A Theory of Reading

Kenneth Moorman, Ashwin RamAAAI 1994

Abstract

Traditional models of reading comprehension often focus on the processing of standard, everyday scenarios, failing to account for "creative reading"—the cognitive ability to understand texts containing novel concepts, such as those found in science fiction. This paper proposes a comprehensive functional theory of creative reading, implemented through the ISAAC (Integrated Story Analysis and Creative Understanding) system. By decomposing the reading process into a hierarchy of "supertasks," the authors provide a framework for how readers navigate complex narrative demands, manage the suspension of disbelief, and construct mental models of non-existent technologies and alien motivations.

Introduction

While the study of reading has spanned decades, existing theories frequently struggle to explain how humans process information that contradicts their established world model. Most psycholinguistic models are optimized for standard comprehension; however, nearly all reading is inherently creative. We are constantly required to integrate new information and adapt our internal schemas.

This paper addresses the gap between standard comprehension and the processing of novel concepts—such as sentient robots or faster-than-light travel. The authors argue that a theory of reading is fundamentally incomplete if it cannot account for these creative leaps. To bridge this gap, they present a functional model that integrates insights from AI, psycholinguistics, and education to simulate the entire reading process.

The ISAAC Architecture: A Functional Approach

The core of this theory lies in functional analysis, breaking reading down into a series of coordinated "Reading Supertasks." These tasks are not isolated algorithms but a symphony of interacting cognitive processes managed by a central architecture called ISAAC.

The Hierarchy of Supertasks

  1. Metacontrol: Acting as the system's manager, this layer regulates the depth of reading and time management. Crucially, it handles the suspension of disbelief, allowing the system to temporarily accept text that violates its internal world model rather than rejecting it as an error.
  2. Sentence Processing: This involves low-level linguistic functions, including syntactic parsing and pronoun resolution, which serve as the foundation for higher-level understanding.
  3. Story Structure Understanding: This task identifies narrative elements such as plot, genre, and setting, providing the macro-context for the story.
  4. Scenario Understander: This component builds the "who, what, and where" of the narrative. It includes an agent modeler to track character motivations and a device modeler to dynamically conceptualize how novel objects (e.g., a "positronic brain") function based on context.
  5. Explanation and Reasoning: The engine of creative reading, this task manages inferences, belief updates, and "metareasoning"—the system's ability to reflect on its own reading progress and strategies.
  6. Memory Management: This layer handles the storage and retrieval of information across all other tasks, ensuring consistency throughout the narrative.

Creative Reading as a Computational Task

A significant contribution of this theory is the formalization of the "suspension of disbelief." In traditional AI systems, a logical contradiction typically triggers an error or a correction. In the ISAAC framework, however, the system is designed to "accept" contradictions within the context of the story world. This allows the system to build a temporary, alternative model of reality.

Furthermore, the Device Modeler allows the system to move beyond static databases of real-world objects. When encountering a science fiction trope, ISAAC does not fail due to a lack of prior knowledge; instead, it uses the text's context to construct a functional model of the new technology. This suggests that creative reading is not a specialized mode for fiction, but a necessary component of all reading, as readers are always encountering novel ways of describing the world.

Evaluation and Results

The ISAAC system was evaluated using the domain of science fiction, specifically the short story "Men Are Different" by Alan Bloch (1963). This text was selected for its high density of novel concepts, including sentient robots, interstellar travel, and the extinction of humanity.

Qualitative Success

ISAAC successfully demonstrated the ability to "understand" the narrative by navigating its complex conceptual requirements. The system utilized its supertask architecture to build coherent models of the robot protagonist and its motivations, showing that high-level cognitive abilities—previously treated as "black boxes"—can be modeled through the integration of psycholinguistic theory and AI reasoning.

Limitations and Future Directions

While the ISAAC system provides a robust proof of concept, the authors acknowledge several areas for further development:

  • Breadth of Content: Initial testing was limited to a small selection of stories. Future work must expand the dataset to ensure the theory's robustness across diverse genres and narrative styles.
  • Evaluation Metrics: Measuring "creativity" and "comprehension" in a computational model remains a challenge. The authors identify the need for more sophisticated metrics to evaluate creative performance.
  • Knowledge Scaling: The system's success is dependent on its underlying knowledge base. Scaling this to handle arbitrary pieces of fiction without manual knowledge engineering remains a significant hurdle.

Conclusion

The theory presented here moves beyond simple text parsing to offer a functional model of how the mind handles the unknown. By treating reading as a multi-task coordination problem and formalizing the suspension of disbelief, the ISAAC system provides a new lens through which to view both human and artificial intelligence in the realm of language understanding.

Summary updated

BibTeX
@inproceedings{moorman1994theory,
  author = {Kenneth Moorman and Ashwin Ram},
  title = {{A Theory of Reading}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}

Loading related content...

Figure