Category Cooccurrence Restrictiorls 
and the Elimination of Metar1~les 
O. Introduction 
James Kilbury 
Technical University of Berlin 
EIT/NASEV, CIS, Sekr. FR 5-8 
Franklinstr. 28/29 
D-IO00 Berlin 10 
Germany - West Berlin 
This paper builds upon and extends certain ideas 
developed within the framework of Generalized Phrase 
Structure Grammar (GPSG). ill A new descriptive 
device, the Category Cooccurrence Restriction (CCR), 
is introduced in analogy to existing devices of GPSG 
in order to express constraints on the cooccurrence 
of categories within local trees (i.e. trees of depth 
one) which at present are stated with Immediate 
Dominance (ID) rules and metarules. In addition to 
providing a uniform format for the statement of such 
constraints, CCRs permit generalizations to be 
expressed which presently cannot be captured in GPSG. 
Sections l.l and 1.2 introduce CCRs and presuppose 
only a general familiarity with GPSG. The ideas do 
not depend on details of GPSG and can be applied to 
other grammatical formalisms. 
Sections 1.3 - 1.5 discuss CCRs in relation to 
particular principles of GPSG and ass~ne familiarity 
with Gazdar et al. (\].985) (henceforth abbreviated 
'GKPS'). Finally, section 2 contains proposals for 
using CCRs to avoid the analyses with metarules given 
for English in GKPS. 
1. Category Cooecurrence Restrictions (CCRs) 
I,i The Principle of CCRs 
The reasons for proposing CCRs to state 
restrictions on the eooccurrence of categories within 
local trees are analogous to those for introducing 
Inmlediate Dominance (ID) and Linear Precedence (LP) 
rules in GPSG (of GEPS, pp. 44-50). A context free 
rule binds information of two sorts in a single 
statement, namely 
(a) information about which daughters a rook has 
in a local tree and 
(b) information about the order in which the 
daughters appear. 
By separating this information in ID and LP rules, 
GPSG is able to state generalizations of the sort "A 
preceeds B in every local tree which contains both as 
daughters," which cannot be captured in a context 
free grammar (CFG). 
Now consider an ID rule such as the following: 
(i) S --> A, B, C 
The fundmnental motivation for CCRs rests on the 
insight that such an ID rule itself combines two 
different kinds of information in a single statement, 
namely 
(a) information involving immediate dominance 
relations, here that <S, A>~ <S, B>, and 
<S, C> are ordered pairs of categories in 
which the first category inmlediately dominates 
the second and 
(b) information about the cooccurrence of 
categories in a single local tree. 
By distinguishing and separately representing these 
types of information it becomes possible to state 
generalizations of the following sort, which cannot 
be captured in the ID/LP format: 
(2) Any local tree with S as its root must have A as 
a daughter. 
(3) No local tree with C as a daughter also has D as 
a daughter. 
Statements such as (2) and (3) restricting the 
cooccurrence of categories in local trees are 
Category Cooccurrence Restrictions, which are 
expressions of first arder predicate logic using two 
primitive predicates, R(cx, t) 'cx is the root of local 
tree t' and D(~, t) 'a is a daughter in local tree 
t'. \[2\] CCRs have the form Vt: ~, where 1T :is a 
schema and the notion of a possible schema is defined 
as follows: 
(i) (R(a, t)) and (D(~,t)) are of form g; 
(it) if ~ is of form g, then (~) is of form n; 
(iiJ) if Ip and x are both of form I~, then (~0Kr) 
is of form ~, where K C (A, V, D, e}; 
(iv) constants designating categories occur as 
first arguments within all coastituent 
predicate expressions; 
(v) the same variable t bound by the quantifier 
Vt occurs as second argument within all 
constituent predicate expressions; 
(vi) these are all expressions of form ~. 
Parentheses may be omitted following the usual 
conventions in predicate logic. 
A CCR Vt: u may be rewritten in conjunctive 
normal form as Vt: ~ ^ ... A ~ , where each clause 
ot posltlve and predicate <Pi is a disjunction . . n negated 
expressions, which is equivalent to 
V t: \[p\] ^ ... AVt: ~ , i.e. a conjunction of simple 
CCRs. -Let 0}, be an e~pression of form n containing 
\[I\] I wish to thank Gerald Gazdar, Christa 
Hauenschild, William Keller, Daniel Maxwell, 
Manfred Pinkal, and Hans Uszkoreit for their 
comments on earlier versions of this paper. This 
work was carried out under the financial support 
of the BMFT of the German Federal Government. 
\[2\] Interpretations of R(~, t) and D(a, t) in terms 
of the theory of feature instantiation in GKPS 
would be 'the root of local tree t is an 
extension of ~' and 'some daughter in local tree 
t is an extension of ~'. 
50 
only the predicate D; then simple CCRs \[3\] have the 
following forms: 
(4) Vt: R(~, t) \] (0' iff a I\[c0\]l 
Vt: 00' :\] R(~, t) iff Iko\]l 
Vt: ¢0' iff I\[0)\]1 
iff ~ I\[.--~W\]l 
Quantification is ignored in the notation on the 
right; ~ replaces P(a, t) and ~P(~, t) and ~-~ 
replaces ~P(a, t) giving 0o from (0', where P = R or D. 
The special brackets ' I\[ \]1 ' enclose daughters and 
render the indication of material implication 
superfluous. Using this notation, (2) and (3) may be 
restated as (5) mid (6), respectively: 
(5) S }\[ A \]l 
(6) I\[ C~D \]l 
To reformulate a set of ID rules we thus need (a) 
the definition of a set of branches constituting 
mother-daughter pairs and (b) an appropriate set of 
CURs. The definition of branches is permissive in the 
sense in which ID rules are permissive (cf GKPS, p. 
76): branches with a conmmn mother can be adjoined to 
form a local tree. CCRs, like the LP rules, which 
also apply to local trees, are restrictive and limit 
the class of local trees admitted by the grammar. 
\[4\] How sets of ID rules may be reformulated in this 
manner will be illustrated in the following section. 
1.2 Examples of CCRs 
GKPS (pp. 47-49) exm, ines sets of simple context 
free rules and then proposes strongly equivalent 
descriptions in ID/LP format. One set of ID rules 
resulting from this reformulation is given in (7): 
(7) S -> NP, VP VP -~ V, VP 
S -> AUX, NP, VP VP ~> V, NP 
VP -> AUE, VP VP -> V, NP, VP 
The ID rules of (7) admit local trees whose brancbes 
are among the following: 
(8) <S, NP>, <S, VP>, <S, AUX>, 
<VP, V>, <VP, VP>, <VP, AUR>, <VP, NP> 
Since none of the local trees admitted by (7) has 
more than one occurrence of a given category as 
daughter, we may say that the gran~ar first admits 
any strictly linearly ordered set \[5\] of branches 
\[3\] If categories are assumed to be atomic (e.g. S, 
NP, V) rather than complex for the moment, then 
it is unnecessary to mention more than one root 
category in a given CCR, 
\[4\] Note that the distinction of permissive vs. 
restrictive statements is closely related to 
that of inherited vs. instontioted feature 
specifications in the feature instantiation 
principles of GPSG. The theory would appear to 
gain in simplicity if a way could be found to 
eliminate these distinctions. 
\[5\] In order to simplify the present exposition, 
that share a conmmn mother as a local tree. This set 
of local trees must then be filtered with appropriate 
CURs so as to characterize the same set of local 
trees admitted by (7). 
A single CCR covers the trees with S as root: 
(9) CCR 1: S \]\[ NP ^ VP \]l 
CUR 1 states that NP and VP are obligatory in any 
local tree with S as its root. Since <S, AUX> is also 
a branch, MIX may optionally occur as daughter in 
such a tree. 
To characterize the local trees with VP as root we 
first construct the following function table: 
(I0) line 
1 
2 
3 
4 
5 
6 
7 
8 
9 
10 
ll 
12 
13 
14 
15 
16 
VP AUX VP V 
0 1 l 1 
0 1 1 1 
0 1 i 0 
1 1 i 0 
0 l 0 I 
0 I 0 1 
0 i 0 0 
0 i 0 0 
1 0 1 \] 
1 0 1 1 
0 0 \] 0 
0 0 1 0 
\] O 0 i 
0 0 0 i 
0 0 0 0 
0 0 0 0 
NP 
A "l" under a category (,n the right side of the table 
indicates that the category is a daughter in a given 
local tree; "0" means it is absent. If a local tree 
with the root on the left side of the table and the 
daughters marked "1" in a given line is to be 
admitted by the grarmuar, then a "1" appears under the 
root category in the corresponding line; "0" 
indicates that the tree is not admitted. 
A corresponding CCR of the form VP I\[('~\]\] can now 
be formulated, where (0 is a Boolean expression in 
conjunctive normal form. The terms of co are 
constructed from the lines designating inadmissible 
trees as follows: 
(II) ~(AUX A V) iff (~AUX v ~V) lines I, 2, 5, 6 
~(AUX A ~V A NP) iff (~AIJX v V v ~NP) 3 & 7 
~(AUX A ~VP) iff ('~AIIX v VP) 5 - 8 
~(~AUX A ~V) iff (AUX v V) II, 12, 15, 16 
~(~VP A V A ~NP) iff (VP v ~V v NP) \]4 & 16 
The normalized terms of (Ii) are conjoined in the CUR 
of (12), which is reformulated with conditionals in 
(13) and then simplified in (14)'. 
(12) VP I\[(~AIIX v NV)^ (,-AUXv Vv ~NP) A 
(~AUR v VP) a (AUX v V) ^ (VP v~ V v NP)\]\] 
(13) VP \[\[(AUX D~V) A (hUX ~ (V v~NP)) ^ 
(AUX DVP) A(AUX vV) A (V ~(VPvNP))\]\[ 
(14) VP \[\[(AUR e~V)A (AUXD(VPA~NP))A (V~ (VPv NP))\]I 
local trees are assumed to contain 
than multisets of daughters. 
sets rather 
51 
Next, (14) is reformulated as the three OORs of (15), 
which taken together with (8) and (9) admit the same 
set of local trees as the ID rules of (7): 
(15) CCR 2: VP \[\[ AUg ~ ~V \]J 
CCR 3: VP I\[ AUg ~ (VP ^ ~P)) \]t 
CCR 4: VP \[\[ V D (VP v NP) \]\[ 
The CCRs of (15) have been formulated only on the 
basis of VP trees, however, and therefore fail to 
capture generalizations that apply to all local 
trees. In particular, any local tree with AUX as 
daughter - regardless of its root - must have a VP as 
sister, so CCR 3 may be restated as two simpler CCRs, 
CCR 2' and CCR 4', where CCR 2' does not depend on 
the root category. Furthermore, CCR 4 can be 
rewritten as CCR 5' since V cannot be a daughter of 
S. The following final set of CCRs thus emerges: 
(16) COR I': S \]\[ NP A VP \]\] 
CCR 2': \]\[ AUX ~ VP \]\[ 
CCR 3': VP \]\[ AUX e ~Y \]\] 
CCR 4': VP l\[ AU× D,--NP \]l 
CCR 5': \]\[ V m (VP v NP) \]\] 
It may first appear that the description with CCRs 
in (8) and (16) constitutes no clear gain over the ID 
rules of (7). The latter, however, are highly 
redundant and express none of the generalizations 
achieved in (16). Furthermore, the replacement of ID 
rules with CCRs is the essential prerequisite for the 
elimination of metarules described in section 2. 
\].3 The Complement-Type Principle 
The ~ttempt to replace all ID rules with 
individual CCRs would lead to very complicated 
descriptions. Fortunately, the idea of CCRs can be 
utilized in a general principle that replaces all 
.lexJca\] ID rules (i.e. those which have a head that 
is an extension of a SUBCAT category; cf GKPS, p. 
54), so that only nonleA'ical ID rules need be 
explicitly reformulated with individual CCRs. 
Shieber (1983) and Pollard (1985) have proposed 
that a list- or stack-valued feature (SYNCAT or 
SUBCAT) be introduced whose value contains the 
complements of a head category. This paper uses TYP 
as a syntactic feature with a semantically oriented 
and lexJcally determined semantic type as its value. 
Following the convention of GKPS (p. 189), '<~, B>' 
will be written for <TYP(~), TYP(~)> where ~ and 
are categories. Given the structure of complex types 
in GKPS as single-valued functions, the types may be 
viewed as lists or stacks. 
A Complement-Type Principle (CTP) can now be 
stated which has the form of a schematic CCR with 
conditions on variables: 
(\]7) (a)\]IX\[BAR 0, +H, TYP 
<6 I, <...<Sn_ l, 5n>...>>\] \]1 X\[TYP 5n\] 
(b) \[IX\[BAR 0, +H, TYP <61, <...<6n_l, 6n>...>> \] 
X\[TYP 51\] ^... ^X\[TYP 5n_l\]^ ~X\[TYP 6'\] \]I 
where 
(i) 5' ~ {61 ..... 5n_ l} ; 
(it) the mother X\[TYP 6n\] and head daughter 
X\[BAR 0, +HI are both ~\[CONJ\] ; \[6\] 
(iii) (a) and (b) are simultaneously fulfilled 
for a given assignment of types to 51, 
.... 5n_l, 6 n for 1 < n. 
CTP allows the complements of a head category to be 
read off from its semantic type if its mother is 
known. According to CTP the lexical head category 
V\[SUBCAT 46\] with type <VP\[-AUX, BSE\], <NP, S>> for 
the verb do has complement sisters VP\[-AUX, BSE\] 
and NP if its mother is S but has just the complement 
VP\[-AUX, BSE\] if its mother is VP, which has the type 
<NP, S>. The use of CTP in dealing with metarules 
will be shown in section 2 below, but first another 
general aspect of the metarule problem must be 
discussed. 
1.4 Metaru\]es and Lexical Rules 
GKPS introduces not only metarules, e.g. the 
Passive Metarule (p. 59) and the Extraposition 
Metarule (p. 118), but also related lexical rules 
involving the same phenomena, e.g. the Lexieal Rule 
for Passive Forms (p. 219) and the Lexical Rule for 
Extraposition Verbs (p. 222). The lexJeal rules are 
not fully formalized but all state roughly that if a 
given lexeme has a certain category, translation, and 
semantic type, then a particular form of the lexeme 
has a corresponding category, translation, and type. 
Since lexieal rules do most of the work, and given 
that metarules apply only to Je_vSeaJ ID rules, it 
is unclear why both should be needed for what is 
essentially one job. \[7\] 
CTP in fact allows the reduction of both devices, 
metarules and lexical rules, to one, here termed 
'metalexical' (ML) rules. The latter are schematic 
rules of the form s =& ~, where a and 8 are category 
schemata which may contain variables in feature 
values. Ignoring the semantic translations of 
lexemes for the present, a ML rule states that if the 
lexicon con£ains an entry assigning ~ to lexeme w, 
then it also contains an entry assigning ~ to w; 
morphological rules determine the particular word 
form of w on the basis of syntactic features in the 
category. ML rules thus provide for an inductive 
definition of the lexicon. They handle not only 
phenomena like passive and extraposJtion but also, 
e.g. the subcategorization of sdng with or without 
an indirect object:, transitive or intransitive, etc. 
Examples follow in section 2, but next the entire 
formalism should be briefly summarized. 
\[6\] The restriction that both categories be ~\[CONJ\] 
(i.e. unspecified for CONJ) is necessary for 
coordination. In the structural analysis of 
bought and read books NP is the complement of 
the V dominating bought and read but not of the 
V dominating read. 
\[7\] Uszkoreit (1984, p. 65) has already expressed a 
similar view. 
52 
1.5 Sunmlary of the Formalism 
The syntactic formalism proposed here proceeds by 
describing items (feature names and values, feature 
specifications, categories, and trees) with 
statements restricting the distribution of lower- 
level items within next-higher-level items. Feature 
nantes and atomic values are primitives. Complex 
feature values are categories or semantic types. A 
feature specification is an ordered pair <f, v> 
containing a feature nmne f and value v, where the 
latter is restricted by the feature-value rmlge of 
the former, h category is a set of feature 
specifications such that no feature name is assigned 
more than one value; it is legal iff it fulfills all 
Feature Cooccurrence Restrictions. h local tree is an 
ordered pair consisting of a legal root category and 
a list of legal daughter categories such that (a) the 
Complmnent-Type Principle, (b) tile Category 
Cooecurrence Restrictions, and (c) tile Feature 
Instmltiation Principles (i.e., respectively, 
lexical, nonlexical, and universal statements in the 
form of CCRs) as well as tile Linear Precedence 
statements are fulfilled. \[8\] A tree is an ordered 
pair consisting of a legal root category and a list 
of daughters, where each dau~,ter is either a tree or 
a word form. Word forms and their lexieal categories 
are specified by tile lexicon, defined by a list of 
basic lexical entries and metalexical rules. 
The gramlmar defines two binary relations over 
categories, ID and LP (the latter constitutiag the 
Linear Precedence statements). A binary relation R ~ 
is the extensional closure of R iff for each <~, g> 
in R, R ~ toni:sins every <y, 6> such that y and 6 are 
extensions (oF GKPS, p. 27) of ~ and 6, respectively. 
A local tree with root C and daughters C~, ..., C 
• . .o ~ n must fulfill him condltlons that <C , C.~ E ID= for 
1 < i < n and <C., C.> ~ LP E+ (i.e? t~e transitive %- 1 . 
extensional elo~ure of LP) where 1 < 1 < n-1 and 
j = i+l. 
The proposed formalism utilizes more restricted 
memm tilmt GPSG but offers greater possibilities for 
expressing generalizations. The el iminat :ion of 
metarules and the introduction of CCRs give it a taore 
Ii~uogeneous struct:are and place cooccurrence 
restrictions of various kinds in the center of 
attention. 
For the present it may be best to regard this 
formalism as a particular variant of GPSG since most 
of tile central notions of the latter are retained. 
All that is sought is a simplification of GPSG as 
described in GKPS. Given the ricll palette of 
formal.tams recently proposed for kinds of unification 
gra~mlar~ it seems rather ingenuous to create a new 
name for thin modification of GPSG, as though tile 
multitude of remaining open questions were thereby 
answered. What we need is a metaformalism that will 
relate the insights of all the current formalisms 
through formal invariants preserved under translation 
from one formalism to another, and that will then 
truly deserve a name of its own. 
\[8\] The assmnption here is that any work done by tile 
Feature Specification Defaults (FSDs) of GKPS can 
be accomplished with suitably defined FCRs and 
CC~s. This will he illustrated in section 2 hut 
cannot be shown in general in this paper. 
2. The Elimination of Metarules 
2.1 General Remarks 
GKPS allows metarules to be used in ways that 
intuitively seem undesirable. For example, a metarule 
may simply indicate that. a daughter h of S is 
optional : 
(lS) ( s -> w, h ) -~ ( s--~ w ) 
The metarule is superfluous if A is enclesed in 
parentheses in tile corresponding ID rules: 
(19) S --> (A), B, C 
S --> (A), B, D 
Single optional elements in the RHS of ID rules are 
permitted but have no theoretical status. Here the 
generalization is lost, however, that A is optional 
in all expansions of S. 
The Complement ~,ission Metarule proposed in GKPS 
(p. 124) is similar: 
(20) \[+N, BAR 1\] --b H, W 
\[+N, BAR 1\] .e H 
This metarule can be avoided \[9\] by simply adding tile 
target of the metarule to the set of base TD rules: 
(21) \[,N, Bhl~ 1\] ~ H 
But the formalism of GKPS does not permit more than 
one clement, to be enclosed in parentheses, so the 
following cannot he an II) rule: 
(22) S -~ A, (B, C) 
Aside from the use of parentheses to indicate 
single optional elements, none of tile ahhreviatory 
conventions proposed in Chomsky/Hal \]e ( 1968, pp. 
393-399) are enlployed in GPS(\]. Thus, the rules of 
(19) cannot be abbreviated with braces as in (23): 
(23) s-> (A), B, ( 
C 
) 
D 
Since such abbreviatory conventions for expressing 
coocurrence restrictions are not provided by GPSG, it. 
is not ~.mrprising that timir work is assumed by 
metarules. GEPS in fact ,!~tates that metarules "amount 
to notifing more than a novel type of rule-collapsing 
convention for rules" (p. 66). 
Now that CCRs have been presented above in section 
1.2 for restating a simple GPSG tllat does not contain 
metarules, we Call examine |low they may be used tO 
e\]inlinate metarules fram the GPSG proposed for 
English in GKPS. 
\[9\] Note that tile metarule does not provide for the 
omission of a sJnK.le complement from a Kraal 
ol" money to tile linffuists or gratefu\] to the 
;ttJnJstrV /br ~he money. 
53 
2.2 The Passive Metarule 
GKP8 (p. 59) presents a Passive Metarule (PM) of 
remarkable simplicity and generality: 
(24) VP --> W, NP 
VP\[PAS\] -@ W, (PP\[by\]) 
PM states that for every lexical ID rule expanding VP 
and containing NP and any multiset W of categories in 
the RHS, there is a corresponding lexical ID rule 
expanding VP\[PAS\] and optionally containing PP\[by\] in 
place of NP in its RHS. Although the head V dominated 
by VP\[PAS\] is not mentioned in PM, it must be 
specified <VFOEN~ PAS> in a local tree by virtue of 
the Head Feature Convention. 
As noted in section 1.4, however, PM does only a 
small part of the work for passive, the main task 
"falling to the Lexical Rule for Passive Forms. 
Moreover, some of the predictions of PM are 
incorrect. Thus, PM applies to the lexical ID rule 
introducing V\[SUBCAT 20\], to which bother belongs: 
(25) VP\[AGR S\] --> HI20\], NP 
But the derived ID rule for V\[20, PAS\] incorrectly 
allows a PP\[PFORM by\] complement. \[10\] Furthermore, 
sentences like That Santa Claus exists .is believed 
by Kim. are grammatical, but PM does not: apply to 
the lexical ID rule introducing V\[SUBCAT 40\] for 
beldt:.ve : 
(26) VP --> H\[40\], S\[FIN\] 
Let PAS be a Boolean-valued feature restricted to 
\[+V, -N\] categories. Then we may state the following 
Metalexical Rule for Passive Forms: 
(27) V\[-PAS, AGR 6n, TYP <61,<...<6n_l,<6n,S>>...>>l 
g 
V\[+PAS, AGR 6n_l, TYP <6~I,<61,<...<6n_i,S>...>>> \] 
and 
V\[+PAS, AGR 6n_\], TYP 461,<...<6n_\],S>...>> \] 
where 
(i) 6 n-l' 6 n 6 {NP, S} and 
(ii) :if 6 = NP then 6' : PP\[by\] else 6' = S. 
n n n 
Note that 6 , and 6 are the categories of the 
--1 . n direct, object! and subject of V\[-PAS\], respectively. 
By CTP V\[-PAS\] with mother VP (of type <6 , S>) has 
complements 6 , .... 6 , while V\[+PAS\] nith mother 
VP (of type <~ ,, S>) hnsleomplements 6 ..... , 6 n- t , l n-z 
and, opt*onally, 6 n. 
\[I0\] V\[PAS\] is specified <SUBCAT, 2> in ,h~n was 
bothered ~Y his boss. 
2.3 The 'Subject-Aux Inversion' (SAI) Metarule 
The second metarule for English discussed in GEPS 
is the 'Subjeet-Aux Inversion' (SAI) Metarule (pp. 60-65): 
(28) V2\[-SUBJ\] "--> W 
V2\[+INV, +SUBJ\] --~ W, NP 
This applies to all lexical ID rules expanding VP. 
\[II\] Because of (29), however, local trees are 
admitted only by derived IB rules produced by its 
application to base lexical ID rules expanding 
categories specified VP\[+AUX\]: 
(29) \[+INV\] = \[+AUX, FIN\] (FCR 1) 
Most of the work of this metaru\]e can be taken 
care of simply by the CTP since a lexical head Y with 
the type <6 I, <...<6n, <hiP, S>>...>> has the 
complements 61, ..., 6 if its mother is VP (of type 
<NP, S>) and {he complements &l' '''' 6n, NP if the 
mother is S. Further restrictions must determine when 
V has which mother. In addition to the FCRs of (29) 
and (30), retained from GKPS, the new FCR of (31) is 
introduced: 
(30) \[+INV, BAR 2\] D \[+SUBJ\] (FCR I0) 
(31) \[INV\] ~ \[+V, -N\] 
INV is a HEAD feature subject to the Head Feature 
Convention (cf GKPS, pp. 94-99), so a V 2 mother of 
V\[+INVJ must be specified <INV, +> and therefore also 
<SUBJ, +>. If V is specified <INV, -> (note that (31) 
requires it to have some specification for INV), then 
its mother is *lot an extension of V 2 (providing for 
coordination) or it is specified <SUBJ, -> according 
to the following CCR: 
(32) I\[ V\[-INV\] 31 (~V 2 v \[-SUBJ\]) \[12\] 
Although GKPS provides for ,an embedded inverted 
sentence in What dJd you see? , no embedded nonhead 
S is specified <INV, +>. This fact is captured with a 
CCR: 
(33) (\[ ~S\[-H, +INV\] \]1 
A special Feature Specification Default to account 
for the distribution of INY (cf CKPS, pp. 30-31) then 
becomes unnecessary. 
\[ll\] 
\[12\] 
Recall the use of aliases in GKPS (p. 61) 
whereby 'VP' stands for V2\[-SUBJ\] and 'S' for 
V2\[+SUBJ\]. 
Note that this CCR contains a disjunction of 
root descriptions and thus does not conform to 
the schemata for simple CCRs with atomic 
categories presented in section 1.1 above. The 
disjunction is to he read "the root is not an 
extension of V 2 or it is an extension of 
\[-SUBJ\]." 
54 
2.4 The Extraposition Metarule 
GKPS (p. 118) proposes the following metarule to 
handle extraposit ion: 
(34) X2\[AGR S\] --> W 
# 
X2\[AaR NP\[it\]\] --> W, S 
The metarule correctly predicts sentences like .It 
bothers John that Kint drinks, because it applies to 
the lexical IB rule introducing V\[SUBCAT 20\] for 
bother : 
(35) W\[A(m S\] --> II\[ZO\], Nr 
To allow It" is s_lppa, rent that Kim drinks. , however-, 
it must also apply to the icxical IB rule introducing 
A\[SUBCAT 25\] for apparent : 
(36) Al\[AGll S\] -) H\[25\], (PP\[PFORM to\]) 
Both case;~ can he covered with CTP if the Lexical 
Rule for" Extraposition Verbs of GKPS (p. 222) is 
replaced wit:h the following metalexieal rule: 
(37) \[+Y, BAll O, AGR S, TYP <61,<...<6n,<S,S>>...>>\] 
\[+V, BAR 0, AGR NP\[iL\], 
TYP <61,<...<6n,<S,<NP\[it\],S>>>...>>\] 
The Extraposition Metarul e of GKPS is then 
superfluous. 
2.5 Slash T,~rmination Metarules 
Slash Termination Metaru\]e 1 (STM1) (of GKPS, pp. 
142-1.44) is of particular interest because of its 
generality: 
(38) X -> W, X 2 
1~ 2 
X --> W, X \[+NULI,\] 
It applies to any \]exical ID rule with a category 
specified <tIAR, 2> in the RItS and produces a rule 
with the specification <NULL, +> added to this 
category. 
It turns out that S'1%1\] lnay be eliminated with two 
simple statements. An FCR expresses the fact that a 
category is ~pecified for NUI,L (i.e. NULL takes the 
value + or -) if and only if it also is <BAR, 2>: 
(39) \[NULLI -= \[BAR 2\] 
A CCR then stipulates that a category specified 
<NULL, +> mu'4t have a lexical category as its sister 
in a local tree: 
.(40) It \[+NULL\] m \[BAR 0\] \]1 
This is equivalent to the condition that STN1 - like 
all metarule:~ - may only apply to lexical IB rules. 
Note that a root category is not indicated in (40) 
and that parasitic gaps (of GKPS, pp. 162 if) are 
provided for. 
As in GKPS, an FCR requires that a category 
specified <NUI,L, +> also be specified for SLASH: 
(41) \[+NULL\] ~ \[SLASH\] (FCR 19) 
The distribution of SI,AStI is in turn governed by the 
CAP, HFC, and FFP. GKPS also postulates an FSD for 
NULl, : 
(42) ~\[NULL\] (FSD 3) 
FgD 3 is not required in tills analysis since 
categories specified <BAR, 2> are freely specified 
with values from (4, -} for NULl,, while all other 
categories must he unspecified for NULL according to 
(39). 
The treatment of gaps in GKPS is completed with 
the Slash Termination Metarule 2 (STM2) (cf GKPS, pp. 
160:162) : 
(43) X --> W, V2\[+SUBJ, FIN\] 
X/NP -> W, vP\[-sUBJ\] 
S~'M2 says that for every \]exical ID rule introducing 
V ~ \[+SUB J, FIN\] as a daughter, there is a 
corresponding rule with V2\[-SUBJ, FIN\] in place of 
V 2\[~SUBJ, FIN\] and with the mother specified 
<SLASI{, NP>. 
Examination of the lexical \]D rules proposed for 
English in GKI)S reveals that all @\[FIN\] daughters 
introduced are also specified <SUB J, +>. We may 
therefore reformulate tim types of the \]exical head 
cat.egories so that V 2\[FIN\] complements do not carry 
the specification <SUBJ, +>. The feature SUBJ is then 
freely specified but restricted by the FCR in (44) 
and the CCR in (45): 
(44) \[SUBJ\] r~ \[-~V, -N, lIAR 2\] 
(/15) 1\[ \[BAR 0\] ^ V 2 \[-SI, JI'IJ, FIN\] \]1 X/NP 
The CCR of (4.5) states that a local tree with 
V2\[-SUBJ, FIN\] and \[BAit O\] as daughters must have a 
root specified <SLASII, NP>. As :in the case of STM\], 
the stipulation of a \[BAR O\] sister is the CCR 
counterpart of the requirement that metaru\]es apply 
only £o lexical IB rules. 
Taken together, the two FCRs of (39) and (44) plus 
the two CCRs of (40) and (45) accomplish all the work 
of STM1 and STM2 and result in the same analyses for 
English as adopted in GKPS. 
References

Chomsky, N. / M. Halle (1968): The Sound Pattern of 
English. llarper N Row, New York et al. 

Gazdar, G. / E. Klein / G. Pullum / I. Sag (1985): 
Generalized Phrase Structure Grammar. Blackwell, 
Oxford. 

Pollard, C. (i\[985): "Phrase Structure Grammar without 
Metarules, " Proceedings of the West Coast 
Conference on Formal Linguistics (Los Angeles) , 
Stanford Linguistics Association. 

Shieber, S.M. / tl. Uszkoreit / F.C.N. Pereira / J.J. 
Robinson / M. Tyson (1983): "The Formalism and 
Imp i enlentat ion of PATR-I I, " Research on 
Interactive Acquisition and Use of Knowledge. SRI, 
Menlo Park, California. 

Uszkoreit, H. (1984): Word Order and Constituent 
Structure in German. Ph.D. dissertation, 
University of Texas at Austin. 
