How do Patent Claims Reflect Scientific Knowledge : Insights from Human Genomics

Patent claims are written in legal jargon and primarily read by patent experts. Most scientists and researchers struggle to comprehend scientific knowledge behind patent claims. Drawing upon literature in science-innovation, I ask following questions: 1) how to make patent claims available as early-source information to R&D frontiers? and 2) how to measure the quality of science in each embodiment disclosed within every patent claim? Leveraging on recent extensive curated database of patent citations to scientific articles, I use sentence encoder to reveal the knowledge behind each independent patent claim. I train Bidirectional-Encoder-Representation-Transformer model on patent data in human genomics, then fine-tune with expert validated data, to predict the most similar scientific sentence to each claim based on maximum cosine similarity value. I observe that the first independent claims which reflect quality science are usually shorter than their comparable claims. The extracted scientific sentences can help scientists 1) to detect the presence of simultaneous discovery in order to reduce duplication of R&D expenditures and 2) to locate useful data in large-scale open science resources (e.g. genetic sequences, chemical compounds, etc.).