Study 2: Non-canonical word order

Methods Participants

Twenty-eight French-speaking children (15 female) mean age 22 months (range 1;5.20–2;0.29 years) took part in this experiment in Geneva. Twelve out of the 28 children (43 %) were exposed to multiple languages. Parental questionnaires showed that exposure to French ranged between 40 % and 85 % in multilingual children. Since scores to the “MacArthur Communicative Development Inventory” (French version; Kern, 1999) did not differ between monolinguals and multilinguals (see means in Table 5; all words: t = -0.825, df = 21.864, p = .418; grammatical words: t = -0.906, df = 23.049, p = .374), we treated them as one group.

Eighteen additional children were tested, but excluded due to fussiness (7) or very unstable eye detection (11).

Materials

We created six test sentences (see Table A2 in Appendix A), four with the non-canonical NNproV order as in (2) and two with the canonical NVN order as in (3). The surface structure of the non-canonical construction is OSproV, with the object NP being moved from its original sentence-final to the sentence-initial position, thus preceding both the subject NP and the verb, and the object-clitic le (masculine) or la (feminine) inserted pre-verbally.

(2) Le garçon, la fille le poune.

the boy the girl him pseudo-verbs

‘The boy, the girl is pseudo-verbing him.’

(3) La fille dase le garçon.

the girl pseudo-verbs the boy

‘The girl is pseudo-verbing the boy.’

In Study 2, only two nouns/characters were used, a boy (le garçon) and a girl (la fille), to diminish the number of NPs to process. We used the same pseudo-verbs pouner and daser as in Study 1. Here, pouner meant to bend someone towards the floor and this pseudo-action was

presented with a distractor video illustrating the pseudo-action of moving someone’s body back and forth by pulling the arm. Daser referred to the pseudo-action of rotating someone’s arm and was paired with a distractor video involving the pseudo-action of moving someone’s leg up and down. The contrastive pseudo-actions were chosen in order to lower the similarity of the two visual displays and hence facilitate visual discrimination. Pseudo-actions were transitive and reversible and not typically lexicalized in French child-directed speech. The audio stimuli were again pre-recorded by a female native speaker of French in a child-friendly speech mode using a natural prosody at an average speech rate of 3.5 syllables/second. The non-canonical OSproV test sentences contained an increased prosodic salience of the topicalized object, with a pitch rise on the object (F0 +340 Hz on average) and a 200 ms pause between the object and the subject.

Each test sentence was paired with two causative action videos. In the agent-first videos, the NP mentioned first in the sentence was the agent of the action, illustrating the incorrect SOproV interpretation of non-canonical NNproV sentences, but the correct SVO interpretation of canonical NVN sentences. In the patient-first videos, the NP mentioned first was the patient, illustrating the correct OSproV interpretation of NNproV sentences, but the incorrect OVS interpretation of NVN sentences. The six test sentences were dispatched across two between-subject lists with different pseudo-random orders. Both lists started with non-canonical sentences to maximize the chances that children process them correctly without being primed by a different word order. The presentation of the target was counterbalanced across the left and right sides of the screen.

Procedure

Data were collected with an SMI RED eye-tracker (500 Hz sampling rate, 24 inch monitor). Familiarization involved two steps. We first familiarized participants with the two characters and with the fact that videos would appear to the left and right sides of the screen. Each character was presented separately, waving and smiling to the participant, together with an audio stimulus (e.g., Oh, il y a un garçon. Où est le garçon? Trouve le garçon! ‘Oh, there is a boy!

Where is the boy? Find the boy!’). Second, we introduced intransitive actions (eating and sleeping) and the simultaneous presentation of two videos to train children to visual discrimination (girl and boy eating on the left side vs. girl and boy sleeping on the right side) together with a neutral audio attention getter (Qu’est-ce qui se passe? ‘What is happening?’), followed by a discrimination request (Regarde! Le garçon et la fille mangent. Tu vois? Ils mangent. C’est où?

‘Watch! The boy and the girl are eating. Can you see it? They are eating. Where is it?’). Like in

Study 1, familiarization did not introduce pseudo-verbs or transitive structures. Thus, understanding the test sentences could not be attributed to learning during familiarization.

The test phase involved six trials. A test trial movie lasted 50 s, divided into ten regions (see Table 8, modeled after Gertner et al., 2006). First, each video was presented separately together with a neutral attention getter (Qu’est-ce qui se passe ici? ‘What is happening here?’).

Then, both videos were presented simultaneously three times. Their first simultaneous presentation (in region 17–20 s) served as baseline since they were shown without the critical test sentence. The following blank screen in region 21–27 s introduced the target sentence in the future tense (e.g., le garcon, la fille va le daser ‘the boy, the girl will him pseudo-verb’). The test sentence was then presented twice in the present tense together with the videos in region 28–35 s (e.g., le garcon, la fille le dase ‘the boy, the girl him is pseudo-verbing’), then once in the past tense during the blank screen of region 36–41 s (e.g., le garçon, la fille l'a dasé ‘the boy, the girl him pseudo-verbed’), and finally once again in the present tense with the videos in region 42–50 s.

The offset of the three test sentences in the present tense fell at 30 s, 34.6 s and 46 s, respectively.

Within each video, the action was shown continuously, loop wise. In order to capture children’s attention, we presented the smiling raising sun from the Teletubbies five times throughout the experiment. The whole testing session lasted about 10–15 minutes.

Data Analysis

The analysis was conducted on five critical 4-second time regions: 17–20 s (baseline), 28–31 s (videos + first test sentence in present tense), 32–35 s (videos + second test sentence in present tense), 42–45 s (videos + third test sentence in present tense) and 46–49 s (videos without test sentence). We used the same exclusion criteria as in Study 1, which resulted in the exclusion of 12 out of 112 trials in the non-canonical NNproV condition (11 %) and 6 out of 56 trials in the canonical NVN condition (11 %). Across time regions, we analyzed the proportion of looks to the agent-first video (calculated over the total looking times to both videos). We ran Generalized Linear Mixed Effects Regression models with subjects and items as random effects and sentence (canonical vs. non-canonical) and region as fixed effects, using the “lme4” package (Bates et al., 2015) in R (R Development Core Team, 2016). We used the maximal random-effects structure by subject and by item justified by the data.

Table 8

Test trial procedure in Study 2

Results

The proportion of looks to the agent-first video (PLA) across the five critical time regions is provided in Table 9. As in Study 1, we computed a baseline-corrected PLA showing the time course of children’s looking preference (see Figure 4).

Table 9

Proportion of looks towards the agent-first video with canonical NVN and non-canonical NNproV sentences in Study 2

Note. Significant differences as compared to the baseline are in bold. Standard deviations are in parentheses.

The model (see Output B2in Appendix B) showed a main effect of sentence type: children looked overall less to the agent-first video after hearing canonical NVN sentences (41.8 %) than after hearing non-canonical NNproV sentences (48.6 %, β = -0.791, SE = 0.402, z(596) = -1.967, p = .049). Crucially, the evolution of the PLA shows opposite directions for the two conditions:

while it increases between the baseline and region 46–49 s in NVN (+18.7 %), it decreases in NNproV (-12.3 %), this interaction between sentence type and time region being significant (β = 1.113, SE = 0.572, z(596) = 1.946, p = .052). Post-hoc comparisons using the Wilcoxon signed-rank test showed a higher PLA in region 46–49 s than in the baseline in NVN (53.0 % vs.

34.3 %, respectively, W = 54, p = .058), but a flipped effect with a lower PLA in region 46–49 s than in the baseline in NNproV (39.9 % vs. 52.2 %, respectively, W = 250, p = .059). All other effects were non-significant (all ps > .10).

Figure 4

Baseline-corrected proportion of looks to the agent-first video with canonical NVN and non-canonical NNproV sentences in Study 2

Note. Error bars indicate +/-1 standard error of the mean.

Discussion

Study 2 shows that 22-month-old children exposed to French have a preference for the correct OSproV interpretation of non-canonical NNproV structures with object topicalization, even though test sentences involved pseudo-verbs. This result extends previous studies conducted on 15- and 20-month-old children exposed to English, which showed that they can interpret object-fronted structures with known verbs (Gagliardi et al., 2016; Perkins & Lidz, 2020; Seidl et al., 2003).

Fronted objects involve a double challenge: sentences not only involve a non-canonical position of the complement (the object) with respect to its head (the verb), they also violate the quasi-universal subject-object order found in the canonical word order of over 95 % of all natural languages (Greenberg, 1963). This quasi-universal linguistic property is reflected in psycholinguistic studies in terms of a strong bias for expressing the agent first in sentence production (e.g., Demiral et al., 2008), and for interpreting the first noun of the clause as the agent in sentence comprehension (e.g., Bever, 1970; Bornkessel‐Schlesewsky & Schlesewsky, 2009).

Study 2 shows that 22-month-old children exposed to French are able to overcome this bias.

Study 2 replicates the preference for the correct SVO interpretation of NVN sentences found in Study 1. Nevertheless, the timing of the effect in Study 2 is slower than that found in

Study 1 and other studies using a similar experimental paradigm. Study 2 critically differs from these studies in that children were exposed to the SVO items after having been exposed to four OSproV items. The delayed effect observed here may thus be due to syntactic priming from these preliminary OSproV sentences in which the first noun is the object. Syntactic priming has been widely reported in the adult literature (see Pickering & Ferreira, 2008, for a review), as well as in children from the age of 3 (e.g., Bencini & Valian, 2008). If priming effectively took place such that the first NP of NVN sentences was initially misparsed as the patient, revision is then required to reach the correct parse. A plethora of evidence shows that children struggle with revising an initial parse, possibly due to the involvement of executive functions that only slowly mature until adolescence (e.g., Choi & Trueswell, 2010; Mazuka et al., 2009). We also note that the bias towards the video illustrating the patient-first interpretation of NVN sentences is already present in the baseline, i.e., before the test sentence. We conjecture that this bias is due to noise, since that condition involved only two items.

Studies 1 and 2 show that before her second birthday, the child already demonstrates elaborate, abstract knowledge of her language’s word order and of the option of displacing the object from its canonical position. This knowledge is necessarily acquired on the basis of input;

however, the mechanism underlying its acquisition is under debate. Before discussing the impact of our data on this debate, we consider some distributional facts about word order in French.