Aims and motivation

One of the key prerequisites for successful communication is an appropriate mapping between meaning and form. In spoken language, this mapping implies a connection between different layers of prominence, referring to the salience of a syllable or word in comparison with surrounding syllables or words in an utterance. These layers comprise semantic-pragmatic notions such as discourse prominence, predictability, attention-orienting and importance on the one hand, and prosodic prominence, including gradient highlighting via variations in the signal as well as conventionalized linguistic information such as pitch accents or metrical structure, on the other.

The relationship between these two layers is highly complex, since it involves a wide variety of cues on both sides: The meaning side is determined by both lexical properties (e.g., part-of-speech, word frequency) and the discourse context (e.g., focus, information status), while the prosodic form is not only restricted to phonological categories such as pitch accent types or boundary tones but also covers gradient phonetic parameters like differences in duration or intensity. The meaning-form relationship itself may either be parallel, e.g., increased semantic-pragmatic prominence, as in the case of contrastive focus, is expected to induce increased prosodic prominence, or it may be inverse, e.g., higher semantic-pragmatic prominence in terms of increased accessibility is likely to be marked by less prosodic prominence. The communicative function of the latter relation has been formalized in Aylett & Turk’s [1] Smooth Signal Redundancy Hypothesis, which suggests that the inverse relation of linguistic predictability and prosodic prominence yields an even distribution of information across the signal making communication more robust.

Consequently, some discourse contextual cues will boost, and some will inhibit the degree of prosodic prominence of an element. Moreover, some factors and their scales are assumed to have a greater impact on an element’s prosodic realization than others. The first aim of the paper is thus to provide new insights into possible weighting differences between some semantic-pragmatic and syntactic prominence scales (i.e., referential and lexical newness/givenness, grammatical role, and syntactic position) keeping others (especially lexical factors) constant. By looking at the factors in combination, we aim to answer the question whether prominence marking is additive in nature depending on the (type and number of) scales involved.

A second aim is to look at the meaning-form relationship from a decidedly prosodic perspective. That is, we ask which prosodic cues are primarily employed to express specific semantic-pragmatic and syntactic prominence cues and, importantly, what is the effect on the prosodic prominence relations, accounting for the fact that prominence is an intrinsically syntagmatic concept.

State of the art

Previous empirical studies have shown that both production and perception of prosodic prominence are affected by a large number of factors, which have to be taken into account in an investigation of prosodic strength relations (see, e.g., [2–4]). Probably the most comprehensive study on German to date, albeit from a perception side, is [5], testing the influence of 17 linguistic variables on prominence ratings collected in a Rapid Prosody Transcription task (see [6]) on read declarative sentences. The choice of variables was based on previous work on West Germanic languages and comprised discrete prosodic variables (e.g., accent types), continuous-valued prosodic variables (e.g., duration of lexically stressed syllables), and non-prosodic variables (e.g., part-of-speech, presence of a focus particle). Results showed that all 17 variables had an impact on perceived prominence. At the same time, they revealed considerable but nevertheless systematic inter-individual variation, which suggests that prominence perception is both highly complex and multi-factorial and communicatively robust since it allows for a substantial amount of redundancy.

At the group level, an explorative Random Forest analysis [7] showed that the discrete prosodic variables accent type and accent position were particularly predictive of perceived prominence, with high and rising pitch accents as well as accents in nuclear position (i.e., the last accent in the phrase) being judged as prominent by the largest numbers of participants (see also [8] for a comparable pattern in American English). This result was in line with a study on (nuclear) accent types in German [9], which established a ranking of accents in terms of their perceived prominence level increasing from low through falling and high to rising accents. Within these categorizations of the F0 contour in the form of accent types, three continuous dimensions have been shown to affect the perceived degree of prominence of a syllable or word: the actual pitch height (the higher the more prominent, see [10]), the range and slope of the pitch excursion (the larger and steeper the more prominent, see [11,12]) and the alignment of the F0 peak with the metrically strong syllable (the later the more prominent, see [10]). A more recent measure which combines F0 scaling and alignment is the Tonal Center of Gravity (TCoG, see [13,14]), which will be described in more detail in the section on Data, annotation and measurements.

Further gradient prosodic cues that have been found to have an influence on prominence in production and perception are segmental duration (the longer the more prominent, e.g., [15–17]) and intensity (the higher the more prominent, especially in high-frequency components; operationalized in various measures calculating overall intensity or spectral balance, e.g., [18,19]). However, in [5], these cues were secondary to the discrete and continuous parameters characterizing the F0 contour. Similarly, [20] in a production study including only continuous phonetic variables found F0 to play the most important role in distinguishing accented from unaccented words and broad from contrastive focus.

Compared to the discrete prosodic parameters, both continuous prosodic and non-prosodic factors only played a subordinate role for the participants’ prominence ratings in the study by [5]. Within the group of non-prosodic variables, however, the parameters word frequency and part-of-speech turned out to be particularly important for the listeners. The relevance of these two lexical parameters confirmed previous findings for American English that more frequent words were rated as less prominent [21] and for both German and English that content words are generally perceived (and produced) as more prominent than function words. In turn, nouns and proper names are often rated as more prominent than other types of content words such as verbs and adverbs (see [22] for German; [23] for American English).

Other lexical properties that have been claimed to have an impact on prominence perception as well as on the prominence level of the prosodic output are the semantic weight of a content word and, somewhat related, the (in-)animacy of a referential noun. Bolinger [24] showed for American English that the semantic weight of (in particular) nouns and verbs influences their probability of being accented: the ‘heavier’ the more prominent, i.e., the more likely to attract the nuclear accent (see also [25] for German). The examples in (1) and (2) differ in the weight of the final verb. In (1), the verb make is semantically rather empty, or ‘light’, so that the nuclear accent is placed on the ‘heavier’ noun point (marked by capital letters). In (2), in comparison, point only receives a prenuclear accent (indicated by small capitals) and the nuclear accent falls on the final verb emphasize, due to its greater weight, which makes the word less predictable and attracts more attention. Thus, predictability is inversely correlated with both attention and prosodic prominence. Example (3) is added to show the gradience of the relation, in that a semantically empty referring expression like something is even less likely to be accented in this context. Importantly, Bolinger states that “it is not necessary for the verb to be fully predictable from the noun; what counts is RELATIVE semantic weight” [24, p. 635].

1. (1) I have a POINT to make. (2) I have a point to EMphasize. [24]
2. (3) I have something to EMphasize.

Animate entities inherently attract more attention than inanimate ones (e.g., [26]) suggesting a greater effort in marking them. In Russian, for example, an increase in prosodic prominence on animate referents as compared to inanimate ones has been found [27]. At the same time, animate elements are considered more accessible and thus more easily retrievable in discourse than inanimate entities [28,29], which facilitates an occurrence relatively early in an utterance, often as the subject (see, e.g., [30,31] on German). Such a word order effect and the higher degree of accessibility would suggest a realization with lower prosodic prominence to be more likely. Empirical investigations on the specific relation between (in-)animacy and prosodic prominence are clearly needed, however, in the present study, we restrict ourselves to inanimate referents.

Both grammatical and semantic roles correlate with specific positions in a syntactic clause, which in turn have an impact on a word’s prosodic realization. As for grammatical role, subjects often occur first in a sentence (Subject-First preference; see [32] for an overview), while objects occur later, with the last object in a verb phrase usually receiving the nuclear accent rather than the predicate or an external argument (subject) (see, e.g., [33] for English, [25] for German). This makes a (phrase-final) object prosodically the most prominent element in an utterance, at least in broad focus structures. In terms of discourse prominence, subjects are generally considered more accessible and thus more (discourse) prominent than objects (see, e.g., [34,35]). Assuming an inverse relation to prosodic prominence, objects would be expected to be produced as prosodically more prominent than subjects.

Similarly, in terms of the semantic role hierarchy, agents take precedence over non-agents, which lends them higher discourse prominence (see [34]). As for their position in the syntactic clause, agents are often associated with the subject, which in many languages occurs before non-agentive arguments (Agent-First preference, e.g., [34,36]). Prosodically, agents thus often receive a prenuclear accent, while non-agentive roles, e.g., patients, are in many cases marked by a (structurally more prominent) nuclear accent. Moreover, non-agentive subjects more often serve as focus exponents than agentive subjects in German (see [25]), which makes them prosodically prominent as well.

Taken together, it thus seems that the first and last elements in a domain are the most marked ones, with differences at different levels of linguistic description. Both syntactically and semantically, the initial position is often very salient, i.e., prominent on the discourse level (e.g., [37]) but not necessarily prosodically prominent since it is claimed to be associated with prenuclear accents, which are structurally weaker and less important than nuclear accents, or with no accents at all. Nevertheless, Bolinger [38] proposed a tendency for West Germanic utterances to have one major accent near the beginning, the accent of power, and one near the end, the accent of interest, constructing a rhythmical frame for the utterance. This tendency was confirmed in a recent production study on sentence topics in German, which found a consistent placement of (prenuclear) accents on sentence-initial referring expressions—in addition to the (nuclear) accent towards the end of the sentence [39].

The next set of prominence cues can only be defined with respect to the previous discourse context, namely focus and information status. There is a large amount of studies, especially on West Germanic languages, which investigated the interrelation of these two levels of meaning with prosody as their main formal indicator. Many studies indicate, e.g., that new information is often marked by increased prosodic prominence whereas given or old information is prosodically attenuated (see, e.g., [40–43]). It has also been shown, however, that this mapping is too simplistic and that intermediate levels of cognitive activation should be assumed, accounting for accessible information such as hyponyms and hypernyms, bridging relations etc. (see [44]). Furthermore, two independent levels of givenness have been suggested which potentially influence the prosodic realization of expressions, namely a referential and a lexical level (see the RefLex scheme, [45–47]). The following examples illustrate the relevance of such a division:

1. (4) I went to the dentist yesterday. (5) John got a new car, I’d like to STRANgle the butcher. [48] and also BILL bought himself a car.

r-given r-new l-new l-given

The butcher in (4) is referentially given (r-given), i.e., coreferential with the dentist in the previous sentence, which makes deaccentuation of butcher appropriate, despite its lexical newness (l-new). The nuclear accent is shifted to the adjacent verb (strangle). A car in (5) is likely to get deaccented as well but for a different reason: This time, the referent is not the same as a previous one (thus r-new) but it is expressed by the same word (thus l-given). In sum, both examples exhibit deaccentuation of a referential expression due to givenness, which applies either on the referential level as in (4) or on the lexical level as in (5).

Many papers on focus suggest a correlation between pragmatic importance, increasing from broad via narrow to contrastive focus, and prosodic prominence (see, e.g., [49,50] for an overview). Put in a simplified way, the more pragmatically important an element or phrase, or the smaller the number of their contextually salient focus alternatives (following the widely accepted approach of Alternative Semantics by [51]), the higher is the probability of a prominent marking of the focus exponent, often by prosodic cues. Focus thus exhibits a parallel increase of discourse prominence and prosodic prominence, i.e., a ’boosting’ effect. Although this broad correlation is surely not universal, it is commonly regarded as valid for (West) Germanic languages.

Finally, there are paralinguistic factors such as emotional emphasis or speaking style that influence the prosodic realization of utterances. In general, spontaneous speech has been shown to differ in many respects from read speech—reflecting different levels of control—both at a phonetic and phonological level, but not always displaying consistent differences or similarities (see the collection of papers in [52]). As an example for two relatively controlled speaking styles, clear speech has been found to be manifested in larger pitch range and longer segmental and pause durations than conversational speech (e.g., [53,54] for an overview). Very similar results have been found for TED Talk speakers [4] and other speaking styles that generally involve ’lively’ speech, including the increased use of specific accent types (esp. L+H*) ([55] for American English, [56] for German).

To conclude, there is a large number of both prosodic and non-prosodic factors that determine the strength relations between the words carrying multi-dimensional information. Fig 1 summarizes the a) lexical, b) syntactic, c) semantic-pragmatic and d) paralinguistic factors discussed above as scales, which are predicted to decrease (inhibit) or increase (boost) the words’ prosodic prominence. Note that the scales are not always continuous–they may consist of discrete levels or binary categories, such as part-of-speech, which could further be divided, e.g., into nouns, proper names, adjectives, verbs and adverbs. In each scale, an increase in prosodic prominence from bottom to top is expected, which means that the meaning-form relationship may either be parallel or inverse, the latter being predicted by the Smooth Signal Redundancy Hypothesis [1], as mentioned in the Aims and motivation section above. Previous work has often focused on the influence of a single scale on prosodic prominence production or perception, with the study by [5] taking a more comprehensive approach from the perception side. What has not been done so far is to combine and weight the impact of multiple semantic-pragmatic and syntactic prominence cues on prosodic structure and its strength relations in a production experiment. This is the aim of the present study. In a reading task involving short stories, we manipulate several syntactic and pragmatic variables (i.e., grammatical role, sentence position, referential givenness and lexical givenness), while keeping other parameters constant (e.g., word frequency, part-of-speech and animacy).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1.

Potentially relevant prominence scales for each word: a) lexical, b) syntactic, c) semantic-pragmatic and d) paralinguistic properties; an increase in prosodic prominence from bottom to top is expected.

https://doi.org/10.1371/journal.pone.0299746.g001

Research questions and predictions

In the exploratory setup of our production study, we select four semantic-pragmatic and syntactic factors that have been found to influence the level of an expression’s prosodic prominence: Referential newness/givenness, lexical newness/givenness, grammatical role and sentence position. Their impact will be tested on five occurrences of the same target word in a task-oriented setup. Specifically, we consider the following sets of research questions (RQs):

RQ1: Which semantic-pragmatic and syntactic factors boost and which inhibit the prosodic prominence of discourse referents? Which cues have a larger impact than others? Is prominence marking additive in nature?

RQ2: Which prosodic cues are employed to express prominence triggered by semantic-pragmatic and syntactic factors? In particular, what is the effect on the prosodic prominence relations within an utterance?

With regard to RQ1, we expect that both referential and lexical newness, object role, and medial position (equivalent to ‘final argument’) will lead to an increase in the prosodic prominence of a target word (see Table 1 and reading materials in the Materials and methods section for details). Comparing the predicted prominence level of the target words with their prosodic realizations will provide information about the weighting of the factors tested and their combined effect on prosodic prominence. As a baseline, however, each factor is assumed to have the same weight, i.e., each factor is assigned one prominence point. In fact, we expect an additive effect when the cues combine.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Predicted (probabilistic) distribution of prominence-boosting (red) and prominence-inhibiting (blue) factors tested in our study, resulting in a ranking of target words according to the number of prominence points assigned to them.

https://doi.org/10.1371/journal.pone.0299746.t001

To address RQ2, we will consider several categorical and continuous prosodic correlates of prominence. Most importantly, the placement of the nuclear accent will provide insight into the prosodic strength relations of constituents in an utterance at the syntagmatic level. Prominence-boosting factors will increase a word’s probability to be marked by a nuclear accent, while inhibiting factors will decrease this probability.

Reading material

We designed eight short texts in which the same target word occurs in each text in various conditions (see example story in (6)). Keeping the lexical characteristics constant (word frequency, part-of-speech, animacy) allows us to investigate the influence of selected paradigmatic factors on the relative prominence of each target word. Of course, the paradigmatic comparison is to some extent affected by the syntagmatic environment of the word, an aspect which is nevertheless accounted for by the factor sentence position (initial vs. medial). Target words are common nouns with the same stress pattern (penultimate stress) and equivalent syllable structures (three syllables, stressed open syllable ending on a long vowel, CV:) such as Limone (/liˈmoːnə/, ‘lemon’) or Praline (/pʁaˈliːnə/, ‘chocolate candy’). The target words are generally voiced throughout in order to guarantee a continuous pitch contour for our analysis. They were controlled to have comparable frequency classes following the Wortschatz Leipzig corpus (https://corpora.uni-leipzig.de/de/res?corpusId=deu_news_2022).

(6)

Maria hat im Büro eine Kollegin, mit der sie sehr gut befreundet ist. Fast jeden Tag hat sie ein kleines Geschenk für Maria dabei, das sie ihr in der Mittagspause überreicht. Heute hat die Kollegin eine Praline (Target1) mitgebracht. Maria war noch am Arbeiten, als die Kollegin zu ihr an den Schreibtisch kam. Dann hat die Kollegin die Praline (Target2) ausgepackt. Sie war von einer edlen Marke und Maria hat sich sehr gefreut. Die Praline (Target3) war mit Karamell und Nüssen gefüllt. Als Kind war sie mit Süßigkeiten sehr wählerisch, was auch mit einem Erlebnis auf ihrem achten Geburtstag zusammenhängt. Damals hat Maria eine Praline (Target4) gegessen. Sie hat aber nicht gewusst, dass die manchmal mit Alkohol gefüllt sind. Maria hat sich geekelt und danach sehr lange gar keine Schokolade mehr angerührt. Diese Zeiten sind zum Glück vorbei und heute liebt Maria Schokolade aller Art. Die Praline (Target5) wird sie gleich als Vorspeise essen.

Maria has a colleague in the office with whom she is very good friends. Almost every day she brings a small gift for Maria, which she presents to her during the lunch break. Today, the colleague brought a chocolate candy (Target1). Maria was still working when the colleague came to her desk. After that, the colleague unpacked the chocolate candy (Target2). It was from a noble brand and Maria was very happy. The chocolate candy (Target3) was filled with caramel and nuts. As a child, she was very picky with sweets, which is related to an experience on her eighth birthday. At that time Maria ate a chocolate candy (Target4). But she didn’t know that sometimes they are filled with alcohol. Maria was disgusted and didn’t touch any chocolate for a very long time afterwards. Fortunately, these times are over and today Maria loves chocolate of all kinds. The chocolate candy (Target5) she will eat right away as an appetizer.

For each story, we manipulate four syntactic and discourse-related properties of the target word, which have been found to have an effect on an utterance’s prosody. The example story in (6) and Table 1 in section Research questions and predictions illustrate a combination of these factors. The five mentions of the target word (here: Praline) occur as being (i) referentially given or new, (ii) lexically given or new, (iii) a subject or object, and (iv) in an initial or medial sentence position. The target sentences are further broadly controlled for length and complexity, as well as for the word frequency of the final verb in the ‘sentence-medial’ condition (the verbs displayed similar frequency classes, taken from the Wortschatz Leipzig corpus; see above). We expect that referential as well as lexical newness, object function, and medial position (which is in fact the final argument in the sentence) will lead to an increase in the prosodic prominence of a target word. If we assign one ‘prominence point’ (along the lines of [57]) to each of these predicted prominence-boosting factors, we arrive at the accumulated number of points given in the bottom row in Table 1. Comparing the predicted prominence level of the target words with their actual prosodic realizations will provide us with valuable information about the weighting of the factors tested and their contribution to the resulting strength relations.

Participants and experimental procedure

We collected data from 15 speakers in an interactive reading task (between November 4th and December 6th, 2022). Each speaker produced the same eight stories with five target words in different conditions. Speakers grew up with German as their native language originating from five different federal states of Germany. None of them spoke in a non-standard variety. They were mostly students (at the University of Cologne) at the time of recordings and between the ages of 18 and 31 years (mean: 24 years). Four speakers self-identified as male, eleven as female. Participants provided their written informed consent to participate in this study and received a monetary compensation of eight euros. The Ethics Committee of the German Linguistic Society (DGfS, vote #2020-04-200327) stated that it "has reviewed your application on the basis of the descriptions and information you last submitted on March 26, 2020. Information letters and declarations of consent for conducting behavioral and non-invasive experiments with healthy adult subjects aged 18 to 65 years were submitted and you received a positive vote, as the commission does not currently have any ethical concerns about conducting such a study using procedures established in psycholinguistics. The survey methods do not represent a particular psychological burden for the target age group" (translation from German).

Recordings were conducted in a sound-attenuated booth at the phonetics laboratory in Cologne. Speakers wore a head-mounted AKG C544L condenser microphone. The speech signal was recorded on a computer via Adobe Audition at a sampling rate of 44.1 kHz and a bit depth of 16 bit.

Stories were presented in pseudo-randomized order. Each participant first read each story silently to become familiar with it and then read it out loud as part of a listener-directed reading task. A confederate was present during the recording, whom the speakers believed to be another participant. The speaker was instructed to read the stories in such a way that the confederate would be able to memorize the stories and answer three comprehension questions after each story. The confederate answered the questions in written form so that the speaker did not learn about the correctness of the answers at any point during the experiment. Answers were discarded after the experiment. We also presented the speaker with a picture (see Fig 2) for each story to make the experimental design more appealing and reinforce the playful nature of the task. The whole procedure took approximately 30 minutes.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Picture accompanying the example story in (6) during the reading task.

https://doi.org/10.1371/journal.pone.0299746.g002

Data, annotation and measurements

In total, we recorded 600 utterances (15 speakers * 8 stories * 5 target words). Due to hesitations or repairs, we had to exclude 10 utterances. Thus, 590 utterances entered the analysis. The produced target words were analyzed phonologically (accent position and type, following GToBI; [58], perceived prominence and phrase boundary placement according to DIMA; [59]) and acoustically (e.g., measured in terms of duration, intensity, and Tonal Center of Gravity, the latter incorporating the shape of the contour and the alignment of turning points; see [13,14,60]).

All utterances were annotated by two independent transcribers following the DIMA guidelines [59]. Specifically, phrase boundaries were annotated as strong (%) or weak (-) boundaries (agreement rate: 94%). At the tonal level, accentual (starred) and non-accentual tones (turning points, non-starred) were annotated (agreement rate: 71%). Based on these annotations, we determined the accent status (nuclear, prenuclear, deaccented) of each target word. In addition, pitch accents were classified according to the GToBI system ([58]; agreement rate: 64%). For the statistical analysis, we treated pitch accent type as an ordinal scale. Following the findings in [9], we assumed the following ranking of accent types from most to least prominent: L+H* > L*+H > H* > H+!H* > H+L* >! H* > L* > deaccentuation. Finally, the perceived prominence of words was annotated on a scale of 1 to 3, where 1 indicates weak, often rhythm-dependent prominence, 2 indicates strong prominence that is primarily based on tonal movement and 3 indicates extra-strong prominence, mostly in emphatic realizations (agreement rate: 86%). The absence of any perceived prominence, although not specifically annotated in our dataset, is indicated by a 0 (in DIMA). In cases of disagreement between the transcribers regarding any of the annotated tiers, a consensus was reached with a third expert.

Furthermore, we considered several continuous phonetic parameters. Firstly, we measured alignment and scaling of the Tonal Center of Gravity (TCoG) following Barnes and colleagues [13,14,60]. TCoG measures represent an alternative to traditional measures using turning points and are able to capture both alignment in time and F0 scaling. TCoG alignment constitutes a point in time relative to the beginning of the target word weighted by the F0 contour. Values of TCoG alignment are higher the later the rise and the more ’scooped’ the F0 contour are. Values are lower the earlier the rise and the more ’domed’ the F0 shape are. TCoG scaling is measured as the weighted average F0 over the target word. Crucially, more sonorous regions contribute more to the value of TCoG scaling [60]. In order to get rid of participant differences in these values, we converted them to a semitone scale relative to a speaker-specific baseline. This baseline was determined by extracting the F0 values of the last low tone (i.e., non-accentual L or accentual L*) annotated in each utterance and taking their median for each speaker.

Furthermore, we measured the duration in milliseconds and the mean intensity in decibels of the stressed syllable of the target word. For the statistical analysis, we excluded phrase-final instances of the target words from the duration measures to control for potential effects of final lengthening.

Statistics

We ran Bayesian mixed-effects regression models using the brms (version 2.20.1, [61]) package as an interface to the Stan modeling language (version 2.26.2, [62]) in R (version 4.3.1, [63]). We built one model for each prosodic parameter, with the parameter in question as the dependent variable. For categorical dependent variables, we used binary logistic or cumulative models, for continuous variables we ran linear regression models.

Each model included four predictor variables: Position, grammatical role, referential givenness and lexical givenness. The reference levels for these predictors were initial (vs. medial), subject (vs. object) and given (vs. new) for both levels of givenness. To account for inter-speaker and inter-item variability, we included random intercepts for item and speaker and random slopes for the by-item and by-speaker effects of all predictors. Four sampling chains were run for 4,000 iterations each with a warm-up period of 2,000 iterations, yielding a total of 8,000 posterior samples per model. For the regression coefficients, we specified weakly informative, normally distributed priors with a mean of zero and a standard deviation of ten for the linear regression models and a mean of zero and a standard deviation of one for the cumulative models. We used default priors supplied by brms for the remaining model parameters.

We compared the posterior estimates of each variable to the reference levels. Positive values in these comparisons provide evidence for the predictions specified in the Research questions and predictions section. For each model, we report the regression coefficient β and 90% credible intervals (CIs) under the posterior distributions as well as the posterior probability that β is larger than zero (Pr(β > 0)). When the CIs do not include zero and the posterior probability Pr(β>0) is larger than 0.95, we take this as compelling evidence in favor of our predictions.

For data processing and plotting, we used the tidyverse (version 2.0.0, [64]) and ggdist (version 3.3.0, [65]) packages. For each parameter, we plot the distribution of the measured data by target word number and the posterior estimates from the regression models for each predictor variable. In the model plots, the colored shapes represent the distributions of the posterior estimates, the dot is the mean, the thin horizontal lines indicate 90% CIs and the thick lines 66% CIs. The vertical dashed line marks the position where the estimate β equals zero, i.e., where there is no difference between the levels that are being compared. When the thin black line does not overlap with the vertical zero line, the criterion that zero is not included in the 90% CI is fulfilled. Data and code for analysis are available on the OSF repository under the following link: https://osf.io/s9tmp/.

Discrete prosodic parameters

The first and central parameter we investigate is accent status. Target1 is usually produced with a nuclear accent (in 96% of the cases), as is Target4 (75%, see Fig 3, left). Target2 and Target3 are predominantly produced with prenuclear accents, in 73% and 83% of cases, respectively. Target5 is almost equally likely to receive a nuclear or prenuclear accent (42% and 57%, respectively). Thus, the probabilistic distribution of the target words’ accent status generally confirms our predicted prominence ranking (see Table 1) with one exception: Target2 and Target5 switched their ranks, since Target5 is produced with more nuclear accents than Target2, which runs counter to our expectations.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Results for accent status.

Left: Distribution of accent status across the five target word positions in percent. Right: Posterior estimates of referential givenness, lexical givenness, grammatical role and position for the effects of accent status as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g003

Since deaccentuation occurs in only 12 target words, we treat accent status as a binary variable with the levels nuclear and prenuclear/deaccented and run a binary logistic regression model. A further justification for this decision stems from the fact that in annotation, prenuclear and deaccented targets are often difficult to distinguish, while it is usually a straightforward task to determine the location of the nuclear accent.

The Bayesian mixed-effects regression model confirms that there are compelling effects of referential givenness (β = 3.81, CI = [2.1; 5.69], Pr(β>0) = 1), lexical givenness (β = 5.13, CI = [2.11; 9.46], Pr(β>0) = 1) and grammatical role (β = 1.76, CI = [0.82; 2.74], Pr(β>0) = 1) in the direction we expected (see Fig 3, right). That is, referentially and lexically new objects are more likely to be realized with a nuclear accent than referentially and lexically given subjects. However, there is some evidence that the initial position is more prominent than the medial position, which we did not anticipate (β = -1.49, CI = [-3.22; 0.23], Pr(β>0) = 0.07).

As evident from Fig 4, we observe some inter-item variability, especially in Target2 and Target4. While Target2 is often the least likely to receive a nuclear accent (between 0% and 25% nuclear accents), i.e., the least prominent target word in a story, it is more likely to receive a nuclear accent in the stories containing Banane (‘banana’, ca. 50% nuclear accents) and Gardine (‘curtain’, ca. 75% nuclear accents). Target4 is usually the second most prominent target word following Target1, which is expected since the two conditions differ only in one factor, namely lexical givenness. However, in the stories related to Banane (‘banana’) and Garage (‘garage’) Target4 is equally prominent as Target1 suggesting that lexical givenness does not inhibit the prominence of Target4 in these stories. In the stories containing Gardine (‘curtain’) and Kombüse (‘galley’), Target4 is relatively less prominent than Target5 and less prominent than Target2 in the story containing Gardine. This could indicate that referential newness in Target4 does not boost prominence as much as in the other stories or that lexical and referential givenness in Target2 in the Gardine story does not inhibit prominence as much. There are also some consistencies across stories in that Target1 is always the most prominent target word, and in the initial position, Target5 is usually more prominent than Target3. Both of these results are in line with our expectations for the specific cue combinations of these conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Distribution of accent status in percent across the five target word positions facetted by item/story.

https://doi.org/10.1371/journal.pone.0299746.g004

In terms of pitch accent type, Target3 and Target5 (both in initial position) are realized very similarly, mostly with the very prominent L+H* (ca. 70%) and L*+H (ca. 25%) accents (see Fig 5, left). In the medial position, there is more variability regarding the produced accent types, especially in Target2 and Target4. In Target1, there is a larger proportion of L+H* accents (33%) as compared to Target2 (13%) and Target4 (24%). Target1 and Target4 are also often realized with falling H+!H* accents (both 29%). Target2 often receives L*+H accents (31%).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Results for accent type.

Left: Distribution of accent types across the five target word positions in percent. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on accent type as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g005

According to our regression model, there is some evidence that lexically new target words are produced with more prominent accent types than given targets, however, this evidence is not compelling according to the criteria specified in the Statistics section (β = 0.71, CI = [-0.06; 1.44], Pr(β>0) = 0.94, see also Fig 5, right). Referential givenness and grammatical role do not seem to have an influence on accent type choice in the expected direction (referential givenness: β = -0.41, CI = [-1.22; 0.39], Pr(β>0) = 0.2), grammatical role: β = -0.32, CI = [-0.96; 0.29], Pr(β>0) = 0.19). However, position is clearly a decisive factor for accent type choice, in that the initial position attracts especially strong accents (β = -2.5, CI = [-3.5; -1.4], Pr(β>0) = 0.00).

As evident from Fig 6 (left), most target words are produced with a full pitch accent corresponding to the DIMA prominence label 2. However, there are some weaker productions of target words corresponding to DIMA levels 0 and 1, especially in Target2 and Target4, both occurring in medial position.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Results for DIMA prominence scores.

Left: Distribution of prominence scores across the five target word positions in percent. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on prominence scores as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g006

The regression model suggests that there is compelling evidence to assume that lexically new target words are produced more prominently than lexically given ones (β = 1.52, CI = [0.06; 2.84], Pr(β>0) = 0.95, see Fig 6, right) and some evidence that referentially new target words are produced more prominently than referentially given ones (β = 0.55, CI = [-0.48; 1.53], Pr(β>0) = 0.83). However, contrary to our expectations, there is some indication that grammatical subjects are perceptually more prominent than objects (β = -0.75, CI = [-1.99; 0.49], Pr(β>0) = 0.16). Again, position has a large unpredicted influence in that initial target words are perceived as more prominent than medial target words (β = -1.66, CI = [-2.76; -0.53], Pr(β>0) = 0.01).

Phrase boundaries occur only after Target3 and Target5, i.e., after the target words in the initial sentence position (see Fig 7, left). This is to be expected, as the medial target words occur in the last slot before the utterance-final verb, where a phrase boundary would be highly marked. The lack of boundaries after medial target words can thus not be interpreted as a lack of prominence. Therefore, phrase boundaries as prominence markers can only be investigated in the initial position and the corresponding regression model is run on a subset including only Target3 and Target5. As a consequence, we can include only one predictor, namely grammatical role, since Target3 and Target5 do not differ in any other predictors. The model provides evidence that objects are more likely to be followed by more and stronger phrase boundaries than subjects (β = 1.23, CI = [0.37; 2.02], Pr(β>0) = 0.99, see Fig 7, right).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Results for phrase boundary placement.

Left: Distribution of phrase boundaries across the five target word positions in percent. Right: Posterior estimates for the effects of grammatical role on prominence scores as predicted by the model run on a subset of the data including only Target3 and Target5, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g007

Gradient prosodic parameters

The Tonal Center of Gravity is located farthest from the beginning of the target word in Target3 (mean = 298 ms, sd = 60 ms) and Target5 (mean = 320 ms, sd = 59 ms), i.e., the initial target words (see Fig 8, left). This is due to the large number of rising pitch accents in this position. Target5 exhibits on average slightly higher values, i.e., later peak alignment, than Target3. In the medial position, Target2 has the latest peak alignment (mean = 221 ms, sd = 52 ms), followed by Target1 (mean = 210 ms, sd = 50 ms) and Target4 (mean = 191 ms, sd = 47 ms), although these differences appear to be relatively small.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Results for TCoG alignment.

Left: Distribution of TCoG alignment (in ms) across the five target word positions. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on TCoG alignment as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g008

The regression model confirms that position is the most important factor in determining alignment in that TCoG in initial target words is later aligned than in medial words (β = -11.0, CI = [-28.06; 6.42], Pr(β>0) = 0.15, see Fig 8, right). Unexpectedly, referentially given target words also appear to exhibit later TCoG alignment than new target words (β = -9.31, CI = [-24.31; 5.96], Pr(β>0) = 0.16). However, lexical givenness behaves as expected, with lexically new target words exhibiting a later TCoG than given words, although this evidence is not compelling according to our criteria (β = 7.08, CI = [-6.91; 20.18], Pr(β>0) = 0.8). Grammatical role barely affects TCoG alignment (β = 3.63, CI = [-11.22; 17.57], Pr(β>0) = 0.67).

Similarly to TCoG alignment, TCoG scaling also reflects the large number of rising accents in the initial position with the highest mean values occurring in Target5 (mean = 6.0 st, sd = 2.5 st) and Target3 (mean = 5.8 st, sd = 2.3 st, see Fig 9, left). In the medial position, Target1 has the highest mean value (4.0 st, sd = 2.9 st), followed by Target4 (mean = 3.3 st, sd = 2.4 st) and Target2 (mean = 3.1 st, sd = 1.9 st), although these differences are again very small.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Results for TCoG scaling.

Left: Distribution of TCoG scaling (in st) across the five target word positions. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on TCoG scaling as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g009

The regression model provides support for position as the most important factor in determining TCoG scaling with the initial target words reliably exhibiting higher values (β = -2.83, CI = [-3.69; -1.97], Pr(β>0) = 0.00, see Fig 9, right). In addition, there is compelling evidence that lexically new target words are produced with higher TCoG scaling (β = 0.66, CI = [0.05; 1.27], Pr(β>0) = 0.96). Although referentially given words and words in object role also appear to be higher in scaling, evidence to support these tendencies is much weaker (referential givenness: β = 0.15, CI = [-0.36; 0.64], Pr(β>0) = 0.7, grammatical role: β = 0.12, CI = [-0.25; 0.49], Pr(β>0) = 0.71).

In terms of syllable duration, the accented syllable in Target1 is on average longest (mean = 192 ms, sd = 35 ms), followed by Target5 (mean = 190 ms, sd = 35 ms) and Target3 (mean = 183 ms, sd = 34 ms), and Target4 (mean = 173 ms, sd = 33 ms) and Target2 (mean = 171 ms, sd = 35 ms) are the shortest (see Fig 10, left). According to the regression model, position is again an important factor in determining syllable duration with the initial target words reliably exhibiting longer duration (β = -16.12, CI = [-21.6; -10.51], Pr(β>0) = 0.00, Fig 10, right). Recall that we excluded phrase-final target words from the duration analysis due to potential final lengthening effects. That is, the relatively long durations of the initial target words (Target3 and Target5) cannot be explained by the considerable number of phrase breaks following them (see Fig 7). In addition, there is compelling evidence that lexically new target words are produced with longer syllable durations (β = 14.89, CI = [5.52; 23.02], Pr(β>0) = 0.99) and also strong evidence that objects are longer than subjects (β = 5.09, CI = [-0.87; 11.04], Pr(β>0) = 0.92). Although referentially new target words also appear to be slightly longer, evidence to support this tendency is very weak (β = 1.44, CI = [-6.08; 8.69], Pr(β>0) = 0.63).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 10. Results for syllable duration.

Left: Distribution of syllable duration (in ms) across the five target word positions. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on syllable duration as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g010

Mean intensity is highest in Target3 (mean = 72.1 dB, sd = 3.9 dB), followed by Target5 (mean = 72.0 dB, sd = 3.9), Target1 (mean = 70.7 dB, sd = 4.4 dB), Target4 (mean = 70.11 dB, sd = 4.3 dB) and Target2 (mean = 69.1 dB, sd = 4.4 dB), but all differences are relatively small (see Fig 11, left). The regression model again supports a strong effect of position: initial target words have higher intensity than medial words (β = -2.86, CI = [-3.73; -1.99], Pr(β>0) = 0.00, see Fig 11, right). There is also strong evidence that lexically and referentially new target words are produced with higher intensity than given target words, which is compelling in the case of referential givenness (lexical givenness: β = 0.6, CI = [-0.09; 1.32], Pr(β>0) = 0.93, referential givenness: β = 0.95, CI = [0.27; 1.63], Pr(β>0) = 0.99). Grammatical role does not have a strong effect on intensity (β = -0.07, CI = [-0.74; 0.61], Pr(β>0) = 0.43).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 11. Results for intensity.

Left: Distribution of intensity (in dB) across the five target word positions. Right: Posterior estimates for the effects of referential givenness, lexical givenness, grammatical role and position on intensity as predicted by the model, means, 66% (thick horizontal lines) and 90% credible intervals (thin lines).

https://doi.org/10.1371/journal.pone.0299746.g011