Using LLMs for Semantic Alignment: A Study on Archival Metadata Description

The advantages of aligning custom data schemas with standardised ontologies within their respective knowledge domain have long since been proven in practice. Sharing a common structural representation by mapping concepts and relationships between the schemas is essential to ensure data interoperability (especially on a semantic level), integration, reuse, and the ability to leverage machine-processable and advanced-search capabilities. Archival institutions preserve, manage, and provide access to large amounts of diverse cultural and historical data, demonstrating a high potential to be active contributors to a global knowledge network, should archival data be transformed and offered as linked (open) data. Based on the expert-validated dataset of the mapping (alignment) of the Swedish National Archives schema to the Records-in-Contexts (RiC-O) ontology, the purpose of this study is two-fold. First, to examine whether it is possible to automatically and effectively extend one case (Sweden) to other archival institutions and align new custom schemas to RiC-O, given an expert-curated dataset of this domain. Secondly, using the aforementioned dataset and one more of a few human-evaluated examples of mapping to other cultural heritage ontologies as input, to examine whether an LLM (e.g., GPT-4o) is capable of recommending meaningful alignments for enhanced metadata description to more ontologies within the same domain (CH and archives), but also across other domains. The experiments reveal several challenges and shortcomings of the LLM prompting approach for these tasks, but also possible opportunities to leverage towards this direction.