Introduction
Analogy making is a cornerstone of human intelligence, allowing us to apply knowledge from a familiar "source" domain to a novel "target" domain. However, computational models often face a "representation bottleneck": they typically require source and target domains to be described using nearly identical structures to recognize a mapping. In real-world scenarios, descriptions are rarely perfectly aligned. This paper addresses the re-representation problem—the challenge of automatically transforming domain descriptions until a common structural similarity emerges.
Theoretical Framework: Anti-Unification
The authors propose a formal framework based on anti-unification, a logical process that identifies the most specific generalization of two different terms. While standard anti-unification operates on individual terms, this work extends the concept to "theories" (sets of formulas).
Term and Theory Anti-Unification
In term anti-unification, specific instances like mass(sun) and mass(nucleus) are generalized into a common structure, such as mass(A). This process generates substitutions that map the general variable back to the specific domain constants (e.g., {A → sun}).
The authors define an anti-unifier of theories as a general set of formulas ($G$) that can be transformed into the source theory ($Ax_S$) and the target theory ($Ax_T$) through specific substitutions. To avoid trivial generalizations, the framework seeks the Least General Anti-Unifier (LGAU), ensuring the analogy remains as specific and informative as possible.
Re-Representation via Logical Inference
The core innovation of this approach is using background knowledge and logical inference to drive re-representation. If a direct match between axioms is not immediately apparent, the system uses inference to derive new formulas that might anti-unify more effectively. For example, if the system knows that distance(x, y) is symmetric, it can rewrite a target axiom to match the argument order of a source axiom, uncovering a hidden structural match.
Case Study: The Rutherford Analogy
The framework is demonstrated using the classic Rutherford analogy between the solar system and the atom. The system is provided with axiomatizations for both domains, covering concepts like mass, gravity, charge, and attraction.
Through this process, the system successfully maps sun to nucleus and planet to electron. More importantly, it demonstrates how to overcome conceptual mismatches. Even though gravity and electromagnetism are governed by different physical laws, the system can use background knowledge to derive intermediate formulas that anti-unify, effectively mapping the abstract concept of "central force" across domains.
Key Insights and Future Work
This research suggests that establishing an analogy is a process of discovery rather than simple matching. By looking at everything that can be inferred from axioms, a logic-based approach can find "hooks" for analogy that are invisible in raw data.
However, several challenges remain. Finding the optimal anti-unifier among sets of formulas is computationally expensive. Future work will focus on the trade-off between coverage (how much domain knowledge the analogy explains) and complexity (the simplicity of the required substitutions), as well as scaling the approach to larger, less-structured datasets.
Summary updated