NLP Models for Extracting Knowledge from Museum Provenance Texts

This package provides SpaCy-based Natural Language Processing (NLP) models specifically trained to extract structured data from museum provenance texts. These models were developed using custom-annotated datasets derived exclusively from the provenance records of the Art Institute of Chicago (AIC). These records follow the stylistic and semantic conventions outlined in The AAM Guide to Provenance Research. Training data and model configuration files are provided alongside the models to support reproducibility and further development.

Provenance records document the sequence of ownership and custodianship of cultural objects. In museum contexts, these records are typically written as compact narratives that rely on punctuation, such as semicolons and periods, to separate events. In the case of the AIC, they also include bracketed notes to indicate historical sources. To extract structured data from these texts, the models perform two core tasks. Sentence boundary detection identifies where one event ends and the next begins. Span categorization then locates and labels key information within each event, including people or institutions, dates, places, and methods of transfer. Together, these tasks support the transformation of narrative provenance into structured, machine-readable data.

The two models achieved F1 scores of 0.99 for sentence boundary detection and 0.94 for span categorization on their respective test sets. The annotation scheme used to train these models is introduced in Teaching Provenance to AI: An Annotation Scheme for Museum Data. The training and implementation of the models are detailed in Hidden Value: Provenance as a Source for Economic and Social History. By transforming narrative provenance into structured, machine-readable data, the models support both scholarly analysis and institutional efforts in transparency, and linked open data publication.