Hierarchical prosodic structures in the intonation of syntactically recursive sentences1 Caroline Féry and Fabian Schubö, University of Potsdam
Draft, November 2008, comments welcome
Abstract
The question of recursion of prosodic domains, especially of higher ones, has often been addressed in the literature (see for instance Ladd 1990, 1992, 1996 who propose compound prosodic domains in analogy with compound words like telephone call). In this paper, the prosody of sentences with center-embedded clauses is examined empirically in two languages.
It is proposed for German that syntactically embedded material form prosodic phrases embedded in the matrix clause, which itself forms a superordinate prosodic phrase. In other words, we propose recursion in prosodic structure of the same kind as has been proposed for syntax. Evidence for this proposal comes from the fundamental frequency (f0) register of the matrix clause which is larger if embedded material is present. It is both higher in the high regions and lower in the low regions than in matrix clauses without embedded material.
In Hindi, no embedding of prosodic phrases occur. Prosodic phrases are concatenated and are in a downstep relationship to each other, regardless of the syntactic structure of the sentences.
German and Hindi differ in the tonal and metrical phonology (German is an intonation language with lexical stress and Hindi is a phrase language without lexical stress), as well as in the prosodic phrasing feature, and it may well be the case that the two properties are related.
1. Introduction
1.1 Background
This paper investigates the prosody of main clauses accompanied by embedded sentences in German, and in Hindi. The correlates of prosody used for this investigation are pitch scaling and duration. The results of the experiments presented below may shed some light on the
1This paper is part of the project Prosody in Parsing, financed by the DFG. Many thanks to Umesh Patil who constructed the Hindi sentences and who also performed the recordings in Delhi. Thanks are also due to Caroline Magister, Verena Thießen and Georg Höhn for technical assistance, as well as to Anja Arnhold and Gerrit
issue of recursion at the interface between syntax and prosody (see Ladd 1986, 1990, Wagner 2005, Hunyadi 2006 and Hauser, Chomsky & Fitch 2002 among many others).
German has center-embedded relative clauses and sentential complements of the kind illustrated in (1).
(1) a. Die Jungen, die das Kind, das Angst hat, ärgern, sind böse.
The boys who the child who fear has tease are evil
b. Der Direktor has seine Ansicht, dass die Firma zu große Verluste einbüßt, der The director has his opinion that the firm too large losses suffers the-dat Gewerkschaft mitgeteilt.
union communicated
In such sentences, an embedded clause, or even two in the case of (1a), occur in the middle of the main clause, which, as a result, is divided into two parts. German differs from English, which iterates relative clauses and sentential complements at the end or the beginning of the sentences, as shown in (2).
(2) a. The cat has chased the dog that Mary bought
b. …Mary believes that John claims that Ann believes that Tom had painted the car.
As long as the relative clauses in (2a) are transitive, another relative clause can modify the deepest and final object. In the same way, we take it for granted that a sentence like (2b) can be extended by adding a superordinate sentence like ‘Tom doubts that’ at the beginning of the sentence.
Syntax is unable to restrict the number of embedded clauses, and prosody seems to be unconstrained as well, as it cannot limit the number of prosodic domains of a sentence. A straightforward example comes from a list like (3). If every element of the list is a prosodic phrase, it is always possible to add an element and thus add a prosodic domain.
(3) (Anna made some errands and bought) [a bottle of orange juice]P, [an apple]P, [sugar]P, [butter]P, [a pair of socks]P …
There is however an important difference between the sentences in (1) and the one in (3). In (3), the prosodic domains iterate but do not overlap. In (1), by contrast, a syntactic domain is
contained into another one. The central question addressed in this paper concerns the prosodic structure of such syntactic embeddings. Are they recursive themselves, or do they show prosodic flatness as compared to the richly hierarchical and recursive syntactic structure?
1.2 Hypotheses about prosodic structure and acoustic correlates
In the experiment presented below, sentences containing center-embedded relative clauses like (1a) were investigated for their prosody. A second kind of sentences contained parenthetical expressions, as illustrated in (4). Both types of sentences entail syntactic recursive structures, since a large sentence contains a smaller one.
(4) Die Summe, wie du weißt, wird sich verdoppeln.
The sum as you know will itself double
It is usually taken for granted in the literature that intonational phrases and prosodic structure in general contain sequences of prosodic domains that do not overlap, as required by the Strict Layer Hypothesis (see Nespor & Vogel 1986:7, Selkirk 1984:26). This means that the syntactic center-embedding of (1) is not reflected in the prosody. In this case, the prosodic structure of a sentence like (1a) is as in (5). P stands for prosodic phrase. The domains can be grouped in larger domains, like Major Phrases or Intermediate Phrases, and even higher in a single Intonation Phrase. The important point here is that there is no level of prosodic phrasing at which the prosodic domains are recursive. More formally, there is no level n of prosodic structure such that a prosodic domain n can dominate another prosodic domain n.
(5) [Die Jungen]P [die das Kind]P, [das Angst hat]P, [ärgern]P [sind böse]P
But there is an alternative to (5) in which prosodic structure involves recursion. In this case a prosodic domain of a certain level n can include another prosodic domain of the same level (see Ladd 1986 for an early proposal of recursivity of intonation phrases). Under such an approach, the prosodic structure of (1a) is illustrated in (6).
(6) [[Die Jungen [die das Kind, [das Angst hat]P, ärgern]P sind böse]P
It is important to understand that the chunks of discourse die Jungen, or die das Kind form themselves prosodic domains of a lower category, not indicated in (6). The main point is that the sentences themselves are included in joint prosodic domains, even though they are separated by embedded sentences.
Turning now to the predictions of the tonal structure, there is a widespread assumption in the literature that the second part of a matrix clause interrupted by a relative clause, a sentential complement or a parenthetical, resumes the declination after the interruption at the same melodic height as it was before the interruption, (Cooper & Sorensen 1981, Kutik et al 1983, Ladd 1986 for English). Cooper & Sorenson (1981) found for sentences like (7a) that the declination of the matrix clause was the same when the parenthetical was present as when it was not there (see Figure 1). Kutik et al. (1983) were able to reproduce the results. (7b) comes from Ladd (1986) who adopts a priori the insights of Cooper & Sorensen (1981).
(7) a. The book on the table, it seems to me, was a gift from my mother b. My brother, who is a geologist, lives in Denver
first part of main clause embedded sentence
second part of main clause Figure 1. Assumption of invisibility of the parenthetical expressions
Hunyadi demonstrates the same effect with a few sentences in Hungarian. He uses not only parenthetical expressions, but also center-embedded relative clauses of the same kind as those we use for German.
But, as far as we know, this assumption was never tested for German. A different tonal scaling behavior of the parts of the sentence, compatible with the idea of recursion as well, is possible which appears in Figure 2. In this case, when the main clause is resumed after interruption, the tonal domain is scaled to the last embedded clause.
first part of main clause embedded sentence
second part of main clause Figure.2 No reset
The production experiment described in section 2 was designed to test the prosody of matrix clauses when material like embedded sentences and parentheticals is inserted between the two parts of the main clause.
2. German experiment
2.1 Stimuli
In a production experiment, the prosody of sentences with recursive, center-embedded clauses and parenthetical expressions were tested. The same items were recorded without embeddings and parentheses for comparison. As exemplified in (8a), the recursive sentences involved copulative constructions with a center-embedded relative clause after the subject noun ([CP1]), in which a further relative clause was center-embedded ([CP2]). The counterpart without embeddings appears in (8b).
(8) a. Die Birnen, die an dem Baum, der grün ist, hängen, sind sauer.
the pears which at the tree which green is hang are sour SUBJ [CP1 [CP2 ] ] V ADJ
‘The pears, that hang from the tree, which is green, are sour.’
b. Die Birnen sind sauer.
the pears are sour
‘The pears are sour’
The parenthetical sentences involved a single parenthesis after the subject noun as shown in (9a). The counterpart without insertion is illustrated in (9b).
The set of test items comprised a total of 24 sentences: six recursive and six parenthetical sentences, plus their counterparts without embedding. The complete set is given in Appendix A.
(9) a. Die Summe, wie du weißt, wird sich verdoppeln.
the sum as you know will REFL reduplicate SUBJ [Parenthetical ] Verbal Complex
‘The sum, as you know, will reduplicate.’
b. Die Summe wird sich verdoppeln the sum will REFL reduplicate
‘The sum will reduplicate’
2.2 Recordings
The recordings took place in a soundproof cabin. Altogether, the production experiment involved the audio recordings of 21 female native speakers of German aged between twenty and thirty. The sentences were pseudo-randomized and presented to the participants in a Powerpoint presentation (one item per slide), in which the target items were separated by one or two fillers. The sentences were recorded on a DAT recorder. The participants were instructed to read the sentences at a normal speech rate and as naturally as possible. In case the participants felt that they failed to produce an utterance correctly, they were given the chance to correct themselves. In this case, only the corrected sentence was included in the analysis. A total of 504 utterances were recorded and entered the analysis.
2.3 Measurements
The analysis of the data was carried out with the acoustic analysis software Praat (Boersma &
Weenink). The recorded utterances were subdivided manually into sub-sequences. The subject noun and the first word of the matrix VP were the target words for measurements (underlined in (10)). In order to render the sentences comparable, the target words were kept constant as far as number of syllables and stress are concerned: the subject nouns were disyllabic with stress on the initial syllable and a schwa in the second syllable. All their segments were sonorous. In all sentences with relative clauses, the first word of the matrix VP was the main verb sind ‘are.’
In the parenthetical sentences, the target words are the subject noun and the first word of the matrix VP (underlined in (10)). As with the sentences with relative clauses, the subject noun is disyllabic, initially stressed, it has a schwa in the second syllable and it involves sonorous segments only. The first word of the VP is either an auxiliary or the main verb, in any case starting with a voiced syllable onset.
In the sentences with embedded relative clauses, a boundary was set at each point of embedding and ‘de-embedding’ (Hunyadi 2006) as shown in (10a). In the parenthetical
sentences, the insertion was separated by boundaries (10c). The respective counterparts without the insertion (10b,d) were subdivided into subject NP and VP. In cases where pauses occurred, these were also separated by boundaries, and measured.
(10) a. Die Birnen | die an dem Baum | der grün ist | hängen| sind sauer b. Die Birnen | sind sauer
c. Die Summe | wie du weißt | wird sich verdoppeln d. Die Summe | wird sich verdoppeln
In all sentences, the peak values of the subject noun, and of the first word of the matrix VP were measured with the help of a Praat script. These values corresponded to the H of the bitonal L*H accent realized on the subject NP, and to the beginning of the matrix VP. The targets are underlined in (10), and the red dots in Figures 3 mark the points of measurement.
In nearly all utterances, the accented word in the VP – the adjective sauer in (10a-b) and the verb verdoppeln in (10c-d) – were lower than the auxiliary at the beginning of the VP. In a few sentences, the adjective or the verb were higher, but we decided to measure the boundary between embedded sentence and matrix VP in all cases. We felt that this value is the one which establishes a relation between the two parts of the matrix clause.
If no voicing was present at a particular boundary, the point of voicing onset in the target word was measured.
Furthermore, we measured the peak values of the embedded and parenthetical material. The first part of the first embedding (CP1), the second embedding (CP2) and the second part of the first embedding (CP1) had one L*H accent respectively, whose peaks constituted the points of measurement. These accents were consistently realized on the noun in the first part of embedding 1, the adjective in embedding 2 and the verb in the second part of embedding 1 (see Figure 4). The parentheticals consistently bore one L*H accent, whose peak was measured.
We had to correct manually the outputs of the script in approximately 40% of the cases. In cases of microprosodic distortions, the f0 values were measured at a place nearby.
For each type of items, the values of all sentences were averaged across all speakers.
a. b.
Figures 3. Pitchtracks of the sentences, without (a) and with (b) embedded relative clauses
3. Results for German
3.1. Results of pitch in the main clauses
Table 1 shows the difference between the f0 values of the main clauses in sentences with two embedded sentences, and those without embedded sentences. The difference in pitch between the highest tone of the subject and beginning of the VP differs considerably in the two types of sentences. In the sentences with embedded clauses, the average difference is 67 Hz while it is 18 Hz in the sentences without embedding.
As can be gathered from Table 1, a fair amount of variation is apparent, although the general tendency that the difference is larger in sentences with embedding is always true. Especially sentence 2 deviates from the others with a relatively large difference when embeddings are present (83,74 Hz) and a relatively small one when they are not (9,95 Hz). Variation is due to the fact that some speakers started the second part of the matrix sentence at a much higher F0 level. An example is displayed in Fig. 5. Here the starting point of the second part is higher than the peak of the first part which caused a negative pitch difference in our calculation. We decided to include all values in our measurements, since the general result is not impeded by the few exceptions.
Table 1. Difference between the two measurement points in sentences with and without embeddings
with without
Sentence 1 (hosen) 67,80 19,05
Sentence 2 (jungen) 83,74 9,95 Sentence 3 (birnen) 72,65 29,89
Sentence 4 (bären) 57,73 15,34 Sentence 5 (blumen) 76,82 17,01 Sentence 6 (bauern) 45,05 16,69
Average 67,30 17,99
Die Jungen sind böse
100 400
200 300
Time (s)
0 1.42481
Figure 5. A deviating case.
Table 2 shows the values for the sentences with parentheticals. In this case, the difference between the two points of measurements betrays 75 Hz in the sentences with parenthetical expression, and 35 Hz in those without. Here, too, some special cases showed a higher value at the beginning of the VP.
Table 2. Difference between the two measurement points in sentences with and without parenthetical expressions
with without
Sentence 1 (summe) 91,59 40,65 Sentence 2 (waerme) 68,92 13,63 Sentence 3 (daemme) 83,22 56,75 Sentence 4 (bruder) 87,76 42,44 Sentence 5 (ghana) 45,63 29,46 Sentence 6 (radio) 75,46 25,32
Average 75,43 32,71
The averaged results for the measurements are plotted in a graphic in Figure 5. It is apparent that not only the high value on the subject NP are higher when embedding and parenthetical are added, but also the low values at the beginning of the VP are lower.2
Figure 5. Measured tonal values in all four types of sentences Table 3 sums up the results fort he main clauses in all sentences.
Table 3. Overall average pitch values of the main clauses
With embedding Without embedding With parenthetical Without parenthetical
NP 316 Hz 276 Hz 304 Hz 287 Hz
VP 249 Hz 258 Hz 229 Hz 252 Hz
3.2 Pitch results for the embedded sentences and for the parenthetical expressions
As can be seen in Table 4 and Figure 6 below, the pitch register of the embedded material is much narrower in comparison to the overall pitch register of the sentences. It ranges from 274
2 The difference in the measured values in the two types of main clauses can be correlated to the length of the sentences. The sentences without parentheticals were longer than those without relative sentences (see
Hz on the peak of the first part of embedding 1 to 257 Hz on embedding 2, thus a difference of 17 Hz, whereas the averaged values range over more than 60 Hz in the main clause.
Interestingly, the value of the second part of embedding 1 is higher than the one of embedding 2, but lower than the one of the first part of embedding 1. Even more relevant for this paper are the interplays between downstep and reset between all points of measurements.
Table 4. Average pitch values of the main clauses and embeddings for the relative sentences.
sentence main clause embedding 1 embedding 2 embedding 1 main clause
1 (hosen) 324,89 269,79 245,67 267,56 256,90
2 (jungen) 334,18 300,83 259,60 269,54 250,45
3 (birnen) 317,17 275,02 276,82 280,71 244,51
4 (bären) 309,19 269,54 273,12 269,11 251,46
5 (blumen) 321,41 273,26 249,39 259,15 244,59
6 (bauern) 291,99 254,80 237,89 256,71 246,95
average 316,47 273,88 257,08 267,13 249,14
Figure 6 displays the average measurement points in the sentences with center-embedded relative clauses in form of a diagram. It shows downstep between the beginning of all three clauses, but after that, the second part of the first relative clause is reset. The second part of the main clause is clearly downstepped relatively to all preceding values. The most interesting aspect is the position of measurement 4, which is the second measurement of the higher relative clause, after finition of the most deeply embedded relative clause.
316,47
273,88
257,08 267,13
249,14
200,00 220,00 240,00 260,00 280,00 300,00 320,00 340,00
1 2 3 4 5
Figure 6. Diagram of the average values from Table 4.
The average pitch values of the parenthetical expressions range from 258 Hz to 282 Hz. Table 5 and Figure 7 show that the average peak value of the parentheticals (270,16 Hz) approximates the middle of the overall pitch range (304,11 Hz to 228,68 Hz).
Table 5. Average pitch values of the main clauses and insertions for the parenthetical sentences.
First part of main Parenthetical Second part of main
summe 327,88 273,14 236,28
wärme 306,94 282,26 238,02
dämme 300,93 279,44 217,71
bruder 314,80 262,17 227,05
ghana 281,06 258,17 235,43
radio 293,06 265,81 217,59
Average 304 Hz 270 Hz 229 Hz
304,11
270,16
228,68
200,00 220,00 240,00 260,00 280,00 300,00 320,00 340,00
1 2 3
Figure 7. Diagram of the average values from Table 7.
3.3 Results of the duration measurements
Duration was measured before and after the embedded clauses, as illustrated in (11). Break 1 is at the embedding position of the first relative clause, break 2 at the embedding position of the second relative clause, break 3 at the de-embedding position of the second relative clause, and break 4 at the de-embedding position of the first relative clause.
(11) Die Birnen
|
die an dem Baum|
der grün ist|
hängen|
sind sauer.break1 break 2 break3 break 4
Table 6 shows the averaged break time realized by the speakers at the different positions. A great deal of variation in the insertion of breaks can be observed, but still a few interesting observations observe from the calculations. De-embedding causes a much larger break than embedding, in line with Hunyadi’s (2006) results for Hungarian. Break 4, at the de- embedding position of the first relative clause is the strongest of all four. The second strongest is at the position of de-embedding of the second relative clause, thus the most deeply embedded ones. The other positions revealed only weak breaks.
Table 6. Breaks in sentences with embedded clauses
Break 1 (n=7)
(NP|CP1)
Break 2 (n=5) (CP1|CP2)
Break 3 (n=3) (CP2|CP1)
Break 4 (n=30) (CP1|VP)
hosen 0,068 0,0923 0,1819
jungen 0,1755 0,2633 0,1922
birnen 0,0626 0,1334
bären 0,1425 0,1645 0,1536
blumen 0,0346 0,1293
bauern 0,0436 0,19 0,0825
Sum 0,4152 0,2761 0,5456 0,8729
Sum divided by total number of speakers (n=21)
0,0197 0,0131 0,0259 0,0415
The same appears in Table 7 for parenthetical sentences. These sentences only have two break positions, one between the first part of the main clause and the parenthetical, and the other between the parenthetical and the second part of the main clause. Only break 2 was made on a regular basis, and this not even by all speakers.
(12) Die Summe
|
wie du weißt|
wird sich verdoppeln break 1 break 2Table 7. Breaks in sentences with parenthetical clauses
Break 1 (n=4) (NP|Parenth)
Break 2 (n=16) (Parenth|VP)
summe 0,0827 0,0741
waerme 0,0543 0,0939
daemme 0,0946
bruder 0,0576 0,1725
ghana 0,1155
radio 0,0592
Sum 0,1946 0,6098
Sum divided by total number of speakers (n=21)
0,0092 0,029
3.4 Discussion
The results obtained for the tonal scaling of the two parts of the main clause do not confirm assumption 1, namely that the embedded part of the sentence is invisible for the matrix clause.
Recall that this hypothesis posited that there is no difference in the tonal scaling of the matrix clause when it is continuous and when it is separated by embedded and parenthetical material.
This was illustrated in Figure 1. Indeed, in the German sentences, a clear effect of length was found. The longer sentences, with embedded clauses and with parentheticals started higher and ended lower than the same sentences without embeddings. In the experimental sentences without embedding, the difference in the tonal height between the subject and the verb betrayed an average of 18 Hz, but in the sentences with embedding, the tonal lapse was 67 Hz, a large difference (see Table 1). Also in the sentences with parentheticals the difference was large, with 32 Hz in the simple sentences and 75 Hz in those with a parenthetical expression (Table 2). Perceptively, the effect of cutting out the embedded material renders the remaining sentences bizarre, as if the speaker starts with a narrow focus on the subject and changes her mind in the VP part. Consequently, our results are not in line with results for English (Cooper
& Sorensen 1981, Kutik et al. 1983, Ladd 1986) or for Hungarian (Hunyadi 2006).
Assumption 2, which posited that the second part of the matrix clause is scaled relatively to the embedded part (see Figure 2) is not wrong, but it is insufficient as an explanation of our data. In assumption 2, the difference between the two versions of the sentences – with and without embedding – was expected to be entirely in the scaling of the second part of the sentence respectively to the embedded part. The first part of the main clause was assumed to be identical in the two kinds of sentences. This expectation was not fulfilled. As clearly visible from Figure 5, both the value of the subject and the one of the verb
of the subject is higher than in sentences without embedding, and the second measured point is lower. In other words, the entire range used in the sentences with embedding is larger than the one used in the sentences without embedding.
4. Hindi Experiment
Sentences with two center-embedded clauses were recorded in Hindi, as a part of a larger experiment including different kinds of production sentences. The experiment was carried out as a Powerpoint presentation. Participants were displayed the sentence on the screen.
Participants were instructed to speak out the sentences displayed on the screen as naturally as possible. Once they answered the question without any hesitations the next stimulus was presented. In case of hesitations they were asked to repeat the sentence. Presentation flow was controlled by the experimenter. Participants were allowed to take a few minutes break when they wanted to do so. A set of instructions and two practice examples familiarized the subjects with the process.
The entire data used in this experiment was collected in one experiment, run individually with 30 speakers. All were female students of the University of Delhi, India. They were native speakers of Hindi in their twenties, coming mostly from Delhi and surrounding states. They were paid for their participation. The target sentences were recorded in the University of Delhi on a DAT tape recorder. Out of 30 speakers, data from only 20 speakers was used for analysis. Selection of items and speakers was not based on any criteria, but was completely random. A total of 120 sentences (20 speakers x 6 items) were analyzed.
An example of sentence appears in (13).
(13) Example of Hindi sentence
[Woh naukar]1 [jisko us phalavaale ne]2 [jo bhaag rahaa thaa]3 [jagaayaa]2 [gariib taa]1
That servant whom that fruit-seller ERG who running was woke up poor was
‘That servant whom the fruit seller who was running woke up was poor.’
More measurements were made in the Hindi sentences than in the German ones, as illustrated in Figure 8. As we did for the German data, the recorded utterances were subdivided manually into sub-sequences. In every domain marked with a number, three measurements were performed: the first lowest tone, the highest tone and the following lowest tone, but in many cases, one of the low tones was missing. This happened when the portion considered was
ending in a high tone, a pattern which was pervasive due to the tonal structure of Hindi (see Patil et al 2008, Féry 2009). Analysis of the data was carried out with the acoustic analysis software Praat (Boersma & Weenink), and corrected manually when necessary.
l H l H l H l H l H l H l H l
1 2 3 4 5 6 7
woh naukar jisko us phalavaale ne jo bhaag rahaa thaa jagaayaa gariib thaa
that servant whom that fruit-seller ERG who running was woke up poor was
100 400
200 300
Time (s)
0 4.5856
Figure 8. Pitch track of a Hindi sentence
5. Results for Hindi
Results of the F0 measurements are displayed in Figures 9 and 10. Every peak corresponds to the highest tone in the domains marked in Figure 8.
As is visible from the figures, the pattern differs from the one in German. No reset occurs at the beginning of the second part of the relative clause (the second CP1), even though, the value at this point is as high as the last value of CP2. However, compared to the first value of CP2, this point is clearly downstepped. The same is true of the transition between relative clause 1 and relative clause 2. Here, too, the two adjacent values are more or less equal. But compared to the first value of CP1, the relative clause CP2 is downstepped.
NP CP1
CP1 CP2
CP2 CP1
VP
150,00 170,00 190,00 210,00 230,00 250,00 270,00 290,00 310,00
Figure 9. High and low tones of the Hindi sentences
Fig. 10 Another representation of high and low tones of the Hindi sentences
Breaks are only scarce and do seem to be play any role in the timing of the sentences.
6. Discussion
Because of the difference between the results obtained in the two languages, a unique prosodic structure valid for all cases does not seem to be possible. We suspect that the reset
pattern at the de-embedding location in German as compared to the absence of reset in Hindi reflects a crucial difference in the prosodic organization of the two languages.
We propose that embedding of prosodic domains, as in (6), accounts best for the obtained data in German, but that the structure proposed in (5) accounts best for Hindi. Focusing first on German, the superordinate prosodic domain containing embedded clauses or parenthetical expressions has an enlarged register domain, as shown in Figure 11. In the sentences with relative clauses or with a parenthetical expression, the register of the subordinate sentence is subordinate to the register of the main clause.
Main clause
NP 1st relative clause
CP1 2nd relative clause
CP2 CP1 V
Figure 11. Recursive p-phrasing in German
The fact that the second part of the main clause is not reset higher than the top lines of the relative clauses is explained by the concept of register enlarging. If a non-recursive p-phrase has a range called r, the same phrase containing embedded p-phrases is enlarged by a factor f, and embedded p-phrases have a narrower register, decreased relatively to r by a factor -f.
The idea is that the prosodic domain, containing the main clause in our sentences, is enlarged in order to be able to contain the entire register of the embedded clauses. The embedded clauses themselves, have a reduced register, in order to be able to be contained in the larger one.
Ladd (1996:240) remarks that under the Strict Layer Hypothesis, the only possible differences of boundary strength are differences of boundary types. In the approach proposed here, this is not true. Here, we posit a prosodic structure consisting of prosodic domains that we called p- phrases, and which can be recursive. In doing so, we adopt an emergent proposal regarding the prosodic domains as recursive (see Wagner 2005, Ito & Mester 2008). The motivation for recursive prosodic structure is that it is not always clear what exact level of prosodic structure is relevant. Moreover, as already noticed in the first section of this paper, it is not possible to restrict the number of domains authoritatively, since it is always possible to add an higher, intermediate or lower level of phrasing. The reset pattern that we find in our data, is a third argument for recursivity. It is to be observed though that we may call the largest phrase intonation phrase, the one around the first relative clause major phrase, and the one around the
second relative clause minor phrase. But once we have done that, we predict that it is impossible to add a prosodic domain, because we would be at a loss of naming the domain (see also Féry & Kentner 2008 for name grouping). Moreover, except for the register, which may be wider or narrower, there is no reason to name the prosodic domains with different labels.
Turning now to Hindi, we adopt the structure (5) for its prosodic phrasing, disallowing recursion. Here is sentence (13) with the intended prosodic structure
(14) Prosodic phrasing of Hindi
[Woh naukar]P [jisko us phalavaale ne]P [jo bhaag rahaa thaa]P [jagaayaa]P [gariib taa]P
That servant whom that fruit-seller ERG who running was woke up poor was
‘That servant whom the fruit seller who was running woke up was poor.’
Schematically, the F0 scaling obtained is displayed in Figure 12. Every p-phrase is just in a downstep relation with the preceding one.
Main clause
NP 1st relative clause
CP1 2nd relative clause
CP2 CP1 V
Figure 12. Non recursive p-phrasing in German
In short, the prosodic structure that we propose for German and Hindi respectively differ in one main aspect: German allows recursion and embedding of p-structure and mimics in this way the syntactic structure, whereas Hindi does not allow recursion, but simply iterates its prosodic domain.
It is an open question how this main distinction correlates with the difference in tonal structure and in lexical stress between the two languages.
6. Conclusion
Sentences containing center-embedded relative clauses were explored in a production experiment in German and in Hindi. In German, sentences containing parentheticals were also studied. For German, the complex sentences were compared with the same sentences without embeddings or parentheticals. In Hindi, only the sentences with embedded relative clauses
were recorded. The aim of this study was to understand whether syntactic recursion is mapped to a similar prosodic structure, and whether this is a universal property. The results revealed a significant difference between the two languages.
In German, the values of the initial pitch accents are lower in sentences without embedding than in sentences with embedding, whereas the final pitch accents are higher, showing that the overall register used is reduced in the short sentences. Contrary to what has been described for English and Hungarian, it not the case that the second part of a main clause starts at the same level than in a sentence containing embedded material. On the contrary, when one cuts away the embedded or parenthetical material from a long sentence, the result sounds awkward, as if the subject was pronounced much too high and/or the VP much too low.
The main result of the experiment, however, lies in the interesting downstep and reset pattern in the sentences with embedding, a result that was not reproduced in Hindi. In German, the second part of the first relative clause is reset relatively to the second relative clause, but downstepped relatively to the first part of the first relative clause. And the last part of the main clause is downstepped as compared to all other parts of the sentence. We interpret this result as an indication that the prosodic structure in German is recursive, in the sense that one prosodic phrase of a certain level n can be interrupted by an embedded prosodic phrase of the same level n.
In Hindi, we find a different pattern. Here, it is not the case that the second part of the first relative clause is reset after the second relative clause. All parts of the sentence are in a downstep relationship to each other (considering only the initial high tones of every parts).
We interpret this pattern as non-recursive. The prosodic structure does not map the syntactic structure in the same way as in German.
To conclude, this experiment is in line with the results of Féry & Kentner (2008) in which the prosodic structure of name groupings was examined. There, too, it became apparent that German follows the syntactic pattern in a much more fine-grained fashion than Hindi.
Appendix
Test Items (Translations are missing)
1 Recursive Sentences and counterparts
1a. Die Birnen, die an dem Baum, der grün ist, hängen, sind sauer.
1b. Die Birnen sind sauer.
2a. Die Hosen, die der Mann, der arm ist, trägt, sind billig.
2b. Die Hosen sind billig.
3a. Die Jungen, die das Kind, das Angst hat, ärgern, sind böse.
3b. Die Jungen sind böse.
4a. Die Bären, die im Wald, der naturbelassen ist, leben, sind friedlich.
4b. Die Bären sind friedlich.
5a. Die Blumen, die der Junge, der einsam ist, sammelt, sind rosa.
5b. Die Blumen sind rosa.
6a. Die Bauern, die den Mais, der reif ist, ernten, sind emsig.
6b. Die Bauern sind emsig.
2 Parenthetical Sentences
1a. Die Summe, wie du weißt, wird sich verdoppeln.
1b. Die Summe wird sich verdoppeln.
2a. Die Wärme, wie du spürst, macht uns sehr zu schaffen.
2b. Die Wärme macht uns sehr zu schaffen.
3a. Die Dämme, wie du siehst, drohen bald zu brechen.
3b. Die Dämme drohen bald zu brechen.
4a. Mein Bruder, wie du weißt, lebt bald in Brasilien.
4b. Mein Bruder lebt bald in Brasilien.
5a. In Ghana, wie du weißt, gibt es viele Löwen.
5b. In Ghana gibt es viele Löwen.
6a. Im Radio, wie du hörst, läuft gerade Werbung.
6b. Im Radio läuft gerade Werbung.
References
Cooper, W. & J. Sorensen (1981) Fundamental frequency in sentence production. Heidelberg:
Springer.
Féry, Caroline (2009). The intonation of Indian languages: an areal phenomenon. In: Imtiaz Hasnain and Shreesh Chaudhury (eds). Festschrift to appear.
Féry, C. and G. Kentner (2008) The prosody of grouping names in German and Hindi. Ms.
Potsdam. Submitted.
Féry, C. and H. Truckenbrodt (2005). Sisterhood and tonal scaling. Studia Linguistica 59.2/3.
Hauser, Marc D., Noam Chomsky, W. Tecumseh Fitch (2002) The Faculty of Language:
What Is It, Who Has It, and How Did It Evolve? SCIENCE 298, 2769-79.
Hunyadi, L. 2006. Grouping, the cognitive basis for recursion in language. Argumentum 2, 67-114
Ito, Junko & A. Mester (2008) The extended prosodic word. In Grijzenhout, J. & B. Kabak (Eds.) Phonological Domains: Universals and Deviations. Mouton de Gruyter: Interface Explorations Series.
Kutik, E., W. Cooper & S. Boyce (1983) Declination of fundamental frequency in speakers’
production of parenthetical and main clause. Journal of Acoustics Society of America 73, 1731-1738.
Ladd, D.R. (1986) Intonational phrasing: the case of recursive prosodic structure. Phonology Yearbook 3. 311-340.
Ladd, D. R. (1990). Metrical representation of pitch register. In J. Kingston, & M. Beckman (Eds.), Papers in laboratory phonology 1: between the grammar and the physics of speech (pp. 35–57). Cambridge: Cambridge University Press.
Ladd, D.R. (1992) Compound Prosodic Domains. Occasional Papers of the Linguistics Department. University of Edinburgh.
Ladd, D.R. (1996) Intonational phonology. Cambridge University Press.
Nespor, Marina & Irene Vogel (1986) Prosodic phonology. Dordrecht: Foris.
Patil, U., G. Kentner, A. Gollrad, F. Kügler, C. Féry and S. Vasishth (2008). Focus, word order and intonation in Hindi. Journal of South Asian Linguistics 1. 53-70.
Selkirk, E. O. (1984) Phonology and syntax. The relation between sound and structure.
Cambridge: MIT Press.
Wagner, M. 2005. Prosody and recursion. Doctoral Dissertation, Massachusetts Institute of Technology, Cambridge, Mass.
caroline.fery@googlemail.com fabian.schuboe@googlemail.com
