Annotation and Prediction of Movie Sentiment Arcs

Some narratologists have argued that all stories derive from a limited set of archetypes. Specifically, Vonnegut (2005) claims in his Shapes of Stories lecture that if we graph the emotions in a story over time, the shape will be an instance of one of six basic story shapes. The work of Jockers (2015) and Reagan et al. (2016) purports to confirm this hypothesis empirically using automatic sentiment analysis (rather than manual annotations of story arcs) and algorithms to cluster story arcs into fundamental shapes. Later work has applied similartechniques to movies (Del Vecchio et al., 2019). This line of work has attracted criticism. Swafford (2015) argues that sentiment analysis needs to be validated on and adapted to narrative text. Enderle (2016) argues that the various methods to reduce story shapes to the putative six fundamental types are actually producing algorithmic artifacts, and that random sentiment arcs can also be clustered into six “fundamental” shapes.In this paper I will not attempt to find fundamental (or even universal) story shapes, but I will take the observed story shape for each narrative as is, without trying to cluster them into archetypes. My aim is to perform an empirical validation of how well basic sentiment analysis tools can reproduce a sentiment arc obtained through manual annotation based on narrative text. Rather than considering novels as narratives, I consider movies, since the annotation ofmovies, when done in real time, is less time consuming. In a previous abstract, I considered the task of predicting the annotated sentiment of individual sentences from movie scripts (van Cranenburgh, 2020), and concluded that sentiment analysis tools achieve comparable performance on narrative text as compared to reviews and social media text (pace Swafford 2015). In this abstract I consider the task of predicting the overall sentiment as annotated based on watching the movie. This task is more challenging since the connection between the narrative sentiment and the narrative text is potentially more distant