Computationally Creative Search for Stories

Storytelling is an ubiquitous phenomenon in human culture. We tell stories to entertain, explain, and illustrate. The production of novel stories is undeniably an act of creativity in humans. How can an autonomous computer system create novel story content? What can we learn about human creativity in the investigation of computational models of story generation? Relevant past work in story generation Fictional plot generation can be thought of as a creative design process. The goal is to produce an artifact – in this case a plot, a sequence of temporally ordered events that involve characters in a story world – that has a particular set of properties that are either implicitly required or explicitly given. These properties constrain the space of possible solutions. Examples of implicit constraints are properties of plot that we assume that all " well-formed " plots should include such as logical causal progression, and character believability. Examples of explicit constraints are that the plot achieves some purpose – dramatic impact on the reader, pedagogical validity, etc. – and occurs in a particular given world consisting of particular characters. Plot generation is often under-constrained, giving rise to the possibility of creative solutions. Boden (2004) describes two forms of creativity: creativity through search and creativity through transformation. Creativity through transformation is the notion that an existing, seemingly unrelated concept can be transformed into an unique new concept. Following Boden, plot generation can be considered as search for an artifact – a plot – that is valued (Riedl and Young 2006). Consider the space of all possible plots where a plot is a set of temporally arranged events that change the story world (for completeness we also consider the empty plot). In this space, plots are adjacent if they differ by only one detail: the events comprising the plot and/or their temporal configuration. One can walk the space of all possible plots, starting with the empty plot, by repeatedly selecting an event and adding it to the plot