Frowning Frodo, Wincing Leia, and a Seriously Great Friendship: Learning to Classify Emotional Relationships of Fictional Characters

Proceedings of NAACL-HLT 2019, pages 647–653 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 647 Frowning Frodo, Wincing Leia, and a Seriously Great Friendship: Learning to Classify Emotional Relationships of Fictional Characters Evgeny Kim and Roman Klinger Institut f¨ur Maschinelle Sprachverarbeitung University of Stuttgart Pfaffenwaldring 5b, 70569 Stuttgart, Germany {evgeny.kim,roman.klinger}@ims.uni-stuttgart.de Abstract The development of a fictional plot is cen- tered around characters who closely interact with each other forming dynamic social net- works. In literature analysis, such networks have mostly been analyzed without particular relation types or focusing on roles which the characters take with respect to each other. We argue that an important aspect for the analysis of stories and their development is the emo- tion between characters. In this paper, we combine these aspects into a unified frame- work to classify emotional relationships of fic- tional characters. We formalize it as a new task and describe the annotation of a corpus, based on fan-fiction short stories. The extraction pipeline which we propose consists of char- acter identification (which we treat as given by an oracle here) and the relation classifica- tion. For the latter, we provide results using several approaches previously proposed for re- lation identification with neural methods. The best result of 0.45 F1 is achieved with a GRU with character position indicators on the task of predicting undirected emotion relations in the associated social network graph. 1 Introduction Every fictional story is centered around charac- ters in conflict (Ingermanson and Economy, 2009) which interact, grow closer or apart, as each of them has ambitions and concrete goals (Acker- man and Puglisi, 2012, p. 9). Previous work on computational literary studies includes two tasks, namely social network analysis and sen- timent/emotion analysis, both contributing to a computational understanding of narrative struc- tures. We argue that joining these two tasks lever- ages simplifications that each approach makes when considered independently. We are not aware of any such attempt and therefore propose the task of emotional character network extraction from fictional texts, in which, given a text, a network is to be generated, whose nodes correspond to char- acters and edges to emotions between characters. One of the characters is part of a trigger/cause for the emotion experienced by the other. Figure 1 de- picts two examples for emotional character inter- actions at the text level. Such relation extraction is the basis for generating social networks of emo- tional interactions. Dynamic social networks of characters are ana- lyzed in previous work with different goals, e.g., to test the differences in interactions between various adaptations of a book (Agarwal et al., 2013); to understand the correlation between dialogue and setting (Elson et al., 2010); to test whether social networks derived from Shakespeare’s plays can be explained by a general sociological model (Nal- isnick and Baird, 2013); in the task of narrative generation (Sack, 2013); to better understand the nature of character interactions (Piper et al., 2017). Further, previous work analyses personality traits of characters (mostly) …

Content truncated to 500 words