INTRODUCTION

Acquiring syntax involves assigning words to grammatical categories and understanding the relationship between the categories at the abstract level. The child must go beyond specific instances heard in the input and reach abstract patterns and rules. Such abstract representations are essential for syntactic productivity in children's speech and comprehension. Here we report an empirical study that addressed the question of abstract grammatical representation in young children, using the case of grammatical gender feature agreement.

Grammatical agreement such as number and gender is common in natural languages. For example, in English the subject and the present tense main verb agree in number. In gender-marked languages, the gender features of certain syntactic categories (such as nouns, determiners, generalized quantifiers) agree when conjoined to form phrases or sentences. In French, for instance, gender-marked determiners have separate masculine and feminine forms (e.g., le _MASC/la _FEM ‘the’; un _MASC/une _FEM ‘a’). Most nouns, however, have one form and belong to one of two genders (e.g., chapeau _MASC ‘hat’; rue _FEM ‘street’). The gender of most nouns is semantically arbitrary. Their gender agreement with other categories is strictly a syntactic property (e.g., un _MASCchapeau _MASC ‘a hat’; une _FEMrue _FEM ‘a street’).

There is a growing interest in young children's representation and processing of grammatical agreement. Recent online comprehension studies examined grammatical gender agreement (e.g., Johnson, Reference Johnson, Brugis, Clark-cotton and Ha2005; Lew-Williams & Fernald, Reference Lew-Williams and Fernald2007; Van Heugten & Shi, Reference Van Heugten and Shi2009). Gender knowledge was tested implicitly, as the child's task focused on identifying a named object while two objects were displayed on a monitor. The target nouns were produced in the context of determiners, and gender of determiners and nouns was manipulated across trials. The idea behind the design was that if children did not process gender, their noun recognition should not differ across gender manipulations. In gender-matched trials, children watched two objects of different genders while hearing speech containing a determiner preceding a noun naming the target object. The determiner and noun shared the same gender, e.g., Dutch: de _COMbal _COM ‘the ball’ (Johnson); French: la _FEMbanane _FEM ‘the banana’ (Van Heugten & Shi); Spanish: la _FEMpelota _FEM ‘the ball’ (Lew-Williams & Fernald). Children's responses suggested that they used the determiner gender to efficiently process the noun while hearing the speech. In gender-uninformative trials, the nouns for each object pair shared the same gender so that a determiner with that gender was compatible with both objects, e.g., Spanish: seeing a ball (pelota _FEM) and a cookie (galetta _FEM) while hearing la _FEMpelota _FEM (Lew-Williams & Fernald). Accordingly, performance was less efficient in uninformative trials than in gender-matched trials. Target recognition was most impeded in gender-mismatched trials, in which gender agreement was violated, e.g., French: *le _MASCbanane _FEM (Van Heugten & Shi); Dutch: *het _NEUTbal _COM (Johnson). These studies suggest that two- to three-year-olds process grammatical gender online.

It is now important to determine whether the gender effect in the above studies reflected knowledge that was truly syntactic. Since the nouns were familiar to young children, an alternative account is equally plausible. This account is linked to the specific phrases that children likely encountered in their prior input. Children may have simply represented specific exemplars, with no abstract knowledge. For example, when a gender-matched trial in Van Heugten and Shi (Reference Van Heugten and Shi2009) presented two objects depicting bateau _MASC ‘boat’ and banane _FEM ‘banana’, the word la _FEM in the auditory instruction had 100% probability to predict the target object (banane) and 0% probability to predict the distractor. This is because the specific phrase la banane was likely encountered in children's prior input and should therefore be previously stored, whereas the ungrammatical phrase *la bateau should not. In an uninformative trial the nouns for the two objects (e.g., bateau _MASC ‘boat’, ballon _MASC ‘ball’) both likely co-occurred with the same determiner (le _MASC) in children's prior input (e.g., le bateau; le ballon). That determiner thus had 50% probability to predict the target or the distractor. Finally, the gender-mismatched phrases (e.g., *le _MASCbanane _FEM) presumably never occurred in the prior input, i.e., 0% probability of co-occurrence, which should yield the poorest noun recognition. This probability account therefore would be based on memorized phrases and their co-occurrence probabilities, rather than assuming any abstract representation.

The key evidence for abstract representation is the ability to generalize from previously encountered words to grammatically coherent new combinations (that have never been heard). Abstract representation is most convincingly demonstrated by the ability to apply generalized structures to novel words and novel combinations, a kind of productivity that is characteristic of syntactic knowledge.

The present study aimed at determining whether young children represent grammatical gender abstractly. We tested this question using novel nouns. In particular, we examined whether children represent gender as an abstract feature of grammatical categories and process gender agreement non-adjacently. The previous studies reviewed above all used adjacent gender-marked words, i.e., a determiner immediately followed by a familiar noun. It is unknown if children process non-adjacent dependency of gender-marked words. Preferential listening studies showed that 1·5–2-year-olds track certain non-adjacent dependencies of their native language (Dutch: Van Heugten & Johnson, Reference Van Heugten and Johnson2011; English: Santelmann & Jusczyk, Reference Santelmann and Jusczyk1998; French: Van Heugten & Shi, Reference Van Heugten and Shi2010; German: Höhle, Schmitz, Santelmann & Weissenborn, Reference Höhle, Schmitz, Santelmann and Weissenborn2006) and of briefly trained artificial speech (Gómez, Reference Gómez2002). For example, infants accept is_ing, but not *can_ing dependencies (Santelmann & Jusczyk). French-learning infants accept determiner–auxiliary dependencies such as le_va (Det_SING_Aux_SING) and disfavor violations such as *les_va (Det_PLUR_Aux_SING) (Van Heugten & Shi). However, the acceptable non-adjacent dependencies in existing studies always involved specific items that co-occur frequently in the input, whereas the disfavored non-adjacent elements do not or rarely co-occur in the input. Infants' sensitivity may be driven by their memory of the specific non-adjacent dependent elements stored from the input (e.g., hearing is_ing and le_va frequently but never *can_ing and *les_va). No study has directly examined children's ability to track non-adjacent dependencies at an abstract level, e.g., between grammatical categories.

To clearly assess abstract representation, we used a word-teaching phase, which introduced two new word–object pairs. Each novel word followed a gender-marked determiner. That is, children were simultaneously presented with the new word forms and their genders. The novel words themselves contained no phonological markings of gender. Their genders were strictly assigned by the determiners during the teaching phase. The test phase presented the two objects while one was named with a new determiner. For example, a child was taught a novel noun as masculine (e.g., un _MASCravole ‘a ravole’). Subsequently, ravole would occur with another masculine determiner (le _MASC ‘the’) in a gender-matched test trial, but with a feminine determiner (la _FEM ‘the’) in a gender-mismatched trial. In neutral trials the target noun followed a gender-unmarked determiner les ‘the’. Crucially, because the combinations of each novel noun and a new determiner in the test phase had never occurred during training, nor in the child's prior environment, the probability between the specific determiner and the target novel word in each test trial was 0%. This design thus enabled us to test if children could rely on the abstract representation of grammatical gender (i.e., beyond any memorized specific phrases and the co-occurring probabilities between words) to distinguish different kinds of test trials. Specifically, we asked if children could assign the abstract gender feature in determiners to adjacent novel nouns during the teaching phase and perceive the feature upon hearing new instances involving new determiners and the novel nouns in the test trials.

In addition, our experiment tested whether young children represented non-adjacent dependencies at the abstract grammatical level, beyond the tracking of specific familiar elements shown in previous studies (Höhle et al., Reference Höhle, Schmitz, Santelmann and Weissenborn2006; Santelmann & Jusczyk, Reference Santelmann and Jusczyk1998; Van Heugten & Johnson, Reference Van Heugten and Johnson2011; Van Heugten & Shi, Reference Van Heugten and Shi2010). We inserted a gender-unmarked adjective between the new determiner and the novel noun in all our test trials (Det + Adj + N). To process gender agreement, children had to activate the genders of new determiners as abstract feature classes, and activate the genders of the newly taught novel nouns. Gender agreement was thus assessed across non-adjacent abstract categories (Det, N).

We hypothesized that two-year-olds represent abstract grammatical gender and the gender feature agreement between different categories. We expected them to assign the determiner gender to the co-occurring novel noun during training, and to process the feature agreement between the new determiners and the novel nouns during the test phase. Moreover, gender processing should be incrementally affected while gender-marked elements were being heard.