Glossa Psycholinguistics

2.1 A wide range of alternatives for asking wh-questions in French

Following Beyssade and Marandin (2006), Ginzburg and Sag (2000), and Marandin (2018), interrogation is a speech act that allows a speaker to ask about a missing piece of information. This missing piece of information can be the truth-value of a proposition (polar question with, roughly speaking, a yes or no answer, like : Did you do this thing?). This missing piece of information can also, and more specifically, be an unknown element of a set upon which a given predication is stated (wh- or partial question: Someone took my apple and ate it. Who did this?). Without going into a much more detailed semantic analysis of interrogation as a whole (for foundational work, see Hamblin, 1973; Karttunen, 1977), we will focus on the coexistence of multiple syntactic variants to ask such wh-interrogatives in French.

French wh-interrogatives are an already well-documented point of linguistic variation, with a wide range of corpus studies starting from the mid–20^th century (Adli, 2015; Ashby, 1977; Behnstedt, 1973; Coveney, 2011; Hamlaoui, 2009; Pohl & Straka, 1965; Reinhardt, 2019a; Söll, 1983; Terry, 1970; Thiberge et al., 2021). The examples in (1) are some of the main forms available to French speakers when asking a wh-question. Wh-questions like (1a) will be described as in situ (IS), (1b) as fronting (F), (1c) as fronting with inversion (FINV), (1d) as fronting with est-ce que (FESK), (1e) as clefting, and (1f) as fronting with complementizer.

Corpus-oriented research has highlighted diachronic change in the use of the main question types, with the in-situ variant becoming more and more frequent in French in recent years (totalling more than 50% of productions around the year 2000), and fronting with inversion being used less and less (totalling less than 10% of productions) (Adli, 2015; Hamlaoui, 2009; Thiberge et al., 2021). This, however, contradicts what is still found in reference grammars for French such as Riegel et al. (2014) (i.e., grammars used to develop teaching programs for L1 French children in school curricula), where the default interrogative variant is said to be fronting with inversion while in-situ sentences and all other variants are deemed informal varieties, which are purportedly used in spoken French only. In total, more than 15 pages are dedicated to French interrogative strategies in this book, but so-called “non-standard” variants (i.e., all variants but fronting with inversion) are described in a mere 10 lines of text. This is not just a niche phenomenon being exaggerated; rather, it must be considered in relation to the proportions of variants found in textbooks intended, for example, for L1 learners of the French language, where inversion is highly present (at least 40%) (Reinhardt, 2019b). In these textbooks, inversion is often presented as the “standard” way of forming interrogatives for children learning how to best use their own language.

The reasoning with respect to the use of in-situ variants, in particular, that is presented in Riegel et al. (2014) is that inversion is more complex and, therefore, disfavored when speaking, with a preference for the canonical word order of declarative sentences: Subject-Verb-Object.¹ The question of how internally “complex” each variant is, and of whether inversion (VSO order) is indeed more complex than non-inversion (SVO), has been addressed in many syntactic works. A common analysis is that the canonical word order (in-situ type) is the base order of constituents, from which all other types are derived. Generative grammars propose a “movement” explanation (Chomsky, 1981; Donati, 2006; Falk, 2012; Kayne, 1981; Ross, 1967; Stockwell, 1977). In this perspective, the interrogative element moves from its original position – leaving a trace – into the sentence’s left periphery (specifier of the CP, complementizer phrase), which is assumed to lead to increased complexity of the fronted question types compared to in situ, and even more so when a verb-subject inversion occurs. Other frameworks such as HPSG (Sag, 2010) or unification grammars (Abeillé, 2007), have described the phenomenon without movement, while still representing the non-locality in different ways that may or may not add to processing complexity. Construction grammar (e.g., Goldberg, 1995; Van Trijp, 2014) and, arguably, Combinatory Categorial Grammar (Steedman & Baldridge, 2011) can represent non-local dependencies without necessarily adding cognitive complexity, and, in this view, frequency alone could, for example, explain the use of the in-situ question type. The increased derivational complexity of fronted types postulated by some theoretical frameworks (highest for fronting with inversion; see Jakubowicz, 2011) and the “simplification” that the in-situ type represents in Riegel et al. (2014) are, we believe, relevant here. The form put forth as “standard” could be argued to be cognitively more demanding, which would make it more prestigious to master and use.

2.2 Using acceptability judgments to target the discrepancy between so-called standard French and corpus frequency of the variants

To shed light on the usage of the different interrogative variants, Thiberge et al. (2021) analyzed 617 French wh-interrogatives that were extracted from the ESLO project (Eshkol-Taravella et al., 2011; LLL, 2017). The ESLO corpus gathers around 5 million tokens, for more than 400 hours of recording in a wide variety of contexts (from scripted interviews to primary school classes, public conferences, or street recordings), with two time periods of data collection: one in the 1960s and the other around 2010. A key finding, among others, was that in the more recent period (second period of data collection), speakers aged 35–55 used fewer in-situ sentences than speakers aged 15–25 (~ 55% of productions vs. ~ 70% of productions). Moreover, a very marginal proportion of interrogatives were fronting-with-inversion sentences (less than 9% for both age groups).

This finding was modulated by the context of the recording: the difference between age groups, while visible when comparing interviews of L1 speakers by a researcher (46% of in-situ sentences for 35–55 y.o. speakers, ~60% for 15–25 y.o.), disappears in recordings conducted during family gatherings at meal-time (72% vs. 69% of in-situ sentences, respectively). These two contexts differ with respect to the presence of a researcher in the room during the recording, and also with respect to the topics that are discussed. During interviews, the researcher is asking questions following a scripted questionnaire relating to the personal lives of the interviewees, while during mealtimes, the topics cover daily activities of all members of the family. The formality of the social situation is clearly different between contexts, and we argue that this is a key element to understanding the different linguistic behaviors across age groups: all speakers use fewer in-situ variants in formal contexts (interviews) than in informal contexts (mealtime gatherings), but speakers aged 35–55 seem more prone to adapt their productions to the change in formality (as a reviewer points out, the reverse view might hold, with younger speakers being less sensitive to formality differences and seeing more situations as informal than older speakers). In a view where normative pressure is taken into account, this is consistent with sociolinguistic work on age-grading (Wagner, 2012), according to which this age group is most likely to be affected by linguistic prescriptivism as a consequence of work-related needs to appear well educated and literate.

All in all, recent corpus data show that only a small proportion of the interrogatives found in modern corpora are fronting-with-inversion constructions while in-situ sentences account for around 50% of the productions, or even more depending on the context. This contrasts sharply with the variant that is promoted as the “standard” way of speaking French (i.e., fronting with inversion) in reference grammars of French (e.g., Riegel et al., 2014). Moreover, this discrepancy between corpus data and the prescriptive linguistic norm makes French wh-interrogatives a particularly interesting test case. Acceptability is typically correlated with the probability of the occurrence of an utterance. However, depending on the social situation and the speaker, this prediction may not hold here (Lau et al., 2017).

The experiments we report below shed light on different factors relevant to the acceptability of linguistic variants. Although we acknowledge that reference grammars are far from being a consistently useful tool for linguists, they are used for L1 and L2 teaching and provide insights into what is considered standard in a given language. In fact, by exploring the issue of why a reference grammar would describe a language (here, French) in a view so inconsistent with what is observed in corpus data,² we provide evidence that L1 speakers have the competence to apply a more variable take on the notion of overall acceptability, which is readily available when given the proper tools to express it. Using the acceptability judgment paradigm, we focus on the issue of how speakers perceive and use linguistic variants when, prescriptively, one variant is considered “standard”. With three acceptability judgment tasks, we assessed the preferences of French L1 speakers with regard to the four main variants of interrogatives: in situ, fronting, fronting with inversion, fronting with est-ce que. This approach will also contribute to resolving a dearth of experimental data on the use of interrogative variants in contemporary French. It builds on recent fine-grained corpus work on the usage of French wh-interrogatives in different contexts of interaction (Thiberge et al., 2021).³

Based on the available data, we anticipate that one-dimensional acceptability judgments will translate to higher ratings for the normative standard variant, that is to say fronting-with-inversion-type interrogatives, if the relationship between the prescriptive norm and acceptability is deeply entrenched in the speaker’s mind. If, on the contrary, acceptability assessments correlate with frequency (in a radical approach to frequency, such as Culicover et al. (2022), but see Bader and Häussler (2010) for a more nuanced account, as mentioned above), we would expect the most common form in contemporary French, the in-situ-type interrogatives, to yield higher results. This is what we aimed for in the first task (Section 3), where we contrasted fronting-with-inversion, fronting, and in-situ interrogatives, and where participants only had one general acceptability scale to provide their assessment.

With two other acceptability judgment tasks (Section 4), we aimed at further distinguishing between what “good French” and “suitable French” are for L1 speakers, which we extended to the fronting with est-ce que variant. For this, we used two different scales. The first scale targeted the normative well-formedness of the variants: i.e., “how does a sentence reflect what the ‘rules’ say French should be?”. The other scale targeted the suitability of the variants in a (short) context we provided and manipulated: i.e., “is it okay to use this particular sentence given the social situation?”, in line with Labov et al.’s (2011) sociolinguistic monitor, which assesses the incongruity of linguistic variants in socially situated interactions. We predicted that, with these two different scales, the judgments would better reflect how the linguistic norm weighs on the participants’ preferences. More specifically, we predicted that with these dichotomized tools, the general acceptability pattern exhibited in the first experiment would project onto the well-formedness scale, which corresponds to the prescriptive view of what the standard way of speaking French is or should be, while a more nuanced pattern would be seen on the suitability scale, which better fits with patterns seen in corpus data. As such, fronting-with-inversion sentences should yield higher ratings on the first scale, compared to other variants, while in-situ sentences should yield higher ratings than fronted variants on the second scale, with modulations from external factors such as context (manipulated independent variable) and age group.

3.1 Notes on statistical analyses

With respect to the inferential statistics, for the statistical analyses of all experiments, we ran Bayesian regressions. A more detailed explanation can be found in the Supplementary Materials, but the main motivation for this is that we wanted to stay as close to the data as possible (thus, not working on a null hypothesis to be falsified, as is done in frequentist modeling.) (See Sorensen et al. (2016) for a more detailed view on the advantages of Bayesian statistics for linguists, psychologists and the cognitive sciences in general.) Bayesian cumulative link models were run for each experiment (see the Supplementary Materials for full specification and outputs), to adequately assess the link between our dependent variable (ordinal ratings, in all cases) and our independent variables (type, formality, and age). Maximal models were run with a full-fledged random-effects structure, integrating participants and items as potential sources of variability in the data as well as random slopes (Barr et al., 2013). All independent variables were mean-center coded to allow a more direct interpretation of the model estimates (Brehm & Alday, 2022). 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).

Finally, a crucial part of the Bayesian analysis lies in how credible and how probable it is that the estimates () are higher or lower than 0 (= no effect when looking at the difference in conditions). Since there is no official consensus on what the thresholds should be in Bayesian inferential analyses, and since a high inter-participant variability can be expected with regard to how normative pressure influences acceptability judgments, we decided to take slightly larger thresholds for the interpretation of the output of our models. As such, and inspired by, a.o., Burnett and Pozniak (2021) and Granlund et al. (2019), we will report probabilities (P(β < 0) or P(β > 0)) that are ⩾ .90 as reliably indicative of an effect of the independent variables on the dependent variables (“robust” evidence for an effect) and probabilities that are ⩾ .80 simply as evidence for an effect. Probabilities lower than .80 will be taken as not supporting the existence of an effect.⁵ For each relevant estimate, we will also report 95% Credibility Intervals (95%CrI), which are the values of the posterior distribution calculated by the model between which there is a 95%-chance that the true value for the estimated effect lies.

3.2 Materials, protocol and participants

For this 3 × 2 design, we constructed 30⁶ fictitious dialogues between a person A and a person B, all of which followed the same pattern: person A sets a very short illocutionary context varying in formality, then person B asks a wh-question relevant in this context. Context formality was modulated by three elements, two of which were combined in each A-sentence: dislocation of an element in the sentence (Jean arrive demain vs. Eh, Jean il arrive demain.), colloquial/neutral lexical alternatives (mon frère vs. mon frangin), and absence/presence of a full-fledged negation (Personne ne m’a rien donné vs. Personne m’a rien donné). We created these alternating contexts⁷ to mimic the contrasting contexts examined by Thiberge et al. (2021). In the first context, which we call informal (corresponding to family interactions in the ESLO corpus), there is minimal social distance between interlocutors. This is achieved through the use of colloquial or low-register nouns (e.g., frangin for frère), the breaking of prescriptive norms (such as the absence of full-fledged negation), and orality markers (like subject or object doubling, which is typically avoided in standard French). In this context, we anticipated that non-standard interrogative variants would be less incongruent, as per the sociolinguistic monitor model proposed by Labov et al. (2011). By contrast, in the second context, which we call formal (corresponding to interviews with researchers in the ESLO corpus), no proximity markers are present. Instead, neutral nouns, full-fledged negation, and the absence of doubling create the right conditions for greater stigmatization of the non-standard variants and, conversely, greater adequacy of the standard default form. Table 1 shows an item in all its conditions.

The 30 target dialogues were mixed with 30 filler dialogues with the same A-B structure, containing no wh-interrogative and no variation in formality. We opted for a 1:1 ratio for fillers and target items, so as to not lengthen the experiment and to avoid satiation effects. After 3 practice items, the 60 items were presented in a Latin Square design with full randomization. The experiment was carried out on the Internet via the IbexFarm platform (Drummond, 2016). After a short introduction to the participants’ rights, a description of the procedure, and a short questionnaire about their linguistic profile, explicit consent was required to proceed. The participants were then presented with the 63 (practice+target+filler) items. Below each sentence, participants rated the acceptability of the B-sentence on a 0–10 numeric scale, with each point represented in a numbered box (0-1-2..8-9-10). Acceptability was here defined as both the general understandability of the sentence and its well-formedness, given the short context provided. Once a participant clicked on the scale or entered a digit on their keyboard, the next item was presented. There was no time limit set for each trial, and the whole experiment was estimated to take roughly 15 minutes. Participants were informed that there was no good or bad decision and that we were only interested in their intuitions.

Participants were recruited via the RISC network (CNRS, UMR 3352) and social networks. 44 self-identified L1 speakers of French took part (age: 19–80, mean = 39.8, median = 27.5). The data for target items amounted to 1320 (30*44) ratings.

3.3 Results

Figure 1 presents a violin plot of the answers for the three type conditions, and shows the general pattern of results (error bars = standard error; white lines = quartiles).⁸

Figure 1: Mean acceptability and distributions of ratings from Experiment 1.

Fronting-with-inversion-type interrogatives generally received higher ratings than fronting-type and in-situ-type interrogatives. The Bayesian cumulative-link modeling of the data (see Figure 4 below for more, and model mb1 in the Supplementary Materials) yields robust evidence for an effect of the type predictor here, both for the fronting vs. fronting-with-inversion contrast ( = –2.01, P(β < 0) = 1, 95%CrI = [–2.91,–1.15]) and for the in-situ vs. fronting-with-inversion contrast ( = –1.52, P(β < 0) = 1, 95%CrI = [–2.27,–0.81]).

Figure 2 details the results depending on the two levels of the manipulated formality variable. Here, the overall pattern distinguishing the three syntactic types is similar across conditions, but the contrast between fronting-with-inversion sentences and the other two types is larger in the formal condition than in the informal condition.

Figure 2: Mean acceptability and distributions of ratings from Experiment 1, by context.

While ratings associated with F-type and in-situ-type sentences stay roughly equivalent between contexts, fronting-with-inversion-type sentences are associated with higher ratings in the formal condition than in the informal one. The Bayesian model yields evidence for a type*formality interaction for the fronting vs. fronting-with-inversion contrast ( = –0.37, P(β < 0) = 0.87, 95%CrI = [–1.02,0.29]) and robust evidence for the in-situ vs. fronting-with-inversion contrast ( = –0.45, P(β < 0) = 0.93, 95%CrI = [–1.06,0.17]).

Figure 3 details the results depending on the age group of participants. As an important note, we took the split into a +30 and a –30 group and not based on the median age. This follows from corpus studies, in which age group was a meaningful predictor of interrogative type production (Thiberge et al., 2021), but also more generally from the sociolinguistic literature. Absolute age is often an important predictor of attitudes toward linguistic variation in itself, because it correlates with many other socioeconomic circumstances or group affiliation. As mentioned above, this is an important element of age-grading phenomena: people falling into the age category where professional obligations heavily constrain a person’s relationship to the linguistic norm (Wagner, 2012). However, since the age variable was not controlled for, the groups are not fully balanced in number (N_+30y.o.=19/44). The overall contrast between the three syntactic types of interrogatives remains highly similar across age groups, but the contrast is larger between fronting-with-inversion sentences and the other two types for participants older than 30.

Figure 3: Mean acceptability and distributions of ratings from Experiment 1, by age group.

The difference between mean ratings associated with fronting vs. fronting-with-inversion sentences is larger in the +30 group for which the model yields evidence (for the type*age interaction, ( = –1.57, P(β < 0) = 0.97, 95%CrI = [–3.23,0.04])). This contrast is even larger for ratings associated with in-situ vs. fronting-with-inversion sentences, for which the model yields robust evidence (type*age interaction at ( = –1.67, P(β < 0) = 0.99, 95%CrI = [–3.09,–0.26])).

Figure 4 shows the posterior distributions for the maximal mb1 model fit to the data from Experiment 1. In Bayesian inferential statistics, posterior distributions are simulated for each predictor (parameter) specified in the model, which help characterize the influence of said predictor on the dependent variable. Each posterior distribution is described by its position relative to 0 (= no influence of the predictor on the dependent variable), its range (extreme values and credible intervals, within which the true value of the estimated coefficient lies), and its mean value (estimated coefficient, just as in frequentist models). As such, posterior distributions where the true value of the estimated coefficient has a high probability of being different than 0 give evidence for an effect of the predictor. In addition to the results described above, the model shows robust evidence for a main effect of formality ( = 0.26, P(β > 0) = 0.94, 95%CrI = [–0.08,0.60]), meaning that, overall, the ratings given by participants were higher in the formal than in the informal contexts. From the type*formality interactions described before, this follows from the higher ratings associated with fronting-with-inversion-type sentences in formal contexts.

Figure 4: Posterior distributions for the maximal model ran for Experiment 1.

Perhaps of greater interest is the type*formality*age interaction for fronting vs. fronting-with-inversion sentences ( = –1.04, P(β < 0) = 0.95, 95%CrI = [–2.33,0.26]). This means that there is robust evidence for a higher sensitivity to context variation among the +30 y.o. group: F interrogatives are rated even lower in formal contexts by this population than by the –30 y.o. participants, compared to fronting-with-inversion sentences.

3.4 Interim summary

Experiment 1 offers a deeper look into the role of context, based on the general acceptability of different wh-interrogative variants of French. A general pattern of acceptability, compatible with a normative take on what standard French should be, is confirmed here: fronting with inversion yielded higher ratings than in situ, and in situ yielded higher ratings than simple fronting. These results are compatible with earlier preliminary experimental work (ThiBerge, 2018) on this topic but they are mainly on a par with many prescriptive or reference grammars of French (e.g., Riegel et al., 2014), where the fronting-with-inversion variant is held as the standard for French.

With respect to formality, we found evidence for an influence of context on this pattern: in formal contexts, fronting with inversion yields higher ratings than in informal contexts. This is directly reminiscent of corpus findings, such as Thiberge et al. (2021), where the uses of each variant were different depending on the context (more fronting-with-inversion sentences in formal contexts with an interviewer or at school than at family meals). Preferences were, however, not reversed, with in-situ sentences becoming the preferred option. We also find evidence for an effect of age, with participants older than 30 years being more sensitive to the difference between variants overall, and giving lower ratings than participants younger than 30 to the two prescriptively “non-standard” variants (simple fronting and in situ). A difference in linguistic production between participants aged 15–25 vs. participants aged 35–55 was also observed for the wh-interrogative variants in the ESLO corpus.

Also in line with these corpus findings, we find in Experiment 1 an interaction between formality and the age group of participants. When compared with participants younger than 30, participants older than 30 seem to more strongly contrast some variants (simple fronting and fronting with inversion) when the context is formal than when it is informal. This could come from a different approach to the linguistic norm as a function of age, and possibly as a function of professional circumstances. Figure 5 illustrates how this could work: when considering age as a continuous variable, the in-situ variant seems more acceptable in formal⁹ contexts for participants younger than 30 and older than 60–70. There is no such phenomenon for the simple fronting variant, which becomes less and less acceptable with increasing age; hence, the type*formality*age interaction described above. This U-shaped curve for in-situ sentences (also visible for fronting-with-inversion sentences, but reversed and to a lesser degree) could be linked to the notion of age-grading, as proposed in sociolinguistic works such as Wagner (2012). Speakers aged between 30 and 60/70 are confronted with the necessity of abiding by linguistic prescriptive norms in their everyday professional lives (i.e., “fronting-with-inversion type is the standard variant”), and this may be reflected in their acceptability ratings. The age effect found in this first experiment is also a reason for not overestimating the representativity of studies that generally include a majority of young adults (university students) as participants, in particular, when normativity may play a role.

Figure 5: Mean acceptability ratings from Experiment 1, by age (continuous).

4.1 Materials, protocol and participants (Experiment 2)

With the exact same items as in Experiment 1 (practice, targets and fillers), we investigated the acceptability of different interrogative types with three main changes. First, we changed the modality of stimulus presentation. Experiment 1 was a written text-based experiment, with participants reading sentences (dialogues between person A and B) and then having all the time they needed to develop a judgment and provide it on a unique response scale. In Experiment 2, we opted for an auditory presentation, with the reasoning that it made for a more ecological setting, more consistent with the everyday interactions found in spoken corpora (e.g., ESLO). Instead of the written dialogue, participants saw an audio player that they had to trigger to hear the stimuli. They could replay them as often as they wished.

The second part we adjusted was the response scale, which this time was divided into two different questions:

• Is this sentence “good French”? (Est-ce que c’est du “bon français”?)

• Is this sentence “suitable French”? (Est-ce que c’est du “français adapté”?)

Both sub-scales were defined in the previously displayed instructions as, respectively, “well-formed French” and “you could speak like this in this context”.

A third adjustment relates to the number of response options on the scales. Experiment 1 was run with a 11-point scale (0–10), while this time the scales were both 6-point (1–6) scales. We chose this option for two reasons: Since participants had to provide judgments on two scales, we wanted to make their task easier by providing fewer options. We also wanted to exclude a neutral position to force participants to make a choice. From previous experience, we did not expect these changes to systematically affect our results. Answers were not mandatory, so participants who felt unable to decide on a scale could skip it for a particular item and continue with the experiment.

The auditory stimuli were recorded with 4 adult French L1 speakers (F1, F2, M1, M2), whom we recorded in a soundproof room. During the recording phase, speakers were instructed to go for “reserved naturalness”, without putting too much expressivity in their voice, so as to minimize prosodic contrasts across, and even within, the sounds played by participants in the actual experiment. They read all possible combinations of contexts and sentence types for all items, and we selected the recordings that we judged most “neutral” in terms of retrievable social cues. All sentences in all conditions were recorded and then manipulated in Audacity (AudacityTeam, 1999–2021) to create dialogues absolutely parallel to the written stimuli of Experiment 1. We crossed voices so that all 12 combinations (F1-F2, F1-M1, F1-M2, F2-F1, F2-M1, F2-M2, M1-F1, M1-F2, M1-M2, M2-F1, M2-F2, M2-M1) were presented for all items and all participants of Experiment 2. We also carried out a norming study to assess the “a priori gender” people generally assigned to voices (so as to minimize potential rating differences in terms of how women or men are supposed to speak to each other¹⁰), and to check that no voice was more associated with a colloquial tone than others, so as to not bias the formality variable. Data from this norming study can be found in the Supplementary Materials.

64 self-identified L1 speakers of French took part in total, both via the Internet and in the lab. They were recruited by the RISC network and social networks. Due to some technical issues with the presentation of some audio stimuli for some participants (usually a sound that could not be played by the server), we only kept answers from 54 participants who saw and answered at least 90% of all practice + filler + target items on both scales (age: 19–70, mean = 35.8, median = 30). This amounts to 1603 observations on each scale for target items.

4.2 Results (Experiment 2)

We analyzed data from Experiment 2 in the same manner as Experiment 1 (see 3.1 and the Supplementary Materials on OSF). However, this time we ran two models: one (mb2g) for the first scale, with “well-formedness” ratings as the dependent variable, and one (mb2s) for the second scale, with “suitability to the context” ratings as the dependent variable.

Figure 6 shows a pattern similar to that seen in Experiment 1. On the first scale, fronting-with-inversion-type interrogatives generally received higher ratings than both F-type and in-situ-type interrogatives. The Bayesian cumulative-link modeling of the data (see Figure 12 at the end of the section for more, and model mb2g in the Supplementary Materials) yields robust evidence for an effect of the type predictor here, for both the fronting vs. fronting-with-inversion contrast ( = –3.16, P(β < 0) = 1, 95%CrI = [–3.92,–2.42]) and the in-situ vs. fronting-with-inversion contrast ( = –2.81, P(β < 0) = 1), 95%CrI = [–3.56,–2.11]).

Figure 6: Mean acceptability ratings from Experiment 2 (good French).

However, Figure 7 displays a different pattern. On the second scale, fronting-with-inversion-type interrogatives received ratings similar to F-type interrogatives, and in-situ-type interrogatives yielded higher ratings overall. The Bayesian cumulative-link modeling of the data (see Figure 13 below, and model mb2s in the Supplementary Materials) yields robust evidence for an effect of the type predictor for the in-situ vs. fronting-with-inversion contrast only ( = 0.45, P(β > 0) = 0.97, 95%CrI = [–0.03,0.95]).

Figure 7: Mean acceptability ratings from Experiment 2 (suitable French).

For formality, the two scales also appear to capture different dimensions of the preferences of French speakers. For the “good French” scale, the effects of formality appear to be very similar to those observed in Experiment 1 (Figure 8): fronting-with-inversion sentences are rated higher than both other types, overall, and the contrast is larger in formal contexts (evidence for the fronting vs. fronting-with-inversion contrast ( = –0.35, P(β < 0) = 0.83, 95%CrI = [–1.08,0.36]). The pattern is different on the “suitability” scale (Figure 9), with the three variants yielding ratings around 5/6 on this scale, whatever the context. However, when looking at the results in more detail, the type yielding the highest ratings in the formal contexts is in-situ, when fronting and fronting-with-inversion sentences yield roughly similar ratings, and in the informal contexts, fronting-with-inversion sentences yield the lowest ratings. The model run for this scale yields robust evidence that the fronting vs. fronting-with-inversion contrast is reduced in the formal contexts (type*formality interaction, = –0.42, P(β < 0) = 0.90), 95%CrI = [–1.08,0.24]), and also evidence with respect to the in-situ vs. fronting-with-inversion contrast (type*formality interaction, = –0.46, P(β < 0) = 0.87, 95%CrI = [–1.29,0.36]). One could wonder why judgments on the “good French” scale would be influenced by context formality at all, if all it captures is the participants’ knowledge of prescriptive norms. One possible answer is that these norms allow for some degree of style-shifting, as a reviewer suggests. Another perspective on this would be that formal contexts can sometimes reinforce the perception of well-formedness of the standard form when it is used, with a near ceiling effect in that condition for the fronting-with-inversion type (mean rating of 5.66/6 and a smaller standard deviation than all other type*context combinations, on this scale and on the other scale).

Figure 8: Mean acceptability ratings from Experiment 2, by context (good French).

Figure 9: Mean acceptability ratings from Experiment 2, by context (suitable French).

For the age variable and on the “well-formedness” scale, the overall pattern does not change between age groups, but participants younger than 30 gave higher ratings to fronting-with-inversion sentences, compared to older participants (Figure 10). This results in evidence for a type*age interaction for the in-situ vs. fronting-with-inversion contrast ( = 0.64, P(β > 0) = 0.82, 95%CrI = [–0.78,2.06]). On the “suitability scale” (Figure 11), participants younger than 30 exhibit a stronger dispreference for fronting-with-inversion sentences (rated lowest) than participants older than 30 (who rated fronting-with-inversion types about the same as fronting types, but still lower than in-situ types). This is captured by robust evidence for type*age interactions for both the fronting vs. fronting-with-inversion ( = –0.67, P(β < 0) = 0.96, 95%CrI = [–1.45,0.10]) and the in-situ vs. fronting-with-inversion ( = –0.80, P(β < 0) = 0.96, 95%CrI = [–1.75,0.13]) contrasts.

Figure 10: Mean acceptability ratings from Experiment 2, by age group (good French).

Figure 11: Mean acceptability ratings from Experiment 2, by age group (suitable French).

Figures 12 and 13 show the posterior distributions for the two models run for both scales of Experiment 2.

Figure 12: Posterior distributions for the maximal model ran for Experiment 2 (good French).

Figure 13: Posterior distributions for the maximal model ran for Experiment 2 (suitable French).

4.3 Interim summary

Contrary to Experiment 1, there seems to be no 3-way interaction type*formality*age for either the fronting vs. fronting-with-inversion or the in-situ vs. fronting-with-inversion contrasts on either scale. Presenting participants with explicit scales where they could disentangle their judgments on well-formedness and suitability for a specific context (in terms of formality) seems to have neutralized, in some way, the differences between age groups with regard to formality assessments and, more precisely, with regard to preferences for some variants in different contexts.

Overall, Experiment 2 provides no evidence for an influence of the number of response options on the scales, as seen from the very similar patterns on the well-formedness scale here and on the unique scale used in Experiment 1. On this scale, fronting-with-inversion sentences are preferred to both fronting and in-situ sentences, and this preference is modulated by context formality.

More importantly, this experiment provides evidence for a differentiated approach by participants to the concept of acceptability. Well-formedness does not follow frequency of use in corpus data, and corpus frequency does not correlate with well-formedness alone. The pattern on the first scale (“good/proper French”) is close to what can be observed in a general acceptability task (i.e., Experiment 1) and closely reflects what can be found in a reference grammar of the language. However, this scale alone would miss participants’ ability to select which variant would best suit a “real-life” everyday context of interaction. This is more clearly captured by the second scale (“suitability”), which participants seem to have rightly interpreted as “Would I, myself, use this sentence in this specific situation?”

Interestingly, the type*age interactions are reduced on the well-formedness scale. This could be construed as evidence that participants have a normative view of what proper French should be (with regards to at least, but most probably not only, wh-interrogation). These interactions are still present on the second scale, which would indicate that not all age groups consider all variants as “suitable” for a given context. Although there is no evidence for a three-way interaction with the formality predictor in Experiment 2, this is still reminiscent of Thiberge et al.’s (2021) results, where different age groups use different variants in different contexts. Further exploration of what makes two contexts different and of how different age groups define and interpret formality would be useful for future work.

4.4 A quick look at the est-ce que wh-interrogatives (Experiment 3)

Experiment 2 gave a more fine-grained take on the acceptability of the three main variants available to French speakers for asking a wh-question. Like Experiment 1 however, it ignored the FESK variant, where an est-ce que sequence is introduced between the fronted wh-element and the Subject-Verb sequence. This variant is actually quite frequently used in everyday speech (about as much as the fronting-with-inversion variant in ESLO data, for example). At the same time, it does not seem much more complex than simple fronting, since it does not involve Verb-Subject inversion. We conducted a third experiment on the basis of the paradigm applied in Experiment 2 (two different 11-point slider scales, one for well-formedness and the other for suitability in a context, with written stimuli, however) to pinpoint how this variant compares to the others, by comparing it to in-situ and fronting-with-inversion sentences. The details of this experiment (participants, materials and detailed analyses) are presented in the Supplementary Materials, and, for the sake of clarity, we will only report the main results here, to provide a more exhaustive picture of the alternation phenomenon with respect to interrogative variants in French.

Figure 14 shows a pattern quite similar to that seen in Experiments 1 and 2. On the first scale, fronting-with-inversion-type interrogatives generally received higher ratings than both FESK-type and in-situ-type interrogatives. Interestingly, FESK interrogatives appear in an intermediate position between fronting-with-inversion and in-situ sentences.

Figure 14: Mean acceptability ratings from Experiment 3 (good French).

Figure 15 displays yet a different pattern. On the second scale, fronting-with-inversion-type interrogatives received ratings similar to in-situ-type interrogatives, and F-type interrogatives yielded lower ratings overall. Although they are rated “better French” than in-situ sentences, FESK sentences generally seem not to be suitable for many contexts. It can also be noted that there is no difference here between fronting-with-inversion and in-situ sentences in terms of suitability, which contrasts with what was observed in Experiment 2. A possible explanation could be that the presence of FESK sentences in the stimuli made the difference between fronting-with-inversion and in-situ less salient to participants.

Figure 15: Mean acceptability ratings from Experiment 3 (suitable French).

For formality, and on the “good French” scale (Figure 16), the pattern appears similar again to that observed in the previous experiments: fronting-with-inversion sentences are rated higher than both other types, overall. However, this time, there is no evidence of an interaction type*formality, for both the FESK vs. fronting-with-inversion and the in-situ vs. fronting-with-inversion contrasts. The reason may be that since FESK sentences are in an intermediate position between fronting-with-inversion and in-situ sentences, participants are less sensitive to the differences between contexts when they consider the overall “well-formedness” of the three variants, which differs from Experiment 2 where fronting and in-situ sentences were behaving in a similar way, for example.

Figure 16: Mean acceptability ratings from Experiment 3, by context (good French).

On the “suitability” scale (Figure 17), the pattern is different, with the three variants rated between 6.56/10 (FESK sentences in informal contexts) and 7.74/10 (fronting-with-inversion sentences in formal contexts), so less contrast than on the first scale. Fronting-with-inversion sentences yield higher ratings in formal contexts, and FESK sentences yield lower ratings in informal contexts, while in-situ sentences appear to not be sensitive to differences in formality.

Figure 17: Mean acceptability ratings from Experiment 3, by context (suitable French).

Also to be noted here, the fronting-with-inversion sentences receive similar ratings across contexts on the “good French” scale. On the “suitability” scale, however, the ratings are quite different for this type, which we take as further evidence that, although some style-shifting is allowed (see the results from Experiment 2 above), the first scale is actually more capturing participants’ knowledge of the linguistic norms.

As in Experiment 2, for the age variable and on the “well-formedness” scale, the general pattern does not change between age groups (Figure 18), but participants older than 30 make a greater distinction between fronting-with-inversion sentences and both other types, compared to younger participants. On the “suitability scale” (Figure 19), participants younger than 30 exhibit a slight dispreference for FESK sentences (rated lowest) compared to the two other types, but this difference is larger in participants older than 30.

Figure 18: Mean acceptability ratings from Experiment 3, by age group (good French).

Figure 19: Mean acceptability ratings from Experiment 3, by age group (suitable French).

In addition, there is evidence for a type*formality*age interaction for in-situ vs. fronting-with-inversion sentences ( = 0.42, P(β > 0) = 0.82, 95%CrI = [–0.48,1.32]), on the “well-formedness” scale. This means that participants older than 30 appear to have rated in-situ sentences a bit higher in formal contexts, when compared to fronting-with-inversion sentences, than participants younger than 30. Although this is surprising, because there is no evidence for such 3-way interactions on the “suitability” scale, this is, nonetheless, tentative evidence for different age groups evaluating different variants differently in different contexts.

When aggregating all these results from Experiment 3, it seems FESK interrogatives are in a somewhat intermediate position between fronting-with-inversion and in-situ sentences in terms of well-formedness, but, overall, they are not as suitable for formal contexts in particular. We also observe a different assessment of this variant in different contexts across age groups, which is in line with both previous experiments and corpus work. FESK seems to be a variant that works in any context without being specialized for any, thus making it fairly acceptable in general, but also generally less suitable than the other two variants in their preferred context.