2.1 The prosody of polar interrogatives

As in languages such as English, Dutch, and Italian (for a review, see Shriberg et al. Reference Shriberg, Stolcke and Jurafsky1998), French polar questions correlate with rising intonation (Safarova et al. Reference Safarova, Muller and Prevot2005), and questions are shorter than statements because of their shorter final vowels in particular (Torreira & Valtersson Reference Torreira and Valtersson2015; but see Smith Reference Smith2002). In addition, French questions have a more pronounced final intensity drop compared to statements (Rossi et al. Reference Rossi, Hirst, Di Cristo, Rossi, Di Cristo, Hirst, Martin and Nishinuma1981; Valtersson & Torreira Reference Valtersson and Torreira2014; Torreira & Valtersson Reference Torreira and Valtersson2015) and their pitch contours start at a higher register than those of statements (Torreira & Valtersson Reference Torreira and Valtersson2015). The automatic classification of questions and statements carried out by Torreira & Valtersson (Reference Torreira and Valtersson2015) reveals an increase in accuracy from 60–65% for most individual cues to almost 80% if all the phonetic cues we have just mentioned are present. Concerning the rising intonation associated with polar questions and continuation statements (H*H% in autosegmental terms), Valtersson & Torreira (Reference Valtersson and Torreira2014) propose that, while both SAs make use of a pattern of rising intonation, the phonetic realization of this pattern is conditioned by the interactive and communicative parameters of the contexts in which these utterances are produced.

French interrogatives with different pragmatic interpretations can be identified on the basis of their prosodic properties. For instance, Michelas et al. (Reference Michelas, Portes and Champagne-Lavau2015) have shown that, unlike unbiased questions such as in (6), questions that are uttered in a negatively biasing context have a f₀ peak in the penultimate syllable of the final word of the utterance (a noun, here guéridon).

Another interesting case study concerns rhetorical questions (Delais-Roussarie & Beyssade Reference Delais-Roussarie and Beyssade2019). These authors demonstrate that polar and Wh- interrogatives such as (8)–(9) respectively more often have a rising intonational contour and are spoken faster when they are used as information-seeking questions.Footnote ²

When they are used as rhetorical questions, these interrogatives more frequently exhibit a falling intonational contour.Footnote ³ These authors remark, however, that not all utterances of rhetorical vs. information-seeking questions contain these prosodic cues. According to them, the rise-fall contour typical of rhetorical questions is used by speakers to signal an implicit meaning (Beyssade & Marandin Reference Beyssade and Marandin2006), i.e., that nobody eats spinach in the case of (8)–(9).

2.2 The prosody of French non-inverted interrogatives

Delattre (Reference Delattre1966) identifies ten different patterns for French prosodic units, based on a combination of three acoustic features, i.e., f₀, intensity and duration. With respect to interrogative sentences, he distinguishes between two intonational profiles corresponding, on the one hand, to non-inverted interrogatives (what he ambiguously calls questions, which also seem to include polar interrogatives), such as (10), and Wh- interrogatives (what he calls interrogations) such as (11).

The intonational pattern for polar interrogatives is a rising f₀ curve (Fig. 1), while for Wh- interrogatives, it is characterized by a falling f₀ curve (Fig. 2).

Figure 1. Intonational pattern for polar interrogatives (Delattre Reference Delattre1966: 4)

Figure 2. Intonational pattern for Wh- interrogatives (Delattre Reference Delattre1966: 4)

Even though sentences produced with these intonational contours will certainly sound familiar to native speakers of French, their accuracy and representativeness have been criticized. For example, Di Cristo (Reference Di Cristo2016: 294) points out that a distinction between different intonational patterns in terms of morphosyntactic criteria is missing from Delattre’s approach, and that his repertoire of interrogative sentence-types is incomplete. Following Di Cristo, polar questions can be asked using declaratives with a rising intonation such as (12), subject-verb inverted interrogatives such as (13), interrogatives initiated by the Est-ce que (“is it the case that”) such as (14), with or without a linguistic negation.

He observes that, while some utterances can be understood as questions without having a final rising f₀ curve, the majority of questions in contemporary French are asked using non-inverted interrogatives, typically uttered with rising intonation. In fact, non-inverted interrogatives have overwhelmingly replaced inverted polar interrogatives in casual conversation (this is discussed by Coveney Reference Coveney2011, see e.g., Wall Reference Wall1985 for empirical evidence that inverted interrogatives are rare in casual conversation).

Di Cristo (Reference Di Cristo2016) does not discuss possible prosodic differences in the final contour between the use of a non-inverted modal interrogative such as (15) as a direct question (“I am asking you whether it is the case that you are passing me the keys”) or as an indirect request for action (“I am requesting that you pass me the keys”).

Some non-inverted modal interrogatives can also be used with the illocutionary force of a threat, as (16) illustrates, in which case, again, Di Cristo proposes that no departure from the final f₀ curve is necessary for the threat to be distinguished from the SA of request for information.

This example is particularly interesting, as it suggests that other prosodic features would be used by the speaker to indicate that the utterance is to be interpreted as a threat, e.g., a higher intensity and louder speech.

3.1 Method

For the data collection phase of our study, we recruited 14 French native speakers after posting an announcement on Facebook. Of these participants, 13 self-identified as female (one male). The average self-reported age was 23 (median = 23). Ten speakers originated from Belgium, and the remaining four from France. Each speaker received 7,5€ for participating.

Each speaker was given 24 utterances to read aloud and record (see Appendix 1 for a complete list).Footnote ⁴ Twelve of these had the modal interrogative form “Tu peux + verbal phrase (VP) ?” (”Can you (2^nd person singular) VP?”), e.g., “Tu peux ouvrir cette fenêtre ?” (“Can you open that window?”); these will be referred to as modal interrogatives. Twelve took the form of “My X is Y”, e.g., “Mon verre est vide” (“My glass is empty”); these utterances will be referred to as negative state remarks. Speakers were instructed to say each utterance twice – once as a request and once as a literal question or statement (counterbalanced for order). In order to make the recording process as naturalistic as possible, speakers were given a few words of explanation about a possible situation in which the utterance was produced. We provided them with (minimal) contextual information about the situations in which they could imagine producing the stimuli. Here is an example of contexts for the Tu peux ouvrir la fenêtre? (“Can you open the window?”) stimuli: a) As a question: Your friend broke his arm and you want to know whether he is able to open the window. (Votre ami s’est cassé le bras et vous voulez savoir s’il est capable d’ouvrir la fenêtre.) b) As a request: You feel too warm and you ask your friend to open the window. (Vous avez trop chaud et vous demandez à votre ami d’ouvrir la fenêtre.) They were also given examples at the beginning of the instruction sheet demonstrating how the same sentence could be used either as a request or as a statement (or question). The example sentences did not appear in the target stimuli, and the different versions of each utterance were not spoken aloud to the participants (so that they could not imitate the prosody of the experimenter). Two complementary lists of stimuli were created: in one list, participants were instructed to utter the first sentence as a request first, the second one as a literal SA first, and so on; in the other list, participants were instructed to utter the first sentence as a literal SA first, the second one as a request first, and so on. Speakers were randomly assigned to either list.

Due to the COVID-19 health situation at the time, the speakers could not perform any audio recordings at the university lab, and they were instead asked to make their recordings at home using the microphone of their laptop or phone. Detailed instructions regarding the recording process, including three training items, were provided to them, along with a list of utterances to produce and self-record. Each speaker was instructed to make their recordings in a quiet room with closed windows and door, sitting down, while remaining at a distance of approximately 30cm from their laptop microphone. They were also told to keep their voice intensity constant across the recording process. The recorded data was listened to and verified by the first author, and three speakers were asked to record a few utterances again.

3.2 Predictions

Based on available experimental data for indirect requests in English (Banuazizi & Creswell Reference Banuazizi and Creswell1999; Hedberg et al. Reference Hedberg, Sosa and Görgülü2014; Trott et al. Reference Trott, Reed, Ferreira and Bergen2019, Reference Trott, Reed, Kaliblotzky, Ferreira and Bergen2022), we focused on the seven utterance-level acoustic features identified by Trott et al. (Reference Trott, Reed, Ferreira and Bergen2019), i.e., mean f₀, f₀ range, standard deviation of f₀, duration, mean intensity, standard deviation of intensity, and slope of f₀. We extracted the mean f₀, the range of f₀, mean intensity, and the number of voiced frames (as a proxy for duration). We also included measures of dispersion for both pitch (standard deviation of f₀) and intensity (standard deviation of intensity).Footnote ⁵ Finally, because the non-request interpretation of our modal interrogatives is a yes-no question about the addressee’s ability, and because prototypical French questions have a low-rise pitch contour (e.g., Di Cristo Reference Di Cristo2016), we used the slope of the f₀ component (slope of regressing f₀ against time) to measure the degree to which an utterance exhibits a rising or falling contour (Roche et al. Reference Roche, Morgan, Fissel Brannick and Bryndel2019; Trott et al. Reference Trott, Reed, Ferreira and Bergen2019).

Following the results of Trott et al.’s (Reference Trott, Reed, Ferreira and Bergen2019; Reference Trott, Reed, Kaliblotzky, Ferreira and Bergen2022) studies, we predicted, across all utterances, i.e., for both interrogatives and declaratives, that items with higher mean f₀ would be more likely to be requests. We also predicted, across all utterances, that items with more positive f₀ slopes would be less likely to be requests. As we explained in the introduction, it is very difficult, on the basis of available experimental evidence, to determine whether prosody should play a more important role for one type of construction (e.g., the more conventionalized modal interrogatives) or for the other (the less conventionalized negative state remarks). Therefore, we will not make any specific predictions regarding the possible role of conventionalization on the production and interpretation of prosodic cues disambiguating between the IR and the ‘literal’ uses of modal interrogatives and negative state remarks.

We predicted that, for non-inverted modal interrogatives, requests would have a less positive f₀ slope than their question counterparts, given that the hallmark of information seeking questions in French is the rising intonation (low-rise pitch contour) (Di Cristo Reference Di Cristo2016). We also predicted, as Trott et al. (Reference Trott, Reed, Ferreira and Bergen2019; Reference Trott, Reed, Kaliblotzky, Ferreira and Bergen2022) did for English, that negative state remarks in French should have a longer duration (i.e., contain more voiced frames) when intended as requests than when intended as literal statements. That is, we expected speakers to “signal” a deviation from the default interpretation of these remarks as statements by putting the emphasis on specific words or syntactic constituents, which should be detectable in the number of voiced frames. By contrast, we approached the remaining acoustic features that could help distinguish between the literal and IR uses of our stimuli from a more exploratory perspective.

3.3 Data processing

For each of the 336 recordings (14 speakers producing 12 utterances with two versions each), we used Parselmouth (Jadoul et al. Reference Jadoul, Thompson and De Boer2018), a Python interface to Praat, to extract the seven acoustic features described above. We then z-scored each of these features with respect to each speaker’s mean and standard deviation for that particular feature, to account for considerable inter-speaker variability. The acoustic features and the original audio recordings will be made available on the Open Science Framework.Footnote ⁶

3.4 Results

3.4.1 Analysis of individual acoustic features

To analyse individual acoustic features, we first asked how much independent variance in Meaning (request vs. literal) was explained by each individual feature. To do this, we first constructed a full generalized linear mixed model with a logit link, with Meaning as a dependent variable, and seven acoustic features as predictors (along with random intercepts for each item). We compared this full model to a series of reduced models omitting each feature in turn; model comparisons were performed using log-likelihood ratio tests, and the p-values obtained from each comparison were adjusted for multiple comparisons using Holm-Bonferroni correction (Holm Reference Holm1979).

In each case, a positive coefficient represents a higher likelihood of a request meaning, while a negative coefficient represents a higher likelihood of a literal meaning. We first conducted these analyses across all utterances, then conducted subset analyses focusing on interrogatives and declaratives independently.

When predicting Meaning across interrogatives and declaratives (i.e., all utterances), mean f₀ and f₀ slope emerged as significant predictors after controlling for multiple comparisons [X ² (1) = 35.697, p < .001]. Specifically, items with higher mean f₀ were more likely to be requests (β = 0.92, SE = 0.17), and items with more positive f₀ slopes were less likely to be requests (β = −1.1, SE = 0.17).

For interrogative items only, model fit was improved by the inclusion of only f₀ slope [X ² (1) = 68.07, p < .001]. Consistent with our predictions and with past work (Trott et al. Reference Trott, Reed, Ferreira and Bergen2019), items with more positive f₀ slopes were less likely to be requests [β = −1.57, SE = 0.25] (see the example sentences in Figure 3).

Figure 3. F₀ curves for Tu peux bouger cette armoire ? as a literal question (left) and as a request (right)

Finally, for declarative items only, model fit was improved by mean f₀ (X ² (1) = 28.36, p < .001). As predicted, utterances with higher mean f₀ were more likely to be requests [β = 1.62, SE = 0.35] (see the example sentence in Fig. 4).

Figure 4. F₀ curves for Mon verre est vide as a literal statement (left) and as a request (right)

Model fit was also improved by the number of voiced frames (X ² (1) = 7.53, p = .04]. Consistent with Trott et al. (Reference Trott, Reed, Ferreira and Bergen2019), longer utterances were more likely to be requests [β = 0.56, SE = 0.21].

3.4.2 Machine learning classifier

While the analysis above indicates which features are informative about the meaning of the recorded utterances, it does not specify how much information a particular feature contains, in particular when all seven features are combined.

Building on Trott et al. (Reference Trott, Reed, Ferreira and Bergen2019), one way to address this question is to ask how accurately a classifier can predict Meaning from all seven features combined. We used leave-one-out cross-validation (LOOCV) to quantify the ability of a classifier to generalize from its training set to novel samples. That is, a model was fitted to every item in the dataset but one; the model was then used to classify the held-out test item, enabling us to determine whether the model successfully generalized across to other items (i.e., whether the predicted label from the model matches the actual label for the held-out test item). This leave-one-out procedure was performed for every item in the dataset, ultimately providing an accuracy score, i.e., the percentage of correctly classified held-out items. For each of the 336 splits of the data (336 recorded utterances in total), we thus fitted a logistic regression classifier to the 335 training utterances. The classifier was trained to predict an utterance’s original Meaning from all seven acoustic features and their interaction with Form (interrogative or declarative). This classifier was then used to predict the Meaning of the held-out test item.

The classifier successfully predicted Meaning on 76% of held-out test items, a rate substantially above chance. As shown in Figure 5, held-out test items that were estimated as being more likely to be requests (i.e., a larger value for p(Request)) were more likely to have originally been meant as requests.

Figure 5. A classifier equipped with the seven acoustic features correctly predicted a held-out test item’s Meaning 76% of the time; the figure illustrates the distribution of classifier probabilities over classes (e.g., literal, request), coloured by the original Meaning of a given item.

Figure 6. Z-scored f₀ slope by Form and Interpretation.

Figure 7. Z-scored mean f₀ slope by Form and Interpretation.

3.5 Interim conclusion

To summarize the results of our production study, our data confirm our prediction that, for both modal interrogatives and negative state remarks, the higher the mean f₀ of the utterance, the higher the likelihood that it was intended as a request. In addition, our second prediction according to which both types of utterances with more a positive f₀ slope should be less likely to have been meant as requests was borne out. We also confirmed our third prediction: for negative state remarks stimuli, longer utterances were more likely to be requests. Furthermore, a classifier equipped with seven acoustic features correctly predicted the meaning of the stimuli 76% of the time.

4.1 Method

4.2 Predictions

We first predicted that, as in English, participants’ pragmatic interpretations would be predicted by the speaker’s original intent – that is, participants should interpret requests as requests, and questions (or negative state remarks) as non-requests.

We also had predictions about which acoustic features ought to predict participants’ pragmatic interpretations. For both types of constructions, we predicted that longer duration would increase the likelihood of a request interpretation (Trott et al. Reference Trott, Reed, Ferreira and Bergen2019). In addition, we expected non-inverted modal interrogatives with less positive f0 slopes to be more likely to be interpreted as requests.

4.3 Results

All analyses were conducted in R (R Core Team 2020). Mixed effects models were constructed using the lme4 package (Bates et al. Reference Bates, Maechler, Bolker and Walker2015), and log likelihood ratio tests were compared using the anova function.

Our first question was whether participants’ pragmatic interpretations were reliably predicted by the speaker’s original Intent. To address this question, we constructed a generalized linear mixed effects model with Interpretation (Request vs. Non-Request) as a dependent variable, fixed effects of Intent (Request vs. Non-Request) and Form (Interrogative vs. Declarative), by-subject random slopesFootnote ⁷ for the effects of Intent and Form, by-item random slopes for the effect of Intent, and random intercepts for items, subjects, lists, and speakers. We compared this model to a model omitting only the effect of Intent, and found that the full model exhibited a better fit, as demonstrated by a log likelihood ratio test [X²(1) = 17.78, p < .001]. As predicted, Request interpretations were more likely for utterances originally intended as Requests [β = 1.18, SE = 0.2]; participants responded correctly for 61.38% of items. Altogether, these results indicate that our participants successfully made use of prosodic information to infer a speaker’s intent. Participants’ responses were also influenced by the form of the utterance: the log-odds of Request interpretations was higher for interrogative utterances than declaratives [β = 1.36, SE = 0.41]

Second, we asked whether participants’ pragmatic interpretations could be predicted by any of four acoustic features (Mean f₀, f₀ Slope, and Number of Voiced Frames), and their interaction with Form. Consistent with past work (Trott et al. Reference Trott, Reed, Ferreira and Bergen2019), we constructed a full model with Interpretation as a dependent variable, fixed effects for each of the seven features mentioned earlier, as well as their interaction with Form, and random intercepts for subjects, items, lists, and speakers. We then compared this full model to a series of reduced models, each of which omitted an interaction term between Form and one of the critical four acoustic features; we also compared each of those reduced models to a model omitting the acoustic feature entirely (i.e., to ask about a main effect of that acoustic feature). After correcting for multiple comparisons, we found that model fit was improved by the interactions between Form and f₀ Slope [X²(1) = 45.38, p < .001], as well as a main effect of f₀ Slope [X²(1) = 153.06, p < .001]. The log odds of a request interpretation increased for utterances with a more positive f₀ Slope [β = 0.12, SE = 0.08], but this effect reversed for interrogative utterances [β = −0.62, SE = 0.09] (Fig. 6). In other words, declarative utterances with a more positive f₀ slope were more likely to be interpreted as requests, while interrogatives with a more positive f₀ slope were less likely to be interpreted as requests (i.e., they were more likely to be interpreted as yes/no questions).

Additionally, we found a main effect of Mean f₀ [X²(1) = 141.91, p < .001] (Fig. 7). Utterances with a higher Mean f₀ were more likely to be interpreted as requests, regardless of Form [β = 0.47, SE = 0.07]. Finally, we found an interaction between Number of Voiced Frames and Form [X²(1) = 14.43, p < .001], and a main effect of Number of Voiced Frames [X²(1) = 29.5, p < .001]. Declarative utterances with a larger number of voiced frames were more likely to be interpreted as requests [β = 0.42, SE = 0.06], but this effect reversed for interrogatives [−0.37, SE = 0.1].