The developmental psychology literature has necessarily used simple tasks to study analogical reasoning ability in children. This research has demonstrated that children can complete simple A:B::C:D analogies when the relations are simple and well known (e.g., Gentner & Toupin Reference Gentner and Toupin1986; Goswami & Brown Reference Goswami and Brown1989). An unfortunate side-effect of these studies is that they focused research primarily on factors that influence children's abilities to form correspondences between situations on the basis of relational similarities.
The ability to form relational correspondences is crucial to analogical reasoning, but it is only one subcomponent of the process. Analogies are central to cognitive processing because they allow people to extend their knowledge of one domain by virtue of its similarity to another domain (Clement & Gentner Reference Clement and Gentner1991; Markman Reference Markman1997). This extension of knowledge is accomplished via analogical inferences.
Analogical inference occurs when people take facts about a base domain that are connected to the match between the base and target and posit that those facts are also true of the target domain. Although analogical inference has not been studied extensively in development, it is clear that children draw analogical inferences frequently. For example, in their seminal studies of children's mental models, Vosniadou and Brewer (Reference Vosniadou and Brewer1992) found that children's beliefs about the earth were strongly influenced by simple analogues. For example, some children knew that the Earth is round, but believed it to be round like a pancake, and so they posited a flat round Earth with people living on the top. Other children knew that the Earth was round like a ball, but assumed that the people lived inside the ball with stars painted on the top. In each case, children were using elements of a known base domain (e.g., pancakes and balls) and transferred knowledge from that base to the less well-known domain of the solar system.
The relational priming model is too limited to account for analogical inference. Obviously, as the model stands, it has no mechanisms for making inferences. More importantly, it is not obvious how such mechanisms could be added in a way that would respect what is known about the inferences people make.
It is crucial that analogical inferences are constrained in some way because this prevents analogies from positing that every fact that is true about the base is also true of the target. (Thus, while a child might believe that the Earth is flat like a pancake, that child is unlikely to think that the Earth would taste good with syrup.) In their WWII–Gulf War analogy simulation, the authors coded only the facts that were relevant, and this was critical to the model's success. The simulation also appears to have benefited from some external control structure that always suppressed the appropriate layers and interleaved the appropriate control process at just the right time to ensure that each of the necessary relations was picked out in turn. How else could it be that the model never cycled through the same relation twice, or searched for a nonexistent element and became stuck?
In contrast, structural alignment assumes that inferences involve facts from the base that are connected to matching higher-order relations between base and target (Clement & Gentner Reference Clement and Gentner1991). This systematic relational structure and the preference for systematicity thus provide constraints on inferences such that structural accounts can function even with rich natural concepts and without any external direction. In addition, the inferences can easily be incorporated into the representation of the target domain.
The authors of the target article try to head off criticisms of this variety by suggesting that explicit mappings (and presumably inferences) could be carried out by different processes than the more implicit processes that find correspondences between domains. The authors use the example of semantic priming in language to illustrate their point. If their suggestion turns out to be correct, then it is those processes that could form the basis for a new theory of analogy. Therefore, the theory posited by the authors may help us to understand some of the sub-processes that are recruited during analogical processing, but it is not actually a theory of analogical processing itself. Indeed, it is worth noting that semantic priming is not taken to be a theory of language; rather, it is understood to be a sub-process that is used in language.
If there were no computational models of analogical reasoning that encompassed both mapping and inference processes, and if those models had never been applied to both developmental and adult data, then it might be reasonable to divide these processes into separate components and assume that two distinct models are required to account for them. However, models like the Structure-Mapping Engine (SME) (Falkenhainer et al. Reference Falkenhainer, Forbus and Gentner1989) and Learning and Inference with Schemas and Analogies (LISA) (Hummel & Holyoak Reference Hummel and Holyoak1997) are designed to account for both analogical mapping and inference, and both models are able to make use of higher-order relations in their domain representations. Furthermore, as the target article notes, SME has been applied to developmental tasks (Gentner et al. Reference Gentner, Rattermann, Markman, Kotovsky, Simon and Halford1995). Thus, it seems unparsimonious to assume that analogical reasoning abilities begin with processes that cannot ultimately perform the variety of tasks that are clearly part of the repertoire of older children and adults.
Although a developmental approach to analogy has the potential to offer great value, it must ultimately point the way toward adult analogical competence in order to actually deliver that value. That is, to be a successful developmental account, a theory must begin at a reasonable starting point and demonstrate the path/process through which the system progresses to reach the known end state. The ARP theory does not explain full competence, and cannot, in principle, be extended to do so without it becoming a part of a larger theory.
