A dataset of scientific citations in U.S. patent Office Actions

How can we create a dataset of scientific citations from U.S. patent Office Actions?

Understanding patent citations is critical for analyzing innovation and technology development. Traditional methods have overlooked the nuances of citations present in Office Actions, leading to incomplete insights into the examination process and the relationship between science and technology. By providing access to a richer dataset, we enable researchers to explore examiner behavior and the impact of scientific literature on patent decisions, supporting more accurate citation-based metrics and analyses of knowledge diffusion and technology transfer.

We compiled a dataset from approximately 850,000 citations found in Office Actions issued by the U.S. Patent and Trademark Office, focusing particularly on 265,000 references to scientific literature. We used machine learning techniques to classify, parse, and disambiguate these citations, which we linked to OpenAlex to enhance usability. We conducted the work using publicly available patent data originating from the USPTO 'Enriched Citation API', with careful attention to the classification of citations into 14 distinct types, focusing on scientific references.

• 8% of examined patents had complete overlap between front-page and Office Action citations, while 60% contained extra scientific references only found in Office Actions.
• The matching accuracy of citations using OpenAlex improved from 70% to 96% based on relevance scores.
• 99% of sampled citations combined via the Crossref match were validated as accurate, highlighting the reliability of the consolidated dataset.

• Approximately 850,000 citations from Office Actions issued by the USPTO
• 265,880 parsed scientific references
• Validation checks on 420 OA citations classified into 14 categories

• Citations accurately reflect examiner reasoning and the relevance of prior art during patent prosecution.
• Machine learning models trained on a subset of the data generalize well for classifying all citations.
• The OpenAlex and Crossref databases provide reliable metadata for citation consolidation.