Toward a Data-Driven Theory of Narrativity

Abstract

In this essay, we provide a framework for the empirical testing of narrative theory using the process of machine learning and predictive modeling. Drawing on a collection of over thirteen-thousand passages from an array of different genres, our models suggest that a very small number of features are highly predictive of narrative communication and that these features strongly align with reader judgments. According to our models, narrativity can best be described by what we call the "distant worlds theory," where narrative communication is most strongly identified through the depiction of concretized actions of an agent set at a distance to the teller. These findings raise interesting questions with respect to the deictic and distanciating functions of narration as a cultural practice. Ultimately, we argue that predictive modeling can serve as a valuable tool in the literary critical toolkit to address the problems of theory validation, theory reduction, and theory development. Predictive modeling can help us move past expert opinion as the sole form of validation and gain confidence about the generalizability of our theories about literary behavior in the world.