Stephen Qingyuan Yang

Unstructured textual information is important for understanding innovation but challenging to study. The advent of large-language models (LLMs) relieves this constraint. This paper develops deep-learning predictive models incorporating ChatGPT textual embeddings to access intricate information about patent quality. These models achieve an R-squared score of 42% predicting patent value and improve the identification of the worst and best applications by 10 percentage points. Using patent value captured by the LLMs, a long-short portfolio generates 6.6% yearly abnormal returns. The models also enable a novel, richer measure of patent value, accounting for potential institutional anticipation. Furthermore, LLMs provide an opportunity to enhance corporate policy vis-à-vis patenting, including revising applications, which could flatten differences in skill across firms and lawyers. Such techniques improve writing but not technological quality---two orthogonal components isolated through a novel decomposition. Both significantly drive patent acceptance yet affect different dimensions of later patent and firm outcomes.

We combine advanced language models with broadly applicable perception tests to overcome restrictive assumptions in "causal" machine learning and identify high-dimensional discrimination using unstructured data. Leveraging large-language-model embeddings to effectively control for patent quality (and its nonlinearities and interactions) and regularities in human learning to gain double robustness, we document systematic bias in patenting by the U.S. Patent & Trademark Office against under-represented groups, especially female and black innovators, unexplained by legally purported criteria for patent approvals. Jointly investigating multiple-dimensional bias for the first time, we also discover a novel, large affiliation bias---individuals are severely disadvantaged relative to employees at large, public companies---that dominates other disparities. Furthermore, factors such as innovation quality and location can either mitigate or compound discrimination, and the interactions of various disparities manifest significantly in approval decisions. For example, racial disparities disappear among public firm employees, suggesting a previous underestimation of discrimination against individual inventors. Existing theories (e.g., that of homophily) do not fully rationalize the findings, but a parsimonious model of correlation neglect does, with direct policy implications.

We develop a model of collective choice that introduces utility to individual voting and study polarized, partisan legislatures. Party allegiances bring about lower social welfare via inefficient policy selection. Bonding across intra-party factions results in multiple voting equilibria and prevents the formation of optimal inter-party coalitions. However, if the legislature is sufficiently polarized and representatives patient, the majority party fractures with significant implications for future outcomes. Intra-party disputes within the majority also lead to a fracturing equilibrium. As a case study, we examine the ousting of Speaker McCarthy and explain why the minority party voted with the polar majority bloc.