This study tests how prosodic boundary strength (i.e., categorical differences between Accentual Phrase, AP, and intermediate phrase, ip, boundaries) per se affects the syntactic parsing of spoken utterances in French. Two forced-choice perception experiments demonstrated that French listeners use prosodic boundary strength (either AP or ip boundaries) at the end of noun phrases (e.g., La nana du sauna ‘The girl who manages the sauna’) to choose whether NPs are likely to be followed by a prepositional phrase (e.g., d’Héléna ‘of Héléna’) or instead by a verb phrase (e.g., déconseille ‘advises against’). Experiment 1 employed fragments extracted from natural speech stimuli, while Experiment 2 made use of resynthesized speech, in which fundamental frequency and durational cues to AP and ip boundaries were independently manipulated. In Experiment 1, results show that listeners prefer PP completions following AP boundaries and VP completions after ip boundaries. Experiment 2 shows that preboundary duration cues consistent with the presence of an AP boundary successfully guide listeners to prefer PP completions, while both fundamental frequency and duration cues consistent with an ip boundary are necessary to induce VP completions. We hence argue that prosodic boundary strength at the right edge of an utterance fragment influences syntactic parsing decisions in French.
Over the past 50 years, a growing number of studies has investigated the role of prosodic boundary placement in the syntactic parsing of spoken utterances. The conclusion warranted by these studies, independent of the language under investigation, is that the presence vs. absence of a prosodic boundary is able to guide syntactic parsing. For instance, Kjelgaard and Speer (1999) compared early closure sentences (e.g., When Roger leaves the house is dark) versus late closure sentences (e.g., When Roger leaves the house it’s dark) under three prosodic conditions in which the placement of an Intonation Phrase (IP) boundary was manipulated, i.e., cooperating prosody (e.g., for early closure: [When Roger leaves]IP [the house is dark]IP), conflicting prosody (e.g., for early closure: [When Roger leaves the house]IP [is dark]IP), and baseline prosody (e.g., for early closure: [When Roger leaves the house is dark]IP) to investigate the role of Intonation Phrase (IP) boundaries in syntactic parsing decisions in American English. Using four different tasks (a grammaticality judgment task, a sentence comprehension task, and two cross-modal naming tasks), they found faster reaction times in the cooperative prosody condition compared to the baseline condition. The results also showed a processing disadvantage for early closure syntactic structures in the baseline and conflicting conditions, which disappeared when utterances were presented with cooperating prosody. Taken together these results suggest that prosody informs the syntactic parsing process and that cooperating prosody can help resolve temporary syntactic ambiguities. More interestingly, in the two cross-modal naming tasks, the facilitation effect of cooperating prosody and the interference effect of the baseline and conflicting prosody were observed at the point of syntactic disambiguation (i.e., the word is for early closure). This shows that the prosodic organization influences syntactic structuring at an early stage of processing by guiding the generation of syntactic parsing. In other words, the main contribution of Kjelgaard and Speer’s findings is in showing that facilitation and interference due to the presence of an IP boundary do not take place once the end of the sentence is reached, but right at the point of syntactic disambiguation. This is in contrast with earlier models of syntactic parsing (see Pynte and Prieur 1996) arguing that the presence of a prosodic boundary contributes only to the re-analysis of an initially determined syntactic parse, hence once the entire utterance is perceived.
Moreover, Kjelgaard and Speer’s results also reveal that a lower-level boundary, i.e., the intermediate phrase break, or ip, can have similar effects in resolving temporary syntactic ambiguity. In fact, in this study, facilitation and interference effects due to the presence of IP boundaries were reproduced with ip boundaries at the critical boundary locations. This is in contrast with the predictions of previous syntactic parsing models, such as the one proposed by Marcus and Hindle (1990). In this model, only IP boundaries can separate speech into chunks in the first stage of parsing, while no other prosodic boundary type can inform syntactic parsing decisions at this stage.
Although the role of IP and ip boundary placement in syntactic parsing in English is now well established (e.g., Schafer 1997; Wagner and Watson 2010 for reviews), their role in French is less clear. It is also important to note that French and English present both similarities and important differences in their prosodic structure. First, both IP and ip, which are generally admitted for English, also exist in French (Jun and Fougeron 2000, Jun and Fougeron2002), though their phonological and phonetic specifications are different. Second, as in English, the highest constituent in the prosodic hierarchy (i.e., the IP) is marked by a right-edge tone that can be either Low (L%) or High (H%), is associated to the strongest degree of preboundary lengthening in French, and is optionally characterized by the presence of a silent pause (Jun and Fougeron 2000; Michelas and D’Imperio 2012). Figure 1(a) shows the French utterance fragment La nana du sauna with the last syllable associated to an IP right boundary. Despite their similarities, French and English differ in a number of ways regarding their prosodic structure. First, although the two languages include an ip-level, this level is not phonologically and phonetically cued in the same way. In English, the ip is typically cued by a phrase accent that is responsible for blocking recursive downstep of subsequent H peaks within the ip. As a consequence, the first H peak after the ip boundary is set to a fundamental frequency (F0) level that is comparable to the level shown by the first H peak of the IP (Beckman and Pierrehumbert 1986). By contrast, in French, an H-phrase accent cueing the right edge of the ip blocks downstep on the last H peak within the phrase, hence resetting the peak before (and not after) the ip boundary (D’Imperio and Michelas 2014). This leads to a complete pitch reset before the ip-right boundary, given that the preboundary peak is scaled at the same level as the first accentual peak in the IP (see Figure 1(b)). In Figure 1(b), the dotted line represents the register level set by the first accentual H peak (LH*) of the utterance fragment.
In addition, the degree of preboundary lengthening at the end of the ip is weaker than the one observed in IP preboundary position (Michelas and D’Imperio 2012). That is why the last syllable in the IP in Figure 1(a) is longer than the last syllable in the ip in Figure 1(b). Moreover, different from English, a third level of prosodic constituency is necessary to account for the French prosodic hierarchy, i.e., the Accentual Phrase, or AP. This level is tonally defined by the presence of an obligatory final rise (LH*) associated with the last syllable of the AP (Jun and Fougeron 2000, Jun and Fougeron2002; Welby 2006). Since stress is implemented at the phrasal level in French, AP right boundaries always coincide with a final accentual prominence. Figure 1(c) shows the utterance fragment La nana du sauna with the last syllable associated to an AP right boundary. In this figure the last LH* rise within the fragment is downstepped relative to the preceding LH* rise since no H-phrase accent intervenes in order to block downstep between subsequent H peaks. Finally, AP-boundaries are cued by the weakest degree of preboundary lengthening (Michelas and D’Imperio 2012), hence smaller than the one observed at the right edge of the ip. Note that the last syllable of the fragment is in fact shorter in Figure 1(c) than in Figure 1(b), due to its association to an AP and not to an ip boundary.
Only a small number of preceding studies have addressed the question of the role of prosodic boundaries in syntactic parsing decisions in French. Among these studies, Pynte and Prieur (1996) showed that IP boundaries influence syntactic parsing decisions in French. In four word-monitoring tasks in which participants were presented with similar sentences and were instructed to detect a pre-specified target word (which was always the final word of the sentence), they investigated the ambiguous attachment of a prepositional phrase, PP, in sentences such as Les espions informent les gardes de la conspiration ‘The spies inform the guards of the conspiracy’, in which the PP de la conspiration ‘of the conspiracy’ could either be attached to the Verb Phrase, VP, or the Noun Phrase, NP. Their findings show that an IP break inserted after the verb ([The spies inform]IP[the guards of the conspiracy]IP) favors a VP-attachment interpretation, confirming that French listeners are able to use the presence of IP breaks to perform syntactic analyses of spoken utterances. In a series of more recent studies, Millotte et al. (2007) and Millotte et al. (2008) showed that a prosodic unit lower than the IP can also constrain syntactic parsing in French. In these studies, the authors investigate temporary syntactic ambiguities due to word homophony (e.g., the word/mɔʁ/can either be a verb, as in [Le petit chien]PPh [mord]PPh ‘the little dog bites’, or an adjective as in [le petit chien mort]PPh ‘the dead little dog’) to investigate whether the presence of a Phonological Phrase (PPh, Nespor and Vogel 1986) boundary aligned with the NP/VP break influences syntactic parsing decision in French. They consistently found that French participants interpreted the ambiguous word as a verb when a PPh boundary was placed at the NP/VP break, while they interpreted it as an adjective when no prosodic boundary was present. The authors hence showed that boundaries that are potentially lower than IP boundaries are used on-line to constrain syntactic parsing decisions. However, because in Millote et al.’s studies, PPh boundaries were mainly defined according to a syntactic algorithm typical of the Prosodic Phonology framework, it remains unclear whether both lower-level boundaries, i.e., AP and ip boundaries, would exhibit the same pattern of results in syntactic parsing decisions in French.
In this study, we hence focus on the role of both lower-level boundaries (AP and ip) in French, and test whether and how they influence syntactic parsing decisions in ambiguous sentence fragments. We exploited the fact that the presence of an ip boundary is highly correlated with the presence of a major syntactic break, such as the Subject NP/VP break in French (D’Imperio and Michelas 2014), as opposed to the AP level cueing lower-level syntactic breaks, such as NP-internal breaks. Hence, we used two-forced-choice completion tasks (prepositional phrase or PP-choice and verb phrase or VP-choice), in which French participants heard sentence fragments ending either with an ip or an AP boundary. In Experiment 1, we used natural speech stimuli, which were previously tonally transcribed, to test whether prosodic boundary strength (i.e., categorical differences between AP and ip breaks) could be used by French participants to perform a syntactic completion task.
In a second experiment, we focused on the phonological and acoustic properties of the breaks in order to better understand the interplay of these cues in boundary strength perception and its influence in parsing. It is in fact now widely acknowledged that prosodic boundaries are abstract phonological entities and that the manner in which they are produced and perceptually exploited is language specific. For instance, uninitiated native speakers of English have been known to misinterpret the long closure interval of Japanese geminate stops as a pause and thus to mishear a word-internal syllable sequence as split across a strong syntactic boundary (Beckman 1996). What is more, the way these abstract prosodic entities are cued in the speech stream is also language specific and requires a complex interplay of phonetic/acoustic cues. In early psycholinguistic studies, only pauses were employed in order to mimic the presence of a prosodic boundary (e.g., Pynte and Prieur 1996). In contrast, contemporary models of sentence processing have begun to control for several acoustic/phonetic indices cueing the different boundaries, including duration and tonal cues known to be two of the main relevant cues for boundary marking (Schafer 1997; Kjeelgard and Speer 1999). However, little is known about how the interplay between these two cues affects syntactic parsing decisions. In fact, in French, the choice of a particular degree of boundary strength can create ambiguities in the phrasal parse. For example, when speech rate is fast, it is difficult to identify different degrees of preboundary lengthening, and thus it can be difficult to differentiate an AP from an ip (Michelas and D’Imperio 2012). Hence, here we ask: when durational cues are ambiguous, are French listeners able to use tonal cues only (i.e., the difference between the pitch height of an AP-final LH* accent and the pitch height of an H-phrase accent) to syntactically parse auditorily presented sentence fragments?
In order to better evaluate the role and the interplay of different phonetic and phonological cues to the AP/ip boundary perceptual difference, in Experiment 2 we artificially manipulated the duration of the last syllable (for the purpose of creating different preboundary lengthening effects) and pitch height (to reproduce specific tonal cues) of both AP and ip boundaries to investigate the effect on parsing. To sum up, we addressed two main questions in this study: (1) are French listeners able to make use of boundary strength indices relative to AP and ip boundaries in the syntactic parsing of sentence fragments, and (2) how does the complex interplay between duration and tonal cues at the boundary location influence French listeners’ morphosyntactic decisions? Answers to these questions will shed light on the role of prosodic constraints on the syntactic parsing process, and more specifically on the interface between phonetic cues and abstract prosodic structure as well as between abstract prosodic structure and the syntactic parsing process.
2 Experiment 1: natural-speech experiment
Given that ip boundaries are generally inserted at the right edge of a maximal syntactic projection in French (i.e., the top node projection of the head, XP, such as a subject NP), our hypothesis is that participants will use boundary strength (AP vs. ip) to choose whether a sentence fragment beginning with a Subject NP would likely to be followed by a VP or instead by an NP-internal PP completion. Specifically, our hypothesis is that participants will interpret an ip boundary as cueing a major syntactic break, i.e., by choosing a VP after perceiving the break, while they will interpret an AP boundary as cueing a minor break, hence preferring an NP complement completion, such as a PP. In line with this hypothesis, we expect more PP responses (e.g., d’Héléna ‘of Héléna’) than VP responses (déconseille ‘advises against’) when an AP boundary rather than an ip boundary is present in the stimulus at the end of the NP La nana du sauna.
Thirty-six French native speakers took part in the experiment. They were all students at Aix-Marseille University and had been living exclusively in France. Their age ranged from 18 to 25.
Twenty pairs of NPs were chosen for the study. Within each pair, the NPs were segmentally identical up to the sixth syllable, the only difference being the prosodic boundary strength associated to the last syllable. In fact, for each pair, the target noun phrase could either be associated to an AP or an ip right boundary (see ppendix). Sentence fragments were always NP subjects, and the target syllable had consistently a CV structure.
In order to induce two levels of prosodic boundary at the right edge of the target syllable, we manipulated syntactic complexity within the Subject NP, which was either simple branching (ip condition, e.g., La nana du sauna ‘The woman who manages the sauna’) or double branching, hence with the NP followed by two PPs (AP-condition, e.g., La nana du sauna d’Héléna ‘The woman who manages Helena’s sauna’). Thus, the target syllable (which was always the last syllable of the first branching constituent) was aligned with an NP/VP break only in the first, i.e., ip, condition. In the AP condition, since the head was double branching, the target syllable was followed by another branching constituent. Constituents were always composed of a two-syllable head noun preceded by a determinant or by the preposition de ‘of’. In the ip condition, target NPs were always followed by a VP beginning with the consonant /d/ (e.g., déconseille ‘advises against’) in order to minimize possible coarticulation effects between the last syllable of the target AP and the following phrases. Syntactic and prosodic structures of the utterance fragment La nana du sauna in our two AP and ip conditions are shown in Figure 2.
A French native speaker (the first author) produced the experimental sentences once. She repeated the utterance in case she misread part of the text. The speaker read the sentence aloud from a list, at a self-selected normal rate. Each utterance was then prosodically labeled by the authors. To do this, the sound files were segmented and each utterance was saved as a separate file. F0 curves and spectrograms were then generated using Praat (Boersma and Weenink 2009), and AP, ip, and IP boundaries were manually labeled through an AM-based prosodic annotation scheme (Jun and Fougeron 2000; D’Imperio and Michelas 2014). The results of the annotation revealed that all target NPs were produced with the expected phrasing pattern (i.e., two APs including two LH* accents for the AP-boundary condition and two APs with two LH* accents and an H-phrase accent for the ip-boundary condition; see Figure 2). The acoustic properties of the last syllable of the first and the second AP were also subsequently analyzed. Specifically, preboundary lengthening and tonal markers were selected as main indicators of prosodic structure. Each utterance was hence phonetically segmented by means of Easy Align (Goldman 2011). Phoneme boundaries were then hand-corrected by inspecting both waveforms and spectrograms. In each of our target NPs, the vocalic nucleus of the last syllable (which was coextensive with the syllable rhyme, i.e., V1 in Figure 3La naNAdu sauna) was segmented and its duration was measured. Durations of target vowels were then automatically extracted using Praat scripts. As expected, duration analyses revealed that the final vowel of the second AP, i.e., V2 in Figure 3 (e.g., La nana du sauNA) was shorter in the AP-boundary than in the ip-boundary condition (t(37.86)=–8.226, p<0.001, d=–2.96), indicating that the degree of preboundary lengthening was stronger in the ip-boundary than in the AP-boundary condition. Moreover, the length of the final vowel of the first AP (V1) was not different in the AP-boundary and ip-boundary conditions [t(37.994)=0.7175, p=0.477], showing that V1 was always associated with an AP boundary independent of the prosodic condition. Finally, V1 and V2 durations were not different for the AP-boundary condition [t(37.994)=–0.717, p=0.477], confirming that both V1 and V2 were associated to an AP boundary in the AP-condition.
We also examined maximum F0 values associated to V1 and V2 in the AP-boundary and the ip-boundary conditions. F0 maxima (H1 and H2) were automatically labeled and extracted using Praat scripts. As expected, H2 was significantly lower than H1 in the AP condition [t(27.886)=–8.392, p<0.001, d=–2.93], while it was not in the ip condition [t(36.992)=–1.057, p=0.2974]. This confirmed that an H-phrase accent, responsible for blocking downstep between the first two LH* rises, is responsible for H2 pitch reset in the ip-boundary condition.
Finally, we also checked whether both completions (VP or PP) were plausible for each given sentence. To do so, a group of 10 French participants read all experimental sentences and judged their overall plausibility on a 0 (completely implausible) to 7 (highly plausible) scale. PP completions obtained the same result as VP completions (mean of 6.1 for PP completions and of 6.2 for VP completions, t(31)>1).
Two lists of stimuli were created so that each participant heard only one member of each sentence pair, while being presented with both AP-boundary and ip-boundary conditions. Within each list, 40 distractors were randomly cut within the sentence, at a word boundary, in order to make sure that participants could not anticipate the length of the fragment they had to complete. Among these 40 distractors, 20 had only a one-AP Subject (e.g., La page ‘the sheet’), 10 had a two-AP Subject (Le frère de James se retrouvera ‘James’s brother will come’), and 10 had a three-AP Subject (Le poster de la chambre de Marion se détache ‘The poster in Marion’s room is tearing off’). The two completions were plausible and consisted of various syntactic structures. The order of presentation of the trials was randomized for each subject.
Participants were seated in front of a computer, wearing headphones, and were individually tested in the soundproof room of the Laboratoire Parole et Langage of Aix-Marseille University. Auditory stimuli were stored at a sampling rate of 16 kHz and presented over Sennheiser HD 212Pro headphones at a comfortable listening level. Participants were instructed to listen to the stimuli and to complete them by choosing one or the other possible response. Specifically, they were instructed to press either the right or the left button of a response box to indicate their choice. Participants were aware that both choices were plausible but were instructed to choose the most appropriate response. The instructions given were as follows: ‘We asked one person to read a set of sentences. You will hear the beginning of these sentences over your headphones and then see two possible completions for each corresponding sentence. You have to complete the beginning of the sentence by guessing the following sequence, which was actually produced by our speaker. Give your response as soon as you can’. Before the experiment began, participants completed a 9-item practice. A trial began with the presentation of an auditory sequence. Two seconds later, two possible completions appeared on the screen. The entire procedure was controlled using E-prime Software (Psychological Software Tools, Pittsburgh, PA, USA).
In line with our main hypothesis, a greater number of PP responses was obtained for the AP condition while more VP responses were obtained for the ip condition (see Table 1).
|Response type||Prosodic boundary|
To test the statistical significance of prosodic boundary strength, mixed-effects logistic regression modelling was employed (lme4, and languageR packages; Baayen 2008; Bates et al. 2015) implemented in R (R Development Core Team). p-values were estimated using the summary () function as implemented in the languageR package (Baayen 2008). For a glmer model, p-values are (as is typical in the context of generalized linear models) asymptotic Wal test statistics. Our binary dependent variable was response type (PP response=0, VP response=1). We first introduced Boundary Strength and List and their interaction as fixed factors. In addition to fixed factors considered in simple linear regressions, linear mixed-effects models account for random variation induced by participants and items. We hence introduced by-participant random intercepts, by-participant random slopes for the fixed effect estimated within participants (not for the List effect), by-item random intercepts, and by-item random slopes, in the following model: ResponseType~Boundary Strength * List + (1 + Boundary Strength|Participant) + (1 + Boundary Strength|Item) (see Barr et al. 2013). The model was then simplified by excluding the List factor which was not significant and whose inclusion did not improve the model fit (z=0.501, p=0.61607). This model included 720 observations.
In line with our hypothesis, the model showed a main effect of Boundary Strength on the response type yielded by the participants (Table 2). Specifically, participants gave significantly more VP responses for ip-boundary stimuli than for AP-boundary stimuli (z=6.397, p<0.0001). Given that both our dependent and explanatory variables were binary, the results also showed that participants gave significantly more PP responses when an AP boundary rather than an ip boundary was present in the stimulus. Note that participants did not provide chance-level responses. Indeed, the p-value of the Intercept (p<0.001) indicates that the probability of occurrence of a VP response significantly differed from chance.
|Coefficient estimate||Standard error||z||p|
Our results show that participants employed prosodic boundary strength to process the depth of syntactic embedding. Specifically, AP-boundary stimuli induced listeners to interpret sentence fragments as likely to be followed by a minor, NP-internal PP break, while ip-boundary stimuli induced listeners to interpret the fragment edge as a major syntactic break (i.e., a NP/VP break). This result is consistent with the view that the strength of a prosodic boundary influences French syntactic parsing decisions. However, because we used naturally-produced stimuli in this experiment, we cannot precisely describe which acoustic/perceptual cues directly present in the speech signal were employed by participants in order to perform the task. Early studies (Lehiste 1972; Warren 1985) on the role of duration and pitch in signaling prosodic boundaries in English, to the purpose of syntactic disambiguation, show that duration cues are stronger predictors of disambiguation performance in certain structures, but that pitch cues can be essential for others. Hence we asked: do participants need a combination of both preboundary lengthening and tonal cues to identify prosodic boundary strength cueing an AP or an ip break and then use it for the purposes of syntactic parsing? What happens in the case of ambiguous prosodic breaks that are marked by only one type of cue rather than by a combination of different boundary cues? Thus, we carried out a similar experiment in which the prosodic cues located at the boundary location were artificially manipulated through resynthesis, as described in the next section.
3 Experiment 2: resynthesized-speech experiment
In Experiment 2 we used the same basic design as in Experiment 1, while introducing an additional experimental factor. We artificially manipulated acoustic correlates (duration and F0) of preboundary lengthening and tonal specification of AP/ip boundaries to obtain a set of stimuli marked either by only one type of cue or by a combination of the two. Specifically, we manipulated (i) the degree of preboundary lengthening associated to the last syllable of the fragments and/or (ii) pitch downstep/reset of the last LH* pitch accent located on the fragment. Our goal was to test the effect of these modifications on listeners’ syntactic parsing of a sentential fragment.
Given the results obtained with natural speech stimuli in Experiment 1, we first hypothesized that when both preboundary/duration cues and tonal/F0 cues are congruent with the presence of an AP boundary, participants would interpret those indices as cueing a minor syntactic break, while when the same cues are congruent with the presence of an ip boundary, they would be employed as cues to a major syntactic break. Specifically, our hypothesis is that, when both duration and F0 are modified in an original AP fragment so that their values are equal to the mean acoustic values for an ip fragment, we expect listeners to choose VP and not PP responses. We also expect listeners to yield more VP responses when both acoustic cues are coherent with the presence of an ip boundary than when only one type of cue (either stronger preboundary lengthening or higher tonal value for the edge tone) is present.
Moreover, when either one or the other of the acoustic cues is present in the signal, two outcomes are possible. On the one hand, the presence of only one type of cue (either duration or pitch height) might not be sufficient for participants to correctly process the depth of the syntactic break at the end of the fragment. On the other, it is possible that only one cue might be sufficient in itself for participants to both perceive and employ the break for syntactic parsing decisions. Since prosodic boundaries are multiparametric and cued by a complex interplay between several acoustic cues, we favored a version of the first hypothesis stating that participants would need a combination of both duration and tonal cues to choose between a PP and a VP completion of the fragment just heard.
Twenty NPs were selected from the corpus employed in Experiment 1, in which the target syllable was associated to an AP boundary. We hence artificially modified F0 and/or duration cues associated to the last syllable of each fragment using PSOLA (Pitch Synchronous Overlap and Add; Moulines and Charpentier 1990) in Praat (Boersma and Weenink 2009) in order to manipulate boundary strength (Table 3) for a total of 120 stimuli (20 NPs × 2 boundary levels × 3 acoustic combinations).
|Boundary strength||Acoustic properties of boundaries|
|Tonal cues||Duration cues|
|AP||both||LH*||AP preboundary lengthening|
|dur||Ambiguous tonal cues||AP preboundary lengthening|
|ton||LH*||ambiguous duration cues|
|ip||both||LH*H-||ip preboundary lengthening|
|dur||Ambiguous tonal cues||ip preboundary lengthening|
|ton||LH*H-||ambiguous duration cues|
The 120 stimuli thereby created differed only in the duration of the last syllable and in the F0 height of the second LH*. Stimuli mimicking an ip boundary were created on the basis of an AP fragment, by adding mean duration values for the last syllable of natural ip fragments, and by superposing a final F0 value scaled at the height of the first H peak (Figure 5). Apart from the modified durational and tonal properties of the last syllable, the stimulus sequences were strictly identical in all six conditions.
For the AP-ton and ip-ton conditions, the F0 value was either equal to the AP or the ip level, while the duration of the last syllable was altered to make it acoustically ambiguous between an AP and an ip boundary. Specifically, for each NP, we adjusted the duration of the syllable to a value that was equal to the mean duration for both AP and ip natural stimuli (see also Ladd and Schepman  for previous use of this methodology). Duration modifications were carried out by adding additional periods to the last vowel in the fragment. Given that (1) our NPs always ended with a CV syllable and (2) preboundary lengthening preferentially affects the vowel nucleus for primary stress in French (Astésano 2001), a modification of the final vowel duration was made to adequately reflect the degree of lengthening associated to a specific prosodic level.
In the 40 durationally modified NPs, the pitch contour was subsequently manipulated. In the AP condition, we simply stylized the pitch contour of the original AP fragment. For the ip-ton condition, the pitch height of the final vowel was manipulated in order to simulate the presence of an H-phrase accent, hence raising the original F0 values so as to obtain the same F0 value for the first H peak of the fragment. In both the AP-dur and the ip-dur conditions, the duration of the vowel within the last syllable was adjusted to a value corresponding either to that of the original AP condition or to that of the original ip condition, while rendering F0 values ambiguous between the AP and the ip level. Target vowel height was also rendered tonally ambiguous by adjusting the final F0 value to the average between the heights of the first and the second pitch accent in the original stimulus. Finally, in the ip-both and AP-both conditions, both the F0 height and the duration of the target vowel were made to reproduce both acoustic cues of the original stimulus.
In order to ensure that the stimuli thereby created did not sound artificial, we asked five French native speakers to judge their naturalness on a 1 (completely natural) to 7 (highly unnatural) scale. Listeners judged the 120 resynthesized stimuli as natural as the original NP fragments read by our speaker (mean of 5.1 for the natural NPs; mean of 5.158 for the artificial NPs; t(290.855)=0.311, p=0.6667). We also checked whether the resynthesized stimuli judged in the six conditions differed in their naturalness rating. The effect of condition on ratings was not significant (F(5,25)=0.926, p=0.480; mean for AP-both=5.15, mean for ip-both=4.68; mean for AP-dur=5.85, mean for ip-dur=4.82; mean for AP-ton=5.56, mean for ip-ton=4.978). On the basis of the 120 stimuli created, six lists of sentences were constructed so that the stimuli created from the same utterance base would appear in a different list. Each list contained 20 stimuli and 40 fillers. The fillers were kept the same in all the lists and were randomly cut at a word boundary. Within each list, the stimuli were presented in random order. Each participant listened to only one list.
Sixty monolingual French native speakers, all students at Aix-Marseille University, took part in the experiment. They had all been living exclusively in France. Their age ranged from 17 to 24.
The same procedure as in Experiment 1 was employed.
Percentage of PP responses for the two prosodic boundary levels (AP and ip) and the three acoustic levels (duration alone, tonal cues alone, or both) is shown in Figure 6.
To test the statistical relevance of the results, we employed a mixed-effects logistic regression model. As in the logistic regression analysis used in Experiment 1, the dependent variable was response type (PP response=0, VP response=1). We first tested a model including Prosodic Boundary Strength, Acoustic Levels, and List and their interaction as fixed factors, and Participant and Item intercepts as random factors: ResponseType~Boundary Strength * Acouctic Properties * List + (1|Participant) + (1|Item). The model was then simplified by excluding the List factor, which was not significant, as the likelihood ratio test showed (χ2=32.534, p=0.3431; see Pinheiro and Bates 2000: 83). The model we finally selected included Prosodic Boundary Strength, Acoustic Levels, and their interaction as fixed factors. By-participant random slopes for the two fixed effects and by-item random intercept were also included: Response Type~Boundary Strength * Acoustic Properties + (1 + Boundary Strength|Participant) + (1 + Boundary Strength|Acoustic Properties) + (1|Item). This model, including 1,200 values, shows a significant interaction between Boundary Strength and Acoustic Properties (Boundary(ip) * Acoustic Properties(dur): p<0.05; Boundary(ip) * Acoustic Properties(ton): p<0.001). The significant interaction between Boundary Strength and Acoustic Properties was confirmed by a likelihood ratio test (LRT) between models with and without interaction: χ2=11.831, p<0.01). This interaction indicates that the difference in occurrence of a VP response between the AP-boundary condition and the ip-boundary condition decreases depending on the acoustic properties. As shown in Figure 7, the difference between the probability of occurrence of VP responses between AP-boundary and ip-boundary conditions was stronger when boundaries were cued by both indices than when they were cued only by duration cues, and in turn stronger when cued by duration indices than when cued only by tonal indices.
To interpret the significant Boundary Strength × Acoustic Properties interaction, the 2 × 3 factor design was converted into a 1 × 6 factor design including the maximal random effect structure. The contrasts of interest were obtained using the multcomp package. In line with our main hypothesis, the multiple comparisons showed that when prosodic boundaries were cued by a combination of lengthening and tonal cues, participants gave significantly more VP responses in the ip-both rather than in the AP-both condition (z=6.531, p>0.0001, see Table 4).
|Coefficient estimate||Standard error||z-value||p (</z/)|
|AP-both vs. ip-both||1.44211||0.22079||6.531||<0.0001|
|AP-dur vs. ip-dur||0.47631||0.21149||2.252||<0.05|
|AP-ton vs. ip-ton||0.11412||0.20980||0.544||0.5865|
When prosodic boundaries were cued by duration indices only, participants gave more VP responses in the ip-dur than in the AP-dur condition (z=2.252, p<0.05). By contrast, the number of VP Responses was not different for the two boundary strength conditions when the breaks were cued only by tonal indices (z=0.544, p=0.5865).
To sum up, results from the re-synthesized speech experiment confirmed and extended the results obtained in the natural-speech experiment. When AP and ip boundaries were cued by a combination of tonal and lengthening cues, French listeners were able to use the strength of the prosodic breaks associated with the last syllable they heard to perform a syntactic analysis of the utterance. However, when the two prosodic levels under investigation were signaled by only one type of acoustic cue (either tonal or lengthening cue), the results were different for the two prosodic levels. For the AP-boundary condition, participants interpreted stimuli as cueing a minor syntactic break when duration cues alone were present at the point of ambiguity (though to a lesser extent than when both cues were present). By contrast, participants did not interpret the AP boundary as a minor break when the boundary was encoded only by tonal cues. This provides evidence that lengthening cues are sufficient for French listeners to perceive an AP boundary, which appears to be exploited as indicating the presence of an upcoming minor syntactic break. By contrast, participants did not preferentially give more VP responses than PP responses in either the ip-dur or the ip-ton conditions. These results support the idea that participants needed the presence of both types of cues (tonal and lengthening cues) to identify an ip break, and hence to correctly complete the sentence after this boundary level.
4 General discussion
The combined results from the experiments presented here form a consistent picture showing that prosodic boundary strength is employed by listeners to constrain syntactic parsing. Across the two experiments, results show that when both AP and ip boundaries are cued by a combination of lengthening and tonal cues, participants preferentially complete fragments with a PP modifier when the last syllable is associated with an AP boundary, while they prefer a VP completion when an ip boundary is present. This is evidence that French listeners use prosodic boundary strength per se to process the depth of syntactic breaks. This result goes beyond previous psycholinguistic research on the influence of the presence (vs. absence) of a prosodic boundary on the syntactic parsing of spoken French (see Pynte and Prieur 1996; Millotte et al. 2007; Millotte et al.2008). Here, we demonstrate that the processing of prosodic boundaries is more complex than previously acknowledged in these studies, since participants also used the level (i.e., strength) of a prosodic boundary in order to syntactically parse utterance fragments. Let us spell out in more detail the implications of our results for syntactic parsing models.
In line with previous research on prosody and sentence processing (for English: Speer et al. 1996; Schafer 1997; Kjelgaard and Speer 1999; for French: Millotte et al. 2007; Millotte et al. 2008, among others), our results suggest that prosodic information rapidly influences syntactic structuring in processing. Many studies have explored the role of prosody in guiding the choice between alternatives for syntactic structure at a point of temporal syntactic ambiguity. In our study, the data did not contain a global syntactic ambiguity, though our stimuli could either be the initial part of a complex NP Subject or the end of the NP itself. Our results suggest that immediately upon identifying the NP fragment, listeners are capable of using prosodic boundary strength in order to choose whether the likely completion is an NP-internal PP or a VP.
Can this result be reconciled with depth-first parsing models positing a delay in the use of prosodic information (Marcus and Hindle 1990 for English; Pynte and Prieur 1996 for French)? If prosodic information contributes only to the reanalysis of previous syntactic commitments (Pynte and Prieur 1996), one might expect that facilitation due to prosodic information would appear at a later stage in the utterance, but not directly at the point of appearance of a prosodic break. In our stimuli, the use of the ip break in sentences such as La nana du sauna]ip déconseille l’utilisation du savon noir should play a role only if the morphosyntactic category of the phrase following the boundary is inconsistent with the predictions made on the basis of prosodic boundary strength. For instance, in a fragment such as La nana du sauna]ip d’Héléna, the NP d’Héléna is inconsistent with an ip-prosodic boundary located immediately before it, since French listeners would expect the presence of an AP boundary (and not an ip boundary) aligned with a minor syntactic break (D’Imperio and Michelas 2014). Similarly, a model positing the early use of boundary tones and at the same time the delayed use of an intermediate level of phrasing (Marcus and Hindle 1990) would predict an effect of AP and ip levels later in the utterance than immediately after the presence of the target prosodic boundary. By contrast, in our study participants were able to successfully predict the occurrence of an upcoming VP based on the strength of the prosodic break, immediately after the prosodic break. Hence, our results appear to be in line with models that posit an immediate use of both boundary tones and French accents.
However, because in our task the response was provided after the alternatives are given, it is not clear whether participants were only building expectations about the morphosyntactic structure of the upcoming phrase. Instead, they could just be comparing the well-formedness of two completed possibilities, and in this case, parsing models positing a delay in the use of prosodic information can explain our data. Thus, the next step of this research will be to employ online measures to show that French listeners make immediate use of the difference between AP and ip boundaries to guide syntactic parsing decisions at their initial steps. Event-related potential studies are a useful way of studying the time course of integration of prosodic information for utterance parsing (see Bögels et al. 2011 for a review). Thus, research must be further developed to determine whether false AP/ip boundaries, i.e., prosodic boundaries that are mismatched with syntactic breaks (e.g., Steinhauer et al. 1999), are sufficient to mislead the listener’s syntactic sentence processor.
Previous findings also differ about the way prosodic boundaries are processed. While some authors (Marcus and Hindle 1990; Watson and Gibson 2005) argue that prosodic boundaries mark points of disjuncture in the syntactic structure, others consider that prosodic units act as processing units (Speer et al. 1996; Kjelgaard and Speer 1999; Schafer et al. 2000). This debate is highly connected with the two possible approaches of syntax-phonology mapping that have been proposed in the literature. On the one hand, the direct approach of the syntax-phonology interface postulates that acoustic events in the speech signal refer directly to the depth of syntactic structure (Gee and Grosjean 1983; Kaisse 1985; Monnin and Grosjean 1993; Keller et al. 1993; inter alia). According to this view, duration and pitch level would be directly interpreted by listeners to process the depth of syntactic breaks. On the other hand, the indirect approach claims that acoustic-phonetic correlates of phrasing relate to an autonomous prosodic structure, which would not be exclusively constrained by the syntactic structure (Nespor and Vogel 1986; Selkirk 1986).
Following this latter view, degrees of preboundary lengthening and scaling properties of LH* rises would encode different boundary strengths of an abstract prosodic structure. Is there a way to distinguish between these two interpretations on the basis of the present set of results? If acoustic events in the speech signal directly refer to the depth of syntactic breaks, with no intervening phonological level, participants should directly use gradual/acoustic prosodic information to make syntactic parsing decisions in a simple relation in which stronger cues, either longer duration or higher pitch at the relevant location, should directly induce the identification of a higher-level boundary. In this case, there need not be phonological abstraction prior to syntactic parsing. By contrast, in Experiment 2, we found that participants needed a combination of both lengthening and tonal cues to correctly complete the sentence fragment ending with an ip break (hence not merely stronger duration cues). This suggests that, at least for the ip level, French listeners needed a complex representation of the boundary in order to correctly construct the abstract prosodic structure that would in turn provide the initial domains for syntactic parsing decisions. Just adding longer duration does not appear to be sufficient to perceive a higher-level boundary in French. This result speaks also against the widespread use of simple duration or pause indices to mark boundaries in several psycholinguistic studies on the syntax-prosody interface.
However, we also know that prosodic boundaries are not only locally cued (i.e., in the direct vicinity of the boundary), and that a prosodic constituent unfolds and reveals itself throughout the beginning of an utterance. This is particularly true for the ip boundary in French. In fact it appears that the height of the H peak found at the ip boundary is not locally defined but is calculated relative to the height of the first accentual peak of the IP (D’Imperio and Michelas 2014). In addition, it appears that the presence of an ip-edge is licensed only if the phrase is long enough (i.e., made up of at least two APs; Michelas and D’Imperio 2012). Thus, in order to process an ip boundary as a cue to a major syntactic break, French listeners might have to analyze tonal cues from the beginning of the IP up to the ip-break. This is in line with the idea that listeners need to construct a full prosodic representation, which in turns provides relevant information for syntactic structuring (Kjelgaard and Speer 1999; Schafer 1997).
To sum up, the hypothesis of a direct relationship between lengthening cues and prosodic boundary levels unmediated by tonal information is not supported by our study. However, we may wonder why duration cues alone were sufficient for listeners to syntactically parse fragments ending in an AP boundary. Remember that, in the framework of the indirect approach of the syntax-phonology interface, it is generally agreed that the relevance of acoustic cues to prosodic boundary marking is language dependent. Also, the fact that the parsing of prosody is language specific (Beckman 1996), and that in French an ip boundary appears to be signaled by the presence of both preboundary lengthening and tonal reset, might explain why both tonal and duration cues were necessary in order to obtain more VP responses, while this was not the case for PP responses. Moreover, a number of authors have previously underlined the importance of duration over pitch in the prosodic marking of French boundaries (Jun and Fougeron 2000; Michelas and D’Imperio 2012; inter alia). In line with these studies, our results support the idea that duration indices are more relevant than tonal cues to AP-boundary marking and its use in predicting the morphosyntactic category of an upcoming phrase. More recently, an interesting study has also underlined the important role of duration cues compared to tonal cues for the purpose of turn-taking in conversation in Dutch (de Ruiter et al. 2006). Although it is difficult to infer the strength of prosodic boundaries at which prosodic manipulations were done in this study (since it did not focus on prosodic phrasing), it also supports the idea that duration cues prevail over tonal cues to give listeners the indices they need in order to predict when a current speaker’s utterance is going to end. Nevertheless, our results also suggest that the relative weight of each type of cue depends on the level of the prosodic constituent in the hierarchy. In fact, although duration indices were sufficient for French listeners to correctly complete a sentence fragment ending with an AP boundary, both duration and tonal indices were necessary to perceive and exploit the presence of an ip boundary in French. This highlights the dependency of prosodic cue marking on boundary strength in an abstract prosodic hierarchy.
Results from two utterance-completion studies provide evidence that listeners use prosodic boundary strength to perform syntactic parsing in French. These results can be interpreted through a very early, automatic, and context-dependent use of information of a phonological prosodic representation during the syntactic parsing process. In addition, the acoustic manipulation of our stimuli showed that duration cues alone were sufficient only for cueing an AP break, signaling an upcoming NP-internal break in French. On the other hand, additional tonal cues were necessary for French listeners to identify an ip break, and hence choose a VP completion. These results lead to a better understanding of the specific role of prosodic boundary strength in the syntactic parsing process, while also underlying the dependency of acoustic cues in prosodic boundary marking on both language type and prosodic level within the prosodic hierarchy. Thus, future research will have to provide a more complete picture of the mapping between fine acoustic details and abstract prosodic boundaries of different strengths, and between these boundary levels and the depth of syntactic break. In both cases, an analysis of language-specific effects will be necessary to integrate prosodic boundary strength in syntactic parsing models in a more general manner.
This work was made possible by a doctoral fellowship from the French Ministry of Education and Research to Amandine Michelas and financial support from l’Institut Universitaire de France to Mariapaola D’Imperio. Institutional support from the Labex BLRI (ANR-11-LABX-0036) in the framework of the Investissements d’Avenir A*MIDEX project (ANR-11-IDEX-0001-02) is also acknowledged. We also thank Stéphane Rauzy for helpful comments and statistical help.
Experimental material. For each target noun phrase, written in capital letters, two sentences were created: in (a) sentences, the target NP is associated with an AP boundary while it is associated with an ip boundary in (b) sentences. Sentences were cut right after the end of the second AP.
1a.LA NANA DU SAUNA d’Héléna déconseille l’utilisation du savon noir.
1b.LA NANA DU SAUNA déconseille l’utilisation du savon noir.
2a.LE SCHÉMA DU TROIS-MÂTS de Nicolas doit être terminé avant demain.
2b.LE SCHÉMA DU TROIS-MÂTS doit être terminé avant demain.
3a.L’AGENDA DU PATRON d’Héléna demeurait introuvable.
3b.L’AGENDA DU PATRON demeurait introuvable.
4a.LE PARVIS DE LA MAIRIE de Paris devrait être entièrement restauré l’an prochain.
4b.LE PARVIS DE LA MAIRIE devrait être entièrement restauré l’an prochain.
5a.LA CORVÉE DE CLASSEMENT des documents durera environ deux heures.
5b.LA CORVÉE DE CLASSEMENT durera environ deux heures.
6a.LA MAISON DE LA CHIMIE de Bordeaux dispose de gros moyens.
6b.LA MAISON DE LA CHIMIE dispose de gros moyens.
7a.LES AMIS DE RÉMY de Dijon décideront des monuments qu’ils souhaitent visiter.
7b.LES AMIS DE RÉMY décideront des monuments qu’ils souhaitent visiter.
8a.LES PROJETS DU CINÉMA de Bordeaux déclenchent la polémique.
8b.LES PROJETS DU CINÉMA déclenchent la polémique.
9a.LE NIVEAU DE JAPONAIS de Marina devient meilleur chaque année.
9b.LE NIVEAU DE JAPONAIS devient meilleur chaque année.
10a.LES CADEAUX DE BIENVENUE du maire deviennent de plus en plus originaux.
10b.LES CADEAUX DE BIENVENUE deviennent de plus en plus originaux.
11a.La villa de Lola de Marseille domine tout le paysage environnant.
11b.La villa de Lola domine tout le paysage environnant.
12a.LE MUSÉE DE L’ARMÉE de Nancy dérange en raison de son exposition sur les nazis.
12b.LE MUSÉE DE L’ARMÉE dérange en raison de son exposition sur les nazis.
13a.LE PALAIS DES CONGRÈS du Pharo doit accueillir 200 visiteurs demain à l’occasion des journées de la mode.
13b.LE PALAIS DES CONGRÈS doit accueillir 200 visiteurs demain à l’occasion des journées de la mode.
14a.LE GÉRANT DU RESTAURANT de Paula discute avec le président depuis plus d’une heure.
14b.LE GÉRANT DU RESTAURANT discute avec le président depuis plus d’une heure.
15a.LE DÎNER DE SARAH de ce soir débutera vers 19 heures.
15b.LE DÎNER DE SARAH débutera vers 19 heures.
16a.LES DÉTENUS DE LA PRISON de Versailles désirent obtenir de meilleures conditions de détention.
16b.LES DÉTENUS DE LA PRISON désirent obtenir de meilleures conditions de détention.
17a.LE PRÉSIDENT DE L’OPÉRA de Paris décide de la programmation.
17b.LE PRÉSIDENT DE L’OPÉRA décide de la programmation.
18a.LES INTÉRÊTS DU CRÉDIT de la maison diminuent d’année en année.
18b.LES INTÉRÊTS DU CRÉDIT diminuent d’année en année.
19a.LE PRÉSIDENT DE L’ASSEMBLÉE de Tunisie démissionnera sous peu.
19b.LE PRÉSIDENT DE L’ASSEMBLÉE démissionnera sous peu.
20a.LE GALA DE FIN D’ANNÉE de Nina dure généralement deux heures.
20b.LE GALA DE FIN D’ANNÉE dure généralement deux heures.
Astésano, Corine. 2001. Rythme et accentuation en français. Invariance et variabilité stylistique. Paris: Editions l’Harmattan. Collection Langue et ParoleSearch in Google Scholar
Barr, Dale, RogerLevy, ChristopheScheepers & Harry J.Tily. 2013. Random effects structure for confirmatory hypothesis testing: Keep it maximal. Journal of Memory and Language68(3). 255–278.10.1016/j.jml.2012.11.001Search in Google Scholar
Bates, Douglas, MartinMaechler & BinDai. 2015. The lme4 package. [Computer software manual]. Retrieved from http://cran.r-project.org/web/packages/lme4/lme4.pdf (accessed 28 January 2015).Search in Google Scholar
Boersma, Paul & DavidWeenink. 2009. Praat: doing phonetics by computer (Version 5.1. 05) [Computer program].Search in Google Scholar
Bögels, Sara, HerbertSchriefers, WietskeVonk & Dorothee J.Chwilla. 2011. Prosodic breaks in sentence processing investigated by event-related potentials. Language and Linguistic Compass5(7). 424–440.10.1111/j.1749-818X.2011.00291.xSearch in Google Scholar
de Ruiter, Jan-Peter, HolgerMitterer & N. J.Enfield. 2006. Projecting the end of a speaker’s turn: A cognitive cornerstone of conversation. Language82. 515–535.10.1353/lan.2006.0130Search in Google Scholar
Goldman, Jean-Philippe. 2011. Easy Align: An automatic phonetic alignment tool under Praat. In Proceedings of the 12th Annual Conference of the International Speech Communication Association, 3240–3. Red Hook, NY: Curran Associates.10.21437/Interspeech.2011-815Search in Google Scholar
Jun, Sun-Ah & CécileFougeron. 2000. A phonological model of French intonation. In AntonisBotinis (eds.), Intonation: Analysis, modelling and technology, 209–242. Boston, MA: Kluwer Academic.10.1007/978-94-011-4317-2_10Search in Google Scholar
Kaisse, Ellen. 1985. Connected speech: The interaction of syntax and phonology. San Diego, CA: Academic Press.Search in Google Scholar
Keller, Eric, BrigitteZellner, StefanWerner & NicoleBlanchoud. 1993. The prediction of prosodic timing: Rules for final syllable lengthening in French. In DavidHouse & PaulTouati (eds.), Proceedings ESCA workshop on prosody, 212–215. Sweden: Lund.Search in Google Scholar
Kjelgaard, Margaret M. & ShariSpeer. 1999. Prosodic facilitation and interference in the resolution of temporary syntactic closure ambiguity. Journal of Memory and Language40. 153–194.10.1006/jmla.1998.2620Search in Google Scholar
Ladd, D. Robert & AstridSchepman. 2003. Sagging transitions” between high pitch accents in English: Experimental evidence. Journal of Phonetics31(1). 81–112.10.1016/S0095-4470(02)00073-6Search in Google Scholar
Marcus, Mitchell, & DonaldHindle. 1990. Description theory and intonation boundaries. In GerryAltmann (eds.), Cognitive models of speech processing: Computational and psycholinguistic perspectives. Cambridge, MA: MIT Press.Search in Google Scholar
Michelas, Amandine & MariapaolaD’Imperio. 2012. When syntax meets prosody: Tonal and duration variability in French accentual phrases. Journal of Phonetics40(6). 816–829.10.1016/j.wocn.2012.08.004Search in Google Scholar
Millotte, Séverine, AliceRenéRogerWales & AnneChristophe. 2008. Phonological phrase boundaries constrain the on-line syntactic analysis of spoken sentences. Journal of Experimental Psychology: Learning Memory and Cognition34. 874–885.10.1037/0278-7322.214.171.1244Search in Google Scholar
Moulines, Eric & FrancisCharpentier. 1990. Pitch-synchronous waveform processing techniques for text-to-speech synthesis using diphones. Speech Communication9(5). 453–467.10.1016/0167-6393(90)90021-ZSearch in Google Scholar
Nespor, Marina & IreneVogel. 1986. Prosodic phonology. Dordrecht: Foris.Search in Google Scholar
Schafer, Amy. 1997. Prosodic parsing: The role of prosody in sentence comprehension. Amherst, MA: University of Massachusetts Amherst dissertation.Search in Google Scholar
Schafer, Amy, ShariSpeer, PaulWarren & DavidWhite. 2000. Intonational disambiguation in sentence production and comprehension. Journal of Psycholinguistic Research29(2). 169–182.10.1023/A:1005192911512Search in Google Scholar
Speer, Shari, MargaretKjelgaard & KateDobroth. 1996. The influence of prosodic structure on the resolution of temporary syntactic closure ambiguities. Journal of Psycholinguistic Research25. 247–268.10.1007/BF01708573Search in Google Scholar
Steinhauer, Karsten, KaiAlter & Angela D.Friederici. 1999. Brain potentials indicate immediate use of prosodic cues in natural speech processing. Nature Neuroscience2(2). 191–196.10.1038/5757Search in Google Scholar
Warren, Paul. 1985. The temporal organisation and perception of speech. Unpublished doctoral dissertation.Search in Google Scholar
Watson, Duane & EdwardGibson. 2005. Intonational phrasing and constituency in language production and comprehension. Studia Linguistica59(2–3). 279–300.10.1111/j.1467-9582.2005.00130.xSearch in Google Scholar
©2015 by De Gruyter Mouton