1.1 Prosody-syntax interactions in focus marking and clefts

The literature on information structure frequently describes interactions between prosody and morpho-syntax in its marking. In line with the assumption that the language system is efficient (Reinhart, 2006), researchers often derive prosodic focus marking from morpho-syntactic focus marking or vice versa (e.g., Büring, 2010; Büring & Gutiérrez-Bravo, 2002; Féry et al., 2007; Hamlaoui, 2007; Szendrői, 2017; Vallduví & Vilkuna, 1998; Vander Klok et al., 2018). For example, Kratzer and Selkirk (2020) describe prosodic phrasing as influenced by a syntactic focus feature. Conversely, Féry (2013) motivates the prosodic, syntactic and morphological focus marking strategies of various languages as a universal preference for focus to be aligned with the edge of a prosodic phrase. Samek-Lodovici (2005) suggests that languages differ in whether they preferentially use prosody or syntax for marking information structure, depending on whether syntactic or prosodic constraints are ranked higher.

While these accounts concentrate on modelling the availability of different means of focus marking–whether or not a language is able to use changes in the position of accents or constituent order at all–cross-linguistic differences in the frequency with which means of focus marking are used have mostly been discussed with respect to cleft constructions. Clefts, as illustrated in (1) below, can mark the clefted constituent as the focus, as indicated by the square brackets, and the rest of the sentence as background (although they do not have to do so, this is often considered their prototypical use, see Akmajian, 1970; Atlas & Levinson, 1981; Cassarà et al., 2022; Chomsky, 1971; Hedberg, 2000; Jackendoff, 1972; Karssenberg et al., 2019; Kiss, 1999; Lambrecht, 2001; Rochemont, 1986). As already suggested by Jespersen (1937, p. 85), it is often observed that clefts are more frequent in languages with fixed constituent order, such as English and French, than in languages which can use their flexible constituent order to mark focus, such as German and Italian (de Cesare, 2014; Di Tullio, 2006; Gundel, 2008; Lambrecht, 2001; Skopeteas & Fanselow, 2010). Moreover, languages with less flexible prosody, like French and Spanish, use clefts more often than languages like English, whose flexible prosody includes a large inventory of pitch-accents that can be placed variably within a sentence to express pragmatic contrasts (Gundel, 2006, 2008; Lambrecht, 2001; Sánchez-Alvarado, 2020; Skopeteas & Fanselow, 2010; van Valin & LaPolla, 1997; Wehr, 2005; but see Dufter, 2009 and Gundel, 2006, 2008, who argue that this cannot completely explain cross-linguistic differences in cleft frequency). Similarly, clefts are more frequent in written language, where prosodic focus marking is unavailable, than in spoken language (Collins, 1991; Declerck, 1988, p. 226; Tönnis et al., 2016).¹

A question that has not been considered with respect to clefts and prosody-syntax interactions is whether using a cleft together with prosodic focus marking constitutes a redundancy that speakers try to avoid or attenuate. Since clefting can already mark the information structure by itself (as it does in written language), additional prosodic marking may be redundant.² If speakers avoid redundancy, they could be expected to either not use prosodic focus marking in clefts or to use it less extensively than in morpho-syntactically unmarked utterances, where prosody is the only means of marking information structure. Whether this kind of prosody-syntax trade-off can indeed be observed has not been investigated previously (with the exception of Arnhold, 2021).

Instead, several authors have used observations about the prosodic realization of clefts as arguments for particular syntactic, information structural or discourse analyses, frequently based on corpora of spontaneous speech (Cassarà et al., 2022; Collins, 2006; Delin, 1995; Frascarelli & Ramaglia, 2013; Geluykens, 1984; Hedberg, 1990; Huber, 2006; Pinelli et al., 2020; Prince, 1978; Van Praet & O’Grady, 2018). Experimental studies on prosody-syntax interactions in clefts have mostly focused on cases where prosody and morpho-syntax do not align to mark the clefted constituent as focused: A series of experiments by Calhoun, Yan and colleagues tested cases where prosody and morpho-syntactic focus marking are in conflict and compared them to equivalents where they do align or where only prosodic marking is present, finding that languages differ in whether prosody or morpho-syntax takes precedence (Calhoun et al., 2021; Yan et al., 2020; Yan & Calhoun, 2019, 2020). Greif and Skopeteas (2021) tested if the ability of four languages to prosodically mark object focus in subject cleft sentences predicts the acceptability of subject clefts with object focus in a rating task with written stimuli. They found that such clefts with focus following the clefted constituent (e.g., It is John that bought the [bicycle]_F) are more acceptable in English and German, which have flexible nuclear-accent placement, than in French and Mandarin. Finally, Pinelli et al. (2020) showed very similar prosodic realizations for two syntactically different cleft types in Italian. More in line with the present study’s concerns is the approach taken by Kember et al. (2019), who tested participants’ memory of words that were clefted, made prominent prosodically, both or neither. For Korean listeners, the authors found a stronger memory advantage of clefting than of prosodic marking, while for English, both effects were equivalent and additive if combined, i.e., there were no trade-offs between prosody and syntax (also see the rating study in Arnhold, 2021). In contrast, Blything et al. (2021) found no additional effect of clefting on pronoun resolution in English once the effect of prosodic focus marking was accounted for.

With the exception of the most recent studies, research considering the prosody of clefts has entirely concentrated on Germanic and Romance languages.

1.2 Mandarin clefts and prosodic focus marking

Mandarin clefts, see (2), below, are also called shi…de constructions, since they contain the copula 是 shi4 and the toneless syllable 的 de0, which are not present in syntactically unmarked equivalents; compare (2) to (3). These constructions occur in different variants, with shi4 and de0 both being able to appear by themselves, whose syntactic and semantic analysis has been debated (Cheng, 2008; Hole, 2011; Hole & Zimmermann, 2013; Lee, 2005; Liu & Shi, 2022; Paul & Whitman, 2008; Simpson & Wu, 2002; Xie, 2012). As indicated in (2), in transitive cleft sentences de0 can appear either between the verb and the object or sentence-finally, conditioned by dialect. Speakers of southern Mandarin dialects exclusively place de0 sentence-finally, whereas in northern dialects, both placements are possible, but only verb-adjacent de0 is compatible with a narrow focus interpretation (Paul & Whitman, 2008). Moreover, Simpson and Wu argue that de0 is undergoing a re-analysis as a past tense morpheme in northern dialects (also see Lee, 2005). For these reasons, clefts without de0, also called bare shi constructions, were used in the studies presented here (as also done in Greif & Skopeteas, 2021; Liu & Yang, 2016).

The prosodic realization of Mandarin clefts is not very well researched. The literature discussing semantic and syntactic analysis contains some remarks about possible locations of prominence, see especially Paul and Whitman (2008). In particular, Paul and Whitman state that in bare shi constructions, the clefted subject is only focused when it is made prominent prosodically; otherwise, “the truth of the entire sentence is strongly asserted, with a meaning comparable to ‘it is (really) that S’ or ‘it is because S’” (p. 424). The latter meaning seems to correspond to verum focus, i.e., focus on the truth of the proposition (Höhle, 1992; Lohnstein, 2016). Liu and Shi (2022) also show that Mandarin clefts can convey verum focus, but state that in these cases, the copula shi4 “receives a focal pitch accent” (p. 4). Both descriptions are based on speaker intuition and provide no further discussion of the acoustic correlates of prominence in either subject-focus or verum focus-clefts. Note, however, that Liu and Shi’s description parallels verum focus prosody in many other languages, e.g., English: My aunt IS steaming mushrooms (with capitals marking the location of nuclear accent; also see e.g., Han & Romero, 2014; Féry & Arnhold, 2019; and Gutzmann et al., 2020, on prosody of verum focus in various languages). Crucially, verum focus notably differs from broad focus, to which subject focus is compared in the present study, also semantically: Broad focus is focus on the entire proposition, whereas verum focus is focus specifically on its truth value. As the present experiments systematically induced either subject focus or broad focus, but not verum focus, the potential verum focus reading of Mandarin clefts should not be relevant here, but we will return to this issue in discussing the results of the perception study (Section 3.3).

There are no corpus studies considering prosody of Mandarin clefts and only one production experiment published so far: In Greif and Skopeteas (2021), 16 speakers produced four transitive sentence items in four conditions each: unmarked syntax with subject focus, subject cleft with subject focus, unmarked syntax with object focus, subject cleft with object focus (256 utterances total). Statistical analysis evaluated f0 measurements, looking for interactions between the factors focus (subject vs. object focus), construction (subject cleft vs. unmarked) and time (multiple measurements per syllable). It found an interaction between focus and time, indicating more rapid f0 rises for focused constituents than for pre- or post-focal ones (most target syllables carried the rising lexical tone 2). It also found an interaction between construction and time. The authors explain this as a delay in the implementation of tonal targets on subjects when preceded by shi4 rather than suggesting a difference in prosodic focus marking between clefts and unmarked syntax.

Other acoustic measures were not analyzed, but the f0 results are in line with previous research on (Mainland) Mandarin prosodic focus marking, finding expanded f0 ranges and raised f0 on focused constituents, as well as post-focal compression (Y. Chen & Gussenhoven, 2008; Ouyang & Kaiser, 2015; B. Wang et al., 2017; B. Wang & Xu, 2011; T. Wang et al., 2020; Xu, 1999). Previous research also found focused constituents to have higher intensity and longer durations compared to pre- and especially post-focal constituents (Y. Chen & Gussenhoven, 2008; Ouyang & Kaiser, 2015; B. Wang et al., 2017; T. Wang et al., 2020; Xu, 1999), but this has not been confirmed with a production study for clefts yet. In their perception studies, Yan and colleagues used stimuli produced by a single speaker, which included higher mean f0, larger f0 range, longer duration and higher intensity for focused constituents than for pre- and post-focal ones (Yan et al., 2020; Yan & Calhoun, 2019, 2020).³ Like Greif and Skopeteas (2021), they did not include a broad focus baseline. Finally, there is evidence that voice quality plays a role in Mandarin focus marking. Chen and Gussenhoven report that some of their participants produced more creak for the low-dipping tone 3 in focus than pre-focally (creak is associated with low f0 and, thereby, with tone 3 more than with the other lexical tones, e.g., Huang, 2020; Kuang, 2017). Similarly, Cao and Zhang (2008) found that creaky voice contributes to the perception of tone 3 words as focused. Zheng (2006) showed that focus leads to more realizations with non-modal voice quality, but only for tones 2 and 3, based on visual inspection of waveform and spectrogram. Likewise using manual classification, Huang et al. (2018) found an increase in creaky voice only for tone 4. However, this finding did not replicate with the acoustic measures they analyzed (harmonic-to-noise ratio, cepstral peak prominence and two spectral tilt measures). Wang et al. likewise did not find an effect of focus on the two phonation measures they investigated—cepstral peak prominence and H1–H2—for any of the four lexical tones. None of these studies on voice quality included cleft sentences or other syntactic focus marking.

Regarding the role of prosody in the perception of Mandarin clefts, studies by Yan and colleagues showed that when prosody conflicts with syntax, as in a subject cleft with prosodic focus marking on the object, prosody wins: Prosodic focus marking, not clefting, determines appropriateness ratings of answers to questions inducing either subject or object focus (Yan et al., 2020). Similarly, only prosodic focus marking is able to prime alternatives (Yan & Calhoun, 2019) and facilitate the rejection of false alternatives (Yan & Calhoun, 2020; also see S. H. Chen et al., 2012). Interestingly, clefting the focused constituent never improved appropriateness ratings compared to the unmarked syntax equivalent when the focused constituent was marked prosodically. However, clefting a background constituent lowered ratings (Yan et al., 2020). This is only partially in line with the findings of Greif and Skopeteas (2021). Their participants also rated subject clefts lower than unmarked syntax in object focus, i.e., where the clefted constituent was part of the background. However, in subject focus, subject clefts received higher ratings than unmarked equivalents. This difference between the studies is likely due to Greif and Skopeteas using written stimuli. As they point out, subjects in unmarked SVO sentences are typically interpreted as topics in Mandarin, in line with a cross-linguistic tendency which explains why non-minimal marking, such as clefts, is used more often for subject than for non-subject focus. It is therefore likely that their participants applied the default topic-comment interpretation to the unmarked sentences, which conflicts with subject focus contexts, whereas prosodic focus marking on the subject would have prevented this default interpretation in the studies by Yan and colleagues. Again, none of these studies used a broad focus baseline.

1.3 Hypotheses

The present study will investigate cleft sentences and syntactically unmarked equivalents in Mandarin to test for syntax-prosody trade-offs in focus marking. In line with the common assumption that language is efficient and avoids redundancy, such trade-offs are hypothesized to occur in both production and perception. With respect to production, the hypothesis is that prosodic focus marking will be less pervasive in clefts than in syntactically unmarked equivalents (vs. simultaneous and additive use of prosodic and syntactic focus marking that should occur in the absence of trade-offs). In other words, the effect of focus on acoustic measures associated with prosodic focus marking is hypothesized to either be present only with unmarked syntax (strong version) or be smaller in clefts than with unmarked syntax (weak version), see (4), below. The strong version of the hypothesis could be predicted based on generative accounts deriving prosodic focus marking from syntactic focus marking, or vice versa–based on the assumption that grammar avoids redundancy, and that deriving one from the other is more efficient than computing prosodic and syntactic structures with focus marking independently and then having to relate them to each other (as spelled out explicitly e.g., by Reinhart, 2006). A functional approach could derive both the strong and the weak version of the hypothesis, paralleling van Heuven’s (2017a, b) discussion of phonological (categorical) and phonetic (gradual) differences between syntactic sentence types.

Both versions of the hypothesis will be tested in a production experiment measuring f0 range, f0 maximum, f0 minimum, duration, intensity and the use of non-modal voice quality (operationalized as a binary variable based on manual annotation following Zheng, 2006, and Huang et al., 2018, the only previous production studies of Mandarin finding an effect of focus on voice quality). Based on the existing research summarized in Section 1.2, prosodic focus marking can maximally be expected to manifest as subject focus showing larger f0 range, higher f0 maximum and minimum, longer duration, higher intensity and more non-modal voice quality on the narrow focus subject (focus expansion) and conversely a smaller f0 range, lower f0 maximum and minimum, shorter duration, lower intensity and less non-modal voice quality on the verb and object (post-focal reduction of prominence) than in broad focus. The hypothesis in (4) can reasonably be considered confirmed if at least half of the evaluated measures, i.e., three out of six, show the expected prosodic focus marking in terms of focus expansion, post-focal reduction or both. Note also that the hypothesis is more tentative for voice quality, since most previous studies have shown effects of focus on voice quality only for words with tones 3 and 2 (except Huang et al., see Section 1.2), whereas the present study included only tones 1 and 4 (see Section 2.1.2).

A second experiment will test the corresponding hypothesis of trade-offs in perception, see (5), below, by asking participants to rate the fit between a question inducing subject focus or broad focus on the one hand and, on the other hand, an answer with subject focus marked via prosody, clefting, both or neither. The hypothesis assumes that participants will rate the answer to a question as more appropriate when they judge it to have suitable focus marking, whereas they will give lower ratings in the absence of the focus marking they expected. Similarly, an absence of suitable focus marking could lead to processing delays, resulting in longer reaction times. The strong version of the trade-off hypothesis predicts that when prosodic and syntactic focus marking are combined, either one of them is redundant and will therefore have no effect. Thus, either prosodic or syntactic focus marking is enough to make an answer fit the context, and adding a second type of concurrent marking will not make a difference. The weak version of the hypothesis predicts that the redundant effect is still significant, but also significantly smaller than when information structure is marked only by either prosody or syntax. That is, using two concurrent types of focus marking has a stronger effect than using only one type of marking, but this combined effect is not as strong as simply adding together the effect of both occurring separately.

The hypothesis of trade-offs as presented in (5) thus contrasts with the alternative hypothesis that prosodic and syntactic focus marking are additive, which would predict that the combined effect of concurrent syntactic and prosodic focus marking would be at least the same as adding together the individual effects of prosodic focus marking and clefting. For example, in subject focus contexts, the use of subject focus prosody alone should increase ratings and speed up reaction times by a certain amount x compared to the use of broad focus prosody (unmarked syntax with subject focus prosody vs. unmarked syntax with broad focus prosody), while the use of subject clefts by itself should increase ratings and speed up reaction times by a certain amount y (clefts with broad focus prosody vs. unmarked syntax with broad focus prosody). Under the assumption that prosodic and syntactic focus marking are additive, their combined effect in terms of improvement in ratings and reduced reaction times (clefts with subject focus prosody vs. unmarked syntax with broad focus prosody) should correspond to at least x+y, whereas the hypothesis of trade-offs as given in (5) predicts that the combined effect is smaller than x+y.

2.2 Results

Out of 1008 target trials (28 participants * 36 trials), 71 (7.0%) were discarded because of slips of the tongue or disfluencies (N = 29), because the participant deviated from the scripted text (N = 36), or both (N = 6). The remaining 937 target utterances contained 1916 subject syllables, 1166 verb syllables, 1384 object syllables and 304 instances of the cleft marker shi4. After removing the instances of shi4, 4466 syllables were retained for statistical analysis of duration, intensity and voice quality. Due to non-modal voice quality or extreme reduction, no reliable f0 values could be obtained for 170 syllables. Thus, the analyzed data sets for f0 measures only consisted of 4296 data points.

For almost all evaluated dependent measures, the best linear mixed-effects model contained a significant interaction between condition and constituent. As suggested by the editor, the subsections below only describe differences between conditions within constituents (i.e., paradigmatic focus effects). Readers who are interested in a fuller picture of the prosodic realization of Mandarin cleft sentences are referred to the supplementary materials, Section 3, for differences between constituents within conditions (i.e., syntagmatic focus effects), as well as all other effects revealed by the best models (also see Section 2.3 below for a brief summary).

As an overview before discussing results for individual measures, Table 2 gives means and standard deviations for the five numeric acoustic measures and percentage of non-modal realizations by condition and constituent. Even though the best model for voice quality did not show a significant interaction between condition and constituent, data is presented in parallel fashion for comparability. Additionally, Figure 2a gives an illustration of the three f0 measures. It displays average f0 contours for all conditions based on measurements at 10 equidistant points for each syllable (also see the example raw f0 contours in Figure 3). For the sake of comparability, shi4 is omitted and points based on less than 10 individual measurements have been trimmed. Note also that while participants in the production experiment were not asked to produce cleft sentences in broad focus contexts, this combination was recorded for the stimuli used in the perception experiment, see Figure 2b.

Figure 2: Time-normalized average f0 contours based on participants’ realizations in the different conditions for the production experiment (panel a) and stimuli for the perception experiment (panel b) by constituent and syllable in the different syntactic and prosodic conditions.

Figure 3: Example renditions for all three conditions in the production experiment. Item 17 教授卖日历。‘A professor is selling calendars’ with unmarked syntax in broad focus (panel a) and subject focus (panel b), and item 16 卫诺借电话。‘Weinuo is borrowing a phone’ in subject focus cleft (panel c), all spoken by participant s10.

2.2.1 F0 range

The best linear mixed-effects model of f0 range included an interaction between condition and constituent, an interaction between constituent and tone, as well as the predictors syllable and participant gender in the fixed effects (p = .02 for gender; p < .001 for all other effects and interactions).

Figure 4 illustrates the significant interaction between condition and constituent by showing estimated marginal means (plot symbols) and 95% confidence intervals (error bars) together with a compact letter display indicating the results of pairwise comparisons based on the model (letter combinations above the error bars). The letters in the compact letter display are not abbreviations; in fact, they have no meaning themselves. They are instead used as symbols: What is relevant is only whether each combination of letters differs from the others in the same figure. Factor combinations sharing one or more letters cannot be shown to differ significantly from each other, whereas factor combinations that do not share any letter differ significantly.

Figure 4: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of f0 range for different conditions by constituent.

Figure 4 indicates that the conditions only differed significantly from each other on subject constituents: Subjects in both subject focus conditions (subject focus cleft, letter ‘d’, and subject focus with unmarked syntax, letters ‘cd’) had significantly larger ranges than those in broad focus (labelled ‘b’, a letter that is not shared by either of the other two conditions; for clefts: β = –.36, p < .001; for subject focus with unmarked syntax: β = –.21, p = .01). The two subject focus conditions did not differ significantly from each other (as indicated by the fact that their letters both include ‘d’; p = .27). On verbs and objects, the three conditions showed no significant differences (within both constituents, all conditions share at least one letter; all p > .1).

2.2.2 F0 maximum

For f0 maxima, the best model contained interactions between condition and constituent, between constituent and tone, as well as syllable and gender as predictors (p < .001 for all).

As illustrated in Figure 5, the model indicated that subjects had significantly higher maxima in both subject focus conditions compared to broad focus (broad focus, subject – cleft, subject: β = –.32, p < .001; broad focus, subject – unmarked subject focus, subject: β = –.26, p = .01), whereas verbs and objects had lower maxima in the two subject focus conditions than in broad focus (broad focus, verb – cleft, verb: β = 1.61 p < .001; broad focus, verb – unmarked subject focus, verb: β = 1.40, p < .001; broad focus, object – cleft, object: β = 2.09, p < .001; broad focus, object – unmarked subject focus, object: β = 1.76, p < .001). There were no significant differences between the two subject focus conditions on subjects and verbs (all p > .1), but objects showed lower maxima in clefts than in subject focus with unmarked syntax (β = –.32, p = .01).

Figure 5: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of f0 maximum for different conditions by constituent.

2.2.3 F0 minimum

The best model of f0 minimum contained interactions between condition and constituent, between condition and tone, as well as syllable and gender as fixed effects (p = .02 for condition*tone; p < .001 for all other effects and interaction condition*constituent).

Condition did not show any significant effects on subject constituents themselves (all p > .1), but f0 minima of verbs and objects were significantly lower in both subject focus conditions compared to broad focus; see Figure 6 (broad focus, verb – cleft, verb: β = 1.89, p < .001; broad focus, verb – unmarked subject focus, verb: β = 1.41, p < .001; broad focus, object – cleft, object: β = 1.94, p < .001; broad focus, object – unmarked subject focus, object: β = 1.57, p < .001). Additionally, verbs and objects had significantly lower minima in clefts than in subject focus with unmarked syntax (cleft, verb – unmarked subject focus, verb: β = –.48, p < .001; cleft, object – unmarked subject focus, object: β = –.37, p < .01).

Figure 6: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of f0 minimum for different conditions by constituent.

2.2.4 Duration

For duration, the best model contained an interaction between condition and constituent (p < .001), with tone and syllable as additional predictors (p < .01 and p < .001; note that removing the interaction or tone from the best model resulted in convergence issues).

Figure 7 indicates a trend towards subjects having longer durations, and verbs and objects having shorter durations in the subject focus conditions compared to broad focus. However, only two pairwise comparisons between conditions were significant: Subjects had longer durations and objects shorter durations in clefts compared to broad focus (broad focus, subject – cleft, subject: β = –9.38, p = .01; broad focus, object – cleft, object: β = 15.67, p < .0001).

Figure 7: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of duration for different conditions by constituent.

2.2.5 Intensity

The best model of intensity included interactions between condition and constituent (p < .001), between constituent and tone (p = .02) and between condition and tone (p = .02), as well the predictor syllable (p < .001).

Condition did not significantly affect subjects (all p > .1), whereas verbs and objects had lower intensity in the two subject focus conditions than in broad focus; see Figure 8 (broad focus, verb – cleft, verb: β = 1.86, p < .001; broad focus, verb – unmarked subject focus, verb: β = 1.11, p < .001; broad focus, object – cleft, object: β = 2.70, p < .001; broad focus, object – unmarked subject focus, object: β = 1.81, p < .001). Verbs and objects also showed significantly lower intensity in clefts than in subject focus with unmarked syntax (cleft, verb – unmarked subject focus, verb: β = –.74, p < .001; cleft, object – unmarked subject focus, object: β = –.88, p < .001).

Figure 8: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of intensity for different conditions by constituent.

2.2.6 Use of non-modal voice quality

Of the 4466 syllables in the dataset, 855 (19.1%) were realized with non-modal voice quality partially or completely. The best-fitting binomial model of whether or not a syllable was realized with non-modal voice quality contained an interaction between condition and tone (p = .03), as well as constituent and syllable as predictors (p < .001 for both). Unlike for the acoustic measures analyzed above, this model was not significantly improved by including an interaction between condition and constituent (p = .45, also note that adding this interaction to the best model resulted in convergence issues).

Regarding the effect of condition, Figure 9 shows a similar pattern for both tones: Non-modal realizations were least frequent in broad focus and most frequent in subject focus with unmarked syntax (broad focus: 15.9%, clefts: 18.4%, subject focus with unmarked syntax: 23%). The differences between the conditions were, however, more pronounced for tone 1 items, where subject focus with unmarked syntax differed significantly from both other conditions (broad focus – unmarked subject focus: β = –1.10, p < .001; cleft – unmarked subject focus: β = –.52, p = .04; the difference between cleft and broad focus was marginal: β = –.57, p = .05). For tone 4, only the difference between broad focus and subject focus with unmarked syntax was significant (broad focus – unmarked subject focus: β = –.47, p = .02; p > .1 for both broad focus – cleft and cleft – unmarked subject focus).

Figure 9: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) of use of non-modal voice quality for different conditions by tone (T1 = tone 1, T4 = tone 4). (More) negative values on the y-axis indicate fewer realizations with non-modal voice quality, positive or less negative values more realizations with non-modal voice quality.

2.3 Summary and discussion

The production study showed prosodic focus marking that can be described as focus expansion on subjects and post-focal reduction on verbs and objects, as summarized in Table 3. Focus expansion resulted in focused subjects having larger f0 ranges, higher f0 maxima and, in clefts only, longer durations compared to their realizations in broad focus. Post-focal reduction was even more pervasive, showing significant effects for all measures except f0 range. Thus, post-focal verbs and objects had lower f0 maxima and minima, as well as lower intensity than in broad focus, with objects in clefts additionally showing shorter durations compared to broad focus.

Results for non-modal voice quality do not clearly fit the classification into focus expansion and post-focal reduction, since there was no interaction between condition and constituent. Instead, sentences with unmarked syntax showed overall more frequent use of non-modal voice quality in subject focus than in broad focus. Clefts also showed marginally more frequent non-modal voice quality than broad focus, but significantly less than unmarked subject focus for tone 1. For tone 4, clefts did not differ significantly from the other two conditions.

With respect to the hypothesis of prosody-syntax trade-offs in (4), its strong version (4)a is falsified by each instance of prosodic focus marking on cleft sentences (‘cleft’ to the left of a > and ‘broad focus’ to its right in the cells of Table 3). As shown in the table, clefts showed prosodic focus marking, i.e., significant differences in the expected direction from the broad focus baseline, on at least one constituent with respect to all measures except the use of non-modal voice quality. Moreover, again with the exception of voice quality, in every instance where clefts did not differ significantly from broad focus, neither did unmarked equivalents. Worse with respect to (4)a, significant focus marking via duration appeared only in clefts. In sum, the strong version of the hypothesis can clearly be rejected.

The weak version of the hypothesis, (4)b, would be upheld if both clefts and their unmarked equivalents showed prosodic focus marking, but less so in clefts than with unmarked syntax. This is not supported by the data, either. In all instances where the two conditions differed significantly, again with the exception of voice quality, clefts showed stronger post-focal reduction (bolding in Table 3). Together with the instances where clefts showed focus marking and unmarked equivalents did not, this suggests that not only can the hypothesis be rejected even in its weaker form–if anything, clefts show clearer focus marking than syntactically unmarked sentences.

The only prosodic correlate showing effects in the expected direction was voice quality: Non-modal voice quality appeared significantly more frequently in subject focus with unmarked syntax than in broad focus, whereas clefts only differed marginally from broad focus for tone 1 items. As clefts did thus not reliably mark subject focus via voice quality, these results are in line with hypothesis (4)a. Even taking into account the marginal effect for clefts, at the very least hypothesis (4)b can be upheld with respect to voice quality, as sentences with unmarked syntax had significantly more non-modal realizations than clefts for tone 1 items (and insignificantly more for tone 4 items, as well). However, in the face of all other prosodic cues contradicting the hypothesis in (4), it should not be upheld based on the findings for voice quality alone.

Two additional points speak for this conclusion: First, the hypothesis in (4) was more tentative for voice quality than for the other measures, as detailed in Section 1.3. Second, the findings for the five acoustic measures more clearly match previous research than those for voice quality. Prosodic focus marking observed in the present study generally echoes previous research for the five acoustic measures, although prosodic focus marking in terms of duration and intensity had previously only been found in terms of syntagmatic comparisons between constituents within a sentence (recall Section 1.2). In addition to differences between conditions, the present data also showed differences between the constituents, described in Section 3 of the supplementary materials. For the five acoustic measures, these observed differences between constituents, including findings of syntagmatic focus effects, are in line with previous research. They indicate a downtrend in f0 and intensity over the course of an utterance, reflected in subjects having significantly higher f0 maxima, f0 minima and mean intensity than verbs and objects (also see Xu, 1999; Yuan & Liberman, 2010, on declination in Mandarin). Verbs had significantly higher f0 maxima and mean intensity than objects only in subject focus, indicating that the declination was steeper post-focally than in broad focus, replicating previous research on post-focal compression as summarized in Section 1.2. Similarly, differences between the constituents in f0 range were only significant for tone 4, not for tone 1 (also see Y. Chen & Gussenhoven, 2008; B. Wang & Xu, 2011; and Xu for differences between lexical tones with respect to post-focal compression). In contrast to these findings echoing existing research, the present study indicated that the use of non-modal voice quality increases significantly later in the sentence, which has not been reported for Mandarin before. The present finding of prosodic focus marking via increased use of non-modal voice quality does agree with previous research, but was not localized to any particular constituent, unlike in previous studies. A follow-up study considering all lexical tones and varying focus location would be needed.

To summarize, with the possible exception of the use of non-modal voice quality, the results of the present production study clearly contradict the hypothesis of prosody-syntax trade-offs given in (4). As pointed out by an anonymous reviewer, this could be because the task of reading sentences as answers in imagined dialogue situations is not as engaging as a more interactive or naturalistic scenario. However, two observations speak against this explanation. First, while significant differences often did not appear between the two subject focus conditions, significant differences between broad focus and subject focus appeared for all dependent variables. This suggests that the experimental manipulation was effective in engaging participants at least to a certain extent, since they did produce systematic prosodic focus marking (which, as pointed out, matches expectations based on previous research). Second, some significant differences did appear between the two subject focus conditions, and these almost exclusively indicated stronger prosodic focus marking for clefts than with unmarked syntax. Therefore, the present findings indicate that prosodic focus marking may even be stronger in clefts than in sentences with unmarked syntax.

3.2 Results

All final models and full results of pairwise comparisons are reported in the supplementary materials, Section 6. Below, only estimates and p-values are reported.

3.2.1 Ratings

Altogether 2448 rating responses were evaluated (102 participants * 24 items). The GAMM with the best fit to the data contained only an interaction between context and prosody, as well as a smooth for trial and by-participant and by-item random smooths for context (p < .01 for trial; p < .001 for smooths and interaction context*prosody). Other interactions, as well as including syntax or list, did not significantly improve model fit (p = .08 for context*prosody*syntax; p = .16 for context*syntax; p = .86 for list; adding syntax or prosody*syntax decreased model fit, increasing AIC by 1.85 and 0.46 respectively). Thus, there was no evidence that participants rated clefts differently from unmarked syntax, nor that either syntactic structure was more appropriate in either context or prosodic realization.

Answers to subject focus questions were rated as more appropriate than those to broad focus questions; see Figure 10. This difference was significant with both broad focus and subject focus prosody on the answers themselves (broad focus context, broad focus prosody – subject focus context, broad focus prosody: β = –2.27, p < .001; broad focus context, subject focus prosody – subject focus context, subject focus prosody: β = –3.77, p < .001). Moreover, context affected the effect of prosody: In broad focus contexts, answers with broad focus prosody were rated significantly higher than those with subject focus prosody (β = 1.08, p < .001), whereas the opposite was the case in subject focus contexts (β = –.43, p < .001).

Figure 10: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) for rating by context and prosody. Since the y-axis reflects the transformed continuous dependent variable used in statistical modelling, dashed lines in light blue and bold numbers on the right indicate boundaries relating these to the original rating categories.

3.2.2 Reaction times

Of the 2448 rating responses, reaction times were discarded for 59 responses that occurred more than 500 ms before the end of the stimulus sound and for 67 responses that occurred more than 10,000 ms after its end (5.15% of the data total), and leaving 2322 trials for analysis. Reaction times from these trials (from the end of the stimulus sound, in ms) were log transformed after adding the largest negative reaction time to all reaction times, since negative numbers cannot be log transformed. This resulted in a distribution closely resembling a bell-shaped curve. Note that in Figure 11 below, both the log transformation and the addition of the largest negative reaction time are reversed, i.e., model results are simply displayed on the original ms scale.

Figure 11: Estimated marginal means, 95% confidence intervals and results of pairwise comparisons (compact letter displays: different letters indicate significant differences, shared letters that no significant different difference could be shown) for reaction time (in ms) for stimuli with different prosody (panel a) and presented in different contexts (panel b) and receiving different ratings (panel c).

The best model of reaction times contained the predictors prosody, context and rating, a smooth for trial, a by-item smooth for context and a random smooth for participant (p = .01 for prosody; p < .001 for context, rating and smooths). Neither the effect of syntax, list, nor any interactions were significant (adding context*prosody*syntax, context*syntax, prosody*syntax, context*prosody, syntax or list decreased model fit, increasing AIC by 1.02, 1.07, 1.42, 1.32, 0.64 and 0.26 respectively).

Figure 11a shows that participants reacted slightly faster to stimuli with subject focus prosody than to stimuli with broad focus prosody (β = –.06, p < .001). Additionally, as shown in Figure 11b, reaction times were smaller for stimuli with subject focus contexts than for those with broad focus contexts (β = –.155, p < .001). Finally, reaction times were smaller with more extreme ratings, whereas the ratings in the middle of the scale were associated with longer reaction times (cf. Figure 11c and Table 20 in the supplementary materials).

3.3 Summary and discussion

The results of the perception study showed a clear effect of prosodic focus marking on ratings, which interacted with context in the expected way: Participants rated target sentences higher when their prosody matched the information structure induced by the preceding context. In contrast to context and prosody, syntax did not significantly affect ratings or reaction times and did not interact significantly with either factor, contradicting both versions of the hypothesis of trade-offs in (5). The weak version of the hypothesis (5b) predicted that when syntactic and prosodic focus marking are combined, effects of both prosody and syntax would be detectable, but participants would be less affected by one of them compared to the effect this marking would have when occurring on its own. In contrast to this prediction, the results showed a total absence of syntax effects, whereas the effect of prosody persisted even when combined with concurrent syntactic marking. The strong version of the hypothesis (5a) predicted that both syntax and prosody would affect ratings and reaction times on their own, but that in conditions with concurrent prosodic and syntactic marking, either syntax or prosody would be redundant and its effect would not be detectable. However, a syntax effect was absent not only when syntactic and prosodic marking concurred, but also when they did not. Thus, the absence of a syntax effect in conditions with concurrent syntax and prosody cannot be explained as a trade-off. Put differently, in contrast to the hypothesis, there was no evidence for an effect of syntax that could have been reduced or neutralized by concurrent prosody.

The absence of any effect of syntax constitutes a difference from a parallel study on English, where an effect of syntax appeared with the same methods (Arnhold, 2021). This difference between the languages is in line with previous research showing that clefting has weaker effects than prosodic focus marking on perception and processing in Mandarin (Yan et al., 2020; Yan & Calhoun, 2019, 2020) and may also be relatively less important for perception and processing in Mandarin than in English (S. H. Chen et al., 2012; but see Yan & Calhoun, 2020). However, it is interesting that the present experiment showed no evidence that clefting affected the ratings at all. Previous studies on Mandarin also did not find clefting to induce lexical priming (Yan & Calhoun, 2019), speed up the rejection of false alternatives (Yan & Calhoun, 2020) or improve ratings compared to unmarked syntax, even when focus marking matched the context (Yan et al., 2020). Still, the syntactic manipulation affected all these tasks: It slowed reaction times overall in lexical decision and false alternative rejection tasks,⁴ as well as lowering ratings when the clefted constituent was not the one that the context indicated to be focused (Yan et al., 2020; Yan & Calhoun, 2019, 2020). Importantly, all of these studies lacked a broad focus baseline. Therefore, either prosodic and syntactic focus marking indicated narrow focus on the same constituent (e.g., (6a) for subject focus and (6b) for object focus) or they provided conflicting clues, each pointing to a different constituent being in narrow focus (e.g., (6c) with subject cleft, but prosodic focus marking on the object, or (6d), an object cleft with prosodic focus marking on the subject; underlining indicates prosodic focus marking in these examples). As Yan et al. acknowledge, this mismatch in itself could have lowered ratings. Actually, if mismatch cases are discounted, as well as conditions differing only marginally from each other, the only significant effects of the syntactic manipulation in Yan et al.’s experiment that remain appeared in object focus contexts: Subject clefts were rated lower than unmarked syntax when prosody marked the subject as focused, i.e., clefting a background constituent additionally lowered ratings (e.g., following ‘What did the captain put on?’, (6a) was rated even lower than (6e)). However, when prosody marked the object as focused in accordance with the context, object clefts were also rated significantly lower than unmarked syntax, though still higher than (mismatch) subject clefts with the same prosody (e.g., (6f) was preferred over (6b), which was in turn preferred over (6c) as an answer to ‘What did the captain put on?’). This could indicate that it is not only object clefts that are less frequent and, accordingly, less readily acceptable, than subject clefts in Mandarin (as stated by Yan et al., based on their own results and previous literature), but that clefts are generally less acceptable in object than in subject focus. This could be because the appropriate cleft in object focus would be an object cleft or, more likely, causality could run the other way around, in line with cross-linguistically common subject/non-subject asymmetries mentioned in Section 1.2. It is therefore possible that the present experiment did not detect any effects of syntax because it contained neither prosody-syntax mismatches of the type in (6c&d) nor object focus conditions.

The present experiment did not include items where prosody and clefting indicated narrow focus on different constituents, as unlike for Yan et al. (2020), the research question was not whether prosody or syntax would prevail in such a mismatch. Instead, the purpose of the present study was to test whether effects of syntactic and prosodic focus marking would diminish when combined. Therefore, conditions where prosody and clefting indicated narrow focus on the same constituent were compared to conditions where only one of them did, while the other was neutral and did not indicate narrow focus on any constituent. Of course, it can be questioned whether prosody can ever be truly neutral. While the unmarked syntactic structures used in the present study are equally compatible with broad focus and subject focus (and with other information structures), broad focus prosody might be considered infelicitous following a subject focus question and therefore not truly be neutral–at the very least, the results of the present study indicate that Mandarin listeners find it significantly less appropriate than subject focus prosody in this context. Based on this, one might argue that clefts with broad focus prosody, as used in the present study, also constitute a mismatch between prosody and syntax. However, the present results suggest that this is not in itself enough to lower ratings significantly in the way that the mismatches in Yan et al. did. It appears that even if broad focus prosody is not neutral, it does not conflict with subject clefting as much as object focus prosody does or, alternatively, that the particular conflict tested by Yan et al. is just particularly unacceptable in Mandarin. This is in line with Greif and Skopeteas’ (2021) conclusion that Mandarin subject clefts with object focus, even if supported by a suitable context, are “marginal in language use” (p. 17).⁵

It may also be relevant that bare shi clefts can be interpreted as verum focus (recall Section 1.2), although it is not agreed whether this possibility exists whenever prosody does not mark subject focus (Paul & Whitman, 2008) or, in line with cross-linguistic trends, only with prosodic focus marking on the copula shi4 (Liu & Shi, 2022). Liu and Shi’s description is compatible with the idea that the clefted constituent is focused per default, but this default can be overwritten with specific verum focus prosody. Paul and Whitman’s account, on the other hand, suggests that Mandarin clefts are a means of marking focus on the clefted constituent only when prosody concurs. However, Greif and Skopeteas (2021) found that subject clefts were rated as more appropriate than unmarked syntax in subject focus contexts–even though their study used written stimuli. While participants likely supplied implicit prosody, as mentioned in footnote 1, it is interesting that their implicit prosody for clefts apparently marked the subject as focused. Alternatively, participants may have interpreted the subject clefts as verum focus but still judged this as a better fit with subject focus contexts than the default interpretation for unmarked syntax, which is that the subject is a topic (see Section 1.2).

At any rate, the possibility of interpreting subject clefts as marking verum focus does not explain the absence of a syntax effect in the present experiment. Verum focus could perhaps be seen as more similar to broad focus than subject focus, but it is not the same (e.g., My aunt IS steaming mushrooms would be an odd answer to a broad focus question such as, What’s happening?). Thus, in the broad focus contexts, it makes sense that clefts with broad focus prosody (indicating verum focus according to Paul & Whitman, 2008) were rated more highly than clefts with subject focus prosody due to prosody alone–for the same reason that sentences with unmarked syntax were rated more highly with broad focus prosody than with subject focus prosody in these contexts (e.g., (7a) was preferred over (7b), just like (7c) was preferred over (7d)). However, it is still unexpected that among sentences with broad focus prosody, clefts like (7a) were not rated lower than unmarked syntax equivalents like (7c) in broad focus contexts. Even with a possible verum focus interpretation, a cleft should not be an appropriate answer to a broad focus question. It is possible that participants accommodated an unprompted insistence on the truth of the utterance, but if so, it is surprising that this affected neither ratings nor reaction times. Even if participants ultimately judged the fit between question and answer as high, accommodation of an initially less expected interpretation could be expected to lead to processing delays.

Regarding subject focus contexts, when prosody marks the subject as focused, higher ratings would be expected for clefts like (7b) than for unmarked syntax as in (7d) if prosodic and syntactic focus marking were additive. The fact that no significant differences appeared could be taken as support for the strong version of the hypothesis in (5), i.e., the effect of syntax being undetectable when combined with the effect of prosody. However, as stated above, given that effects of syntax appeared nowhere else in the present experiment, it seems more likely that clefting simply did not affect ratings at all. A potential explanation for this is explored in the general discussion (Section 4).

A final finding that should be discussed, though it is not directly relevant to the question of trade-offs, is that context itself had a significant effect in the present study, with answers to subject focus questions receiving higher ratings than answers to broad focus questions. Arnhold (2021) observed the same context effect in a parallel experiment on English and speculated that the more specific narrow focus questions may have induced the perception of a better fit between question and answer in these cases. In line with this interpretation, in the present experiment participants responded significantly faster in subject focus contexts than in broad focus contexts. Additionally, reaction times were also significantly smaller for stimuli with subject focus prosody compared to broad focus prosody. Interestingly, however, the English results showed a ceiling effect where all answers to subject focus questions were rated so highly that no difference could be observed between the two prosodic conditions (or the syntactic ones, which likewise differed significantly in broad focus contexts). This contrasts with the present findings for Mandarin, which showed a significant difference between the ratings of the two prosodic realizations in both context conditions. It is possible that the context effect was weaker in Mandarin than in English, or that the effect of prosodic focus marking was stronger in Mandarin. The latter explanation seems more likely, given that previous studies on the perception of clefts also show a stronger effect of prosodic focus marking in Mandarin than in English (S. H. Chen et al., 2012; Yan & Calhoun, 2020).

To sum up, this experiment did not provide evidence for the hypothesis of prosody-syntax trade-offs.