Computational Creativity in Narrative Generation: Utility and Novelty Based on Models of Story Comprehension

Narrative is one of the fundamental modes of creative expression. Creativity plays a central role in the effective design of film, novels, oral stories, computer games and most other narrative media. Beyond the creation of narratives intended to entertain, narrative is used as a powerful mode of communication in a range of other contexts. Narrative and its creative design plays a role in education and training (Gee, 2003), the communication of scientific ideas (Hoffman, 2005), characterizations of business practices (Thomas, 1999), communicating cultural heritage (Collins et al, 2003) and other contexts. There is growing awareness of the foundational role of narrative in our understanding of not only fictional worlds (Gerrig, 1993) but also the real world around us (Bruner 1991). While it is clear that the process of generating a narrative is systematic, and sometimes formulaic, the challenge of generating an interesting narrative is at the brink of our scientific and scholarly understanding (Graesser, Pomeroy, & Craig, 2001). That is, there is an insufficient explicit understanding of the means by which creativity shapes the design of a narrative artifact that has an interesting plot, an illuminating point, and a pacing that is captivating. We do know that much of the research on creativity shares a common notion of its nature: An artifact that is at once fulfilling a specific notion of utility while doing so in a way that is novel or unexpected (Perkins, 1981; Sternberg, 1997). Computational approaches to creativity have often focused on methods for encoding the creation process as search through a design space (Lindsey, et al, 1980; Langley, et al, 1987). Rules or heuristics are used to encode notions of creativity by directing search through the design space in ways that focus the search on creative solutions or by modifying the search space to create designs that are unexpected or novel. Structural constrains on the space ensure that solutions found within it all meet the utility criteria, while directed search or careful redefinition of the space (Lenat 1983) give rise to solutions that satisfy the novelty constraint. Recent approaches to the generation of narrative have employed similar search techniques. While these methods have shown initial successes, most are essentially focused inward, that is, they reason about a narrative’s structure rather than how it will be experienced by its readers or viewers. Narrative is a communicative medium and much of its structure is composed by story tellers to change the mental state of those who experience it. A narrative is considered creative due directly to inferences made regarding its structure by viewers. A system seeking to build a creative narrative can exploit psychological models of narrative comprehension to produce stories that are both coherent (that is, they satisfy the utility constraint) and surprising (that is, they intentionally give rise to expectations that are later violated, thus satisfying the novelty constraint) (Graesser et al., 1998). This paper describes very early work in progress whose objective is to develop new, cognitively informed computational models of creativity in the context of the generation of cinematic narrative – narrative that is told within 3 dimensional (3D) virtual environments. Motivated by narrative theoretic models of narrative structure (Polti, 1977; Chatman, 1980; Rimmon-Kenan, 1983) and psychological models of narrative comprehension (Graesser et al., 1998; Graesser & Klettke, 2001), we are developing techniques for creating stories and the cinematic techniques needed to convey them to users. These techniques will use these models to search for narratives that are at once coherent and surprising, satisfying these two constraints to generate creative narratives.