Text and Networks: Mapping changes in the expertise diversity of patent teams

Alina Lungeanu, Ryan Whalen, Jasmine Yutong Wu, Leslie DeChurch, Noshir Contractor

Contact: jasminewu@u.northwestern.edu

Innovation in science and technology is increasingly the province of teams. Tackling complex scientific problems requires teams to have members with specialized expertise in diverse areas. In order for these teams to spark new ideas, they need the ability to bridge knowledge that resides in multiple individuals. This study addresses that shortfall by developing a measure of team expertise diversity. Unlike prior research that typically examined members’ broad-based disciplinary affiliations or keywords of their publications, we leverage the full textual data of research outputs themselves to map individuals’ expertise and utilize those to measure expertise diversity across individuals in a team. We illustrate the value of our novel measure of team expertise diversity to address two research questions across all patented scientific inventions: How has team expertise diversity changed over time? Does expertise diversity of teams vary across different scientific domains? To do this, we create measures of expertise and diversity. Specifically, we draw on the USPTO patent data that include descriptions of over 6 million U.S. patents from 1976 to 2018, data on the citations between them, as well as demographic information about inventors. We use these data to identify team members’ research expertise and the degree to which their expertise was diverse. We apply vector space modeling to represent texts in an n-dimensional space. To measure expertise for each individual, we identify all the patents on which he or she is listed as an inventor. Then we use Doc2Vec—a neural network approach—to process the text of each patent to situate it within an n-dimensional vector space. Patents that are textually similar are situated near one another in this vector space. An inventor’s expertise is situated in this same vector space at the centroid of the locations in vector space of all the scientific patents created by that inventor. The expertise similarity between two inventors on a team is computed as the cosine similarity between the locations of their expertise in vector space. The expertise similarity between individuals is represented as a network link between them. Finally, team expertise diversity is computed as the standard deviation of the dyadic measures of expertise similarity between pairs of inventors. Preliminary results show the expertise of team members became more diverse from 1976 until 1995. However, after 1995 team expertise diversity has remained relatively constant with a slight increase. Furthermore, the overall pattern is similar across various patent domains but the magnitude of expertise diversity varied: Human Necessities and Chemistry have the lowest team expertise diversity, whereas New Technological Developments have the highest team expertise diversity. In sum, this paper provides network insights into the expertise diversity of scientific teams. By relying on the embedded text of prior inventions, we develop new methodological techniques to analyze scientific teams and the networks that underlie them. It helps guide future research by both providing novel empirical insights into expertise diversity, as well as methodological approaches that help future researchers understand the evolution of scientific networks over time and across scientific domains.

← Schedule