Prominence in Triconstituent Compounds: Pitch Contours and Linguistic Theory

1.1 Prominence in triconstituent compounds: hypotheses

It is generally assumed that prominence assignment to triconstituent compounds depends on the branching direction of the compound. Left-branching compounds have highest prominence on the left-most constituent, whereas right-branching compounds, that is, compounds with a right-hand complex constituent, have highest prominence on the second constituent of the whole compound. The LCPR proposed by Liberman and Prince (1977) has been evoked to account for this generalization. The LCPR labels metrical trees on the basis of strong–weak relations between two sister constituents. Hence, one constituent is always strong (S), that is, more prominent, in relation to its weaker sister constituent (W). In particular, the LCPR makes a prediction as to which constituent of a binary construction will be prosodically strong.

In a configuration [A B]: if C is a lexical category, B is strong if it branches. (Liberman & Prince, 1977, p. 257)

The predictions of the LCPR are illustrated in Figures 1 and 2. In addition to the labels ‘S’ and ‘W’ (strong and weak, respectively), the ‘N-level’ refers to the level of the embedded NN compound and ‘IC-level’ refers to the level between the complex and the single constituent, with ‘IC’ meaning ‘immediate constituent’. The tree in Figure 1 shows a left-branching compound (labeled ‘L’), while Figure 2 shows a right-branching compound (labeled ‘R’).

OPEN IN VIEWER

Figure 1.

Metrical tree of a left-branching compound.

OPEN IN VIEWER

Figure 2.

Metrical tree of a right-branching compound.

Sproat (1994) revises the LCPR in such a way that prominence assignment is seen as the result of deaccentuation: according to his np system, the second constituent of a compound retains its accent only if it is complex.

For each node C dominating [A B], if B is not complex, and if C is N⁰, then B is deaccented; else both A and B retain their accents. (Sproat, 1994, p. 84)

Yet, even Sproat’s revised version of the LCPR is only poorly empirically supported with respect to prominence assignment in complex compounds. In fact, available studies dealing with prominence assignment to NNN constructions (e.g., Berg, 2009; Giegerich, 2009; Kösling & Plag, 2009; Kvam, 1990) have shown that the LCPR fails to predict the correct prominence pattern for a considerable number of compounds.

For instance, Kvam (1990) investigated 40 NNN constructions in a production experiment. Even if his methodology remains somewhat unclear, Kvam found that the majority of the investigated compounds, namely 30 out of 40, were produced either exclusively or by the majority of the experimental subjects with prominence on constituent N2. Yet, Kvam points out that only 10 of these compounds were also clearly right-branching, that is, the group of compounds that should indeed have this stress pattern. Hence, only in 10 cases out of 40, stress assignment could be directly related to the branching direction of the compound.

Apart from Kvam’s study, additional evidence towards variation in the prominence assignment of NNN constructions is provided by Berg (2009). Taking an explorative approach by looking at a total of 642 NNN combinations taken from the British National Corpus (Berg, 2009, p. 87), Berg finds that 57.2 percent of the combinations are prominent on constituent N2, and 26.5 percent on N1. Thus, Berg’s findings go in the same direction as Kvam’s results, revealing a general tendency for triconstituent compounds to be in their majority most prominent on the second constituent, be they left- or right-branching. In addition to that, Berg also provides information about a number of right-branching compounds with prominence on constituent N1 and N3, as well as left-branching compounds with prominence on N3. Similar counter-examples are also provided in a more recent approach by Giegerich (2009). His study investigates the traditional English Compound Stress Rule, and, in contrast to Kvam and Berg, explicitly argues against the LCPR and its predictions.

Finally, a considerable number of counter-examples to the LCPR were also documented in a recent corpus study by Kösling and Plag (2009), who tested the predictions of the LCPR by analyzing about 500 compounds taken from the Boston University Radio Speech Corpus (Ostendorf, Price, & Shattuck-Hufnagel, 1996). Their analysis of the corpus data revealed a general trend for left-and right-branching compounds to behave as predicted by Liberman and Prince’s LCPR, that is, left-branching compounds tended to be prominent on constituent N1, and right-branching compounds on constituent N2. Crucially, however, their corpus data also revealed a significant proportion of compounds violating the rule, both at the level of the immediate constituents (IC-level) and at the level of the final nodes of the tree (N-level). Table 1 lists some of these N-level and IC-level violations. Note that some of the examples may appear to allow different branching interpretations than the one given in the table. Yet, the context in which the triconstituent compounds occurred in the corpus clearly shows that the given interpretation of the internal structure is the most plausible one.

OPEN IN VIEWER

Table 1.

Lexical Category Prominence Rule violations.

Branching direction	Violation at N-level	Violation at IC-level
Left-branching	[science fíction] shocker	[Boston area] commúnities
	[capital gáins] tax	[weekend] séries
Right-branching	state [health prógram]	commúnity [meeting hall]
	school [drug úse]	crédit [scoring system]

The examples listed in Table 1 raise the question of which factors are responsible for their aberrant prominence behavior. ² Based on their data, Kösling and Plag (2009) argue that the violations at the N-level and the IC-level may both be explained by the presence of right-prominence in NN compounds. Previous studies have shown that there is a host of right-prominent NN constructions, and their frequency has been estimated to be roughly 30 percent of all NN constructions (cf. Bell & Plag, 2012; Kunter, 2011; Plag, 2010; Sproat, 1994).

Kösling and Plag (2009) propose that the violations at the N-level arise due to this presence of embedded right-prominent NN compounds: for example, science fiction shocker has prominence on the second constituent because the complex left constituent consists of the right-prominent science fiction; the same assumption is also put forward by Giegerich (2009). This prominence pattern is not predicted by the LCPR or the np system, because the existence of right-prominent compounds is not possible under the formulation of the LCPR: in Liberman and Prince’s (1977) approach, right-prominent NNs (apparently including science fiction) are simply assumed to be syntactic phrases, and the prominence pattern of phrases is assumed to follow a different prominence rule than the LCPR. The same problem is present in the np system, which also hinges on a clear distinction between compounds and phrases. For example, Sproat (1994) claims that there is a structural difference between Párk Street (which is considered a compound) and Madison Ávenue (which is considered not a compound but a phrase; Sproat, 1994, p. 83). His explanation of this difference appears to be based on the prominence pattern alone, thus leading to a circular argument: his np system will only yield right prominence for Madison Avenue if it is a phrase, but the claim that Madison Avenue has to be a phrase seems to be derived from the stress pattern alone. Indeed, it has been shown (e.g., in Bauer, 1998; Giegerich, 2009) that the distinction between compounds and phrases is very difficult to draw on formal grounds, and Park Street and Madison Avenue are a case in point.

If we apply the np system to triconstituent compounds, the patterns in Table 1 involving right-prominent NN compounds are unpredicted by the np system. For science fiction shocker, the algorithm would deaccentuate fiction (unless it is considered a phrase, but there is no independent evidence for this), and also deaccentuate shocker, because it is not complex, yielding scíence fiction shocker with prominence on the first constituent, not on the second as in the corpus data. Likewise, if we assume drug use to be a compound, use would be deaccentuated because it is not complex, so the resulting prominence pattern would be school drúg use, which conflicts with the attested school drug úse.

With reference to the IC-level violations, Kösling and Plag (2009) suggest that the same factors responsible for prominence variation in NN compounds are also responsible for prominence variation in NNN compounds. The literature on prominence variation in NN compounds usually discusses three groups of factors, namely structure, semantics and analogy. Experimental and corpus studies testing these factors (e.g., Arndt-Lappe, 2011; Kunter, 2011; Plag, 2006; Plag, Kunter, Lappe, & Braun, 2008) found that, in particular, semantics and analogy play a role, with structure being less significant (a more detailed account of these factors can be found in section 2.1 below).Hence, right-prominent NN compounds are responsible for many instances of aberrant prominence patterns at the IC-level, because the same mechanisms operating in biconstituent compounds, that is, at the N-level, may also operate at the IC-level of a given compound.

The present analysis remains agnostic about this ongoing debate on the factors affecting the prominence patterns of compounds at the IC-level, and concentrates on the N-level violations instead, for which we can formulate explicit predictions. We want to test Kösling and Plag’s (2009) and Giegerich’s (2009) hypothesis regarding the violations at the N-level with a new methodological instrument. We focus on the question as to whether the prominence pattern of the embedded NN compound affects prominence assignment in triconstituent compounds, as suggested by these authors. In particular, we want to test whether left-branching compounds with embedded right-prominent NN compounds are prominent on constituent N2, and whether right-branching compounds with embedded right-prominent NN compounds are prominent on constituent N3. These two cases would be unpredicted by the LCPR. Furthermore, we want to test whether compounds with embedded left-prominent NN compounds are prominent on constituent N1 (in the case of left-branching compounds), or on constituent N2 (in the case of right-branching compounds), as it is predicted by the LCPR. We may reformulate these assumptions as two related hypotheses, which, taken together, will be referred to as the Embedded Prominence Hypothesis (EPH, proposed by Kösling, 2011) for the rest of this paper.

So, in effect, the EPH in (3) claims that in a triconstituent compound that has highest prominence on the complex constituent, the highest prominence falls on the constituent that is also prominent in the embedded compound. Accordingly, the hypothesis makes predictions for four different types of compounds: on the one hand, left-branching compounds with either an embedded left-prominent NN compound or an embedded right-prominent NN compound, and on the other hand, right-branching compounds with either an embedded left-prominent NN compound or an embedded right-prominent NN compound. These four types will be referred to as ‘L/N1’, ‘L/N2’, ‘R/N2’ and ‘R/N3’ in the remaining sections of this paper. The labels ‘L’ (left) and ‘R’ (right) indicate the branching direction of the NNN compound at the IC-level. The labels ‘N1’, ‘N2’ and ‘N3’ refer to the constituent in the compound that the EPH predicts to receive highest prominence in the compound. For example, right-branching compounds with a left-prominent complex constituent are expected to have highest prominence on the left-hand constituent of the embedded complex constituent, which is, as seen in Figure 2, constituent N2. This type of compound is therefore labeled ‘R/N2’. In contrast, ‘R/N3’ compounds are right-branching compounds with embedded right-prominent NNs, as the EPH predicts that the right-prominence is preserved under embedding, and highest prominence should therefore fall to constituent N3. Similarly, ‘L/N1’ and ‘L/N2’ compounds are left-branching compounds with left- and right-prominent complex constituents, respectively. Table 2 summarizes the four NNN compound types.

OPEN IN VIEWER

Table 2.

Classification of different types of branching direction and their embedded prominence patterns predicted by the Embedded Prominence Hypothesis.

Label	Branching direction	Prominence pattern of embedded NN	Expected prominence pattern of NNN	Example
L/N1	Left	[ŃN]	[ŃN]N	[háy fever] treatment
L/N2	Left	[NŃ]	[NŃ]N	[science fíction] book
R/N2	Right	[ŃN]	N[ŃN]	business [crédit card]
R/N3	Right	[NŃ]	N[NŃ]	family [Christmas dínner]

In order to empirically test the EPH, one does not only need large amounts of pertinent compounds, but also a reliable method of determining the actual prominence pattern for each of these compounds. This is not a trivial task, and this paper will put forward a new methodology that is able to overcome certain shortcomings of previous approaches.

1.2 Pitch and prominence

The autosegmental-metrical framework of intonational phonology (for a summary, see, e.g., Ladd, 1996) proposes that prominence differences in multi-word constructions are linked to the presence of pitch accents on some of the words in such a construction. In principle, pitch accents are tonal targets that are aligned with stressed syllables and that shape the pitch contour of an intonational phrase. In English, these tonal targets can be either high (H*) or low (L*), or in the case of bitonal pitch accents, a combination of both (L*+H, L+H*). Perceptually, words that are accented are perceived as more prominent than unaccented words, and the last accent in a phrase (the nuclear accent) has highest prominence. In addition to pitch accents, tonal targets can also be aligned with the right edge of phrases. These phrase accents and boundary tones are not assumed to be prominence-lending, but provide acoustic cues to the phrasing of larger utterances, as well as the pragmatic coherence of the phrases (see Beckman & Edwards, 1994; Beckman & Pierrehumbert, 1986; Gussenhoven, 2004; Pierrehumbert, 1980, for detailed descriptions of the intonational system of English, and Hirschberg, 2004, for a summary of the link between intonation and pragmatic functions).

The autosegmental-metrical account assumes a particular relation between perceived prominence and notable changes in pitch height, and the temporal distribution of these changes (see Arvaniti, Ladd, & Mennen, 1998, and Dilley, Ladd, & Schepman, 2005, for discussions of the timing of pitch contours): listeners are attuned to the shape of the pitch contour, and perceive those elements as particularly prominent during which the pitch contour shows particular events. This link between pitch and prominence has been supported by the findings in studies such as Rietveld and Gussenhoven (1985) and Terken (1997): syllables received increasingly higher prominence ratings with increasing pitch excursion sizes. Rietveld and Gussenhoven (1985) have shown that for Dutch, a pitch excursion of 1.5 ST is a sufficient cue to prominence; Terken (1997) accordingly concludes that higher prominence ratings indeed appear to be proportional with the size of pitch excursions. It is noteworthy, though, that this proportional relation relates only to single pitch peaks, that is, to situations in which there is just a single pitch accent, but it does not account for the relative perceptual prominence in sequences of pitch accents.

Indeed, it has been shown (e.g., in Gussenhoven, Repp, Rietveld, Rump, & Terken, 1997; Gussenhoven & Rietveld, 1988; Ladd, Verhoeven, & Jacobs, 1994; Rietveld & Gussenhoven, 1985) that in a sequence of two pitch peaks, the second peak is perceived to be as prominent as the first peak even if the second pitch excursion is lower than the first one, which accounts for the claim that the nuclear accent in an intonational phrase is perceived as most prominent.

Thus, in sum, there is empirical support for the claim made in intonational phonology that accentuation plays a crucial role in determining the distribution of prominence in multi-word sequences in languages such as English or Dutch. By extension, this claim can be assumed to apply not only to phrases, but also to compounds. Gussenhoven (2004), for instance, proposes an Optimality Theory-based based description of accentuation in English that he assumes to apply to both phrases and compounds. However, until recently there was only weak empirical support for the assumption that prominence patterns in compounds can also be attributed to accentuation differences.

Farnetani et al. (1988) present an early investigation of the acoustic correlates that distinguish left-prominent and right-prominent NN compounds in English. The authors show that duration, intensity and maximum pitch show significant differences for the two prominence patterns. With respect to pitch, the difference between maximum pitch measurements in the first and second elements is found to be only small in right-prominent compounds, but large in left-prominent compounds. Farnetani et al. interpret this in terms of pitch accent distribution and conclude that in the former case both elements are accented, while in left-prominent compounds, only the first element carries an accent. Later, Kunter (2011) explicitly tests this interpretation in an analysis of pitch, intensity, duration and spectral balance in NN compounds. His analysis provides further robust support for the hypothesis that prominence patterns in NN compounds are distinguished by the presence or absence of a pitch accent on the second element, while the first element is always accented.

Turning to the four different types of triconstituent compounds described in Table 2, we would predict the following distribution of pitch accents. L/N1 compounds should have only one accent on the first constituent, while the other constituents should be unaccented. L/N2 should have an accent on N1 and on N2, because the complex constituent has the same distribution of accents as a right-prominent NN compound. The same pattern should be found in R/N2 compounds, but for slightly different reasons: the complex right constituent is a left-prominent NN compound, and therefore, N2 should be accented. At the same time, N1 should also be accented, because at the IC-level, R/N2 compounds are expected to be right-prominent, and in right-prominent compounds, the left constituent should be always accented. Finally, R/N3 compounds should have an accent on all three constituents: N2 and N3 together form a right-prominent NN compound (with accents on both constituents), and N1 should be accented in R/N3 compounds due to the same reasons as in R/N2 compounds.

In sum, previous studies have shown that there is indeed a relation between pitch and prominence in compounds, and that this relation can be described in terms of accentuation. As the relation between prominence and pitch can be quantified, it offers an opportunity to test statistically whether different types of compounds are produced with different prominence patterns. Using acoustic data obtained from speech corpora and production experiments, this methodology has been employed in recent studies of NN compounds (Farnetani et al., 1988; Plag, 2006; Plag et al., 2008; Štekauer, Zimmermann, & Gregová, 2007) and NNN compounds (Kösling, 2011; Kösling & Plag, 2009). However, the way that pitch is measured in all of these studies suffers from an identical weakness, namely that the researchers looked at rather crude abstractions of the pitch contour in the investigated compounds. For instance, Farnetani et al. (1988) measure the peak pitch in each compound constituent, Plag (2006) measures pitch at the mid-point of the syllables with primary stress, while Štekauer et al. (2007), Plag et al. (2008), Kösling and Plag (2009) and Kösling (2013) measure pitch averages in each constituent. Any information not captured by these measurements is ignored in the analysis, which is why these measurements run the risk of ignoring potentially systematic variation in the pitch contour that may provide perceptually important cues to prominence.

The movement direction of the pitch contour, that is, whether pitch is falling or rising, may be one of these cues. Terken and Hermes (2000), who have investigated how far pitch contributes to the perceived prominence of syllables, have found that a fall is perceived as more prominent than a rise of the same excursion size. Yet, all pitch measurements employed in the earlier literature on compound prominence are one-dimensional representations of the pitch contour in the sense that they describe only localized pitch events, and as such are incapable of registering any kind of pitch movement. ³ Furthermore, only parts of the pitch contour are considered in these studies, for instance the nucleus (Štekauer et al., 2007) or the rime (Plag et al., 2008) of the stressed syllables in each compound constituent. However, as Arvaniti et al. (1998) have shown, the tonal targets of bitonal pitch accents are not always aligned with the accented syllable itself. Thus, it might be possible that the second tonal target of L+H* accents does not fall within the measurement intervals used in the aforementioned studies. In such a case, the potentially prominence-lending high pitch of this accent type would be missed in the analysis.

These complications show that a methodology that considers the complete pitch contour in compounds may be more successful in detecting prominence differences than previous approaches, which were restricted to portions of the contour. Accordingly, the present paper presents a statistical model of the pitch contour in the four different types of English triconstituent compounds described in the previous section. In this context, we understand the intonation contour of a compound as the combination of the contours of the involved pitch accents (plus any boundary tones and phrase accents). For the task of modeling this pitch contour, we will use GAMs, a flexible and powerful statistical technique that is particularly suited for the problem at hand. This new method circumvents the limitations of previous studies and, at the same time, allows one to detect statistically significant differences between the contours that may be interpreted as different prominence patterns. The details of the statistical analyses will be explained below.

2.1 Target compounds

For the production experiment, a total number of 40 NNN compounds was devised, 10 of each type described in Table 2. Thus, in order to test whether left-branching compounds have highest prominence on constituent N1 and right-branching compounds on constituent N2 if the complex constituent is left-prominent, we constructed 10 compounds of the L/N1 type and 10 compounds of the R/N2 type. Furthermore, in order to test whether a right-prominent embedded NN compound causes left-branching compounds to have highest prominence on constituent N2 we constructed 10 compounds of the L/N2 type. Finally, in order to test whether a right-prominent embedded NN compound causes right-branching compounds to have highest prominence on constituent N3, we constructed 10 compounds of the R/N3 type. The appendix lists all 40 target compounds.

The prominence pattern of the embedded NN compounds was controlled by means of various American English dictionaries: Oxford Student’s Dictionary of American English (Hornby, 1983), Longman Dictionary of American English (Longman, 2002), Longman Advanced Dictionary of American English (Summers, 2000), Oxford Advanced Learner Dictionary CD (Hornby, 2000) and The Oxford Spanish-English dictionary (Carvajal & Horwood, 1996, as provided in Teschner & Whitley, 2004). Only those NN compounds whose prominence pattern was attested in at least one of these sources qualified as potential complex constituents for our triconstituent compounds. NN compounds that were attested in more than one source but with different prominence patterns were not considered as potential complex constituents either. Furthermore, compounds that have been reported in the literature to exhibit variable prominence patterns (e.g., íce cream versus ice créam, Bloomfield, 1933) were also not considered.

The left- and right-prominent two-word compounds were combined with another noun to form an NNN compound. If the third noun was attached to the right, this resulted in a left-branching NNN; if the third noun was attached to the left of the two-word compound, this resulted in a right-branching NNN. The choice of the third noun was constrained by several factors. Firstly, all resulting NNN compounds had to be unambiguously left-branching or unambiguously right-branching in order to exclude the possibility that potential prominence variation would be caused by structurally ambiguous compounds. The problem with structural ambiguous compounds, such as kitchen towel rack or police dog trainer, is that they are argued to be prominent either on constituent N1 or on constituent N2, depending on their interpretation (e.g., Visch, 1999; Warren, 1978). For example, the compound kitchen towel rack may be interpreted as either ‘a rack for kitchen towels’ (left-branching) or ‘a towel rack located in the kitchen’ (right-branching), and the compound police dog trainer may be interpreted as either ‘a trainer of police dogs’ or ‘a dog trainer working for the police’. In the case of a right-branching interpretation, the LCPR predicts the compound to be prominent on constituent N2, whereas a left-branching interpretation will lead to prominence on constituent N1. Given this variable prominence behavior of ambiguous compounds and given the fact that it is difficult to control which interpretation the speaker may choose, compounds were constructed in such a way that they left as little room as possible for different interpretations of their branching structure.

In order to test the role of the embedded compound’s prominence it is necessary to control for potentially intervening prominence relations at the IC-level. The study by Kösling and Plag (2009), for example, found left-branching compounds being prominent on constituent N3 and right-branching compounds prominent on constituent N1, which may be the result of prominence determinants working at the IC-level. For example, it has been claimed that semantic categories trigger the occurrence of right prominence in NN compounds (e.g., Bell, 2008; Fudge, 1984; Ladd, 1984; Liberman & Sproat, 1992; Olsen, 2000, 2001; Sproat, 1994). Examples of these semantic subgroups are, for instance, the group of copulative compounds (e.g., poet-translátor), compounds with temporal (summer níght) or locative modifiers (Boston márathon) or compounds with an ingredient and material as N1 (silk tíe, chocolate púdding) (cf. Kösling & Plag, 2009). Another such factor in compound prominence assignment is analogy (e.g., Arndt-Lappe, 2011; Liberman & Sproat, 1992; Plag, 2006, 2010; Schmerling, 1971; Spencer, 2003). Compounds that share the same left or right constituent tend to have the same prominence pattern. An example often mentioned in this context is that of street versus avenue compounds: while street compounds are generally left prominent (Óxford Street), avenue compounds (Madison Ávenue) are consistently right-prominent (e.g., Plag, 2003).

In order to ensure that the above-mentioned factors would not influence the present analysis, we avoided all semantic relations and categories at the IC-level of the compounds that have been found to trigger rightward prominence in NN compounds in earlier empirical studies (e.g., Kunter, 2011; Plag, Kunter, & Lappe, 2007; Plag et al., 2008). In addition to that, we also avoided lexical items as heads, which are generally mentioned in the literature to trigger rightward prominence in NN compounds (e.g., pie, avenue). By applying these different criteria, the role of the embedded NN compound on the prominence pattern of the entire compound could be tested without any of these factors potentially intervening at the IC-level of the compounds. Other factors, such as frequency or the degree of lexicalization, were not taken into consideration.

For the production experiment, the 40 target compounds were embedded in short declarative sentences that were adapted to be compatible with the semantics of the respective compounds. In order to avoid any potential effects that sentence position may have on the prominence pattern of compounds, all targets were placed in the object position of the sentence, followed by a two-word adverbial. The adverbial was added to keep both the nuclear accent of the sentence and the sentence-final boundary tones out of the target compound, as these tonal events may be expected to be assigned to an utterance due to intonational regularities that are outside of the scope of prominence assignment in compounds. Furthermore, we avoided potential environments in which contrastive meaning was evoked, such as He read about a coffee table designer, not a coffee mug designer. As indicated in the example, contrastive emphasis can quite freely change the canonical prominence pattern of a given compound (cf. Bauer, 1998). The compounds were presented in the sentence in such a way that they always expressed new information, as to avoid any effect of information structure on the placement of tonal events in the target compounds (cf. for instance Hirschberg, 2002). Examples of carrier sentences are given in (4).

2.2 Procedure

For the experiment, seven female and six male undergraduate and graduate students of the University of Toronto were recorded, aged between 18 and 27 years. All were monolingual native speakers of North American English. The majority grew up in the province of Ontario; three speakers originated from other Canadian provinces, and one speaker was originally from Massachusetts, and had been living in Toronto for two years at the time of the recordings. None of the participants, who were paid for their participation, reported any speaking or hearing disorder, and none of them was aware of the purpose of the experiment during the recordings.

Before each recording session, the participants were asked to read the instructions provided to them on a sheet of paper. In addition, the subjects were instructed to read aloud five training sentences that were used to adjust the loudness level of the recorder to the subjects’ individual voices and to give the subjects an opportunity to familiarize themselves with the test situation and the process of reading out loud. The training sentences differed from the target sentences of the experiment in that they contained no compounds.

In addition to the 40 target sentences containing the constructed compounds, the same number of filler sentences was added in order to distract the participants from the actual purpose of the experiment. The filler sentences differed from the target sentences in that they contained no triconstituent compound. Instead, the object slot was filled by a long noun phrase. To control for order effects, the total set of 80 sentences were pseudo-randomized to create three lists, each list with a different order of the 80 sentences. Pseudo-randomization was done on the basis of blocks of 10 sentences and systematically varying the order of these blocks. The sentences were presented to the participants on four separate sheets of papers.

The recordings were taken in a soundproof booth at the University of Toronto using a Marantz PMD660 portable solid state audio recorder at a sampling rate of 44.1 kHz. An external microphone was placed a distance of about 50 cm in front of the participants. Depending on the subjects’ reading speed, each recording took between 15 and 25 minutes. A researcher was present in the soundproof booth during each recording, and asked participants to repeat sentences during which mispronunciations, reading disfluencies or noisy interferences (such as rustling papers or loud chair movements) occurred.

2.3 Acoustic pitch measurements

Using the phonetic software Praat (Boersma & Weenink, 2012), the beginning and end of each clause in which a target compound was embedded were manually annotated, as were the beginning and end of the compound itself, and the two boundaries between the three constituents. The usual criteria for segmentation were applied (see, for instance, Ladefoged, 2003; Turk, Nakai, & Sugahara, 2006) to decide on the location of these boundaries.

For each sound file containing one target clause, the Praat auto-correlation algorithm (Boersma, 1993) was used to extract a pitch object. Pitch floor and pitch ceiling were set according to the sex of the speakers: for female speakers, the pitch range was 100–500 Hz, and for male speakers, it was 75–300 Hz. All other parameters of the pitch tracker were left at their default values.

Any gap in the contour of the resulting pitch object was filled by interpolating linearly between the closest available pitch measurements to the left and the right of the gap. This approach assumes that the pitch contour of an utterance is, in principle, uninterrupted, and that the overall shape of a pitch contour in an utterance is not affected by breaks introduced, for instance, by unvoiced speech segments. This assumption is in line with other representations of pitch contours, such as that in Hermes (2006), who argues that pitch contours are perceived as continuous, and that interruptions in the contour caused by unvoiced speech segments are not noticed by listeners as affecting the overall shape of the pitch contour.

By applying a smoother to the derived pitch contour, the transitions between interpolated and measured segments of the pitch contour receive a more natural shape, and localized misreadings of the pitch tracker algorithms are eradicated. In addition, the effect of microprosodic changes in the pitch contour, such as the lowered F0 that is frequently associated with voiced obstruents (see, for instance, Kingston & Diehl, 1994) is reduced, which is considered advantageous as the present study focuses on the macroscopic scale of the pitch contour that relates to prosodic prominence.

The effects of these modifications are shown in Figure 3. Figure 3(a) shows the unmodified pitch contour of the sentence She read about a gene therapy technology last night as produced by female speaker SP49. There are numerous gaps in the contour that can be associated with unvoiced segments; for instance, the two interruptions in therapy are linked to the voiceless segments /θ/ and /p/. In Figure 3(b), these interruptions are filled by interpolating between the surrounding portions of the pitch contour. Figure 3(c) shows the contour after application of Praat’s ‘Smooth’ function with a bandwidth of 10.0 Hz. The unnatural spikes that were created by the interpolation, for instance that in gene, are flattened. Likewise, small changes in the pitch contour, such as the tiny spikes in the middle portion of technology, which presumably are too small to be noticed by listeners and thus are probably not affecting the perception of the overall intonation contour, are removed.

OPEN IN VIEWER

Figure 3.

(a) Unmodified, (b) interpolated and (c) smoothed pitch contours.

Only the interval of the pitch contour associated with the target compound (indicated by the left-most and right-most dotted line around gene therapy technology in Figure 3) is considered for the present analysis. This target interval is split up into 100 parts of equal length, and the pitch is obtained from the smoothed contour. To account for the non-linear nature of the Hertz scale, all measurements are transformed to the linear semitone (ST) scale relative to 1 Hz. Every pitch measurement is stored together with the number of that part of the interval at which it was taken (ranging from 0 to 99). This number will be used as the time variable in the different smooth terms.

Very rarely, the automatic pitch tracker failed to obtain pitch measurements for parts of the compound, even after interpolation and smoothing. This could occur, for instance, if the end of the respective target sentence was produced with a creaky voice. Non-modal phonation has been shown to occur with very irregular F0 pulses (see, for instance, Blomgren, Chen, Ng, & Gilbert, 1998, for an analysis of the properties of non-modal phonation), which is why automatic pitch trackers frequently fail to detect a periodic signal. The interpolation process described above attempts to compensate for this failure by supplying linear approximations of the pitch contour during problematic parts of the recordings, but for 426 out of the 13 × 40 × 100 measurement points (0.8 percent), no pitch information could be obtained. Thus, the total number of observations was 51,574.

What is notable here is that an equal number of measurements is taken for each compound token, irrespective of any duration differences either between different compound types, or between tokens of the same compound type. For example, the compound type gene therapy technology shown in Figure 3 has, on average across all speakers in the experiment, a longer duration than the compound type day care center (the average durations are 1.413 and 0.784 s, respectively). Speaking rate differences between individuals may lead to duration differences between tokens of the same type: for instance, the duration of the token in Figure 3 is 1.434 s, but observed duration of this type ranged from 1.277 to 1.599 s.

The underlying assumption for using the same number of measurements for compounds of various lengths is that the pitch contour of different compound types is predominantly determined by their branching type. For instance, the LCPR will associate specific tonal events with specific constituents of the NNN compound, but this association is expected to be indifferent to the length of each constituent, or the time a speaker requires to produce the compound. Thus, by taking the same number of pitch measurements from each compound, irrespective of the absolute duration of the item, durational differences are normalized, resulting in a time-averaged series of 100 measurements from each item. ⁴ Yet, in order to relate the subsequently obtained model of the pitch contour to the linguistic structure of the compound types used as stimuli, each pitch measurement was stored together with information on whether the measurement was taken from the left, middle or right constituent of the compound.

2.4 Generalized additive models

GAMs (Wood, 2006, 2011) are an extension of the generalized linear model that provides flexible tools for modeling non-linearities. GAMs take the form

y = X β + f i ( x 1 , x 2 , … ) + … + ε ,

where y is the response variable, X β is a linear predictor and f_i are smooth functions of the covariates x_i. The central concept motivating GAMs is that the dependence of the response variable is decomposed into two submodels. The first submodel, represented above by X β, is a parametric model exactly as in classical linear regression, with β representing a vector of coefficients, one for each of the predictors in the X matrix. The second submodel is non-parametric, and provides one or more smooth functions for one, or several, predictor variables. A smooth function for a single predictor variable makes it possible to model wiggly lines in the plane spanned by the predictor and response variable. A smooth function for two or more predictors defines a wiggly surface in the (hyper)space spanned by the predictor variables and the response variable. By combining the two submodels, it is often possible to obtain a more parsimonious and insightful model compared to a fully general model stating that the response variable is some smooth function of all covariates jointly. In other words, the aim of generalized additive modeling is to provide the simplest possible yet adequate model for the data, allowing non-linearities where necessary, but using standard linear terms wherever possible.

In what follows, we make use of the mgcv package (Wood, 2012) for R (R Development Core Team, 2011). For smooths of single predictors, we used restricted cubic splines, henceforth referred to in tables with the abbreviation s(). For smooths in more than one dimension, we made use of tensor products (abbreviated in tables to te()). Tensor product smooths estimate a wiggly surface (or hypersurface) from two (or more) basis functions. The smoothness of the basis functions (typically cubic splines) determines the amount of wiggliness of the regression surface. The mathematics of tensor products are complex; an informal way of describing a tensor given its marginal basis functions is that the tensor surface is constructed in such a way that in each dimension it is as faithful as possible to the shape of its marginal function in that dimension. Random-effect factors (such as subject and item) can be brought into the model as well, leading to a generalized additive mixed model. The problem of estimating the appropriate smoothing parameters can be solved in various ways: we have used the default, generalized cross-validation. Similar results were obtained with relativized maximum likelihood. Importantly, the optimal degree of smoothness is not determined by the user, but is estimated as part of the model fitting, along with the parametric coefficients and random intercepts for subject and item.

Significance of parametric terms is evaluated by means of the usual t-tests, and/or by means of analysis of deviance tests. Significance of non-parametric terms is evaluated by means of the Bayesian p-values recommended by Wood for smooths. The evaluation of significance for GAMs is more difficult than for standard linear models, hence, it should be kept in mind that probability values are approximate.

Of central interest to the following analysis is how pitch develops over time for the four branching conditions (L/N1, L/N2, R/N2, R/N3). We therefore included in our model a restricted cubic spline for time (t), which we allowed to be optimized for each branching condition separately. A series of model comparisons (see below for further details) clarified that models with less than four branching conditions (obtained, for instance, by collapsing L/N2 and R/N2), provided significantly inferior fits. Branching was also included in the parametric submodel in order to test for (time-independent) differences between the intercepts for the four branching conditions.

Various other predictors were included as controls. Firstly, random intercepts were included for the random-effect predictors speaker and item (i.e., compound). Secondly, the sex of the speaker was included, in order to control for potential differences between female and male speakers. As these differences are expected primarily over time, we included separate temporal smooths for females and males.

Thirdly, position was included as a numeric predictor, as different average pitch measurements may be expected for the left, middle and right constituent (encoded as 0, 1 and 2, respectively) due to the overall declination of pitch. ⁵

Fourthly, the length of a constituent (in phonemes) was included, to control for possible consequences of length for the planning of a constituent’s pitch declination. We also considered an interaction of constituent length by position, to allow for systematic changes in pitch declination as a function of position in the compound. An alternative measure that might be considered is constituent length in milliseconds. In the present study, we have not added such an additional measure to our set of predictors, for two reasons. Firstly, a durational measure would be highly correlated with length in phonemes, and increase multicollinearity. More importantly, as our measure of time in the analysis is a relative one, normalized to the same interval for all compounds, measures of constituent duration would require further normalization with respect to the durations of the other constituents. We leave this issue to future research.

Finally, the lemma frequency of each constituent was also included as a predictor, as frequency of occurrence has been found to have an influence on acoustic durations (Bell et al., 2003). We therefore sought to rule out the possibility that the effects of branching might be confounded with an effect of frequency on pitch. For each constituent in our test items we obtained its lemma frequency from the British National Corpus, and this frequency was added as a covariate to the pitch measurements taken for this constituent. Thus we could explore whether the effect of frequency might vary over time by means of a tensor product smooth.