Towards a Formal Model of Narratives

Louis Castricato, Stella Biderman, Rogelio E. Cardona-Rivera, David Thue

TL;DR

This paper introduces a formal, model-theoretic framework for narrative, bridging the gap between symbolic logic and data-driven AI. By treating story understanding as a process of belief dynamics—expansions, contractions, and revisions—the authors provide a mathematical language to measure story coherence, plot relevance, and the accuracy of information flow from narrator to reader.

What Problem Does This Paper Address?

Narrative AI currently exists in two separate worlds: data-driven approaches (like LLMs) that excel at pattern matching but struggle with deep causal reasoning, and symbolic approaches (like automated planning) that are logically rigorous but often too rigid.

The authors argue that the field lacks a shared foundational framework to reconcile these methods. Specifically, we lack a formal way to describe how a "Narrator" communicates a story world to a "Reader," and how that Reader’s mental model evolves over time. Without this, it is difficult to quantify subjective qualities like "coherence" or to identify which events are truly "relevant" to a plot.

Methods

The authors turn to Model Theory—a branch of mathematical logic that deals with the relationship between formal languages and their interpretations.

Key components of their approach include:

  • Story-World Models ($S$): A formal "world" consisting of a domain of objects (characters, places) and an interpretation function that determines which propositions (e.g., "Jay wears blue") are true.
  • The Fabula ($F$): Defined not just as a graph of events, but as a "distillation" of the story world—a set of explicit beliefs a reader holds that are essential for constructing the model.
  • Commutative Diagrams: Borrowed from category theory, these diagrams (see Figure 1) model the communication process. A story is "accurately conveyed" if the path from the Narrator’s mental model to the Reader’s mental model is consistent, meaning the Reader forms the beliefs the Narrator intended.
  • Weak Filters: Since first-order logic doesn't allow for a standard probability distribution over infinite sets of models, the authors use "filters" to represent a reader's uncertainty—capturing what a reader has "actively in mind" without requiring them to know every single detail of the world.
  • Information Entropy: The authors use Shannon entropy to define Plot Relevance. An event is relevant if knowing its outcome significantly reduces the reader's uncertainty about other parts of the story.
Figure 1: Commutative diagram outlining storytelling

Datasets

This is a theoretical foundational paper; it does not utilize empirical datasets or introduce new benchmarks. Instead, it provides the mathematical definitions and algorithms that future researchers can use to build such benchmarks.

Key Insights and Ideas

  • Narrative as Communication: The paper treats storytelling as a lossy compression process. The Narrator compresses a complex world ($S_N$) into a discourse/fabula ($F_N$), which the Reader then attempts to decompress into their own mental model ($S_R$).
  • Formalizing Literary Theory: The authors provide a mathematical basis for Roland Barthes’ concepts of Kernels (major plot points that shift beliefs) and Satellites (minor events that support kernels). A Kernel is formally defined as an event that causes a "drastic" change in the reader's set of plausible worlds.
  • Coherence as Agreement: Coherence isn't just a vibe; the authors define it through two new metrics:
  • Entropy of World Coherence (EWC): Measures how much the various "plausible worlds" in a reader's head agree on static facts.
  • Entropy of Transitional Coherence (ETC): Measures how well the reader understands the "laws" of the story world—i.e., how consistently they can predict how the world changes over time.

Key Results

The primary result is the formal framework itself, which yields explicit algorithms for:

  1. Measuring Accuracy: Determining if the essential character of a story was successfully transmitted from Narrator to Reader.
  2. Quantifying Relevance: Using entropy to calculate the relevance of Question B to Question A (e.g., "Does Harry being the Heir of Slytherin explain why he can speak Parseltongue?").
  3. Measuring Coherence: The EWC and ETC metrics provide a way to computationally score how "confusing" or "consistent" a narrative discourse is.

Limitations and Future Work

  • Discrete Time: The current model treats story progression as a series of discrete steps. The authors note that for media like film, a continuous time model might be more appropriate.
  • Author Intent: The framework relies on a definition of "author intent," which is notoriously difficult to pin down. The authors acknowledge there is no single best way to define this.
  • Implementation: While the math is laid out, the paper does not provide a software implementation. The authors suggest that the temporal evolution of these models could be a fruitful area for Reinforcement Learning research in natural language understanding.
  • Human Validation: The conjectures about how human readers respond to under-specification (using "ultraproducts" to reconcile beliefs) remain experimental questions for future cognitive science research.

Summary updated

Abstract

In this paper, we propose the beginnings of a formal framework for modeling narrative \textit{qua} narrative. Our framework affords the ability to discuss key qualities of stories and their communication, including the flow of information from a Narrator to a Reader, the evolution of a Reader's story model over time, and Reader uncertainty. We demonstrate its applicability to computational narratology by giving explicit algorithms for measuring the accuracy with which information was conveyed to the Reader and two novel measurements of story coherence.
BibTeX
@misc{castricato2021towards,
  author = {Louis Castricato and Stella Biderman and Rogelio E. Cardona-Rivera and David Thue},
  title = {{Towards a Formal Model of Narratives}},
  year = {2021},
  eprint = {2103.12872},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2103.12872},
}
Figure