What We Can Expect and What is Unexpected: Psychological and Computational Experiments

Unexpected events often occur in our everyday lives, yet little research has gone into how we might think about unexpected events occurring in the future. A review of the literature shows that research often looks at what we expect or plan for our futures, and only considers the unexpected retrospectively. In this thesis, we advance a theory of prospective unexpected events; that is, how we think about unexpected futures. In several behavioral studies, we find that people are negatively biased about the future unexpected, but that this is dependent on aspects of the preceding events. Namely, people take into consideration the a priori valence and controllability of current events when thinking about what unexpected events could occur next. If events are a priori negative or uncontrollable, we see a reduction in the number of negative unexpected events produced. We use state-of-the-art transformer language models to model human behavior in this domain, and discuss the implications for both human behavior and the capabilities of these large language models. We find that a classification model appears to rely on the negative-sentiment terms in classifying events as unexpected, mirroring findings in the behavioral studies. Moreover, when we ask GPT-2 to generate unexpected events, people are sensitive to the level of unexpectedness produced. We discuss the findings of these studies and their implications for cognitive science as well as that of natural language processing.