Word-order variation and coherence in German infinitival complementation

Abstract

This study provides a synthesis of corpus-based and experimental investigations of word-order preferences in German infinitival complementation. We carried out a systematic analysis of present-day German corpora to establish frequency distributions of different word-order options: extraposition, intraposition, and ‘third construction’. We then examined, firstly, whether and to what extent corpus frequencies and processing economy constraints can predict the acceptability of these three word-order variants, and whether subject raising and subject control verbs form clearly distinguishable subclasses of infinitive-embedding verbs in terms of their word-order behaviour. Secondly, our study looks into the issue of coherence by comparing acceptability ratings for monoclausal coherent and biclausal incoherent construals of intraposed infinitives, and by examining whether a biclausal incoherent analysis gives rise to local and/or global processing difficulty. Taken together, our results revealed that (i) whilst the extraposition pattern consistently wins out over all other word-order variants for control verbs, neither frequency nor processing-based approaches to word-order variation can account for the acceptability of low-frequency variants, (ii) there is considerable verb-specific variation regarding word-order preferences both between and within the two sets of raising and control verbs under investigation, and (iii) although monoclausal coherent intraposition is rated above biclausal incoherent intraposition, the latter is not any more difficult to process than the former. Our findings indicate that frequency of occurrence and processing-related constraints interact with idiosyncratic lexical properties of individual verbs in determining German speakers’ structural preferences.

1 German infinitival complementation

German infinitival complementation provides an intriguing challenge for contemporary linguistic theory as infinitive-embedding verbs and their complements can show considerable variability with respect to their linearization patterns. Infinitival complements can either be placed at the right periphery of the embedding verb, a word-order variant referred to as ‘extraposition’ (1a), they can be intraposed, i.e. placed between complementizer and matrix verb (1b), or they can be split into a discontinuous infinitival construction referred to as the ‘third construction’ (1c) (Haider 2010).

(1)	a.	extraposition
		dass	Fred	versucht	[den	Hund	zu	streicheln]
		that	Fred	tries	the	dog	to	pet
	b.	intraposition
		dass	Fred	[den	Hund	zu	streicheln]	versucht
		that	Fred	the	dog	to	pet	tries
	c.	third construction
		dass	Fred	[den	Hund]	versucht	[zu	streicheln]
		that	Fred	the	dog	tries	to	pet
		‘that Fred tries to pet the dog'

Intraposed infinitives as in (1b) were first recognized as potentially ambiguous by Bech (1955), who noted that some verbs build a single clausal domain with their infinitival complement while others select infinitives that form a clausal domain independent of the governing predicate. Bech (1955) refers to these two infinitive types as ‘coherent’ (i.e., monoclausal) and ‘incoherent’ (i.e., biclausal) infinitival constructions, respectively.¹ In more recent work, incoherently construed infinitives are usually assumed to be CP constituents as shown in (2a), whereas coherent infinitives are thought to build a verbal complex with the matrix verb as indicated in (2b).

(2)	a.	dass	Fred	[CP	PRO	den	Hund	zu	streicheln]	versucht
	b.	dass	Fred		den	Hund	[VP	zu	streicheln	versucht]
		that	Fred		the	dog		to	pet	tries
		‘that Fred tries to pet the dog'

Despite extensive discussion of these syntactic variation patterns in formal linguistic theory (Haider 2010; Reis 2001; Wurmbrand 2001, among others), complemented by findings from a small number of experimental studies (Bayer et al. 2005; Schmid et al. 2005; Bader and Schmid 2009; Bosch et al. 2022), the factors influencing speakers' preferences for individual syntactic variants are not yet fully understood. Factors which may lead speakers or writers to choose one variant over another include lexical properties of individual verbs or verb classes, the frequency of occurrence of individual linearization patterns, and processing-related factors.

The current study investigates how the frequency distributions of different word-order patterns relate to these patterns' perceived acceptability, and whether German speakers' previously reported preference for a coherent construal of intraposed infinitives can be attributed to processing-related factors. Additionally, our study contributes to the discussion of whether subject raising and control verbs can be divided into clearly distinct subgroups according to their word-order behaviour, as has been proposed within formal linguistic theory (e.g., Reis 2001; Wurmbrand 2001; Haider 2010).

Following a brief review of the theoretical and experimental literature, we will report the results from (i) a synchronic corpus study of present-day German seeking to establish the frequency distributions of different word-order patterns, (ii) two acceptability judgement experiments to assess the degree to which German speakers accept different word-order variants and coherent versus incoherent construals, and (iii) a self-paced reading experiment examining the processing of intraposed infinitival complements.

2 Theoretical and empirical approaches to German infinitival complementation

2.1 Usage-based approaches

One possible approach to word-order variation is represented by usage-based frameworks, which argue that frequency of occurrence determines language performance and representation (e.g., Bybee 2006; Bybee and Beckner 2010). Bybee (2010), for example, regards properties of languages and their grammars as emergent from the frequency of input and the way the human language faculty deals with the experience of language use (e.g., Ellis and Larsen-Freeman 2006). With respect to word-order variation, usage-based approaches to language performance would expect language users’ acceptability of alternative syntactic variants to be determined their frequency distributions across the language.

Newmeyer (2003), in contrast, argues that language performance and representation are not predictable from usage-based facts about language. Although he acknowledges the value of statistical information for the investigation of language variation, use, and change, he points out that “it is a long way from there [corpus-based frequency analyses—authors’ note] to the conclusion that corpus-derived statistical information is relevant to the nature of the grammar of any individual speaker” (695). Accordingly, Newmeyer (2003) believes that infrequent or generally dispreferred structures may be part of speakers’ grammatical representations as ‘latent’ structures and may thus be judged as acceptable even though they are rarely used or attested in corpora.

Many previous studies have combined synchronic corpus analyses with experimental methods such as acceptability judgements to test predictions derived from usage-based accounts (e.g., Gries et al. 2005; Schmid et al. 2005; Kempen and Harbusch 2008; Baroni et al. 2009; Adli 2011; Radford et al. 2012; Ford and Bresnan 2013). Interestingly, several of these studies observed imperfect correlations between corpus probabilities and rating patterns: Highly acceptable alternatives predominate corpus frequencies, while less acceptable variants might not be attested in corpus data at all (Featherston 2005; Kempen and Harbusch 2005; Bader and Häussler 2010). Also, patterns which are rarely used may nevertheless elicit relatively high acceptability scores, a finding that is in line with Newmeyer’s (2003) proposal.

In order to assess frequency distributions and speakers’ acceptance of different word-order variants in German infinitival complementation, Bayer et al. (2005) conducted a multi-methodological study that included a corpus search and an acceptability rating experiment. Their search of the Mannheim COSMAS corpus of newspaper texts and literature found extraposed infinitives to be much more frequent than intraposed ones: Less than 4% of infinitival complements occurred in an intraposed position, and less than 20% of all verbs in the sample of control verbs they examined occurred with an intraposed infinitive. These findings were taken to indicate that intraposition represents a marked (although grammatical) option. In contrast, the acceptability rating data revealed that subject control verbs with infinitival complements functioning as direct objects elicited fairly favourable ratings for intraposition (a mean rating of 2.1 on a 5-point Likert scale, with 1 meaning ‘acceptable’ and 5 meaning ‘unacceptable’). This imperfect correlation between corpus data and experimental findings suggests that frequency of occurrence alone cannot predict acceptability, particularly with respect to infrequent word-order patterns. However, note that Bayer et al. (2005) only considered two of the three possible word-order variants, and their frequency data was not systematically related to their judgement data on a per-verb basis.

Also combining corpus search and experimentation, Bosch et al. (2022) additionally included the third construction (1c) in their investigations. Focusing on a small set of control verbs, we compared the frequency of occurrence of the three linearization patterns with German speakers' scalar acceptability judgements for each individual verb. Again, an imperfect alignment was observed between frequency counts and acceptability ratings, with high-frequency patterns matching corpus probabilities but low-frequency patterns sometimes being judged as more acceptable than their statistical distribution would have suggested. Both Bayer et al. (2005) and Bosch et al. (2022) found that the extraposition pattern was the most frequently attested one and also elicited the most favourable acceptability ratings.

The present study builds on and extends the above investigations by carrying out a large-scale corpus search covering different corpora and genres, by additionally including a set of subject raising verbs to examine whether these do indeed differ categorically from subject control verbs with regard to their word-order behaviour and coherence properties and by systematically testing these verbs' word-order preferences across all three variants (extraposition, intraposition, and third construction) for their predictability from frequency distributions.

2.2 Processing-based approaches

Constraints imposed by the language processing system are another factor to take into account when examining the question of why certain syntactic variants are preferred over others. These may include, for example, the avoidance of clausal centre-embedding, a preference for the structurally simplest analysis, ambiguity avoidance, and the minimization of dependency length. From a processing perspective, extraposition should be favoured over other word-order variants as it avoids centre-embedding and does not give rise to any local or global structural ambiguities (compare, e.g., Bader and Schmid 2009). It also helps to keep the distance between the matrix verb and its subject as short as possible. In a spoken production experiment, Bosch et al. (2022) found that extraposition was strongly preferred over all other word-order variants for the control verbs under investigation, corresponding to this pattern's relative corpus frequency.

Regarding intraposition, Bayer et al. (2005) report experimental evidence showing that incoherent construals are dispreferred by speakers of German, a finding that can be accounted for by structural economy constraints: Coherent construals involve less complex syntactic representations than incoherent construals and will thus be an economy-driven parser's preferred analysis choice. Bader and Schmid (2009) note that from the perspective of left-to-right incremental processing, coherently analysed intraposed infinitives avoid two well-known processing problems: the increased processing cost typically associated with clausal centre-embedding, and the need for structural reanalysis due to a clausal boundary that was initially overlooked (1461).

Bader and Schmid (2009) report results from speeded grammaticality judgements showing that intraposed incoherent infinitives are indeed difficult to process in comparison to extraposed infinitives, a finding consistent with what the authors refer to as the Clause-Union Preference Hypothesis. At the same time, however, their results provide evidence in favour of another hypothesis: According to Bader and Schmid's (2009) Verb-Cluster Complexity Hypothesis (VCCH) a coherent analysis incurs more processing cost than an incoherent one at the point where the verb cluster is formed, reflecting the cost associated with argument structure unification. Considering these findings, it would appear that neither a coherent nor an incoherent analysis of intraposed infinitives represents an optimal clausal structure from a processing economy perspective. Both Bayer et al. (2005) and Bader and Schmid (2009) raise the question of why intraposed infinitives exist at all in German, given that extraposition is available as an alternative. Bader and Schmid (2009) suggest that the verb-final property of German is the reason that intraposed infinitives are (still) available.

Note that from a left-to-right processing perspective the VCCH predicts a processing disadvantage for coherently versus incoherently construed infinitives when the verb cluster is encountered but not during earlier parts of the infinitival complement. Collecting end-of-sentence acceptability judgements did not allow Bader and Schmid (2009) to test this prediction directly, however. Their conclusion regarding the VCCH moreover rests on grammaticality judgements of long passive constructions, whose grammatical status is at best controversial. Our Experiment 3 tests the above prediction more directly by using a word-by-word self-paced reading task and different types of stimulus sentences.

Neither of the above studies investigated the acceptability or processing of third constructions, a word-order pattern which involves neither verb cluster formation nor clausal centre-embedding, but which may nevertheless allow for a coherent construal in that no clause boundary needs to be postulated (Wöllstein-Leisten 2001). Comparing whole-sentence reading comprehension times for all three word-order patterns, Bosch et al. (2022) found that intraposed and extraposed infinitival complements were about equally difficult to comprehend and significantly easier to process than third constructions. We argued that third constructions are hard to process because this linearization pattern not only breaks up the dependency between the matrix subject and matrix verb but also the infinitival complement itself, and may additionally create a temporary ambiguity that requires reanalysis. However, as the intraposed infinitives in Bosch et al. (2022) were not disambiguated towards a coherent or incoherent construal, their results do not tell us anything about the relative processing difficulty of coherent versus incoherent infinitives.

2.3 Formal linguistic approaches

The syntactic type of infinitival complement has been argued to limit word-order variation in present-day German. As illustrated by (3a) versus (3b) below, extraposition is assumed to be available only for verbs such as beabsichtigen ‘intend’ that select CP complements (resulting in a biclausal incoherent construal), whilst it is precluded for verbs such as scheinen ‘seem’ that require monoclausal coherent construals.

(3)	a.	weil	Fred	niemals	beabsichtigte	[CP	PRO	seinen	Goldfisch
		because	Fred	never	intented			his	goldfish
		auszusetzen]
		to.release
		‘because Fred never intended to release his goldfish’
	b.	*weil	Selma	gestern	schien	[einen	Essay	zu	schreiben]
		because	Selma	yesterday	seemed	an	essay	to	write
		‘because Selma seemed to write an essay yesterday’

Intraposition, on the other hand, allows for the positioning of the infinitival complement in the matrix clause's middle field. If no other material intervenes between the infinitival and matrix verb, intraposed infinitival complements are often ambiguous with respect to coherence, particularly with control verbs, such as versuchen ‘try’, which allow both monoclausal coherent and biclausal incoherent construals, as illustrated in (2) above.

However, intraposition may also yield patterns which unambiguously signal either coherent or incoherent configurations. In (4a) below, for instance, non-verbal material (the adverb niemals ‘never’) intervenes between the infinitive and the matrix verb, thus precluding the formation of a verbal complex and triggering an incoherent structure. The scrambling out of infinitival material into the matrix domain, on the other hand, unambiguously signals a monoclausal structure. A classical diagnostic pattern for this is pronoun fronting as in (4b).

(4)	a.	weil	Fred	[CP	PRO	seinen	Goldfisch	auszusetzen]	niemals
		because	Fred			his	goldfish	to.release	never
		beabsichtigte
		intended
		‘because Fred never intended to release his goldfish’
	b.	weil	ihni	das	Leben	ti	zu	überholen	drohte
		because	him.ACC	the	life		to	pass.by	threaten
		‘because life threatened to pass him by’

Other diagnostics for the distinction of monoclausal and biclausal structures include their behaviour with respect to the scope of negation, adverbial modifiers and binding; see Haider (2010, 311) for an exhaustive list of diagnostics with pertinent examples. The status of the third construction pattern is more controversial and has not been studied very extensively with the exception of Wöllstein-Leisten (2001) (for arguments of third constructions as coherent structures see also Reis (2001, 140)).

Traditionally, it has been argued that the distribution of (in-)coherent infinitival constructions is lexically driven. A well-defined group of verbs obligatorily build verbal complexes with the non-finite verb they select, thus requiring the infinitival complement to be intraposed. This class includes, for example, raising verbs such as scheinen ‘seem’ and pflegen ‘be in the habit of’, which do not select an external argument and thus in general do not impose any semantic restrictions on their grammatical subject. Control verbs, in contrast, assign thematic roles to both their external and internal arguments and have been argued to select CP complements, which may either be extraposed (5a) or intraposed (5b).

(5)	a.	weil	Selma	ihm	vorwirft,	[ihren	Goldfisch	vernachlässigt
		because	Selma	him.DAT	reproaches	her	goldfish	neglected
		zu	haben]
		to	have
		‘because Selma reproaches him for having neglected her goldfish’
	b.	weil	Selma	ihm	[ihren	Goldfisch	vernachlässigt	zu
		because	Selma	him.DAT	her	goldfish	neglected	to
		haben]	vorwirft
		have	reproaches
		‘because Selma reproaches him for having neglected her goldfish’

A subset of control verbs select both clausal and non-clausal infinitival complements, with no construction-specific meaning difference between these two complement types (Haider 2011, 278). Two prominent examples include the subject control verbs versuchen ‘try’ and beschließen ‘decide’. These verbs are traditionally referred to as optionally coherent verbs (Haider 1994) and not only allow for both intraposition (6a) and extraposition (6b) but also for the third construction pattern (6c).²

(6)	a.	weil	Fred	diesen	Boxkampf	zu	gewinnen	versucht
		because	Fred	this	fight	to	win	tries
	b.	weil	Fred	versucht,	diesen	Boxkampf	zu	gewinnen
		because	Fred	tries	this	fight	to	win
	c.	weil	Fred	diesen	Boxkampf	versucht	zu	gewinnen
		because	Fred	this	fight	tries	to	win
		‘because Fred Tries to win this boxing match'

An alternative approach to the classification of infinitive-embedding verbs has been put forward by Wurmbrand (2001) who uses the notion of graded (in-) coherence, distinguishing between four instead of three types of verbs selecting infinitival complements. In her view, control verbs that are transparent for coherence phenomena are not of the same type as raising verbs. Instead she distinguishes between lexical and functional predicates, yielding what she calls lexical restructuring and functional restructuring infinitives, respectively (see Table 1 for an overview of verbal classifications according to Haider (2010) and Wurmbrand (2001)). In functional restructuring contexts, the matrix verb instantiates a functional head (AUX) and the non-finite verb is the main verb, whereas in lexical restructuring contexts the matrix verb is the head of a VP which in turn embeds a non-finite VP complement. Wurmbrand’s (2001) approach differs from Haider's (1994) particularly in the way it accounts for the distribution of third constructions. According to Wurmbrand (2001), non-focus scrambling constructions such as the third construction are unavailable for functional restructuring verbs such as scheinen and pflegen but are available for lexical restructuring verbs such as versuchen. However, while lexical restructuring verbs allow for the third construction as well as pronoun fronting, others prohibit the third construction, constituting the class of reduced non-restructuring verbs, among which she includes, e.g., beschließen. In Wurmbrand’s (2001) view genuinely non-restructuring verbs are those that block any type of coherent structures including pronoun fronting. Typical examples for non-restructuring verbs are factive verbs such as bedauern ‘regret’ and vorwerfen ‘reproach’ (see Table 1).³

Table 1 Comparison of verbal classification according to Haider (2010) and Wurmbrand (2001)

Another attempt to conceptualize word-order variation in German infinitival complementation as a graded phenomenon was put forward by Grosse (2005). Grosse claims that individual members of infinitive-embedding verbs that are compatible with both coherent and incoherent structures exhibit varying degrees of coherence, and thus allow different word-order variants to varying degrees.

To shed more light on the question of how the set of control verbs allowing monoclausal construals is defined, Schmid et al. (2005) collected judgement data to examine the effects of structural factors such as argument structure and control properties on the acceptability of various types of coherent and non-coherent infinitival constructions for subject and object control verbs. Schmid et al.'s (2005) results revealed substantial verb-specific variation patterns with regard to coherence, with the verb versuchen scoring highest on the authors' coherence scale. Schmid et al. (2005) do not discuss these verb-specific findings any further, however.

None of the above-mentioned experimental studies have considered raising verbs, which in present-day German are assumed to behave homogeneously in permitting coherent intraposed infinitival complements only. The present study focuses on a small set of subject raising and subject control verbs and examines these verbs' word-order behaviour systematically across both corpus and experimental data, including their coherence preferences.

3 The present study

Our study has two main objectives: (i) to test hypotheses derived from usage-based, processing-based and formal linguistic approaches to word-order variation, and (ii) to test the hypothesis that processing economy constraints should disfavour an incoherent construal of intraposed infinitives. Usage-based approaches predict that speakers' preferences for certain word-order variants should be determined by frequency distributions, whilst processing-based approaches predict that certain word-order options or analysis choices should be preferred due to processing economy constraints, other things being equal. Formal approaches make predictions as to what structural patterns a given type of verb should or should not permit.

As a first step, we take a corpus-based perspective on German infinitival complementation and examine how different word-order variants in combination with particular infinitive-embedding verbs are distributed across corpora of present-day German. Our corpus analysis serves to establish frequency distributions of the word-order patterns under investigation and thus provides a baseline for assessing the predictions made by usage-based approaches to language performance and representation.

Secondly, we take an experimental perspective on word-order behaviour in German infinitival complementation by carrying out two untimed acceptability rating tasks and an online self-paced reading task. The acceptability data is then evaluated against our corpus findings, to assess whether and to what extent corpus frequency distributions of the different linearization patterns can predict their acceptability. We also examine the acceptability of monoclausal coherent versus biclausal incoherent construals of intraposed infinitives, to establish a baseline as to whether incoherent word-order patterns are less readily accepted than coherent ones, and to see to what extent the sets of raising and control verbs under investigation behave homogeneously with respect to these linearization patterns. Our reading-time task then builds on that baseline and additionally probes the real-time processing of coherent versus incoherent intraposed infinitival complements of subject control verbs.

The current study focuses on the two prototypical German raising verbs scheinen ‘seem’ and pflegen ‘be in the habit of’, and on the four subject control verbs versuchen ‘try’, beschließen ‘decide’, ankündigen ‘announce’, and bedauern ‘regret’. The latter four verbs share the same control properties and argument structure and differ only in their semantic interpretation: While according to Wurmbrand (2001) ankündigen and bedauern are classified as factive/propositional verbs, versuchen and beschließen represent irrealis verbs.

3.1 The corpus study

In order to establish the frequency distributions of the different word-order variants across our selected verbs, we performed a systematic analysis of the following corpora: a selection of newspaper articles from DeReKo, the German Reference Corpus⁴ (Kupietz et al. 2010), a German Twitter corpus⁵ (Scheffler 2014) and a subcorpus of the DWDS-Kernkorpus 21 including fiction and scientific texts⁶ (Geyken 2007). By including different genres, and in particular a source that, although it contains both curated and spontaneous content, shows an “oral-like” style (Scheffler 2014, 2288), i.e. the Twitter corpus, we aimed at expanding the range of different corpus types in order to obtain a more representative input sample.

The newspaper corpus included articles from the year 2016 including two national newspapers, Die Zeit and Süddeutsche Zeitung, and consisted of 31,236,121 tokens. The Twitter corpus comprised 360,331,744 tokens and the DWDS 4,641,197 tokens. Hence, in total, our corpus comprised 396,209,062 tokens.⁷

3.1.1 Methods

For each corpus we extracted infinitival complements governed by the verbs scheinen, pflegen, versuchen, beschließen, ankündigen, and bedauern in sentences with a closed sentence bracket, that is, either verb final subordinate clauses (7) or main clauses with a complex predicate (8).

(7)	…daß	irgendwer	versucht,	sie	zu	erreichen
	…that	someone	tries	her	to	reach
	‘…that someone tries to reach her’
						(DWDS: Beyer, Spione, 93)

(8)	Er	würde	aber	versuchen,	alles	zu	verzögern
	he	would	however	try	everything	to	delay
	‘However he would try to delay everything’
					(DWDS: Kopetzky, GrandTour, 460)

The hits were then annotated as to the position of the infinitive as follows:

(9)	a.	extraposition
		Wenn	sie	beschließen,	[die	Dinge	selbst
		if	they	decide	the	things	themselves
		in	die	Hand	zu	nehmen]
		in	the	hand	to	take
		‘If they decide to take things into their own hands’
					(DeReKo: Z16/FEB.00275 Zeit, 11.02.2016, 58)
	b.	intraposition
		Wie	Firmen	[diese	Gemengelage	zu	nutzen]
		how	companies	this	hotchpotch	to	use
		versuchen
		try
		‘How companies try to use this hotchpotch’
					(DeReKo: U16/FEB.00930 SZ, 06.02.2016, 57
	c.	third	construction
		Dass	die	gegnerischen	Fans	[dich]	versuchen
		That	the	opposing	fans	you.ACC	try
		[abzulenken]
		to.distract
		‘That the opposing fans try to distract you’
					(DeReKo: Z16/OKT.00119 Zeit, 06.10.2016, 58)
	d.	intraposition	coherent
		wie	[ihn]	Ernst	Nolte	1963	[zu
		as	it.ACC	Ernst	Nolte	1963	to
		etablieren]	versuchte
		establish	tried
		‘As Ernst Nolte in 1963 tried to establish it’
					(DeReKo: Z16/SEP.00736 Zeit, 29.09.2016, 53)
	e.	intraposition	Incoherent
		Er	war	seit	langer	Zeit	der
		he	was	since	long.DAT	time	the
		erste	Mensch,	den	belogen	zu
		first	person	whom	lied.to	to
		haben	sie	wirklich	bedauerte.
		have	she	really	regretted
		‘He was he first person in a long time whom she really regretted to have lied to’
					(DWDS: Kopetzky, GrandTour, 500)

In the case of intraposition, if any evidence for either a coherent or an incoherent structure was found, this was further specified. For example, in case pronoun fronting was attested, intraposition was further annotated as coherent as in (9d). If the intraposed infinitive was not directly adjacent to the matrix verb but interrupted by non-verbal material as in (9e), intraposition was coded as incoherent. Example (9e) represents a case of a ‘pied-piping’ construction in which the infinitival complement directly follows the relative pronoun den ‘whom’ and is located at the left periphery of the relative clause in front of the matrix subject sie ‘she’.

3.1.2 Results

The six chosen verbs differed in their absolute frequencies of occurrence in combination with infinitival complements. While the raising verb scheinen and the control verb versuchen occurred most frequently with an infinitival construction, the other four verbs were attested considerably less often.⁸ However, as expected, opposite distributions of word-order patterns were found between raising and control verbs, such that the two raising verbs scheinen and pflegen occurred with an intraposed infinitive in 100% of the cases, while for our control verbs, extraposition was the most frequent variant (see Table 2). A Fisher’s Exact Test on the absolute counts confirmed that verb type and distribution of attested word-order patterns were significantly associated (p < .001). No significant differences were found between the two raising verbs scheinen and pflegen (p = 1, using a Fisher’s Exact Test).

Table 2 Corpus frequencies for raising and control verbs

While neither scheinen nor pflegen were attested with an incoherent biclausal structure, three examples of pronoun fronting, exemplified in (10) and indicating a monoclausal coherent structure, were found with the verb pflegen.

(10)	wie	[es]	die	gastronomische	Literatur	bis	zur
	as	it.ACC	the.NOM	gastronomical	literature	until	to.theGEN
	Übersättigung	ihrer	Leser	[zu	tun]	pflegt.
	oversaturation	itsGEN	readers	to	do	is.in.the.habit.of
	‘as the gastronomical literature usually does, until its readers are over-saturated’
				(DeReKo: Z16/MAR.00507 Zeit, 17.03.2016, 51)

Across all four control verbs extraposition was clearly the most frequent word-order pattern, ranging between 74.4 and 100% of occurrences. However, while for the three verbs ankündigen, bedauern and beschließen intraposition and third construction are rarely or not at all attested (see Table 2), most variability is observed with the verb versuchen, such that all three word-order patterns are attested: extraposition is dominant with 74.4%, intraposition follows with 23.3% and the third construction is attested with a rate of 2.3%. Pairwise comparisons between the control verbs revealed that versuchen differs significantly from the other verbs (vs. bedauern: p < .01, vs. ankündigen: p < .001, vs. beschließen: p < .001, using a Fisher’s Exact Test).

Of the intraposed infinitives with the verb versuchen, eight instantiations show evidence for a coherent construal. Example (9e) above is the only attested case of a pied-piped intraposed infinitive, and thus the only instance of unambiguously incoherent intraposition. The little variability within individual verbs also suggests that genre did not play a major role in determining word-order preferences, with one exception: The verb versuchen not only showed a more variable word-order behaviour than the other control verbs but also differences between the sub-corpora.

The third construction, exemplified in (9c) above, is almost exclusively found in the Twitter corpus (N = 59/65), compared to only six instances in the other sources (see Table 3). This result is not surprising, however, since the third construction is sometimes considered as a marked option in the written language (Wöllstein-Leisten 2001). Bosch et al. (2022) found that this pattern is indeed more frequent in spoken than in written corpora, and that Twitter data patterns with spoken data in this respect. In addition, Table 3 shows that the relative frequency of intraposed infinitives also differs among sub-corpora, with Twitter showing the smallest proportion of intraposed infinitives (12%) compared to the other written sources (30% in newspapers and 40% in fictional and scientific texts). Whilst the near-absence of the third construction from written corpora might at least partly reflect prescriptive norms, such norms cannot explain why intraposition is very rare as well, and especially rare in our Twitter corpus, thus suggesting that additional factors play a role in determining word-order preferences.

Table 3 Distribution of word-order patterns attested with the verb versuchen per individual corpus

3.1.3 Discussion

The frequency distributions of the different word-order variants confirm Bayer et al.’s (2005) and Bosch et al.'s (2022) observation that intraposed infinitival complements only occur infrequently with subject control verbs, indicating that intraposition represents a marked word-order option. Differences among the investigated sub-corpora further suggest that this is even more so in spoken-like styles. Additionally, we conclude that word-order behaviour is in part lexically restricted, as witnessed by the observed difference between raising and control verbs: While raising verbs consistently occur with intraposed infinitives only, subject control verbs do not appear as a homogenous group, with versuchen exhibiting more variability than the other three control verbs.

Given that the intraposition was rarely attested for the control verbs under investigation, and the examples we found were almost always ambiguous between coherent and incoherent construals, we proceeded to test German speakers' coherence preferences experimentally. Considering that third construction patterns were hardly attested at all, collecting acceptability judgement should help us find out to what extent the third construction is considered a permissible word-order option by present-day German speakers.

3.2 Acceptability judgements

We carried out two complementary acceptability judgement experiments on the word-order behaviour of German infinitive-embedding verbs. These will allow us to test predictions derived from usage-based frameworks (e.g., Bybee 2006; Bybee and Beckner 2010) by examining the extent to which acceptability rating patterns align with corpus frequency distributions. We will also consider to what extent lexical properties of individual verbs or verb classes can account for participants' rating patterns. Experiment 1 compares extraposition, third construction, and coherent versus incoherent intraposition structures, and Experiment 2, intended as a partial replication of the former, examines both ambiguous and non-ambiguous intraposition patterns so as to allow for a better comparison of the judgement and corpus data.

Our corpus counts predict that we should find clear differences in the judgement patterns for raising compared to control verbs. Since the raising verbs scheinen and pflegen exclusively occurred with intraposed infinitival complements in our corpus data, they are expected to elicit high acceptability ratings for intraposed structures but poor ratings for all other word-order variants. We would also expect the raising verbs scheinen and pflegen to pattern alike from the point of view of formal linguistic approaches: Raising verbs should elicit high acceptability ratings for monoclausal intraposition structures and poor ratings for all other word-order variants.

In contrast, given that subject control verbs occurred most frequently with extraposed infinitival complements, they are expected to yield higher acceptability ratings for this condition than for all other word-order variants. The third construction pattern, which was attested only for the verb versuchen, is expected to elicit the poorest ratings overall. A similarly graded acceptability pattern would be expected from a processing perspective, with the added prediction that coherent or ambiguous intraposition should be favoured over incoherent intraposition.

With respect to verb-specific variation patterns, corpus frequency distributions predict a two-way distinction for control verbs: While the subject control verbs ankündigen, bedauern, and beschließen occurred hardly at all with intraposed infinitival complements or third constructions in our corpus data, the verb versuchen sits apart as it appeared with all possible word-order variants. According to our corpus data, all subject control verbs other than versuchen are expected to elicit poor ratings for both intraposition and third constructions.

According to Haider's (1993, 2010) binary classification system of control verbs in German, the obligatorily biclausal verbs ankündigen and bedauern should pattern together in eliciting better ratings for extraposed and incoherent intraposed relative to coherent intraposition. The optionally monoclausal verbs versuchen and beschließen, in contrast, should exhibit the highest degree of variability across both monoclausal and biclausal construals, with all word-order patterns being deemed reasonably acceptable. Wurmbrand's (2001) three-way distinction of German subject control verbs, on the other hand, predicts that the lexical restructuring verb versuchen should pattern differently from the reduced non-restructuring verb beschließen, with versuchen eliciting better ratings than beschließen for both monoclausal coherent infinitives and for the third construction.

3.2.1 Experiment 1

3.2.1.1 Participants

Experiment 1 included 39 participants (3 male, 36 female) aged between 18 and 44 years (mean: 25.9 years). All participants were native speakers of German living in Germany at the time of testing and did not report any language-related or other behavioural or neurological disorders.⁹ The majority of them were students at different universities across the country who were recruited online, either via the SONA participant pool provided by the University of Potsdam or via e-mail contact. Participants were offered a small fee of 4€ as a reimbursement.

3.2.1.2 Materials

The experimental materials were constructed using the six chosen matrix verbs, which were combined with infinitival complements exhibiting different word-order patterns so that each verb appeared in four experimental conditions as shown in (11a–d).¹⁰ To make sure that infinitival complements occurred in a verb-final (SOV) structure, the critical verbs and their infinitival complements were presented in subordinate clauses. Examples (11) and (12) illustrate item sets for Experiment 1 with the infinitive-embedding verbs versuchen and scheinen.

(11)	a.	extraposition
		Julia	sagt,	dass	Fred	versucht	[ihn	zu	streicheln].
		Julia	says	that	Fred	tries	him	to	pet
	b.	third construction
		Julia	sagt,	dass	Fred	[ihn]	versucht	[zu	streicheln].
		Julia	says	that	Fred	him	tries	to	pet
	c.	intraposition coherent
		Julia	sagt,	dass	ihn	Fred	[zu	streicheln]	versucht.
		Julia	says	that	him	Fred	to	pet	tries
	d.	intraposition incoherent
		Julia	sagt,	dass	Fred	[ihn	zu	streicheln]	oft	versucht.
		Julia	says	that	Fred	him	to	pet	often	tries
		‘Julia says that Fred tries to pet him (often).’
(12)	a.	extraposition
		Julia	sagt,	dass	Fred	scheint	[ihn	zu	streicheln].
		Julia	says	that	Fred	seems	him	to	pet
	b.	third construction
		Julia	sagt,	dass	Fred	[ihn]	scheint	[zu	streicheln].
		Julia	says	that	Fred	him	seems	to	pet
	c.	intraposition coherent
		Julia	sagt,	dass	ihn	Fred	[zu	streicheln]	scheint.
		Julia	says	that	him	Fred	to	pet	seems
	d.	intraposition incoherent
		Julia	sagt,	dass	Fred	[ihn	zu	streicheln]	oft	scheint.
		Julia	says	that	Fred	him	to	pet	often	seems
		‘Julia says that Fred seems to pet him (often).’

For all experimental sentences, a picture illustrating the entity referred to by the personal pronoun ihn in the subordinate clause was presented above the actual test sentence so as to provide an appropriate antecedent for the pronoun (see (13)).

In the extraposition condition (11a and 12a), the infinitival complement follows the matrix verb and is located at the right periphery of the sentence. The third construction (11b and 12b) represents a discontinuous construction, in which the accusative object of the infinitive ihn appears left-adjacent to the matrix verb while the infinitive zu streicheln appears in an extraposed position to its right. Conditions (11c, 12c) and (11d, 12d) both contain intraposed infinitives. Condition (11c, 12c) represents an unambiguously coherent structure as the object pronoun (ihn) has been scrambled out of the infinitival complement in front of the subject of the subordinate clause (here, ‘Fred’). In contrast, condition (11d, 12d) represents an unambiguously incoherent construal, with the infinitival complement projecting its own CP, which is signaled by the lack of adjacency between the infinitive and the subcategorizing verb: The infinitival complement and the matrix verb are separated by an intervening adverb (oft 'often'), which indicates the presence of a clause boundary between infinitive and matrix verb.

Twenty-four critical item sets were created by placing each of the six critical verbs in four different semantic sentence contexts and combining them with the four above-mentioned word-order patterns. Each sentence began with an introductory sentence fragment such as Julia sagt, dass… (‘Julia says that…’) followed by the remainder of the subordinate clause. The subordinate clauses all contained one of the critical matrix verbs combined with an infinitival complement. The infinitival verbs all appeared with a pronominal object in the dative or accusative case. This allowed us to use pronoun fronting as a coherence signal in (11c) and (12c) whilst ensuring that all four experimental conditions were close minimal pairs. Full lists of our experimental materials are available at the Center for Open Science Framework website (https://osf.io/a7rqd/).

In addition, Experiment 1 included 40 filler sentences which did not contain any infinitival complements and differed from our experimental items both in their lexical material and syntactic structure. We included both fully acceptable fillers such as Klaus glaubt, dass die Erklärungen des Lehrers nicht richtig sind. (‘Klaus believes that the teacher's explanations are incorrect.’) and fully ungrammatical fillers such as *Lena prüft, was verursacht beim Unfall für Schäden sie hat. (lit. ‘Lena checks what caused in the accident for damage she has.’). This served to mask the ultimate purpose of the experiment and allowed us to verify whether participants read the stimulus items attentively, and to encourage them to make use of the full range of our rating scale.

All experimental sentences were pre-tested in a plausibility rating task in order to make sure that all of them were easily readable and interpretable. In this pretest, 30 German native participants were recruited from among the student communities of the University of Potsdam and provided with an online questionnaire via Google Forms which asked them to rate the plausibility of sentences with intraposed (raising verbs) and extraposed (control verbs) infinitival complements on a scale from 1 (= ‘highly plausible’) to 5 (= ‘not plausible at all’). To make sure that participants were not presented with highly plausible sentences only, we created a set of 24 implausible sentences (e.g., Vera glaubt, dass Sven die Nudeln Gedichte sagen hört.—‘Vera believes that Sven hears the noodles recite poems.’) and mixed them with the critical and filler items from Experiment 1. Mean plausibility ratings per item were calculated, and any critical items or fillers, which elicited a low mean plausibility rating (> 2), were adapted for Experiment 1 so as to make them more plausible. This was the case for two experimental items. Finally, four different presentation lists were created using a Latin square design. Each list contained 24 critical items and 40 filler sentences presented in pseudo-randomized order, with not more than three critical items appearing in a row. This design made sure that each list contained four experimental sentences per critical verb, each in a different context and word-order variant. Hence, in order to avoid repetition effects, each participant saw each matrix verb in all experimental conditions but in different sentence contexts which were not repeated within a single list.

3.2.1.3 Procedure

The questionnaire was implemented via Google Forms and administered via the world-wide web. Upon opening the link to the experiment, participants were first asked to answer some biographical questions and to provide their consent to participating in our study. After reading the instructions, participants received three practise items to familiarize themselves with the experimental task. Then the main experiment started with the presentation of the first trial.

Participants were instructed to read each stimulus sentence carefully and then rate its acceptability on a 5-point Likert scale. Sentences were presented individually in Arial 20pt in black colour against a white background, with the judgement scale appearing right underneath each test sentence. The scale ranged from 1 labelled as “vollkommen akzeptabel” (‘totally acceptable’) to 5 labelled as “völlig inakzeptabel” (‘totally unacceptable’).¹¹ Points 2, 3, and 4 were not labelled on every single scale, but were introduced as 2 meaning “eher akzeptabel” (‘rather acceptable’), as 3 meaning “nicht sehr akzeptabel” (‘not very acceptable’), and as 4 meaning “eher inakzeptabel” (‘rather unacceptable’). All experimental trials were presented as illustrated in (13).

Here the picture of the dog provides a referent for the personal pronoun ihn, which occurs before the subordinate clause subject Fred. This visual illustration should make our stimulus sentences are more easily interpretable and more plausible for participants.

Participants could take as much time as they needed for each stimulus sentence. Once they had provided their acceptability judgement for a given sentence, the next trial started automatically. On average, Experiment 1 took approximately 25 min to complete.

3.2.1.4 Data analysis

The data from two participants were excluded from statistical analysis since they uniformly rated all experimental and filler items either with best (= 1) or worst (= 5) scores. For the remaining rating data we employed cumulative link mixed models (CLMM) for ordinal regression (Christensen 2018) from the ordinal package in R (R Core Team 2017). The factorial structure of our experiment was reflected in the structure of our models, which were held constant across by-participant and by-item random effects. As categorical fixed effect variables, the models included the factors ‘Verb Type’, comparing raising versus control verbs, the factor ‘Condition’, comparing the four word-order patterns under investigation, and for the follow-up analyses the factor ‘Critical Verb’ in order to examine effects of word-order variation across individual infinitive-embedding verbs. Also, interactions between these factors were included. Finally, the categorical factor ‘Experimental List’ and the continuous predictor ‘Trial Position’ (i.e., the rank of items in the task) were also included in the analyses, in order to control for task-related effects and to remove auto-correlation of residuals (Baayen and Milin 2010). Before including the continuous predictor ‘Trial Position’ in the models, it was centred around its mean. For determining the best-fit random slopes structure, we started with the maximal model including random intercepts and slopes for factors and their interactions. For models which failed to converge, we iteratively removed random slopes by participant or by item which explained the least variance (Barr et al. 2013) until the model did not improve any further according to its AIC value (Venables and Ripley 2002).¹² The best-fit model for each analysis is reported in the results section. For the overall model, we employed treatment contrasts to factors with extraposition as reference level for the factor ‘Condition’ and raising as reference level for ‘Verb Type’. For effects within factors we also employed treatment contrasts and the comparisons of interest were obtained by relevelling factors and refitting the model.

3.2.1.5 Results

Overall, raising verbs elicited the highest mean acceptability ratings for infinitival complements in the intraposed coherent condition (2.15), while all other complement types were rated considerably worse (extraposition: 3.41, intraposition incoherent: 4.05, third construction: 3.53). In sharp contrast to this, control verbs yielded the best ratings for extraposition compared to all other word-order variants (mean rating: 1.36). However, intraposed coherent complements, as well as those in the third construction, still elicited fairly acceptable ratings in the intermediate scalar range (intraposition incoherent: 3.51, intraposition coherent: 2.84, third construction: 2.63).

A CLMM including the factors ‘Verb Type’ and ‘Condition’ revealed significant effects of both predictors as well as a significant interaction (Table 4).

Table 4 Summary of the statistical analysis of Experiment 1 (asterisks indicating significant effects at α= 0.05).

Given the interaction, further statistical analyses were conducted on raising and subject control verbs separately by including the factors ‘Condition’ and ‘Critical Verb’. For raising verbs (Fig. 1), these analyses revealed significant effects of the factor ‘Condition’: scheinen and pflegen elicited significantly better ratings for the intraposition coherent condition relative to the ratings of any other word-order pattern (intraposition coherent vs. extraposition: Estimate  =  1.268; SE  =  0.22; z  =  5.640; intraposition coherent vs. intraposition incoherent: Estimate  =  1.926; SE  =  0.24; z  =  8.112; intraposition coherent vs. third construction: Estimate  =  1.436; SE  =  0.23; z  =  6.279). Extraposed infinitival constructions and third constructions did not statistically differ, but showed significantly better ratings than the intraposition incoherent condition (extraposition vs. intraposition incoherent: Estimate  =  0.658; SE  =  0.216; z  =  3.043; third construction vs. intraposition incoherent: Estimate  =  0.486; SE  =  0.32; z  =  2.240).

Fig. 1

Mean acceptability ratings and standard errors per condition for the verbs scheinen and pflegen, Experiment 1

Additionally, we obtained significant interactions of the factors ‘Condition’ and ‘Critical Verb’, and thus reliable verb-specific variation patterns between the two raising verbs: While for the verb scheinen, the intraposition coherent condition yielded significantly better ratings than the other word-order variants (intraposition coherent vs. extraposition: Estimate  =  2.086; SE  =  0.30; z  =  6.830; intraposition coherent vs. intraposition incoherent: Estimate  =  2.343; SE  =  0.32; z  =  7.319; intraposition coherent vs. third construction: Estimate  =  1.808; SE  =  0.29; z  =  6.239), pflegen did not only produce similar ratings for intraposition coherent and extraposed infinitival complements (intraposition coherent vs. extraposition: Estimate  =  0.355; SE  =  0.25; z  =  1.413), but also elicited significantly better ratings for third constructions relative to scheinen (Estimate  =  0.764; SE  =  0.19; z  =  4.087).

Participants' acceptability ratings also showed different patterns across our subject control verbs (Fig. 2). CLMMs including the factors ‘Condition’ and ‘Critical Verb’ firstly revealed significant effects of ‘Condition’. Extraposed infinitives were rated significantly better than all other word-order variants (extraposition vs. intraposition coherent: Estimate  =  1.811; SE  =  0.29; z  =  6.245; extraposition vs. third construction: Estimate  =  1.36; SE  =  0.29; z  =  4.728; p=  .032; extraposition versus intraposition incoherent: Estimate  =  1.96; SE  =  0.29; z=  6.745), whereas intraposition coherent and third constructions did not elicit any statistically different rating patterns (Estimate  =  −0.45; SE  =  0.26; z  =  1.71), but were both judged as more acceptable than intraposition incoherent complements (intraposition coherent vs. intraposition incoherent: Estimate  =  0.571; SE  =  0.143; z  =  3.985; third construction vs. intraposition incoherent: Estimate  =  0.602; SE  =  0.14; z  =  4.218).

Fig. 2

Mean acceptability ratings and standard errors per condition for subject control verbs, Experiment 1

In addition, significant effects of the factor ‘Critical Verb’ and significant interactions between the factors ‘Condition’ and ‘Critical Verb’ were obtained, suggesting modulations of word-order behaviour across different individual control verbs.

Whilst extraposition was rated best across the board, planned comparisons between individual verbs also revealed significant differences: The verb versuchen elicited significantly better ratings for third constructions than all other verbs (beschließen: Estimate  =  5.508; SE  =  0.77; z  =  7.158; bedauern: Estimate  =  5.858; SE  =  0.869; z  =  6.742; ankündigen: Estimate  =  5.294; SE  =  0.81; z  =  6.535). Additionally, the significant interactions between the factors ‘Condition’ and ‘Critical Verb’ for the four subject control verbs also show reliable verb-specific variation patterns with respect to their preferences for monoclausdal versus biclausal construals. In particular, versuchen yielded significantly better ratings for the intraposition coherent condition (relative to intraposition incoherent) than beschließen (Estimate  =  1.001; SE  =  0.48; z  =  2.052), while beschließen, in turn, elicited significantly better ratings for coherent intraposition compared to both bedauern (Estimate  =  0.923; SE  =  0.46; z  =  1.997) and ankündigen (Estimate  =  0.761; SE  =  0.366; z  =  2.080). The latter two verbs did not significantly differ, however.

3.2.1.6 Summary

The raising verb scheinen elicited the highest acceptability ratings for intraposed coherent infinitives, while all other linearization patterns were rated considerably worse. The verb pflegen, however, elicited equally acceptable ratings for intraposed coherent and extraposed infinitival complements. These verb-specific variation patterns for scheinen and pflegen were neither predicted by corpus frequency distributions nor by formal linguistic assumptions about their coherence properties.

Control verbs yielded best ratings for extraposed infinitival complements compared to all other word-order variants. The verb versuchen showed substantial variability across alternative variants, however, eliciting significantly better ratings for third constructions than all other control verbs. This finding corresponds to our corpus frequencies, according to which versuchen was the only one of our four control verbs that occurred with third construction patterns. Recall, however, that versuchen only appeared in combination with third constructions in 2.3% of all cases in our corpus data, which was considerably less compared to its occurrence with intraposed structures (23.3%). Our finding that for versuchen the third construction was rated rather favourably (eliciting a mean rating of 1.72), and better numerically than either intraposition condition, was thus unexpected. These judgement patterns show that frequency of occurrence cannot fully predict the relative acceptability of German infinitival complements. In addition, the graded pattern of verb-specific acceptability for monoclausal relative to biclausal construals indicates substantial variability in coherence preferences across our four subject control verbs.

3.2.2 Experiment 2

3.2.2.1 Participants

Experiment 2 included 56 native speakers of German (34 female, 22 male) aged between 20 and 42 years (mean: 32.6 years). None of these participants had any background in Linguistics, reported any language-related or other behavioural or neurological disorders or were informed about the ultimate purpose of the study before testing. All of them were recruited from among the student and working communities in and around Potsdam and/or Berlin and reported not to be speakers of any particular German dialects.

3.2.2.2 Materials

Our experimental materials were similar to those used in Experiment 1 and were constructed around the same six infinitive-embedding verbs. In order to be able to directly compare our experimental data with the corpus frequency counts for intraposition, which were largely based on examples whose coherence status could not be determined, Experiment 2 included an ambiguous intraposition condition, replacing the third construction condition that was tested in Experiment 1. The four experimental conditions thus included extraposed infinitival complements and three intraposition conditions (ambiguous, coherent, and incoherent), as illustrated by example (14) for the verb versuchen.

(14)

a.

extraposition

Finn

sagt,

dass

der

Junge

versucht

[sich

die

Hände

Finn

says

that

the

boy

tries

himself

the

hands

zu

waschen].

to

wash

b.

intraposition ambiguous

Finn

sagt,

dass

der

Junge

[sich

die

Hände

zu

Finn

says

that

the

boy

himself

the

hands

to

waschen]

versucht.

wash

tries

c.

intraposition coherent

Finn

sagt,

dass

[sich]

der

Junge

[die

Hände

zu

Finn

says

that

himself

the

boy

the

hands

to

waschen]

versucht.

wash

tries

d.

intraposition incoherent

Finn

sagt,

dass

der

Junge

[sich

die

Hände

zu

Finn

says

that

the

boy

himself

the

hands

to

waschen]

vergeblich

versucht

wash

in.vain

tries

‘Finn says that the boy tries to wash his hands (in vain).’

The reflexive pronoun sich was used throughout (instead of a non-reflexive personal pronoun in Experiment 1) so as to allow us to use pronoun fronting as a coherence signal in condition (14c) whilst eliminating the need for adding a context picture. Experimental condition (14b) is structurally ambiguous between a coherent and an incoherent construal. In contrast, condition (14c) represents an unambiguously monoclausal structure as the reflexive pronoun sich has scrambled out of the infinitival complement to the front of the subordinate clause subject (here, der Junge ‘the boy’), whereas condition (14d) represents a biclausal construal indicated by the fact that the infinitival complement and the matrix verb are separated by the intervening adverb vergeblich ‘in vain’.

We created 24 critical item sets by placing each of the six critical verbs in four different sentence contexts. Each sentence began with an introductory clause such as Finn sagt, dass… (‘Finn says that…') followed by the rest of the subordinate clause which included the infinitive.

In addition, 16 filler sentences were added to the item set in order to mask the true purpose of the experiment and to encourage the full use of the judgement scale. Our filler sentences differed from the experimental sentences in their lexical material and syntactic structure, such that they mostly consisted of simpler sentences with conjunctions. We included both fully acceptable filler sentences and ungrammatical fillers such as *Den Pullover zu flicken Gisela nie (lit. ‘The sweater to patch Gisela never.’).

Critical items and fillers were distributed across four presentation lists in a Latin square design, such that each list contained 24 critical items plus 16 filler sentences, presented in pseudo-randomized order with no more than three critical items presented in a row.

3.2.2.3 Procedure

The questionnaire was again administered via the internet using Google Forms and generally followed the same procedure as Experiment 1. As Experiment 2 only included reflexives but no personal pronouns in the critical test items, no context pictures were presented along with the stimuli, however. Experiment 2 took approximately 15 minutes to complete.

3.2.2.4 Data Analysis

An error in the stimulus material in one of the four lists led us to exclude one of the experimental sentences containing the verb bedauern in condition (14c). This affected 14 data points, corresponding to an exclusion rate of 1.89%. The statistical data analysis followed the same procedure as in Experiment 1.

3.2.2.5 Results

Overall, raising verbs elicited the most favourable mean acceptability ratings in the intraposed ambiguous condition (1.96). This highly acceptable rating was closely followed by intraposed coherent infinitivals (2.05), while all other complement types were rated considerably worse (extraposition: 3.19, intraposition incoherent: 4.36). In contrast, subject control verbs yielded best ratings for extraposed infinitival complements (1.63) compared to all other word-order variants. However, ambiguous and coherent intraposed complements both elicited moderately acceptable ratings (intraposition incoherent: 3.75, intraposition coherent: 2.9, intraposition ambiguous: 2.8).

A CLMM including the factors ‘Verb Type’ and ‘Condition’ revealed a significant effect of ‘Condition’ as well as significant interactions of conditions across verb types (Table 5).

Table 5 Summary of the statistical analysis of Experiment 2 (asterisks indicating significant effects at α  =  0.05)

To follow up on these interactions, statistical analyses were conducted on raising and control verbs separately. For raising verbs (Fig. 3), significant effects of the factor ‘Condition’ were obtained: both verbs elicited significantly better ratings for ambiguous and coherent intraposition relative to the intraposition incoherent condition (intraposition ambiguous vs. intraposition incoherent: Estimate  =  −  2.212; SE  =  0.291; z  =  −  8.048; intraposition coherent vs. intraposition incoherent: Estimate  =  −  2.212; SE  =  0.284; z  =  7.764). The ambiguous and coherent intraposition conditions did not significantly differ (Estimate  =  0.130; SE  =  0.26; z  =  0.489). Interestingly, extraposed complements elicited significantly better ratings for raising verbs than incoherent infinitival complements (extraposition vs. intraposition incoherent: Estimate  =  −  1.09; SE  =  0.28; z  =  −  3.886). Additionally, significant interactions of the factors ‘Condition’ and ‘Critical Verb’, and thus reliable verb-specific variation patterns between the two raising verbs were obtained: While for the verb scheinen extraposition was rated significantly worse compared to the intraposed ambiguous and coherent conditions (extraposition vs. intraposition ambiguous: Estimate  =  1.921; SE  =  0.238; z  =  8.068; extraposition vs. intraposition coherent: Estimate  =  1.939; SE  =  0.25; z  =  7.695), for the verb pflegen extraposed infinitival complements were rated as good as the intraposition coherent condition (Estimate  =  0.336; SE  =  0.234; z  =  1.433).

Fig. 3

Mean acceptability ratings and standard errors per condition for the verbs scheinen and pflegen, Experiment 2

Our participants showed different rating patterns for subject control verbs (Fig. 4). CLMM including the factors ‘Condition’ and ‘Critical Verb’ firstly revealed significant effects of ‘Condition’. Extraposed infinitives were rated significantly better than all other word-order variants (extraposition vs. intraposition ambiguous: Estimate  =  1.893; SE  =  0.31; z  =  6.096; extraposition vs. intraposition coherent: Estimate  =  2.315; SE  =  0.311; z  =  7.439; extraposition vs. intraposition incoherent: Estimate  =  2.761; SE  =  0.31; z  =  8.784), while the ambiguous and coherent intraposition conditions did not elicit significantly different ratings (Estimate  =  0.422; SE  =  0.288; z  =  0.231).

Fig. 4

Mean acceptability ratings and standard errors per condition for subject control verbs, Experiment 2

In addition, we found significant effects of the factor ‘Critical Verb’ as well as significant interactions between the factors ‘Condition’ and ‘Critical Verb’: While the verb versuchen elicited significantly better ratings for intraposed coherent infinitives than beschließen (Estimate  =  −  1.306; SE  =  0.281; z  =  4.644), beschließen still showed significantly better rating patterns for this monoclausal construal than ankündigen (Estimate  =  0.822; SE  =  0.26; z  =  3.161) and bedauern (Estimate  =  1.200; SE  =  0.329; z  =  3.647). The latter two verbs did not statistically differ, however.

3.2.2.6 Summary

Similarly to Experiment 1, raising verbs elicited the highest acceptability ratings for intraposed coherent and intraposed ambiguous complement types. However, in contrast to scheinen, the verb pflegen elicited comparable acceptability ratings for extraposed and intraposed coherent infinitives in both experiments. This rather striking difference between the two raising verbs' rating patterns was unexpected from corpus frequency distributions and indicates that only scheinen clearly favours monoclausal construals, whereas pflegen allows much more variability regarding the type of infinitival complement it selects. Control verbs, in contrast, received better ratings for extraposed infinitival complements compared to all other word-order variants. Intraposed ambiguous infinitival complements yielded the same acceptability ratings as intraposed coherent infinitives. Intraposed coherent structures elicited a graded pattern of verb-specific acceptability, with versuchen yielding the best and ankündigen/bedauern the worst ratings of this condition.

3.3 Self-paced Reading Experiment

The results from Experiments 1 and 2 showed that coherent or ambiguous intraposition was generally rated more favourably than incoherent intraposition even for control verbs, which are thought to select CP complements. A previous reading-time experiment on intraposed infinitival complements showed that intraposition serves as a coherence trigger (Bayer et al. 2005), such that incoherent construals elicit more processing cost in speakers of German, even for control verbs which resist clause union. To further test these previous claims to the effect that coherent construals are preferred for processing reasons, we carried out a self-paced reading cum paraphrase judgement task, which—unlike experimental methods that provide offline or end-of-sentence measures only—allows us to chart incremental word-by-word processing as it occurs and to pinpoint the source of potential increases in processing cost for intraposed infinitival complements.

3.3.1 Participants

We tested 55 German native speakers (seven male; mean age 24.64 years, range 18–38 years) from the Potsdam and Berlin area. They were recruited via the University’s participant database and social media contact. All participants reported to have grown up with only German being spoken at home and were not speakers of any non-standard German dialects. All participants had normal or corrected to normal vision and did not report any language-related or other behavioural or neurological disorders. All participants gave their voluntary written consent and received a small fee for their participation.

3.3.2 Materials

In order to investigate the processing of coherent versus incoherent intraposed construals, we focused our investigations on subject control verbs exclusively. In addition to the four control verbs used previously, we added the subject control verbs planen (‘to plan’) and vorschlagen (‘to propose’) to our set of critical verbs for more variety. We created 24 minimal sentence pairs, four for each matrix verb, as illustrated in (15) below.

(15)

a.

Ambiguous

Malte

hat

das

Buch

zweimal

zu

lesen

versucht,

aber

Malte

has

the

book

twice

to

read

tried,

but

ist

dabei

eingeschlafen.

is

thereby

fallen.asleep

‘Malte tried to read the book twice but fell asleep doing so.'

b.

Incoherent

Malte

hat

das

Buch

zu

lesen

zweimal

versucht,

aber

Malte

has

the

book

to

read

twice

tried,

but

ist

dabei

eingeschlafen.

is

thereby

fallen.asleep

‘Malte tried twice to read the book but fell asleep doing so.'

Both versions of our stimulus sentences contained an intraposed infinitival complement and one of the six chosen matrix verbs. Condition (15a) is globally ambiguous in that the infinitival complement can either be analysed as a clausal unit or as being integrated with the matrix clause via head merging with the subcategorising verb (i.e., versucht). The adverb zweimal can take scope either over the infinitive or the matrix verb. In the latter case, the accusative object das Buch must have undergone scrambling, thus indicating a coherent construal (i.e., the absence of a clause boundary). Condition (15b), in contrast, is disambiguated towards an incoherent construal as signalled by the adverbial being interposed between the infinitive and the matrix verb, which indicates the presence of a clause boundary.

In both conditions (15a) and (15b), from the point of view of left-to-right incremental processing, there is no evidence for the existence of a clause boundary at the point at which the DP das Buch is processed. At the infinitival marker zu, it becomes clear that an embedded infinitive must be integrated into the emerging sentence representation, with the (predicted) finite verb in press later on. In condition (15b), coming across the adverbial signals the need for the post-hoc insertion of a clause boundary, the insertion of a null subject and the establishment of a control relation. These processes may increase sentence processing cost in comparison to condition (15a). Ambiguous infinitival complements as in (15a) allow the reader to adopt a monoclausal analysis by forming a verbal complex of the infinitive zu lesen and the matrix verb versucht. A monoclausal coherent construal would render any structural reanalysis unnecessary and provide an elegant way for the parser to stick with its initial analysis. Although our set of matrix verbs also included a subgroup of verbs which are claimed to resist clause union (i.e., ankündigen and bedauern), we do not expect different RT patterns across matrix predicates based on our acceptability rating data, which did not elicit better ratings for intraposed incoherent construals relative to intraposed coherent complements for these verbs, as well as previous research which showed that readers did not revise an initial coherent analysis towards an incoherent construal when encountering a non-clause-union verb (Bayer et al. 2005).

However, Bader and Schmid’s (2009) VCCH predicts that the cost incurred by verb cluster formation should in fact lead to a processing disadvantage of coherent vs. incoherent construals at the point at which the two heads are merged. From an incremental processing perspective, the VCCH thus predicts increased processing cost for condition (15a) compared to condition (15b) when the matrix verb is encountered.

To obtain an indication of participants' interpretation of our stimulus sentences (15a,b) and to help ensure that they actively read for meaning, we combined the online reading task with an end-of-sentence paraphrase judgement task. We created two alternative paraphrases for each sentence. One of these corresponded to the adverbial taking infinitival scope and the other to the adverbial taking matrix scope, as shown in (16a,b).

(16)	a.	Infinitive scope paraphrase
		Malte	hat	versucht	das	Buch	zweimal	zu	lesen.
		Malte	has	tried	the	book	twice	to	read
		‘Malte has tried to read the book twice.’
	b.	Matrix scope paraphrase
		Malte	hat	zweimal	versucht	das	Buch	zu	lesen.
		Malte	has	twice	tried	the	book	to	read
		‘Malte has twice tried to read the book.’

For the ambiguous intraposition condition (15a), both paraphrases should be acceptable in principle. Whilst acceptance of the infinitive scope paraphrase (16a) would be compatible with either a coherent or an incoherent construal, acceptance of the matrix scope paraphrase (16b) would indicate a monoclausal coherent construal of the preceding sentence as this reading should be blocked if the infinitival complement is analysed as a CP. For the incoherent intraposition condition (15b), only the matrix scope paraphrase (16b) is appropriate as the adverbial directly precedes the matrix verb and thus should not be able to take infinitival scope.

The experimental sentences were distributed across four presentation lists in a Latin square design, pseudo-randomized and mixed with 36 filler sentences, resulting in 60 sentences per list, with not more than three critical items in a row. A subset of fillers (n  =  12) resembled our experimental items in that they also included infinitival complements (with the reflexive pronoun sich), and the corresponding paraphrases rephrased the infinitival clause. Additional fillers represented different word-order variation phenomena. Half of the filler sentences were followed by appropriate, the other half by inappropriate paraphrases.

3.3.3 Procedure

The experiment was designed as a web-based study implemented via the experimental platform Ibex Farm (Drummond 2013), and participants received a link to access the experiment. We used a word-by-word, non-cumulative self-paced reading paradigm (Just et al. 1982), which allows readers to determine the presentation duration of each word using button presses. The presentation of each item began with a fixation cross. Pressing the space bar triggered the presentation of a stimulus sentence's first word. Each word was presented at the centre of the computer screen and was replaced by the next word when participants pressed the space bar. The final word in each sentence was presented with a full stop, and when participants pressed the space bar again, a paraphrase sentence appeared whose appropriateness participants had to judge by clicking on the ‘yes’ or ‘no’ button shown on the computer screen. The computer recorded participants' word-by-word reading times, their end-of-trial responses, and their response times.

The main experiment was preceded by a set of biographical questions and the presentation of three practice trials in order to familiarize participants with the experimental task. Both experimental and filler items were presented in black letters (30pt Lucida Grande font) against a light grey background, and there were four pre-programmed breaks during the experiment. The whole experiment could be completed in about 30 minutes. A progress bar shown above the stimulus sentences allowed participants to keep track of their progress during the experiment.

3.3.4 Data cleaning and analysis

One participant was excluded prior to any data analysis as it turned out that he/she was an early bilingual speaker.

Prior to the analysis of the reading time data, we excluded both extremely long reading times above 2000 ms and excessively short reading times below 200ms, which affected 0.44% of the data. Raw reading times were transformed into residual reading times (RRTs) by linear modelling on the log-transformed reading times to eliminate effects of word length. RRTs were analysed for each of the five words of the infinitival complement as well as the matrix verb separately, adding up to six regions of interest. In addition, full sentence reading times were compared so as to obtain a measure of global processing difficulty. Statistical analyses were conducted on RRTs making use of linear mixed-effects modelling (Bates et al. 2015) including the fixed factor Condition (2 levels: ambiguous vs. incoherent), using treatment contrasts with ambiguous as reference level, as well as crossed random effects for participants and items (Baayen et al. 2008). Post-hoc comparisons between optionally coherent and obligatorily incoherent verbs additionally included the factor Verbal Class (2 levels: optionally coherent vs. incoherent). As for Experiments 1 and 2, for determining the best-fit random slopes structure, we started with the maximal model and iteratively removed random slopes which explained the least variance (Barr et al. 2013) until the model’s AIC value did not improve any further (Venables and Ripley 2002). The RRT at the pre-interest region was added as a co-variate to the best-fit model of each region of interest to control for potential spill-over effects from the preceding region (Bartek et al. 2011).

Regarding the end-of-trial paraphrase judgements, we analysed both the proportions of ‘yes' responses to paraphrases and participants’ response times. Paraphrase judgements were coded with 0 (= no) and 1 (= yes) and were analysed with generalized linear mixed-effects models (family = binomial) using the lme4 packages of R (Bates et al. 2015). In addition to the fixed factors Condition (ambiguous vs. incoherent) and Paraphrase (matrix vs. infinitival scope), we also included crossed random effects for participants and items (Baayen et al. 2008). To determine the best-fit random slopes structure, we employed the same method as described above. For main effects and overall interactions, we employed sum-coded contrasts to the factors, for effects within Condition and Paraphrase we employed treatment contrasts. Excessively long response times to paraphrases above 10000 ms were removed from the dataset (affecting 1.82% of datapoints) before they were log-transformed and analysed with linear mixed effects models using the lme4 package (Bates et al. 2015), with crossed random effects for participants and items (Baayen et al. 2008). The analysis of response times included the same fixed factors as the analysis of paraphrase judgements. The best-fit random slope structure was determined by using the same procedure as above.

3.3.5 Results

3.3.5.1 Reading times

Figure 5 displays participants’ log-transformed RRTs across the infinitival complement and matrix verb for both conditions. Statistical analyses revealed that full reading times for sentences containing intraposed ambiguous infinitivals in condition (15a) and intraposed incoherent infinitivals in condition (15b) did not significantly differ (t =  −  0.036).

Fig. 5

Log RRTs for the infinitival complement and matrix verb for both experimental conditions, Experiment 3 (AMB = ambiguous, IC = incoherent)

Summaries of statistical word-by-word RRT analyses are available at the Center for Open Science Framework website (https://osf.io/a7rqd/). Reading times at the DP (e.g., das Buch, Segments 3 and 4) did not differ statistically between the two conditions. However, the appearance of the infinitival marker zu led to a significant increase in RRTs in both conditions, at Segment 5 in the incoherent condition (15b) (t = 3.985) and at Segment 6 in the ambiguous condition (15a) (t =  −  2.827). At Segment 7, the appearance of the adverbial (e.g., zweimal) in condition (15b), signalling a clause boundary between the infinitive and the matrix verb, triggered a significant RRT increase (t = 2.809). On the matrix verb (e.g., versucht) at Segment 8, however, RRTs did not differ significantly, due to RRTs in the ambiguous condition (15a) showing a relatively greater increase from Segment 7 to Segment 8 compared to condition (15b). Post-hoc comparisons between optionally coherent and incoherent matrix verbs did not yield statistically significant effects (see OSF link).

3.3.5.2 Paraphrase judgements

Table 6 provides an overview of participants’ proportions of ‘yes’ answers and their response times (RTs) for each of the four paraphrase conditions. Overall, participants judged matrix scope paraphrases as more appropriate than infinitival scope paraphrases for both ambiguous (15a) and incoherent (15b) infinitives.

Table 6 Proportions of 'yes' responses and mean RTs (in ms) per condition, with SDs in parentheses, Experiment 3 (AMB = ambiguous, IC = incoherent)

The model output of the overall statistical analysis of participants' responses is shown in Table 7. We obtained an effect of Condition, such that significantly more ‘yes' responses were obtained for paraphrases following incoherent construals, an effect of Paraphrase, such that matrix scope paraphrases received statistically more ‘yes' responses than infinitival scope paraphrases, and an interaction of both fixed factors.

Table 7 Summary of the statistical analysis of participants’ responses to paraphrases, Experiment 3 (asterisks indicating significant effects at α  =  0.05)

To further investigate the interaction, we compared the two paraphrases for each of the two sets of conditions individually. As expected, the statistical analysis of the two paraphrase conditions for incoherent sentences (15b) revealed significantly more ‘yes' responses for the matrix scope paraphrase compared to the infinitival scope paraphrase (z  =  4.07). Interestingly, the proportions of ‘yes' responses to matrix scope paraphrases following ambiguous sentences (15a) was also significantly higher compared to the those for infinitival scope paraphrases (z  =  2.13). These judgement patterns indicate that participants allowed a matrix scope reading of the adverbial more readily than an infinitival scope reading, which in the case of ambiguous infinitives (15a) indicates that a monoclausal construal was easy to obtain. The analysis of mean response times yielded a similar picture (Table 8). While response times for paraphrases following incoherent sentences in condition (15b) did not statistically differ (t  =  1.22), matrix scope paraphrases indicative of a monoclausal coherent interpretation of ambiguous sentences (15a) were responded to significantly faster than infinitival scope paraphrases (t  =  2.13).

Table 8 Summary of the statistical analysis of participants’ response latencies, Experiment 3 (asterisks indicating significant effects at α  =  0.05)

3.3.6 Summary and discussion

Experiment 3 yielded three main findings. Firstly, full sentence reading times for our ambiguous and incoherent conditions did not statistically differ, which shows that incoherent intraposition does not necessarily incur greater processing cost compared to intraposed infinitives that can be construed coherently. Secondly, incoherent and ambiguous intraposition elicited different word-by-word processing patterns. Encountering the infinitival marker zu, which signals the presence of an embedded infinitival phrase and a corresponding delay in the arrival of the finite matrix verb, increased processing cost in both conditions. Another source of local processing difficulty was the appearance of an adverbial following the infinitive in condition (15b), which forced an incoherent construal of the intraposed infinitive. However, the relative processing advantage of the ambiguous over the incoherent condition disappeared when the matrix verb was encountered, at which point reading times increased disproportionally in the ambiguous condition. The latter observation supports the VCCH (Bader and Schmid 2009), which claims that the formation of a verbal cluster of the infinitive and the matrix verb is computationally costly. Thirdly, participants' paraphrase judgements and response times confirmed that ambiguous intraposition is readily construed coherently. The end-of-trial judgement results should be treated with some caution, however. Although matrix scope paraphrases were accepted relatively more frequently than infinitival scope paraphrases for the incoherent condition (15b), as was expected, participants' acceptance of infinitival scope paraphrases was unexpectedly high. Participants may have had a general bias towards accepting the paraphrases of critical items, possibly due to the difficulty of the task, in particular the difficulty to remember exactly where the adverbial had been placed in the preceding sentence. Additional statistical analyses on our filler sentences showed, however, that participants did not have a general positive response bias. They responded to filler paraphrases with an overall accuracy of 92.94%, with an equal distribution of 'yes' (50.89%) and 'no' responses (49.11%).

Nonetheless, the patterns we observed in the judgement data—more surface scope readings for incoherent and fewer surface scope readings for the ambiguous condition—went in the expected directions. The results from Experiment 3 will be further discussed in the next section, together with those from Experiment 1 and 2.

4 General discussion

Our study's first aim was to test how well the frequency of occurrence of different word-order patterns, processing-related factors and/or lexical verb class membership might be able to account for speakers' acceptability judgement patterns. Secondly, we investigated the hypothesis that coherent intraposition is preferred over incoherent intraposition on processing economy grounds. Our main findings can be summarized as follows:

• For verbs that allow for incoherent construals, extraposition elicited the highest acceptability ratings.

• Word-order patterns that were rarely attested in our corpora were rated as more acceptable than their low frequency of occurrence would have suggested.

• Although raising verbs were rated best for coherent or ambiguous intraposition, the two raising verbs under investigation did not pattern alike.

• Our four control verbs also showed noticeable verb-specific variation, with the verb versuchen behaving markedly differently from the rest.

• While incoherent intraposition was considered less acceptable than coherent or ambiguous intraposition, the results from our reading-time experiment showed that incoherent intraposition is not necessarily harder to process compared to intraposed infinitives that allow for a coherent analysis.

In the following, we will discuss the implications of our findings for different theoretical approaches to German infinitival complementation.

4.1 Frequency versus acceptability

Usage or frequency-based approaches to language performance predict that word-order behaviour is determined by frequency patterns as reflected in corpus probabilities. Our findings show this prediction to be only partially borne out: On the one hand, we obtained a ceiling match effect, such that the most frequent word-order variants for all verbs were also rated best (Bader and Häussler 2010). Intraposed ambiguous (together with intraposed coherent) infinitives received the most favourable ratings for the two raising verbs scheinen and pflegen (Experiment 2), while extraposed structures were rated most acceptable for subject control verbs. Consequently, for high-frequency word-order variants, frequency-based frameworks of language performance and representation can correctly predict speakers’ performance patterns.

On the other hand, word-order patterns which are attested relatively infrequently in synchronous corpora, such as extraposed infinitives for raising verbs, and intraposition and third construction for control verbs, exhibit substantial variability in rating patterns across different verbs—an effect previously referred to as a floor mismatch effect (Kempen and Harbusch 2008; Bader and Häussler 2010; Bader 2019). According to Bader and Häussler (2010, 316), a floor mismatch effect is observed if two or more syntactic variants with zero or near-zero frequencies still differ in terms of their perceived well-formedness and thus elicit significantly different acceptability ratings. Our findings provide ample evidence for a floor mismatch effect: While for scheinen, for example, all infrequent word-order patterns, i.e. intraposition incoherent, extraposition, and third construction, received ‘unacceptable’ ratings, the verb pflegen unexpectedly elicited intermediate ratings for highly infrequent extraposition and third constructions. Other infrequent or unattested word-order variants also received much more favourable ratings than their frequency counts would have predicted. Third constructions, for example, were not attested at all in our corpora for the verbs beschließen, bedauern, and ankündigen but nevertheless elicited fairly acceptable ratings in the intermediate scalar range. In the case of the verb versuchen, for example, third constructions elicited the second highest ratings after extraposition, even though third constructions were attested much less frequently with versuchen than intraposition in our corpora. The floor mismatch effects we observed for low-frequency word-order patterns suggests that frequency-based approaches to language performance cannot fully predict speakers’ acceptability rating patterns.

The recurring observation that infrequent and/or unattested constructions may nevertheless be deemed acceptable has been discussed both in the experimental psycholinguistic and theoretical linguistic literature. Newmeyer (2003), for example, argued that rarely used syntactic structures may be judged as acceptable alternatives because they exist in the speakers’ grammar representations as latent structures. In a similar spirit, Adli (2011) argues that the acceptability of optional syntactic variants is a necessary but not sufficient condition for their actual usage. He also classifies potential output candidates that are theoretically available from a speaker’s grammar but hardly ever used in language production as ‘latent constructions’ (see also Barbiers 2005 for a discussion of possible versus unrealized structures in the domain of verb clusters). Culicover (2013) compares such constructions to a concept from epidemiology. He argues that these constructions are present in the language of a population much in the same way as viruses are. “They may reside in a speaker without overt symptoms, that is, they are accepted as ‘possible’ or ‘acceptable’, but they are not produced” (252). The mismatch we observed between acceptability judgements and corpus probabilities supports the differentiation of the human grammar system and language use.

Featherston’s (2005) Decathlon Model provides an attempt to relate judgement data and frequency counts to one another systematically. It assumes two distinct functions to evaluate potential candidate structures: A constraint application function which assigns a well-formedness value to each potential output candidate, corresponding to the relative acceptability of each candidate. Frequency distributions obtained through corpus analysis, on the other hand, essentially reflect repeatedly made speaker choices. With each choice, the speaker evaluates a set of alternative, but potentially equally suitable, linearization patterns. This output selection process results in the forced selection of the optimal candidate, i.e. of the member of the set of alternative expressions that is associated with the highest acceptability value in a given context. Hence, any asymmetry between the outcome of such repeatedly produced forced-choice selections (i.e., corpus data), and speakers’ intuitions of the acceptability of alternative variants is indicative of a power-law distribution: While the winner candidates are selected consistently, less optimal candidates—although potentially felicitous choices—are rarely selected (see also Verhoeven and Temme 2017). This means that the strongest candidates—in our case, the extraposition pattern for control verbs and the intraposition pattern for raising verbs—will predominantly appear in corpus data since they reliably win the output selection competition. At the same time, corpus frequencies for all other candidates may tend to zero. In scalar judgement data, however, weaker candidates are not excluded by any output selection process. Therefore, the floor-mismatch effect we observed follows from the idea that corpus data can only identify forms that are good enough to be produced, whereas judgement data can also distinguish between variants that are not good enough to be used.

4.2 Processing-related factors

From a left-to-right incremental processing perspective, the extraposition pattern clearly wins out over all other word-order patterns. Unlike biclausal incoherent intraposition, extraposition avoids clausal centre-embedding, a well-known source of processing difficulty. Extraposing an infinitival complement also clearly signals its status as a clausal unit, in contrast to intraposition, where the presence of a clause boundary may initially be missed. The post-verbal placement of the entire infinitival complement moreover precludes an initial misanalysis of the embedded verb's object as an object of the matrix verb, which may happen with discontinuous third constructions as a result of this pattern's temporary ambiguity. Bosch et al. (2022) whole-sentence reading-time experiment provided evidence that the third construction is indeed the most difficult-to-process linearization pattern. Taking into account the claim that coherent construals of intraposed infinitives should be preferred over incoherent ones on processing economy grounds (Bayer et al. 2005), processing considerations led us to expect acceptability ratings to align with the scale in (17).

(17)	Extraposition > coherent Intraposition > incoherent Intraposition > Third Construction

For control verbs, extraposition was indeed rated best. Interestingly, although this linearization pattern was expected to be unavailable for raising verbs, extraposition was also considered acceptable here to some degree, in particular for the raising verb pflegen. We will discuss this finding further in the next section.

While coherent and ambiguous intraposed infinitives were rated second best for control verbs, in line with our processing scale in (17), we did not observe generally better ratings for incoherent intraposition compared to the third construction pattern (compare Fig. 2). That is, the third construction was not rated as poorly as might have been expected from a processing perspective. We follow Bosch et al. (2022) in concluding that the third construction pattern is grammatical but represents a latent construction that is rarely produced, possibly due to processing constraints and prescriptive norms both disfavouring it, the latter being supported by our finding that the third construction is more frequent in more informal spoken-like corpora, such as the Twitter corpus. Note that even though the third construction pattern may be comparatively difficult to comprehend, its temporary ambiguity should be irrelevant in language production as speakers (or writers) know what the message is they wish to convey.

The results from our Experiment 3 are in line with the hypotheses discussed by Bader and Schmid (2009) for monoclausal coherent vs. biclausal incoherent analyses of intraposed infinitives. Our reading-time data not only showed evidence of local processing difficulty when readers realised that an incoherent construal is required, but also evidence supporting the hypothesis that verb cluster formation incurs a processing cost. Together, these observations indicate that different processing-related factors interact in determining global processing difficulty: Although incoherent vs. ambiguous intraposed infinitives showed different word-by-word processing profiles, there was no difference in total processing times. Again, we see a discrepancy between the results from processing tasks and acceptability, albeit in the opposite direction from what we noted above: Incoherent intraposition is rated less favourably than coherent or ambiguous intraposition even though the former is not necessarily more difficult to process than the latter. Taken together, these findings show that the acceptability of different word-order variants cannot be fully predicted from processing-related factors (or vice versa). Processing constraints by themselves also cannot account for verb-specific variation in word-order behaviour, a finding we discuss below.

4.3 Verb-specific variation

Formal linguistic theory categorically distinguishes raising from control verbs. Unlike control verbs, raising verbs are assumed to be obligatorily coherent, thus disallowing biclausal construals. This general distinction was supported both by our corpus and acceptability data. Raising verbs showed the highest ratings for monoclausal coherent construals, whilst control verbs favoured extraposition, i.e. biclausal construals. However, individual verbs within each set did not behave homogeneously. Whilst the verb scheinen clearly favoured coherent or ambiguous intraposition, the verb pflegen elicited similarly acceptable ratings for both intraposed coherent and extraposed infinitival complements. The observed differences between the two raising verbs were unexpected both from a formal linguistic perspective (Wurmbrand 2001; Haider 2010) and from the results of our corpus analysis, according to which both scheinen and pflegen occurred exclusively with intraposed infinitives.

One potential explanation for the observed difference between scheinen and pflegen might be rooted in the lexical ambiguity of the verb pflegen. In contrast to scheinen, the verb pflegen can be used both as a subject raising verb (‘be in the habit of’) or as a transitive verb (‘to care for’) selecting a nominal direct object and also assigning a thematic role to its subject (e.g., Der Junge pflegt seinen kranken Vater. ‘The boy cares for his sick father.’). Transitive pflegen can also appear in the sense of ‘to maintain’ (e.g., Er pflegt einen teuren Lebensstil. ‘He maintains an expensive lifestyle.’). This reading appears to be conceptually close to pflegen as a raising verb (e.g., Er pflegte einen teuren Lebensstil zu führen. ‘He used to lead an expensive lifestyle.’). When considering the absolute frequency of occurrence of scheinen and pflegen in combination with an infinitival complement, pflegen occurs considerably less often, i.e. in only 137 attested sentences, than scheinen, which is attested in 1637 cases (see Table 2). Due to the argument structure ambiguity and low frequency of pflegen as a subject raising verb, participants might have been more tolerant of extraposition in combination with the verb pflegen compared to scheinen. Future research might further explore the conditions under which extraposition is deemed possible for raising verbs, a word-order option that was grammatically licensed in earlier stages of the history of German, but which gradually disappeared for the verbs scheinen and pflegen, probably as a consequence of increasing normative pressure (De Cesare 2021).¹³ Looking at individual participants' rating patterns also revealed evidence for considerable inter-individual variability: A subset of participants rated extraposition for the verb pflegen as good as, or even better than, coherent intraposition (Experiment 1: 24 of 39 participants; Experiment 2: 35 of 56 participants). In each rating task, these included six participants who rated extraposition best for all verbs under investigation. This finding could be taken to suggest that for some speakers the raising verb pflegen was not categorically different from control verbs.

For subject control verbs, verb-specific variation patterns were expected from the perspective of several linguistic approaches. Haider's (1994) binary grouping of German control verbs into obligatorily biclausal and optionally monoclausal verbs predicted that ankündigen and bedauern should pattern together vs. versuchen and beschließen. Our acceptability judgement results are more in line with Wurmbrand's (2001) three-way classification system, however. According to Wurmbrand (2001), the lexical restructuring verb versuchen would be expected to allow both monoclausal coherent construals and the third construction pattern more readily compared to the reduced non-restructuring verb beschließen. While the verb beschließen received significantly better ratings for monoclausal coherent intraposition compared to the verbs ankündigen and bedauern, it elicited significantly worse ratings for the coherent intraposition than the verb versuchen.

The judgement patterns we observed for the third construction in Experiment 1 are also consistent with Wurmbrand’s (2001) classification system. The verb versuchen elicited significantly better ratings for third constructions than the verb beschließen, which patterned with the obligatorily biclausal verbs ankündigen and bedauern in this regard. A similar pattern was seen in the corpus frequency distributions: The third construction only occurred with versuchen but was not attested at all for any of the other three control verbs.

Note, however, that the number of control verbs we examined here was very small. This is a limitation of the present study which leaves open the possibility that, if a larger set of control verbs were tested, evidence for a more continuous scale of coherence preferences along the lines suggested by Grosse (2005) might emerge.

5 Concluding remarks

The present study combined corpus-based and experimental investigation to test hypotheses derived from usage-based, processing and formal linguistic approaches to word-order variation in German infinitival complementation. Our findings revealed a complex picture, emphasizing the usefulness of combining different types of data in order to gain more comprehensive insights than can be obtained from corpus analysis or experimentation alone. We found that neither frequency distributions nor processing-related factors can fully account for the acceptability of different word-order patterns in syntactic variation contexts. Acceptability ratings of low-frequency variants aligned imperfectly with our corpus counts, showing that word-order variants that are strongly dispreferred in language production may nevertheless be deemed acceptable. The acceptability of the third construction pattern was also unexpected from a processing perspective as this word-order variant had previously been shown to be difficult to process. On the other hand, we found that incoherent intraposition, although barely attested and rated less favourably than coherent intraposition, was not any more difficult to process than the former. This finding demonstrates the importance of testing claims about syntactic variants' relative processing difficulty by gathering appropriate experimental data.

We also observed that verbs belonging to the same major class, i.e. raising vs. control verbs, did not necessarily elicit uniform judgement patterns across different word-order and coherence conditions. Our judgement data suggest that these verbs’ word-order behaviour with respect to monoclausal versus biclausal construals might best be captured by a gradient approach to coherence. However, limiting our investigation to a small set of infinitive-embedding verbs does not allow for any firm conclusions to be drawn in this regard. We hope that our study will inspire future studies of this kind that look at larger sample of verbs so as to obtain a more comprehensive and nuanced picture of their coherence properties and of the factors that influence these.

Our findings suggest that while usage frequency, processing economy, and lexical properties of individual verbs all play a role in determining which word-order variants win out over others, their relative importance or weighting might vary from task to task, and possibly also from person to person. Systematically examining the issue of inter-individual variability in word-order preferences might be another worthwhile objective for future research.