Dutch Cross Serial Dependencies in HPSG 
Gerrit Rentier* 
Institute for I,anguage Technology and Artiticial \]intelligence 
'rilburg University, PO Box 90153, 5000 LE Ti\]burg, The Netherlands 
rentier(@kub.nl 
Abstract 
We present an analysis of Dutch cross serial depen- 
dencies in Head-driven Phrase Structure Grammar 
(\[P&S(1994)\]). Wc start out from the assumption that 
causative and perceptual verbs, like auxiliaries, can 
lexically 'raise' the arguments of the non-finite verbs 
they govern to their own list of arguments through 
"argument composition" (\[H&N(1989)\]). 
1 Introduction 
Dutch cross serial dependencies (DCSDs), well-known 
from (1) and (2), still challenge computational linguis- 
tics for an efficient treatment. 
(1) dat ik~ haar~ de nijlpaardeu~ zag~ voeren~ 
that I her the hippos saw feed 
"that I saw her feed the hippos" 
(2) dat ik~ Henk~ haar 3 de nijlpaarden a zag~ 
that I Itenk her the hippos saw 
helpena voeren 3 
help feed 
"that I saw tIenk help her feed the hippos" 
The problematic aspects of I)CSDs are of course 
the bounded discontinuous relation between the NPs 
and the verbs of which they are arguments, indi- 
cated in (1) and (2) by the subscripted integers, 
and the recursiveness of the phenomenon. The con- 
struction is only licensed by members of two closed 
classes of verbs, the class of perceptual verbs like 
zicn( "see" ), hoven( "hear" ) and voelen( "feel" ), and the 
class of causative verbs like laten('qet/make") and 
helpen("help"). In the analysis put forward here wc 
emphasize this lexical aspect of the phenomenon; in 
our analysis DCSDs are strictly tied to the subcate- 
gorization and semantics of perceptual and causative 
verbs. We analyze them as verbs which select, apart 
from their subject, a nonfinite V-projection which de- 
notes an event. More particularly, as is proposed for 
German auxiliaries in \[H&N(1989)\], they subcatego- 
*Sponsored by EC projects FJSI'IHT P5254 (PLUS), \],\].ql'Rrl' 
P6665 (DANDELION) and two travel grants by the Netherlands 
Organization for Scientific Research (NWO). Many thanks to 
Bob Borsley, Jo Calder, Bart Geurts, 3os~e IIeemskerk, John 
Nerbonne, Paola Monachesi, Ivan Sag, Wietske Sijtsma and 
Craig Thiersch for detailed ornaments and sound advice. Er- 
rors are, of course, completely my own. 
rize for the arguments of the verb they govern, a mech- 
anism frequently referred to as argument composition 
or argument inheritance. 
Recently DCSDs have been analyzed in a non- 
standard version of tIPSG 1 in \[P~eape(fc.)\]. In his so- 
called sequence union approach, the standard concept 
of phrase structure grammar (i.e. that a string is de- 
fined as the terminal yield of a phrase structure tree) 
is abandoned. Our analysis is more standard, in the 
sense that we only need to refer to the lexicon and 
the HPSG-mechanism of structure sharing3 Our pre- 
ferred explanatory mechanism, argument composition, 
is not so much an additional mechanism as an effect 
which derives from careful specification of structure 
sharing, and structure sharing is already at the theo- 
retical core of HPSG. 
Furthermore, argument composition is indepen- 
dently motivated, because Dutch is like German with 
rcspect to the phrase-structural behaviour of auxil- 
iaries, and argument composition in German construc- 
tions with auxiliaries is well-motivated (\[H&N(1989)\]). 
So we have good reason to assume argument composi- 
tion present in the theory, regardless of DCSDs. 
2 Event Semantics in HPSG 
The choice of semantics in terms of a theory of events, 
known from \[Davidson(1967)\], offers interesting ad- 
vantages and explanations of logical and linguistic phe- 
uomena, motivating tt, e development of a constraint- 
based version of it. 3 So, in the spirit of event seman- 
tics we propose that main verbs like voeren("feed') 
in (3) should denote a discourse referent, which is in 
fact a very natural assumption. In (3) and throughout 
the paper, recurring \[-~'s indicate structure sharing, 
that is token-identity of information, as is common us- 
agc in HPSG. Note also that we follow \[Borsley(1987)\] 
in representing subjects as values of SlJB3 and follow 
\[P&S(1994)\] (chapter 9) in representing non-subject 
1 Head-driven Phrase Structure Grammar, which the reader 
is p ......... d to b ......... less familiar with (see \[P&S(1994)\]). 
2In fact our analysis differs from many previous analyses of 
DCSDs in that we do not refer to any 'additional' (often pow- 
erful) mechanisms (sequence union, head wrapping, repeated 
rightward head movement). 
3The combination of" IIPSG with (shallow) event seman- 
tics and translation to an event-based logical fomn originates 
with work on the EC-sponsored PLUS project, EsPaiT P52fi4, 
a "Pragmatics-based Language Understanding System". 
818 
arguments as wducs of C()MI'~. 
(3) 
"PHON ( voe,'en ) 
IIEAD \[VFORM ~L~s\],~ 
COMPS ( NP\[A(:o'\]:\[ 3\] ) 
GOV ) 
CONT -DNT I,~VI,:N'V 
PARA \[~ 
I,EX + 
q?he constraint-based event semantics of the base form 
verb voeren as it is depicted in (3), with the quasi- 
determiner V~W.:N% should be interpreted as an existen- 
tially quantified event with a parameter \[a I which is 
restricted to involve a relation of feeding, an argumcvt 
with the role of agent which is associated 4 with a sc-- 
mantle content \[2 I and an argument associated with a 
semantic content \[~ which is the theme. '~ 
Here the valne of I)1,7\[' is a 'shallow' representation 
of a quantifier, (5 and the value of PAHA, which is an 
abbreviation for 'parameter', is structure shared with 
the value of a feature/NST which is short for 'instance'. 
We will suppose that the value of I'AI(A corresponds 
with a discourse referent in the discourse represen- 
tation associated with a i, atural langnage expression, 
without formally defining this relation here. The value 
of HI,;STIq which abbreviates 'restrictions': is a set of 
constraints on the value of this parameter. 
3 An Argument Colnposition Analysis 
We assume that the clause structure of 1)CSI)s is one 
where we have a binary left-branching verbal cmnplex. 
This verbal complex then locally selects the sum of the 
arguments of the verbs which constitute it. We fecl 
that a binary branching analysis is empirically moti- 
vated by auziliary flip in the same way as auxiliary 
flip motivates a binary right-branching structure for 
the German verbal complex, following \[H&N(1989)\]. 
4Here mid throughout the paper, "(}:~" memm "feature 
tltrllcLure • with as C()NTI,\]N'I'-value ~". 
BWe ~lSSllllJ.e thttt our constraint-based event st21llalltie8 is itt- 
duetively translated to a level of underspecified logical fortxl, 
and that this ULF-level then can be mapped to a level of logical 
form and a model-theoreLic interpretation. The attxiliary levels 
are not defined here, but of. \[llentier(ms.)\]. 
The eoztcept of semantics we will outline here will b e shallow 
for instance ht the sense that we do not dismiss quantification as 
it is COlIt~non-place in formal semantics. IIowever, of. chapter 8 
of \[P&S(1994)\] for discussion of a treatment of quantifier scope 
which could he cornhlned with ore' approach, if so desired. 
A governing auxiliary will apply argument composi- 
tion and raise all the complements from the governed 
verb(s) to become arguments of the auxiliary, as pro- 
posed in \[H&N(1989)\]. We assume that causative and 
perceptuM verbs syntactically behave just like auxil- 
iaries in this respect. 
Tile difference between auxiliaries on the one hand 
and perceptual and causative verbs on the other we 
view as basically semantic. We take it that auxiliaries 
semantically more or less operate on events, affecting 
features for tense and aspect or modality. Causative 
and perceptual verbs on the other hand will be ana- 
lyzed as events themselves, events which take other 
events as their argument~ in general as a theme (viz, 
a valne of AIt(~, cf. the entry in (7) below). 
ht chapter 9 of \[P&S(1994)\] the approach to local se-- 
lection from \[Borsley(1987)\] is developed fltrther and 
leads to the Valence t)rinciple, which refers to the va- 
lence features SiJI-~d and C()MPS through 'F': 
(4) Valence Prineiph, Chapter 9, \[P&S(1994)\] 
In a headed phrase, h)r each valence featnre F, 
the t,' wdue of the head-daughter is the con-. 
catenation of the phrase's F value with the list 
of SYNSI,;M values of tt, e F-daughters' value. 
The general effect of the principle on a phrase which 
is headed by some sign is that this headed sign can 
only become 'complete' (or "saturated') if it is com- 
bined with the appropriate arguments. For example, 
in the case of a transitive verb, such a verb must find 
a subject NP (selected through Sl;B./) and some oh. 
ject (selected through (:()Ml'S). If we assume a lint 
clause structure analysis of Dutch and we furthenm)re 
assume lcxical signs like (3) and (7), then the imme~ 
diate dominance statements (5) and (6) will suffice to 
describe tilt: constructim, of Dutch we are concerned 
with here. r Here the H,S and C indicate that the 
daughters of the phrase include a head, a subject and 
complements, not necessarily in that order (eft chap- 
ter 9 of \[P&S(1994)\] for details). Note that in addition 
to the wdency features suJLI and COMPS, we also as- 
sume the t)resence of the G()V-feature, ranging over 1 
complement:S 
(5) XP\[ll,:x-q -~ S,C1,.., C=,lI\[<ov(), u,:x-t\] 
(6) XiH.:x I\] ~ II\[(,w ((:~), ,,,,:×+\] , Ci 
'Phe second schenta is in a sense not a "phrase" 
structure schema but is instead a "cluster-formation'- 
schema. This is because normally the combination of 
two or more words leads to a sign which is I,I.;x-, a 
phrasal sign, but here it leads to a 'complex word' 
which is I,\],;X F. Also (6) is strictly binary: it takes one 
argument, namely the argument which is the value of 
fActually, our analysis also presupposes the IIead Featm'e 
Principle and Semantics Principle from \[P&S(1994)\]; cf. Figures 
1 and 2 for informal illustration. 
aFollowing discussions of Weheltmth, Ackerman, Sag and 
Pollard at WCCFI, XlII, suggesting tiffs for German, and Chang 
of Ohio State University originally suggesting this for Korean. 
819 
COY. We arrange the lexicon so that any value of coy 
will always be an unsaturated base form verb which is 
defined as LEX+ as well. By the Valency Principle, 
this selection requirement of the governing verb will 
be appropriately 'cancelled' after string concatenation 
during parsing. 
Central to our analysis of the case-markings of 
NPs in the Dutch Mittelfeld is the assumption from 
\[Pollard(fc.)\] that base forms of verbs do not assign 
any case to their subject. The value for the subject- 
NP's CASE-feature in (3), "c^s~", is the supertype in 
the type hierarchy for those atomic types that are ap- 
propriate values of the feature CASE. So, the value 
cast." is the supertype of NOM and Ace in Dutch and En- 
glish, and in German also of DA'r and G~N. The result 
of assigning the subject-NP this supertype for case in 
practice bolls down to giving this NP some kind of 
"any"-value for case; the case-value CASE of such an 
NP will unify with any other possible case value. 
In our analysis, the discontinuous relation between 
arguments and verbs in DCSDs is brought about firstly 
by lexically defining finite pereeptuals like zag (and 
finite causatives) as argument composition verbs, along 
the following lines: 9 
(7) PHON(~o~) 
\[MAJOR ~ \] 
HEAD VFORM bUN 
I SUBJ ( NP\[NOM\]I~ ) 
COMPS ~NP\[~ocl ) ~ \[\] 
~OV ( "V\[I\]ASE\] \]) 
SUBJ ( ~\]NP ) 
COMPS \[\] 
CONT \[\] 
LEX + 
CONTENT 'bET ~VnNT 
PARA \[\] 
N / / \[\] / 
\[\] J 
LEX + 
The finite argument composition verb zag selects a 
singular nominative NP through its suB J-feature. As 
non-subject arguments it selects through its COMPS- 
feature first the NP tagged as \[\] which is unified with 
the SUB J-value of the governed verb(s), and secondly 
the list \[~\] of zero or more non-subject arguments 
of the governed verb(s). And crucially, being a gov- 
erning verb, zag selects through cov a governed base 
form verb, 1° with as SUB J-value "\[~', as COMPS-value 
9In this entry and throughout the paper, @ stands for con- 
catenation of arbitrary-length lists of arguments. 
1°One base form verb, or a base form verb-headed verbal clus- 
"\[~2' and as semantics "\[\]'. Note that, since the gov- 
erned v\[nsi.:\] is selected as missing a subject and a list 
of complements, it must not 'find' this subject or these 
complements, which it indeed doesn't (cf. the tree in 
Figure 1). 
As it were in passing, the governing perceptual verb 
(or causative verb alike) imposes accusative case on 
the NP which denotes the suloject-argument of the 
governed verb. The unification of \[CASE CASK\] and 
\[CASE Acc\] will be forced through the structure-sharing 
indicated in (7) as "\[\]', and will result in the more 
specific restriction \[CASE ,cc\]. This accounts for the 
accusative case-marking on hasp ("her") in examples 
(1) and (2), and in general on all non-subject argu- 
ments in such constructions. 
The second and crucial step in our account of the 
discontinuity is accounting for the linear order in the 
verb cluster with DCSDs. The linear order of the verb 
cluster in Dutch we account for through (8): 
(8) Linear Precedence l=tule Dutch Verb Clusters 
\[c~ov ( x )\] < x 
(9) Linear Precedence Rule German Verb Clusters 
X < \[(~ov ( x )\] 
By these LP-rnles, in each part of the binary branch- 
ing verb cluster the governing verb will appear head- 
initial in Dutch, and head-final in German. n It is 
straightforward to show that the above approach has 
the desired effect also for the sentence (2) mentioned 
in the introduction if we define a lexical entry for the 
causative helpen with a syntax and semantics along the 
same lines as the perceptual zag. The only difference 
must be that such nonfinite entries do not assign NOM 
to their subject, but "CASl.:". Other than that, there 
will just be additional embeddings in the semantics as 
well as in the verb cluster. Thus, by the ID-rule in (6) 
and the lexical entries for causatives and perceptuals, 
we account for the recursiveness of the phenomenon, 
cf. the tree in Figure 2. 
4 Conclusion 
We extended the \[H&N(1989)\]-analysis of German 
to Dutch, accounting for the difference, resp. nested 
vs. cross serial dependencies, through one single LP- 
parameter. Also, we argued that such an argument 
composition approach is to be preferred over several 
alternative approaches, since argument composition is- 
n't an 'additional' mechanism. Further linguistic ad- 
vantages of this approach, i.e. accounts of irregular 
case assignments and constraints on double infinitives, 
are discussed in \[Rentier(1994)\]. We are able to derive 
verb second constructions by standard application of 
ter; due to the ID-schema in (6) either will be LEX-\[-, so that 
we are able to recltrsively build up bigger and bigger LEX+- 
complexes. 
u LP-rules like these are common in HPSG, cf. for instance 
the rule XP < SUBJ ( XP ) which orders subjects before VPs 
in English (\[Borsley(1987)\]). 
820 
Figure 1: The discontinuous relation: Valence Principle, schema's (5) & (6), entries (3) & (7), LP-rule (8). 
S 
NP\[NOM\]:\[~\] NP\[A(;(;\]: 2~ NP\[ACC\]: a~ 
ik haar n. 
suBa NP\[NOMI:~\] ) 
COMPS NP\[ACC\]:~, NP\[Accl:\[~ ) 
GOV ) 
-I)ET EVENT 
PA I1,A \[6~ 
(X)NT gES,pH ,/ \[ : >)~A")~ NT 
m,;x + 
'V\[l,qN\] 
S/'BJ ( NP\[NOM\]:~\] ) 
'Vbsg 
GOV (~\] ('OMPS(~)) 
(\]ON'F \[\] 
m,:x + 
"DET EVENT 
PAISA \[-6J /P 
CON'\[' INST \[ 
Ll~x q- 
I zag 
~3 
"v\[ns,,:\] 
SU\[IJ 
COMPS 
GOV ( 
C()NZ ~ 
m,:x + 
N*':XI > 
NPbcc\]: a~j ) 
DET EVENT 
PARA 
f\[7 'J\] 
I voeren 
821 
Figure 2: Recarsion in the Verb Cluster (Sentence (2)). 
SUBJ ( \[a\]NP\[NOM\] ) 
COMPS ( \[b\]NP\[hc:c:\] , \[c\]NP\[ACC\] , \[d\]N1)\[hcC\] ) 
r~F,X ÷ 
"V\[I,'IN\] \] 
COMPS ( \[b\]NP\[*cc\] ) 0 \[~ 
boy<> / 
L~.~x + J 
LEX + 
zag 
\[V\[13ss\] 1 \[SUBJ ( NP\[cASI,:\] ) 
\[qI~-OM~..~ (\[~\]Nt'\[*~(:\], \[a\]NP\[,,oq ) 
\[~,ov ( > 
LL~X + 
\[ 
"v\[sss\] \[~ 
s,,,. ( NP\[~,,~,,:\] ) 
COMPS ( \[cJNP\[Acc\] ) (t) 
\[v\[,~s~\] \] 
Is.,,. < )/ < E#OM S 
L,,,,:x ÷ J 
I,EX + \] 
helpen 
J_ 
\[V\[~sl,:\] 
\[21\]oo~P~ ( \[a\]NP\[~.c\] 
kU,:× .I- 
voeren 
the Dutch versions of the extraction lexical rules (see 
\[Rentier(1993)\]) to the verbs at the lexical level. 
References 
\[Borsley(1987)\] R. Borsley, "Subjects and Comple- 
ments in HPSG", CSLI Report 107~ Stanford 
University, USA 
\[Davidson(1967)\] O. Davidson, "The Logical Form of 
Action Sentences", reprinted in "Essays on Ac- 
tions and Events", Clarendon Press, Oxford, 1980 
\[H&N(1989)\] E. Hinrichs, T. Nakazawa, "Flipped out: 
AUX in German", in "Proceedings of the 25th Re- 
gional Meeting of the Chicago Linguistic Society"~ 
CLS, USA 
\[Pollard(fc.)\] C. Pollard~ "On Head Non-Movement', 
in: Bunt, H. & van Horck, A. (eds.), "Proceed- 
ings of the Symposium on Discontinuous Consti- 
tuency", Mouton-de Gruyter, Germany 
\[P&S(1994)\] C. Pollard, I.A. Sag, "tlead-driven 
Phrase Structure Grammar"~ University of Chi- 
cago Press and CSLI Publications, USA 
\[Reape(fc.)\] M. Reape, "Getting Things in Order", in: 
Bunt, H. ~z van Horck, A. (eds.), "Proceedings of 
the Symposium on Discontinuous Consfituency", 
Mouton-de Gruyter, Germany 
\[Rentier(1993)\] G. Rentier, "Dutch Object Clitics, 
Preposition Stranding and Across-the-Board Ex- 
traction", in Sijtsma, W. & Zweekhorst, O. (eds.), 
'(Papers from Computational Linguistics in the 
Netherlands (CLIN) III, 1992", the Netherlands 
\[Rentier(1994)\] G. Renticr "A Lexicalist Approach to 
Dutch Cross Serial Dependencies" in: "Proceed- 
ings of the 30th Regional Mecfing of the Chicago 
Linguistic Society", Chicago, CLS USA 
\[Rentier(ms.)\] G. Rentier, "IIead-driven Phrase Struc- 
ture Grammar and Underspecified Logical Form", 
ms., ITK, Tilburg University, the Netherlands 
822 
