Leech et al. propose relational priming as the mechanism for the development of analogy making in children, and they provide simulation results that follow several known trends observed in children's early analogical reasoning. However, there are limitations in the present model that undercut its feasibility as a general explanation for the development of analogy. First, the authors claim that rather than explicitly representing relations as structured representations, it may be sufficient to conceptualize relations as transformations on objects. A consequence of this approach is that it appears to leave the model unable to integrate multiple relations when making analogies. For example, it is unclear how the model could sort out the mappings between kicks (x, y) and kicks (y, z) and hits (b, c) and hits (a, b). Although the authors acknowledge that their current model does not explain all aspects of relationally complex adult analogical reasoning, in fact the ability to represent integrated relations is crucial to a developmental model as well. By 5–6 years of age, children are well above chance in tasks requiring integration of multiple relations (Andrews & Halford Reference Andrews and Halford2002; Holyoak et al. Reference Holyoak, Junn and Billman1984; Richland et al. Reference Richland, Morrison and Holyoak2006), and even children ages 3–4 show some ability to integrate two relations (e.g., Rattermann & Gentner Reference Rattermann and Gentner1998a; Richland et al. Reference Richland, Morrison and Holyoak2006).
Second, all developmental changes in the model derive from knowledge accretion, but knowledge has been shown unable to predict all behavioral patterns of development (see Goswami et al. Reference Goswami, Leevers, Pressley and Wheelwright1998; Richland et al. Reference Richland, Morrison and Holyoak2006). In Richland et al. (Reference Richland, Morrison and Holyoak2006), children were asked to map analogical correspondences between sets of two pictures in which the same relation was represented (e.g., picture A showed dog chases cat chases mouse; picture B showed mom chases boy chases girl). A test of the stimuli showed that children ages 3–4 had sufficient relational knowledge to identify the primary relation at 90% accuracy across all the pictures, indicating that adequate relational knowledge could not be the explanation for condition or age related differences in performance. Although 3–4 year olds could solve simple one-relation analogies between the scenes (e.g., dog chases cat; boy chases girl), the same participants (with the same relational knowledge) showed significantly lower relational mapping accuracy when there was an additional processing load. This was either in the form of an object similarity distracter in the target scene, or when the mapping task required a higher level of relational complexity. All pictures were counterbalanced across conditions. With age, children improved in these conditions in timing that coincides with maturation of working memory and inhibitory control systems. These data indicate that while acquisition of adequate relational knowledge is likely a prerequisite to relational reasoning, it is not sufficient to explain established developmental patterns.
The authors motivate the approach of using priming rather than explicit relational representations because no current models account for how explicit relational representations can be learned. This is, however, an oversight. The Discovery of Relations by Analogy model or DORA (Doumas & Hummel Reference Doumas, Hummel, Bara, Barsalou and Bucciarelli2005a; Reference Doumas, Hummel, McNamara and Trafton2007; Doumas et al. Reference Doumas, Hummel and Sandhofer2008) is a connectionist model that learns explicit representations of relations from unstructured examples. Beginning with simple distributed representations of objects as feature vectors, DORA uses comparison to bootstrap learning explicit representations of object properties and relational roles. DORA then uses mapping to link sets of relational roles into complete multi-place relations. Importantly, DORA uses time to dynamically bind these representations of object properties, roles, and relations to arguments (see, e.g., Doumas et al. Reference Doumas, Hummel and Sandhofer2008; Hummel & Holyoak Reference Hummel and Holyoak1997; Reference Hummel and Holyoak2003). In other words, DORA learns explicit representations of relations that explicitly specify bindings between relational roles and fillers. The resulting representations allow DORA to account for numerous phenomena from children's and adults' relational reasoning including all those accounted for by the target model, as well as numerous phenomena the present model cannot account for (such as relational integration; see Doumas et al. Reference Doumas, Hummel and Sandhofer2008).
