through digital interfaces.