Full Text

Why are some narratives more successful? Although this question has ancient roots, studying it empirically has been challenging. We suggest that semantic progression (i.e., semantic similarity between adjoining portions of a narrative) might shape audience responses but that this role changes over the course of a narrative. Specifically, although slower semantic progression (i.e., greater semantic similarity between adjoining portions) is beneficial at the beginning of narratives, faster semantic progression is beneficial toward the end. To test this possibility, we used natural language processing and machine learning to analyze over 40,000 movie scripts. Consistent with our theorizing, deep-learning-based embeddings find that movies with slower semantic progression early and faster semantic progression later are evaluated more positively. Analysis of over 10,000 TV episodes finds similar results. Overall, these findings shed light on what makes narratives engaging, deepen understanding of what drives cultural success, and underscore the value of emerging computational approaches to understand human behavior.

Narratives are an integral part of everyday life. Early humans shared epic poems and stories of the hunt, and today we read books, watch movies, and consume content online.

But why are some narratives more successful than others?

Everyone from psychologists and philosophers to linguists and narrative theorists have long been interested in quantifying properties of narratives (e.g., Carroll, 1990; Cutting, 2016; Pennebaker, 2018; Propp, 1968). In Poetics (trans. Aristotle & Heath, 1996), Aristotle proposed that stories tend to have common structures, and Freytag (1900) later suggested a dramatic arc or pyramid of rising and falling action. Campbell (1949) theorized that there was an archetypical hero’s journey, and Kurt Vonnegut’s rejected master’s thesis (Vonnegut, 2010) argued that stories could be divided into eight common shapes. Recent work has begun to address this topic empirically, examining variation in sentiment (Reagan et al., 2016) and categories...