Artificial grammar learning in children: abstraction of rules or sensitivity to perceptual features?

We sincerely thank an anonymous reviewer for this nice metaphor and argument.

It may be worth explaining why it was not possible to include another test (such as a recognition test, for instance) following the implicit generation test. Recognition is very likely to elicit intentional information recollection processes (Gardiner & Java, 1993), with the result that such a test is a much more explicit test than an implicit generation test. Several authors have pointed out the danger of contamination between implicit and explicit tests when they are jointly administered to the same participants (Gebauer & Mackintosh, 2007; Seger et al., 2000; Shimamura, 1985). Counterbalancing the order between tests does not resolve the problem. A recognition test cannot be administered first because half of the items presented during this test are non-grammatical. However, exposing participants to errors (non-grammatical information) is known to be prejudicial to implicit learning (e.g., Baddeley & Wilson, 1994; Perruchet, et al. 2006) and this can affect performance in a subsequent test. The same reasoning holds if the generation test is presented first. Such a test inevitably leads participants to generate false (non-grammatical) items (partially or completely).

This does not mean that the children could not evoke the presence of adjacent or non-adjacent repetitions. As we saw in a small number of children, if the experimenter explicitly asked "Were the flags always made of different colors, or were some colors sometimes repeated inside the flags?", they were able to recognize that some colors were sometimes repeated. The explicit knowledge that could be expressed by children in the question phase was highly dependent on the type of questions asked by the experimenter.

Perruchet and Reber (2003, p. 129) argued that "if trained control subjects do not perform at chance level, this means that some undetected biases are still in operation". The below-chance production of repetitions in the random group probably expressed a natural bias against repetitions. Because the additivity assumption claims that non-specific variables have similar effects on the experimental and control groups trained in very close conditions, if repetition avoidance led to below-chance productions in the control group, it was very likely that sensitivity to repetitive structures raised performance in the experimental groups. However, this increase reached only chance level, not more. For a more direct demonstration of such learning, we computed the ratios between observed and theoretical proportions, and run an ANOVA with Type of series as a between-subjects factor. Type of series yielded significance for the non-adjacent and adjacent repetitions, F(2, 117) = 7.05, p < .01 and F(2, 117) = 18.25, p < .01 respectively. Planned comparisons showed that the non-adjacent group produced more non-adjacent repetitions than the adjacent and random groups, F(1, 117) = 9.46, p < .01 and F(1, 117) = 11.58, p < .01 respectively, as did the adjacent group for the adjacent repetitions, F(1, 117) = 26.85, p < .01 and F(1, 117) = 27.9, p < .01. Thus, there was a specific increase of adjacent or non-adjacent repetitions only in the groups where specific learning was expected.

The results suggest that the frequent initial bigrams were not learned, as their mean frequency was very close to the theoretical probability. However, it is worth pointing out that the second color in these bigrams was the first color of the repeated adjacent unit (e.g., BYY). Learning the initial bigram (BY) may have been prevented, or at least slowed down, by the formation of the repeated unit (YY). It would appear to be very interesting to study this issue in greater detail because it reveals the limitations of a purely statistical view of learning.