Developing sociolinguistic evaluative competence: the case of French WH- interrogatives

2.1 The emergence of a sociolinguistic evaluative competence

Labov (1964) develops a very prominent model of how children acquire and develop their sociolinguistic overall abilities and can be summarized as a 4-stage sequential process. In this model, children under the age of 5 (basic grammar stage) first acquire the core grammatical rules and the basic lexicon of their linguistic system, based mainly on the input they receive from their parents. Between age 5–12, children build up their vernacular variety (local dialect) by receiving input from their expanding peer group. In the early adolescence stage, children then begin to develop some degree of consciousness of the social significance of their local dialect when they are confronted to other dialects, mostly at school and by crossing paths with more adults. In the late adolescence stage (also called stylistic variation stage), children progressively learn how to adapt their productions according to the requirements of particular interactions, such as using standard variants in formal situations. This model is inspired by a grammatical framework where general rules precede variable rules, and although it has been criticized and/or amended by decades of following research, this sequential view is still a very prevalent take on how children develop language.

As Chevrot (2024) points out, there is however still a general dearth of empirical investigations into how children perceive sociolinguistically charged language. In an overview of research ranging from 1958 to 2013, Nardy et al. (2013) found 39 studies investigating the production of sociolinguistic variants in children. These studies have almost exclusively focused on the production of phonetic/phonological variants – mostly in English –, probably as a consequence of a long-standing tradition of sociophonetic studies in sociolinguistics. This line of research, however, has provided solid evidence for an earlier ability in children to produce sociolinguistic variants in distinct contexts. Roberts (1994) for example found that children aged between 3;2–4;11 are able to alternate between non-standard/vernacular -in and standard -ing verbal endings according to their interlocutor (more standard productions when directed to an adult). In French, works such as Chevrot et al. (2011) have linked the production of optional liaison by children aged from 2;3 to 6;0 to their socioeconomic status (SES). Older children produce more of these non-mandatory phonological alternations (e.g. gros éléphant ‘big elephant’ which can be pronounced either [gʁozelefɑ̃] with the liaison or [gʁoelefɑ̃] without it) overall, in a target-like fashion, but children from higher SES backgrounds use them more proficiently, and earlier, than children from lower SES backgrounds.

Such works have brought forth evidence for a more nuanced picture than that proposed in Labov (1964), with heavy influences from usage-based approaches to language use, and in particular Bybee (2006)’s exemplar theory. Influenced by their respective linguistic input (which is highly constrained by socioeconomic factors), children’s early linguistic productions are a first window into their sensitivity to sociolinguisic variation. To consistently tease apart contexts where one variant is more suited than another, they have to assess the social properties of said contexts much earlier than what Labov postulated in his earliest proposals.

Some works have studied the “evaluative” component of sociolinguistic competence more directly. Buson (2009) and Buson and Billiez (2013) for example assessed the conscious evaluative capacity of children with regard to socially charged language. By asking children aged 9–11 to comment on answering machine messages with different French registers, these works establish that “metapragmatic comments” are already possible at around that age, with explicit judgments such as “this isn’t a proper way to speak French”. With a few English sentences that are not sociolinguistically marked, Vaughn and Becker (2024) documented the emergence of the ‘social-semiotic landscape’ of 94 American English speaking children aged between 5 and 12 y.o. The authors ran questionnaires that explicitly targeted how the children felt about a person they had just heard, but also why they felt that way. They binned the answers according to what type of knowledge or perspective was used: for instance, did the child compare what was what said to what he/she likes to do, did the child place the person in a social type of macro-demographic category, etc. The authors show that the proportion of ‘public’- and ‘evaluative’- type answers increase with age. These comments either contain explicit reference to an ideological schema, i.e. what some categories of people are supposed to talk like, or at least contain evaluations or descriptions of the speech signal itself, denoting a judgment driven by how the person spoke. Independent of sociolinguistic variation, these results illustrate how children become able to make social assessments over language and speakers in a progressive way.

For a more implicit assessment of younger children’s evaluative sociolinguistic ability, Rosenthal (1974) used boxes speaking either African American Vernacular English (AAVE) or Standard English to elicit sociolinguistic attitudes in preschool children (age 3;0–5;11). The results point to a difference in perception with regard to niceness and other social traits (e.g. need for a gift) depending on the variety of English the children heard. Comparing Hawaiian Creole English to Standard English, Day (1980) found with the same protocol that these attitudes shifted between kindergarten and first grade, with influences from SES-related factors. A more recent work (Kinzler and DeJesus 2013) made use of a matched-guise-like paradigm where, as inspired by Lambert et al. (1960), children aged 5–6 had to locate different people on a variety of social scales after hearing some statements which were either native-accented (American English) or foreign-accented (French-accented English). A first experiment found that when the statements contained neutral sentences, American-accented speakers were rated higher on some scales (friendliness, niceness, ‘living around here’) but not on others (smartness, ‘in charge’). In a second experiment with socially charged statements (American-accented speakers saying mean things, foreign-accented speakers saying nice things), foreigners became nicer, friendlier and smarter than American English speakers. In yet another matched-guise inspired experiment, Dollinger et al. (2024) investigated the perception of foreign-accented English with Canadian English speaking children aged from 7 to 12. They found that younger children (aged 7) seem more tolerant with regard to accentedness than older children (who rate accented Englishes harder to understand than Standard Canadian English), which the authors take as first evidence for the apparition of some degree of prejudiced attitudes against non-native accented language.

This – very scarce – body of research on child evaluative abilities has paved the way for a more systematic approach to investigating sociolinguistic abilities in children, both in production and evaluation of socially situated language. Our work aims at contributing to this line of research by assessing the latter in French, by focusing on a syntactic point of variation (French WH- interrogatives) and by expanding the social properties that are being assessed. To get a better view of how children develop their understanding of the relationships between language and social characteristics, we argue it is necessary to go beyond the in-group proximity measures that focus on friendliness (for instance by asking children to also assess prestige-related traits), but also to have a more quantitative approach across age groups (with more children taking the tasks, which requires a rapidly and easily deployable experimental paradigm). To go beyond the traditional focus on sociophonetic points of variation, we argue that the system for WH- interrogation in French is an ideal case study.

2.2 WH- interrogatives in French and their acquisition

As mentioned in the introduction, there are different linguistic variants for French WH- interrogatives (1). In adult French, many different works have analyzed this alternation phenomenon, from different linguistic perspectives. From the syntactic point of view, it is generally established that the so-called in situ form (“Tu arriveras quand ?”) is the base form from which all other variants derive, with different descriptions and constraints depending on the authors’ theoretical frameworks (Abeillé 2007; Ross 1967; Sag 2010). Semantico-pragmatic approaches have documented the influence of information structure on the production of the different variants (Beyssade 2006; Chang 1997; Cheng and Rooryck 2000). Phonotactic studies have presented empirical evidence for the number of syllables of both the interrogative and non-interrogative parts of the question being a predictor of the variant used in modern French (Hamlaoui 2011), while both macro- and micro-diachronic changes have been documented with respect to the preferred variants (Larrivée 2019; Thiberge et al. 2021). Finally, different sociolinguistic approaches have linked the different uses of interrogatives to diatopic (Guryev and Delafontaine 2015; Mathieu 2009) and diastratic (Adli 2006; Coveney 2011; Quillard 2001) variation.

This amount of work illustrates how complex the WH- interrogative system is in French, yet only the syntactic factors seem to have permeated to acquisition research. It has been widely documented that in situ variants appear first in L1 French children productions, around age 2 but with considerable individual variation (Hulk 1996; Prévost 2009; Zuckerman and Hulk 2001). The derivation of other variants from the in situ one has led researchers to build a “complexity metric” according to which complexity increases gradually across variants so that they become more difficult to process and produce by typically developing children – but also by atypically developing children and L2 learners (Jakubowicz 2011; Scheidnes and Tuller 2010).

Another account (Faure and Palasis 2021; Palasis 2013) proposes a diglossic approach to the acquisition of the French WH- interrogation system. In this proposal, children acquire two different grammars one after the other: the first one is that of a language used with family and peers, while the other one is that which one uses outside these circles. This view is, in a sense, very similar to Labov (1964)’s, albeit with slightly different stages and with a heavier weight of syntactic constraints.

3.1 Design and materials

We devised an auditory matched-guise experiment where participants had to listen to recordings of dialogues between two speakers (A and B) on a computer. Person A sets a small discursive context and person B asks a WH- question. After hearing each recording, participants had to assess six different social characteristics of person B (only) with six corresponding 7-point scales. The scales targeted dimensions such as Age, Richness, Sociability, Geographical Origin, Education, and Hobbies. For a better comparison of both the adult and the child experiments, we used pictographic representations of social properties rather than textual descriptions of the dimensions participants had to assess. The pictures used for this were normed in a separate experiment (see Supplementary Materials in the OSF repository at https://osf.io/7r6kv/). These six scales were selected among the nine that were normed because they yielded the best results on the social dimensions that were targeted, and because they were more reusable with children. Figure 1 illustrates a typical item (The text under the audio players reads: “The last person who speaks is more like…”).

Figure 1:

Example item from MGT-A as seen by participants.

As for the experimental design, we manipulated two elements, with a TYPE variable for the syntactic variant of interrogative (3 levels: in situ, fronting with inversion, fronting) and with a FORMALITY variable for the context (2 levels: informal or formal), resulting in 2 × 3 = 6 different experimental conditions. 30 target items were created, with context formality being normed in a separate experiment. 30 filler items not containing any interrogative were added, as well as 3 practice items. Four L1 adult speakers of French (F1, F2, M1, M2) were recorded in a soundproof room while they read all the items in all the conditions, with a naturalistic tone. The recordings were then rearranged in Audacity (AudacityTeam 1999-2021) into the A/B dialogues so as to get all possible combinations of voices (F1-F2, F1-M1, F1-M2, F2-F1, F2-M1, F2-M2, M1-F1, M1-F2, M1-M2, M2-F1, M2-F2, M2-M1). In another norming study we assessed the a priori gender of the voices so the drawings used for the scales would be coherent with them, and we checked that all four voices did not differ too much in terms of how colloquial they sounded to avoid introducing biases with regard to the formality variable. An example item in all its conditions can be found in Table 1, and a complete list of the materials can be found in the Supplementary Materials.

3.2 Analysis and predictions

Previous works (see above) have shown that, for continental French at least, fronting with inversion is associated with a more standard way of speaking French and is more used in formal contexts and by people older than 30 y.o. As a consequence, with participants from a relatively homogeneous group of participants (from France), we expected this specific variant to give rise to higher ratings on age, richness, urban background, longer education and frequent reading scales. On the sociability scale, we expected some kind of interaction between TYPE and FORMALITY, with fronting with inversion making the person sound a bit more ‘out-of-place’ or weird in more formal contexts.

We analyzed the respective and combined influences of our manipulated variables (TYPE, FORMALITY) on the ratings given by participants on the six different social scales. Ratings were given on 7-point scales, which we coded from -3 (drawing on the left) to +3 (drawing on the right). We used Bayesian cumulative-link models (one per scale) to investigate the relationship between our independent and dependent variables, since they are best suited for ordinal data.^[1] Just like classical (frequentist) inferential statistics, these models yield numeric estimated coefficients (so-called estimates, β ˆ ) that help assessing the influence between variables, but they do not yield binary p-values. Rather, they generate a probability distribution of many simulated estimates that help evaluate how trustworthy the yielded estimate is for a given predictor (or combination of predictors). The probability of the estimated coefficient to be greater, P(β > 0), or lesser, P(β < 0), than 0 indicates how reliable the evidence from the dataset is regarding the existence of an effect of the predictor under investigation. Given the number of datapoints in this experiment, we will report probabilities higher than 0.90 as “robust evidence” for the existence of an effect, and probabilities where 0.80 < P < 0.90 as indicative of a trend. We will also report 95 % Credibility Intervals (95 % CrI), which are the numerical values of the posterior distribution between which there is a 95 % chance of finding the true value of the estimated coefficient.

In line with other works on French WH- interrogatives, we hypothesize that age group might also be predictive of some differences between the uses and ratings of different WH- interrogative variants, so we included it in the analyses as a binary predictor (+/− 30 years old) although we did not control for it during participants’ recruitment. This predictor, as well as the TYPE and FORMALITY ones were mean-center coded to allow for a more direct interpretation of the outputs (Brehm and Alday 2022) (–0.5/+0.5 coding for FORMALITY and age, with +0.5 being formal contexts and more-than-30-year-old respectively; for the TYPE variable the reference level was the FINV variant). In the present study, we used the brms package (Bürkner 2017, 2018; Carpenter et al. 2017) in R version 4.2.1 (Posit Team 2023; R Core Team 2023). To account for possible variation across items and participants, we included these in the random effects structure of maximally-specified models, including suitable random slopes (Barr et al. 2013). The full technical specifications and outputs of the models can be found in the Supplementary Materials.

3.3 Participants and procedure

52 people from continental France took part in the experiment. They were recruited via the RISC network (CNRS, UMR 3352), social networks and among university students. For this experiment, age was not a controlled variable and the adult population was artificially considered homogeneous in its approach to the linguistic variants, to which the children data could then be contrasted. As such, participant age ranged from 18 to 77 (mean = 26.8, median = 22.5). The experiment was programmed on the IbexFarm platform (Drummond 2016) and hosted on university servers, but experimental sessions took place in the lab. Participants listened to the stimuli in a soundproof room and rated the speakers on the scales on a computer. Before the experiment began, we informed participants of their rights, of the general nature of the task, and we presented them with the drawings they would see for the scales (12 different sets of contrasting pictures that explicitly introduced the targeted social dimension, see Figure 2, where the texts read: “Here is Jade. // Here is Elsa. // Who is richer?”).

Figure 2:

Familiarization screen from MGT-A.

After, this, participants were also invited to locate themselves on a paper-printed version of the scales, to familiarize them with the materials and to get a better grasp of how they felt about their meanings. Then, the real experiment began. A session was approximately 30 minutes long and participants were compensated accordingly.

With 30 target items, 30 filler items and 3 practice items, we received 63 × 52 = 3,276 answers per scale. The analyses we present below focus on the 1,560 answers per scale for target items only.

3.4 Main results

3.4.1 Correlations

We first checked for correlations between answers. Figure 3 offers a global view of the network of correlations between scales, and thus between social dimensions, that were calculated.

Figure 3:

Correlation matrix for the different scales from MGT-A.

The numbers on the bottom left part of the graph are the Spearman’s ρ rank correlation coefficients for the different scales (a cross means the correlation was not statistically significant), with a more visual representation on the top right part (blue means a positive relation between values, the bigger the circle the higher the coefficient). All in all, and just as in the norming experiment where we assessed the social properties associated with all the pictures we used (see Supplementary Materials), it appears there are some associations between scales. For example, answers on both the age and the richness scales are somewhat correlated (ρ = 0.42), but less so than the hobbies and education scales, where high education and reading seem to be quite often associated, ρ = 0.57.

3.4.2 General results

Figure 4 compiles all answers from participants. Each point from a given colored polygon represents the overall mean of answers for a given scale in a given condition. More than the precise numeric values for each point, it is the different shapes of the polygons that stand out, with a similar pattern for the fronting and in situ types and a different one for the fronting with inversion one.

Figure 4:

Overall results from MGT-A.

An in-depth inferential statistical analysis (see Table 2, and the Supplementary Materials for a more complete view of the models outputs) yields robust evidence for virtually all of these differences, on all scales, that set the fronting with inversion type apart from the other two, with:

1. F and IS sentences being associated with lower ratings on the “age” scale than FINV sentences (for both, P(β < 0) = 1).

2. F and IS sentences being associated with lower ratings on the “richness” scale than FINV sentences (for both, P(β < 0) = 1).

3. F and IS sentences being associated with higher ratings on the “sociability” scale than FINV sentences (respectively, P(β > 0) = 0.90 and P(β > 0) = 1).

4. F and IS sentences being less associated with ratings pointing to an urban background on the “geographical origin” scale than FINV sentences (respectively, P(β < 0) = 1 and P(β < 0) = 0.85 – only a trend for IS sentences).

5. F and IS sentences being associated with lower ratings on the “education” scale than FINV sentences (for both, P(β < 0) = 1).

6. F and IS sentences being less associated with ratings pointing to reading on the “hobbies” scale than FINV sentences (for both, P(β < 0) = 1).

Table 2:

Main results from the models for all scales.

Scale	Contrast	β ˆ	P(β > 0)	P(β < 0)	95 % CrI
Age	F versus FINV	–0.63		1	[–0.93, –0.32]
	IS versus FINV	–0.54		1	[–0.87, –0.21]
Richness	F versus FINV	–0.70		1	[–1.02, –0.37]
	IS versus FINV	–0.59		1	[–0.97, –0.20]
Sociability	F versus FINV	0.17	0.90		[–0.09, 0.43]
	IS versus FINV	0.56	1		[0.19, 0.94]
Geo. origin	F versus FINV	–0.42		1	[–0.73, –0.11]
	IS versus FINV	–0.15		0.85	[–0.42, 0.13]
Education	F versus FINV	–0.91		1	[–1.29, –0.55]
	IS versus FINV	–0.74		1	[–1.07, –0.40]
Hobbies	F versus FINV	–1.12		1	[–1.50, –0.72]
	IS versus FINV	–1.01		1	[–1.42, –0.61]

The combination of these suggests there is a specific social persona associated with using fronting with inversion: that of an above-average educated, older, richer person who reads books more often. On the other hand, using the fronting with inversion variant is generally perceived as socially weird.

4.1 Design and materials

Since it was important that even very young children were able to complete the task, we made sure the child’s task was simpler, shorter and more accessible than the adult one. Overall the paradigm was still the same (i.e. still a matched-guise task programmed on the IbexFarm platform), in which child participants had to listen to speech productions and then associate them to a “scale” embodying some social characteristics. However:

1. The number of items was reduced. From 4 practices + 30 fillers + 30 target items, we selected 4 practices + no filler + 18 target items from the adult experiment which seemed the easiest for children to understand (see complete list in the Supplementary Materials).

2. The number of conditions was reduced, and thus item length was reduced. We removed the formality condition and kept only the core aspect of the experiment: the TYPE condition (with the same three levels: fronting with inversion, fronting, in situ).

3. The number of response scales per item was reduced. We focused on 3 dimensions we thought would work best with children: richness, education and hobbies.

4. Since we wanted the task to be enjoyable and fast-paced enough to maintain a child’s attention throughout, we replaced the 7-point scales by binary choices (between property pairs which were the drawings used as scale limits in the adult experiment: poor/rich, cashier/doctor, sports/books). Figure 5 illustrates how it looked (the texts read: “The voice is rather the voice of…”).

Figure 5:

Example items from the child experiment (“richness” scale).

Overall, the experiment was a simple 1 × 3 design, with children hearing either a F/FINV/IS version of an interrogative sentence ( Où tu vas ? / Où vas-tu ? / Tu vas où ? – “Where are you going?”). Then they had to choose between two characters, picking the one who they believed fit better with what they heard. Since we wanted a simple task with simple items, only one property pair was shown at a time. For this reason, we built 3 blocks in each experimental list, with each block containing a different variant of each target sentence, which was associated to a different property pair. Each list was thus 18 × 3 = 54 target items, preceded by the 4 training items. To counterbalance everything (all interrogative variants being presented with all three social dimensions for all items, and with at least 2 different voices – 1 male and 1 female), we created 6 lists (see Supplementary Materials). Item presentation inside a block was randomized.

4.2 Variables, analyses, and predictions

Our dependent variable was thus the choice for either the picture on the left (poor/cashier/sports) or the picture on the right (rich/doctor/books), which we recoded 0/1 for the analyses. Accordingly, we ran Bayesian logistic regression using the brms package. Bayesian models were particularly adapted here, given the high variability we could expect from an experiment with children. We ran one model for each dimension, and we will report results for each separately. Contrary to the adult experiment, given the exploratory nature of this research and the noisiness of the data we collected, we will report probabilities higher than 0.85 as “robust evidence” for the existence of an effect, and probabilities where 0.70 < P < 0.85 as indicative of a trend.

Our main independent variable was the syntactic TYPE of the interrogative (reference level: fronting with inversion). Since we are interested in the evolution of the associations between interrogative type and social properties over time, we assigned children to different GROUPS that roughly encapsulate both age and actual class group (as given by the school were the experiment took place – see below for details). This variable was recoded numerically, from 1 to 4. We also included random intercepts for participants and items, with random slopes (see Supplementary Materials for full specifications).

Based on current knowledge of how social perception develops in children, we could anticipate different things:

1. Children seem to be able to make “metapragmatic comments” on sociolinguistic variation, that somewhat resemble adult judgments, at around age 9–11 (Buson 2009). We thus expected children from the corresponding age categories to exhibit a similar pattern as the one exhibited by adults on the scales we selected.

2. More precisely, we expect a progressive build up of this pattern, with the youngest children making different associations – or making no apparent association – between linguistic forms and social meanings, with the oldest children making associations in an adult-like way. In this view, in-between children would exhibit a different pattern from the youngest group but still not as solidified as their older peers.

3. Given that children progressively build their ability to explicitly establish how they feel about a person depending on how they talk (Vaughn and Becker 2024) but also that very young children exhibit production patterns similar to that of adults for at least some sociolinguistic variables (Roberts and Labov 1995), we also anticipate potentially different trajectories for the associations between language and social properties, with some associations established very early on while some might take longer to appear.

4.3 Participants and procedure

Participants were 136 children from a mostly monolingual community from a small town in Western France. The study was prepared long beforehand, with a convention signed between the experimenter’s university and the school, an ethics committee approval (Uni. Paris Cité, IRB number 00012019-19), and signed consent from all the children’s parents. Children who did not want to take part on the days the experiment took place were free to do so (but nearly all the children at the school participated).

The task was presented as a mini-game the children had to play during teaching hours. For the youngest children (age 3–5), the experimenter, who had previously been introduced to the children, played the game together with them on a tablet computer, under supervision by a trusted adult (school staff) in a quiet room. For children aged 6 and above, groups of 5–6 children played the game independently at the same time, on school computers in the same small quiet room with a trusted adult, after a quick introduction of the task by the experimenter. The experimenter then went from child to child to ensure everything was working fine and to answer questions. During the game, children were told they had to help researchers who had mixed audio files, by assigning them to people they thought had uttered them. At the end of each experimental block, a small screen showed that progress had been made and that the scientists were becoming more and more relieved. At the end of the game, a small diploma was given to all the children to thank them, even to those who did not participate. The task took a few minutes for the oldest children, and session times varied greatly for younger children, up to 20 minutes.

To ensure maximum protection of the children, no personal information except their age and their “classgroup” was linked to the data. Tables 5 and 6 give an overview of these dimensions (mean age: 7;06, median age: 7;00). Given that the school had multiple overlaps between age and class groups, for the analyses we merged class groups higher than CP in two “supergroups” based on input from school teachers relating to the levels of pupils: the first group (CE1-CM1) mixes children from CE1 and CE2 classes but also a few “advanced” CP children put in a majority-CE1 class and a few CM1 children who were in a majority-CE2 class, the second group regroups the vast majority of CM1s and CM2s (see Table 7 for the detailed grouping).

Out of the 54 × 136 = 7,344 possible answers, we discarded 1,761 because the experimenter noticed that, when, because of randomization parameters, an item was followed by another one with the same answer scale (i.e. two different sentences but the same pictures to assign to them), children often just re-clicked on the same picture they had just chosen, thinking it was a bug. For similar reasons, we also removed answers when children took less than 2 s to answer (N = 308). In total, we kept 5,275 answers (1,786 on the richness scale, 1,743 on the education scale, 1,746 on the hobbies scale). Despite this conservative approach to the data, Bayesian modeling still allows for good generalization over these results.

4.4 Main results

We first present results for each scale individually, before moving to the global picture.

4.4.1 Results for the ‘richness’ scale

Figure 6 provides a visualisation for the proportions of choices from the children for the ‘Rich’ picture, depending on the sentence TYPE they heard (color lines) and the class/age GROUP they belong to (x-axis).

Figure 6:

Results on the richness scale depending on age/class group (MGT-C).

A model run with these variables (see full specifications in the Supplementary Materials) yields robust evidence for an overall effect of type for the IS versus FINV contrast ( β ˆ  = 0.58, P(β > 0) = 0.92, 95 % CrI = [–0.21, 1.41]), meaning that, independent of the children’s age, IS sentences are consistently more associated with the rich person than FINV sentences. The model also yields tentative evidence (trend) for a distinction between F and FINV sentences that would go in the opposite direction (F sentences less associated with the rich person overall, β ˆ  = –0.25, P(β < 0) = 0.73, 95 % CrI = [–1.05, 0.55]).

These associations are not constant over time however, and the model also yields evidence for TYPE*GROUP interactions for both the IS versus FINV (robust, β ˆ  = –0.15, P(β < 0) = 0.87, 95 % CrI = [–0.42, 0.11]) and F versus FINV (trend, β ˆ  = 0.08, P(β > 0) = 0.73, 95 % CrI = [–0.18, 0.34]) contrasts. Both correspond to the difference between FINV sentences and the other two types reducing over time (IS being less associated to richness and F slightly more as children age).

4.4.2 Results for the ‘education’ scale

Figure 7 provides a visualisation for the proportion of choices from the children for the ‘Educated’ picture (doctor, as opposed to the stereotypically less educated cashier), depending on the TYPE and GROUP variables.

Figure 7:

Results on the education scale depending on age/class group (MGT-C).

Here, the overall picture appears to be different from the previous one. The model run for this scale yields evidence for an effect of TYPE only for the IS sentences compared to the FINV ones ( β ˆ  = –0.48, P(β < 0) = 0.88, 95 % CrI = [–1.30, 0.33]). This is evidence for a constant distinction over time between the two types, and the absence of interaction with the GROUP variable further highlights that this difference stays consistent from the early stages on. There is no conclusive evidence for a difference between F and FINV sentences, either with or without taking age/class into account.

4.4.3 Results for the ‘hobbies’ scale

Figure 8 gives the results for the proportions of choices from the children for the ’books’ picture, depending on sentence TYPE and children GROUP.

Figure 8:

Results on the hobbies scale depending on age/class group (MGT-C).

Yet again, the overall picture is different from the other two, with the model run for this scale yielding some evidence (trend, β ˆ  = –0.10, P(β < 0) = 0.78, 95 % CrI = [–0.37, 0.17]) for a TYPE*GROUP interaction for the IS versus FINV contrast only. There does not seem to be an overall difference between the three sentence types, but when looking at it more precisely, FINV sentences do in fact become more associated with books and reading than IS sentences when the children become older.

4.4.4 Overall picture and discussion

Figure 9 compiles the results for all scales in the child experiment (on the left), with a screenshot of all associations between linguistic forms (color lines) and social stereotypes (x-axis) over all 4 developmental stages (1 cell per group), and compares it to the same three scales from the experiment with adults (on the right, but it should be noted that the dependent variable and scale is not the same). This allows for a more intuitive comparison of the network of associations between forms and social meanings in children, at four different stages of development, to the same network in adult speakers.

Figure 9:

Overall results from the child matched-guise (left) compared to the adult experiment (right).

A first observation is that the clear-cut pattern exhibited by adults, with in situ and fronting with inversion being really differentiated, is seen in older children but not in younger ones. This supports our first predictions (1 and 2), with a progressive build-up of the pattern seen in adults as children get older. It should be noted however that, although older children do indeed make a distinction between fronting with inversion and in situ, and if the associations between these forms and social meanings go in the same directions as in adults (i.e. fronting with inversion is more associated with educated persons, doctors, than with cashiers, and more associated with books than with sports, compared to in situ), the simple fronting case is different. In older children (ages 9–11) it behaves very similarly to fronting with inversion, while in younger ones the two variants are substantially different. One explanation might be that both variants are actually not very prevalent in spontaneous adult French, thus making them both marked enough from a sociolinguistic perspective that children perceive them as ‘apart’ from the in situ variant, which is much more frequent (more than 60 % uses on adult corpus data). It is possible that the adult-like pattern solidifies after the age of 12, when children are confronted with more situations which allow them to better distinguish between the situational uses of both marked variants.

A second observation is that all three associations between form and social meaning, i.e. the three scales (richness, hobbies, education), behave in very different ways from a developmental perspective. On the richness scale, whatever differences between variants are seen in younger children disappear over time; on the hobbies scale the difference observed in adult data (fronting with inversion more associated with books) only appears in older children; and on the education scale the degree of association between fronting with inversion and the doctor/educated person (similar to that of the fronting variant, but greater than for the in situ variant) appears very early on and remains constant over time. This supports our third prediction regarding how different form-social meaning associations might follow different developmental paths because of how children may view the world very differently at different stages of their upbringing. Nearly all children go to the doctor from their first years on, which, according to a usage-oriented approach à la Buson (2009) and Buson and Billiez (2013), they can make use of very early on to building up a repertoire of many exemplars where stereotypical interactions with the physician involve interrogatives with fronting (and maybe a verb-subject inversion) such as Comment vas-tu ? or Où as-tu mal ? (‘How are you? Where does it hurt?’). For the hobbies scales, children learn how to read over the years, and it is through this progressive access to literature that they can access more complex books where they are exposed to sentences with fronting with inversion variants. For the richness scale, children get more knowledge of how the world operates as years go by and, from very limited interview data (post-session discussions with the children where the experimenter asked them what kind of person they had in mind when asked about ‘rich individuals’ they knew), we observed that young children conceive richness as mostly related to fame (with answers along the lines of ‘that football player/that famous singer I saw on TV the other day’). They do not really relate it to social prestige as an adult internalizes it, hence the disconnect between the two scales in children, which was not seen in adults. Furthermore, for the most part, children did not treat the education scale as a measure of what adults would consider education, and the link between years of education and professions like doctor versus cashier is not apparent to them. So the results on that scale should probably be thought of in terms of what situations the two drawings (cashier vs. doctor) embodied: going shopping and talking to someone behind the register on one hand, and going to the doctor on the other hand.