Transforming Scientific Literature into Machine-Actionable Knowledge

Abstract

Scientific literature is a vast repository of knowledge, yet much of its valuable data remains trapped in unstructured formats, limiting its accessibility for machine learning applications. Extracting and structuring information from figures and tables in PDFs presents a significant challenge due to their complex and information-dense nature. We introduce MERMaid (Multimodal aid for Reaction Mining), an end-to-end knowledge ingestion pipeline that leverages vision-language models to extract, interpret, and integrate visual data into structured knowledge graphs. MERMaid achieves 87% accuracy in converting reaction conditions from electrochemistry literature into machine-actionable formats, demonstrating advanced chemical context awareness, self-directed context completion, and robust coreference resolution. Its modular and topic-agnostic design allows adaptation across diverse scientific domains. By unlocking structured insights from literature at scale, MERMaid bridges the gap between human knowledge and computational analysis, enabling more efficient knowledge discovery and accelerating data-driven scientific research.