💬 NLP

Computational Representations of Character Significance in Novels

This paper proposes a novel six-component structural model of character that incorporates discussion by other characters, demonstrating through computational analysis of 19th-century British realist novels that this approach offers new insights into literary theories of character centrality and gendered dynamics.

Original authors: Haaris Mian, Melanie Subbiah, Sharon Marcus, Nora Shaalan, Kathleen McKeown

Published 2026-01-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Haaris Mian, Melanie Subbiah, Sharon Marcus, Nora Shaalan, Kathleen McKeown

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a novel not just as a story, but as a bustling city. For a long time, literary critics and computer scientists tried to figure out who the "important" people in that city were by simply counting how many times they walked down the street (appeared in a scene) or how many times they shook hands with others (had a conversation).

This paper argues that this method is like judging a person's importance in a city only by how often they are seen in public. It misses the fact that someone might be the most talked-about person in town, even if they stay home all day.

Here is a breakdown of what the researchers did, using simple analogies:

1. The New "Six-Point" Scorecard

Instead of just counting "street appearances," the team (a mix of computer scientists and literature experts) created a new six-part scorecard to measure a character's "space" in the story. Think of it like a report card for a character's presence:

Name (N): How many times is their name written down? (Like seeing their name on a mailbox).
Communication (C): How much do they speak or write? (Like the volume of their voice in a town square).
Interiority (I): How much do we get inside their head to hear their thoughts and feelings? (Like reading their private diary).
Action (A): What physical things do they do? (Like building a house or running a race).
Discussion by Others (DC): How much do other characters talk about them? (This is the new big one. It's like how often people at a party whisper about someone who isn't even there).
Narrator Description (DN): How much does the storyteller describe them? (Like a narrator pausing to say, "Look at his sad eyes").

2. The "Gossip Map" (The New Network)

The researchers built a special kind of map (a graph) based on the "Discussion" part of the scorecard.

Old Maps: Previous maps showed who stood next to whom (co-occurrence) or who talked to whom (dialogue). These are like two-way streets.
The New Map: This map shows who is talking about whom. This is a one-way street.
- Analogy: Imagine a town where Alice never speaks, but everyone else is constantly talking about her. In an old map, she might look invisible. In this new "Gossip Map," she is the center of the universe because everyone is pointing at her.

3. The Experiment: Teaching Computers to Read Like Critics

The team tested two ways to get computers to fill out this six-part scorecard for 19th-century British novels (like Pride and Prejudice and Jane Eyre):

Method A (The "BookNLP" Pipeline): A specialized, rule-based tool designed specifically for literature. It's like a seasoned librarian who knows exactly where to look for specific types of sentences.
Method B (The "LLM" Approach): Using powerful AI chatbots (like GPT-4) to read the text and guess the counts.

The Result: The specialized librarian tool (BookNLP) was much more accurate and consistent than the AI chatbot. The chatbot often got confused, especially with minor characters or when the story was told from a first-person perspective (where the narrator is also a character). The researchers decided to use the "librarian" tool for their big study.

4. What They Discovered

Using this new scorecard and the "Gossip Map," they analyzed 64 classic novels and found some surprising things:

The "One vs. The Many" Theory: A famous theory suggests novels are built around one main hero and a crowd of minor characters fighting for space. The researchers found this is mostly true, but the "One" is often actually a small cluster of 2 or 3 people who share the spotlight, rather than just a single hero.
Being Talked About vs. Talking: In Pride and Prejudice, Mr. Darcy is the most "talked about" character (he has the highest "In-Degree" on the gossip map). However, Elizabeth Bennet is the one doing the most talking about others (highest "Out-Degree"). This shows that being the topic of conversation is different from being the speaker.
Gender and Gossip: They looked at who talks about whom based on gender. They found that women in these novels talk about other women more than men talk about other men. Also, women were more likely to be the ones discussing others, while men were often the ones being discussed. This revealed a hidden social layer that the old "who stood next to whom" maps completely missed.

Summary

The paper introduces a new way to measure character importance that goes beyond just "who shows up." By counting how much characters think, act, and—crucially—how much they are talked about by others, the researchers created a more nuanced map of a novel's social world. They proved that computers can do this, but they need the right tools (specialized software rather than just general AI) to get the details right. This allows us to see that in these classic stories, the "main character" is often a complex web of relationships, not just a single hero.

Technical Summary: Computational Representations of Character Significance in Novels

Problem Statement

Traditional computational approaches to character analysis in literature have largely relied on modeling characters based on their physical presence in scenes, primarily through co-occurrence or dialogue exchange. These methods often conflate "presence" with "significance," implicitly prioritizing the main character who appears most frequently. However, this interaction-centric view contradicts literary theories such as Alex Woloch's "the one vs. the many," which posits that narratives are structured around a central individual and a host of subordinate figures competing for limited narrative space. Furthermore, existing methods often neglect the distinction between the narrator and the character, and they fail to account for how characters are discussed by others—a component that reveals interest and social positioning without requiring physical presence.

The authors address the need for a more nuanced, multi-dimensional framework to quantify character significance that aligns with contemporary literary theory and captures relational dynamics beyond simple interaction.

Methodology

The Six-Component Structural Model

The core of the methodology is a structural model of character significance developed in collaboration with literary scholars (Marcus, 2025). This model decomposes character presence into six distinct components:

Name (N): Instances of the character's proper names (excluding pronouns).
Communication (C): Sentences of the character's quoted speech or writing.
Interiority (I): Verb phrases expressing the character's thoughts, feelings, intentions, or perceptions.
Action (A): Verb phrases conveying physical actions (excluding speech or interiority).
Discussion by other characters (DC): Sentences of communication from other characters in which the target character is discussed.
Description by the narrator (DN): Sentences in the narrative voice describing the character's appearance, manner, or dress.

Data and Annotation

The authors constructed a dataset by manually annotating all chapters of Pride and Prejudice (Jane Austen) and Jane Eyre (Charlotte Brontë) to establish ground truth counts for these six components at the chapter level. They note that Jane Eyre presents specific challenges due to its first-person narrator, where the narrator and protagonist are the same, complicating the distinction between interiority and narrator description.

To enable scaled analysis, the authors curated a corpus of 64 major 19th-century British realist novels (1800–1900), excluding first-person narratives where the narrator is a character, as preliminary results indicated poor performance on such texts.

Computational Approaches

The paper evaluates two primary pipelines for operationalizing this model:

BookNLP-based Pipeline: A transformer-based pipeline using entity, quote, supersense, event, and coreference packages. It parses text to identify spans meeting component criteria (e.g., using VerbNet classes for actions and interiority).
LLM-based Approaches:
- Chapter-level counts: Prompting LLMs (gpt-4o-mini) to generate JSON counts for all components and characters per chapter.
- Span-level annotations: Prompting LLMs to identify specific text spans for individual components (e.g., resolving pronouns in dialogue to determine "Discussion by other characters").

Network Construction

The authors construct three types of character networks:

Co-occurrence Network: Undirected edges based on paragraph-level co-presence.
Dialogue Network: Undirected edges based on spoken turn-taking (Component C).
Discussion Network: A novel directed graph where edges $u \to v$ represent character $u$ discussing character $v$ (Component DC). This allows for the distinction between "being talked about" (in-degree) and "talking about others" (out-degree).

Key Contributions

Theoretical Framing and Dataset: The paper introduces a six-component model of character significance to computational literary studies and releases an accompanying dataset of expert-annotated component scores for Pride and Prejudice and Jane Eyre.
Computational Pipeline: The authors develop and release a pipeline capable of labeling these components across 19th-century novels. They demonstrate that while LLMs can be used for labeling, the BookNLP-based pipeline offers superior reliability and lower error rates for scaled corpus analysis, particularly for components like Action and Interiority.
New Network Representation: The paper proposes the "Discussion Network" (DC), a directed graph that captures how characters are positioned relative to one another through discourse, offering a new lens on character centrality distinct from co-occurrence or dialogue.
Corpus-Level Analysis: The authors apply these representations to a corpus of 64 novels to answer specific literary questions regarding character centrality and gendered dynamics.

Results

Labeling Accuracy

BookNLP vs. LLM: The BookNLP-based pipeline achieved a Mean Absolute Error (MAE) of 1.7 and a Pearson correlation of 0.75 against gold counts for Pride and Prejudice. In contrast, LLM-based methods, even with granular prompting, exhibited higher MAE and inconsistent performance, particularly for minor characters and components like Action (A) and Interiority (I).
First-Person Narratives: The pipeline struggled with Jane Eyre due to coreference resolution issues between the narrator and protagonist. Without manual correction, the MAE was 4.35; with correction, it dropped to 3.00. Consequently, first-person narratives were excluded from the main corpus analysis.

Network Analysis and Literary Insights

Centrality Discrepancies: The study found that different network constructions yield different rankings of character prominence. For example, in Pride and Prejudice, Mr. Darcy ranks first in "in-strength" (being discussed), while Elizabeth ranks first in "out-strength" (discussing others). This highlights that "being talked about" and "doing the talking" are distinct dimensions of social prominence.
Woloch's "One vs. Many": Analysis of inequality metrics (Gini coefficient) across the corpus reveals that narrative attention is highly concentrated but not monolithic. The "one" is not a single individual but a centralized cluster of several characters, with multiple degrees of "many" constituting the periphery.
Gendered Dynamics: The Discussion Network revealed significant gendered structures not visible in undirected networks:
- Female characters are overrepresented among the most central discussants (out-degree) and the most discussed (in-degree).
- There is an asymmetric cross-gender attention pattern: edges from women to men occur significantly more frequently than from men to women ( $F \to M / M \to F = 1.26$ ).

Significance and Claims

The paper claims to offer a new computational lens for literary analysis that moves beyond simple interaction metrics. By integrating a component-based theory of character with network analysis, the authors demonstrate that:

Character Significance is Multi-faceted: A comprehensive view of a character requires accounting for interiority, description, and discussion, not just physical presence or dialogue.
Directed Graphs Reveal New Social Layers: The Discussion Network uncovers relational dynamics (who discusses whom) that are invisible to traditional undirected co-occurrence or dialogue networks.
Gender Shapes Narrative Space: The computational analysis provides empirical evidence for gendered dynamics in character discussion, showing that women are more likely to be the topic of discussion and to discuss others, a nuance missed by prior network studies.

The authors conclude that while their methods are currently limited to third-person 19th-century English novels due to annotation costs and coreference challenges in first-person narratives, this framework provides a robust foundation for future work in scaling literary theory to large corpora. They explicitly state that their goal is to facilitate future research rather than to provide a definitive, error-free solution for all narrative forms.