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

Abstract

Characters are fundamental to literary analysis. Current approaches are heavily reliant on NER to identify characters, causing many to be overlooked. We propose a novel technique for character detection, achieving significant improvements over state of the art on multiple datasets.

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)

“…Him included as a place. The other clusters are messier, but still in- (Vala et al, 2015). In any case, Brown clustering works fairly well for common names, but for rarer ones, the clustering is haphazard.…”

“…Though there are a few novels which have been tagged for characters (Vala et al, 2015), we wanted to test our system relative to a much wider range of fiction. To this end, we randomly sampled texts, sentences, and then names within those sentences from our name-segmented Project Gutenberg corpus to produce a set of 1000 examples.…”

“…Named entity recognition (NER) of person names is generally the first step in identifying characters; locations are also a prevalent NE type, and can be useful when tracking different plot threads (Wallace, 2012), or trends in the settings of fiction. There are not, to our knowledge, any NER systems that are specifically targeted at literature, and most related work has used Stanford CoreNLP as an off-the-shelf solution Vala et al, 2015). In this paper, we show that it is possible to take advantage of the properties of fiction texts, in particular the repetition of names, to build a high-performing 3-class NER system which distinguishes people and locations from other capitalized words and phrases.…”

“…In the context of literature, the most closely related task is character identification (Vala et al, 2015), which is itself an intermediate task for character speech identification , analysis of characterization , and analysis of social networks (Elson et al, 2010;Agarwal et al, 2013;Ardanuy and Sporleder, 2015). In addition to NER, character identifica-tion also involves clustering multiple aliases of the same character, and discarding person names that don't correspond to characters.…”

“…In addition to NER, character identifica-tion also involves clustering multiple aliases of the same character, and discarding person names that don't correspond to characters. Vala et al (2015) identify some of the failures of off-the-shelf NER with regards to character identification, and attempt to fix them; their efforts are focused, however, on characters that are referred to by description rather than names or aliases.…”