Mr. Bennet, his coachman, and the Archbishop walk into a bar but only one of them gets recognized: On The Difficulty of Detecting Characters in Literary Texts

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 769–774, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Mr. Bennet, his coachman, and the Archbishop walk into a bar but only one of them gets recognized: On The Difficulty of Detecting Characters in Literary Texts Hardik Vala1, David Jurgens1, Andrew Piper2, Derek Ruths1 1School of Computer Science 2Department of Languages, Literatures, and Cultures McGill University, Montreal, QC, Canada hardik.vala@mail.mcgill.ca, jurgens@cs.mcgill.ca andrew.piper@mcgill.ca, derek.ruths@mcgill.ca Abstract Characters are fundamental to literary analysis. Current approaches are heav- ily reliant on NER to identify characters, causing many to be overlooked. We pro- pose a novel technique for character detec- tion, achieving significant improvements over state of the art on multiple datasets. 1 Introduction How many literary characters appear in a novel? Despite the seeming simplicity of the question, precisely identifying which characters appear in a story remains an open question in literary and nar- rative analysis. Characters form the core of many computational analyses, from inferring prototypi- cal character types (Bamman et al., 2014) to iden- tifying the structure of social networks in literature (Elson et al., 2010; Lee and Yeung, 2012; Agar- wal et al., 2013; Ardanuy and Sporleder, 2014; Jayannavar et al., 2015). These current approaches have largely assumed that characters can be reli- ably identified in text using standard techniques such as Named Entity Recognition (NER) and that the variations in how a character is named can be found through coreference resolution. However, such treatment of character identity often over- looks minor characters that serve to enrich the so- cial structure and serve as foils for the identities of major characters (Eder et al., 2010). This work provides a comprehensive exami- nation of literary character detection, with three key contributions. First, we formalize the task with evaluation criteria and offer two datasets, in- cluding a complete, manually-annotated list of all characters in 58 literary works. Second, we pro- pose a new technique for character detection based on inducing character prototypes, and in compar- isons with three state-of-the-art methods, demon- strate superior performance, achieving significant improvements in F1 over the next-best method. Third, as practical applications, we analyze liter- ary trends in character density over 20 decades and revisit the character-based literary hypothesis tested by Elson et al. (2010). 2 Related Work Character detection has primarily been performed in the context of mining literary social networks. Elson et al. (2010) extract character mentions from conversational segments, using the Stan- ford CoreNLP NER system to discover character names (Manning et al., 2014). To account for vari- ability in character naming, alternate forms of a name are generated using the method of Davis et al. (2003) and merged together as a single char- acter. Furthermore, the set of aliases for a char- acter is expanded by creating coreference chains originating from these proper names and merging all coreferent expressions. Agarwal et al. (2013) also rely on the CoreNLP NER and coreference resolution systems for character detection; how- ever for …

Content truncated to 500 words