Abstract This paper presents MACARTHUR, a computational model designed to simulate how human readers adjust their inferences based on extra-textual prompts. By utilizing a hierarchy of "explanation shapes," the system demonstrates how initial, parsimonious interpretations can be abandoned in favor of more complex, cohesive narratives when challenged by external questioning. The research highlights the directional nature of inference search and the role of context in shaping textual understanding.
Introduction
In cognitive psychology and early Artificial Intelligence, understanding how humans interpret text often focused on finding a single, logical meaning. However, Richard H. Granger, Jr.'s 1981 research suggests that interpretation is a dynamic process influenced by "extra-textual" factors. For instance, being asked "Are you sure?" can force a reader to re-evaluate a story, shifting from a simple explanation to a more complex one. This paper introduces MACARTHUR, a program designed to model this "inference pursuit."
The Problem of Static Interpretation
Early natural language processing (NLP) systems typically sought the most direct logical path to explain a text. Granger identifies a gap here: human readers are flexible. If a story describes an official making an unscheduled stop, a reader might initially assume it is for leisure. If prompted to reconsider, the reader might pivot to theories of a secret meeting or mechanical failure. MACARTHUR was built to bridge this gap, modeling how the purpose of reading changes the content of understanding.
The MACARTHUR Model: Explanation Shapes
MACARTHUR operates on the Parsimony Principle, a heuristic suggesting that the best interpretation accounts for the most actions with the fewest goal inferences. To manage this, the system uses Conceptual Dependency (CD) to represent underlying meanings and organizes them into four primary "explanation shapes."
These shapes serve as templates for understanding character motivations:
- Pursue-desired-state: Direct goal-seeking behavior.
- Avoid-undesired-state: Preventative actions.
- Accident-reaction: Responses to unexpected hindrances.
- Cover-stories: Deceptive explanations used to mask secret goals.
As shown in the table above, MACARTHUR maps specific actions (like PTRANS or INGEST) to these shapes. For example, a PTRANS (physical transfer) might be interpreted as "pursuing a function" in a Pursue-state, or a "secret meeting" in a Cover-story context.
Inference Pursuit and Re-probing
The core novelty of MACARTHUR is its "re-probing" logic. When the system is asked the same question twice, it interprets the repetition as a signal that its current parsimonious explanation is insufficient. This triggers a search for a new "shape" that might connect story events more cohesively, even if the new explanation is more complex.
This process was tested using "toy" stories:
- The Pakistani Ambassador Story: Demonstrates shifting from a "vacation" goal to a "secret meeting" (cover story).
- The Dr. Fitzsimmons Story: Shows a shift from a man seeking milk to sleep (Pursue-state) to seeking coffee to stay awake (Avoid-state) after re-probing.
- The Movie Theater Story: Illustrates how later information (e.g., drug deals) can override initial, simpler inferences (e.g., watching a movie).
Key Results and Insights
The research demonstrates that inference is directional, not just deep. "Deep" reading involves changing the direction of the search rather than just making more inferences. MACARTHUR successfully generated different English-language answers based on its current state of pursuit. In the Dr. Fitzsimmons example, the program's output evolved from a simple explanation of seeking rest to a more nuanced realization that the character might be trying to stay awake.
Limitations and Future Directions
Granger acknowledges that MACARTHUR faces a "knowledge bottleneck." To handle more complex extra-textual factors—such as distinguishing between a news report and a mystery novel—the system would require vast amounts of world knowledge. Future work suggests modeling how questions asked during the reading process affect interpretation and conducting psychological experiments to validate if MACARTHUR's "explanation shapes" truly align with human cognitive categories.
Summary updated