Abstract
This paper presents a multi-pronged computational framework for analyzing literary characters, moving beyond simple visualizations to high-dimensional feature spaces. By applying network science and natural language processing (NLP) to thousands of texts spanning four centuries, the authors track the "vanishing periphery" of dramatic casts and reveal how linguistic markers of "interiority" have migrated from 19th-century heroines to modern male science fiction protagonists. The study demonstrates that computational methods can reveal large-scale historical shifts in characterization and social structure that are difficult to capture through traditional qualitative analysis.
Introduction
Literary character study has traditionally been a qualitative endeavor, focusing on specific "types" or the psychological depth of individual figures. However, characters are "semiotically abundant"—proper names and pronouns account for roughly 16% of all words in a novel. This sheer volume makes it difficult for traditional "close reading" to capture large-scale historical shifts or the structural "sociability" of characters across vast corpora.
The authors address three specific gaps in current literary scholarship:
- Structural Evolution: How the social morphology of plays has changed over centuries.
- Defining Interaction: Moving beyond the assumption that "interactions" are self-evident to recognize them as complex verbal constructions.
- High-Dimensional Characterization: Utilizing more than simple word counts to understand how agency, behavior, and description define a character’s identity across genres.
Methodology and Datasets
The research synthesizes three distinct projects, each employing different technical approaches. To study network topology, the authors analyzed 3,439 English plays from the Chadwyck Healy Drama Corpus (1500–1920). They employed Gini Coefficients of Eigenvector Centrality—originally an economic tool—to measure the distribution of social power among characters. They also used Betweenness Centrality to measure "mediatedness," or how characters act as bridges between social groups.
For characterization, the authors developed the "Emma" tool, a custom pipeline implementing a modified version of BookNLP and the Stanford dependency parser. This tool extracts 28 distinct features, including positionality (agency in sentences), modality (behaviors like cogitation and perception), and centrality. Comparative datasets included a 19th-century novel corpus and a modern science fiction corpus to track the evolution of these traits.
The Structural Evolution of Drama
A primary finding of the study is the "vanishing periphery" in English drama. Over 400 years, characters such as servants, guards, and messengers have steadily disappeared as plays moved from the public "throne room" to the more intimate "drawing room." This shift is reflected in the concentration of social power.
In early modern drama, power was typically concentrated in a single protagonist. However, by the 17th century, this function began to disperse across multiple central characters, creating more complex, multi-core social networks. As shown in the figure below, the percentage of characters occupying the top quartile of social importance increased over time, indicating a more democratic distribution of narrative focus.
The data also reveals a massive structural break between 1650 and 1700. The "Puritan shuttering" of theaters created a lasting discontinuity; the drama that returned during the Restoration was fundamentally different in its social structure than what preceded it.
Characterization and Interiority
Beyond network structure, the research highlights a "gendered migration of interiority." Historically, linguistic markers of interiority—focusing on a character's thoughts, perceptions, and internal states—were strongly associated with 19th-century heroines. The "Emma" tool reveals that these same markers have now been transposed onto the male heroes of modern science fiction. This suggests that the way we represent "depth" in characters is not fixed but shifts across genres and eras.
Conclusion
While computational tools offer powerful insights, they face significant hurdles. Machine learning models for interaction extraction currently peak at an F1 score of 0.61, highlighting that social interaction in prose is far more linguistically ambiguous than previously thought. Future work will explore "social investments" in literature, moving from quantifying interactions to understanding the qualities of social relationships that generate the most reader engagement.
Summary updated