Adjuncts and the Processing of Lexical Rules 
Gertjan van Noord and Gosse Bouma 
BCN RUG Groningen 
{vannoord, goss e} 0let. rug. n\]. 
Abstract 
The standard HPSG analysis of Germanic verb 
clusters can not explain the observed narrow- 
scope readings of adjuncts in such verb clus- 
ters. 
We present an extension of the HPSG analy- 
sis that accounts for the systematic ambiguity 
of the scope of adjuncts in verb cluster con- 
structions, by treating adjuncts as members 
of the sul)eat list. The extension uses power- 
Nl reeursive lexical rules, implemented as com- 
plex constraints. We show how 'delayed eval- 
uation' teehMques from constrMnt-loglc pro- 
gramming can be used to process such lexical 
rules. 
1 Problem Description 
1.1 Dutch Verb Clusters 
Consider the following Dutch subordinate sen- 
tences. 
(1) dat Arie wil slapen 
that Arie wants to-sleep 
(2) dat Arie Bob wil slaan 
that Arie Bob wants to-hit 
that Arie wants to hit l/ot) 
(3) * (lat Arie Bob wil slapen 
that Arie Bob wants to-sleep 
that Arie wants to sleep lloh 
(4) * dat Arie wil Bob slaan 
(5) dat Arie Bob cadeautjes wil geven 
that Arie Bob presents want to-give 
that Arie wants to give presents to Bob 
(6) * dat Arie Bob wil cadeautjes geven 
dat Arie wil Bob ca(leautjes geven 
(7) dat Arie Bob zou moeten kunnen willen 
knssen 
that Arie Bob should must can want 
to-kiss 
that Arie should be able to want to kiss 
Ilob 
The examples 1-3 indicate that in l)utch the 
arguments of a main verb can be realized to 
the left of an intervening auxiliary verb, such 
as a modM verl). Furthermore the sentences 
in d-6 indicate that in such constructions the 
arguments must 1)e realized to the left of the 
auxiliary verbs. In 7 it is illustrated that there 
can he any numl)er of auxiliaries. 
1.2 The IIPSG analysis of verb- 
clusters 
The now standard analysis within t\[PSG of 
such verb-clusters is based oil ideas from Cat- 
egorial Grammar (cf. for example Moortgat 
(1988)) and defined within the HPSG flame- 
work by IIinrichs and Nakazawa (1989). In 
this analysis auxiliary verbs subcategorize for 
an unsaturated verb-phrase and for the com- 
pleinents that are not yet realized by this verb- 
l)hrase. In other words, the arguments of 
the embedded verl)-phrase are inherited by the 
auxiliary. 
For example, the auxiliary 'wil' might be de- 
fined as in Iigure 1. If we assume an ai)plica- 
tion rule that produces flat vp-structures, then 
we obtain the derivation in figure 2 for tim in- 
finite verb-phrase 
(8) ... Arie boeken wil kunnen geven 
1.3 Problems with the scope of ad- 
juncts 
A major problem that this analysis faces is 
the possibillity of narrow-scope readings ill tile 
250 
\[ v;;;UlA;'ant(Subj,able(Subj,give(Subj,books,arie))) \] 
lez : arie le~: : boeken le:r, : wit 
dir : left j ,//,':left l, sc : (P, R,L, J) sc : (R,I,, J) .~c : (L,J) dir : right p dir: right 
R 
Figure 2: The parse tree for tile verb-phrase %'ie t)oeken wil kunnen geven'. 
stem( 
VERBAL \[ 1 
seml,~clqfso. : argl : Subj 
soa-arg : Obj 
\[ w,ntnAt, 1 sere : Obj 
sc : ( s~tbjlsemlindez : Sjlx IA) 
sc:A 
Figure 1: The modal auxiliary 'wil'. 
case of adjuncts. For example, the follow- 
ing Dutch subordinate sentences are all sys- 
tematically ambiguous between a wide-scope 
reading (adjunct modilies tile event introduced 
by the auxiliary) or a narrow-scope reading 
(adjunct modi%s the event introduced by the 
main verb). 
(9) dat Arie vandaag Bob wil slaan 
that Arie today Bob want to-hit 
that Arie wants to hit IIob today 
(10) dat Arie hot artikel op tljd probeerde, op 
te stllren 
that Arie the article on time tried to send 
that Arie tried to send the article in time 
(11) dat Arie Bob de vronwen met een 
verrekljker zag bekljken 
that Arie Bob the women with the 
telescope saw look-at 
that Arie saw Bob looking at the women 
with the telescope 
Firstly note that tile treatment of adjuncts 
as presented in Pollard and Sag (in press), can- 
not be maintained a.s it simply fails to derive 
any of these sentences because the introduc- 
tion of adjuncts is only possible as sisters of 
saturated elements. The fact that ~trguments 
and adjuncts can come interspersed (at least in 
languages such as Dutch and German) is not 
accounted for. 
A straight forw~ml solution to this problem is 
presented in Kasper (in prepar~tion). Here ad- 
jmwts and arguments are all sisters to a head. 
Tim arguments should satisfy the subcat re- 
quirements of this hea.d - the adjuncts modify 
the semantics of the head (via a recnrsively 
defined a.djuncts principle). 
The main I)rol)lem for this treatment of ad- 
.iuncts is that it cannot explain the narrow- 
scope readings observed above, if adjuncts 
modify the. head of the phrase they are part of 
then we will only obtain the wide-scope rea,d- 
ings. 
If we assume, on the other hand, that ad- 
.jnncts are oil the subcat list;, then we will ob- 
tain both readings straightforwardly. In tile 
narrow-scope case tile adjunct is on the snbcat 
list of the embedded w~rb, and then inherited 
by the matrix w.~rb. In the wide-scope case 
tilt adjunct simply is on the subcat list of the 
matrix verb. in the next section we present a 
treatment of adjuncts in which each adjunct is 
subcategorized for. By me,ms of lexical rules 
we are able to obtain the. effect that there can 
be any mmfl)er of adjuncts. We also sketch 
how the semantics of modification might be 
delined. 
257 
2 Adjuncts as Arguments 
2.1 Adding adjuncts 
The previous section presented an argument 
that VP modifiers are selected for by the verb. 
Note that this is in line with earlier analyses 
of adjuncts in HPSG (Pollard and Sag, 1987) 
which where abandoned as it was unclear how 
the semantic contribution of adjuncts could be 
defined. 
IIere we propose a solution in which ad- 
juncts are members of the subcat list, just like 
ordinary arguments. The difference between 
arguments and adjuncts is that adjuncts are 
'added' to a subcat list by a lexical rule that 
operates recursively. 1 Such a lexical rule might 
for example be stated as in figure 3. 
Note that in this rule the construction of 
the semantics of a modified verb-phrase is still 
taken care of by a rood feature on the adjunct, 
containing a val and arg attribute. The arg 
attribute is unified with the 'incoming' seman- 
tics of the verb-phrase without the adjunct. 
The val attribute is the resulting semantics of 
the verb-phrase including the adjunct. This al- 
lows the following treatment of the semantics 
of modification 2, cf. figure 4. 
We are now in a position to explain the ob- 
served ambiguity of adjuncts in verb-cluster 
constructions. Cf.: 
(12) dat Arie Bob vandaag wil kussen 
that Arie Bob today wants to-kiss 
In the narrow-scope reading tim adjunct is first 
added to the subeat list of 'kussen' and then 
passed ou to the subcat list of the auxiliary 
verb. In the wide-scope reading the adjunct is 
added to the subcat list of the auxiliary wM~. 
The final instantiations of the auxiliary 'wil' 
for both readings are given iu figure 5. 
2.2 Discussion 
A further problem concerning the syntax of ad- 
juncts is posed by the fact that adjuncts can 
take part in unbounded dependency construe- 
lions. Lexical treatments of the kind presented 
in Pollard and Sag (in press), chapter 9 assume 
that a lexlcal rule is responsible for 'moving' 
lcf. Miller (1992) for a similar suggestions concern- 
ing French. 
2inspired by Kasper (in preparation) 
an element from the subcat list to the slash 
list. Such an account predicts that adjuncts 
cau not take part in such unbounded depen- 
dency constructions. In Pollard and Sag (in 
press), chapter 9 a special rule is introduced 
to account for those cases where adjuncts do 
take part in UI)Cs. '\['he treatment that we 
propose for adjuncts obviates the need for such 
an 'ad-hoc' rule. 
Clearly many details concerning the syntax 
of adjuncts are left untouched here, such as the 
quite subtle restrictions in word-order possibil- 
ities of certain adjuncts with respect to argu- 
ments and with respect to other adjimcts. In 
the current framework linguistic insights con- 
cerning these issues could be expressed as con- 
straints on the resulting subcategorization list 
(e.g. by means of LP-constraints). 
lit should also be stressed that treating ad- 
juncts and arguments on a par on the level 
of subcategorization does not imply that ob- 
serve<\[ differences in the behavi<)r of adjuncts 
and arguments could not be handled in the 
proposed framework. For example the differ- 
ence of adjuncts and arguments in the case of 
left dislocation in Dutch (exemplified in 13-16) 
can be treated by a lexica\] rule that oper~tes 
on the subcat list before adjuncts are added. 
(13) De voorstelling duurt een uur 
Tim show takes an hour 
(l,l) l';en uur, dat duurt de voorstelling 
(15) Arieen Bob wandelen een uur 
Arie and Bol) wall< an hour 
(16) * l"en uur, dat wandelen Arieen Bob 
3 Processing Lexical Rules 
3.1 Lexical Rules as Constraints on 
Lexical Categories 
Rather than formalizing the 'add-adjuncts' 
rule as a lexical rule we propose to use re- 
cursive constraints on lexical categories. Such 
lexical constraints are then processed using de- 
layed ewduation techniques, a 
Such an approach is more promising than an 
off-line approach that precomputes the effect 
aRefer to Carpenter (1991) for a proof of TurilLg 
equivalence of simple eategorial grammar with recur- 
slve lexical rules. 
252 
VERBAL 1 
sc : P • S :~ 
sere : Sere0 
VI,;RBAI, 
sc:l'.( wod : ar~l : Sem~) ) . S 
va/: Sere 
.N(?'Gq, : ~OIll 
Figure 3: A lexieal rule that adds a single adjunct to the sul)cat list of >t verb. In the. case of n 
~uljuncts the rule applies n times. 
RI,~S T I1.~.1) V E ll.n I A L 
arglnuc : \[ 
~ttod : 
val\[nuc : \[ 
resh" 
qI*oa : Q \] 
O P_.AI)Vt,;RIIIA L 
arg : Soa 
m~d : va"nucl I 
,'~,~t," : 0 
Figure 4: A restrictive, adverbial and an olmrator a(lverl)ial. Restrictive adverbials (such as locatives 
and time adverbials) will generally be encoded as presented, where R0 is a meta-wu'iable that is 
instantiated by the restriction introduced by the adjunct. Operator ~ulverblals (such as causatives) 
on the other hand introduce their own quantified state of aft'airs. Such mlverbials generally are 
encoded as in the following examph', of the adverbial 'toewdlig' (accidentally). Adw.,rbials of the first 
type add a restriction to the semantics of the verb; ;Ldwn'bials of the second type introduce a new 
scope of modification. 
of lexical rules by compil~tion of the lexicon, 
as it is unclear how recursive lexical rules can 
be treated in such an architecture (especially 
since sOllle recursive rules e:ut easily lead to 
an infinite number of lexical entries, e.g. tlle 
adjuncts rule). 
Another alternative is to consider lexical 
rules as 'ordinary' unary rules. If this tech- 
nique, is applied for the lexical rules we have 
envisaged here, then (unary) derivations with 
unbounded length have to be <:mlsidet'ed. 
\]f we formalize lexieal rules as (oomph,x) 
constraints on lexical categories then we are 
able to use delayed evaluation techniques for 
such constraints. 
Assume that the 'underlying' feature struc- 
ture of a verb is given by a definition of 'stem' 
(e.g. as the example of 'wil' abow'., or as the 
example of a simple transitive verb such as 
'kussen' (to-kiss) in figure 6). 
Such a feature-structure is not the actual 
category of the verb -- rather this category 
is define.d with complex eonstraints with re. 
speet to this base form. \]lere the constraint 
that adds adjuncts to the subc:tt list has our 
VEItBAI, /\[ "°U" \] 
sc : s,:,,, : a., ) \[ ou. \] 
subj : sere: A I 
scml,,,clqf.soa : kisser :A1 
kissed A'2 
l;'igure (i: (h~.tet,;ory for 'kussen' (to Idss) 
special attention, but there, is also a constraint 
tha.t adds a snbj<'.ct to the subeat list (as part 
of the in\[lection constraint for finite verbs) and 
a constraint that pushes an element from the 
subeat list to slash (to trea,t utll)ounded depen- 
dencies along the lines of eha.pter 9 of Pollard 
and Sag (in press)), etc. Thus a \]exical entry 
might be defilmd as in ligure 7. 
\],exical rules are regarded as (complex) con- 
strah~ts ill l;his framework because it allows an 
imple.mentation using delayed evaluation tech- 
niques from logic progrannning. The idea is 
253 
VERBAL 
sc : (E, tI) 
sc : ( lea: : kussen ' 
dir : right 
lee: : wil 
VERBAL 
VERBAL 
sc : (II) 
sc : ( lex : kussen 
dir : right 
lea wil 
lea : vandaag , lea: : bob , lez : arie 
dir : left E dir : left tl dir: left 
} \[N° N 1 \[ °uN \] , lez : vandaag , lex : bob , lez : arie dir: left dir: left lI dir: left 
) 
Figure 5: The finM instantiation of the moda.l for both the narrow- and the wide-scope reading of 
the sentence ~Arie Bob vandaag wil kussen'. In tit(', narrow-scope rending the adverbial occurs both 
on the subeat list of the embedded verb and on the subeat list of the matrix verb -- indicating that 
the embedded verb introduced the adjunct. In the wide-scope reading the adverb only occurs on 
the subcat list of the matrix verb. 
lexicM_entry(A) : - 
stem(B), add_adj(B, C), 
inflection(C, D), push_slash(D, A). 
inflection ( 
VEnBAL "\] FINITE 
phon : P phon : P (9 "t" 
se : Se ' sc : Sc . (Subj) )' 
subj : Subj subj : Subj 
Figure 7: A lexical entry is defined with re- 
spect to a base form using complex constraints. 
Subject addition is a constraint associated 
with finite inflection. 
that a certain constraint is only (partially) 
evaluated if 'enough' information is available 
to do so successfully. As a relatively simple 
example we consider the constraint that is re- 
sponsible for adding a sul)ject as the last el- 
ement on a subcat list of finite verbs. As a 
lexical rule we might define: \[FINITE 1 
subj: Subj ==> \[ sc : St. (Subj) \] 
8C : Sc 
If we use constraints the definition can be given 
as in figure 7, as part of the constraint associ- 
ated with finite morphology. Note that the two 
approaches are not equivalent. If we use lexical 
rules then we have to make sure that the add- 
subject rule should be applied only once, and 
only for finite verbs. As a constraint we sim- 
ply call the constraint once at the appropriate 
position. 
The concatenation constraint (associated 
with the ~dot' notation) is defined as usual: 
concat (0 , A, a). 
eoneat((BIC), A, (II\]D)) : - 
concat(C, A, l)). 
If this constraint api)lies on a category of which 
the subcat list is not yet fully specified (for ex- 
ample because we do not yet know how many 
adjuncts have been added to this list) then we 
cannot yet compute the resulting subcat list. 
The constrMnt can be suimessfully applied if 
either one of the subcat lists is instmttiated: 
then we obtain a finite miml)er of i)osslble so- 
hltions to the. colistraint. 
The relation add_adj recursively descends 
through a subcategorization list and at each 
position either adds or does not add an ad- 
junct (of the appropriate type). Its definition 
is given in figure 8. Note that it is assumed 
in this definition that the scope of (operator- 
type) adverbials is given by the order in which 
they are put in in the subcategorization list, 
i.e. in the obliqueness order. 4 
4Cf. Kasper (i,, preparation) for discussion of this 
point, also in rehttion with adjm, cts that introduce 
qmuttiflers. Note that in our approach dilR.'rent pos- 
sibilities can be defined. 
254 
I SIGN \] SIGN 1 
add_adj( sc : A sc : J 
sere : B ' sere : K ) :- 
subj : Subj subj : Subj 
add_adj(A, J, B, K). 
~dd-~dj(0, 0, A, a). 
add_adj((CID), (ClE), A, n):- 
add_adj(D, E, A, B). 
add_adj(A,( rood: arg : B 
w,l : l,; 
add_adj(A, D, I,;, C). 
ID), 1~, C):- 
Figure 8: Definite clause specification of 
'add_adj' constraint. 
3.2 Delayed evaluation 
For our current purposes, the co-routining fa- 
cilities offered by Sicstns Prolog are power- 
ful enough to implement a delayed evaluation 
strategy for the cases discussed al)ove. For 
each constraint we declare the conditions for 
evMuating a constraint of that type by means 
of a block declaration. For example the con- 
cat constraint is associated with a declaration: 
'- block coneat(-,?,-). 
This declaration says that evaluation of a c~dl 
to concat should be delayed if both the Iirst and 
third arguments are currently variable (unin- 
stantiated, of type "toP). It is clear fr<>m 
the definition of concat that if these argu- 
ments are instantiated then we can evahm.te 
the constraint in a top-down manner with- 
out risking non-termination, l!',.g, the goal 
concat((A, B), C, D) succeeds by insta.ntiating 
D as the list (A,I\]\]C). 
Note that block declarations apply recur- 
sively. If tit(: third argument to a call to con- 
eat is instantiated as a list with a wu'iahle tail, 
then the evaluation of the recursive al)l)lication 
of that goat might be blocked; e.g. ewduation 
of the goat co,~.~(A, (S j), <nit>)s.e,'.oeds ei- 
ther with both A and C instantiated as the 
empty list and by unifying Sj ;rod B, or with 
A instantiated as the list (l\]\[l)) for which the 
constraint concat(D, (Sj), C)has to be satis- 
tied. Similarly, for each of the other constraints 
we declare the conditions under which the con- 
stra.int can be ewluated. For the add_adj con- 
straint we define: 
"- block add_adj(?, -, 7, 7). 
One may wonder whether in such a,n archi- 
tecture enough information will ever become 
available to allow the evaluation of any of the 
constraints, hi general such a prol)lem may 
surface: the parser then finishe.s a derivation 
with a large collection of constraints that it 
is not ~dlowed to evaluate - and hence it is 
not clear whether the sentence associated with 
that derivation is in fad; gram m~tical (as there. 
may 1)e no solutions to these constraints). 
The strategy we have used successfitl/y so- 
far is to use the structure hypothesized by the 
parsm' as a 'generator' of information. For ex- 
ample, given that the parser hypothesizes the 
al)plication of rules, and hence of certain in- 
stmttiations of the sul)cat list of the (lexicM) 
head of such rules, this provides information 
on the subcat-list of lexical categories. Keep-- 
ing in mind the definition of a lexical entry as 
in figure 7 we then are able to ewfluate each 
of the constraints O)l the wdue of the subcat 
list in tl,rn, starting with the push_slash con- 
straint, up through the inflection and add_adj 
constraints. Thus ra.ther than using the con- 
sir.tints as q)uilders' of subcat-lists the con- 
straints :~re evaluated by checking whether a 
subcat-list hypothesized by the parser can be 
related to a sat)cat-list provided by a verb- 
stein, in other words, the \[1GW of information 
in the definition of Ie:~:ical_entry is not as the 
order of constraints might suggest (froln top 
to 1)ottom) but ratht, r the other way around 
(from hottom to top). 
4 Final remarks 
We illustrated that recursive lexic~d con- 
straints might be useful from a linguistic per- 
spectiw~. If lexlc~d rules are formalized as 
cotni)lex cot/strahlts ol) \]exica\] categories then 
methods from logic l)rogl:amtning can be used 
to imtflement such constraints. 
Note that complex CG,lstraints and delayed 
eva.huttion techniques are also useful in other 
areas of linguistic desciptlon. For example. 
we used the same methods to deline and pro- 
255 
cess IIPSG's FOOT FEATURE PRINCIPLE. The 
method may also be applied to implement 
IIPSG's binding theory. 
As a testcase we improved upon the IIPSG 
analysis of (Germanic) verb clusters and ad- 
juncts by treating adjuncts as categories that 
are on the subcat list by virtue of a complex 
constraint. The fragment that has been imple- 
mented with the methods described is much 
larger than the discussion in the previous sec- 
tions suggest, but includes treatments of extra- 
position, ipp, modal inversion, participium in- 
version, the third construction, partial-vp top- 
icalisation, particle verbs, verb-second, subject 
raising, subject control, raising-to-object, ob- 
ject control and clitic climbing in Dutch. 
References 
Bob Carpenter. The generative power of cat- 
egoriM grammars and head-driven phrase 
structure grammars with lexical rules. Com- 
putational Linguistics, 17(3):301-313, 1991. 
Erhard Hinrichs and Tsuneko Nakazawa. 
Flipped out: AUX in german. In Papers 
f~vm the 25th Annual Regional Meeting of 
the Chicago Linguistic Society, pages 187- 
202. Chicago Linguistics Society, Chicago, 
1989. 
Robert Kasper. Adjuncts in the mittelfeld. Ill 
John Nerbonne, Klaus Netter, and C~trl Pol- 
lard, editors, German Grammar in IIPSG, 
Lecture Note Series. CSLI, Stanford, in 
preparation. 
Philil) Miller. Clitics and Constituents in 
Phrase Structure Grammar. Garland, New 
York, 1992. 
Michael Moortgat. Categorial Investigations. 
PhD thesis, University of Amsterdam, 1988. 
Carl Pollard and Ivan Sag. Information Based 
Syntax and Semantics, Volume 1. Center 
for the Study of Language and Information 
Stanford, 1987. 
Carl Pollard and Ivan Sag. l\[ead-driven 
Phrase Structure Grammar. Center for the 
Study of Language and Information Stun- 
ford, in press. 
256 
