Detecting Scenes in Fiction: A new Segmentation Task

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 3167–3177 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 3167 Detecting Scenes in Fiction: A new Segmentation Task Albin Zehe∗◦ Leonard Konle∗◦ Lea D¨umpelmann• Evelyn Gius† Andreas Hotho◦ Fotis Jannidis◦ Lucas Kaufmann◦ Markus Krug◦ Frank Puppe◦ Nils Reiter‡ Annekea Schreiber† Nathalie Wiedmer⋄ ∗Equal Contribution ◦University of W¨urzburg •University of Heidelberg †TU Darmstadt ‡University of Cologne ⋄University of Stuttgart ◦{zehe, hotho, krug, puppe}@informatik.uni-wuerzburg.de, {leonard.konle, fotis.jannidis}@uni-wuerzburg.de, lucaskaufmann@gmx.de †evelyn.gius@tu-darmstadt.de, annekealuise.schreiber@stud-mail.tu-darmstadt.de •lea.duempelmann@gmx.de ‡nils.reiter@uni-koeln.de ⋄nathalie.wiedmer@ims.uni-stuttgart.de Abstract This paper introduces the novel task of scene segmentation on narrative texts and provides an annotated corpus, a discussion of the lin- guistic and narrative properties of the task and baseline experiments towards automatic solu- tions. A scene here is a segment of the text where time and discourse time are more or less equal, the narration focuses on one action and location and character constellations stay the same. The corpus we describe consists of German-language dime novels (550 k tokens) that have been annotated in parallel, achiev- ing an inter-annotator agreement of γ = 0.7. Baseline experiments using BERT achieve an F1 score of 24 %, showing that the task is very challenging. An automatic scene segmentation paves the way towards processing longer narra- tive texts like tales or novels by breaking them down into smaller, coherent and meaningful parts, which is an important stepping stone to- wards the reconstruction of plot in Computa- tional Literary Studies but also can serve to im- prove tasks like coreference resolution. 1 Introduction Text segmentation is a long standing issue in the area of natural language processing (NLP) encom- passing different tasks like segmenting a text into sentences or finding the boundaries between dif- ferent topics. In this paper, we introduce the task of scene segmentation. A scene can be understood as a segment of a text where the story time and the discourse time are more or less equal, the nar- ration focuses on one action and space and char- acter constellations stay the same. Scenes can be found predominately in narrative texts like novels or biographies, which can be understood as a se- quence of segments, where some of the segments are scenes and others are not. Scene segmenta- tion is of great interest for the high-level analysis of longer texts, for example the reconstruction of plot, but also for many areas of NLP that deal with longer narrative texts, since even modern methods struggle with processing text longer than a cou- ple of sentences or paragraphs. As an example, the memory requirements of state of the art coreference resolution models scale with O(n4) (input length n) (Lee et al., 2017; Joshi et al., 2020) and their performance deteriorates on longer texts. There- fore, it is very helpful to break down the texts into smaller pieces where the character constellation remains the same, enabling us to perform corefer- ence resolution within a scene and then match the characters identified across multiple …

Content truncated to 500 words