Patci — a tool for identifying scientific articles cited by patents

Abstract

Scientific research increasingly drives innovation and development of new technologies, and patent-to-paper citations can be used to trace this diffusion of knowledge and measure these science-to-technology spillover effects . However, the so-called “non-patent citations” in USPTO records do not contain authoritative identifiers, nor do they adhere to a standard format. They are strings written in free-form, often much too free, which makes it harder to systematically identify the articles or pieces of work cited. Here, we introduce Patci -- a tool that takes a citation string and probabilistically identifies matching records from a set of bibliographic databases. It currently permits matching to biomedical literature (21.5M PubMed records) and computing/information sciences literature (3.2M DBLP records). It uses a probabilistic model trained on USPTO records but works well for citations originating from outside the patenting sphere. The algorithm extracts and weighs several hundred predictive features and does not rely on punctuation as delimiters of fields. A match probability as attached to each source link ID (e.g., PMID) which permits setting application-appropriate level of match stringency and permits sensitivity analysis. All 16M citations listed in granted USPTO patents (1975-present) have been processed and is available as a separate dataset.