Modelling Suspense in Short Stories as Uncertainty Reduction over Neural Representation

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 1763–1788 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 1763 Modelling Suspense in Short Stories as Uncertainty Reduction over Neural Representation David Wilmot and Frank Keller Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB, UK david.wilmot@ed.ac.uk, keller@inf.ed.ac.uk Abstract Suspense is a crucial ingredient of narrative fic- tion, engaging readers and making stories com- pelling. While there is a vast theoretical litera- ture on suspense, it is computationally not well understood. We compare two ways for mod- elling suspense: surprise, a backward-looking measure of how unexpected the current state is given the story so far; and uncertainty reduc- tion, a forward-looking measure of how unex- pected the continuation of the story is. Both can be computed either directly over story rep- resentations or over their probability distribu- tions. We propose a hierarchical language model that encodes stories and computes sur- prise and uncertainty reduction. Evaluating against short stories annotated with human sus- pense judgements, we find that uncertainty re- duction over representations is the best predic- tor, resulting in near human accuracy. We also show that uncertainty reduction can be used to predict suspenseful events in movie synopses. 1 Introduction As current NLP research expands to include longer, fictional texts, it becomes increasingly important to understand narrative structure. Previous work has analyzed narratives at the level of characters and plot events (e.g., Gorinski and Lapata, 2018; Martin et al., 2018). However, systems that pro- cess or generate narrative texts also have to take into account what makes stories compelling and enjoyable. We follow a literary tradition that makes And then? (Forster, 1985; Rabkin, 1973) the pri- mary question and regards suspense as a crucial factor of storytelling. Studies show that suspense is important for keeping readers’ attention (Khrypko and Andreae, 2011), promotes readers’ immersion and suspension of disbelief (Hsu et al., 2014), and plays a big part in making stories enjoyable and in- teresting (Oliver, 1993; Schraw et al., 2001). Com- putationally less well understood, suspense has only sporadically been used in story generation sys- tems (O’Neill and Riedl, 2014; Cheong and Young, 2014). Suspense, intuitively, is a feeling of anticipation that something risky or dangerous will occur; this includes the idea both of uncertainty and jeopardy. Take the play Romeo and Juliet: Dramatic suspense is created throughout — the initial duel, the meet- ing at the masquerade ball, the marriage, the fight in which Tybalt is killed, and the sleeping potions leading to the death of Romeo and Juliet. At each moment, the audience is invested in something be- ing at stake and wonders how it will end. This paper aims to model suspense in computa- tional terms, with the ultimate goal of making it deployable in NLP systems that analyze or generate narrative fiction. We start from the assumption that concepts developed in psycholinguistics to model human language processing at the word level (Hale, …

Content truncated to 500 words