NASA has developed an AI tool called the GCMD Keyword Recommender (GKR) to assist with tagging scientific data. The GKR model recommends standardized keywords to help data providers assign the right tags automatically. This not only reduces the workload on human curators but also ensures high-quality metadata. By making data more accessible and organized, the GKR model simplifies the process of finding relevant Earth science datasets for researchers, students, and the public.
The upgraded GKR model addresses the challenge of extreme multi-label classification, where it must select multiple keywords from a set of thousands to accurately tag datasets. To enhance its performance, NASA’s team implemented focal loss, a technique that focuses on underrepresented labels rather than common ones. With access to a larger dataset of over 43,000 records from NASA’s Common Metadata Repository, the GKR model now considers over 3,200 keywords, enabling more precise predictions and improved accuracy.