CAREER: Automated, Quality, and Trustworthy Scientific Information Extraction from Massive Text Data

Virginia Polytechnic Institute And State University, Blacksburg VA

Investigators

Abstract

Scientific knowledge is growing at an unprecedented rate, with millions of research articles published annually. However, accessing and organizing this vast amount of information remains a major challenge, slowing the pace of discovery and innovation. This project aims to develop automated technologies that allow computers to extract high-quality, trustworthy scientific information from massive text collections without requiring extensive human annotation. By enabling easier access to accurate scientific knowledge, this work will accelerate breakthroughs in fields such as biomedicine, chemistry, and engineering. It will also foster transparency and trust in science and inspire future generations of researchers through education and outreach efforts focused on diversity and inclusion. This project proposes a new paradigm for scientific information extraction that eliminates the need for costly expert annotations while achieving expert-level accuracy. The first thrust focuses on co-optimizing fine-grained typing of diverse scientific entities and relationships within large textual contexts by integrating multiple weak supervision signals, enabling scalable and domain-specific extraction. The second thrust introduces retrieval-augmented techniques that leverage multimodal domain knowledge, such as scientific knowledge bases and related data, to enhance understanding beyond the text. The third thrust develops methods for joint label and explanation generation to build trustworthiness in automated extraction without human labeling. These innovations are implemented using scalable algorithms based on optimal transport theory and Token Turing Machine architectures, with planned deployment of a beta extraction system on PubMed for real-world evaluation. Comprehensive evaluation and ethical safeguards ensure the methods are practical, reliable, and widely applicable across scientific disciplines. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.