UC Santa Cruz

Stories are the most valuable currency of human communication; a fact manifested in our social lives, cultural identities and prevalent forms of entertainment. Seeking the reasons behind this level of influence takes us to a journey in human cognition. Stories are also the currency of our situated understanding of events and experiences. This deep cognitive link not only speaks to the reasons behind the influence of stories, but it also outlines specific cognitive processes involved in storytelling between humans. Awareness of these cognitive processes can enable a storyteller to tell better stories, to the point that Herman’s Storytelling and the Sciences of the Mind recommends that cognitive scientists would benefit from studying narratology and narratologists would benefit from studying cognitive science. More specifically, Keith Oatley discusses how a storyteller uses a process of theory-of-mind to tell a story tailored for the perception by a listener. Thus, when evaluating generated stories in the field of computational story generation, we also need to focus on the cognitive processes involved in the perception of stories. Crucially, the contexts in which the generated stories are used, and the approaches with which they are generated, introduce a strong influence over how this evaluation can be performed. If the semantics of the domain in which the stories are generated are known, such as it is the case in games, then a much wider set of approaches become available to both generation and evaluation of stories. However, given the advances in story generation and potential expanded use cases of it in the future, such as interactive sociable agents, it is increasingly inefficient to assume the semantics of a domain and perform knowledge engineering before generating stories. To this end, I focus on open story generation, in which such a priori semantic models are not assumed. It is decidedly more challenging to generate stories through open story generation, and it is particularly more challenging to evaluate them. I believe that a focus on the perception of stories should be an integral part of this generation and evaluation, and I see open story generation to be the most compatible approach with future use cases. To this end, in this dissertation, I offer a consolidation of literature review and an evaluated expanded theory of perceived interestingness in stories. I then report on an evaluated approach to generating stories without assuming a priori semantics and using the event sequences of past interactions. Further, I will introduce an evaluation metric for the perception of stories that focuses on predictive inference and consequently cognitive interest, and show this measure to correlate with human judgment. Lastly, I will report on "stories in the wild", the tales of two prototypes developed and evaluated in the domains of music listening and online shopping, that use story generation techniques while incorporating aspects of story interestingness theories.