Abstract
Boloh and Ibernon (2013) use the acquisition of the French gender system to argue for a ‘dual mechanism’ model of language (e.g. Marcus, Brinkmann, Clahsen, & Wiese, 1995; Ullman, 2004), in which the masculine is computed by ‘default’ whereas the feminine (or specific case) is rote-learned in an associative memory system. In contrast, in ‘single mechanism’ models there is no strict division between grammar and lexicon; productivity depends on (type) frequency and transparency of a pattern, and children learn both regular and irregular patterns through distributional learning from the input, with representations only gradually becoming more abstract (e.g. Bybee, 2001; Tomasello, 2003). Recent studies have shown that children possess powerful distributional and statistical learning capacities (e.g. Gómez, 2002), and that syntactic categories may be inferred from the input (e.g. Mintz, Newport, & Bever, 2002).
At first sight, French children seem to have excellent opportunities for bootstrapping the grammatical gender system based on a distributional analysis of (morpho)-phonological regularities, as nouns may have typical feminine (e.g. la bicyclette ‘the bicycle’) or masculine (le bateau ‘the boat’) endings. Indeed, Karmiloff-Smith (1979) has argued that French children may use such lexically based ‘phonological’ strategies to assign gender to a novel word.
In the current study, Boloh and Ibernon (henceforth ‘B&I’) claim that French-learning children determine the grammatical gender of a novel word on the basis of a ‘masculine as default’ strategy rather than phonology. This study is the last of a series of three studies replicating previous experiments by Karmiloff-Smith (henceforth ‘K-S’). In the first study (Boloh & Ibernon, 2010, K-S exp. 8), gender was assigned based on phonological cues alone (e.g. ‘deux plichettes’ should become ‘la plichette’, and ‘deux bicrons’ should become ‘le bicron’). K-S found that the suffix-congruent choice was dominant for both endings. Although scores were higher for the masculine words, scores for feminine endings were all well above 50% (except for the scores of the 9-year-olds). In contrast, B&I found near ceiling scores for masculine and much lower scores for feminine endings (16% vs 52%), which was interpreted as evidence for a masculine default. In their discussion of the previous results by Karmiloff-Smith, Boloh & Ibernon (2010, p. 4) note that scores for masculine endings were ‘virtually always at ceiling for all ages, while scores for feminine endings showed asynchronous zigzag curve shapes with important variations between adjacent age classes (e.g. 72% at age 4 to 100% at age 5 and then to 57% at age 6 for the -ine ending)’. However, taking a closer look at the original data (Karmiloff-Smith, 1979, p. 160), it appears that while scores for one masculine suffix (-on) are indeed close to ceiling, scores for another (-ier) show variation (from 100% at age 3 to 71% at age 4 to 79% at age 5 to 85% at age 6), comparable to the variation found for feminine -ette. In fact, if we compare the scores across the 3- to 6-year-old age range, the average score for the masculine -ier (84%) is lower than the average score for -ette (89%).
In their second experiment, Boloh, Escudier, Royer, & Ibernon (2012, K-S exp. 9) investigated whether supplying an article would override the suffix-based response (e.g. ‘un plichette’ should then become ‘le plichette’ and ‘une bicron’ should become ‘la bicron’). Here, the differences between the two studies are even more striking; while K-S found suffix-based responses for both masculine and feminine endings for the younger children (up to 63% for masculine endings at age 4 and up to 81% for feminine endings at age 5), Boloh et al. found that suffix-based responses for masculine words were at chance and very low for feminine words (ranging from 5% to 13%).
In the current experiment (which replicates K-S exp. 10), B&I investigated whether a natural gender cue would override the phonological cue. Here, K-S again shows suffix-based responses to be dominant (ranging from 54% at age 9 to 84% at age 5 for masculine endings, and from 20% at age 9 to 87% at age 8 for feminine endings). Only at age 10 and 11 did natural gender-based responses prevail. In contrast, B&I find a high rate of suffix-based responses for masculine endings (ranging from 83% at age 4 to 91% at age 12, after which they drop below chance), and a low rate of suffix-based responses for feminine endings (ranging from 13% at age 12 to 40% at age 4, dropping to zero for 15- and 17-year-olds).
A first question that arises is how two very similar sets of experiments can yield such very different findings, an issue that is not addressed by B&I. First, there are differences in age groups and group sizes. For instance, K-S tested up to nine different age groups (ranging from 3;2 to 11;6); her youngest participants were 24 3-year-olds (M = 3;5) and 43 4-year-olds (M = 4;6), while the youngest children in B&I (2010) were 50 4-year-olds (M = 4;0) and 50 7-year-olds (M = 7;0). In the current study, B&I also use slightly older children than K-S (e.g. 47 4-year-olds, M = 4;5, ranging from 3;8 to 5;0). However, K-S (1979) divided participants across five different experiments, which means group sizes were small (ranging from four to five for the youngest group to nine to ten for the older groups). There were also some differences in methodology, possibly resulting in additional processing load for the B&I experiments, in which three rather than two colour-coded objects were coupled with three bags of different sizes in a story context, while K-S only had children comment on a simple action. Another likely reason for the different findings is the difference in stimuli, as seven endings were not used in the replication (e.g. -ien, -elle), while seven additional endings were used by B&I (e.g. -a, -ise). Such differences may have led to different (type) frequencies of endings. Furthermore, B&I chose stems that are rare or non-existent in French, which means that analogical effects might have been weaker in their studies. However, note that such input-related differences between the studies are not expected under the ‘masculine-as-default’ approach, as the suffix-based response should be at ceiling for masculine and at chance for feminine always, regardless of the phonological make-up of the stimuli.
Second, one could ask whether the authors are justified in interpreting the fact that scores for feminines were at or below chance as evidence for a ‘default’. Surely, a classic ‘unusual’ circumstance in which a default should apply is a novel word, which does not have a lexical trace (e.g. Marcus et al., 1995). As shown in their Table 2, the percentage of suffix-congruent choices for 4-year-olds is higher for masculine suffixes (83%) than feminine suffixes (39%). However, the fact that 4-year-olds produce a feminine article in almost 40% of the cases (even though the puppet had a male appearance) indicates that the default view is problematic. While the masculine gender is clearly the dominant choice, one should also be able to account for the apparent influence of phonology (i.e. lower rates of masculine responses due to feminine endings). More evidence for a phonological strategy in the authors’ own studies is found when we look at individual ‘strategies’. For instance, B&I (2010, exp. 1) find that there were more 4-year-olds who relied exclusively on feminine suffixes (11/50) than 4-year-olds showing the opposite pattern of almost never supplying a feminine response (10/50), which does not support their theory. In the current study too, quite a large number of 4-year-olds (10/47) treated the feminine as the dominant response, although a larger number (21/47) predominantly used the masculine. Furthermore, the general finding that the suffix-based response is predominant among the youngest age group indicates that they are least likely to apply a masculine default strategy. However, the classic notion of a default entails that it is phonologically and lexically unrestricted, used in overgeneralizations and used by young children rather than older children (e.g. Clahsen, 1999; Marcus et al., 1995). When the authors state that ‘Critically, scores for feminine suffixes were never at a level high enough to support a viable version of the phonological scenario’, one wonders when levels would be high enough. If an (early) analogical route for ‘feminine’ non-words is allowed to some extent, it becomes very difficult to distinguish the ‘default’ and the ‘phonological’ account on empirical grounds. All in all, the evidence seems less clear-cut than the authors present it to be, in line with a more gradient notion of productivity based on input patterns.
Could the dominance of the masculine gender be explained by its higher type frequency (approx. 60/40)? A default analysis has also been proposed for the Dutch two-gender system (e.g. Blom, Polišenská, & Weerman, 2008), but here input frequency conspires to make the common gender dominant, as it is twice as frequent as the neuter. Nevertheless, Blom et al. argue that children’s knowledge is grammatical rather than lexical, as it generalizes to other gender domains such as attributive adjectives – which may possibly provide a way of disentangling the two approaches. B&I acknowledge that input statistics may be important (i.e. ‘a possible initial, although not subsequent, role for type frequency’) to set the masculine as the default gender. However, given that some kind of statistical learning apparently needs to take place in order to yield the masculine as a default, why wouldn’t such a mechanism stay in place? This is especially puzzling given the fact that at a later stage, sensitivity to endings reappears, as it ‘would be what children might eventually induce upon acquiring grammatical gender rather than the way they learn it’. To conclude, advocates of a dual mechanism model should make very clear why rules or abstract features are needed in addition to probabilistic patterns in associative memory, if the latter alone might suffice.