Introduction
While artificial intelligence has made significant strides in generating coherent sequences of events, creating narratives that evoke specific emotional responses remains a complex challenge. Most existing narrative generators focus on the "fabula"—the raw chronological set of events in a story world. However, the emotional impact of a story, particularly suspense, often depends on the "sjuzhet"—the specific selection and ordering of those events as presented to the reader.
This paper proposes Suspenser, an AI framework designed to algorithmically manipulate the level of suspense in a narrative. By treating the reader as a cognitive "problem-solver," Suspenser uses automated planning techniques to control the reader's perception of the protagonist's success, thereby modulating the narrative tension.
The Cognitive Theory of Suspense
The core of the Suspenser framework is the "Reader as a Planner" hypothesis. This theory suggests that as a story unfolds, readers are not passive consumers; they are active cognitive agents performing "plan recognition." They constantly evaluate the protagonist's situation and attempt to find potential paths to a successful outcome.
Suspense, in this model, is defined by the reader's perception of a protagonist's dwindling options. If a hero is trapped in a room with multiple exits, the suspense is low. If there is only one exit and the key is missing, the suspense peaks. Formally, the system defines the Suspense Level ($SL$) as being inversely proportional to the number of successful paths ($success$) the reader can identify toward the protagonist's goal:
$$SL \propto \frac{1}{\text{number of solutions}}$$
To account for human limitations, the model introduces a reasoning bound ($R$). This acknowledges that human readers do not possess infinite computational power; they only search for solutions within a specific cognitive limit. This bound allows the system to simulate a more realistic reader experience.
The Suspenser Architecture
The framework operates within a tripartite narrative model that separates the story's content from its presentation:
- The Fabula: The complete "story world" generated by an AI planner called Crossbow. This includes every possible action, causal link, and temporal constraint within the narrative universe.
- The Sjuzhet: The narrative structure. This is where the Suspenser module operates, selecting a specific subset of events from the fabula and ordering them to manipulate the reader's experience.
- The Discourse: The final medium of the story (e.g., natural language text or cinematic shots). While the discourse layer is the final output, this paper focuses primarily on the transition from fabula to sjuzhet.
Narrative Manipulation Techniques
Suspenser builds a story through a two-phase process designed to maximize or minimize suspense based on the desired output:
Phase 1: Skeleton Building
The system first identifies "core" events—the essential actions and causal links required for the story to remain coherent and logical. These form the backbone of the narrative.
Phase 2: Structure Building
Once the skeleton is established, Suspenser adds or hides information based on two primary strategies:
- Increasing Obstacles: The system can add actions to the sjuzhet that negate the protagonist's goal literals. By introducing new problems, the system reduces the number of viable solutions available to the protagonist.
- Information Control: Suspense is often achieved by withholding information. The system may defer telling the reader about helpful events or existing solutions. By keeping the "perceived" solution count low, the system maintains a state of high cognitive anxiety in the reader.
Theoretical Hypotheses
As a system-description and theoretical framework, the paper focuses on formalizing the mechanics of suspense rather than providing empirical performance metrics. The author proposes two central hypotheses for future validation:
- Hypothesis 1: The addition of actions that negate the protagonist's goal literals will significantly increase the calculated suspense level.
- Hypothesis 2: Removing or deferring the presentation of actions that assist the protagonist (unifying goal literals) will result in a higher suspense level compared to a straightforward chronological telling.
Limitations and Future Directions
The current model focuses exclusively on plot-suspense—the cognitive anxiety derived from problem-solving—rather than action-suspense, which is the visceral thrill of high-speed or physical conflict. Furthermore, the model does not yet account for dramatic irony, where the reader knows more than the protagonist, which is another significant source of narrative tension.
The most immediate limitation is the lack of human subject evaluation. The effectiveness of the Suspenser heuristic relies on the assumption that the mathematical model accurately mirrors human cognition. Future work involves conducting trials to compare Suspenser’s output against human-written stories and refining the reasoning bound ($R$) to better match the cognitive limits of human readers. Ultimately, the goal is to integrate this structural planning with natural language generation to produce fully realized, engaging, and suspenseful prose.
Summary updated