Speakers
Description
Heliophysics papers frequently reference mission data informally rather than through formal citations, leaving data usage often untracked. We have created a pipeline that uses large language models to automatically extract data references from heliophysics papers. These data references consist of missions, instruments, and observed time ranges. For instance, a paper may describe using SOHO/LASCO observations to research CMEs, but not provide any clear citations. In this case, our system would identify that the paper uses the SOHO mission, LASCO instrument, and a time range used in the paper. The system has been manually validated against 5,000+ claims, achieving an aggregate precision of 99.88%. While initial extraction relied on proprietary models (GPT-5.X), our benchmarking found that open-source alternatives (gpt-oss-120B) perform comparably on this task. This has allowed us to scale to over 40,000 papers at one-tenth the cost. The extracted metadata has also been integrated into mission landing pages on the HelioData site, allowing for a high-level overview of the publications that use a mission’s data.