Bayesian Modelling of the Well-Made Surprise

The “well-made” surprise is a narrative pattern of setting up and executing a surprise in a way that is generally perceived as enjoyable and rewarding. It leverages biases in human cognition to manipulate the audience’s state of belief, and is commonly found in western culture as early as Aristotle’s Poetics. We propose a novel framework to model the audience’s beliefs of a narrative world using approximate Bayesian inference over Markov Logic Networks. We operationalise three qualitative attributes of the well-made surprise (consistency, divergence and certainty) as quantitative functions of the outputs of inference. This work follows the paradigm from computational narrative of operationalising qualitative concepts from literary theory in order to model and generate narratives, either autonomously or cooperatively with a human author. We demonstrate the proposed framework on ten short narratives, and test it with a study on 91 participants. We find that for consistency and divergence, a change in the model’s prediction corresponds with a significant change in the participants’ rating. Our results suggest that the proposed framework may have meaningful predictive power and potential for future applications to narrative generation, plot analysis, and computer-aided creativity.