Grammatical gender attribution,input statistics and the default gender in French children: A reply to Kerkhoff

In her valuable commentary, Kerkhoff (henceforth, K) raises several issues, both empirical and conceptual. In particular, she argues that morphophonological regularities represented in associative memory might suffice for the acquisition of French grammatical gender. She thus questions whether a default implemented as a rule is necessary and whether it is even mandated by our data. According to K, a distributional analysis of morphophonological regularities might bootstrap the gender system (i.e. given the reference to Karmiloff-Smith, 1979, gender attribution). There are reasons to be sceptical about this, whatever children’s statistical learning capacities. French has around 50 nominal suffixes and although some have high gender-predictive values, most are only probabilistic. Those values cannot be known in advance: They must be progressively induced on the basis of the proportion of Det_masc–N pairs vs Det_fem–N pairs for each suffix. Why, then, would children bother purposely inducing dozens of probabilistic phonological patterns and then rely on them to eventually decide on the gender form of the determiner if the availability of the latter is the logical prerequisite of the former? If they did so, gender attribution errors would abound, which is not the case. So, gender attribution must be deeply rooted in rote learning of Det–N pairs throughout lexical development (Boloh & Ibernon (B&I), 2010, p. 21; Matthews, 2010, p. 881).

Based on this, a tacit, purposeless registration of the probabilistic gender values of phonological clusters might also take place as a byproduct of gender learning. This might eventually ease the storage and retrieval of the gender of previously learned nouns. However, this plainly could not guarantee the initial determination of nouns’ gender, even if a weak form of productivity can be demonstrated. Our production data precisely show this. We indeed found a moderate amount of suffix-based responses to nonce nouns with feminine suffixes. However, neither children nor adults supplied accurate suffix-based responses despite suffixes’ exceptionally high predictive values (above .90). Results would thus be worse for the many suffixes with lower predictive values (the type frequency issue is addressed in B&I, 2010). K spotted a subset of K-S’s (Karmiloff-Smith, 1979, experiment 8) results that contradict our findings. However, as there was no consistent developmental trend in her data there is no reason to discard particular age groups, especially the 9-year-old group. The average score is thus 74% for two of the most predictive feminine suffixes and the reasons for the differences with our results are discussed in B&I (2010, pp. 16–17).

Still, one could consider that near ceiling scores for nouns with masculine suffixes, combined with even a moderate amount of suffix-based scores for nouns with exceptional feminine suffixes, would provide acceptable evidence for the phonological scenario. The key question here is whether suffix-congruent responses for masculine nouns are true suffix-based responses or default responses. We found evidence for the latter. First, there were frequency effects for feminine suffixes but not for masculine ones (B&I, 2010). Second, while the masculine and the feminine are almost equally probable in the input (P&S input: 57/43, RJ data: 54/46; Lexique 2 data: 55/45; Matthews, 2010, p. 887; New, Pallier, Brysbaert, & Ferrand, 2004), they are not so in children’s representation: when provided with nonce nouns devoid of any phonological cue, children overwhelmingly supply the masculine (on at least 89% of trials) both for animates (current experiment 2) and for inanimates (in preparation). Third, the masculine applies despite putative contrary evidence, keeping suffix-based responses for nouns with exceptional feminine suffixes to 50% or lower.

K’s counter-argument here is that the default should be lexically and phonologically unrestricted. This requirement might hold for the English past tense issue but it is too strong for French grammatical gender. Considering that the masculine determiner might operate as a default does not entail that we should definitively dispense with any form of associative learning. So, either everyone admits that suffixes can eventually be associated to a gender value (i.e. nouns, including nonce ones, may have a morphological trace) or no one does, in which case the phonological scenario is ruined.

K further argues that our data show a general decrement of suffix-based responses for feminine nouns, which does not fit the default account. However, there was no significant difference between 4- and 7-year-olds in the current study. High suffix-based scorers were indeed more numerous among the 4-year-olds than the 7-year-olds in B&I’s (2010) experiment 1 but not in experiments 2 and 3. More generally, there was no consistent decrement of suffix-based responses for feminine nouns across the five experiments reported in B&I (2010) and Boloh, Escudier, Royer, and Ibernon (2012).

That being said, we agree with K that further data would be necessary to disentangle the issues. In a further experiment (in preparation), we assigned the ‘ending alone’ conditions from B&I and the ‘discordant’ conditions from B et al. to 204 children aged 4–11. We computed correlations between scores to feminine nouns in the former and scores to masculine nouns in the latter as both sets were originally found to be around 50% or below, with some inter-individual variation. Assuming that children do not segregate a priori between feminine and masculine suffixes, the hypothesis that suffix-congruent scores for masculine suffixes are actually suffix-based ones predicts a positive significant correlation between the two distributions: we found a significant negative one at each age. Participants who were high suffix-based scorers both for feminine suffixes in the ‘ending alone’ condition and for masculine suffixes in the ‘discordant’ condition were fewer than 2%. Suffix-congruent responses for nouns with masculine suffixes are thus qualitatively different from suffix-based responses for nouns with feminine suffixes: they do not seem to stem from a tacit online estimation of suffixes’ gender values, which suggests that they should better be regarded as default responses.

To conclude, morphophonological patterns are neither necessary nor sufficient and French children need not labour in vain to extract their probabilistic values. The default might not have been necessary on computational grounds but there is evidence that children organize the input that way, which might stem from the logic of the discovery of gender classes (B&I, 2010, pp. 22–23; B&I, the current article). Note that there is nothing puzzling in the role that phonological information plays at that step: while the input must provide a subset of suffixes with high predictive values in each subclass, the child need not have induced them yet. He or she only has to take advantage of this to make different noun types similar through their shared suffix.