Looking for Cluster Creepers in Dutch Treebanks. Dat we ons daar nog kunnen mee bezig houden.

Abstract

In Dutch V-final clauses the verbs tend to form a cluster which cannot be split up by nonverbal material. However, Haeseryn et al. (1997) as well as other studies on the phenomenon list several cases in which the verb cluster may be interrupted by cluster creepers. The most common examples are constructions with separable verb particles, but examples with nouns, adjectives, and adverbs are attested as well. Since the majority of the data in previous studies is collected by introspection and elicitation, it is interesting to compare those findings to corpus data. The corpus analysis is based on data from two Dutch treebanks (CGN and LASSY), which allow to take into account regional and/or stylistic variation. This is an important aspect for the analysis, since cluster creeping is reported to be a typical property of spoken and regional variants of Dutch. The goal of this corpus-based investigation is on the one hand to provide insight in the frequency of the phenomenon, and on the other hand to classify the types of cluster creepers. Besides the linguistic analysis, methodological issues regarding the extraction of the relevant data from the treebanks will be addressed as well.

Full text

Computational Linguistics in the Netherlands Journal 4 (2014) 149-170 Submitted 06/2014; Published 12/2014
Looking for Cluster Creepers in Dutch Treebanks.
Dat we ons daar nog kunnen mee bezig houden.
Liesbeth Augustinus liesbeth@ccl.kuleuven.be
Frank Van Eynde frank@ccl.kuleuven.be
Centre for Computational Linguistics, KU Leuven, Belgium
Abstract
In Dutch V-final clauses the verbs tend to form a cluster which cannot be split up by nonverbal
material. However, Haeseryn et al. (1997) as well as other studies on the phenomenon list several
cases in which the verb cluster may be interrupted by cluster creepers. The most common examples
are constructions with separable verb particles, but examples with nouns, adjectives, and adverbs
are attested as well.
Since the majority of the data in previous studies is collected by introspection and elicitation,
it is interesting to compare those findings to corpus data. The corpus analysis is based on data
from two Dutch treebanks (CGN and LASSY), which allow to take into account regional and/or
stylistic variation. This is an important aspect for the analysis, since cluster creeping is reported
to be a typical property of spoken and regional variants of Dutch.
The goal of this corpus-based investigation is on the one hand to provide insight in the frequency
of the phenomenon, and on the other hand to classify the types of cluster creepers. Besides the
linguistic analysis, methodological issues regarding the extraction of the relevant data from the
treebanks will be addressed as well.
1. Introduction
1.1 Dutch Clause Structure
There are two fixed positions in the Dutch sentence. Those positions are known as poles. In verb-
initial clauses, such as example (1a), the finite verb heeft ‘has’ occupies the first pole, while the past
participle gedronken ‘drunk’ is in the second pole.1In subordinate clauses, such as example (1b),
the complementizer dat ‘that’ takes up the first pole, while the verbal elements beschouwd wordt
‘is considered’ occupy the second pole. (1b) shows that the second pole may consist of multiple
elements, but it can also be empty, as in example (1c) (Haeseryn et al. 1997, pp.1225-1226).
(1) a. Z’n
his
broer
brother
heeft
has
altijd
always
al
already
graag
gladly
een
a
glas
glas
bier
beer
gedronken.
drunk
‘His brother has always enjoyed a glass of beer.’
b. (Het
it
blijkt)
seems
dat
that
hij
he
zowat
almost
overal
everywhere
ter
in the
wereld
world
als
as
een
an
autoriteit
authority
beschouwd
considered
wordt.
is
‘(It seems) that he is considered to be an authority almost all over the world.’
c. Z’n
his
broer
brother
drinkt
drinks
graag
gladly
een
a
glas
glass
bier.
beer
‘His brother likes to drink a glass of beer.’
The poles divide sentences into topological fields: The voorveld is the part before the first pole, the
middenveld is the part between the poles, and the naveld is the part after the second pole.
The sequence of verbs in the second pole is called the werkwoordelijke eindgroep ‘lit: verbal
end group’ or verb cluster. The nonverbal elements appear before or after the second pole (2a-b).
1. Dutch verb-initial clauses comprise V-first and V-second clauses.
c
2014 Augustinus and van Eynde.

The intrusion of nonverbal material in the verb cluster, as in (2c), is considered ungrammatical.
Canonically, nonverbal elements are not allowed in the verb cluster (Haeseryn et al. 1997, p.1355).
(2) a. (Hij
he
beweerde)
claimed
dat
that
hij
he
het
it
gisteren
yesterday
aan
to
de
the
leraar
teacher
had
had
verteld.
told
b. (Hij
he
beweerde)
claimed
dat
that
hij
he
het
it
gisteren
yesterday
had
had
verteld
told
aan
to
de
the
leraar.
teacher
‘(He claimed) that he told the teacher about it yesterday.’
c. * (Hij
he
beweerde)
claimed
dat
that
hij
he
het
it
gisteren
yesterday
had
had
aan
to
de
the
leraar
teacher
verteld.
told
1.2 Cluster Creepers
Although the impenetrability of the verb cluster is the norm in most constructions, there are some
exceptions. Example (3) shows a construction in which the verb cluster is interrupted by the adjective
schuldig ‘guilty’, but which is nonetheless well-formed.
(3) De
The
verdachte
suspect
ontkent
denies
tot
until
op
at
heden
present
zich
himself
aan
to
zwendel
fraud
te
to
hebben
have
schuldig
guilty
gemaakt.
made
‘Until today the suspect denies being guilty of fraud.’
According to Haeseryn et al. (1997), instances of cluster creeping occur more often in Belgian Dutch
compared to Dutch spoken in the Netherlands.
1.2.1 A typology of cluster creepers
Haeseryn et al. (1997) mention three types of cluster creepers:
1. The most typical cluster creepers are inherent parts of the verb phrase, such as predicative
adjectives and nonverbal parts of idiomatic expressions. Usually, those elements occur just
before the second pole, as in (4a), but they can also be included in the verb cluster, as in (4b)
(Haeseryn et al. 1997, p.1358). Note that (3) is also an example of this type.
(4) a. ...
...
dat
that
hij
he
zich
himself
niet
not
bang
afraid
zal
will
laten
let
maken.
make
b. ...
...
dat
that
hij
he
zich
himself
niet
not
zal
will
laten
let
bang
afraid
maken.
make
‘... that he will not be frightened.’
2. A second category of cluster creepers consists of stranded adpositions, often being the second
part of pronominal adverbs. Canonically those adpositions are realised before the verb cluster
(5a), but they may also occur within the cluster (5b).
(5) a. ...
...
dat
that
hij
he
daar
there
nog
still
aan
on
moet
must
denken.
think
b. ...
...
dat
that
hij
he
daar
there
nog
still
moet
must
aan
on
denken.
think
‘... that he still needs to think about that.’
This type of adposition stranding within the cluster is considered typical of Belgian Dutch
(Haeseryn et al. 1997, p.1362).
3. A third type that is also typical of Belgian Dutch but less common than adposition stranding
is cluster creeping by an object or an adverbial modifier (Haeseryn et al. 1997, p.1362):
150

(6) a. ...
...
dat
that
de
the
Rode
Red
Duivels
Devils
nog
still
twee
two
doelpunten
goals
moeten
must
scoren.
score
b. ...
...
dat
that
de
the
Rode
Red
Duivels
Devils
nog
still
moeten
must
twee
two
doelpunten
goals
scoren.
score
‘... that the Red Devils still need to score two goals.’
Haegeman and van Riemsdijk (1986) discuss several constructions for West-Flemish, a regional
variant of Dutch spoken in Belgium, such as (7a). Most speakers consider the corresponding
construction in (Standard) Dutch ungrammatical (7b). What differentiates (7b) from (6b) is
the presence of a determiner: While cluster creeping by bare nominals is more common, NPs
with a determiner are rarely used in the verb cluster.
(7) a. WF . . . da
. . . that
Jan
Jan
wilt
wants
een
a
hus
house
kopen.
buy
b. DU * . . . dat
. . . that
Jan
Jan
wil
wants
een
a
huis
house
kopen.
buy
‘. . . that Jan wants to buy a house.’
Besides genuine cases of cluster creeping, Haeseryn et al. (1997) mention several constructions
that look like cluster creeping but should not be treated as such. For example, separable verb
particles (svps) are not considered as cluster creepers if they occur within the verb cluster. They
argue that in the case of svps, constructions in which the svp is realised in front of the verb cluster
(8a) are less preferred than constructions in which the svp is realised within the cluster (in front of
the main verb or as a part of it), as in (8b) (Haeseryn et al. 1997, pp.1357-1358).2
(8) a. ...
...
dat
that
hij
he
haar
her
op
up
moet
must
bellen.
call
b. ...
...
dat
that
hij
he
haar
her
moet
must
opbellen.
up-call
‘... that he must call her.’
The fact that Haeseryn et al. (1997) do not treat svps as real cluster creepers as opposed to inher-
ent sentence parts leads to classification problems, since the distinction between svps and inherent
parts of the sentence is often hard to draw (Haeseryn et al. 1997, p.1359). Consider for exam-
ple koffiedrinken ‘drink coffee’ versus champagne drinken ‘drink champagne’. Are those examples
separable verbs or regular combinations of a verb and a noun?
In order to avoid this uncertainty, we will treat both svps and inherent parts of the verb phrase
as cluster creepers, which is in line with amongst others Evers (2003) and Wurmbrand (2005).
1.2.2 Position of the cluster creepers
Cluster creeping is only possible if the main verb does not occur at the front of the cluster, since
the nonverbal element cannot occur after the main verb, as shown in (9).
(9) a. * . . . dat
. . . that
hij
he
gedronken
drunk
koffie
coffee
heeft.
has
Intended: ‘. . . that he has drunk coffee.’
b. * . . . dat
. . . that
hij
he
drinken
drink
koffie
coffee
wil.
wants
Intended: ‘. . . that he wants to drink coffee.’
2. Haeseryn et al. (1997) consider constructions like (8a) typical of spoken (Netherlandic) Dutch.
151

Therefore, cluster creeping occurs more often in infinitive constructions than in constructions with a
participle, since infinitives are usually realised at the end of the verb cluster, as opposed to participles
(Haeseryn et al. 1997, pp.1355-1356). See also Hoekstra (2010, pp.178-179) for a discussion on the
relation between verb order within the cluster and cluster creeping.
The canonical position of a cluster creeper is just before the main verb, but in clusters with
more than two verbs it may also occur more to the front of the verb cluster, as in (10) (Haeseryn
et al. 1997, p.1357).
(10) ...
...
dat
that
hij
he
haar
her
had
had
op
up
moeten
must
bellen.
call
‘... that he had to call her.’
2. Goals
The methodology used for this research is corpus-based, in the sense that treebanks (i.e. syntactically
annotated corpora) will be used to verify the claims about cluster creeping. This corpus-based
investigation will provide insight in the frequency of the phenomenon, which makes it possible to
compare constructions that are theoretically possible to the constructions that are actually used.
More specifically, we will classify the types of cluster creepers according to their syntactic function
and their phrasal category or part-of-speech (pos) in order to investigate whether the types of cluster
creepers mentioned in Haeseryn et al. (1997) are reflected in the corpus data, or whether the data
reveal other categories, aiming at a more complete description of the possible cluster creepers in
Dutch.
Furthermore, we will consider the occurrence of cluster creepers in spoken versus written lan-
guage, as well as their occurrence in clusters containing participles versus clusters with infinitives.
Specifically, we aim at extracting and investigating corpus examples like the following:
(11) a. we
we
hebben
have
zo
so
nog
another
ne
a
politieker
politician
die
that
ons
us
daar
there
altijd
always
ook
also
doet
does
aan
on
denken.
think
‘we have another polititian of that kind who always reminds us of that. [CGN, fvc701156_222]’
b. . . . aan
. . . to
iedereen
everyone
die
that
toen
then
de
the
toekomst
future
van
of
dit
this
land,
country
van
of
de
the
huidige
present
en
and
toekomstige
future
generaties
generations
hebben
has
veilig
save
gesteld.
put
‘. . . to everyone who back then has saveguarded the future of this country, of the current
and future generations.’ [LASSY, dpc-vhs-000745-nl-sen.p.13.s.3]
Section 3 gives a formal definition of the concept ‘verb cluster’, and it provides an overview of
the relevant constructions, i.e. constructions containing a verb cluster (which may contain cluster
creepers). Section 4 describes the two treebanks used for the corpus study (CGN and LASSY).
Section 5 explains how the queries are constructed in order to extract the relevant constructions.
Section 6 presents and discusses the results of the treebank investigation. Those results largely
confirm the claims made in Haeseryn et al. (1997), but they also contain a surprise, i.e. multiple
cluster creepers, as in (12).
(12) ...
...
dat
that
we
we
ons
us
daar
there
nog
still
kunnen
can
mee
with
bezig
busy
houden.
keep
‘... that we can still keep ourselves busy with that.’
Section 7 sums up the conclusions and points out some topics for future research.
152

3. Defining the verb cluster
In order to retrieve constructions with a cluster creeper, we first have to define precisely what a verb
cluster is. Generalizing from the examples in section 1.1, we define a verb cluster as a sequence of two
or more verbs in the second pole of the clause. The sequence is ordered in two ways. One concerns
the order of selection. In zou hebben gedronken ‘would have drunk’, for instance, the finite modal
auxiliary zou selects a bare infinitive, i.e. hebben ‘have’, which in turn selects a past participle, i.e.
gedronken ‘drunk’. The last verb in this chain is the ‘main’ verb. (13) defines the selection order in
the cluster in general terms.3
(13) (Vfinite) (V(te )inf )∗(Vpsp)
In words, a cluster has at most one finite verb (modulo coordination), followed by 0, 1 or more
bare and/or te infinitives, followed by at most one past participle (modulo coordination). Table 1
provides some examples of verb clusters. In verb-initial clauses (Vinitial) the cluster only contains
non-finite forms, since the finite verb is in the first pole. In verb-final clauses (Vf inal) the cluster
also contains the finite verb. The ‘main’ verb is a past participle, a bare infinitive or a te-infinitive.
The clusters are in italics.
Vinitial Vfinal
Past Part inf+psp finite+inf+psp
Hij zou gisteren koffie hebben gedronken. . . . dat hij gisteren koffie zou hebben gedronken.
‘He would have drunk coffee yesterday.’ ‘. . . that he would have drunk coffee yesterday.’
Bare inf inf+inf finite+inf+inf
Hij zal morgen koffie willen drinken. . . . dat hij morgen koffie zal wil len drinken.
‘He will drink coffee tomorrow.’ ‘. . . that he will drink coffee tomorrow.’
te-inf inf+te-inf finite+inf+te-inf
Hij heeft gisteren koffie proberen te drinken. . .. dat hij gisteren koffie heeft proberen te drinken.
‘He has tried to drink coffee yesterday.’ ‘. . . that he has tried to drink coffee yesterday.’
Table 1: Verb clusters
The second way in which the sequences are ordered is the linear order. This order canonically
coincides with the order of selection, as in zou hebben gedronken ‘would have drunk’ and the other
examples in Table 1. Alternative orders are also possible, though. The finite verb may also occur
as the last element in the cluster, as in gedronken hebben zou. The past participle can occupy any
position within the cluster, e.g. zou gedronken hebben ‘would drunk have’, gedronken zou hebben
‘drunk would have’.
Clauses with a te-infinitive are tricky, since this infinitive may either be the last member of the
cluster or the first member of the naveld. The former is invariably the case if its selector is a so-
called Infinitivus Pro Participio (ipp), i.e. an infinitive which is selected by the perfect auxiliary.4
Relevant examples are given in the last row of Table 1. Notice that the selector of the te-infinitive
is the ipp proberen, which in its turn is selected by the auxiliary of the perfect. These examples can
be contrasted with those in (14), where the perfect auxiliary is combined with the (expected) past
participle.
(14) a. Hij
he
heeft
has
gisteren
yesterday
geprobeerd
tried
koffie
coffee
te
to
drinken.
drink
‘He has tried to drink coffee yesterday.’
3. For a comprehensive list of the verbs which can take a nonfinal position in the cluster, see Augustinus and
Van Eynde (2012).
4. The name IPP captures the fact that such auxiliaries normally require a past participle.
153

b. . . . dat
. . . that
hij
he
gisteren
yesterday
heeft
has
geprobeerd
tried
koffie
coffee
te
to
drinken.
drink
‘. . . that he has tried to drink coffee yesterday.’
In these clauses the te-infinitive is in the naveld.
Independent evidence for the distinction is provided by the reordering possibilities. While te-
infinitives which are part of the verb cluster may appear in other positions than the last one, the
te-infinitives in the naveld must follow those which are part of the cluster.
(15) a. . . . dat
. . . that
hij
he
gisteren
yesterday
koffie
coffee
proberen
try
te
to
drinken
drink
heeft.
has
‘. . . that he has tried to drink coffee yesterday.’
b. * . . . dat
. . . that
hij
he
gisteren
yesterday
koffie
coffee
geprobeerd
tried
te
to
drinken
drink
heeft.
has
This criterion is also applicable to combinations in which the selector of the te-infinitive is a finite
verb, as in (16).
(16) a. . . . dat
. . . that
hij
he
koffie
coffee
probeerde
tried
te
to
drinken.
drink
‘. . . that he tried to drink coffee.’
b. * . . . dat
. . . that
hij
he
koffie
coffee
te
to
drinken
drink
probeerde.
tried
The ungrammaticality of the second clause shows that the te-infinitive is in the naveld.
4. Data set
For the corpus study we use the CGN Treebank and LASSY Small. Those treebanks for respectively
spoken and written Dutch each contain ca. one million tokens. As the corpora are more or less equal
in size, they are suited for comparing written to spoken language data.
4.1 CGN
The Corpus Gesproken Nederlands (CGN) (Oostdijk et al. 2002) is an annotated corpus of spoken
Dutch.5It consists of recorded speech which is orthographically transcribed, resulting in a corpus of
ca. ten million words, of which one million is syntactically analysed. That syntactically annotated
part of CGN will be referred to as the CGN treebank.
Two thirds of the corpus data consists of Dutch spoken in the Netherlands, whereas one third
of the data comprises Dutch spoken in Flanders, the Dutch speaking part of Belgium. The corpus
contains both dialogues and monologues, and is further divided into specific genres. The division into
subcorpora allows to investigate stylistic variation (e.g. by comparing spontaneous conversations to
news reports), as well as regional variation (by comparing Dutch spoken in Belgium to Dutch spoken
in the Netherlands).
4.1.1 Contents
Table 2 presents the contents of the CGN treebank. The label nis used to refer to the Dutch data,
while the label vrefers to the Flemish data. The labels ato orefer to the different types of speech
that the corpus comprises. The parts ato hcontain dialogues, whereas the parts ito oconsist of
monologues. # Sentences refers to the number of sentences (or utterances) in each subcorpus; #
Words refers to the number of words (excluding punctuation).
5. http://lands.let.ru.nl/cgn
154

Components # Sentences # Words # Sentences # Words # Sentences # Words
N V TOTAL
A. Spontaneous conversations 50,239 302,828 22,881 147,418 73,120 450,246
(’face-to-face’)
B. Interviews with teachers of Dutch 2,484 25,724 4,289 34,158 6,773 59,882
C. Telephone conversations 11,649 70,084 3,142 19,984 14,791 90,068
(recorded via a switchboard)
D. Telephone conversations 0 0 929 6,309 929 6,309
(recorded on MD)
E. Simulated business negotiations 3,123 25,524 0 0 3,123 25,524
F. Interviews/discussions/debates 6,290 75,167 2,617 25,122 8,907 100,289
(broadcast)
G. (Political) discussions/debates/ 1,166 25,125 543 9,009 1,709 34,134
meetings (non-broadcast)
H. Lessons recorded in the classroom 3,064 26,004 1,395 10,116 4,459 36,120
I. Live (sports) commentaries 2,251 25,002 1,026 10,147 3,277 35,149
(broadcast)
J. Newsreports (broadcast) 2,259 25,084 536 7,686 2,795 32,770
K. News (broadcast) 1,923 25,353 558 7,306 2,481 32,659
L. Commentaries/columns/reviews 1,857 25,082 601 7,431 2,458 32,513
(broadcast)
M. Ceremonious speeches/sermons 444 5,190 107 1,894 551 7,084
N. Lectures/seminars 593 14,921 701 8,159 1,294 23,080
O. Read speech 0 0 3,256 44,144 3,256 44,144
Complete corpus 87,342 671,088 42,581 338,883 129,923 1,009,971
Table 2: Contents of the CGN treebank
The word and sentence counts in Table 2 are based on the CGN Treebank version 2.0.1, converted
to the Alpino-XML data format.6
Each sentence in the corpus has a unique identifier, e.g. [fva400392_6] for the sentence in (17).
(17) awel
well
’k
I
ga
go
ne
a
keer
time
een
a
typisch
typical
voorbeeld
example
geven.
give
‘well, I’ll give a typical example.’ [CGN, fva400392_6]
The sentence ID refers to the origin of the fragment (in this case V, for the Flemish part), the
component (in this case A, for the subcorpus containing spontaneous conversations), the fragment
number (400392), and the sentence number (6).7
4.1.2 Linguistic annotations
The CGN Treebank contains pos tags (Van Eynde 2004) as well as syntactic annotations (Hoekstra
et al. 2003). The resulting syntactic structures can be represented as tree structures, cf. Figure 1.
6. http://www.let.rug.nl/vannoord/Lassy/alpino_ds.dtd
7. In the official release, it is not encoded in the identifier whether the sentence occurs in the Dutch or the Flemish
data. This information was added afterwards (based on the information in the corpus).
155

Figure 1: Tree representation of a CGN sentence (fva400392_6)
The annotations of the CGN treebank are manually corrected, which makes the treebank a
high-quality resource for linguistic research. The annotations on sentence level have an accuracy of
97.53% (Fersøe et al. 2006).
4.2 LASSY
The LASSY treebank (Large Scale Syntactic Annotation of written Dutch) (van Noord et al. 2013)
is a corpus of syntactically annotated sentences.8The project resulted in the construction of two
treebanks: LASSY Small and LASSY Large. For the purpose of this research LASSY Small is used,
since it is complementary to the CGN treebank.
4.2.1 Contents
LASSY Small is a one million word corpus of written Dutch. Table 3 provides an overview of the
contents of the LASSY Small treebank.
The word and sentence counts in the table are based on version 1.1 of the LASSY Small treebank.
Each sentence in the corpus has a unique ID, e.g. [dpc-bal-001239-nl-sen.p.15.s.2] for the sentence
in (18).
(18) Laat
let
ik
I
een
an
voorbeeld
example
geven.
give
‘Let me give an example.’ [LASSY, dpc-bal-001239-nl-sen.p.15.s.2]
The sentence ID refers to the subcorpus (in this case DPC-bal), the text number (001239), and the
location within the text (page 15 sentence 2). The division into subcorpora allows to investigate
stylistic variation (e.g. by comparing newspaper articles to law texts).
8. http://www.let.rug.nl/~vannoord/Lassy
9. Paulussen et al. (2006), http://www.kuleuven-kulak.be/DPC
156

Treebank Contents # Sentences # Words
DPC Dutch side of the Dutch Parallel Corpus [dpc]911,716 193,029
Wikipedia Dutch Wikipedia pages [wiki] 7,341 83,360
WR-P-E E-magazines [WR-P-E-C], news letters [WR-P-E-E], 14,420 232,631
Teletext pages [WR-P-E-H], Web sites [WR-P-E-I],
Wikipedia pages [WR-P-E-J]
WR-P-P Books [WR-P-P-B], brochures [WR-P-P-C], guides 17,691 281,424
and manuals [WR-P-P-E], law texts [WR-P-P-F],
newspapers [WR-P-P-G], periodicals and
magazines [WR-P-P-H], policy documents [WR-P-P-I],
proceedings [WR-P-P-J], reports [WR-P-P-K],
surveys [WR-P-P-L]
WS-U auto cues [WS-U-E-A], news scripts [WS-U-T-A], 14,032 184,611
texts for the visually impaired [WS-U-T-B]
LASSY Small Complete treebank 65,200 975,055
Table 3: Contents of LASSY Small
4.2.2 Linguistic annotations
LASSY Small is manually corrected after automatic parsing with the Alpino parser (van Noord
2006),10 a dependency parser for Dutch. The general lay-out of the treebank is very similar to the
CGN treebank, as it contains the same pos tags, and almost the same syntactic annotations (van
Noord et al. 2011). The main annotation difference is the use of indexed nodes, as illustrated in
Figure 2. Since ik ‘I’ is both the subject of laten ‘let’ and the embedded verb geven ‘give’, it is also
included as the subject of the verbal complement (vc) in the form of an index node.
Figure 2: Tree representation of a LASSY sentence (dpc-bal-001239-nl-sen.p.15.s.2)
10. http://www.let.rug.nl/vannoord/alp/Alpino
157

Because of the corrections, LASSY Small is a high-quality resource: The annotations on sentence
level have an accuracy of 97.8%; the accuracy of the syntactic annotations on node level is 99.8%
(Jongejan et al. 2011).
5. Querying the treebanks using GrETEL
Both the CGN Treebank and LASSY can be queried with XPath, a W3C standard query language
for XML trees.11 This can be done using the GrETEL search engine.12 In this search tool the user
has two ways of entering a syntactic query. The first approach is called Example-based Querying
(Augustinus et al. 2012, Augustinus et al. 2013) which consists of a query procedure in several steps,
starting from a natural language example and resulting in an automatically generated XPath query,
which is then used to query the treebanks. The matching sentences are returned to the user, who
can inspect them in more detail. The second approach consists of directly formulating an XPath
query that describes the syntactic pattern the user is looking for, which is then processed in the
same way as in the first approach.
For the research presented here, we started off from XPath queries generated using the example-
based method, which were then manually refined by adding more constraints in order to look for
more specific constructions. For example, the input construction in (19) was used to automatically
derive the query in (20a).13 (20b) is a visual representation of the query, i.e. a subtree of the parsed
example in (19).
(19) ...
...
dat
that
hij
he
koffie
coffee
wil
want
drinken.
drink
‘... that he wants to drink coffee.’
(20) a. //node[@cat="ssub" and
node[@rel="hd" and @pt="ww"] and
node[@rel="vc" and @cat="inf" and
node[@rel="hd" and @pt="ww"] ] ]
b. ssub
hd
ww
vc
inf
hd
ww
11. http://www.w3.org/TR/xpath
12. Greedy Extraction of Trees for Empirical Linguistics, http://nederbooms.ccl.kuleuven.be/eng/gretel
13. In order to derive the XPath query, we indicated pos for the verbs in the example sentence in the GrETEL engine.
158

The query in (20a) extracts V-final constructions (ssub) with a verb (ww) as head daughter (hd)
and a verbal complement (vc) in the form of a bare infinitive (inf). The XPath engine does not
take into account the order of the nodes; for the query in (20a) it also returns constructions in which
the verb follows the infinitive.
Note that the XPath engine performs a greedy search,14 i.e. queries like (20a) do not only
return constructions where a finite verb and a bare infinitive cluster in the second pole, but also
the constructions where another element intervenes between the finite verb and the second pole. So
constructions like the ones in (21) are included as well.15
(21) ssub
hd
ww
X vc
inf
hd
ww
ssub
hd
ww
vc
inf
X hd
ww
The XPath expressions can be further specified or generalized by adding or removing constraints.
For example, by adding the constraint @wvorm="pv" (for ‘persoonsvorm’) to the node of the selecting
verb, we state that the selecting verb should be a finite form. Greedy search furthermore means that
the query in (20b) returns all matches containing at least a verb and a bare infinitive, so it will also
return constructions with more than two verb forms. In order to keep control of the cluster length,
we added a constraint stating that the vc node should have no more than one verbal daughter, using
the not()-function. The resulting query is shown in (22).
(22) //node[@cat="ssub" and
node[@rel="hd" and @pt="ww" and @wvorm="pv"] and
node[@rel="vc" and @cat="inf" and
node[@rel="hd" and @pt="ww"] and
not(node[@rel="vc" and (@cat="inf" or @cat="ti" or @cat="ppart" or
@pt="ww")]) ] ]
The queries above look for non-terminal vc nodes, i.e. the vc nodes containing more than one
daughter, e.g. a verb and a direct object, such as koffie drinken ‘drink coffee’ in (19). If the vc
node consists of one word, it is represented as a terminal node in the treebank. To retrieve the
constructions with terminal vcs, the query in (23a) is used.16 The query tree is presented in (23b).
(23) a. //node[@cat="ssub" and
node[@rel="hd" and @pt="ww" and @wvorm="pv"] and
node[@rel="vc" and @pt="ww" and @wvorm="inf"] ]
b. ssub
hd
ww (pv)
vc
ww (inf)
The queries in (22) and (23a) were used to extract constructions of the type ‘Vfinal, finite +
infinitive’, i.e. category f in Table 4 (see infra). The other clustering constructions are found by
means of adaptations and extensions of the queries presented in this section. Constructions with
14. The notion greedy is used in a similar way as pattern matching with regular expressions, see a.o. Jurafsky and
Martin (2009, p.56); XPath expressions are greedy in the sense that they match with as much of a tree pattern
as they can.
15. ‘X’ stands for any sequence of nodes that may occur in that position.
16. The not() condition need not be stated here, since the terminal vc node cannot have any embedded vcs.
159

more than two verb forms in the cluster have another vc node embedded under the vc. V-initial
constructions can be retrieved by changing the label ssub to smain for V-second clauses or to sv1 for
V-first clauses. The label ppart is used for non-terminal past participles (i.e. participial phrases),
whereas the Dutch label vd (for ‘voltooid deelwoord’) is used for terminal nodes.
For example, the query in (24) returns V-initial constructions with a finite verb, a bare infinitive
and a past participle, i.e. category d in Table 4 (see infra).17
(24) //node[(@cat="smain" or @cat="sv1") and
node[@rel="hd" and @pt="ww" and @wvorm="pv"] and
node[@rel="vc" and @cat="inf" and
node[@rel="hd" and @pt="ww"] and
node[@rel="vc" and @cat="ppart" and
node[@rel="hd" and @pt="ww"] ] ] ]
6. Results
6.1 Identifying the clusters
Even though the treebank annotations do not contain a separate tag for clustering verbs, it is possible
to automatically extract clustering constructions using the relevant queries (see section 5). Table 4
presents the treebank counts for the constructions with at least two verb forms in the cluster. For
each construction, the total number of occurrences is the sum of the queries for non-terminal vcs
and terminal vcs.
As motivated in section 1.2.2, we want to separate the constructions that potentially contain
cluster creepers from the constructions that do not. Since cluster creeping is excluded in construc-
tions in which the main verb occurs at the beginning of the cluster, the results were split up into
two categories: Clusters in which the main verb is not the first verb in the cluster (mv 6=1), and
clusters in which it is (mv = 1).
In section 3 it was already mentioned that constructions with te infinitives are not necessarily
clustering. Those constructions can be split up into constructions where the te-infinitive is a part of
the cluster, as in (25), and constructions in which it is not, as in (26) and (27).
(25) ’k
I
ben
am
blij
happy
dat
that
ik
I
zo
so
veel
much
belangstelling
interest
heb
have
weten
known
te
to
wekken.
raise
‘I am glad that I have been able to raise so much interest.’ [CGN, fnf007126_142]
(26) en
and
ik
I
denk
think
dat
that
men
one
daarin
there-in
moet
has
trachten
try
het
the
juiste
right
evenwicht
balance
te
to
zoeken.
search
‘and I think that one has to try to find the right balance in that.’ [CGN, fvg600012_38]
(27) Nu
now
pas
only
kunnen
can
de
the
bedrijven
companies
proberen
try
wat
something
terúg
back
te
to
verdienen.
gain
‘Only now the companies can try to gain something back.’[LASSY, WS-U-E-A-0000000042.p.31.s.8]
In constructions with ipp, such as (25), the te-infinitive is part of the verb cluster. In (26) the
cluster consists of a finite verb and a bare infinitive, whereas the te-infinitive is in the naveld. (26)
has thus the same type of cluster as the constructions in category f. (27) does not contain a verb
cluster: the verb proberen ‘try’ is the only verb in the second pole, whereas the te-infinitive is in
the naveld. Since we are interested in constructions with at least two verbs in the second pole, such
constructions were removed from the data set.
17. Also in this case the not() function need not be stated. As the past participle is the last element of the cluster,
it does not matter whether it has any embedded (extraposed) vc nodes.
160

Since constructions with a te-infinitive in the naveld are tagged similarly to clustering construc-
tions (i.e. both constructions received a vc tag in the treebanks), we have limited the set of clustering
constructions containing a te-infinitive to the set of ipp constructions (containing at most one te-
infinitive), as those constructions are always clustering.
cgn lassy
Cluster Type mv 6=1 mv = 1 Sum (#) mv 6=1 mv = 1 Sum (#)
a) Vfinal, finite + past part 1664 1626 3290 3544 1519 5063
b) Vfinal, finite + inf + past part 152 138 290 443 262 705
c) Vfinal, finite + inf + inf + past part 11 11 22 20 10 30
d) Vinitial, inf + past part 127 356 483 830 532 1362
e) Vinitial, inf + inf + past part 10 18 28 37 29 66
f) Vfinal, finite + inf 3472 43 3515 2989 6 2995
g) Vfinal, finite + inf + inf 438 0 438 298 0 298
h) Vfinal, finite + inf + inf + inf 14 0 14 5 0 5
i) Vinitial, inf + inf 1715 1 1716 653 0 653
j) Vinitial, inf + inf + inf 49 0 49 9 0 9
k) Vfinal, finite + inf (ipp) + te inf 3 0 3 10 0 10
l) Vinitial, inf (ipp) + te inf 14 0 14 19 0 19
Sum (#) 7669 2193 9862 8857 2358 11215
Sum (%) 77.76 22.24 100 78.97 21.03 100
Table 4: Clustering constructions in CGN and LASSY
We have found 9862 clustering constructions in CGN and 11215 in LASSY. Neither of the tree-
banks contains clusters with more than four verbs. In LASSY, the majority of the clustering con-
structions contains a past participle (categories a-e), whereas in CGN, the clusters containing bare
infinitives occur more frequently (categories f-j).
The results show that the proportion of clusters that potentially contain cluster creepers, i.e. the
clusters in which the main verb is not the first verb in the cluster (mv 6=1), is more or less equal in
both treebanks, i.e. 77.76% in CGN and 78.97% in LASSY.
6.2 Cluster creepers
After having collected the set of clustering constructions, we extracted the constructions with cluster
creepers, i.e. constructions in which nonverbal elements occur between the verbs in the second pole.18
Due to the treebank design, it is not possible to extract all constructions with cluster creepers in
that way, however. Separable verb particles (svps) are only tagged separately if they are written as
a separate word, but not if they are written as a part of the verb, as in example (8b). In the LASSY
treebank they can be extracted in another way, but not in CGN, as will be explained in section 6.3.
Therefore, this section focuses on the cluster creepers that are written as a separate word.
Since the set of constructions with cluster creepers is low in comparison to the set of all clustering
constructions, the results were manually verified after the automatic extraction.
Even though the quality of the annotations in both LASSY and CGN is very high, the treebanks
contain some annotation errors that are problematic for this research. For example, sentences that
are erroneously tagged as V-final whereas they are V-initial.
In (28), for instance, the clause after ‘uh’ is tagged ssub instead of smain.
18. For the extraction of cluster creepers, we started from the XPath queries defined in section 5. Since it is hard to
determine the linear order of the nodes in an elegant way using XPath, we have used XQuery scripts in which we
defined constraints for extracting the constructions in which nonverbal elements occur between the verbs. As an
example, the XQuery script which was used to find cluster creepers in two-verb clusters with a finite verb and an
infinitive is included as an appendix to this paper.
161

(28) dan
then
kan
can
ik
I
uh
uh
ik
I
kan
can
’m
him
in
in
de
the
keuken
kitchen
nergens
nowhere
inpluggen
plug in
vrienden.
friends
‘then I can’t plug it in in the kitchen, friends.’ [CGN, fna000573_58]
Besides the elimination of annotation errors, two types of false positives were filtered out semi-
manually. The first type concerns constructions with stopgaps, corrections, and/or interruptions,
such as the examples in (29). Those constructions were mainly encountered in CGN.
(29) a. maar
but
wat
what
wij
we
merkten
noticed
in
in
Frankrijk
France
was
was
dikwijls
often
dat
that
ge
you
’s
at
middags
lunchtime
soms
sometimes
zeer
very
goede
good
menu’s
menu’s
kondt
could
gebr-
use.interrupted
allee
go.stopgap
eten
eat
dus
thus
hè.
hè
‘What we often noticed in France was that you sometimes could use- well eat very good
menu’s at lunchtime.’ [CGN, fva400295_400]
b. enfin
well
ik
I
weet
know
niet
not
hoe
how
ik
I
het
it
moet
must
uh
uh
omschrijven
describe
uh.
uh
‘well I don’t know how I have to uh describe it.’ [CGN, fva400534_85]
The second type of false positives is the occurrence of punctuation marks within the verb cluster.
Those examples were exclusively found in LASSY.
(30) Het
it
is
is
dus
thus
niet
not
zo
so
dat
that
deze
this
tanks
tanks
al
already
eerder
before
“gekannibaliseerd”
“cannibalised”
waren
were
om
for
er
there
bruikbare
usable
onderdelen
parts
uit
out
te
to
halen.
get
‘It is thus not the case that these tanks were “cannibalised” before to get useful parts out of
it.’ [LASSY, WR-P-E-I-0000013937.p.4.s.235]
Table 5 presents the results for both treebanks. To compare the amount of cluster creepers to
the set of clustering constructions that may allow cluster creepers, i.e. the constructions in which
the main verb is not the first verb of the clusters, the numbers for those constructions are included
in this table as well (mv 6=1).
Cluster Type cgn lassy sum
a) Vfinal, finite + past part 23 11 34
b) Vfinal, finite + inf + past part 2 0 2
c) Vfinal, finite + inf + inf + past part 1 0 1
d) Vinitial, inf + past part 1 1 2
e) Vinitial, inf + inf + past part 0 0 0
f) Vfinal, finite + inf 79 7 86
g) Vfinal, finite + inf + inf 20 0 20
h) Vfinal, finite + inf + inf + inf 1 0 1
i) Vinitial, inf + inf 49 0 49
j) Vinitial, inf + inf + inf 4 0 4
k) Vfinal, finite + inf (ipp) + te inf 0 2 2
l) Vinitial, inf (ipp) + te inf 3 3 6
sum 183 24 207
mv 6=17669 8857 16526
Table 5: Frequency of cluster creepers in CGN and LASSY
162

Compared to the large amount of clustering constructions, the results in Table 5 show that cluster
creeping is a very infrequent phenomenon in both CGN and LASSY. In CGN, we have encountered
183 constructions with cluster creepers, whereas in LASSY we have only found 24. So, cluster
creeping occurs more frequently in the spoken data (CGN) than in the written data (LASSY). The
constructions account for 2.4% of all clusters that potentially allow cluster creepers (mv 6=1) in
CGN, and for less than 0.3% of those constructions in LASSY.
6.2.1 Single cluster creepers
Despite the low number of corpus examples, the constructions with cluster creepers show a large
variety of cluster creepers, both in category and syntactic function. The three types mentioned in
Haeseryn et al. (1997) are all present in the data: The sentences in (31) show cluster creeping by a
predicative adjective (31a) and by a part of a fixed expression (31b). (32) is an example of adposition
stranding within the cluster. In (33a) the cluster is interrupted by an object, and in (33b) by an
adverbial modifier.
(31) a. de
the
dokters
doctors
zeggen
say
wel dat
that
‘t
it
gaat
goes
goed
good
komen.
come
‘The doctors say that it will be fine.’ [CGN, fva400370_6]
b. ’k
I
zeg
say
dat
that
gaat
goes
moeten
must
beginnen
begin
op
on
gang
pace
komen
come
hè.
‘I say that should start to get going.’ [CGN, fva400643_87]
(32) De
the
plicht
duty
die
than
hem
him
nu
now
roept,
calls
kan
can
hem
him
straks
later
de
the
mooiste
most-beautiful
baan
job
kosten
cost
waar
where
een
a
Beier
Bavarian
kan
can
van
of
dromen.
dream
‘The duty that calls him now can cost him the most beautiful job a Bavarian can dream of.’
[LASSY, WR-P-P-I-0000000033.p.21.s.4]
(33) a. als
if
ze
they
moeten
must
teksten
texts
schrijven
write
dan
then
schrijven
write
ze
they
die
them
met
with
de
the
PC.
PC
‘If they have to write texts then they write them with a PC.’ [CGN, fvb400165_130]
b. maar
but
normaal
normally
moet
must
ge
you
dat
that
kunnen
can
zo
so
regelen
arrange
dus
thus
dat
that
dat
that
wegblijft
away-stays
dus
thus
dat
that
‘t
it
niet
not
verschijnt.
appears
‘But normally you have to arrange that in such a way that it stays away so that it does
not appear.’ [CGN, fva400079_264]
An overview of all creeper types is provided in Table 6. The labels in the columns indicate the
syntactic function (dependency relation): Separable verb particle (svp), prepositional complement
(pc), direct object (obj1), predicative complement (predc), location or direction complement (ld),
indirect object (obj2), modifier (mod), and predicative modifier (predm). The left part of the table
concerns complements selected by the verb, whereas the right part concerns modifiers.
The labels in the rows indicate the lexical categories (pos) at the top half of the table and
the phrasal categories at the bottom part of the table. 14 instances of cluster creeping show a
combination of several categories. They are not included in Table 6, but will be discussed in this
section as well (see 6.2.2).
163

svp pc obj1 predc ld obj2 mod predm sum (#) sum (%)
prep 12 37 0 2 12 0 7 0 70 36.27
adj 13 0 0 20 0 0 11 0 44 22.80
n5 0 16 0 0 0 0 0 21 10.88
adv 5 0 0 0 2 0 6 1 14 7.25
pron 0 0 4 1 1 0 5 0 11 5.70
pp 4 1 0 2 7 1 5 0 20 10.36
np 0 0 8 0 0 0 1 0 9 4.66
ap 0 0 0 1 0 0 2 0 3 1.55
advp 0 0 0 0 0 0 1 0 1 0.52
sum (#) 39 38 28 26 22 1 38 1 193
sum (%) 20.21 19.69 14.51 13.47 11.40 0.52 19.69 0.52 100
Table 6: Types of cluster creepers in CGN and LASSY
As expected, the largest category consists of cluster creepers where an svp occurs within the cluster,
as in (34).
(34) Ik
I
heb
have
mijn
my
agenda
diary
niet
not
hoeven
need
om
down
te
to
gooien
throw
om
to
die
that
man
man
te
to
kunnen
can
ontvangen
receive
(...)
I did not have to completely change my schedule to be able to receive that man (...)’ [LASSY,
dpc-rou-000479-nl-sen.p.10.s.14]
As mentioned above, the results do not include the cases of cluster creeping with separable verbs in
which the svp and the verb are written as one word.
Another major group are the prepositional complements. They include the cases of adposition
stranding illustrated in (32).
The other frequently occurring creeper types are also mentioned in Haeseryn et al. (1997), i.e.
predicative adjectives (31a), direct objects (33a), and modifiers (33b).
More remarkable examples in the data set are the constructions in which a full phrase occurs
within the cluster, such as the prepositional indirect object in (35a) and the location complement
in (35b).
(35) a. (...)
(...
’k
I
weet
know
’k
I
ik
I
niet
not
of
or
dat
that
’k
I
ik
I
dat
that
nu
now
moet
must
laten
let
weten
know
aan
to
hem
him
of
or
dat
that
’k
I
ik
I
dat
that
eerst
first
moet
must
aan
to
mijn
my
kot
student’s apartment
vragen
ask
(...)
(...) )
‘(...) I don’t know whether I should let him know or that I should ask (the people of)
my student’s apartment first.’ [CGN, fva400507_4]
b. dat
that
die
that
nu
now
moet
must
in
in
de
the
Verenigde Staten
United States
blijven
stay
in
in
Miami
Miami
bij
with
de
the
familie
family
(...)
(...)
‘that he now has to stay in the United States in Miami with his family (...)’ [CGN,
fvj600261_9]
A final note on Table 6 concerns the four instances of phrasal svps. Those constructions all
contain fixed expressions, such as the example given in (31b). Nonverbal parts of fixed expressions
are tagged as svps in the treebanks, but one could also classify those constructions as pcs.
6.2.2 Multiple cluster creepers
The 14 constructions that are not included in Table 6 form a heterogenous group that is not en-
countered in the literature on cluster creeping. Those examples contain multiple cluster creepers.
164

It is hard to draw any generalizations over this kind of constructions. Out of the 14 instances,
10 cluster creepers consist of a modifier, combined with a direct object, a predicative complement,
a prepositional complement or a locational/directional complement. With regard to the syntactic
category of the complex cluster creepers, any combination of lexical and phrasal categories seems to
be possible. Some examples are given in (36).
(36) a. (...)
(...)
den
the
dokter
doctor
heeft
has
eerst
first
moeten
must
tien
ten
minuten
minutes
die
those
twee
two
vrouwen
women
kalmeren
calm-down
voor
before
ie
he
het
the
onderzoek
examination
kon
could
doen.
do
‘(...) The doctor first had to calm down those two women for ten minutes before he could
do the examination.’ [CGN, fvn400019_191]
b. (...)
(...)
alhoewel
although
dat
that
ik
I
er
there
wel
indeed
’ns
some time
zou
would
graag
gladly
aan
on
meedoen.
participate
‘(...) although I would like to participate in that.’ [CGN, fvb400165_191]
c. (...)
(...)
als
if
je
you
zeg
say
maar
but
homo
gay
bent
are
en
and
dan
then
uh
uh
ja
yeah
gewoon
just
nie
not
ja
yeah
je
you
weet
know
niet
not
hoe
how
je
you
het
it
met
with
je
your
ouders
parents
moet
must
’t
it
erover
there-over
hebben
have
(...)
(...)
‘(...) for example if you are gay and you don’t know how you should talk about it with
your parents.’ [CGN, fna000541_298]
In (36a) the cluster contains a temporal modifier and a direct object NP. (36b) is a combination
of an adverbial modifier and adposition stranding. In (36c) not only the preposition occurs within
the cluster, but the pc as a whole is realised in situ. Moreover, the cluster is interrupted by the
direct object as well. Not surprisingly, all instances of such complex creeping constructions occur in
the spoken data (CGN).
6.2.3 Position of the cluster creepers
Another aspect regarding cluster creeping is the position of the nonverbal elements. In section 1.2.2
it was said that in clusters with more than two verbs, the nonverbal element typically occurs right
in front of the main verb.
In the data set, there are 30 cases of cluster creeping in constructions with three or four verb
forms. In 18 cases, the cluster creeper occurs just in front of the main verb, as in (37a), whereas in 12
constructions, they occupy a more leftward position in the cluster, as in (37b). The numbers confirm
the statement of Haeseryn et al. (1997), but the amount of relevant examples in the treebanks is
very low.
(37) a. (...)
(...)
iemand
someone
die
who
zich
himself
heeft
has
weten
know
binnen
in
te
to
werken
work
in
in
kringen
circles
met
with
een
a
hoog
high
sociaal
social
aanzien
standing
(...)
(...)
‘(...) someone who has managed to work his way up into high society (...)’ [LASSY,
dpc-ind-001652-nl-sen.p.11.s.1]
b. dus
so
dat
that
huisje
house.dim
wat
what
we
we
daar
there
hebben
have
neer
down
laten
let
zetten
put
(...)
(...)
‘so that little house that we got built over there.’ [CGN, fni007330_43]
165

6.2.4 Language-internal variation
Haeseryn et al. (1997) state that cluster creeping is more typical in Belgian Dutch compared to
Netherlandic Dutch. Since CGN contains meta-information on the origin of the data, it is possible
to verify that aspect in the treebank results as well. Out of the 183 occurrences of cluster creeping
in CGN, 145 constructions are part of the Belgian data set, while the remaining 38 constructions
occur in the Netherlandic data, so the data indeed show that cluster creeping is more common in
Belgian Dutch. In section 4 it was mentioned that CGN contains twice as much Netherlandic data
as Belgian data. If we normalise the data, it turns out that cluster creeping occurs 7.6 times more
often in the Belgian data compared to the Netherlandic part of the corpus.
6.3 A note on separable verbs in LASSY
As mentioned in section 6.2, separable verbs may be written as one word if the svp occurs next to
the verb. In those cases the svps are not individually tagged in the treebanks.
It is possible, however, to detect the clusters containing an svp by extracting the root forms of
the verbs in the clustering constructions in LASSY. In the root tag of the verb the root and the
svp are separated by an underscore, e.g. bel_op for the verb opbellen ‘call’. The numbers are given
in Table 7.19
mv 6=1 mv = 1 sum
Separable verbs 2556 390 2946
Non-separable verbs 6301 1968 8269
sum 8857 2358 11215
Table 7: Distribution of separable verbs within clusters in LASSY
The results show that there are 2556 occurrences of cluster creeping by an svp in the LASSY
treebank, indicating that such constructions are relatively frequent, in contrast to the observations
in Table 6. The creeping constructions account for 22.8% of all clustering constructions in LASSY,
and for 86.8% of all separable verbs in clusters in LASSY.
Note that separable verbs are only represented as such in the root forms but not in the lemmas.
It is possible to retrieve svps in this way in the LASSY treebank, but not in CGN, since the CGN
treebank only includes lemmas but no root forms. It is thus not possible to compare the results in
Table 7 to the frequency of separable verbs in CGN. Exploring alternative ways of retrieving those
constructions in CGN remains future work.
7. Conclusions and future work
This paper investigated the occurrence of cluster creepers in the CGN and LASSY treebanks. Since
those treebanks do not contain a specific tag for clustering verbs, we first had to define which
constructions we consider as verb clusters before extracting the relevant constructions. Compared
to the large amount of clustering constructions, the treebanks show that cluster creeping is a low-
frequent phenomenon in Dutch, except in the case of svps. Despite the small set of treebank
results, the variety of the creeper types turned out to be rather large. All categories mentioned in
(Haeseryn et al. 1997) are included in the data. Moreover, a subset of the cluster creepers consists
of a combination of several creeper types. Those constructions are not mentioned in the literature
on the phenomenon, showing that corpus-based research can add additional insights into linguistic
phenomena.
Further work is needed on how to deal with the inconsistent spelling in Dutch regarding separable
verb particles, as well as with the problematic annotation of separable verbs in the treebanks.
19. The results include the examples with the separately tagged svps as well.
166

As we only found some examples of cluster creeping in the data, it would be interesting to
investigate the phenomenon in a larger corpus, for example the SoNaR treebank (Oostdijk et al.
2013). That treebank of written Dutch not only contains more data (500M words), it also covers a
larger variety of text types.
Acknowledgments
We thank the audience of the CLIN conference (Leiden, January 17, 2014) and the anonymous
reviewers for their comments. The research presented in this paper is part of a project on complement
raising and cluster formation in Dutch, sponsored by FWO Vlaanderen (2011-2015, G.0.559.11.N.10).
167

References
Augustinus, L. and F. Van Eynde (2012), A Treebank-based Investigation of IPP-triggering Verbs
in Dutch, Proceedings of the Eleventh International Workshop on Treebanks and Linguistic
Theories (TLT11), Edições Colibri, Lisbon.
Augustinus, L., V. Vandeghinste, and F. Van Eynde (2012), Example-Based Treebank Querying,
Proceedings of the 8th International Conference on Language Resources and Evaluation (LREC-
2012), European Language Resources Association (ELRA), Istanbul, pp. 3161–3167.
Augustinus, L., V. Vandeghinste, I. Schuurman, and F. Van Eynde (2013), Example-Based Tree-
bank Querying with GrETEL - now also for Spoken Dutch, Proceedings of the 19th Nordic
Conference of Computational Linguistics (NODALIDA 2013), NEALT Proceedings Series 16,
Oslo, pp. 423–428.
Evers, A. (2003), Verbal Clusters and Cluster Creepers, in Seuren, P.A.M. and G. Kempen, editors,
Verb Constructions in German and Dutch, John Benjamins, Amsterdam/Philadelphia, pp. 43–
89.
Fersøe, H., S. Olsen, C. Navarretta, and B. Jongejan (2006), Validation Report Corpus Gesproken
Nederlands 1.0 Linguistic Validation, Technical report, Center for Sprogteknologi, University of
Copenhagen.
Haegeman, L. and H. van Riemsdijk (1986), Verb Pro jection Raising. Scope and the Typology of
Rules Affecting Verbs., Linguistic Inquiry 17, pp. 417–466.
Haeseryn, W., K. Romijn, G. Geerts, J. de Rooij, and M. van den Toorn (1997), Algemene Neder-
landse Spraakkunst, second ed., Martinus Nijhoff/Wolters Plantyn, Groningen/Deurne.
Hoekstra, E. (2010), On the interruption of Verb-Raising clusters by nonverbal material, Structure
Preserved. Studies in Syntax for Jan Koster, John Benjamins, Amsterdam, pp. 175–183.
Hoekstra, H., M. Moortgat, B. Renmans, M. Schouppe, I. Schuurman, and T. van der Wouden
(2003), CGN Syntactische Annotatie. 77p.
Jongejan, B., S. Olsen, and H. Fersøe (2011), Validation Report Lassy Corpora Linguistic Validation,
Technical report, Center for Sprogteknologi, University of Copenhagen.
Jurafsky, D. and J. Martin (2009), Speech and Language Processing, 2nd ed., Pearson Education,
New Jersey.
Oostdijk, N., M. Reynaert, V. Hoste, and I. Schuurman (2013), The construction of a 500-million-
word reference corpus of contemporary written Dutch, in Spyns, P. and J. Odijk, editors, Es-
sential Speech and Language Technology for Dutch: resources, tools and applications, Springer,
pp. 219–247.
Oostdijk, N., W. Goedertier, F. Van Eynde, L. Boves, J.-P. Martens, M. Moortgat, and H. Baayen
(2002), Experiences from the Spoken Dutch Corpus Project, in Rodriguez, Manuel Gonzalez
and Carmen Paz Saurez Araujo, editors, Proceedings of the 3rd International Conference on
Language Resources and Evaluation (LREC-2002), Las Palmas, pp. 340–347.
Paulussen, H., L. Macken, J. Truskina, P. Desmet, and W. Vandeweghe (2006), Dutch Parallel
Corpus: a multifunctional and multilingual corpus, Cahiers de l’Institut de Linguistique de
Louvain, CILL 32 (1-4), pp. 269–285.
Van Eynde, F. (2004), Part of Speech Tagging en Lemmatisering van het Corpus Gesproken Neder-
lands, 87p.
168

van Noord, G. (2006), At Last Parsing Is Now Operational, in Mertens, P., C. Fairon, A. Dister,
and P. Watrin, editors, TALN 2006. Verbum Ex Machina. Actes de la 13e conference sur le
traitement automatique des langues naturelles, pp. 20–42.
van Noord, G., G. Bouma, F. Van Eynde, D. de Kok, J. van der Linde, I. Schuurman, E. Tjong
Kim Sang, and V. Vandeghinste (2013), Large Scale Syntactic Annotation of Written Dutch:
Lassy, in Spyns, P. and J. Odijk, editors, Essential Speech and Language Technology for Dutch:
resources, tools and applications, Springer.
van Noord, G., I. Schuurman, and G. Bouma (2011), Lassy Syntactische Annotatie, Revision 19455.
208p.
Wurmbrand, S. (2005), Verb Clusters, Verb Raising, and Restructuring, in Everaert, M. and H. van
Riemsdijk, editors, The Blackwell Companion to Syntax, Vol. V, Blackwell, Oxford, chapter 75,
pp. 229–343.
169

Appendix: XQuery script for cluster creepers
This XQuery script looks for cluster creepers in V-final finite-infinitive clusters:20
(: XPath extracts V-final finite-infinitive clusters in the LASSY small treebank :)
for $xp in db:open("LASSY_ID")/treebank/alpino_ds
//node[@cat="ssub" and node[@rel="hd" and @pt="ww" and @wvorm="pv"] and
node[@rel="vc" and @cat="inf" and node[@rel="hd" and @pt="ww"] and
not(node[@rel="vc" and (@cat="inf" or @cat="ti" or @cat="ppart" or @pt="ww")])]]
(: get sentence ID:)
let $sentenceid := ($xp/ancestor::alpino_ds/@id)
(: get sentence:)
let $sentence := ($xp/ancestor::alpino_ds/sentence)
(: get finite verb and infinitive :)
let $finite := ($xp/ node[@rel="hd" and @pt="ww" and @wvorm="pv"]/@word)
let $infinitive := ($xp/node[@rel="vc" and @cat="inf"]/node[@rel="hd" and @pt="ww"]/@word)
(: get position of the finite verb and the infinitive :)
let $finiteposition := ($xp/ node[@rel="hd" and @pt="ww" and @wvorm="pv"]/@begin)
let $infinitiveposition := ($xp/node[@rel="vc" and @cat="inf"]/node[@rel="hd" and @pt="ww"]/@begin)
(: get cluster creepers :)
(: finite - infinitive :)
let $creepers1 := ($xp/descendant::node[(number(@begin) > number($finiteposition)) and
(number(@begin) < number($infinitiveposition))])
(: infinitive - finite :)
let $creepers2 := ($xp/descendant::node[(number(@begin) < number($finiteposition)) and
(number(@begin) > number($infinitiveposition))])
(: only return constructions with cluster creepers :)
where ($creepers1 or $creepers2)
(: return sentences, verb cluster, cluster creepers (word, syntactic function and POS tag) :)
return
if (number($finiteposition) < number($infinitiveposition))
then <match>{data($sentenceid)}#{data($sentence)}
#FINITE-INFINITIVE#{data($finite)}-{data($infinitive)}
#{data($creepers1/@word)}#{data($creepers1/@rel)}#{data($creepers1/@pt)}</match>
else
<match>{data($sentenceid)}#{data($sentence)}
#INFINITIVE-FINITE#{data($infinitive)}-{data($finite)}
#{data($creepers2/@word)}#{data($creepers2/@rel)}#{data($creepers2/@pt)}</match>
20. Comments are put between (: and :).
170