Reproduce “Automating the search for a patent's prior art with a full text similarity search” on OpenResearch

About this paper

Lea Helmers, Franziska Horn, Franziska Biegler, Tim Oppermann, Klaus-Robert Müller

Researchers from Technische Universität Berlin and Pfenning, Meinig & Partner mbB developed an automated full-text similarity search system for patent prior art, leveraging machine learning and natural language processing. This approach demonstrates improved accuracy in identifying relevant prior art, particularly by reducing false negatives through semantic understanding provided by neural network language models, outperforming current manual processes and official citation records.

Key insights

• The current manual patent citation process shows low precision (0.43) and moderate recall (0.78) when compared against expert-derived relevance labels, indicating significant flaws.
• Automated full-text similarity scores correlated more strongly with expert relevance judgments (Spearman's ρ = 0.652) than with official patent citations (ρ = 0.501), suggesting a better alignment with actual relevance.
• Neural network language models provide crucial semantic understanding, enabling the detection of relevant prior art that employs different or more abstract terminology, thus addressing a key limitation of keyword-based methods.

Summary by alphaXiv.