A Bayesian Mixed Effects Model of Literary Character

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 370–379, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics A Bayesian Mixed Effects Model of Literary Character David Bamman School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA dbamman@cs.cmu.edu Ted Underwood Department of English University of Illinois Urbana, IL 61801, USA tunder@illinois.edu Noah A. Smith School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA nasmith@cs.cmu.edu Abstract We consider the problem of automatically inferring latent character types in a collec- tion of 15,099 English novels published between 1700 and 1899. Unlike prior work in which character types are assumed responsible for probabilistically generat- ing all text associated with a character, we introduce a model that employs mul- tiple effects to account for the influence of extra-linguistic information (such as au- thor). In an empirical evaluation, we find that this method leads to improved agree- ment with the preregistered judgments of a literary scholar, complementing the results of alternative models. 1 Introduction Recent work in NLP has begun to exploit the potential of entity-centric modeling for a vari- ety of tasks: Chambers (2013) places entities at the center of probabilistic frame induction, show- ing gains over a comparable event-centric model (Cheung et al., 2013); Bamman et al. (2013) ex- plicitly learn character types (or “personas”) in a dataset of Wikipedia movie plot summaries; and entity-centric models form one dominant approach in coreference resolution (Durrett et al., 2013; Haghighi and Klein, 2010). One commonality among all of these very dif- ferent probabilistic approaches is that each learns statistical regularities about how entities are de- picted in text (whether for the sake of learning a set of semantic roles, character types, or link- ing anaphora to the entities to which they refer). In each case, the text we observe associated with an entity in a document is directly dependent on the class of entity—and only that class. This re- lationship between entity and text is a theoreti- cal assumption, with important consequences for learning: entity types learned in this way will be increasingly similar the more similar the do- main, author, and other extra-linguistic effects are between them.1 While in many cases the topi- cally similar types learned under this assumption may be desirable, we explore here the alterna- tive, in which entity types are learned in a way that controls for such effects. In introducing a model based on different assumptions, we provide a method that complements past work and pro- vides researchers with more flexible tools to infer different kinds of character types. We focus here on the literary domain, exploring a large collection of 15,099 English novels pub- lished in the 18th and 19th centuries. By account- ing for the influence of individual authors while in- ferring latent character types, we are able to learn personas that cut across different authors more ef- fectively than if we learned types conditioned on the text alone. Modeling the language used to de- scribe …

Content truncated to 500 words