What's in a Mental Model? On Conceptual Models in Reasoning with Spatial Descriptions

Göran HagertIJCAI 1985

Introduction

The nature of human reasoning with spatial descriptions—such as "the fork is to the left of the spoon"—is a central debate in cognitive science. A prominent theory, championed by Philip Johnson-Laird, posits that humans construct "mental models": internal, image-like structural representations. This view is largely supported by the observation that people process "determinate" descriptions (those defining a single layout) differently than "indeterminate" ones (ambiguous descriptions allowing multiple layouts).

This paper challenges the necessity of a unique, image-based representational format. Instead, it proposes a computational framework where these reasoning patterns emerge from abstract logic and flexible processing strategies. The "mental model phenomenon" is framed not as a change in data storage, but as a consequence of processing effort and strategy selection.

Methodology and Computational Framework

The author develops a theory grounded in Logic Programming, distinguishing three hierarchical levels of reasoning:

  1. Domain Level: Fundamental spatial facts and axioms (e.g., transitivity of "left-of").
  2. Strategy Level: The mechanisms for searching and combining these facts.
  3. Planning Level: The meta-cognitive process of selecting appropriate strategies for a given task.
Logical rules defining spatial relations and transitivity within the proposed computational framework.

To validate this model, a behavioral experiment was conducted with 15 subjects using 12 spatial reasoning tasks adapted from Mani and Johnson-Laird (1982). These tasks included both determinate examples (e.g., "The spoon is on the left of the knife. The plate is on the right of the knife") and indeterminate ones (e.g., "The spoon is on the left of the knife. The plate is on the right of the spoon"). Subjects were tasked with either drawing the resulting scene from memory or evaluating descriptions based on their resemblance to the original text.

Key Insights: Representation vs. Process

The core thesis is that the perceived difference between a "mental image" and a "logical statement" may be an artifact of processing effort. In indeterminate cases, the logical model fails to integrate disparate fragments into a single, unified structure, resulting in a "partial composite representation."

While indeterminate tasks do not necessarily take longer to read, they are significantly harder to manipulate. The reasoner must maintain multiple partial structures, increasing cognitive load. Crucially, Hagert demonstrates that a propositional system can still exhibit "structural similarity" to the physical world through its conceptual linkages, achieving the functional benefits of a mental model without requiring an internal "image."

Experimental Results

The behavioral data supported the computational model's predictions:

  • Processing Time: There was no significant difference in reading times between determinate and indeterminate descriptions (~105 seconds). However, drawing indeterminate scenes took significantly longer (57.3s) than determinate ones (41.1s).
  • Handling Ambiguity: In indeterminate drawing tasks, 80% of correct responses (12/15) involved subjects explicitly mapping the ambiguity (e.g., placing two objects in the same relative position to a third) rather than arbitrarily choosing one possible layout.
An example of a spatial layout produced during the reasoning tasks, illustrating the structural relationships between objects.

The computational model successfully replicated this "partial representation" phenomenon, suggesting that a single, abstract logical form is sufficient to explain the data previously used to argue for specialized mental models.

Conclusion and Future Directions

The study demonstrates that the "mental model" effect can be explained through logic-based processing strategies. However, limitations remain, including the small sample size (n=15) and the need for a more detailed "Planning Level" to explain how humans dynamically switch reasoning strategies. Future research should explore how these conceptual models are modified during active reasoning and how these multi-level processes apply to non-spatial, abstract domains.

Summary updated

BibTeX
@inproceedings{hagert1985what,
  author = {Göran Hagert},
  title = {{What's in a Mental Model? On Conceptual Models in Reasoning with Spatial Descriptions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1985},
}
Figure