Associations of Automated Metadata Extraction and Model Uncertainty with Dataset Reuse in Open Science Repositories

Abstract

The proliferation of open science repositories has democratized access to primary research data, yet the actual reuse of these datasets remains severely hindered by inconsistent, incomplete, or absent metadata. As the volume of deposited data scales exponentially, manual metadata curation becomes an intractable bottleneck. Consequently, automated metadata extraction utilizing advanced natural language processing has emerged as a critical solution. However, the inherent ambiguities in scientific language and the variability in dataset documentation introduce significant predictive challenges. This paper investigates the phenomenon of dataset reuse by analyzing the interplay between automated metadata extraction systems and their corresponding model uncertainty. By quantifying both epistemic and aleatoric uncertainty during the extraction phase, we propose a comprehensive framework that evaluates how the confidence of algorithmic metadata generation influences the discoverability and subsequent utilization of scientific datasets. The study systematically examines extensive repository logs, applying extraction algorithms to unstructured documentation while continuously monitoring uncertainty metrics. The findings demonstrate that dataset reuse is not solely a function of metadata presence but is deeply correlated with the reliability of the extraction process, as encapsulated by low model uncertainty. This research provides a foundational understanding of how uncertainty-aware machine learning can optimize data infrastructure, ultimately fostering a more robust and transparent open science ecosystem.