MEXICA: a computer model of a cognitive account of creative writing.

Abstract

📜 Abstract

MEXICA is a computer model that produces frameworks for short stories based on the engagement-reflection cognitive account of writing. During engagement MEXICA generates material guided by content and rhetorical constraints, avoiding the use of explicit goals or story-structure information. During reflection the system breaks impasses, evaluates the novelty and interestingness of the story in progress and verifies that coherence requirements are satisfied. In this way, MEXICA complements and extends those models of computerised storytelling based on traditional problem-solving techniques where explicit goals drive the generation of stories. This paper describes the engagement-reflection account of writing, the general characteristics of MEXICA and reports an evaluation of the program

Description

✨ Summary

The paper presents MEXICA, a computational model for generating frameworks of short stories about the Mexicas, the historical inhabitants of what is now Mexico City. The system is designed not merely as a practical story generator, but as an executable exploration of the engagement–reflection account of creative writing. The authors explicitly distinguish MEXICA from a model of human cognition: it is a tool for testing the adequacy and completeness of a cognitive theory of writing.

The proposed writing process alternates between two complementary activities. During engagement, the system produces a sequence of narrative actions using content and rhetorical constraints, without relying on explicit story goals or a predefined plot structure. During reflection, it reviews the developing story, resolves impasses, checks coherence, evaluates novelty and interestingness, and modifies the constraints that guide subsequent generation.

This design contrasts with planning-oriented story generators in which explicit goals and goal-directed problem solving drive plot construction. MEXICA instead treats constraints, memory retrieval, and iterative evaluation as the primary mechanisms for generating narrative material.

MEXICA learns its narrative knowledge from a user-supplied collection of previous stories. These stories are transformed into several kinds of representations: abstract structures connecting story-world contexts to possible subsequent actions, representations of changing dramatic tension, and concrete sequences of previously used actions. Users define the primitive story actions, their preconditions, post-conditions, and associated text using dedicated specification languages.

Each action can alter the story world through emotional relationships, dramatic tensions, and character locations. These consequences are stored in a separate story-world context for each character. Because characters may have different locations and experiences, the same event can be interpreted differently by different characters. The contexts serve as retrieval cues for long-term memory and thereby determine which actions can plausibly follow.

Action retrieval uses exact matching, inclusive matching, and a dynamic strategy that transforms the retrieval cue when direct matching fails. Candidate actions are instantiated with particular characters, allowing the system to reuse existing characters or introduce new ones while preserving narrative continuity. Filters remove actions that violate general constraints or current guidelines.

Coherence is handled primarily during reflection. Each action has preconditions, such as a character needing to be injured before another character can heal them. When preconditions are not satisfied, MEXICA inserts additional events to establish the missing circumstances, including events that bring characters to the same location.

Novelty is assessed by comparing the current story with the previous stories. The system classifies the developing story as adequate, similar to a previous story, or a copy, and correspondingly applies normal, high, or strict novelty guidelines. These guidelines influence which actions remain eligible for later generation.

Interestingness is modeled through the pattern of dramatic tension. MEXICA treats stories as interesting when they contain degradation–improvement processes: tension rises as obstacles or dangers develop, reaches a climax, and decreases as conflicts are resolved. Reflection compares the current tension pattern with patterns found in the previous stories and can impose guidelines to increase, maintain, or decrease tension.

The authors evaluated four MEXICA configurations through an Internet questionnaire completed by 50 participants from 12 countries. Participants rated narrative flow and coherence, narrative structure, content, suspense, and overall quality on a five-point scale. The configuration combining engagement, reflection, filtering, and dynamically updated guidelines received the highest ratings in every category. It also received the highest suspense rating jointly with the human-generated comparison story. Versions in which tension was deliberately kept low received substantially lower suspense and overall-quality ratings.

The evaluation also compared MEXICA with MINSTREL, GESTER, and a human-generated story framework. The results suggested that reflection improved narrative structure and coherence, while the tension model contributed to suspense. A separate comparison of themes indicated that MEXICA could generate stories with themes distinct from those in its training collection, although novelty was defined primarily as content difference from the supplied previous stories.

The paper’s central contribution is a concrete architecture showing how constraint-driven generation can be combined with reflective revision and evaluation. It demonstrates that a story generator need not be organized entirely around explicit goals: retrieval from structured memory, character-specific world contexts, dramatic tension, and iterative filtering can collectively produce coherent and comparatively interesting narrative frameworks.

The authors also identify significant limitations. MEXICA generates story frameworks rather than fully realized natural-language narratives; its knowledge of the world, audience, environment, and emotion is highly restricted; and its evaluation depends on hand-authored actions, prior stories, parameters, and predefined notions of tension and novelty. The system therefore demonstrates the plausibility of the engagement–reflection approach within a constrained domain rather than establishing a general theory of creativity.

Subsequent work directly extended MEXICA’s engagement–reflection architecture. MEXICA-impro used multiple MEXICA-like agents for collaborative narrative generation, with agents alternately advancing a shared plot using separate knowledge bases and social beliefs. (sciencedirect.com) The authors later described MEXICA as part of a broader research program that included collaborative narrative generation, social norms in stories, and the SLANT system, which integrates MEXICA with other computational narrative components. (rafaelperezyperez.com)

MEXICA has also been incorporated into later academic treatments of narrative generation and computational creativity, including a 2023 book chapter by Pérez y Pérez and Sharples that revisits the model and discusses its collaborative and self-evaluative capabilities. (academic.oup.com) It is cited in subsequent computational-creativity literature as an example of author-centered story generation in which reflection is interwoven with narrative production rather than limited to initial planning. (robsaunders.net)