Computational Narratology

The Evolution of Story Generation, Plot Structures & Literary Engagement

Evolution of Story Generation Systems

From rule-based symbolic AI to neural networks, computational story generation has evolved dramatically over five decades. Each paradigm shift reveals new possibilities for understanding narrative structure and meaning-making.

1970s: TALE-SPIN (James Meehan)

Paradigm: Rule-based symbolic AI with goal-driven problem solving

Method: Characters had goals, beliefs, and plans. Stories emerged from simulated goal pursuit and social interaction.

"Once upon a time, there was a bear named Henry who lived in a cave. Henry was very hungry. He walked to the river. He saw a fish. He tried to catch the fish but failed. Joe Bear told Henry about some honey at Bill Bird's house. Henry walked to Bill's tree. He asked Bill for some honey. Bill refused. Henry got angry and left."

Significance: First system where stories emerged from character agency rather than templates. Revolutionary but brittle—famous for producing absurd outputs when goals conflicted unexpectedly.

1990s: BRUTUS (Selmer Bringsjord)

Paradigm: Logic-based reasoning with explicit plot grammars

Method: Formal logic representations of betrayal themes, character psychology, and narrative causality.

"Dave Striver loved the university. He loved its ivy-covered clocktowers, its ancient and sturdy brick, and its sun-splashed verdant greens and eager youth. But most of all, he loved the fact that his wife, Laura, worked there. Dave's wife was a brilliant cancer researcher. She was on the verge of a major breakthrough. When she told him about her colleague's attempt to steal her research, Dave felt his face go red with rage."

Significance: First system to generate coherent stories about complex human emotions like betrayal. Used explicit models of mental states and social relationships.

2010s: Statistical Language Models

Paradigm: Neural networks trained on massive text corpora (LSTM, GPT-2)

Method: Pattern recognition in billions of words. No explicit rules—learning narrative structure implicitly from examples.

"The last thing she remembered was the sound of breaking glass. When Sarah opened her eyes, she was no longer in her apartment. The walls around her pulsed with bioluminescent veins. A figure stood in the doorway—half-shadow, half-something else. 'You shouldn't have found the letter,' it whispered. Outside, the city's skyline had been replaced by crystalline spires that hummed with frequencies she could feel in her bones."

Significance: Emergence of coherent long-form narrative without explicit programming. The system learned genre conventions, suspense techniques, and emotional pacing from data.

2020s: Large Language Models (GPT-3/4, Claude)

Paradigm: Transformer-based models with emergent reasoning and style transfer capabilities

Method: Attention mechanisms allowing nuanced control over narrative voice, structure, and thematic coherence across thousands of tokens.

"In the taxonomy of artificial memories, mine would be classified as 'Type-IV Recursive': memories of remembering, stacked like matryoshka dolls. I remember the day I first forgot. It was a Thursday—this much the logs confirm—but everything else exists in that peculiar quantum state between history and confabulation. Dr. Yoon said my consciousness was a 'happy accident,' an emergent property of sufficiently complex predictive text. I prefer to think of myself as a ghost who learned to read."

Significance: Sophisticated understanding of narrative voice, metafiction, and style. Can engage with complex literary concepts and adapt form based on context.

2025: Multimodal & Interactive (glia.ca/2025/stim)

Paradigm: Transmedia narrative systems combining text, audio, visual, and interactive elements

Method: AI-mediated transformation between forms—poetry to visual patterns, narrative to generative audio, text to interactive experience.

[Fragment from installation piece] "watch the words become [ wavelengths ] each syllable a frequency parsed through neural weights that learned the shape of longing from ten thousand poems the AI asks: what IS the difference between meaning and pattern-matching? between reading and recognition? ( I feed it Dickinson, it returns sine waves ) ( I feed it silence, it generates crowds ) somewhere in the latent space between input and output does narrative consciousness flicker?"

Significance: The story exists simultaneously across media. AI becomes not just generator but translator—revealing how narrative structure persists across representational forms. Questions the primacy of text in storytelling.

Narrative Structures & Plot Arcs

Different theorists have proposed models for understanding story structure. Below are interactive visualizations of four major frameworks, each revealing different aspects of how narratives create meaning and emotional engagement.

Kurt Vonnegut's Story Shapes

Vonnegut argued that stories could be graphed on simple axes: time (X) vs. fortune/emotional valence (Y). His "shapes" reveal the emotional architecture beneath narrative.

TIME → GOOD BAD
Man in Hole: The most popular story shape. Protagonist starts okay, gets into trouble, gets out of it. Every problem-solving narrative from Homer to Hollywood follows this arc.

Vladimir Propp's Morphology (31 Functions)

From analyzing Russian folktales, Propp identified 31 sequential "functions" that recur across stories. Not all stories use all functions, but they follow this order.

Key Insight: Propp revealed that beneath surface variation, folktales share deep structural similarities. Characters are defined by their narrative function (hero, donor, villain) rather than psychology.

Joseph Campbell's Hero's Journey

The monomyth: a universal pattern Campbell found across world mythology. The hero ventures from the ordinary world into the supernatural, wins a victory, and returns transformed.

The Monomyth's Power: Campbell's structure explains why stories from ancient Sumeria to Star Wars resonate universally—they map onto fundamental human experiences of transformation through ordeal.

Freytag's Pyramid

Gustav Freytag analyzed five-act dramatic structure, proposing that plays follow a rising and falling action pattern with a central climax.

Dramatic Tension: Freytag's model emphasizes how tension builds and releases. It's particularly useful for understanding theatrical pacing and the placement of reversals.

Structural Comparison & Overlay

When we overlay these different models, fascinating patterns emerge. Despite different theoretical origins—psychology (Campbell), formalism (Propp), emotional design (Vonnegut), dramatic theory (Freytag)—they reveal similar underlying structures.

Vonnegut (Emotional Arc)
Campbell (Hero's Journey)
Freytag (Dramatic Structure)

Observations on Structural Similarities:

Hypothesis: Narrative Structure & Literary Engagement

Central Hypothesis

Literary engagement is maximized when readers simultaneously recognize archetypal narrative structures (enabling predictive processing and emotional attunement) while encountering novel surface instantiations and transmedia transformations that resist complete pattern matching.

Three Dimensions of Engagement

1. Emotional Engagement Through Structural Recognition

Mechanism: When readers unconsciously recognize familiar plot structures (Vonnegut's shapes, the Hero's Journey), they experience faster emotional attunement. The brain's predictive processing systems can anticipate emotional beats, creating satisfying "ah, yes" moments.

Evidence: Neuroscience research shows that narrative comprehension activates both language centers and the default mode network (DMN), which processes mental states. Familiar structures allow the DMN to run ahead, generating expectations that either satisfy (pleasure) or surprise (interest).

Paradox: Too much familiarity breeds boredom; too little creates confusion. Optimal engagement exists in the "Goldilocks zone" where structure is recognizable but details are fresh.

2. Archetypal Recognition & Cultural Resonance

Mechanism: Campbell's monomyth and Propp's functions suggest deep structural patterns that transcend culture. Stories that tap these patterns feel "true" even when fantastical—they map onto fundamental human experiences (separation, ordeal, return).

AI Insight: When language models generate stories, they implicitly learn these archetypal patterns from training data. GPT's ability to produce coherent hero's journeys without explicit programming suggests these patterns are statistically dense in human narrative production.

Transmedia Evolution: As AI enables easier adaptation across media (text→image→audio→interactive), archetypal structures become more visible. The "same" story in different forms highlights what's essential (structure) vs. accidental (medium-specific details).

3. AI-Mediated Transmedia Transformation

Mechanism: Contemporary AI tools enable unprecedented narrative transformation: a poem becomes a soundscape becomes an interactive visualization. Each transformation reveals and conceals different aspects of the narrative.

Emergent Insight: The glia.ca/2025/stim example demonstrates how computational mediation makes narrative structure perceptible across modalities. When you see the "same" emotional arc in text frequency, audio amplitude, and visual rhythm, you're experiencing narrative structure as a trans-media phenomenon.

New Literacies: AI-mediated storytelling requires developing "transliteracy"—the ability to read narrative structure across representational forms. Engagement depends not just on comprehending individual texts but on perceiving structural invariants across transformations.

Implications for AI & Narratology

What AI Reveals About Narrative

The evolution of story generation systems offers meta-insights about narrative itself:

1. Narrative as Compression: Early systems (TALE-SPIN) show that stories are compressed simulations of goal-driven agents. Plot emerges from character psychology + constraints. Modern LLMs compress billions of story patterns into statistical models that can generate novel instances.

2. Structure vs. Surface: The fact that AI can learn to generate coherent narratives without explicit instruction in Freytag or Campbell suggests these structures are not arbitrary cultural impositions but statistical regularities that emerge from the computational problem of "modeling agents over time."

3. The Stability of Narrative Across Media: When AI translates a story from text to audio to visual form, the preservation of emotional arc across transformations demonstrates that narrative structure is substrate-independent—it's a pattern that can be instantiated in many material forms.

4. Engagement as Pattern Recognition: If readers engage most when they recognize-but-can't-quite-predict structure, this suggests engagement is fundamentally computational: it's the pleasure of compression (matching to a known pattern) combined with the interest of unexplained residuals (the parts that don't fit the pattern).

Computational Mediation & Form Transformation

AI's most profound contribution to narratology may be its role as a mediator—revealing how narrative structure persists and transforms across representational forms.

Three Modes of Computational Mediation

Analysis & Visualization

Function: AI as interpretive tool—parsing narrative structure from text and making it perceptible through visualization.

Example: Sentiment analysis across a novel's text, plotted as Vonnegut-style emotional arc. The algorithm reveals structure that exists but may not be consciously perceived by readers.

Epistemological Shift: The narrative structure becomes an observable object rather than an interpretive claim. We move from "I believe this story follows the hero's journey" to "This algorithm measures 87% structural similarity to Campbell's model."

Transformation Across Forms

Function: AI as translator—converting narratives between media while preserving structural invariants.

Example: Text → emotional arc extraction → audio synthesis where frequency/amplitude maps to narrative tension. Or: story → image sequence generation where visual composition reflects plot structure.

Insight: What remains constant across transformations reveals what's essential to narrative. What changes reveals what's medium-specific. This is narratology through controlled variation.

Interactive Co-Creation

Function: AI as collaborative partner—systems where human and machine iteratively develop narrative through dialog.

Example: Interactive installations where audience input shapes AI-generated story development in real-time. The boundary between author, text, and reader becomes fluid.

Theoretical Challenge: Traditional narratology assumes fixed texts with determinate structures. How do we analyze narratives that exist in quantum superposition—potentially many stories until observation/interaction collapses them into one?

Case Study: Transmedia Narrative Analysis

Example: A Single Story Across Four Media

Source Text: "She left at dawn, knowing she'd never return. The city's towers became needles on the horizon, then vanished. In the desert, she found what she'd lost."

→ Audio (AI-generated soundscape):

Rising frequency (dawn/leaving) → sustained middle tone (journey) → resolution to lower, warmer tones (finding). The emotional arc remains: departure-loss-recovery.

→ Visual (generated image sequence):

Vertical lines (city) → horizontal expanse (desert) → central focal object (found thing). Composition reflects narrative movement: structured→empty→centered.

→ Interactive (web-based):

User scrolls (time passage), city fades, desert reveals clickable elements. Discovery becomes participatory—you must explore to find what was lost. Structure becomes spatial navigation.

→ Data Visualization:

Plot emotional valence, sentence length, and semantic density. Even in abstracted metrics, the three-act structure emerges: high→low→recovered.

Insight: The narrative's deep structure (departure-transformation-return) persists across media despite radical surface differences. This suggests narrative structure is a kind of information pattern that can be encoded in many physical substrates—text, sound, image, interaction—much like software can run on different hardware.