Recent advances in machine learning, combined with the increased availability of large natural language online datasets, have opened up new opportunities for understanding human behavior. With these new methods, it is now possible to observe what people read and talk about, and thus think and feel about, a wide range of common objects and events. The goal of this project is to study how these novel methods and datasets can be combined with existing psychological theory to predict and understand human judgment, with the primary application to the domain of risk perception. The computational techniques proposed as part of this project allow for the automatic, large-scale analysis of the perception of naturalistic risks, and thus can be used to identify problems and develop interventions involving risk communication and risk management. More generally, by applying cutting edge methods in data science to the study of human behavior, this project develops novel technologies for predicting and understanding attitudes, beliefs, and preferences. By doing so, it facilitates a wide array of policy and commercial applications not feasible using existing empirical methodologies in the behavioral sciences.
How can we uncover and quantify rich mental representations for all of the objects and concepts that are the target of everyday judgment? How can we use these rich representations to predict judgments and judgment errors, to study domain and individual-level differences in judgment, and to characterize the complex web of associations that underlie judgment? Finally, how can we study differences in mental representations and judgments across cultures and time periods? This project will use "semantic vectors" -- high dimensional representations for words obtained from large-scale language data -- to address these questions. Semantic vectors provide a good proxy for the structure of knowledge and association in people's minds and can be combined with various machine learning algorithms to predict judgments for naturalistic objects and concepts, such as sources of risk. These vectors can also identify the key associates of judgments, facilitating directed hypothesis tests for a diverse array of psychological dimensions, including emotions, moral concepts, and personality. Finally, when trained on different types of language datasets, semantic vectors can shed light on the complex interaction between judgment and culture, history, and society. Note that semantic vectors can be used in this manner for nearly any judgment domain that involves common objects, individuals, and events. Thus one key goal of this project is to evaluate the applicability of semantic vectors for predicting and understanding the many different types of judgments that that people make on a day-to-day basis (including not only risk judgment, but also, for example, health judgment and consumer judgment).
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
