Metadata completion using Large Language Models

Abstract

A large number of datasets and metadata are published via data catalog sites. The availability of metadata properties such as title and description is high, but metadata properties like temporal and spatial coverage are often difficult to obtain. Temporal and spatial coverage are key properties for federated search, but assurance of metadata poses challenges. This study attempts to explore the completion of temporal and spatial coverage in metadata using LLMs and aims to reveal its effectiveness and limitations. Additionally, we propose the "JSON-LD fill-in-the-blank method" as a prompt design for metadata completion.