Predicting the effects of suspenseful outcome for automatic storytelling

Abstract

Automatic story generation systems usually deliver suspense by including an adverse outcome in the narrative, in the assumption that the adversity will trigger a certain set of emotions that can be categorized as suspenseful. However, existing systems do not implement solutions relying on predictive models of the impact of the outcome on readers. A formulation of the emotional effects of the outcome would allow storytelling systems to perform a better measure of suspense and discriminate among potential outcomes based on the emotional impact. This paper reports on a computational model of the effect of different outcomes on the perceived suspense. A preliminary analysis to identify and evaluate the affective responses to a set of outcomes commonly used in suspense was carried out. Then, a study was run to quantify and compare suspense and affective responses evoked by the set of outcomes. Next, a predictive model relying on the analyzed data was computed, and an evolutionary algorithm for automatically choosing the best outcome was implemented. The system was tested against human subjects' reported suspense and electromyography responses to the addition of the generated outcomes to narrative passages. The results show a high correlation between the predicted impact of the computed outcome and the reported suspense.