Can believable characters act unexpectedly?

Unexpectedness is a major factor controlling interest in narratives. Emotions, for instance, are felt intensely if they are associated with unexpected events. The problem with generating unexpected situations is that either characters, or the whole story, are at risk of being no longer believable. This issue is one of the main problems that make story design a hard task. Writers face it on a case by case basis. The automatic generation of interesting stories requires formal criteria to decide to what extent a given situation is unexpected and to what extent actions are kept believable. This paper proposes such formal criteria and makes suggestions concerning their use in story generation systems. 1. The unexpectedness-believability dilemma Interest in narratives crucially relies on the author’s ability to design unexpected situations. The exercise requires however a bit of caution, as the following example illustrates. The Knife Story. John and Mary are true lovers. Over time, their love is growing. On that Tuesday, Mary has breakfast with John as usual. She stands up, goes to the kitchen, grabs a knife, returns and stabs John in the back. This example illustrates the mutual exclusion between unexpectedness and believability. Mary’s action does not make sense at this point of the story. Readers are in a desperate need for an explanation that will restore Mary’s rationality. In the absence of such explanation, or if it comes too late, Mary’s character is at risk of appearing non believable, and the story’s attractiveness will suffer as a result. Few authors addressed the issue on a general basis. The narrative generator Prevoyant (Bae and Young, 2008) is an attempt to generate flashback and foreshadowing, specifically targeted at the evocation of surprise in the reader’s mind. Surprise is mainly aroused by the manipulation of temporal structures in the narrative. Foreshadowing provides the reader with expectations and flashback provides the reader with an explanation of the surprising event. Using a reader model, the system evaluates both the presence of an unexpected event and the fact that the story structure as a whole will hang together and make sense to the reader. Suspenser (Cheong and Young, 2008) is a framework that determines narrative contents intended to arouse high level of suspense in the reader. This system relies on the idea that a reader’s suspense level is affected by the problems that characters must face and by the number of solutions available to them. The system manipulates the story events in order to increase or decrease the chances of success and also proposes a measure of the level of suspense. Both studies, Prevoyant and Suspenser, address the question of generating coherent stories that arouse an effect in the reader. However, the problem of creating interesting situations by generating surprise or suspense is addressed considering only some specific aspect of these notions. No general theoretical framework that would serve as guidance for the generation of interesting and believable narratives is provided.Reconciling unexpectedness and believability is generally considered as a skill that often makes the difference between good and average quality fiction. The automatic generation of interesting stories requires some formal criterion to decide to what extent a given situation is unexpected and whether actions are kept believable. The aim of this paper is to offer such a formal criterion and to show how it could be implemented in story generation systems. In what follows, we will first illustrate with an example (the Grail Story) how unexpectedness is used by authors to raise interest. They often bring characters to the frontiers of believability. To reproduce this skill, a theoretical framework is needed. We consider such a framework, Simplicity Theory, in a further section. After a brief presentation of the theory, we will show how it can be used to formally characterize two notions: surprise and intentionality. Thanks to these definitions, we will be able not only to express the unexpectedness-believability dilemma, but also to get around it. We will then go further, showing how the theoretical definitions can lead to practical calculus and, eventually, could be used in story generation systems.