Abstract
This paper investigates the cognitive mechanisms underlying analogy and similarity through the lens of the Structure-Mapping Engine (SME), a computational simulation of human thought. It specifically addresses the "access-soundness paradox": the observation that while humans find deep, relational analogies most logically sound, their memories are more frequently triggered by shallow, surface-level similarities. By employing distinct rule sets to simulate "True Analogy" and "Mere Appearance" processing, the authors demonstrate that SME successfully replicates human psychological data, reinforcing Dedre Gentner’s Structure-Mapping Theory as a robust model of human cognition.
Introduction: The Access-Soundness Paradox
A fundamental challenge in cognitive science is understanding how we process similarity. Human studies reveal a curious split between accessibility—the ease with which one event reminds us of another—and soundness—the perceived quality or logical power of an analogy. For instance, a person might be reminded of a past event simply because it shares surface features (e.g., both involve a specific animal), yet when asked to evaluate the quality of the match, they will favor a deep structural analogy that shares underlying causal logic. This paper seeks to determine if a computer program, constrained by cognitive rules, can replicate this human behavioral pattern.
Methodology: The Structure-Mapping Engine (SME)
The researchers utilized the Structure-Mapping Engine (SME), a domain-independent tool designed to simulate analogical mapping. SME operates as a "tool-kit" where researchers can apply different rule sets to simulate various cognitive constraints:
- True Analogy (TA) Rules: These prioritize the Systematicity Principle, which suggests that humans prefer mapping systems of connected knowledge (higher-order relations) over isolated facts.
- Mere Appearance (MA) Rules: These focus on object descriptions and first-order relations, simulating surface-level attribute matching.
The simulation involved encoding five sets of stories (e.g., "Karla the Hawk") into predicate logic. These stories included "True Analogy" targets (shared relational structure) and "Mere Appearance" targets (shared object attributes).
Key Results and Empirical Corroboration
The simulation results closely mirrored human data, confirming that the rules governing memory retrieval differ significantly from those governing logical evaluation.
Soundness and Systematicity
When configured with "True Analogy" rules, SME's evaluation scores consistently aligned with human soundness ratings. For the "Karla the Hawk" story set, the structural match (a story about a country) received a high score of 22.36, while the surface match (a story about an eagle) scored only 16.82. This matches human preferences for deep relational alignment.
Accessibility and Surface Similarity
Conversely, when SME utilized "Mere Appearance" rules, the scores flipped, mirroring human recall patterns. In the same "Karla the Hawk" set, the surface-level eagle story scored 7.70, while the structural country match scored only 6.41. This suggests that memory "access" is a computationally "cheaper" process triggered by surface features, whereas "soundness" judgment requires more intensive structural alignment.
Predicting Human "Errors"
A notable finding occurred in Story Set 21 ("Acme, IRS"), where both humans and SME failed to prefer the "True Analogy." Analysis revealed that this story required a "many-to-one" object mapping, which violated the structural consistency rules of both the theory and the program. This failure highlights the sensitivity of human analogical power to internal knowledge representation.
Discussion and Conclusion
The success of SME in replicating these psychological phenomena provides strong evidence for the Systematicity Principle. It suggests that our preference for deep analogies is driven by a cognitive bias toward connected systems of knowledge. However, the "spectre of tailoring" remains a challenge. Future work will focus on developing universal structural evaluators and more flexible mapping algorithms to handle complex "many-to-one" scenarios. Ultimately, this research demonstrates that computational simulations can provide deep insights into the paradoxical nature of human memory and reasoning.
Summary updated