SYN"I'ACT I (7 FUNCTIONS IN f~PSG 
ka\[el OLIVA 
Facu\].ty of ~la-\[heme.tics and Physics 
Charles University 
~slostransk6 n,~.m~stl 25 
{\]S-118 O0 Praha I - ~al/~ Strana 
C zechoslovakia 
Abstract : 
This paper motivates and T~roposes 
a~J(\]irl~ ~ new feature of s sj,'ntactio function to 
the feature system of CPSG. Later , it shows 
its necessity in number of synt~ctic con- 
structs , such as passivization, extraposi- 
-\[;Jon , coordination etc. But ~ddlng such fea~ 
ture is not understood as a mere technicali- 
t.~ , an~ thus some implications for the expla- 
n~vtor)< power of the theory are also discus- 
se(! o 
1. Remarks on the Notation of PP 
T}~e notion of Prepositional Phrase has a 
fJ. rm and long standing tradition in the An~ 
/:J o-.laxor, linguistic framework , including 
r~P\[~C~. The classical Continental tradition, on 
the other h~ind, does not use this term for 
\]~nf<uage description in majority of frame- 
~.;or\]:s developed and treats prepositions pre- 
4omJnantl,v as markers of nouns. In this pa- 
y, er, I will adhere to this approach. 
The most important reason for doing 
this is the fact ~hat , on the assumption of 
existence of PP, it is impossible to generate 
s~ich sentences as (1) in the GPSG context.- 
free framework. 
(I) I went there with John and with Paul, 
who met each other before. 
The point is, that in (I), on the assumption 
~f the existence of PP, we find a PP-coordi-- 
r~tion, resulting in a PP, rather than an NP 
one. Consequently, the relative clause must 
be sister of this PP rather than of an NP. 
lint this is (at \].east in my opinion) hardly 
a~ acceptable solution, both intuitively and 
formally. Even an attempt to cope with this 
oroblom by means of rightward "Across-the- 
Poard" extraposition would not make things 
better, since such a solution would presume 
phrases as "John who met each other" to be 
correct English NP's. I do not wish to pursue 
this further here, the problem is more tho- 
roughly described in (01iva,88) from both 
intuitive and formal viewpoints. What I real- 
ly wanted to do was to motivate the decision 
of avoiding the term PP in this paper (and, 
possibly, the necessity of omitting it from 
the CPSG framewor\]e altogether), which will be 
reflected in the notation of rules. Thus , 
assuming otherwise the notation familiar from 
(Oazdar,Klein,Pullum and Sag,85) , the catego- 
ry PP {\[PFOR~ z\]} will he reoasted as 
~Pf\[PREP z \]} , which should road as "NP mar- 
ked with preposition ~ " As a special value 
of the feature PREP, the value "nil" ought to 
be introduced, with the meaning that there 
is no preposition marking the NP in question 
(i.e., the category NP~\[PREP nil\]} in the new 
494 
notation is exactly equivalent to the simple 
NP in the notation of (Gazdar ,~(lein ,Pullum 
and Sag ,85)). 
2° The Szn__tactic Function Feature 
In this paper, I shall try to show 
some of the problems connected with omitting 
the traditional notion of (syntactic) func- 
tion in GPSG, and their possible solutions° 
The payoffs of the notion mentioned can 
be seen e.g. on the classification of adver-~ 
bials: each adverbial modifier, whether ex- 
pressed by an NP {PREP} , by an adverb, by a 
subordinated clause or otherwise, has its 
more specific meaning ("function") in the 
sentence. With specific lexlcal units, e.g. 
with the verb "put", this is so obvious that 
a "semantic" classification was proposed even 
in some of the syntactical approaches, cfo 
(Bresnan and Grimshaw,78), and later trans- 
ferred into GPSG without change, as can be 
seen on the presence of the feature L0C and 
related lexical rule introducing "put" (2). 
(2) VP----~ H\[6\] ,NP,PP\[+LOC} 
Further, two more such features, viz. MANNER 
and TEMP, were, unfortunately again with no 
systematic motivation, proposed as "possible 
solution" for some problems of coordination 
in (Sag,Gazdar,Wasow and Weisler,85). These 
features LOC,MANNER,TEZP are, of course, no- 
thing else than function markers on adverbial 
modifiers (of location, of manner, of time)° 
Now the time has come to introdfice these 
modifiers more systematically; but I add two 
arguments in favour of such systemization 
before. First, the rule (2), explicitly sti- 
pulating the complement as PP (or, alterna- 
tively, as NP {PREP\] ) , does not provide the 
theory with the power to generate any of (3). 
(3)a. I have put my spectacles somewhere, 
but now I cannot recall where it was. 
b. She put it where I wanted it. 
The second argument comes again from coordi-- 
nation; neither under the traditional "cate- 
gorial harmony" nor under the newly proposed 
theory of (Sag,Gazdar,Wasow and Weisler,85), 
the contrast in (4) could be explained. 
(4) a. She was killed by her husband 
and by his mistress. 
bo She was killed by a hammer 
and by an axe. 
c. *She was killed by her husband 
and by a hammer. 
Under the "functional" approach, however, we 
find in (4)a the coordination of a_~nts 
("deep subjects"), in (4)b the coordination 
of instruments; but in (4)c an attempt to 
coordinate a~ and instrument is involved, 
which makes this string ungrammatical. 
Thus, it seems to be advantageous to add 
one more item to the feature system of GPSG, 
namely the feature of function (let it be 
called FUNC, hence) of the phrase in the sen- 
tence. Its values will mark the adverbial 
modifiers of place (lee), direction (dir), 
manner (mod), instrument (instr), purpose 
(purp), ss well as other functions of phra- 
ses, such as subject (subj), direct object 
(dobJ), indirect object (iobj), predicative 
(prd), resu\]t (eff), agent of passive senten- 
ces (ag) etc., as is usual in the Prague 
Linguistic School. Further, the example (4) 
and the related discussion guides us to add 
FUNC to the set of Head-features. 
Next, we shall consider the following 
triads o~ examples, all generated by the 
standard GPSG grammar. 
(5) a. Kim gave Sandy the book. b. 
The book was given Sandy (by Kim). 
Co Sandy was given the book (by Kim). 
(6) a. We considered him a friend. 
be He was considered a friend (by us). 
e.*A friend was considered him (by us). 
(7) a. The king made him a duke. 
he }~e was made a duke (by the king). 
e.*A duke was made him (by the king). 
The problem is clear: the Passive Metarule 
(8) applied on rules introducing NP's as sis- 
ters of V 0 even in cases these NP:s were not 
direct objects, which was not always correct. 
(8) Passive Metarule 
'4P -----> W ,NP II 
V~'{pas\]- ~--~-~ W,(~P{\[ P~,P by\]) ) 
Similarly, (8) can apply to the ID-rules 
(9)a,(10)a :introducing "there"-oonstructions 
and clefts , respectively, yielding (9)b , 
(10)b, which is again obviously wrong. 
(9)a. V.~{\[AGR NP(there,\[PLUR Z\]}\]} 
, ~r\[221 ,NP {\[PLUR 4} 
b, VP{\[AGR NP{there,\[PLUR /'\]},pas} 
H b2\] ,(NP {\[PRF, P by\]}:) 
(10)a. VP (+it} --'---> H \[4/4\] ,NP ,S{+R} 
b. VP {+it,pas\] H\[Zl4\] ,S{+R} 
,(NP{\[PREP by\]} ) 
One more such example can be constructed if 
we apply the Extraposition Metarule (tl) to 
the rule (12)a and let the Passive Metarule 
(8) apply to the result, as sketched in 
(12)a ,b ,Co 
(11) Extraposltlon Metarule 
~ (\[AGR S\]} ------> W 
x2 {lAiR ~P{it\]\]} ---%w,s 
(12)a. VP{\[AGR S\]} .... *Hb0\] ,NP 
That Lee was elected bothered Kim. b. 
VP \[\[AGR NP(it}\]} -----~ Hi20\] ,NP,S 
It bothered Kim that Lee was elected. 
s. VP {\[AGR NP{it~,pa~ 
--~ tt be\] ,s ,(NP {\[P~P by\]} ) 
*It was bothered (by Kim) 
that Lee was elected. 
Note that in (12)c, "it" serves as a dummy 
subject, i.e. it is not a personal pronoun 
which could be replaced by the referred noun 
as in "It (=the child) was bothered that Lee 
was elected.".* And even if this were not the 
case, the presence of the optional 
NP {\[PREP by\]} would in any ease rule the 
grammatJeality of such sentences out. 
Thus, some strings are generated by the 
standard GPSG grammar that are ungrammatical 
while, on the other hand, such sentences as 
(13) cannot be generated by the grammar, sin- 
9e the verb "hope" is introduced by an ID- 
rule as in (14), where the role of direct 
object is played by a subordinated clause 
rather than by an accusative NP, and thus the 
construction cannot undergo passivization via 
the Passive Metarule (8). Similar situation 
occurs with "concede" etc., cf. (15). 
(13) That he would sing 
was hoped by everybody. 
(14) VP--~ II,S{fin) 
(15) VP ~H,(~P{\[PI~m ) to\]} ) ,~\]{fin\] 
Thus we arrive to the basic insight concer- 
ning passivization: that, generally, such 
verbs may be passivized which s~bcategprize 
~gr direct object, irrespectively to the 
means by which this diFeet..pb~ect is realised 
(NP S ...): 
Another important point in "the j)assivi- 
zatlon process (as viewed in transformational 
grammars) is that the direct ob'e~ in active 
construction becomes the sub~@_~% in the pas- 
sive one. In the GPSG framework, this shift 
must be reflected in the passive counterpart 
of the active ID-rule , otherwise the grammar 
will generate (16)b as the passive counter- 
part of (16)a, while, again, i% will not be 
possible to generate (13), for the FSD11, 
assumed in (Gazdar,Klein,Pullum and Sag,85) 
and repeated here as (17), will force the 
\[AOR NPffNF0m~ norm\]}\] to be Instantiated on 
the V O and thus will rule out the possibility 
of non-NP subject occurring in {he sentence I . 
(16)a. Everybody hoped that he would sing. 
b.*He was hoped (by everybody). 
(17) FSD11: 
{iv,+\] ,\[.BAR ,0\]} ~ {\[A~R NP{\[NFOI~ norm\]}\]} 
To make the transition of (active) direct ob- 
Ject to (passive) subject explicit, we first 
mark all the direct objects in the lexical 
ID-rules in the metagrammar wlth the feature- 
value pair \[FUNC dobj\] , and, second, aug- 
ment the Passive Netarule with the require- 
ment that the passive VP agree with the cate- 
gory of active direct object. Thus we arrive 
to the shape (18) of Passive Metarule for 
Direct Objects. 
495 
(18) Passive Metarule for Direct Objects 
VPf(\[AGR X\])) -------> W,Y {\[FUNC dobj\]} 
VP{pas,\[AQR Y\]}-----~ W,t(X){\[FUNC a~\]}) 
Several remarks concerning (18): first, the 
"variable" X stands for the same category in 
the whole me~arule, and this is also 'the case 
with Y. Second, the bracket pairs "(" and ")" 
mark off mutually bound optionality when used 
around the same category - in this rule -the 
case is that of (\[AGR X\]) and (X), i.eo con- 
oerning \[AGR X\] , either there is \[AGP X\] in 
the mother of the pattern rule and then, con- 
sequen½1y, X {\[FUNC ag\]} occurs as the licens-- 
ing category of optional agent complement in 
the target rule, or there is no \[AO}~ X\] in 
the pattern rule ana then just {\[FUNC ag\]}is 
the licensing category of the optional agent, 
all other features \].eft to later instantiat- 
ion. This is important to allow for (19)a,b; 
note, however, that the verbs in %hls example 
are of different subcategorization classes 
V\[2\] and V\[20\] , respectively 2. 
(19)a. Lee was bothered by K im. 
b. Lee was bothered 
(that) Kim was elected. 
To prevent strings like those in (20) to be 
~enerated, we have to introduce FSD (21) 
which marks all categories containing 
\[~'T~NC ag\] as NP{\[PREP by\]\] ; if this does not 
contradict some other principle of feature 
instantiation (which is not the case with 
"bother" in (19)b, because here the agent is 
already marked as S by the licensing rule) 3 
(20)a. *Lee was bothered Kim. 
b. *Lee was bothered after Kim. 
(21) FSD: 
{\[FUNC ag\]}~\[N,+\], \[V,-\] ,\[PREP by\]} 
Further, to prevent (22) %o be generated, it 
is necessary to add FCR (29) to the metagram- 
mar. This move, however, together with exam- 
Dies (24) and analogs ((24)b will be of par- 
ticular interest), will lead us to a profound 
reconsideration of the roles and mutual rela- 
tions of features COMP/PREP and FUNC. The 
matter is that, given the metagrammar in its 
current shape, (24)b cannot be generated, be- 
cause FUNC is considered to be a Head feature 
and -thus, in virtue of the Head Feature Con- 
vention, \[FUNC subj\] will be forced on the 
second subject sentence; but since this will 
be lacking the \[COMP that\] specification, it 
will bemarked as unacceptable by (23). 
(22)a. *For us to be elected bothers Sandy. 
b. *Kim was elected bothers Sandy. 
(23) FCJ~: s 
{\[r~c sub j\]} ~{\[COMP that\]} 
(24) a. ThatKim was elected and 
that Sandy was dismissed bothered Lee. 
b. That Kim was elected and 
Sandy was dismissed bothered Lee. 
IIow to solve this problem ? For inspiration, 
\].et us turn to the classical and well-known 
analogy between verbs with complementizers 
and nouns with prepositions , shown in 
(25) ,(26). 
(25)a. That Sandy was dismissed bothered I, ee. 
bo*Sandy was dismissed bothered Lee° 
(26)ao Kim waited for Leslie. 
b.~Kim waited Leslie° 
At first glance, there is some disci.epancy 
between (24)b and (25)b, in respect to the 
basic idea of coordination in (Gazdar,Klein, 
Pullum and Sag,85), namely that each member 
of a coordinated construction should be able 
to appear alone in place of the entire coor- 
dination structure. The same case will be, 
naturally, that of (26)b and (27)° 
(27) Kim waited fox. Lee and Leslie. 
Of course, the matter is that in (24)b and 
(27) the COMP and. PREP, respectively, belong 
to the whole coordination, not to its first 
member solely, and -thus extracting the second 
member to stand alone as shown in (25)b,(26)b 
is wrong - it is, in fact, not the whole 
second member being extracted; one its indi- 
visible part gets lost somehow, and thus it 
is only natural that the resulting string is 
ungrammatical. The second observation %o be 
made is by far more important in its scope: 
from the comparison of examples (25)a vs. 
(25)b and (26)a vs. (26)b, it i.s clear that 
the role of constituent having some function 
in the sentence can be played only by a ful- 
i/f__~eeified phrase of the appropriate type 
(i.e., in (25), the role of (sentential) sub- 
ject by a sentence containing "that"-comple- 
ment, in (26), the role of (prepositional) 
object by a noun phrase containing the con- 
cerned preposition). And the same holds also 
vice versa, i.e. no underspeclfied phrase can 
play the role of the constituent having the 
function in question. In other words, when- 
ever the complement or preposition is expan.- 
ded in the generation process, the "rest" of 
the phrase must not be any more marked as 
having the function the "whole" phrase has. 
This is just what is expressed in the COMP 
and PREP expansion rules (28)a,b. 
(28) Z~ Inil 
l\[coMP *U \[\[coMP ni~/ 
b. N~\[FU'C~\]I--~{\[S,~CAT ~J}, N~iF~NC hi@ 
\[\[PREP ~\]\] I\[PREP ni\]l\] 
Not only that these rules cope correctly wlth 
all problems which we mentioned in this res- 
pect, but (more importantly) they shed light 
on the mutual relations between oom~lementi- 
zers ~ositions and syntactic functions~ 
that th_~ar~ust mirror images of eaoho%- 
which cannot exist ~. 
496 
3. Conclusions 
It was shown that a new feature denoting 
syntactic function of a phrase should probab- 
ly be added to the feature system of GPSG, 
which will help make the theory more ade- 
quate. This feature proved to be useful not 
only in some technicalities, but even in the 
intuitive and explanatory background of the 
theory. The bulk of support material for 
this new feature was taken from passiviza- 
%ion of constructions with direct object in 
this paper, but similar evidence for the new 
feature can be drawn also from other types of 
passivization (of constructions with indirect 
objects, p:ceposltional complements ("A doctor 
was sent for.") or adjuncts ("This bed has 
been slept in.")), as well as from other 
types of extraposition (cf. the impossibility 
of generating sentences like "It w~s hoped by 
everybody that he would sing." in the stan- 
dard GPSG) and yet other sources. 
Adding the syntactic function feature to 
"the GPSG (mete)grammar goes in certain as- 
pects in parallel with the independently mo- 
tivated efforts to eliminate metarules from 
GPSG altogether and/or to shift their work to 
lexical redundancy rules, e.g. (Pollard,85), 
(Kilbury,86). This fact only confirms the old 
wisdom that, given some linguistic informat- 
ion, the Question of dividing it between 
grammar and lexicon may be decided rather ar- 
bitrarily. The i~ortant point is that the 
proposals made here do not require such radi- 
cal changes in the GPSG framework as the "le- 
xlcal" approaches do, i.e., they keep better 
the basic spirit of the current GPSG (while 
solving the problems shown at least equally 
well). The only points in which a really sub- 
stantial change occurred in this paper was, 
first, the structure of PP, and, second, the 
stipulation that purely morphological chara- 
cteristics of constituents is insufficient 
for the description of English syntax. Put I 
hope that in the light of the data presented 
these moves appear to be fully Justified. 
Notes: 
I am indebted to Eva Haji~ov&, Petr Sgail, 
Alexandr Rosen and an unknown Coling referee 
for fruitful comments on the first version of 
this paper, as well as to Laura Janda for a 
lot of the English examples. 
I) The status and validity of this FSD is 
however, somewhat unclear: it forces 
\[AGR NP {INFORM norm\]}\] to all verbal predica- 
tes which are ,lot marked otherwise in the 
respective ID-rule. Consequently, such verbs 
as "bother" must be listed in the lexicon 
twice', each time with different subcategori- 
zatlon class, but in both cases subcategorlz- 
ing fox' identical complements (and, hence, 
there have to be two almost identical lexical 
rules in the grammar, one for each of the 
classes). This, in turn, makes the (otherwise 
isimple and clear) sentence (i) ambiguous, 
while on the other hand, due to this FSD none 
of the sentences in (ii),(iil) can be gene- 
rated. 
(i) Kim was bothered. 
(li) That John would not come was obvious. 
That Kim was elected 
gave Sandy the satisfaction. 
(iii)It was obvious 
that John would not come. 
It gave Sandy the satisfaction 
that Kim was elected. 
On closer observation, a lot of verbs can 
(appropriate pragmatic conditions provided) 
take a sentential subject: the ~brute force" 
method for coping with this fact would be to 
increase the size of rules and lexlcal en- 
tries , as performed with "bother" and ana- 
logs in (Oazdar,Klein,Pullum and Sag,85). 
More appropriate solution (at least in my 
opinion) would be to soften the FSD11 to the 
form (iv), and simultaneously to stipulate 
that only verbs 'taking exclusively sententlal 
subject (if they exist) or exclusively nomi- 
nal subject (e.g. "believe","hope") would be 
marked as such° 
(iv) FSD11 : 
{\[v ,+\] ,\[ BA~ ,o1} ( {\[AGR NP{\[NF0im norm\]\]\]t or {\[AGR S\]} ) 
Generating (iii), however, requires further 
replacing the \[AGE S\] specification in the 
mother of the pattern rule in the Extraposi- 
tion Metarule wlth~ \[AGR NP\] specification. 
(v) Extraposition Metarule 
X2{~\[A~R NP\]} ~ w 
x a I\[A~R NP {itl}l} ---+ w ,s 
2) But see the previous comment for some dis °- 
cusion of this point. 
3) However, to keep consistency with the 
proposals from I), we had better to introduce 
a FCR (i) instead of FSD (21) 
(1) FGR {\[FUNC ag\]} 
~(NP {\[PREP by\]} or S {\[CO~rP that\]} ) 
References
Bresnan J.W. and Grimshaw J.: The syntax of 
free relatives in English, 

Linguistic Inquiry 9, pp.331-391 , 1978 

Gazdar G. , Klein E., Pullum G. and Sag I.: 
Generalized Phrase Structure Grammar, 
Basil Blackwell, Oxford 1985 

Kilbury J.: Category Cooccurrence Restric-~ 
tions and the Elimination of Metarules, 
in Proceedings of Coling'86, Bonn, 1986 

Olive K.: GPSG from a Praguian Viewpoint, 
Part I, in The Prague Bulletin of Mathema- 
tical Linguistics 49, Prague, 1988 

Pollard C.: Phrase Structure Erammar without 
Metarules, 'in Proceedings of the 4th West 
Coast Conference on Formal Linguistics, 
Stanford University, 1985 

Sag I., Wasow T. , Gazdar G. and Weisler S.: 
Coordination and How to Distinguish Catego- 
ries, in Natural Language And Linguistic 
Theory 3, 1985. Also appeared as CSLI Re- 
port No. 3, Stanford, 1984 
