Heritage connector: A machine learning framework for building linked open data from museum collections

As with almost all data, museum collection catalogues are largely unstructured, variable in consistency and overwhelmingly composed of thin records. The form of these catalogues means that the potential for new forms of research, access and scholarly enquiry that range across multiple collections and related datasets remains dormant. In the project Heritage Connector: Transforming text into data to extract meaning and make connections, we are applying a battery of digital techniques to connect similar, identical and related objects within and across collections and other publications. In this article, we describe a framework to create a Linked Open Data knowledge graph from digital museum catalogues, perform record linkage to Wikidata, and add new entities to this graph from textual catalogue record descriptions (information retrieval). We focus on the use of machine learning to create these links at scale with a small amount of labelled data, and models which are small enough to run inference on datasets the size of museum collections on a mid‐range laptop or a small cloud virtual machine. Our method for record linkage against Wikidata achieves 85%+ precision with the Science Museum Group (SMG) collection, and our method for information retrieval is shown to improve NER performance compared with pretrained models on the SMG collection with no labelled training data. We publish open‐source software providing tools to perform these tasks.