The structure-mapping engine is a mechanism intended to implement ideas pertaining to Gentner’s structure mapping theory of analogy. The basic principle of this theory is that analogy is concerned with systems of connected knowledge, rather than collections of independent attributes, so the amount of common higher-order relational structure is the criterion for ranking possible matches.

The authors discuss the SME algorithm, which builds all structurally consistent interpretations of comparisons between a base and a target, at length. For each of the algorithm’s four conceptual steps they provide theoretical background, examples, and complexity analysis. They even supply LISP code for the three types of rules (literal similarity, analogy, and mere appearance) that guide the construction of a match.

Aware that many programs are restricted to working on some carefully chosen examples, Falkenhainer, Forbus, and Gentner dedicate a section of their paper to demonstrating SME’s generality and flexibility. They also compare their work to other results in the field. They claim that SME is the only mechanism that generates all structurally consistent analogical mappings without heuristic search. They suggest that structural constraints are sufficient to counterbalance exponential explosion. As always, however, semantic problems cannot be evaded. Semantic implicit information has to be correctly encompassed in the graphs SME uses. This task is difficult, especially when the system has to take into account analogies that involve lots of attributes and their connections. How is this task fulfilled if, for example, one wants to find analogies between the reasons for the fall of great historical empires? And how will the engine perform in such a case?

The authors point out directions for future development. This comprehensive paper does not require any special background, and anyone concerned with learning by analogy or learning in general should read it.