1.1 On the nature of allegedly “syntactic” variations

The syntactic alternatives for both i) the ditransitives (double object versus prepositional dative; Anttila et al. 2010) and ii) the genitive attributes (genitive case vs. prepositional genitive; Shih et al. 2015) involve different constructions that are semantically quasi-equivalent alternatives but that do not share a syntactic relationship. It is therefore not entirely clear whether the reportedly prosody-driven choice between these options affects syntactic operations per se, or whether it is better conceived as the selection between two precompiled syntactic frames, as advocated in construction grammar (Goldberg 1995; Kay 2002). In the latter case one might instead assume a prosodic effect on the selection from the “lexicon of constructions”, as it were, not so much a direct prosodic effect on syntactic computation or on syntactic relations.

Even when word order is affected by linguistic rhythm, it is debatable whether this constitutes a prosodic effect on syntactic computation. Agbayani & Golston (2010; 2016) consider hyperbaton in Classical Greek and Latin to be a type of word order alternation that involves phonological constituents rather than syntactic ones, and cannot be explained in purely syntactic terms. In this regard, their notion of “prosodic movement” (Agbayani et al. 2015; Bennett et al. 2016) embodies a conception of word order variation as phonological rather than syntactic, and this notion may well apply to cases like (2) in which the syntax is largely indifferent to the word ordering.

Finally, the prosody-driven inclusion or omission of optional elements need not be a prosodic effect on syntax, but merely a case of phonological ellipsis. This may be true of optional that in English complement clauses as there is no obvious syntactic difference between clauses with vs. without that other than the presence or absence of the complementiser.

With these considerations in mind, the studies reviewed so far suggest that the observable rhythmic effects on sentence structure do not necessarily touch upon intrinsically syntactic relations or computations. Instead, they may affect either the resulting phonological representation (the mere linearisation of syntactic constituents), or the choice among syntactically unrelated, possibly precompiled, constructions that happen to be quasi-synonymous.

However, there are cases of syntactic variability that appear to involve processes that may be more properly construed as syntactic in nature and that cannot be relegated either to the phonology or to the “lexicon of constructions”. One example is the alternation between introduced (4a) and unintroduced (4b) complement clauses (CC) in German, which is the object of the study to be presented here. These structures are the German equivalent to English complement clauses with or without that.

In contrast to its English analogue, this kind of alternation does constitute a difference in word order (verb-final subordinate clause versus verb-second structure), but the word order difference is, crucially, not reducible to phonology, as it affects syntactic constituents, not phonological ones. Also, the word order difference is not a simple one, as the structure in (4a) requires the complementiser dass which is not licensed in verb-second structures like (4b), i.e. there is a complementary distribution of subordinating conjunctions and verb-second order. On the other hand, one might say that (4a) and (4b) are different and independent constructions; but in contrast to the cases discussed above (i.e. the English dative alternation or genitive construction choice), there is a systematic syntactic relationship between the two sentence structures. The conventional wisdom on these German structures holds that the verb-final order of a subordinate clause is the underlying word order. From this basic word order, other orders may be derived via verb movement and topicalisation (4) (see, among many others, e.g. Thiersch 1978; Wöllstein-Leisten et al. 1997).

The alternation between complement clauses like (4a) and (4b) therefore constitutes a test case for the introductory question whether linguistic rhythm affects syntax in language production. Applying this question to the alternation concerning complement clauses in German is furthermore motivated by the above-mentioned language production experiment by Lee & Gibbons (2007) who found that the rhythmic environment affects the inclusion or omission of the optional complementiser that in English.

Section 2 provides more background on the syntax, semantics and pragmatics of introduced versus unintroduced complement clauses. In Section 3, we report on four experiments that largely fail to produce compelling evidence for a rhythmic effect on the choice between introduced and unintroduced complement clauses in German. We consider potential reasons for this in Section 4; specifically, we present a synoptic discussion of studies concerned with ostensible rhythmic effects on sentence structure and argue that rhythmic effects are restricted to sub-clausal domains of structure building. Section 5 concludes the paper.

3.1 Experiment 1

Experiment 1 is a conceptual replication of the language production experiment by Lee & Gibbons (2007), adapted to German, that closely follows their experimental protocol. Also, the sample size of this and the following experiment (i.e. the number of participants and items) is modeled on the example of Lee & Gibbons. Participants read sentence pairs silently in order to produce them afterwards in response to a recall cue. Sentence pairs consisted of a filler sentence as first stimulus and an experimental target sentence as second stimulus. Following Lee & Gibbons, target sentences were simple matrix clause – complement clause structures that license the optional complementiser dass in German like (5) and (6). The presence or absence of dass (together with the word order of the complement clause) and the stress quality of the surrounding syllables were systematically varied.

Assuming that participants vary their use of dass when producing the memorised sentences, we predicted that they do so in favour of an alternating rhythm. Thus, in sentences with an embedded trochaic subject (e.g. NADja), the unstressed dass should be produced more frequently than in those with an iambic one (e.g. NaDINE) as the latter would result in a stress lapse (i.e. a sequence of two unstressed syllables, deviating from the optimal rhythmic alternation of stressed and unstressed). Additionally, we predicted that an unstressed final syllable of the main verb would lead to more omissions of dass, again in order to avoid a stress lapse. Conversely, the stressed monosyllabic embedding verb should result in more frequent use of the unstressed complementiser in order to avoid a stress clash (i.e. a sequence of two stressed syllables).

3.2 Experiment 2

Experiment 2 was a replication of Experiment 1 with a few alterations to reduce memory load; this was done to achieve a higher number of valid productions. As in Experiment 1, participants read sentences silently in order to produce them afterwards in response to a three-word recall cue. Unlike Experiment 1, stimuli weren’t sentence-pairs, but a combination of a target sentence as first stimulus and a simple arithmetic task.

3.3 Evidence against rhythmic effects on complementiser use? A model selection approach via Bayes factors

In contrast to the study by Lee & Gibbons (2007) on English, the two language production experiments on German fail to provide evidence in favour of rhythmic effects on complementiser use. The posterior distributions for both rhythmic effects, NameStress and VerbStress, and their interaction, are centered relatively close to 0 in both experiments. Using Bayes factors, it is possible to ascertain the degree to which this absence of evidence provides evidence of absence of these effects (Kass & Raftery 1995). In general, Bayes factors can be used for hypothesis testing, as they express the extent to which the data support one hypothesis (as formulated in terms of a model for the data) over another hypothesis (i.e. a competing model). By means of Bayes factors, we compare the likelihood of the observed data given the “full” models (as reported in Tables 3 and 6) and the likelihood of the data given models from which factors of interest are eliminated, i.e. assuming a point null value for these factors (=models conforming to the null hypothesis).

Specifically, we compare the likelihood of the data given the full models for Experiment 1 (Exp1), Experiment 2 (Exp2), and a model for the combined data set (Exp1 + Exp2)³ with i. the likelihood of the data given a “reduced” model and ii. a “null” model. The reduced model assumes that the factor NameStress does not contribute to the manifestation of the dependent variable, i.e. to dass-mention; consequently, the coefficient of NameStress and the interaction involving NameStress are culled from both the fixed effects and the random effects terms of these models. The null models assume that both rhythmic factors, NameStress and VerbStress, have no influence on complementiser use; that is, both factors and the interaction are eliminated from these models.

Bayes factors are affected by the prior distributions assumed for the model coefficients. Following suggestions by Vasishth et al. (2018), we report Bayes factors under different priors that are all centered around 0 but have different spreads. The spread σ of the prior around 0 expresses the assumption that the null model is Normal (0, σ). Thus, smaller values of σ express more prior certainty that the true mean is 0. Apart from the weakly informative prior with sd = 1 (under which posteriors with means greater than 1 on the log odds scale are assumed to be quite likely),⁴ we also chose tighter priors with sd = .5 and sd = .2 for the rhythmic effects NameStress and VerbStress.

The Bayes factors reported in Table 7 indicate the degree to which the data support the reduced models (top) or null models (bottom) over the full model. Bayesian convention (Jeffreys 1998; Lee & Wagenmakers 2014) holds that Bayes factors greater than 3 indicate moderate evidence, and Bayes factors greater than 10 constitute strong evidence in favour of the reduced/null models over the full model (the corresponding numbers are marked in bold in Table 7); Bayes factors close to 1 are deemed inconclusive, and Bayes factors smaller than .3 constitute evidence in favour of the full model (the latter are italicised in Table 7).

In the case of Experiment 1, the Bayes factors do not indicate a preference for the reduced model over the full one, and in fact some evidence against the reduced or null models. However, given the relative sparseness of the data from Exp 1, we suspect that the success of the full model in Experiment 1 is due to overfitting.⁵ With accumulating data in Exp 2 and the combined data set, the reduced or null models are clearly supported over the more complex models. The comparison of the models for the combined data set do support the reduced and the null models, that is the models that do without the rhythmic coefficients NameStress and VerbStress are more plausible given the data than models including these factors. Note that the support for the reduced or full models is weaker the tighter the priors are. This is expectable given that the reduced or null models are relatively similar to a model that assumes a narrow prior around 0 for the effect in question.

In sum, for the combined data analyses (which have all the data available), the model comparisons using Bayes factors suggest that the evidence either favors the null/reduced models, or there is strong evidence for the null/reduced models. That is, local linguistic rhythm at the clause boundary does not appear to affect the syntactic choice between introduced and unintroduced complement clauses in spoken German.

3.4 Experiment 3 – spoken language corpus

In order to validate the results of the language production experiments, we examined the dgd archive (http://dgd.ids-mannheim.de), the largest collection of spoken German corpora.⁶ For our purpose, we chose sub-corpora that contained transcriptions of unscripted speech by native speakers of Standard German only, namely the FOLK corpus and the Freiburg corpus, which comprise 2.6 million word tokens in 270 hours of speech in total. The dgd archive does not provide syntactic annotations other than part-of-speech (POS) tagging. Therefore, the search for relevant structures turned out to be rather time-consuming because the data had to be sifted, and the relevant prosodic features annotated, by hand.

3.5 Experiment 4 – written language

The previous experiments do not hint at an effect of linguistic rhythm on the choice between introduced and unintroduced complement clauses in spoken German. However, we know that certain types of language use, styles or registers are especially prone to observe constraints on prosodic well-formedness. A host of recent research suggests (see e.g. the collections of papers in Frazier & Gibson 2015; Kentner & Steinhauer 2017) that prosody plays a significant role in the processing of written language, in line with the Implicit Prosody Hypothesis (Fodor 1998; 2002). This may seem paradoxical because written language generally lacks explicit cues to prosodic structure (e.g. Chafe 1988), but see Evertz & Primus (2013) and Ktori et al. (2018) on sublexical cues to word stress. In any case, psycholinguistic research strongly suggests that readers have immediate access to prosodic features like stress (Ashby & Clifton 2005) when reading words, and they make use of this information when parsing sentences (Bader 1996; Breen & Clifton 2011; 2013; Kentner 2012; Kentner & Vasishth 2016). Writers are likewise known to structure their text in a way that aligns prominent words with sentence positions that are likely to receive a (nuclear) accent when the sentence is spoken (Bolinger 1957; Anttila et al. 2018). Furthermore, the process of writing itself appears to be accompanied by prosody. This is at least suggested by Fuchs & Krivokapić (2016) who show that pauses between key strokes during spontaneous writing are correlated with prosodic breaks in a read rendition of the same text. Similarly, in handwriting, writers leave greater spaces between letters belonging to different prosodic constituents (syllables, feet, phonological words) than between letters that belong to the same prosodic constituent (Domahs et al. 2016; Bronner et al. 2018).

Given the role of prosody in written language, we ask whether the choice between introduced and unintroduced complement clauses in the written modality remains unaffected by rhythmic-prosodic features of the words surrounding the clause boundary, as in the studies on spoken speech (Experiments 1, 2, and 3). This would suggest that this specific syntactic decision is a very stable one and cannot easily be undone in favour of phonological well-formedness – in contrast to other types of syntactic variation. On the other hand, an effect of linguistic rhythm in the written modality would not make us assume bidirectional interaction between syntactic encoding and phonological encoding in language production. Writers can revise their wording and the final text does not give away the spontaneous syntactic choice writers make, but only the end-result of a process that may well involve revisions. Still, an effect would at least show that writers do not tie themselves down to their initial syntactic decision concerning the form of the complement clause but consider the local rhythmic environment at the clause boundary when formulating complement clauses.

4.1 Lack of rhythmic effect on dass-mention: Potential reasons and implications

The two sentence production experiments and the two corpus studies presented here largely fail to support our initial predictions concerning the role of rhythm on the choice between introduced and unintroduced complement clauses in German. The only evidence pointing into the predicted direction is the rather uncertain effect of VerbStress found in Experiment 4: writers apparently favour the complementiser dass, and hence a verb-final complement clause, when the embedding verb ends in a stressed syllable. We interpret this effect to reflect the propensity for rhythmic alternation in written language. We hasten to add that this effect does not necessitate the possibility of direct interaction between grammatical encoding and phonological encoding in models of sentence production (as advocated in e.g. Vigliocco & Hartsuiker 2002; Shih 2014). Rather, this effect is explicable with recourse to the notion of a monitoring loop (Levelt 1983; Hartsuiker & Kolk 2001) that checks the output after the formulation stage of language production has been completed, and may trigger stylistic repairs.

Apart from the uncertain effect of VerbStress in Experiment 4, the lack of a rhythmic effect in the direction predicted by Anttila et al. (2010), Shih (2014) or Vogel et al. (2015) in general, and Jaeger (2006) or Lee & Gibbons (2007) in particular, is consistent across the four experiments. While it is problematic to argue on the basis of “absence of evidence”, this consistency warrants commentary (but note the suggestive evidence of absence obtained from the Bayes factor analysis in Section 3.3).

As alluded to in the introduction, the kind of syntactic variability studied here is qualitatively different from the kinds of syntactic variability studied by other authors. Specifically, the choice between a complement clause with vs. without the overt complementiser in German necessarily involves the choice between verb-final structures (with complementiser (4a)) vs. verb-second structures (without complementiser (4b)). No such word order difference is involved in the choice between English complement clauses with vs. without complementiser.

The fact that, in English, the difference between the two constructions merely affects the presence or absence of that invites the assumption that this seemingly syntactic difference is, in essence, a phonological one: that is, both complement clause variants may provide a structural position for the complementiser which, in the case of that-less complement clauses, simply remains unpronounced. If the presence or absence of that is indeed regulated by the phonological processing module, a rhythmic effect on complementiser use, as reported by Lee & Gibbons, is explicable and expectable without assuming a bidirectional interaction between grammatical and phonological encoding.

In the following, we submit an admittedly ad hoc and speculative, but testable, approach that accounts for the difference in susceptibility to phonological influences between complement clause selection in German (putatively no phonological effect), and other kinds of syntactic variation that have been shown to be affected by phonological constraints.

We argue that the encoding of clauses (or propositions for that matter) is less prone to be affected by rhythmic constraints than the encoding of phrases below the clause; we do so on the basis of the following assumption: the syntax of clauses needs to be specified earlier in sentence production than phrasal categories smaller than a clause; we conjecture that the specification of a coarse syntactic skeleton of clauses is in fact the prerequisite for phonological encoding, and that is why we find complement clause structure to be largely immune to influences of stress and rhythm.

This assumption is in line with a proposal by Ferreira (2000) who suggests that the formulation stage in language production involves, in a first step, the generation of basic syntactic frames or “elementary trees” (Frank 1992) that consist of a simple proposition, i.e. a predicate plus its extended projection including its arguments, in short: a clause; only once the argument slots are syntactically specified, e.g. with the corresponding NP treelets, their phonological encoding may begin. Evidence for this kind of hierarchical planning comes from a study by Bock & Cutting (1992) who found subject-verb agreement errors (8) triggered by the local noun books to be more likely within the work space of a clause (8a) than across clause boundaries (8b).

This suggests that sentence production “proceeds in hierarchical rather than sequential fashion, with the planning of clauses preceding the sequencing of words” (Bock & Cutting 1992: 122).⁹

Applied to the case at hand, we assume that the grammatical encoding of the complement clause is triggered once the embedding verb in the matrix clause is syntactically encoded (in the parlance of the language production literature: at the stage of “lemma selection”). However, the syntactic specification of the complement clause is encapsulated from the subsequent phonological encoding of the embedding verb.

Likewise, the syntactic skeleton of the complement clause is specified without regard to the phonological properties of the complement clause subject. That is, phonological imperfections that arise when the two clauses are fused cannot easily become undone because, at this point, the production system has already decided upon their syntactic form.¹⁰

If true, this explanation makes strong predictions about what kinds of syntactic variability are susceptible to phonological well-formedness conditions and what kinds are relatively immune. Testing this prediction is beyond the scope of this paper. However, based on the available evidence concerning rhythmic-phonological effects on syntactic encoding, we can at least provide a plausibility check. To this end, we give a synopsis of relevant studies and categorise the available evidence as follows (see Table 11).

The first set of studies (upper section of Table 11) suggests a rhythmic effect on the realisation of optional elements (e.g. the optional infinitive marker to in English that is often omitted in the context of unstressed syllables).¹¹ As discussed above, a phonological effect on the mention of such elements is explicable without assuming bidirectional information flow between phonological and syntactic encoding; instead, we assume that these elements are part of the syntactic representation even when they are phonologically empty. That is, the presence/absence of these elements is exclusively regulated in the phonological encoding stage.¹² Accordingly, the fact that the phrasal level affected is seemingly a clause in the case of Lee & Gibbons (2007), does not constitute counter-evidence to our proposal.

The second set (middle section in Table 11) consists of studies in which speakers/writers exploit word order flexibility in order to achieve a felicitous rhythmic representation. The most prominent phenomenon in this set is the ordering of (mostly NP/DP) conjuncts, e.g. salt and pepper vs. pepper and salt (McDonald et al. 1993; Benor & Levy 2006; Lohmann 2014). The word order variations do not touch upon the semantics of these constructions; however, some of them may signal a higher degree of expressiveness (e.g. the determiner inversion in Schlüter 2005; Kentner 2018) or possibly involve stylistic mannerisms (e.g. the verb cluster ordering reported in Vogel et al. 2015). The syntax appears to be indifferent to at least some of these word order alternations: For example, determiner inversion engenders split constituents (quite a long report ∼ [a [[quite long] report]]) and thus violates rules concerning phrasal integrity.

Nevertheless, we assume that the phenomena in this set are relevant for the stage of grammatical encoding that is concerned with linearising the syntactic structure (the “positional level” of sentence formulation according to Bock & Levelt 2002). In this set, the phrasal levels affected by linguistic rhythm are invariably below the clause level.

Finally, the third set (bottom section of Table 11) contains cases in which speakers/writers consider linguistic rhythm when choosing a particular syntactic construction. This set involves studies on genitive (Shih et al. 2015) and dative construction choice in English (Anttila et al. 2010) and the case of negated attributive adjectives (avoidance of stress clash: ^*/?a NOT POpular person but ^✔a NOT very POpular person, Schlüter 2005). These phenomena involve syntactic decisions that go beyond the mere ordering of constituents (possibly touching upon the “functional level” of sentence production, cf. Bock & Levelt 2002). The phrasal categories affected by rhythm in this set are, again, smaller than a clause.

All in all, given this synopsis, it seems plausible that effects of linguistic rhythm on syntactic encoding are restricted to phrase categories that are smaller than a clause. The lack of a rhythmic effect in our experiments is consistent with this assumption. However, the nature, and the sparseness, of the data in Table 11 do not allow firm conclusions to be drawn. For one thing, there are, as far as we are aware, no other studies directly testing effects of linguistic rhythm on the production of higher syntactic levels such as the clause. The present study on complement clause structure in German appears to be the only one. Moreover, several studies only consider written material. As mentioned above, written corpora are not very informative about the process of language production, as they merely reflect the result of the process, and they do so in a modality that is often considered secondary to the spoken modality. Nevertheless, in the studies that consider written and spoken data, the findings are largely consistent. On a more general note concerning the current state of science communication, there is a very strong bias for positive results to be published, and negative or null results are hardly reported (Open Science Collaboration 2015). Therefore, the list in Table 11 has to be taken with some caution.