Jump to ContentJump to Main Navigation
Show Summary Details
More options …


An Interdisciplinary Journal of the Language Sciences

Editor-in-Chief: Gast, Volker

IMPACT FACTOR 2018: 1.066

CiteScore 2018: 0.97

SCImago Journal Rank (SJR) 2018: 0.384
Source Normalized Impact per Paper (SNIP) 2018: 1.409

See all formats and pricing
More options …
Volume 57, Issue 5


Clustering and stranding in Dutch

Frank Van Eynde
Published Online: 2019-06-26 | DOI: https://doi.org/10.1515/ling-2019-0023


This paper has both a descriptive and a theoretical aim. The descriptive one is to demonstrate that the phenomenon of clustering is not limited to verbs, but that it also affects adpositions. The theoretical one is to develop a formal analysis that captures the common properties of verb clustering and adposition clustering. For that purpose we employ the framework of Head-driven Phrase Structure Grammar. Both the descriptive and the theoretical part are backed up with quantitative data about the use of adposition clusters in a sample consisting of one million words of spoken Dutch and one million words of written Dutch.

Keywords: circumposition; particle; adposition stranding,adposition cluster; verb cluster; complement raising; complement extraction,Head-driven Phrase Structure Grammar

1 Introduction

Van Eynde (2017) argues that Dutch not only allows the formation of verb clusters but also of adposition clusters. The latter take the form of stranded circumpositions and have a number of properties in common with verb clusters. This paper provides an update of that descriptive study (Sections 2 and 3) and demonstrates that the phenomena of clustering and stranding are interrelated, in the sense that they result from the same syntactic processes, i.e. complement raising and complement extraction.1 The demonstration takes the form of an analysis which is cast in the notation of Head-driven Phrase Structure Grammar (Section 4). The analysis is compared with other monostratal treatments of clustering and stranding (Section 5), and the main findings are summarized in the conclusion (Section 6).

Both the descriptive and the theoretical part are backed up with quantitative data about the use of verb clusters and adposition clusters in a sample consisting of one million words of spoken Dutch and one million words of written Dutch.2

2 Circumpositions

Dutch adpositions may precede their complement, as in (1), or follow it, as in (2).


‘In August 1864 Sheridan started a raid through the valley’



‘All day long people are coming in order to sign the register’


Adopting standard practice, we call them prepositions and postpostions respectively. A third possibility for the Dutch adpositions is to surround their complement, as in (3).


‘Against the earlier agreements Milosevic gets speaking time’


We call them circumpositions. Some other examples are door ... heen and achter ... aan, as used in (4–5).


‘Throughout the years those treaty funds have not been handled well’



‘They are after the chickens’


This section provides an inventory of the Dutch circumpositions (2.1), a set of criteria to differentiate them from separable verb particles and predicative adpositions (2.2), and an analysis of the internal structure of circumpositional pps (2.3).

2.1 An inventory of circumpositions

Inventories of circumpositions can be found in descriptive grammars, such as Haeseryn et al. (1997: 526–530) and Broekhuis (2013: 50–51). The former lists 32 circumpositions, the latter 36, each exemplified by self-made sentences. To collect some data about usage and frequency we did a search for circumpositions in the two million word sample. This is facilitated by the fact that the annotation manual prescribes the use of a separate dependency label for the second part of a circumposition (@rel=“hdf”) (Hoekstra et al. 2003). Making use of GrETEL 2.0, a tool for example-based treebank querying (Augustinus et al. 2012), we derived XPath queries which retrieve all pps which consist of an adposition, an object complement and another adposition, in that order.3 They yield 270 hits for lassy small and 571 for the cgn treebank. Table 1 provides a survey of the relevant hits.4 They are classified according to the rightmost adposition. This yields 11 classes. Most frequent are the combinations with toe (333 hits) and heen (183 hits). Some of the circumpositions consist of more than two adpositions, as in (6).

Table 1:

The Dutch circumpositions in the sample.


‘Hoorn was involved in whale hunting nearly from the beginning’


They are listed separately below the double line. The combinations in bold do not figure in either Haeseryn et al. (1997: 526–530) or Broekhuis (2013: 50–51). A common property of these circumpositions is that they exclusively occur in the Spoken Dutch Corpus, mostly in the components which consist of spontaneous face-to-face conversations (fna and fva). An example is naar ... heen, as used in (7).


‘if you are going to Mexico’


The absence from Haeseryn et al. (1997: 526–530) and Broekhuis (2013: 50–51) is probably due to the fact that they focus on Standard Dutch.5

Conversely, some of the circumpositions that are listed in Haeseryn et al. (1997: 526–530) and Broekhuis (2013: 50–51) do not occur in the sample. Several of them concern combinations with langs and om, which are the least common in the sample with respectively 3 and 7 hits (boven/onder/voor ... langs and achter/voor ... om). Also missing is a number of combinations with uit (achter/tussen ... uit) and vandaan (achter/om/tussen/uit/voor ... vandaan). Their absence is probably due to data sparseness.

2.2 Identifying circumpositional pps

Not every [pnpp] sequence is a circumpositional pp. To illustrate this, let us compare the [vannpaf] sequences in (8–9).


‘The Japanese had expected right from the start that ... ’



‘...the House Wittelsbach gave up its claims to the throne of the emperor’


In (8) af is part of the circumpositional pp van het begin af, but in (9) it is the separable particle of the verb afzien ‘give up’, which takes a pp-complement that is introduced by van. To differentiate them we use three diagnostic tests. The first concerns topicalization: While a circumpositional pp can be topicalized, as in (3–4) and (8), a sequence of a pp and a separable verb particle cannot, as shown in (10).


This suggests that a circumpositional pp is a constituent while a sequence of a pp and a separable verb particle is not. Confirming evidence is provided by a second test: While a circumpositional pp can be used adnominally, as in (11), a sequence of a pp and a separable verb particle cannot.


‘The distribution of this picture throughout the years has ...’



A third test is based on the linear order of the adpositions. In a circumpositional pp that order is fixed. It is not possible to put p2 before p1, as shown by the contrast in (13).


‘The Japanese had expected right from the start that ... ’


In a sequence of a pp and a separable verb particle, by contrast, the order is not fixed. The particle may follow the pp, as in (9), but it may also precede it, as in (15).


‘...the House Wittelsbach gave up its claims to the throne of the emperor’

The first two tests are also used in Broekhuis (2013: 52–66) (his numbers IV and V) and the the third one is a generalization of his pp-over-v test (III). The two remaining tests in Broekhuis (2013) are based on the omissibility of the [pnp] sequence (I) and on its substitutability by a locative pronoun (II), but the application of these tests is accompanied by so many ifs and buts that we leave them aside.

More relevant at this point is the fact that the three diagnostic tests can also be used to differentiate circumpositional pps from other sequences of a pp and an adposition that do not form a constituent. A relevant case is (16), where af is the predicative complement of moet ‘must’.6


‘she must indeed get rid of her rubbish’


That this is not an instance of the circumposition van ... af is clear from the fact that the [vannpaf] sequence cannot be topicalized nor used in adnominal position.


Confirming evidence is provided by the fact that af may precede the pp complement of the verb, as in (19).


‘Also the two politicians wanted to get rid of that community nonsense’


Summing up, circumpositional pps are constituents that can be topicalized and used adnominally, and the linear order of the adpositions in a circumposition is fixed. This differentiates them from sequences of a pp and a separable verb particle or a predicative adposition.

2.3 The internal structure of circumpositional pps

Within a circumpositional pp the np complement is more tightly related to the first adposition than to the second one. This is clear from the fact that the [p1np] sequence can be separated from p2, as in (20), whereas the [npp2] sequence cannot be separated from p1, as shown in (21).7


‘which parade did the children run after ?’


Some instances of p2 stranding from the sample are (22–23).8


‘that 3000 books had not made it through the conversion process’



‘I then run around behind Miranda’


To capture this we assume that circumpositional pps have a left branching structure, as in (24).


Further evidence for this structure is provided by the fact the sample contains several instances in which p2 scopes over a sequence of [p1np] conjuncts, as in (25), but none in which p1 scopes over a sequence of [npp2] conjuncts.


‘we have been reading against the clock and each other’


The binary structure is also applicable to circumpositions with more than two members, as in (26) and (6), repeated in (27).


‘all disappearance files are taken out from under the dust’



‘Hoorn was involved in whale hunting nearly from the beginning’

The relevant structures are given in (28).


The structure at the left applies to combinations with two prepositions and one postposition. It is not only attested by van onder ... vandaan (1 hit), but also by tot aan ... toe (1 hit) and tot in ... toe (3 hits). Evidence for treating the lowest pp as a complement of the first adposition is provided by the fact that van and tot belong to the small class of prepositions which can take a pp complement, as in van voor de oorlog ‘from before the war’ and tot aan de grens ‘till at the border’.9 The structure at the right applies to combinations with one preposition and two postpositions. Evidence for treating van ... af as a constituent is provided by the fact that it also occurs without aan, as in (29).


‘... which rejected the terror version from the beginning’


Taking stock, circumpositional pps have a binary branching structure in which the rightmost adposition heads the higher pp and takes the lower pp as its complement. As usual in such combinations, the selecting adposition constrains the choice of the adposition in the selected pp. Af, for instance, selects a pp which is introduced by bij, op or van.

3 r-pronouns and adposition stranding

A peculiar property of the Dutch adpositions is that they do not combine with a number of singular neuter pronouns. The demonstrative dat ‘that’, for instance, cannot be used as the complement of an adposition, as in (30). In that combination it must be replaced by the r-form daar ‘that.r’. Moreover, it must precede the adposition, as in (31).


‘that she often thinks of that’

The same holds for dit ‘this’, het ‘it’ and wat ‘what’.10 For the quantifying pronouns iets ‘something’, niets ‘nothing’ and alles ‘everything’, the replacement is optional: (32) and (33) are equally well-formed.


‘that she simply thinks of nothing from time to time’


‘that she simply thinks of nothing from time to time’

Table 2 provides a survey of the pronouns which show the [–/+ r] alternation.

Table 2:

The Dutch pronouns with an r-form.

The alternation has no effect on the form of the adposition in (30–33), but for some of the adpositions there is an effect. Met ‘with’ and tot ‘till’, for instance, invariably precede their complement and are, hence, not compatible with an r-pronoun. In that combination they are replaced by respectively mee and toe, as in daar mee/*met and er toe/*tot.

Characteristic of the r-pronouns is that they may be realized out of the local pp, leaving the adposition stranded, as in (34).


‘that she often thinks of that’

This phenomenon has been studied extensively. Descriptive surveys are provided in Haeseryn et al. (1997) and Broekhuis (2013), transformational treatments in Van Riemsdijk (1978) and Bennis (1986), and monostratal treatments in Rentier (1993), Bouma (2000) and Van Eynde and Augustinus (2014).

3.1 Stranded circumpositions

Of special interest in this paper is the use of r-pronouns in combination with a circumposition. This use is unexceptional: In the combination of het ‘it’ with naar ... toe, for instance, the pronoun takes the r-form and precedes its selector, as in (35).


‘... come to it by one’s own means’


Moreover, it may be realized out of the local pp, as in (36).


‘still breast cancer specialists go there every year’


The sequence naar toe in (36) is, hence, a stranded circumposition.

To find out which of the circumpositions can be stranded and to collect some data about frequency and usage we turned again to the two million word sample. Unfortunately, retrieving the relevant instances was less straightforward than anticipated, for while the distinctive dependency label for the second part of a circumposition (@rel=“hdf”) is assigned with considerable accuracy if the circumposition surrounds its complement, it is not assigned consistently if the circumposition follows its complement, especially if it is separated from it, as in (36). For that reason we ignored the annotation and resorted to string search by regular expressions. A second complication is the inconsistency of the orthography: While the two parts of a stranded circumposition are canonically treated as separate words, as in (35–36), there is a tendency to amalgamate them when they are adjacent, as in (37).


‘in case you might want to go there.’


In spite of what the spelling suggests, the syntactic and semantic relations between er, naar and toe in (37) are the same as in (35). Making matters worse, the inconsistency of the orthography also affects the r-pronouns: While they are canonically treated as separate words, there is tendency to incorporate them in the adposition when they are adjacent, as in (38–39).


‘what kind of audience comes to that?’



‘the people who were obliged to go there’


Table 3:

Adposition clusters in the sample.

Also here, there is no difference in meaning or syntactic function between the instances with an incorporated pronoun and those in which the pronoun is treated as a separate orthographic unit. For that reason each query had to be applied in four variants. Doing that for the combinations in Table 1 as well as for the twelve non-attested ones (listed in the last paragraph of Section 2.1), we got more than 1700 hits. Given the crude nature of string search it is not surprising that many of these are false hits. They were filtered out manually.11 Table 3 provides a survey of the relevant hits, sorted on the basis of the rightmost adposition. We find the same 11 groups as in Table 1, except for the fact that two are missing, i.e. those with om and vandaan. Examples can be constructed, though, such as (40), or found in larger corpora, such as (41), which is extracted from the SoNaR corpus (Oostdijk et al. 2013).


‘we just about managed to get around it’


‘he came from behind it a couple of times’

The sample does not contain any combinations with more than two adpositions either, but they can be constructed as well.12


‘from under which a woman was taken alive’


‘what did they take those files from under ?’

The only circumpositions for which the stranding is impossible are those which are introduced by an adposition that is invariably prepositional. They include tot toe and its ternary variants tot aan toe and tot in toe. They cannot be stranded, since tot cannot be preceded by its complement.

3.2 Clustering

At this point we are ready to describe the similarity between stranded circumpositions and verb clusters. An example of the latter is given in (44).


‘that we had better not taken that train’

The auxiliary in this clause takes a participial vp as its complement and that complement is headed by genomen ‘taken’, but the latter is separated from its object complement die trein ‘that train’ by the adjunct beter niet ‘better not’. As a result, the auxiliary and the participle form a cluster that is separated from the non-verbal complements of the verbs.

This mirrors the relations in (45).


‘she can get everywhere by bike.’


The postpositional toe in (45) takes a pp complement and that complement is headed by naar ‘to’, but the latter is separated from its complement overal by the adjunct op de fiets ‘on the bike’. As a result, the adpositions form a cluster that is separated from the non-adpositional complements of the adpositions.

There are, of course, also differences between verb clusters and adposition clusters. The linear order in adposition clusters, for instance, is rigid, while the one in verb clusters is more flexible.13 Besides, verb clusters are much more common than adposition clusters. A count in the same sample as the one that is used in this paper yields a result of 19,074 clusters of two verbs (Augustinus 2015: 127), which is 44 times more than the total of 432 clusters of two adpositions. From a structural point of view, though, the phenomena are so much alike that it is worthwhile to explore whether a uniform account makes sense. This is the aim of the next section.

4 Analysis

There are many proposals in the literature for the treatment of verb clustering, on the one hand, and adposition stranding, on the other hand, but there are no proposals yet for the treatment of adposition clustering. The purpose of this section is to fill this gap. Starting from the observation that adposition clusters have much in common with verb clusters, we aim for an analysis that is sufficiently general to deal with both types of clusters. For this purpose we employ the framework of Head-driven Phrase Structure Grammar (hpsg), more specifically the constructional version developed in Ginzburg and Sag (2000). Since this is a monostratal framework the phenomena of stranding and clustering are not modeled in terms of movement. There is, for instance, no transformation that moves the complement of a verb or adposition out of its local projection into some given landing site. Instead, the non-locally realized complement is related to its head by means of the sharing of selection requirements. To show how this works, we first discuss the local realization of complements (4.1) and then the two types of non-local realization that trigger clustering, i.e. complement raising (4.2) and complement extraction (4.3). In a last step we argue why it is important to differentiate complement raising from complement extraction (4.4).

4.1 Complement realization

The subordinate clause in (46) contains a verb (geeft ‘gives’) which selects two complements, i.e. a direct object (water ‘water’) and an indirect object (die plant ‘that plant’).


‘it seems that she never waters that plant’

To model this selection hpsg employs the valence feature comps. It is assigned to all signs (words as well as phrases) and its value is a list that contains the elements which the sign selects as a complement.14 The verb geeft ‘gives’, for instance, has two nominals on its comps list: one for the direct object and one for the indirect object. When combined with a nominal, such as water ‘water’ in (46), the resulting phrase has still one nominal on its comps list. The addition of an adjunct, such as nooit ‘never’, does not affect the comps list, but the addition of the indirect object does, yielding a phrase with an empty comps list, as shown in (47).15


To model this in general terms we use the implicational constraint in (48).16

What this constraint states, is that in signs of type head-complement-phrase

  • -

    the syntactic and semantic properties of the non-head daughter match the requirements that the head daughter imposes (1);

  • -

    the saturated comps requirement is absent from the comps list of the mother;

  • -

    the other elements on the comps list, if any, are shared with the mother (A).

Since (48) does not impose any constraints on the part of speech of the daughters, it also subsumes phrases with a circumposition, as in (49).


The adposition tegen ‘against’ selects a nominal complement (2) and when that complement is added, the result is a pp with an empty comps list. That pp is in its turn selected by the adposition in ‘in’ (3) and the result is again a pp with an empty comps list.

A minor complication for this treatment is that complements are not always overtly realized. The direct object complement of eaten, for instance, is not overtly realized in we have not eaten yet. The canonical way to model this in hpsg is to make the np on the comps list of the verb optional, as in [comps <(np)>]. This intransitive use is also possible with certain adpositions. In (50), for instance, the complements of voor ‘for’ and tegen ‘against’ are not overtly realized.


‘Who is in favor (of it)? And who is against (it)?’

Unsurprisingly, the phenomenon also affects the circumpositions, as in (51).


‘as entremets there was the sjellik with a spoonful of soup over it


To model this we allow the np complement of over to be optional. Technically, this means that its comps value is allowed to be the empty list, as in (52).


Of the 432 adposition clusters in the sample there are 12 (2.78 %) with an omitted complement (i.e. without overt complement).

A more serious complication for the treatment of complement realization is the fact that complements are not always realized in the local projection of their selector. To model this we adopt some terminology from topological field theory, as described amongst others in Engel (1970) and Höhle (1986). In that theory clauses are partitioned in two poles and three fields. The first pole, also known as the linke Satzklammer, is filled by the finite verb in main clauses and by the complementizer in subordinate clauses. The second pole, also known as the rechte Satzklammer, is filled by the verb cluster; it contains the non-finite verbs (if any); in subordinate clauses it also contains the finite verb. The space before the first pole is the Vorfeld, the space between the two poles is the Mittelfeld, and the space after the second pole is the Nachfeld.

Making use of these distinctions we differentiate between complement raising and complement extraction. The former concerns the realization of a complement outside of the local projection of its selector, but inside the Mittelfeld. The latter concerns the realization of a complement in the Vorfeld.

4.2 Complement raising

To model complement raising we take a cue from the analysis of the Dutch verb clusters in Van Eynde and Augustinus (2013).17 It is briefly presented in 4.2.1 and extended to the analysis of adposition clusters in 4.2.2. Vacuous complement raising is discussed in 4.2.3. In the sample complement raising accounts for the formation of 308 of the 432 adposition clusters (71.30 %).

4.2.1 Complement raising out of verbal projections

If we replace the simple present geeft ‘gives’ in (46) with the present perfect heeft gegeven ‘has given’, we get a combination in which the participle is separated from its complements by the auxiliary.


‘it seems that she never watered that plant’

Assuming that the auxiliary selects a participial complement, the application of the constraint in (48) yields (54).


This, however, is not what we want. Since the comps requirements of the participle are unsaturated, they should be propagated up the tree, till the point at which the selected nominals are added, as in (55).


To model this we add the implicational constraint in (56), quoted from Van Eynde and Augustinus (2013).

What (56) says, is that in signs of type headed-phrase, the unsaturated comps requirements of the non-head daughter (A) are added to those which the mother inherits from the head daughter (B).18

To understand the effect of this constraint, it has to be considered in tandem with the constraint on head-complement phrases in (48). While (48) has the effect of removing elements from the comps list, (56) has the effect of adding elements to the comps list. As a consequence, since a head-complement phrase is by definition a headed phrase –-in hpsg parlance the former is a subtype of the latter–- it is possible for a comps list to be extended with new members while at the same time losing other members. In fact, this is what happens in the node which immediately dominates the verb cluster in (55). Its comps list no longer contains the requirement for a past participle, but is extended with the nonsaturated comps requirements of that participle.

4.2.2 Complement raising out of adpositional projections

Since there are no constraints on the part of speech of the daughters in (56), the latter also models the raising out of adpositional projections, as in (57), where the r-pronoun is raised out of the pp.


‘they could not look over it’

Employing the constraints in (48) and (56), the analysis of (57) looks as follows:


The comps requirement of kijken (2) is immediately saturated and so is the one of heen (3), but the requirement of over for a nominal complement (4) is not. It is added to the comps list of the pp, inherited by the vp and propagated up the tree, just as in (55). The formation of the adposition cluster in (58) is, hence, the result of the same phenomenon as the one that triggers the formation of the verb cluster in (55).

4.2.3 Vacuous complement raising

An issue arises when the complement of a verbal or adpositional cluster immediately precedes its lexical selector, as in (59–60).


‘it seems that she sent a card then’


‘the crowd that was attracted by this’


For the analysis of such combinations one can adopt a left branching structure in which the nominal complements are sisters of respectively gestuurd ‘sent’ and op ‘up’, as in (61–62).


het schijnt dat ze toen [[een kaartje gestuurd] heeft]


de menigte die [[hier op] af] kwam

Alternatively, one can assign a right branching structure in which the complements are sisters of the respective clusters, as in (63–64).


het schijnt dat ze toen [een kaartje [gestuurd heeft]]


de menigte die [hier [op af]] kwam

The second option looks more complex, but it has the advantage of providing a more uniform account. In the case of (59), for instance, the structure in (63) is also valid if the participle follows the auxiliary, as in een kaartje heeft gestuurd ‘a card has sent’, while the structure in (61) is no longer applicable. By analogy, we assume that (64) is preferable to (62) for the adposition cluster. More specifically, we assume that the r-pronoun in (60) is raised out of the lower pp and realized as a complement of the cluster, as in (65).


Confirming evidence is provided by the fact the r-pronoun may scope over a conjunction of clusters, as in (66).


‘whether you run under or over that’

In principle, it would be possible to treat the r-pronoun as raised out of the higher pp as well and, hence, as a dependent of the verbal projection. This, however, is neither necessary nor desirable, since the combination of the r-pronoun and the adposition cluster is treated as a constituent in the case of extraction, as illustrated in (67).


‘In between Joan la Barbara and Susan Narucki sang Cage’s songs’


It is also treated as a constituent in the absolute met-construction in (68).


‘The columns consist of concrete with a layer of bronze around that’


Of the 308 instances of complement raising in the sample, 156 concern vacuous raising.

4.3 Complement extraction

To model complement extraction we adopt the treatment of unbounded dependencies in Ginzburg and Sag (2000: ch. 5). We first discuss its role in the formation of clusters, both verbal and adpositional ones (5.3.1), and then treat the phenomenon of topic drop (5.3.2). Of the 432 adposition clusters in the sample 112 (25.92 %) result from complement extraction.

4.3.1 Complement extraction out of verbal and adpositional projections

An example of complement extraction is given in (69).


‘what would they have promised?’

The direct object of beloofd ‘promised’ is not realized in its local projection. Instead, it appears in the Vorfeld, preceding the first pole, which is filled by the finite auxiliary zouden ‘would’. To relate the extracted complement to its selector Ginzburg and Sag (2000) does not rely on the sharing of comps values. Instead, the comps requirement that is not locally saturated is subtracted from the comps list of the selector, stored in another feature, called slash, and propagated up the tree, as in (70).19


The subtraction is modeled in terms of the Argument Realization Principle (Ginzburg and Sag 2000: 171).20

The subtracted comps requirement is an object of type gap-synsem and such objects have the defining property that their local value, which contains their syntactic and semantic properties, is added to their slash set (Ginzburg and Sag, 2000: 170).

In other words, if the selector requires, say, an accusative np complement, then this requirement is stored in the slash value. To model its propagation we assume that the slash set of the mother equals the union of the slash sets of the daughters. This stops when the slashed clause (zouden ze beloofd hebben) is combined with the extracted complement (wat). This is modeled by the constraint on phrases of type head-filler in (73), adapted from Ginzburg and Sag (2000: 174).21

In signs of type head-filler-phrase, the head daughter is required to be a verbal projection whose slash set contains an element of type local that matches the loc value of the non-head daughter. This checks whether the requirements which the selector of the extracted element imposes are met. In (70), for instance, the extracted element must be an accusative np.

Extraction out of adpositional phrases can be modeled in the same way, as illustrated for the circumpositional pp in (74–75).


‘where you can send your complaint to’



The adposition naar saturates the comps requirement of toe (1), but its own requirement for a nominal is subtracted from its comps list, added to its slash set, and propagated up the tree till the point where the r-pronoun is added.

4.3.2 Topic drop

Extracted complements are not always overtly realized. The phenomenon is known as topic drop and is exemplified by such elliptical clauses as doen we ‘do we’ and ken ik niet ‘know I not’, where the omitted topic is a pronoun like dat ‘that’. It also affects the extracted complements of circumpositions, as in (76).


‘Yes ok, but I’ll get over it, don’t you think so?’


The omitted topic is understood to be daar ‘that.r’ and the omission is optional: Adding it in the position of the first dash yields a well-formed combination.

It is worth stressing that topic drop is distinct from complement omission. There is, first of al, a clear stylistic difference. While complement omission occurs both in spoken and in written Dutch, including formal registers, topic drop is typical of informal spoken Dutch: All instances in the sample are from the spoken Dutch treebank (15 from the Netherlands and 1 from Belgium). Another difference concerns the omitted material: While the omitted complement of intransitively used adpositions is er ‘it.r’, the omitted element in the case of topic drop is daar ‘that.r’.

To model topic drop, we treat it as a filler-gap mismatch: The requirement of the verb or adposition for a nominal complement is subtracted from the comps list, added to the slash set and propagated up the tree in the usual fashion. The only unusual bit is that the filler is not overtly realized. Another instance of a filler-gap mismatch in the sample is (77).


‘to Euro Disney I’d like to go to some time as well’


While the extracted complement is an np, the filler is a pp. Just like topic drop, this mismatch is stylistically marked. In fact, most speakers find it ill-formed.

4.4 Constraints on complement raising and complement extraction

There is a lot that complement raising and complement extraction have in common. In both cases there is a complement that is not realized in the local projection of its selector and in both cases there is a device of sharing selection requirements in order to relate the non-locally realized complement to its selector. The only difference, it seems, is the position of the complement: While raised complements appear in the Mittelfeld, extracted complements appear in the Vorfeld. The purpose of the section now is to show that this is not the only difference. More specifically, we will show that there are complements which can be raised but not extracted (4.4.1) and we will argue that raising and extraction are subject to different constraints (4.4.2).

4.4.1 Major vs. minor

Most of the complements that can be raised can also be extracted, and vice versa, but there is an important exception in the case of the pronouns. For many of its pronouns Dutch has both a full form and a reduced form. The former have a clear vowel, as in jou ‘you’, and mij ‘me’, while the latter usually have a mute vowel, as in je ‘you’ and me ‘me’. Both the full and the reduced forms can be raised out of the verb cluster, as in (78), but as the contrast in (79) shows, only the full forms can be extracted.


‘they should have fired you immediately’


‘they should have fired you immediately’

This correlates with some other differences: While the full forms can take dependents, such as alleen ‘alone’, and can be conjoined, the reduced forms cannot.


‘we saw only you’


‘Peter stood between you and me’

The distinction is also relevant for the r-pronouns, where the full form daar contrast with the reduced forms d’r and er. As expected, both can be raised, as in (82), but only the full form can be extracted, as shown in (83).


‘you can send your complaint there.’


‘that you can send your complaint to’

It may be worth stressing that the distinction concerns a syntactic property, rather than a phonological one, for while it is true that most of the reduced forms have a mute vowel as their nucleus, this is not always the case. The reflexive pronoun zich, for instance, has a clear vowel, but shows the defining characteristics of reduced forms: When topicalized, modified or conjoined, it must be replaced by the full form zichzelf, see Van Eynde (1999: 146). For this reason we make the distinction in the category value of the signs, rather than in the phonology value. More specifically, we assume that the values of the category feature come in two types, which we will call major and minor, after Van Eynde (1999: 141).


The distinction is orthogonal to the part-of-speech distinction. Technically, this follows from the fact that part-of-speech labels are values of the head feature, rather than of the category feature. Empirically, it is motivated by the fact that the distinction between major and minor is not only relevant for various types of pronouns (personal, reflexive, demonstrative, ...), but also for adpositions, as shown in Van Eynde (2004), and for determiners, as shown in Van Eynde (2006). Employing this distinction, we can formulate the relevant constraint as a restriction on the arguments which are subtracted from the comps list.

(85) requires the arguments which are subtracted from the comps list to be major. Minor signs can, hence, not be extracted. Since this constraint does not hold for complement raising it provides empirical evidence for the differentiation between complement raising and complement extraction.

When assessing the relevance of this constraint for other languages, it should be borne in mind that the major/minor distinction is subject to cross-linguistic variation. The fact that a word is minor in one language does not necessarily imply that its translational equivalents are minor as well. The Dutch pronoun het ‘it’, for instance, is minor, but its English equivalent is not, as demonstrated by the fact that it can be conjoined and modified, as in (86).


you might be tempted to read it and it alone, fanatically, ... (quote on the cover of Don Delillo’s Underworld)

It should also be borne in mind that the distinction does not coincide with the one between tonic and clitic pronouns: While minor signs are independent lexical units, clitics are often incorporated in another element, usually a verb.

4.4.2 Initial vs. final

Another reason for differentiating between complement raising and complement extraction is that there are languages which allow complement extraction but not complement raising. One of them is English. That it allows complement extraction is well known. (87) provides examples of extraction out of verbal and adpositional projections.


that topic I never want to discuss __ again


what did you say she was speaking about __ ?

Complement raising, however, does not seem to exist in English. In fact, it is explicitly ruled out by the Empty Comps Constraint, as spelled out in Ginzburg and Sag (2000: 33).22

This constraint makes it impossible for a phrase to inherit unsaturated comps requirements of its daughters.

For Dutch, which does allow complement raising, the constraint in (89) is obviously too strong, but that does not mean that its complements are allowed to be raised anywhere. For a start, they must be realized in the Mittelfeld. The first pole is, hence, a barrier for complement raising. To model this we need a way to define the first pole. Given that it contains either a complementizer or a verb, we introduce a feature that is assigned to both. We call it position and declare its values to be initial and final.23


The complementizers are invariably initial, but the verbs show more variation: The non-finite forms are final, the imperative forms are initial and the other finite forms are underspecified. The latter’s position value is resolved to initial if they occur in the first pole and to final if they occur in the second pole. Assuming that position is a head feature, its value is propagated throughout the local projection, as in (91).


The non-finite gegeven is the head of the vps water gegeven and die plant water gegeven, and hence shares its position value (1) with them. The finite auxiliary, by contrast, is the head of the vp heeft die plant water gegeven, and shares its position value (2) with the top node.

Employing the position feature, the barrier status of the first pole for complement raising can be formulated as follows:

What (92) says is that a headed phrase with the position value initial must have an empty comps list. An example is the top node in (91). The requirement that such phrases have an empty comps list implies that complements cannot be raised out of a v-initial vp, nor out of a cp. Raising out of a v-final vp, however, is allowed. Formally, (92) is similar to the Empty Comps Constraint in (89). The right hand side of the implication is, in fact, identical. The difference concerns the left hand side.

Intriguingly, the contrast between v-initial and v-final vps is also relevant for pps. Notice, for a start, that there are adpositions which invariably precede their complement, such as met ‘with’ and tot ‘till’,24 as well as adpositions which invariably follow their complement, such as af, mee and toe. Besides, there are adpositions which are underspecified in this respect, such as over ‘about’ and aan ‘on’. This strongly suggests that the distinction is relevant for adpositions too. Acting on that hint, let us assume that the position value is also assigned to adpositions and their projections. To model this we make a distinction between nominal and non-nominal parts of speech, where the latter comprise the verbs, the complementizers and the adpositions, and declare the position feature for the non-nominal parts of speech.25


non-nominal : [position position]

Given the constraint in (92) it follows that not only cps and v-initial vps are barriers for complement raising, but also p-initial pps. This implication turns out to be correct. To show this let us compare (95–96) with (97–98).


‘I read a book about that’


‘I read a book about that’


‘I contributed to a book about that.’


Both in (95) and (97), the position value of over is resolved to final, since it is preceded by its complement daar ‘that.r’. By contrast, the position value of aan in (97) is resolved to initial, since it precedes its np complement. Given that p-initial pps are a barrier for complement raising, this accounts for the fact that daar can be raised out of the pp in (95), but not out of the pp in (97).

Turning now to complement extraction, it is clear that it is not constrained by (92). Neither cps nor v-initial vps are barriers for extraction, as illustrated in (99), where the extracted complement is separated from its selector by two complementizers and the v-initial verb denk ‘think’.


‘what do you think she said she wants to buy?’

p-initial pps, however, do block extraction. This holds both for pps wich are headed by an inherently initial adposition, such as met ‘with’, and for pps which are headed by an adposition whose position value is resolved to initial, such as aan in (101).


To model this we assume that adpositions whose position value is of type initial must have an empty slash set.26

This constraint applies both to adpositions and their projections. It is similar to a constraint that is proposed for the Dutch adpositions in Tseng (2005): “we can say that only [+post] Ps can put their complement in slash, or equivalently, that all the slash set [sic] of all prepositions must be empty.”

Returning to the main topic of this section, the fact that the constraint on complement raising in (92) differs from the constraint on complement extraction in (102) provides further evidence for differentiating the two phenomena. Moreover, zooming out and taking a broader perspective, it has long been recognized in generative grammar at large that scrambling and extraction are sufficiently different to require different treatments.

4.5 Summing up

This section has shown that the factors which trigger the formation of adposition clusters are the same as those which trigger the formation of verb clusters, i.e. complement raising and complement extraction. Both phenomena have been given a fully explicit analysis which is cast in the notation of hpsg. Besides, the differentiation between complement raising and complement extraction has been backed up with empirical evidence.

5 A comparison with other monostratal treatments

The previous section has provided an analysis of the phenomena of clustering and stranding which is applicable to both verbs and adpositions. In this respect, it departs from the common practice in hpsg to treat clustering and stranding as unrelated, with the former being confined to verbs and the latter to adpositions. Possibly as a consequence of this assumption of unrelatedness, the treatments are very different. To illustrate this we start from the scheme in Table 4. While hpsg treatments of verb clustering commonly ignore the distinction between complement raising and subject raising (Section 5.1), hpsg treatments of adposition stranding commonly ignore the distinction between complement raising and complement extraction (Section 5.2). We will briefly present these proposals and show why it is preferable to treat complement raising as distinct from both subject raising and complement extraction.

Table 4:

Raising vs. extraction.

5.1 Verb clustering and generalized raising

The canonical way to model verb clusters in hpsg is based on generalized raising. It was originally proposed for a treatment of German verb clusters in Hinrichs and Nakazawa (1989) and Hinrichs and Nakazawa (1994), and got adopted and adapted by various authors to deal with similar phenomena in other languages, such as Dutch verb clusters (Bouma and van Noord 1998) and clitic climbing in French (Abeillé et al. 1998) and Italian (Monachesi, 1998). As an illustration of how it works, let us take (103).


‘It seems that she sent him a card’

To model the dependencies in this clause the auxiliary heeft ‘has’ is claimed to inherit the argument requirements of its participial complement. These include both the subject and the complement requirements. The auxiliary in (103), hence, selects four arguments: its own participial complement, augmented with the three arguments of that participle.


In this analysis subject raising and complement raising are treated in the same lexicalist head-driven way. In our treatment, by contrast, the subject requirement (1) of the participle is inherited by the perfect auxiliary, but its complement requirements (2 and 3) are propagated directly to the cluster and beyond. Technically, the difference only shows in the comps value of the auxiliary. The rest of the structure is identical to (104).


Conceptually, the difference is more important. What it intends to capture is the assumption that subject raising is indeed amenable to the lexicalist head-driven analysis that is familiar from the treatment of subject raising in English, but that complement raising requires another approach.

There are at least four arguments in favor of this differentiation. First, subject raising is a phenomenon that affects a limited number of lexical signs, more specifically the subject-to-subject raisers (modals, temporal and aspectual auxiliaries, ...) and the subject-to-object raisers (perception verbs, causatives, ...). These include mostly verbs and –-in some languages–- predicative adjectives, such as likely and bound Ginzburg and Sag (2000: 21). Complement raising, by contrast, is a much more general phenomenon. In the languages which allow it, such as Dutch and German, it occurs in all sorts of verbal and adpositional constructions, as illustrated in Section 4.2.

Second, subject raising and complement raising are mutually independent, in the sense that there are instances of complement raising which do not involve subject raising, and that there are instances of subject raising which do not involve complement raising. Starting with the former, subject control verbs, such as the Dutch willen ‘want’ and proberen ‘try’, obviously do not belong to the subject raising verbs, but they do allow complement raising, as shown in (106–107), where the bracketed complements of stoppen ‘stop’ and bellen ‘call’ are raised out of the infinitival vp complements of willen ‘want’ and geprobeerd ‘tried’.27


‘that he had not wanted to stop the bloodshed’



‘I tried to call her but ... ’


Similarly, adpositions which introduce a pp-complement are assumed to lack a subj requirement in hpsg and, hence, do not qualify as subject raisers, but they do allow complement raising, as illustrated in Section 4.2. Conversely, there are instances of subject raising which do not involve complement raising. English, for instance, has subject raising verbs and adjectives, but does not allow complement raising, and in languages which allow both, such as Dutch, one can have subject raising without complement raising, as in (108), where the nonreferential subject of the infinitival complement of lijkt ‘seems’ is raised, but where the complement of the infinitive is realized in situ.


‘there seems to be a problem with the water supply’

Third, the non-lexicalist treatment of complement raising is also appropriate for the analysis of clauses which contain both a stranded adposition and a verb cluster, as in (109).


‘that they had to give all their money to that’

The unsaturated comps requirement of aan ‘to’ is added to the comps value of the vp, in conformity with the constraint on complement raising in (56). No extra stipulations are needed to model this.


In the lexicalist treatment of generalized raising, though, there is a problem, since the unsaturated comps requirement of aan has to be included in the comps list of the modal, even though the modal is not a sister of the adposition.

Fourth, the non-lexicalist treatment of complement raising also deals with raising out of pp adjuncts, as in (111).


‘… that we suffered heavy losses because of that’

The unsaturated comps requirement of the adposition (3) is added to that of the verb (2), yielding a vp that selects two nominal complements, as in (112).


This is possible since the constraint on complement raising applies to all headed phrases. In the generalized raising analysis, by contrast, there is a problem, since adjuncts are not lexically selected by their head sister.

Summing up, there are at least four arguments for differentiating subject raising from complement raising. In fact, this list could be extended. Van Eynde and Augustinus (2013), for instance, provides two further arguments, relating to the interaction with binding and passivization, but we leave these aside, since the presentation would lead us too far astray from the central topic of this paper.

5.2 Adposition stranding and extraction

Treatments of adposition stranding in hpsg tend to ignore the distinction between complement raising and complement extraction. Rentier (1993) and Müller (1995), for instance, take it for granted that adposition stranding invariably results from extraction and focus on how it can be implemented. Bouma (2000) does consider the option of differentiating between complement extraction and argument inheritance (which includes complement raising), but then dismisses it and argues for a uniform extraction analysis: “Although argument inheritance plays an important role in the syntax of Dutch verb clusters an approach based on argument inheritance seems highly unlikely for r-pronouns” (Bouma 2000: 69). The dismissal is based on four arguments which we will present and discuss one by one. This discussion is an update of Van Eynde and Augustinus (2014).

First of all, prepositions which do not allow extraction (such as met) cannot be associated with an r-pronoun in the Mittelfeld either. If two different mechanisms are used to account for these two phenomena, such generalizations are easily lost. (Bouma 2000: 69)

Our treatment indeed employs different devices for complement raising and complement extraction, but this does not cause a loss of generalization, since the constraints on these devices have a different range of application: While the constraint on complement raising in (92) extends to head-initial vps and cps, the constraint on complement extraction in (102) does not.

Second, as argument inheritance normally involves the composition of two comps lists, r-pronouns would have to be allowed on comps, even though they can, apart from a few exceptional cases, never appear in a position following the preposition. (Bouma 2000: 69)

Our treatment indeed allows r-pronouns on comps lists. That they never appear in a position following the adposition does not by itself imply that this is an illicit assumption, since many of the Dutch adpositions allow their complement to precede them, as shown in (113).


‘that he drove into the garage backwards’

In fact, the inherently final adpositions, such as af, mee and toe, even require their complement to precede them. Yet, this is not by itself a valid argument for not allowing the np or pp which they select in their comps list. Indeed, given that Dutch has v-final vps and a-final aps, as shown in (114) and (115), it is only natural to assume that it also has p-final pps.


‘that he has a big mouth’


‘that he is fed up with her antics’

There is, hence, nothing anomalous about the presence of r-pronouns on comps lists of adpositions.

Third, the set of argument inheritance verbs must now not only contain auxiliaries and modals, but all verbs which select a (prepositional) complement. Examples such as Kim is er tevreden mee ‘Kim is happy about it’ introduce further complications for an argument inheritance approach, as it suggests that predicative adjectives and nouns must be argument inheritors as well. (Bouma 2000: 69)

This objection is indeed valid for the generalized raising treatment, but not for our treatment, since we do not add the unsaturated comps requirement of the adposition to those of its head sister. Instead, it is propagated directly to the ap and the vp, as illustrated in (116).


The unsaturated comps requirement of mee is shared with the ap tevreden mee and the vp tevreden mee is, but it is not shared with the adjective nor with the copula.

Finally, ... in an argument inheritance approach, the relationship between valence and syntactically realized arguments has to be one-on-one, and thus there is no room for amalgamation of syntactic functions. (Bouma 2000: 69)

An example of amalgamation is (117).


‘There is a prince in it.’

In Gosse Bouma’s analysis er simultaneously fulfills two syntactic functions in (117). It is both the expletive subject of the clause and the complement of the stranded adposition. This amalgamation, he claims, is impossible to model in terms of argument inheritance, since that device does not allow for a one-to-many relation between syntactically realized arguments and valence. While this may be a valid criticism for the argument inheritance approach, it is not necessarily a problem for our treatment.

For a start, notice that in our treatment er has only one syntactic function. It is the expletive subject of the clause, but it is not the complement of the stranded adposition. In fact, it cannot be the complement of the adposition, since the constraints on raising and extraction do not allow this. More specifically, the complement raising barrier in (92) blocks the raising out of a v-initial vp, and the constraint on extraction in (85) blocks the extraction of minor complements. (117) is, hence, not an instance of amalgamation. Instead, we assume that it is an instance of complement omission.28 Evidence for this assumption is provided by the fact that the subject er is expletive, while the non-realized complement of the adposition is not. This is a problem for the amalgamation analysis since it has the undesirable consequence that the same word is required to be expletive and referential. The complement omission analysis does not have this problem, since er has only one function. Another piece of evidence is provided by infinitival clauses that are introduced by the complementizer om ‘for’, as in (118).


‘that I would like to go to’


The unrealized complement of the stranded circumposition is understood to be er ‘it.r’, but there is no overt er elsewhere in the sentence with which it could be claimed to be amalgamated. The omission treatment does not have this problem.

Taking stock, Bouma (2000) lists four arguments against the differentiation between argument inheritance and complement extraction, but closer scrutiny reveals that none of them sticks. Conversely, the discussion of the constraints on complement raising and complement extraction in Section 4.4 has shown that these constraints are sufficienlty different to motivate a treatment which differentiates complement raising from complement extraction.

6 Conclusion

This paper has accomplished three things. First, it has demonstrated that Dutch not only has verb clusters, but also adposition clusters. These are stranded circumpositions which are headed by a postposition and which have two or three members in a rigid linear order. That part is based on Van Eynde (2017). Second, it has shown that stranding and clustering, both of verbs and adpositions, are the result of the same syntactic processes, i.e. complement raising and complement extraction. The analysis of these phenomena is cast in the notation of Head-driven Phrase Structure Grammar, more specifically in the constructional version of Ginzburg and Sag (2000). Third, it has argued that the resulting treatment is more adequate than the existing treatments of verb clustering and adposition stranding in hpsg. The former treat complement raising in the same lexicalist way as subject raising. The latter treat complement raising in the same way as complement extraction. Building on Van Eynde and Augustinus (2013) and Van Eynde and Augustinus (2014) we have argued that these reductions lead to problems and that complement raising had better be differentiated from both subject raising and complement extraction.


Many thanks are due to the editor and the anonymous reviewers of Linguistics. Preliminary versions of this paper were presented at the annual conference of the Linguistic Society of Belgium (Louvain-la-Neuve, May 13, 2016), the European Workshop on Head-driven Phrase Structure Grammar (Paris, March 24–25, 2017) and a seminar of the research group on Formal and Computational Linguistics (Leuven, June 27, 2017). I thank the participants for their comments and suggestions.


1. The queries to retrieve the circumpositions which surround their complement are of the folllowing form:


//node[@cat=“pp” and node[@rel=“hd” and @pt=“vz” and number(@begin) < ../node[@rel=“obj1”]/number(@begin)] and node[@rel=“obj1” and number(@begin) < ../node[@rel=“hdf” and@pt=“vz”]/number(@begin)] and node[@rel=“hdf” and @pt=“vz”]]

The values of @rel are dependency relations, the values of @cat are phrasal categories, and the values of @pt are lexical categories.29 < is the linear precedence relation. It is used to require the first adposition to precede the complement and to require the complement to precede the second adposition. The query was generated automatically by GrETEL on the basis of a relevant example. The query yields 270 hits for lassy and 571 for cgn. 774 of these appear in Table 1. The remaining 67 can be divided in three groups. The largest one (50 tokens) concerns the combinations with op ... na, as used in (120).


‘Africa was ... colonised by European powers, except for two regions’


We do not treat this as a circumpositional pp, since the meaning of na in this combination is not related to that of the adposition na ‘after’, as used in na Pasen ‘after Easter’. Instead it is related to that of the adjective na ‘near, close, dear’, as used in alle kinderen zijn me even na ‘all children are equally dear to me’. This adjective can be declined, as in naë bloedverwanten ‘close.dcl relatives’, and has comparative and superlative counterparts (nader ‘near.cmp’ and naast ‘near.sup’). The combination op ... na is, hence, an ap which is headed by an adjective that selects a pp complement that is introduced by op. Other such adjectives are belust ‘keen’ and gesteld ‘keen’, as used in op wraak belust ‘keen on revenge’ and op luxe gesteld ‘keen on luxury’. A second group involves the combinations with met ... mee (5 tokens) and op ... uit (1 token), as used in (121–122).


‘your partner remigrates together with you’



‘... a mayor that is eager for peace’


They do not qualify as circumpositional pps, since the order of the adpositions may be changed without any effect on well-formedness or meaning, as in (123–124).


‘your partner remigrates together with you’


‘... a mayor that is eager for peace’

The third group (11 tokens) concerns annotation errors and dysfluencies.

2. The queries to retrieve the stranded circumpositions take the form of regular expressions over strings, such as /naar toe/ and /naartoe/. To retrieve those with an incorporated r-pronoun we used more complex expressions, such as /∖S{2,}naar toe/ and /∖S{2,}naartoe/.30 False hits were filtered out manually. They can be divided in three groups. The first consists of sequences in which the second adposition is the separable particle of a verb. In (125), for instance, af is part of the verb afwijken ‘diverge’.


‘the dialects of the surrounding places clearly diverge from it’


The second group consists of sequences in which the second adposition forms a unit with the words that follow it, rather than with the first adposition. This is, for instance, the case when it is a preposition followed by an np, as in (126), or when it is a complementizer followed by a vp, as in (127).


‘Pietje meets Jantje on the camping’



‘he only came out at night in order to go the pub’


The third group consists of sequences in which the adpositions form a compound. This can be a prepositional compound, as in (128), or an adverbial compound, as in (129).


‘from 1882 (onward) Ensor becomes a member of the art group “L’Essor”’



‘in between he found the time to ...’


Prepositional compounds invariably precede their complement. This differentiates them from the circumpositions which surround their complement if the complement is not an r-pronoun. Adverbial compounds do not take a complement.


  • Abeillé, Anne, Dani&#x2018;ele Godard, Philippe Miller & Ivan A. Sag. 1998. French bounded dependencies. In Sergio Balari & Luca Dini (eds.), Romance in HPSG, 1–54. Stanford, CA: CSLI Publications. Google Scholar

  • Augustinus, Liesbeth. 2015. Complement raising and cluster formation in Dutch. Leuven: Katholieke Universiteit Leuven dissertation. Google Scholar

  • Augustinus, Liesbeth & Frank Van Eynde. 2014. Looking for cluster creepers in Dutch treebanks. Computational Linguistics in the Netherlands Journal 4. 149–170. Google Scholar

  • Augustinus, Liesbeth, Vincent Vandeghinste & Frank Van Eynde. 2012. Example-based treebank querying. In Proceedings of the 8th international conference on Language Resources and Evaluation (LREC), 3161–3167. May, 21–27, Istanbul. Google Scholar

  • Bennis, Hans. 1986. Gaps and dummies. Tilburg: Katholieke Hogeschool Tilburg dissertation. Google Scholar

  • Bouma, Gosse. 2000. Argument realization and Dutch R-pronouns. In Ronnie Cann, Claire Grover & Philip Miller (eds.), Grammatical interfaces in HPSG, 51–76. Stanford, CA: CSLI Publications. Google Scholar

  • Bouma, Gosse & Gertjan van Noord. 1998. Word order constraints on verb clusters in German and Dutch. In Erhard Hinrichs (ed.), Complex predicates in nonderivational syntax. New York: Academic Press. Google Scholar

  • Broekhuis, Hans. 2013. Syntax of Dutch: Adpositions and adpositional phrases. Amsterdam: Amsterdam University Press. Google Scholar

  • Chomsky, Noam. 1970. Remarks on nominalization. In Roderick A. Jacobs & Peter S. Rosenbaum (eds.), Readings in English transformational grammar, 184–221. Waltham, MA: Ginn and Company. Google Scholar

  • Engel, Ulrich. 1970. Regeln zur Wortstellung. In Ulrich Engel (ed.), Forschungsberichte des Instituts für deutsche Sprache, 1–148. Tübingen: Gunter Narr. Google Scholar

  • Ginzburg, Jonathan & Ivan A. Sag. 2000. Interrogative investigations. Stanford, CA: CSLI Publications. Google Scholar

  • Haeseryn, Walter, Kirsten Romijn, Guido Geerts, Jaap de Rooij & Maarten C. van den Toorn. 1997. Algemene Nederlandse Spraakkunst. Groningen and Deurne: Martinus Nijhoff and Wolters Plantyn. Google Scholar

  • Hinrichs, Erhard & Tsuneko Nakazawa. 1989. Flipped out: AUX in German. In Papers from the 25th Regional Meeting of the Chicago Linguistic Society, 193–202. Chicago: Chicago University Press. Google Scholar

  • Hinrichs, Erhard & Tsuneko Nakazawa. 1994. Linearizing AUXs in German verbal complexes. In John Nerbonne, Klaus Netter & Carl J. Pollard (eds.), German in Head-driven phrase structure grammar, 11–37. Stanford, CA: CSLI Publications. Google Scholar

  • Hoekstra, Heleen, Michaël Moortgat, Bram Renmans, Machteld Schouppe, Ineke Schuurman & Ton van der Wouden. 2003. CGN Syntactische annotatie. Utrecht and Leuven. Google Scholar

  • Höhle, Tilman. 1986. Der Begriff “Mittelfeld”, Anmerkungen über die Theorie der topologischen Felder. In Albrecht Schöne (ed.), Kontroversen, alte und neue: Akten des 7. Internationalen Germanisten-Kongresses, Göttingen 1985, vol. 3, 329–340. Tübingen: Niemeyer. Google Scholar

  • Höhle, Tilman. 1997. Vorangestellte Verben und Komplementierer sind eine natürliche Klasse. In Christa Dürscheid, Karl Heinz Ramers & Monika Schwarz (eds.), Sprache im Fokus: Festschrift für Heinz Vater zum 65. Geburtstag, 107–120. Tübingen: Niemeyer. Google Scholar

  • Jackendoff, Ray. 1977. X-bar syntax: A study of phrase structure. Cambridge, MA: MIT Press. Google Scholar

  • Monachesi, Paola. 1998. Decomposing Italian clitics. In Sergio Balari & Luca Dini (eds.), Romance in HPSG, 305–357. Stanford, CA: CSLI Publications. Google Scholar

  • Müller, Stefan. 1995. Scrambling in German. Extraction into the Mittelfeld. In Proceedings of the 10th Pacific Asia conference on language, information and computation, 79–83. Hong Kong: City University of Hong Kong. Google Scholar

  • Müller, Stefan. 2002. Complex predicates. Stanford, CA: CSLI Publications. Google Scholar

  • Oostdijk, Nelleke, Wim Goedertier, Frank Van Eynde, Louis Boves, Jean-Pierre Martens, Michaël Moortgat & Harald Baayen. 2002. Experiences from the Spoken Dutch Corpus Project. In M. Gonzalez Rodriguez & C.P. Saurez Araujo (eds.), Proceedings of LREC 3, 340–347. Paris: European Language Resources Association. Google Scholar

  • Oostdijk, Nelleke, Martin Reynaert, Veronique Hoste & Ineke Schuurman. 2013. The construction of a 500-million-word reference corpus of contemporary written Dutch. In Peter Spyns & Jan Odijk (eds.), Essential speech and language technology for Dutch. Results by the STEVIN programme, 219–247. Berlin and Heidelberg: Springer. Google Scholar

  • Pollard, Carl J. & Ivan A. Sag. 1994. Head-driven phrase structure grammar. Stanford, CA and Chicago: CSLI Publications and University of Chicago Press. Google Scholar

  • Rentier, Gerrit. 1993. Dutch clitic pronouns and preposition stranding. In Computational linguistics in the Netherlands. Papers from the Third CLIN Meeting, 109–120. Tilburg: Institute for Language Technology and Artificial Intelligence, Tilburg University. Google Scholar

  • Sag, Ivan A., Thomas Wasow & Emily Bender. 2003. Syntactic theory. A formal introduction, 2nd edn. Stanford, CA: CSLI Publications. Google Scholar

  • Smessaert, Hans, William Van Belle & Ingrid Van Canegem-Ardijns. 2014. Adpositional constructions of location and motion in Dutch. In Nicole Delbecque, Karen Lahousse & Willy Van Langendonck (eds.), Non-nuclear cases, 127–171. Amsterdam and Philadelphia: John Benjamins. Google Scholar

  • Tseng, Jesse. 2005. Directionality and the complementation of Dutch adpositions. Belgian Journal of Linguistics 18. 167–194. Google Scholar

  • Van Eynde, Frank. 1999. Major and minor pronouns in Dutch. In Gosse Bouma, Erhard Hinrichs, Geert-Jan Kruijff & Richard Oehrle (eds.), Constraints and resources in natural language syntax and semantics, 137–151. Stanford, CA: CSLI Publications. Google Scholar

  • Van Eynde, Frank. 2004. Minor adpositions in Dutch. Journal of Comparative Germanic Linguistics 7. 1–58. CrossrefGoogle Scholar

  • Van Eynde, Frank. 2006. NP-internal agreement and the structure of the noun phrase. Journal of Linguistics 42. 139–186. CrossrefGoogle Scholar

  • Van Eynde, Frank. 2015. Predicative constructions. From the Fregean to a Montagovian treatment. Stanford, CA: CSLI Publications. Google Scholar

  • Van Eynde, Frank. 2017. Adposition clusters in Dutch. In Hilke Reckman, Lisa L.S. Cheng, Maarten Hijzelendoorn & Rint Sybesma (eds.), Crossroad semantics. Computation, experiment and grammar, 77–92. Amsterdam and Philadelphia: John Benjamins. Google Scholar

  • Van Eynde, Frank & Liesbeth Augustinus. 2013. Why and how to differentiate complement raising from subject raising in Dutch. In Stefan Müller (ed.), Proceedings of HPSG 2013, 222–242. Stanford, CA: CSLI Publications. Google Scholar

  • Van Eynde, Frank & Liesbeth Augustinus. 2014. Complement raising, extraction and adposition stranding in Dutch. In Stefan Müller (ed.), Proceedings of HPSG 2014, 156–175. Stanford, CA: CSLI Publications. Google Scholar

  • Van Noord, Gertjan, Gosse Bouma, Frank Van Eynde, Daniel de Kok, Jelmer van der Linde, Ineke Schuurman, Erik Tjong Kim Sang & Vincent Vandeghinste. 2013. Large scale syntactic annotation of written Dutch: Lassy. In Peter Spyns & Jan Odijk (eds.), Essential speech and language technology for Dutch, 147–164. Berlin and Heidelberg: Springer. Google Scholar

  • Van Riemsdijk, Henk. 1978. A case study in syntactic markedness. Lisse: The Peter de Ridder Press. Google Scholar

  • Zwarts, Joost. 1997. Complex prepositions and P-stranding in Dutch. Linguistics 35. 1091–1112. Google Scholar


  • 1

    In the terminology of transformational grammar complement raising is a form of scrambling and complement extraction a form of A-bar-movement. 

  • 2

    For the spoken data we use the cgn treebank (version 2.0.1), described in Oostdijk et al. (2002). For the written data we use lassy small (version 1.1), described in Van Noord et al. (2013). Examples which are extracted from these treebanks have a unique identifier. Those from the cgn treebank have the ‘fn’ prefix if they are from the Netherlands and the ‘fv’ prefix if they are from the Dutch speaking part of Belgium (Vlaanderen). 

  • 3

    The format of the queries is spelled out in the Appendix. 

  • 4

    The sum of the relevant hits is 774, which is 67 less than the total number of hits (270 + 571). The difference is explained in the Appendix. 

  • 5

    In the case of aan ... door it may also be due to the fact that it only occurs in the multi-word unit aan één stuk door ‘uninterruptedly’. 

  • 6

    In contrast to the English modals, the Dutch modals are compatible with a predicative complement. See Van Eynde (2015: 213–214) for examples and discussion. 

  • 7

    Notice that (21) violates the constraint that p2 must follow p1. 

  • 8

    The rightmost adposition in (22) is stranded in the verb cluster. For a treatment of cluster creepers with quantitative data from the sample, see Augustinus and Van Eynde (2014). 

  • 9

    The third member of this small class is voor ‘for’, as in voor bij de koffie ‘for with the coffee’. 

  • 10

    The r-form pronouns have to be distinguished from the homophonous locative adverbs. In dat ze niet van hier zijn ‘that they are not from here’, hier is not the r-form of the pronoun dit, but a locative adverb, meaning ‘here’. As the example shows, the locative adverbs follow the adposition, just like ordinary nps. 

  • 11

    Details about the retrieval and the filtering are given in the Appendix. 

  • 12

    (42) is quoted from Smessaert et al. (2014: 164). 

  • 13

    A possible explanation for this difference is given in Van Eynde (2017: 90). 

  • 14

    Technically, these elements are objects of type synsem, i.e. bundles of syntactic and semantic features. Not included in the synsem objects are phonological and pragmatic features. 

  • 15

    This piecemeal addition of the complements is also proposed for German in Müller (2002). 

  • 16

    A stands for any list, including the empty list. stands for the concatenation of lists. 

  • 17

    It is similar to the analysis of the German verb clusters in Hinrichs and Nakazawa (1989) and Hinrichs and Nakazawa (1994), but there are also some non-trivial differences, see Section 5.1. 

  • 18

    In signs of type non-headed phrase, such as coordinate phrases, the comps list of the mother is identified with the comps lists of each of the conjunct daughters separately, rather than with the concatenation of those lists. In he buys and sells cars, for instance, the coordinate phrase buys and sells has the same comps list as its conjunct daughters buys and sells. 

  • 19

    Empirical evidence for treating extraction differently from raising will be given in section 4.4. 

  • 20

    The constraint will be slightly modified in section 4.4.2. stands for list subtraction. Besides subj and comps, Ginzburg and Sag (2000) uses a third valence feature, called spr, for the selection of specifiers. Nouns, for instance, are claimed to select a determiner as their specifier. We do not use this feature, since we adopt the functor treatment of determiners, as developed in Van Eynde (2006). In that treatment, it is the specifier which selects its head sister, rather than the head which selects its specifier. 

  • 21

    stands for disjoint set union. 

  • 22

    It might make sense to restrict this to headed phrases, since non-headed phrases may be unsaturated for comps. The bracketed coordinate phrase in he [buys and sells] cars, for instance, has a non-empty comps list. 

  • 23

    The assumption that complementizers share properties with verbs is also made for other languages. Höhle (1997) provides evidence for German and Ginzburg and Sag (2000: 23) for English: “the part of speech types (the values of the feature head) associated with verbs and complementizers are subtypes of a common supertype” (p. 23). 

  • 24

    Zwarts (1997: 1094–1095) lists no less than 57 of these adpositions. 

  • 25

    The notion of non-nominal part of speech is inspired by x-bar theory, in which the lexical categories are analyzed in terms of the boolean features n and v. In this analysis nouns are [+n, –v], adjectives are [+n, +v], verbs are [–n, +v], and adpositions are [–n, –v], see Chomsky (1970) and Jackendoff (1977). 

  • 26

    This constraint does not hold for English, as is clear from the wellformedness of (88). 

  • 27

    The distinction between subject control verbs and subject raising verbs is discussed amongst others in Pollard and Sag (1994: 132–145) and Sag et al. (2003: 364–376). 

  • 28

    A similar assumption is made in the transformational treatment of Bennis (1986). 

  • 29

    vz is short for voorzetsel, the Dutch term for adposition. 

  • 30

    ∖S{2,} stands for a sequence of two or more non-white spaces. We choose for non-white spaces rather than for alpha-numeric characters in order to include combinations with d’r. 

About the article

Published Online: 2019-06-26

Published in Print: 2019-09-25

Citation Information: Linguistics, Volume 57, Issue 5, Pages 1025–1071, ISSN (Online) 1613-396X, ISSN (Print) 0024-3949, DOI: https://doi.org/10.1515/ling-2019-0023.

Export Citation

© 2019 Walter de Gruyter GmbH, Berlin/Boston.Get Permission

Comments (0)

Please log in or register to comment.
Log in