Assessing Scientific Knowledge in Patents: A Large Language Model Approach

Abstract Researchers often infer science-technology linkage from the scientific non-patent references (sNPRs) that patents cite, yet this single signal can miss or misclassify scientific knowledge. Treating sNPRs and university-assignee status as observable indicators of an underlying latent construct, science-relatedness, this study introduces an LLM-based measure of the extent to which patents rely on scientific knowledge. The GPT-4o model is employed to score a balanced sample of 2,000 USPTO patents, and the resulting science-knowledge scores correlate strongly with both sNPRs and university-assignee status, confirming convergent validity. Where the GPT score diverges from citation or assignee signals, it reduces false negatives (science-driven patents lacking citations) and false positives (technology-driven patents that merely cite the literature). Prompt experiments show that supplying only title, abstract, and claims captures almost the full signal, whereas adding formal definitions of “science” and “technology” weakens it. Treating GPT scores, sNPRs, and assignee status as reflective indicators advances measurement of science-technology linkage, and the prompt experiments provide practical guidance for large-scale LLM-based patent analysis. Peer Review https://www.webofscience.com/api/gateway/wos/peer-review/10.1162/QSS.a.493