💬 NLP

Revisiting Northrop Frye's Four Myths Theory with Large Language Models

This paper proposes a new character function framework that maps Jungian archetypes to Northrop Frye's four narrative genres, validating its effectiveness through a multi-model LLM study that demonstrates the models' ability to systematically recognize and distinguish genre-specific archetypal roles across diverse narrative works.

Original authors: Edirlei Soares de Lima, Marco A. Casanova, Antonio L. Furtado

Published 2026-02-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Edirlei Soares de Lima, Marco A. Casanova, Antonio L. Furtado

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 you are a chef trying to understand the secret recipes of four very different types of cuisine: Comedy (like a happy wedding feast), Romance (like an epic adventure quest), Tragedy (like a heartbreaking opera), and Satire (like a sharp political cartoon).

For decades, literary critics have argued about what makes these stories work. One famous chef, Northrop Frye, said that every story in these genres follows a specific "seasonal" pattern (Spring for comedy, Winter for satire, etc.). But he mostly looked at the plot—the sequence of events.

This paper asks a new question: What about the characters? If the plot is the recipe, who are the specific ingredients, and how do they change depending on the dish?

Here is the story of how the authors used modern AI to solve this puzzle.

1. The "Psychic" Blueprint

The authors realized that to understand the characters, they needed a universal map. They borrowed a map from Carl Jung, a famous psychologist who believed every human mind has four main parts:

The Ego: The "I" (The main character).
The Persona: The mask we wear (The wise guide).
The Shadow: The dark side we hide (The villain).
The Anima: The soul or emotional connection (The partner).

They decided that in a story, these four parts of the human mind get split up into four different characters:

The Protagonist (The Hero/Ego)
The Mentor (The Wise Guide/Persona)
The Antagonist (The Villain/Shadow)
The Companion (The Partner/Anima)

2. The "Shape-Shifting" Problem

Here is the tricky part: These four roles look very different depending on which "genre" (or season) you are in.

Think of it like a chameleon.

In a Comedy (Spring), the "Hero" is a clumsy young guy in love, and the "Villain" is a pompous, blocking old man.
In a Tragedy (Autumn), the "Hero" is a doomed adventurer, and the "Villain" is the force of order trying to fix the mess.
In a Satire (Winter), the "Hero" is a rebel who doesn't fit in, and the "Villain" is a giant, scary statue of a dictator.

The authors created a "Role Dictionary" with 16 specific names (4 roles × 4 genres). For example, in a Romance, the Companion is a "Faithful Victim," but in a Satire, the Companion is a "Rebellion Partner."

3. The AI Taste Test

Now, the big question: Does this dictionary actually work? Or is it just the authors' imagination?

To find out, they didn't ask human experts (who might argue forever). Instead, they asked six different super-smart AI models (like the brains behind modern chatbots) to play a game.

The Game: They gave the AI 40 famous stories (like Macbeth, 1984, Pride and Prejudice) and said, "Here is a character and a role. Does this fit?"
The Trick: They also gave the AI some fake matches (like saying the villain is actually the hero) to see if the AI would catch the mistake.

4. The Results: The AI Got It!

The results were surprisingly good.

The Score: The AI models got about 82.5% of the answers right. That's like getting an "A" on a very hard literature exam.
Agreement: Even though the AIs were built by different companies, they mostly agreed with each other. This proves that the "rules" the authors wrote down aren't just random guesses; they are real patterns hidden in stories.

However, the AIs struggled with specific genres:

Tragedy & Comedy: The AIs were great at these. The roles are very clear and structured (like a strict recipe).
Romance: The AIs were good at spotting the bad matches but sometimes missed the good ones. Why? Because in Romances, the "Helper" role is often shared between many people (a magical sword, a wise old man, a lucky charm), making it hard to pick just one character.
Satire: The AIs were confused here. Why? Because Satire is designed to be tricky! The "Villain" might also be the "Guide," and the "Hero" might be weak. The authors realized that the confusion wasn't a bug in the AI; it was a feature of the genre. Satire is supposed to blur the lines.

5. Why Does This Matter?

Think of this paper as building a GPS for storytellers.

For Writers: If you are writing a video game or a movie, you can use this map to make sure your characters feel "right" for the genre you are writing.
For AI: If we want computers to write good stories, they need to understand not just the plot, but who the characters are supposed to be. This paper gives the AI a rulebook for character behavior.
For Critics: It proves that ancient literary theories (like Frye's) can be tested and validated using modern technology.

The Bottom Line

The authors took an old, complex theory about stories, broke it down into simple character roles, and used AI to prove that these roles are real, structural parts of how we tell stories. They showed that while stories can be messy, the "skeleton" underneath follows a predictable, universal pattern—like the changing seasons.

1. Problem Statement

Northrop Frye's theory of four fundamental narrative genres (Comedy, Romance, Tragedy, Satire) has been a cornerstone of literary criticism, traditionally analyzed through narrative patterns and seasonal metaphors. However, existing computational approaches to Frye's framework have largely focused on narrative patterns (plot structures) while neglecting character functions.

The authors identify a gap in computational narratology: there is no formalized, computationally testable framework that maps Frye's genre-specific character archetypes to universal psychological functions. Furthermore, it remains unclear whether these theoretical character-role correspondences are systematic enough to be recognized by Artificial Intelligence across diverse narratives, or if they are merely subjective literary interpretations.

2. Methodology

The authors propose a novel Character Function Framework grounded in Jungian analytical psychology and validated using Large Language Models (LLMs).

A. Theoretical Framework Construction

Jungian Mapping: The authors map four universal character functions to Carl Jung's psychic structure components:
- Ego $\rightarrow$ Protagonist
- Persona (transformed into the Wise Old Man) $\rightarrow$ Mentor
- Shadow $\rightarrow$ Antagonist
- Anima $\rightarrow$ Companion
Genre Specialization: These four universal functions are specialized into 16 genre-specific roles based on four prototypical works (one per genre):
- Comedy: Lad in love, Troubleshooter, Pompous blocker, Witty damsel in love.
- Romance: Hero, Donor, Villain, Faithful victim.
- Tragedy: Ill-fated adventurer, Soothsayer, Order restorer, Ill-fated partner.
- Satire: Nonconformist, Inquisitor, Dystopian idol, Rebellion partner.

B. Experimental Validation Design

To test the validity of these correspondences beyond the prototypical works, the authors conducted a multi-model study:

Dataset Construction:
- Positive Samples: 40 diverse narrative works (10 per genre) spanning theater, literature, epics, and film. 160 valid character-role correspondences were generated via LLM-assisted identification and expert validation.
- Negative Samples: 30 invalid correspondences created by introducing two types of errors:
  1. Primary-Secondary Swap: Replacing a main functional character with a minor supporting character.
  2. Inter-role Swap: Swapping characters between different functional roles within the same story.
Evaluation Protocol:
- Models: Six state-of-the-art LLMs (Claude Opus 4.5, DeepSeek R1, Gemini 2.5 Pro, GPT OSS 120B, Llama 4, Qwen 3) were used as analytical instruments.
- Task: Models were asked to judge whether a specific character-role assignment was justified based on the character's structural function in the narrative.
- Metrics: Performance was measured using Recall (identifying valid roles), Specificity (rejecting invalid roles), and Balanced Accuracy (the mean of both). Inter-model agreement was measured using Fleiss' Kappa.

3. Key Contributions

Formalization of Character Functions: The paper bridges Frye's literary theory and Jungian psychology to create a formalized, 16-role character function framework suitable for computational analysis.
LLM-Based Validation of Literary Theory: It demonstrates that LLMs can act as reliable analytical instruments for literary criticism, validating that archetypal patterns are not just subjective human interpretations but recognizable structural patterns in data.
Granular Performance Analysis: The study provides a detailed breakdown of how different genres and specific roles affect computational recognition, revealing that structural ambiguity in certain genres (like Satire) correlates with lower model performance.
Dataset Creation: The creation of a curated dataset of 40 works with validated character-role mappings and 30 negative samples serves as a benchmark for future computational narratology research.

4. Results

The study yielded substantial evidence supporting the framework's validity:

Overall Performance: The six LLMs achieved a mean Balanced Accuracy of 82.5%, significantly exceeding the random baseline of 50%.
- Recall: 76.1% (Models recognized valid roles well).
- Specificity: 86.5% (Models were very good at rejecting invalid roles).
Inter-Model Agreement: There was strong consensus among the diverse models, with a Fleiss' Kappa of 0.600 (substantial agreement) and a mean pairwise agreement of 82.0%. This suggests the patterns are structural rather than model-specific artifacts.
Performance by Genre:
- Tragedy (89.9%) and Comedy (85.0%) showed the highest accuracy, likely due to their rigid, well-defined narrative arcs.
- Romance (78.8%) showed high specificity but lower recall, suggesting valid roles are sometimes distributed across multiple characters (functional distribution).
- Satire (77.1%) had the lowest performance, reflecting the genre's inherent ambiguity and the "disappearance of the heroic," where archetypal distinctions are deliberately blurred.
Performance by Role:
- Protagonist and Antagonist roles were the easiest to identify (high recall and specificity).
- Companion roles (e.g., "Rebellion partner" in Satire at 52.5%) were the most difficult, often due to the relational and thematic nature of these roles rather than plot mechanics.

5. Significance and Implications

Computational Narratology: The paper establishes a method for using LLMs to test and refine literary theories. It moves beyond simple pattern matching to deep structural analysis of character functions.
Narrative Generation: The 16 specialized roles provide explicit functional specifications that can guide LLM-based narrative generation and interactive storytelling. By assigning specific functional roles to agents, systems can generate stories that adhere to specific genre conventions (e.g., ensuring a "Troubleshooter" appears in a Comedy).
Theoretical Insight: The variation in model performance offers new insights into literary theory itself. The difficulty models had with Satire and Romance confirms that these genres rely on more fluid, ambiguous, or distributed character functions compared to the rigid structures of Tragedy and Comedy.
Future Directions: The authors suggest extending the framework to include gender-specific archetypes (addressing Jung's male-centric bias) and incorporating supplementary characters that intensify genre moods (e.g., buffoons in comedy, scapegoats in satire).

In conclusion, the paper successfully demonstrates that Northrop Frye's myth theory, when operationalized through Jungian archetypes, captures systematic structural patterns that are robustly identifiable by modern AI, providing a powerful foundation for the next generation of computational storytelling tools.