Abstract
In this paper, we propose a formal framework for modeling narrative 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, along with two novel measurements of story coherence.
Introduction
In the field of Artificial Intelligence, "story understanding" is often treated as a benchmark for language models. However, a deep divide exists between data-driven approaches (like Large Language Models) and symbolic approaches (like logic-based planning). While LLMs excel at generating plausible text, they often lack the formal consistency required for complex narrative reasoning. This paper proposes a bridge: a formal mathematical framework for narrative based on Model Theory. By treating a story as a communicative act between a Narrator and a Reader, we provide a rigorous way to measure information flow, belief evolution, and plot relevance.
A Model-Theoretic Framework for Narrative
The core of this framework rests on three pillars: Model Theory, Commutative Diagrams, and Information Entropy.
1. Model Theory and Possible Worlds
In logic, a "model" is a world where a specific set of sentences is true. The authors use this to represent the "story world." Because a story never provides exhaustive detail, the Reader's mind maintains a set of many possible models (possible worlds) that are all consistent with the text provided so far.
2. The Communication Diagram
To visualize the transmission of information, the framework employs commutative diagrams from category theory. As illustrated in Figure 1, the process involves a Narrator compressing their internal mental world ($S_N$) into a formal fabula ($F_N$). This fabula is conveyed to the Reader ($F_R$), who then reconstructs their own mental model ($S_R$).
3. Quantifying Coherence and Uncertainty
To handle the inherent uncertainty in reading, the authors use binary entropy to define novel metrics:
- Entropy of World Coherence (EWC): This measures the degree to which the various "plausible" story worlds in a reader's head agree on the implications of the story.
- Entropy of Transitional Coherence (ETC): This measures how well the reader understands the "laws" or dynamics of the story world over time—for instance, understanding if a character's death is permanent based on the established rules of the setting.
Key Insights: Kernels and Relevance
A significant contribution of this work is the formalization of literary concepts like "Kernels" (major plot points). Within this framework, a Kernel is defined as a narrative event that causes a "drastic" shift in the reader's beliefs, effectively collapsing many possible worlds into a few.
Furthermore, the authors argue that Relevance is Truth-Dependent. The relevance of a narrative question depends on the actual truth of the story world. For example, if a character's ability is derived from their heritage, then that heritage is relevant; if it is an accident, the heritage is irrelevant, regardless of what the reader might initially suspect.
Conclusion and Future Work
This preliminary framework provides a set of formal definitions and algorithms to quantify narrative accuracy and coherence. While currently limited to discrete time steps and requiring further integration with probabilistic logic, it offers a path toward AI that understands stories as communicative acts. Future work aims to connect this framework to Reinforcement Learning, where agents might learn language by minimizing the entropy of their internal story models.
Summary updated