You are on page 1of 56

JEAN MARK GAWRON

SITUATIONS AND PREPOSITIONS

"He pretends to dispatch the PP in order to make it


return endlessly,.., and to conclude: it is always t h e r e . . . "
Derrida

INTRODUCTION

I n this paper I propose to explore some issues in the semantics of


prepositions within the framework of an extended situation semantics,
specifically that extended version developed in 'Types, Contents, and
Semantic Objects' (Gawron, 1987, henceforth TC&SO).
Most syntactic theories distinguish at least two kinds of prepositional
phrases, adjuncts or adverbial modifiers, and verbal complements or
subcategorized-for prepositional phrases. It is my aim in this paper to
show that this difference in syntactic function need not be reflected in the
lexical semantics of prepositions, and that a more perspicuous account of
the semantics of verbs and prepositions, as well as of valence alternation
facts, is possible if it is not. I will present an account on which adjunct
PP's correspond to external predicates on described situations (external
predication), while subcategorized-for PP's correspond to facts added
within described situations (internal predication). Yet in both cases the
prepositions themselves will correspond to two-place relations. Indeed,
there will be no reason why the same lexical content cannot function
both as a subcategorized-for complement and an adjunct. This means
that a characterization of prepositions as a semantic class is possible, and
that the same content can expIain the semantics of an obligatory occur-
rence such as that of for in This song is meant for children, and an
optional occurrence such as He did it for the children.
In a semantics based on the kind of type-sorted denotations advocated
in Montague (1974), this result would be difficult, although certainly not
impossible, to achieve. In Montague's system the operations that com-
bined semantic objects are most easily formulated if they conform to the
sorting of the type-theory. In the system sketched in TC&SO, there is no
such sorting, and the operations that compute contents are driven not by
a type-theory, but by syntactic labels. Thus, there is no pay-off in having
differences in ways of combining semantic objects be reflected in
differences in the semantic objects themselves. The primary intuition that
this analysis tries to capture is that various preposition functions can all in

Linguistics and Philosophy 9 (1986) 327-382.


(~ 1986 by D. Reidel Publishing Company
328 JEAN MARK GAWRON

principle be served by the same preposition content, and what may


change from function to function is not the contents themselves, but the
ways of combining those functions.
I will also argue that beyond the distinction between complement and
adjunct, there is a second critical distinction to be made among the
subcategorized-for PP's. Subcategorized-for PP's must be divided into
two semantically distinct sorts: one called argument PP's, in which the
preposition marks a genuine argument of the verb, and another, called
co-predicating PP's, in which the preposition does not mark an argument
of the verb, but introduces a new participant and new situational in-
formation. Again, however, I will conclude that the same preposition
content can serve both functions; the PP's will combine differently with
their heads, but no new preposition contents will be appealed to to
explain the semantic differences. This semantic parsimony will result not
only in a more compact lexical semantics; it will also help explain why a
single grammatical class serves all these syntactic functions, why there
are not, for example, disjoint classes of adjunct and subcategorized-for
prepositions. The end result is a generally sounder combinatorial seman-
tics, because the same forms will give the same contents across a greater
variety of translations.
I will thus distinguish at least three sorts of PP: adjuncts, arguments,
and co-predicators. A fourth type of PP - XCOMP's or controlled
complement PP's, will be recognized and discussed in section 4, but no
specific analysis will be proposed, Since no analysis of control will be
assumed here.
There are two reasons why situation semantics has proved an apt
vehicle for the analysis of prepositions, one theoretical, one technical.
First, an ontological commitment to situations is crucial in formulating
certain of my claims about how heads combine with their arguments. For
example, adjuncts cannot introduce external predications on situations
without situations being objects. Second, the variable-unification style of
semantic composition introduced in Barwise and Perry (1983) and
common to a number of situation semantics fragments (see, for example,
Barwise (1984a) and Cooper (1984a), as well as TC&SO) facilitates the
statement of some of the proposed semantics. For example, in the
account of subcategorized-for "argument" PP's introduced in section 3,
prepositions will be associated with two-place-relations which end up
sharing both their arguments with the head verb. Such wholesale
argument-sharing is certainly possible using lambda-abstraction and the
"curried" functions common to Montagovian fragments; it is just messier
to state.
SITUATIONS AND PREPOSITIONS 329

Neither of these considerations singles out situation semantics as a


unique solution. The distinction between internal and external predi-
cation can be made in a Davidson-style (Davidson, 1967), or Parsons-
style (Parsons, 1984) semantics, in which situations are quantified over in
an ordinary first-order logic. The advantage of variable-unification over
curried function-argument application is at best a technical point. At
worst, the fact that I find certain kinds of semantic combination "easier
to state" using variable-unification may show my lack of cleverness with
lambdas. However, the fact that within situation semantics I am able to
address both of these concerns simultaneously, while presenting what I
hope is a persuasive account of prepositions and verbal valence, may in
some small way argue for the usefulness of situation semantics in ad-
dressing semantic issues of interest to linguists. Beyond that there is
simply the virtue of executing the analysis within a framework which
embraces a broader conception of semantics than more traditional
theories centered on truth and reference. What situation semantics
promises is an account of information, both for sentences and for the
world in which sentence tokens are exchanged.
A few words now about methodology and the importance of lexical
semantics. The principle concern of combinatorial formal semantics is to
show how the meaning of the whole is built from the meaning of the
parts. But vows of compositionality amount to little if the meanings of
the parts are allowed to vary arbitrarily with the theoretical pressures
induced by the whole. Thus Frege may be rightly charged with a breach
of compositional faith when he tells us that sentences denote different
things in propositional attitude contexts; likewise, situation semantics can
be praised as a more strongly compositional theory because it promises
an innocent account of the attitudes.
Compositionality thus goes hand in hand with semantic parsimony.
Ultimately, the analysis of the meanings of the parts takes us down to the
lexical items. It is one of the central assumptions of this paper that any
successful compositional theory must be based on a parsimonious treat-
ment of lexical meaning.
After parsimony, the next most important standard I want to raise is
meaningfulness. Prepositions, even tiny ones, are meaningful. On the
account offered here, no preposition will be given a trivial content like
Ax[x], unless its distribution can be shown to be ungoverned by semantic
factors. Good candidates for such high-handed treatment would be
passive by and the of of nominalizations. But subcategorized-for prep-
ositions in general will have full-bodied lexical meanings, a crucial point
on which this account of prepositions will differ from others (for example,
330 JEAN MARK GAWRON

the account of Dowty (1978) or Gazdar (1983)). This course will have
three virtues:
(1) It will allow us to make a somewhat obvious hypothesis about a
very large class of subcategorized-for PP's, called Argument PP's, that
the class of verbs they subcategorize is a semantically motivated class.
This hypothesis, The Argument Principle for PP's, will be made explicit
in section 3. Take as an example the preposition for in its occurrence
with verbs like wish, hope, pray, ask, long, try, hunger and yearn, and
with nouns like desire, aspiration, search, quest, thirst, and hunger. Of
course there are numerous semantically related predicates that do not
take for. We have, for example, the verb desire, the verbs aspire and
hanker; there is a good deal of indeterminacy in the business of sub-
categorization. It admits of competing processes in the same domain,
frequent gaps, even a kind of blocking, or pre-emption, where one
subcategorization rules out other semantically plausible candidates; sub-
categorization shows many of the same general features of morphological
processes. But it is an important hypothesis of this paper that one
important way in which it differs from at least some morphological
processes is that its semantics is compositional.
(2) Assuming all prepositions have lexical meaning will allow us to
explain the semantics of other appearances of these prepositions, outside
the subcategorized-for class, compositionally. Thus, for example, the
preposition against occurs with a group of verbs having to do with
moving one object forcefully against another: hit, strike, rap, beat,
smash, and tap. Call these impingement verbs. I will argue that against
should be analyzed as marking an argument when it occurs with im-
pingement verbs and that it should thus be subcategorized-for: Against
also occurs with verbs like break, which are outside this class. Yet with
break, it still marks an object against which some other object (the
broken thing) is brought into forceful contact. I will argue that against is
a co-predicating preposition with break, that is, that it does not mark an
argument of the verb. The semantic rules I will propose, together with
the assumption that against has the same content in both occurrences,
will account for the fact that against introduces the notion of forceful
contact with break.
(3) A consequence of (1) is a semantic account of at least some verbal
valences; if we have a general semantic account linking valences with
verbs, then it is quite natural that more than one valence should be
linked with a given verb class. Thus, the road is open to an account of
semantically bounded valence alternations such as Dative-Movement
which makes no appeal to lexical rules. Note that an account of Dative-
Movement along these lines makes no sense unless the same verb content
SITUATIONS AND PREPOSITIONS 331

appears in both valences. When Dowty (1978) accounts for Dative-


Movement with a lexical rule that generates a new verb give with its
argument order permuted, he assumes not only that there are two
different verbal meanings involved (even though they differ trivially), but
also that the preposition to, which occurs with all the Dative-Movement-
verbs, is the identity function on NP-meanings. On this account of
Dative-Movement, it is an accident that the verb remove is not among
the Dative-Movement verbs. One could, of course, assign all the relevant
verbs a common semantic component by meaning postulate, but nothing
will help semantically distinguish dative-movement from any other ad-
vancement of an oblique to direct object position, and nothing will be
said about why a semantically related group of verbs all enter into a
valence alternation crucially involving the preposition to.
Point 3 above raises one of the important issues in the treatment of
preposition semantics, valence alternation. I believe that an elegant
treatment of valence alternation was one of the early motivations for an
analysis of subcategorized-for prepositions that assigned them trivial
meanings. If the to of dative-movement verbs is an identity function,
then it is easy to say that give has the same meaning in both valences. It
is thus important to show that the analysis given here does not preclude
an elegant account of valence change. A general discussion of valence
change is the topic of a separate paper (Gawron, 1985); but I will sketch
the outlines of such an account in section 3. In fact it is quite possible to
give an account that is parsimonious for verbs as well as prepositions,
and to allow the same verb content to appear in two different valences,
with two different prepositions.
A few remarks on the general organization of the paper. Section 1
presents a very preliminary analysis of the prototypical subcategorized-
for PP, the argument PP. In section 2, a distinct sort of subcategorized-
for PP is identified and discussed, and in section 3, an analysis is
presented, in the light of which the analysis of argument PP's is
modified. Section 3 also briefly discusses a semantic theory of valence
and the Argument Principle, and sketches the beginnings of a theory of
bondedness relations, relations like cause and result that are inferred
among the facts in a clause. Section 5 discusses adjunct PP's, and extends
the system presented in TC&SO to account for them.

1. ARGUMENT P P ' s
I assume subcategorized-for prepositions have genuine content. In par-
ticular, I will propose that their contents be situation-types defined by
two-place relations. I will also argue for distinguishing the two types of
332 JEAN MARK GAWRON

subcategorized-for PP's mentioned in the introduction, argument PP's


and co-predicators. As far as I know, this division among sub-
categorized-for PP's is without precedent in the semantics literature.
Syntactically, there does not seem to have been much call to differen-
tiate among the subcategorized-for PP's as PP's; although the distinction
between TO-OBJ and other obliques in LFG, and between 3's and other
oblique functions in Relational Grammar, does cut across PP's, the class
of PP's called TO-OBJ's and 3's is very small, and their status as
exemplars of a unique syntactic function has occasionally been called
into question. The proposed distinction between co-predicators and
argument PP's will differ in two ways: (1) both classes will be very large;
(2) the difference is essentially one of semantic function, not lexical
semantics; numerous prepositions can be the heads of both kinds of pp.1
I will thus proceed in small steps. First I will outline an account, still well
within the working assumptions of this paper, on which the composition
rule for subcategorized-for PP's covers only argument-PP's. Then to
handle some new phenomena we will move on to a fancied-up version.

1.1. First Approximation

We continue to assume that all prepositions have non-trivial meanings.


The general framework of the semantics we have been pursueing then
suggests that they be situation-types defined by lexically specific rela-
tions, just as verbs, nouns, and proper names are. The simplest assump-
tion is that they are simply one-place relations. We begin there: our first
tentative lexical entry for against (to be modified later) will be:
[against: {(LOC $LOC
REL against
P-OBJ $INDO
POL $POL)}].

We have introduced a new grammatical relation, P-OBJ, here, for the


objects of prepositions. For the moment, nothing very vital hinges on its
status as a unique grammatical function. We assume that the objects of
all subcategorized-for PP's will be marked with that function in the
lexicon.
Before writing any rules, we turn to an example. Consider:
(1) Jack hit the stick against the fence.
Note that the PP is obligatory in (1) if direct object is to be interpreted as
SITUATIONS AND PREPOSITIONS 333

"instrument". I hit the stick has only the wrong interpretation. I therefore
assume that hit in the valence with against is a three-place relation with
a lexical entry like this:

[hit: V: {(LOC $LOC0


REL hit
SUBJ $IND0
OBL $IND 1
OBJ $IND2
POL $POLO)}].

We will see below that three-place hit will serve all the needs of the
standard examples. No two-place hit will be necessary. Given the lexical
entries for hit and against, the following additions to the rules of the
fragment in TC&SO will handle the subcategorization dependencies of
hit.

(PP ~ P NP)
[[P NP]] = QI([[NP]] P-OBJ [[P]])
(VP--* V NP PP)
[[V NP PP]] = QI([[NP]] OBJ QI([[PP]] OBL [[V]])).

Note that the PP content is incorporated into the described situation-type


using "QI", the same operation that incorporates NP's. This is possible
because the labeled indeterminate that assigns a subcategorized-for PP
its content is no different in structure from the labeled indeterminate of
an NP; it simply has an extra relation. The labeled indeterminate for
against the fence is:

{(REL the-fence
ARG $IND0)
(LOC $LOC0
REL Against
P-OBJ $IND0
POL $POL)}.

The labeled indeterminate for the fence is the same minus the against
fact. Since both exhibit the function " A R G " , both can be combined in
using the operation "QI". Using the VP rule above, the labeled in-
334 JEAN MARK GAWRON

determinate for Jack hit a stick against the fence is:

(1T) {(LOC $LOC0 (REL the-fence


REL hit ARG $IND1)
SUBJ $IND0 (LOC $LOC1
OBL $IND 1 REL Against
OBJ $IND2) P-OBJ $IND 1)}
(REL Jack
ARG $IND0)
(REL the-stick
ARG $IND2)

An immediate concern raised by (1T) is the semantics of the preposition


relation, against. What does it mean to have the property of against-
ness?
The preposition against occurs, among other places, with a certain
very small class of semantically related verbs: hit, slap, beat, rap, tap,
smash, and strike. Two facts argue for relating these occurrences of
against to the occurrence with hit: first, the clear semantic relatedness of
all the verbs; second, the fact that all of them participate in the same
valence alternation as hit, an alternation discussed in Fillmore (1977a).
Compare (1) with (2), and note that all the verbs in this small class can fit
in both these valences.

(2) Jack hit the fence with the stick.

Call this small group of verbs impingement verbs, and situations involving
impingement verb predications impingement situations. In proposing we
talk this way, I am in fact proposing that the situations associated with
these various verbs exhibit a regularity of contour: an agent A brings one
object B forcefully against another C. I propose that we chose the
semantics of the prepositions to follow this regularity among the verbs:
note that against always marks Object C. with always marks Object B.
Note also that we also have intransitive verbs like knock that use against
(Her hand banged against the table). So we might say that the particular
regularity of contour that concerns the prepositions only involves objects
B and C: one object coming forcefully against another. I propose we
canonicalize that regularity with a two-place relation IMPINGEMENT,
and that we relate IMPINGEMENT to both the prepositions and the
verbs. The following constraints say more or less what I have in mind.
I M P I N G E M E N T Constraints (version 1)
(Ia) (1, a, x, y, z, 1)--~ (l, IMPINGEMENT z, y, 1)
(Ib) (l,/3, y, z, 1 ) ~ (l, IMPINGEMENT y, z, 1)
SITUATIONS AND PREPOSITIONS 335

(IIa) (l, IMPINGEMENT y, z, ! ) ~ (Against y, 1)


(IIb) (l, IMPINGEMENT y, z, 1 ) ~ (With y, 1)

where a ~ {hit3, slap3, knock3, strike3, rap3, tap3, smash3, beat3}


and/3 c {hit2, slap2, knock2, strike2, rap2, tap2, smash2, beat2}.
These constraints are modeled on a format used in Cooper (1984a).
Cooper's interpretation of the arrow involves an appeal to actual situa-
tions: for all actual situations, if the situation contains the fact on the left,
then it also contains the fact on the right. This implies a view on which
"actual" situations are complete: they contain all the facts about some
chunk of reality. Given that some facts are unlocated it may be non-
trivial to precisely identify such chunks. It may also be non-trivial to
specify what completeness means, if a situation containing facts true at
some location can itself be a constituent of facts at that location. Because
of these misgivings about the appeal to actual situations, I want to
interpret the above constraints more along the lines on which constraints
are interpreted in Barwise and Perry (1983): if some factual situation
contains the fact on the left, then some larger factual situation contains
the fact on the right.
Turning now to the substantive claims made by (I) and (II): (I)
essentially identifies IMPINGEMENT as a "less finely grained
regularity" than any of the relations in a or/3. (II) says that if there is a
factual IMPINGEMENT fact then there are factual Against and With
facts. This allows for both with and against to mark less finely grained
regularities than IMPINGEMENT. Thus, this with, for example, might
be identified with the general instrumental with.
The existence of lexical two-place IMPINGEMENT relations (those
on the /3 list) is shown by examples like her hand banged against the
table, and the stick hit the fence. We will have something to say about
"causative" relationship between the o~ and/3 verbs, in particular for hit,
in 3.3.2. For now, the two-place verbs serve to establish that the IM-
PINGEMENT relation which semantically links all the impingement
verbs must be a two-place relation.
To see these constraints do some work, let us consider the semantic
composition for valence of hit used in (2). The relevant lexical entries
will be:
[hit: V: {(LOC $LOC0
REL hit
SUBJ $IND0
OBJ $IND1
OBL $IND2
POL $POLO)}]
336 JEAN MARK GAWRON

[with: P: {(LOC $LOC0


REL with
P-OBJ $IND3)}].
The new entry for hit is exactly like the old, except that the functions
OBL and OBJ have exchanged argument-positions. (Presumably there is
also some other syntactic difference between the entries to indicate the
difference in preposition complement, and its optionality). This means
that the composition rules given above will substitute the direct object
NP and the oblique PP into different slots in the hit relation. So that the
content of (2) is given by:

(2T) {(LOC $LOC0 (REL the-fence


REL hit ARG $IND1)
SUBJ $IND0 (LOC $LOC1
OBJ $IND 1 REL With
OBL $IND2) P-OBJ $IND2)}
(REL Jack
ARG $IND0)
(REL the-stick
ARG $IND2)

Because of the with and against facts, the indeterminates associated with
(1T) and (2T) are not alphabetic variants. Note, however, that the two
lexical entries for hit are. Thus, while (1) and (2) do not have the same
contents, the verbs used in them do. This is not an unreasonable result;
Fillmore points out that in certain cases sentences like (1) and (2) may
have very different functions; if, for example, the valence used in (1) is
used with an animate oblique ("Jack hit the stick against Harry"), the
effect is somewhat peculiar, as if Harry's animacy was being willfully
de-emphasized. Constraints (I) and (II), however, guarantee that the
contents of (1) and (2) are not simply unrelated situation-types. If a
situation realizes the type of (2T), then there must be a larger factual
situation containing an against fact. Schematically, this is because:
Chit)--+ (IMPINGEMENT) ~ (against).
So (2) has (1) as what Barwise and Perry call a "weak consequence". The
argument from (1) to (2) is symmetric:
(hit) --~ (IMPINGEMENT) --~ (with).
Clearly, if defining the "consequence" relation between (2) and (1) were
the only issue, we could have dispensed with the relation IMPINGE-
SITUATIONS AND PREPOSITIONS 337

MENT, and simply collapsed (1) and (11). However, it will be convenient
to speak of IMPINGEMENT situations in general, and in the revised
analysis offered in the next section, the abstractness of IMPINGEMENT
will vanish.
Several loose ends remain. In situation semantics, some troubling
questions about tense and polarity arise immediately when we factor what
is ordinarily thought of as a single main verb predication into distinct
facts. Thus, when we assign past tense to (IT), we will want both the verb
and preposition location to be past. A little worse: when we assign
negative polarity, what we will want in the general case is that at least
one polarity be 0 (this will be more obviously true in the case of the
co-predicators introduced next section). I do not propose to treat these
problems here. If, however, I am right about polarity assignment to verbs
and subcategorized-for prepositions, then polarity assignment cannot
involve simple substitution the way argument-filling operations do. In-
stead, a careful distinction must be drawn between the polarities of
true-predicators like those of the verb and prepositions facts, and those
of other facts (say, those of a singular NP fact). Then polarity assignment
will require quantification over predicator polarities. VP negation would
introduce facts like:
(not-all {POLO, $POL1}, TYPE($POL3, (equal $POL3, 1))).
Here, the second argument of not-all is the type of polarities that equal
"1", the first argument a set of predicator polarities.
Another loose end that may now be tied up is the fate of two-place hit.
Or, what amounts to the same thing: what do we do about the fact that
the with PP is optional with hit? Here we may take advantage of a
peculiar property of types. Relations in lexical types have a set of slots
associated with particular grammatical relations, but they do not have
reality; they are not functions with a particular number of arguments. If a
particular argument is not supplied, the result is still a well-formed type;
the slot corresponding to the missing argument simply has no constraints
placed on it, and the result is like unrestricted existential quantification.
Thus sentences like Rhonda hit Pablo will receive a sensible inter-
pretation. Any situation which is of the situation type determined by the
interpretation will, however, have to instantiate some implement with
which Pablo was struck (even if it is only Rhonda's first).
I am assuming, of course, that the syntax is responsible for matters of
optionality. This is not an entirely satisfactory solution. For one thing, all
the impingement verbs have a common property which I have left it to
the syntax to account for: in the valence that takes against the preposi-
338 JEAN MARK GAWRON

tional complement is obligatory. More troubling perhaps, sentences like


Rhonda hit Pablo have a conventional interpretation for what the missing
instrument is. This tendency to a conventional interpretation may even
be stronger with some other verbs. Compare Jack rubbed the cat's back
with a stick, with Jack rubbed the cat's back. The first sentence does not
have the second as an entailment, because the second entails that Jack
touched the cat. I will suggest a possible treatment for such "understood
arguments" in section 4.
Yet another loose end is what to do with subcategorized-for PP's when
there seems to be more than one of them. The answer is that more rules
and more grammatical functions may be required. Consider hammer, a
verb that seems, as far as meaning goes, to be a perfect impingement
verb. Yet it shows a valence the other impingement verbs lack: Jack
hammered against the door with the stick. Here we have the same
prepositions we had all along, and marking the same participants, but
co-occurring. Semantically, this verb gives strong support for our hypo-
theses about the meanings of both with and against. Combinatorially,
however, it is the occasion for a new composition rule:

Argument Rule with Two Obliques


(VP---~ V PP1 PP2: QI([[PP2]] OBL1 QI([[PP1]] OBL2 [[V]]))).

The existence of such examples suggests one of two possibilities: either


the notion of a single grammatical function OBL is inadequate, and we
should have several or many distinct oblique functions as in Bresnan
(1983), or one of the two PP's with hammer should be analyzed as the
direct object. I will not try to choose between these alternatives here.

2. THE NEW PHENOMENON

Now the claims made by (I) and (II) have been made about relations,
which is to say, on objects in our ontology, independently of any facts
about verb valence. One might then ask whether they are uniformly
reflected by the facts about verb valence. The ideal state of affairs would
be one in which all and only the impingement verbs could take both
with and against to mark particular "roles" in an impingement situation.
That this is not so is shown by some examples discussed in Fillmore
(1977), which have received little attention elsewhere:

(3) John broke the vase with the hammer.


(4) John broke the hammer against the vase.

While these two pairs of sentences bear a surface resemblance typical


SITUATIONS AND PREPOSITIONS 339

impingement pairs, they aren't paraphrases. In (3) it's the vase that
breaks; in (4), however unlikely, it's the hammer. The key point is that
direct object always breaks. Thus break uses with and against in (ap-
parently) just the meanings we have been considering, but with no
explicit valence change. Clearly, if we associate the verb in (4) with a
three-place relation, it can't be the same relation as that associated with
the verb in (3). The alternatives are either (1), to use two different
three-place relations in the semantics of (3) and (4), methodologically the
last resort, or (2), to deny that the participants marked by prepositions in
(3) and (4) are arguments of the verb; the question then is, what are"
they?
We can move closer to an answer by framing the problem somewhat
differently. First, note that (4) involves the same kind of moving into
forceful contact that (1) and (2) do; specifically, it involves movement of
the direct-object, the thing broken, and movement of the broken thing is
not a logistical necessity of breaking (as 3 shows).
How, then, do we introduce the semantics of movement and IM-
PINGEMENT into (4)? One possibility is to use what we've got. There is
no semantic reason why a verb that syntactically selects a direct object
and oblique complement needs to be associated with a three-place
relation. Although the rule that combines the PP with the verb makes
reference to the function OBL, we will show in section 3 that no
incoherent semantics results if we simply allow vacuous substitutions.
Thus, if break the relation were binary, and break the verb was allowed
to optionally subcategorize for against, we would get the following
semantics for (4):

(4T) {($LOC0, break, $A, $B)(Jack, $A)


(the-vase, $C)(the-hammer, $B)
($LOC 1, Against, $C)}.

For readability, I have taken the liberty here of giving the unlabeled
version of the labeled indeterminate, and using lettered, rather than
indexed, irgteterminates: I will follow this practice below whenever we
are not specifically discussing matters of grammatical function. Although
this semantics merits praise for its simplicity, it is still somewhat beside
the point. We are trying to introduce the semantics of impingement, but
nothing we have yet said relates the situation type in (4T) to an
IMPINGEMENT situation. If, however, we amend (IIa) to be a bicon-
ditional, we have:
(IIa') (3z)(l, IMPINGEMENT y, z) ~--~(Against z).
340 JEAN MARK GAWRON

Now the existential quantification here is problematic, a peculiar addition


to our constraint language, but it merely makes explicit some of the
implicit existential entailments in B&P (1983); for a discussion of this
point, see B&P (1984, p. 49), which suggests an "unpacking" for extra
indeterminates on the right hand side of a constraint arrow. If we
adopted (IIa'), any factual instantiation of the type of (4T) would have to
be involved with an impingement situation; thus, the vase in (4) would
correctly be interpreted as an "impingee".
Now this account has the virtue of being parsimonius of translation
rules, but it still falls far short of what we need. (4T) along with (IIb')
correctly implicates IMPINGEMENT in the semantics of (4), and cor-
rectly identifies the grammatical object of the preposition as the physical
object impinged against, but it says nothing about the other crucial role
in impingements. Nothing in (4T) constrains "the hammer" to be the
"impinger". It is not sufficient to say that impingement against 2 intro-
duces impingement; rather, it introduces impingement together with one
new participant, the "impingee", taking for its "impinger" one of the
core participants of the verb. Consider another verb discussed in Fill-
more (1977), c u t . Like break, cut does not normally imply movement of
the direct object participant, yet with the appearance of against, the
"cuttee", while still cut, also moves: John cut his foot against the rock.
This parallels the superimposition of movement against gives with break.
Other transitive verbs showing a superimposition of movement or im-
pingement: sharpen, rattle, push, flatten, crack, and scrape. Moreover, in
all these cases the impingee is the direct object. With intransitive
versions of some of the same verbs, the impingee is the subject. In the
case of hammer, we saw against occurring with an oblique impinger, but
hammer is an impingement verb, a verb that always involves forceful
movement. With non-impingement verbs like break, the impingee is
always either the direct object (transitives), or the subject (intransitives).
These are clause level dependencies that should be captured with clause
level translation rules. They cannot be gotten at via constraints.
For the same reason, that the introduction of impingement involves
two grammatically definable participants, I rule out an analysis of (3)
and (4) in which the PP's are adjuncts, in the sense of adjunct discussed
in section 5. Adjuncts involve external predications on described situa-
tions, and have no access to the participants in those situations. The
possibility that (3) and (4) involve ordinary complement control will be
considered in section 4, along with the possibility that they are what
Bresnan (1983) calls XADJ's.
SITUATIONS AND PREPOSITIONS 341

3. CO-PREDICATION

The semantic regularity uniting (1) and (4) is that both sentences des-
cribe situations in which one object comes into forceful contact with
another. I propose to hang this regularity squarely on the preposition by
making its lexical relation be the two-place IMPINGEMENT relation
Note that physical movement does not seem to be necessary with
instrumental with: The soprano broke the glass with a High C. To
account for this, I propose we associate with with two-place relation.
INCIDENCE, which is vaguer than the IMPINGEMENT relation. The
content of the INCIDENCE relation is that one of its argument acts
directly upon the other. IMPINGEMENT would be one such kind of
action. This analysis allows us to identify with we have been talking
about with general instrumental with. In terms of constraints we have:
I M P I N G E M E N T Constraints (version 2):
(Ia) (or, x, y, z)--~ (IMPINGEMENT z, y)
(Ib) (fix, y) ~ (IMPINGEMENT x, y)
(II) (IMPINGEMENT y, z)---) (INCIDENCE, z, y).
These constraints replace I and II of section 1. Thus, to put it back in
terms of lexical items, (against y, z) will imply (with z, y) but not
conversely. Yet (2) and (1) will still imply each other, because the verb
hit will still imply impingement.
The lexical entries and translation rule needed to handle (3) and (4)
will be:
[against: P: {(LOC $LOC0
REL IMPINGEMENT
P-SUBJ $IND0
P-OBJ $IND1
POL $POL0)}]
[with: P: {(LOC $LOC0
REL INCIDENCE
P-SUBJ $IND0
P-OBJ $IND 1
POL $POL0)}]
[break: V: {(LOC $LOC0
REL break
SUBJ $IND0
POL $POL0)}]
342 JEAN MARK GAWRON

Co-Predication Rules:
(VP ~ V NP PP)
[[V NP PP]] = QI([[NP]] OBJ M([[PP]] P-SUBJ [[V]] OBJ))
(VP--* V PP)
[[V PP]] = M([[PP]] P-SUBJ [[V]] SUB J).
We discuss the preposition entries first. Since the prepositions are now
two-place relations, some function had to label the new argument posi-
tion. I have called it P-SUBJ; P-SUBJ is distinct from both of the main
verb functions SUBJ and OBJ, though, in general it will share its value
with one of those functions.
The composition rules above introduce a new composition operation,
M (for "merge"). M is a slight generalization of the operation QI. Using
B (or "bind"), QI(QUANT FUN HEAD) unifies the value of FUN of
H E A D with the value of A R G of QUANT, and then takes the union of
Q U A N T and HEAD. M does the same thing, except that instead of
being limited to A R G function of QUANT, any label can be specified. M
thus takes four arguments, two labeled indeterminates and two functions.
Its full definition in terms of previously defined operations:
M ( A R G FbrN1 H E A D FUN2) =
U ( A R G B(ARG:FUN1 FUN2 HEAD))
In the two composition rules above P-SUBJ is merged with OBJ and
SUBJ respectively. When the full NP argument then comes along (as it
does in the VP rule), it winds up "controlling" both the P-SUBJ slot and
its conventional argument slot. The content for (4) is given by:

(4T) {(LOC $LOC0 (REL the-vase


REL break ARG $IND2)
SUBJ $IND0 (LOC $LOC1
OBJ $IND1) REL IMPINGEMENT
(REL Jack P-SUBJ $IND 1
ARG $IND0) P-OBJ $IND2)
(REL the-hammer
ARG $IND1)
Or equivalently:
(4T) {($LOC0, break, $A, $B)(Jack, $A)
(the-vase, $C)(the-hammer, $B)
($LOC1, IMPINGEMENT $B, $C)}.
Call PP's that have combined with their head verbs this way co-predi-
SITUATIONS AND PREPOSITIONS 343

cated PP's. In (4T), the co-predicated PP can be thought of as controlled,


because it shares an argument with the main verb. However, the control
involved is not ordinary complement control because the semantic con-
stituent formed by combining the controller and controllee is not an
argument of the verb.
But what about the rule we gave in section 1 to treat prepositions that
marked an argument of the verb? That rule will now need to be
modified:
Argument Rule
(VP ~ V NP PP)
[[V NP PP]] =
QI([[NP]] QI([[PP]] M([[PP]] P-SUBJ [IV]] OBJ))).

The above composition rule actually combines the PP indeterminate in


twice, once with M, and once with QI, but since it does so using set
union, no harm is done; the same PP-indeterminate is unioned in both
times, since substitution operations are only performed on functions
belonging to the head. This new composition rule will render (1) as:

(1T) {(hit, $A, $B, $C)(Jack, $A)


(the-stick, $C)(the-fence, SB)
(IMPINGEMENT $C, $B)}.

Call PP's that have combined with their head verbs this way argument
PP's. The unusual feature of (1T) is that the preposition shares all its
arguments with the main verb. It is more than controlled; it is completely
merged.
We have now introduced a major split among subcategorized PP's, a
split defined by whether the head has an argument position that merges
with the P-OBJ. At the moment is is also defined by two separate
composition operations, but do we really need two? Consider for a
moment the fully unpacked definition of QI:
QI(QUANT FL~q HEAD) =
U(QUANT S(QUANT:ARG HEAD:FUN HEAD))
We haven't yet said anything about what our retrieval operation ":" does
when H E A D doesn't have the function FUN. Suppose we say that it
returns some distinguished value ("nil" or A); anything that cannot be a
constituent of an indeterminate will do; then suppose that S simply
returns H E A D unchanged when its second argument does not occur in
H E A D (this is exactly what the LISP function SUBST does). Then, in
the case where H E A D does not have the function FUN, QI simply
344 JEAN MARK GAWRON

reduces to U(QUANT HEAD). In that case, the composition operation


in the Argument Rule can be used for both argument-PP's and co-
predicators, only in the latter case, the function OBL does not occur in
the verb indeterminate, and the application of QI involving the PP is
vacuous.

3.1. The Argument Principle

Let us turn now to a comparison of (IT) and (4T), and consider what
consequences the differences between them have.
Structurally, the difference between the two contents is that hit, a
three-place relation, shares two arguments with against, where break, a
two-place relation, shares only one. I want to claim that this difference in
translations correlates with an important semantic difference.
In (1T), the IMPINGEMENT relation is in fact redundant, because of
Constraint (I). In (4T), it carries new information, specifying who's
moving towards what. My intention is that the informativeness of the
preposition be the chief semantic diagnostic for distinguishing argument
PP's from co-predicating PP's: the information introdued by the meaning
of an argument-marking preposition is always redundant.
This is worth stating as a principle. To facilitate that statement let us
say that when two relations are related in a constraint the way hit and
IMPINGEMENT are related in (I) in section 3, the n-ary relation R1 on
the right hand side of the arrow is a component of the m-ary relation R2
on the right (n ~ m);

The Argument Principle for Prepositions:


A Subcategorized-for prepositional phrase can only be an
argument-PP if the lexical relation of its head preposition is a
component of the verb's lexical relation.
The descriptive consequences of the Argument Principle can be illus-
trated with the aid of an example cited in the introduction. The prep-
osition for occurs with a class of verbs having to do with desire: wish,
hope, pray, ask, long, try, hunger, and yearn. If we posit a relation
DESIRE which is a component for all these verbs, and use DESIRE as
the lexical relation for one meaning of the preposition for, then for will be
eligible to mark arguments with any of them. The Argument Principle
does not require that it do so; it merely licenses the subcategorization as
a possible valence for the grammar. Note that the verb desire does not
allow for, though it presumably contains the appropriate semantic com-
ponent; note also that the noun reverses this valence decision.
SITUATIONS AND PREPOSITIONS 345

It may seem that we have come full circle. We started out insisting that
prepositions were meaningful, and now we have concluded that some-
times, at least, their meaning is redundant. Small gain, it seems. But the
moral of the story is that anyone who substitutes "meaningless" for
" r e d u n d a n t " in their analysis of against in (1) then has no account of
how the notion of impingement is introduced into (4). (4), on the current
account, is not a problem for our analysis of (1) and (2); it is the strongest
evidence for it.
T h e proposal on the table is to draw a major bifurcation in sub-
categorized for PP complements. Some - the " a r g u m e n t " PP's - are
treated just like NP's; they are quantified in; they are counted among the
verb's lexical "arguments"; they are officially assigned the grammatical
function O B L I Q U E . T h e situational information they introduce is in
effect redundant. Other PP's are not in any semantic sense inherent to
the verb. T h e y add situational information to that of the verb, at the
same time sharing one participant with it. T h e y are, in effect, predicators
all on their own.
Yet, given that the formal status of the division is clear, what is the
empirical status? We saw that analyzing the occurrences of with and
against with break as argument prepositions would multiply the number
of lexical entries for break, because the role the direct object played
depended on which preposition was present. But why not analyze all
prepositional complements as co-predicators, and leave it to lexico-
graphers, long since accustomed to such drudgery, to decide which
prepositions are " c o m p o n e n t s " of their verbs?
One reason is that the Preposition A r g u m e n t Principle makes a strong
- and to a large extent correct - prediction about subcategorization. It
says all subcategorization for argument PP's is semantically motivated. It
thus claims that it is no accident that we find semantically related verbs
like the impingement verbs subcategorizing for the same prepositions -
and that it is no accident that when we find an impingement verb with an
intransitive valence like hammer, the two prepositions chosen are against
and with. T o say there are also co-predicating PP's is to say there are
prepositional occurrences that are not semantically motivated, or at least
not motivated in the same way.
A second motivation for the distinction becomes evident when we turn
to valence change. On the current analysis, valence change can be dealt
with merely as a different assignment of grammatical function to the
argument slots of a lexical relation. An account that never assigned
prepositions to a verbal argument slot would need additional semantic
machinery to map between different relations. Moreover, since there is
346 JEAN MARK GAWRON

an account of which prepositions can mark which arguments, there is no


reason to posit a lexical rule for valence changes like those exhibited by
the I M P I N G E M E N T verbs. One merely takes advantage of something
already provided for by the syntax: an argument that is neither subject
nor object is available to be marked by a preposition. The Argument
Principle makes a claim about which prepositions are appropriate. 3
This non-rule account of valence change still leaves open the question
of which argument gets to be direct-object, but if there are regularities to
be captured about those phenomena, they are not ones that can be
captured by talking about I M P I N G E M E N T verbs alone. The tack taken
in Gawron (1983) in proposing a full-fledged theory of valence was to
assume that there exist semantic constraints on the choice of nuclear
terms for any verb, but that those constraints are incomplete, leaving
many verbs with "valence options". Among the "constrained" verbs are
those involved in change of state (like break), or necessary movement
(like thrust): Hit, falling into neither class, has options. Other accounts
are imaginable; the only crucial point here is that if anything can be said
about the choice of a particular argument for direct object, it can be said
without direct reference to the particular semantic class of impingement
verbs.
Another motivation for argument PP's, though a bit weaker, comes
from the phemomenon of control. Most theories of control have as a
theorem some such principle as "controllers must be arguments", (usually
derived from something stronger). Since a verb like appeal requires the
controller to be inside a PP, within those theories, at least some PP's must
be argument PP's.
Finally, there are semantic arguments that participants marked by
prepositions are often core participants of a verb. For example, the claim
made by this analysis is that hitting situations are inherently impingement
situations. That is, we can always individuate the two crucial participants
of an I M P I N G E M E N T scene in any hitting scene, even if a linguistic
description does not explicitly encode all three participants. This can be
resolved a talk of entailment relations, if one's tastes run that way. So, for
example, Jack hit the vase entails that there exists some z such that z
moved towards the vase. In other words, a hitting always requires an
instrument (which is one way of describing one of the arguments in an
I M P I N G E M E N T scene). On the other hand, breaking situations are not
inherently impingement scenes. Jack can break the vase by dropping it
on the floor, if he likes, but then the floor doesn't move towards the vase,
the floor certainly isn't an instrument. We don't say, "Jack broke the vase
with the floor."
SITUATIONS AND PREPOSITIONS 347

3.2. Other Co-predications


In this section I want to explore some other cases where a co-predication
analysis can be illuminating.
Consider (5), on its most natural reading:
(5) Jack hit the ball against the fence.
The ball gets hit, travels some distance, then strikes the fence. In (1) the
stick moves but doesn't get hit; in (5), the ball moves, but does get hit. In
both cases the movement of the direct object can be explained by
claiming that it enters into an IMPINGEMENT predication. But that
claim doesn't help us make the choice between an argument and co-
predicational analysis of (5). If (1) and (2) exemplify the only available
valences for hit, then the fact that the ball gets hit in (5) that means that
it realizes the valence in (2), with the optional instrument missing; in that
case, the P-OBJ of against is a new participant, who could be accounted
for with a co-predication analysis. Here's the content that results if we
compose the semantics of the above sentence using the co-predication
rule:

(5T) {($LOC1 hit, $A, $B, $IND3)(Jack, $A)(the-ball $B)


(the-fence $C)($LOC2 IMPINGEMENT $B $C)}.

The "Instrument" ($IND3) of the verbal hitting situation has never been
combined in and will be arbitrarily instantiated in instantiating situations.
Meanwhile the ball, like the hammer in (4), has been drafted into a
distinct impingement fact in which it comes in contact with the fence
(distinct because it has a distinct location and distinct parcipants). These
are the right logistics. (1) and (5) have been assigned distinct semantics
that captures an important difference between them - yet without multi-
plying the number of verb or preposition meanings.
Since co-predicators introduce new participants, non-arguments,
nothing prevents them from iterating indefinitely many times. This sug-
gests a very straightforward treatment of so-called "path" PP's, prob-
lematic for most argument-slot filling analyses of PP's. Alongside (5) we
can have examples like Joan hit the ball through the alley between the
buildings into Mrs. Magillacuddy's window. The two PP's, through the
alley between the buildings and into Mrs. Magillacuddy" s window, can be
successively combined with the direct object and unified with the clause
content. The resulting content will simply have two preposition facts in
place of the singleton fact of (5T). Of course, if such multiple instances
of co-predication are all to be sisters of V as we have them in our syntax,
348 JEAN MARK GAWRON

then we would have to allow a Kleene-Star in our co-predication rule:


correspondingly, we would have to modify the composition operations to
be iterative. The technical details of those changes are probably best
deferred until the issue of subcategorization of co-predication, discussed
in section 4.3, is settled.
A somewhat different extension of the co-predication analysis is sug-
gested by examples like (6):
(6) Mary walked under the bridge.
(6) exhibits a well-known ambiguity: on one reading, call it the locative
reading, Mary's aimless strolling takes place under the bridge; on an-
other, call it the goal reading, Mary walks to a point under the bridge.
Cases like (6) have sometimes been pointed to to illustrate a systematic
ambiguity among English prepositions between a locative and a goal
sense; indeed this "ambiguity" exists in other languages as well, although
sometimes it exhibits a formal reflex; in Polish and Russian, for example,
in the locative sense a preposition governs the locative case, in the goal
sense it governs the accusative. What I would like to propose is that, in
English, at least, the ambiguity of (6) not be treated as a lexical
ambiguity at all, but, again, parallel to the difference between (1) and (5),
as an ambiguity that reflects two different ways of combining the same
contents. Suppose we treat the "goal" PP of (6) as a co-predicator; then
we have:
(6T) {($LOC1, walk, $A)(Mary, $A)
($LOC2, under, SA, SB)(the-bridge, SB)}.

Here, the under relation holds between Mary and the bridge. On the
locative reading, on the other hand, the same under relation would hold
between the situation of Mary's walking and the bridge. Details for such
an adjunct semantics are given in section 5.
There is a problem with (6T), of course, as it stands. What it says is
that Mary did some walking at one location, and that Mary is under the
bridge at another. (6T) really says nothing about how Mary got under the
bridge. (6) specifically asserts that it was by walking that Mary got there.
I will defend this underspecified semantics in the next section. For the
remainder of this section I want to discuss some features of this analysis
as it stands.
(1) and (5) can be seen as a minimal pair that is evidence for the
distinction between argument PP's and co-predicators. The ambiguity in
(6) can be seen as similar evidence for the distinction between co-
predicators and adjuncts. At this point it might be useful to examine this
SITUATIONS AND PREPOSITIONS 349

proposed three-way distinction to see how it accords with our pre-


theoretical intuitions.
When I first began thinking about prepositions I thought the first order
of business was to find a semantically principled way to draw the line
between arguments and adjuncts, particularly for optional arguments.
The difficulties of this distinction are notorious, and it is no surprise that I
very quickly gave up the search. At the time, I concluded that what we
were dealing with was a natural cline, that the only difference between a
location preposition like in and the for that came with hope, was that
the locations marked by in were circumstantially compatible with a much
broader class of verbs than the "desired objects" marked by for. There
were no arguments, only participant roles (of which were there lots) and
the classes of situation they were compatible with. The tendency for
certain participants, but not others, to become nuclear terms could be
explained in terms of general semantic preferences for certain sorts of
participants (like agents) to be nuclear terms - preferences that needed to
be accounted for in anybody's theory whether they believed in arguments
or not.
I now believe that this skepticism, while healthy, is wrong, and it is
critical to the analysis put forth here that it be wrong. Anyone who be-
lieves that instrumentals, say, are just participants appropriate for some
verbs but not for others has no way of talking about the different status
of the instrument with hit and break. The view adopted here is that there
is a fundamental change in the combinatorial status of the preposition
with when we move from hit to break. The difference is reflected in the
structure of the semantics; it is not a "fuzzy" distinction (which is not to
say that we theorists will always have an easy time telling which is which).
Furthermore, the way in which an adjunct combines with its head will
differ from the ways both arguments and co-predicators do, and the
resulting content will be quite different in its predicational structure.
Thus the distinctions being made are structurally based; there is no
prediction that the things we call adjuncts ought to combine with a greater
variety of verbs than the things we call arguments.
In fact, I think such "predictions" are wrong. Adjuncts with a very
limited distribution are easy enough to find, once we look for them.
Consider an example of Keenan's: The spy shot the general from the roof.
1 want to suggest that this example is ambiguous in just the way (6) is,
between a co-predicational and an adjunct reading. On the reading on
which the general is propelled from the roof, call it the resultative
reading, the semantics is co-predicational, with the causal connection
between the shooting and the departure from the roof supplied by the
350 JEAN MARK GAWRON

bondedness relations discussed in the next section. On the other reading,


call it the event-site reading, from the roof is an adjunct one of whose
arguments is the entire shooting situation.
This analysis requires a little defending. I think one's initial reluctance
to accept it follows from a general erroneous assumption that adjuncts
ought to be "freely occurring". In fact I think adjuncts are just what the
semantics given in section 5 claims they are: contents with lexical
relations that take the entire described situation of a clause as their
argument. Sometimes that lexical relation will be selectionally un-
demanding, as the location relation is. Sometimes they will be quite
choosy, as I claim from is. What I am suggesting is that there is a sense of
from (perhaps the only sense) which, when it takes a situation as its
"theme" argument, allows only situations with an inherent directionality.
Compare Keenan's example with something like He called to me from the
next room, or She sang the aria from far downstage. The bullet and the
voices play parallel "roles" in these sentences, though both are gram-
matically unexpressed; they are the physical quantities that give the
situation a direction. We can paraphrase Keenan's example: the shooting
situation involving the spy as shooter and the general as shootee extends
outwards from the roof. Compare John Anderson's (Anderson, 1970)
example of a non-motion use of from: The fog extended from London to
Paris. Here, directionality is imposed simply by the location of the
speaker's camera.
Note that in the framework we've adopted an argument analysis of the
event-site reading of Keenan's example just isn't very plausible. Since
shoot is transitive, the from in question would have to be a relation
holding between the general and the roof, quite plausible for the resul-
tative reading, but a bizarre claim for the event-site reading.
What I hope to have shown in this short section is that a three-way
distinction, if it can be defended, fills some outstanding gaps in the
standard accounts of prepositions, and fills them without multiplying
preposition contents. Co-predicating PP's can iterate like adjuncts but
still modify particular participants of clauses, which allows an account of
both path PP's and the systematic "ambiguity" between the location and
goal uses of a class of locative prepositions. Adjuncts are semantically
heads, containing predicates that take the main clause content as a whole
as an argument. As full-fledged predicates on full-fledged semantic
clauses, they have semantic privileges of their own, including the "selec-
tion" of appropriate situations as their arguments. Argument PP's ground
the entire analysis. It is the argument principle that offers a motivated
account of the semantics introduced by co-predicators, by supplying a
SITUATIONS AND PREPOSITIONS 351

distributional rationale for particular preposition contents. Argument


PP's are also critical to a clean account of valence alternation.

3.3 Bondedness Relations

In the last section we noted that the semantics given by a co-predication


analysis of (6) was somewhat underspecified. Parallel considerations call
into question the adequacy of (4T) and (5T). Nothing in the semantics of
(4) says that it is coming against the vase that seals the hammer's fate,
just as nothing in the semantics for (5) says that it is being hit by OBL
that sends the ball on its fateful journey. In fact nothing even determines
the order of these events; as far as the requirement imposed by the
situation-type go, the impingement event in (4) could have happened
long after the breaking of the vase.
I think charges of underspecification are serious business, particularly
since it was to avoid underspecified semantics that I abandoned the
account of section 1 in favor of an account which recognized co-
predication. There, however, the issue could be formulated in terms of
grammatical function; with transitive verbs, the direct object entered
into a fixed relation with a co-predicating PP; with intransitive verbs, it
was the subject. Here, the issue is purely semantic. In the case of (5), for
example, the issue is, do we need an account of causality to be packaged
in with our account of (5)? There are two ways we might do this. First,
we could posit a new verb hit with a new argument slot representing
result, and essentially call (5) a case of complement control. Second, we
could posit a particular translation rule for cases like (5), using some
explicit predicate CAUSE. I will examine some consequences of the first
alternative in section 4. What to do in this section is argue on very
general grounds against the second alternative and some versions of the
first by arguing against an exPlicit predicate CAUSE in our clause
semantics.

3.3.1. What fits in a clause. Consider the "discourse:"

(7) Jack hit the ball. It (the ball) hit the fence.
(8) The ball hit the fence. Jack hit it (the ball).

Now both (7) and (8) are fine, healthy texts, and both allow sequential
interpretations on which the event in the first sentence happened before
the event in the second. Yet the sequential interpretation of (7) makes a
more strongly connected text. Why? Because in (7), the hitting of the ball
352 JEAN MARK GAWRON

can be regarded as an efficient cause (to use Aristotle's term) of the


hitting of the fence. That is, under a fairly popular theory about how
natural laws work, the events in (7) have a natural connection. Now
certainly we can construct a causal connection in (8), for example, a
protective feeling for the fence on the part of Jack, but any such
construction requires building a special context attributing some special
disposition to Jack. So (7), I claim, has a natural causality; (8) does not.
To impose an interpretation is to impose a relationship on the sentences
in the text. This is why a sequence of sentences about events with no
apparent connection, with no shared participants, is hard to interpret.
Causality is an important relation between events in the world. Its textual
image is thus an important relationship in texts about events. What I
want to claim, then, is that the importance of causality in intraclausal
relations follows from its importance in discourse, that we no more need
to introduce a predicate C A U S E into the semantics of the clause, than
we need to introduce it into our theory of discourse interpretation.
The problem with this, of course, is that (7) and (8) are both well-
formed discourses. But the events described in (7) can be described with
(5); the events described in (8) cannot. So we need to say something
special about the clause as a discourse domain. We need to claim that
only certain specially privileged discourses can be fit into a clause; those,
quite naturally, are the most tightly connected.
Here is a scenario for (5): The clause is a small stage which mounts
only one scene per performance. Generally a scene accommodates only
one located fact and the players associated with it. The relation in this
fact comes from the verb. In those very rare instances when two located
facts are allowed onstage, they must be bonded together in one of a very
small number of ways. One such bonding relation is being-an-efficient-
cause-of. Since there is only one way to order the facts in (7) and achieve
efficient causality, the only interpretation of (5) is one in which the hitting
of the ball precedes and in fact causes the hitting of the fence.
Similar remarks apply to (6). Walking can be a way of getting places.
Thus a walk is a natural efficient cause for being under a bridge. This
account attributes the goal/locative ambiguity of prepositions to the
interaction of locative meaning with efficient causality: movement yields
location. Note that the account is asymmetric. The prediction is that all
locative prepositions will be able to be used as goal prepositions, but not
vice versa. This appears to be correct, So to and into are never just
locative. But even on has its goal uses: Tanya threw her food on the floor.
This may do for Jacks, balls and bridges, but will causality account for
everything? If instrumental with is a co-predicator in cases like: Jack
SITUATIONS AND PREPOSITIONS 353

fought the dragon with the magic sword then do we want to say that a
C A U S A L link joins the fighting fact with the incidence of sword and
dragon? Or, again, do we want to say that Jack's breaking the hammer
has a causal connection with its hitting the vase in (4)? The answer in
both cases seems to be no. Rather, the prepositional fact is constitution-
ally part of the verbal fact, in a sense of "part of" which is quite distinct
from our formal "part of" relation between situations. It is in the nature
of relations that two objects can be in more than one relation at once;
sometimes the event associated with one relation is indissolubly part of
the event linked with the other, even when the "part o f " relation is not
a necessary one. Thus, the incidence of the sword and the dragon is one
piece of what we call Jack's fighting the dragon. Call this relation
between facts constituenthood.
How else might facts get bonded together? Presumably, via the
" c o m p o n e n t " relation, as well. We will say one clause fact fl is a
component of another f2 when the relation in fl is a component of the
relation in fz, and the other objects in f~ are identical to the correspond-
ing objects in f2. (1T) is an example. The I M P I N G E M E N T relation is a
component of the hit hit relation, because constraint I says its relation is
a component of hit, and because its location and participants correspond
to the ones in the hit fact. Componency may just be the limiting case of
constituenthood; or better, perhaps, componency may simply be neces-
sary constituenthood. Impingement is a necessary constituent of hitting;
Incidence (or the introduction of an instrument) is only a contingent
constituent of fighting. (See the remarks in section 4.3 on how the notion
of necessity may help us distinguish between co-predicators and
arguments.)
It is not necessary that these be the only bonding relations; there may
be others, though there probably shouldn't be too many. And the others
should only be allowed to compete with causality and constituency in
cases where there are really two understandings about the relatedness of
facts in a single clause.
Because our contents are objects with a great deal of structure, it is not
too difficult to reconstruct the notion in the semantics; the single clauses
we have looked at have been associated with situation-restrictions that
pick out situation-types. Within those situation-restrictions there are
usually several facts, but only those corresponding to verbs and prep-
ositions are located-facts. The bondedness relations that I have talked
about thus far need only apply to located facts. 4 All these constraints, of
course, apply only to top-level facts in a situation-restriction. There may
be embedded situation-restrictions (when a verb takes a clausal corn-
354 JEAN MARK GAWRON

plement) with entirely different casts in different bondedness relations; or


there may be situation-restrictions hidden inside roles.
W e might even make what is a more theoretically ambitious move,
more in keeping with the realist spirit of situation semantics, and explain
our concept of clause in terms of a more constrained definition of actual
situation. As things stand now, anything can be an actual situation: the
fact of Augustus being emperor of R o m e and the fact of my writing this
paper now can, in the formalism, be knit into a single abstract situation.
But perhaps actual situations can only contain facts knitted together in
certain tight relations of causality and constituenthood - our bonded-
hess relations. 5 And perhaps clauses are tailored to describe actual
situations.
In constructing a scenario for bondedness relations I have appealed to
some notions which are going to be difficult - or even impossible - to
make precise, not the least of which is the notion "efficient cause". But
wherever I have sinned an account with an explicit predicate C A U S E
will have to sin at least as grievously. It is precisely the notion of an
efficient cause which the predicate will be intended to capture; it isn't
that Jack hit the ball and then some great chain of events ensued, and
then finally, after a trip on a barge through Detroit, the ball hit the fence.
T o say this and no more is to call it a toss-up. T h e pressing question is
why one would R A T H E R talk in terms of bondedness relations than in
terms of an abstract predicate C A U S E . T h e argument for the bonded-
ness view hinges on one major point: an account which leaves the
ordering of the facts in the clause up to some notion of "natural causal
relation", as the bondedness account does, ultimately appeals to our
shared views of how the world works. It claims that what makes the
hitting the C A U S E of the impingement in (5) is the same thing that
makes (7) a more complete discourse than (8). In contrast, to appeal to a
predicate C A U S E is to commit oneself to an explicit description of the
linguistic patterns that express a notion which will be central in any
theory of discourse. T h e C A U S E account must stipulate the natural
causal ordering for each verb and preposition pair. Compare the hammer
broke against the vase with Joan hit the ball against the fence. In the
breaking case, the preposition fact gives the cause, the verb fact the
result. In the hitting case the tables are turned. T h e C A U S E account has
to stipulate this contrast in two different rules, whether lexical or syntac-
tic. But this fact of English just follows from folk theoretic physics.
Impingements are very natural causes of breaking. A g e n t e d impinge-
ments are very natural causes (via movement) of further impingement.
What really seems to be required is not a sorting by lexical or syntactic
SITUATIONS AND PREPOSITIONS 355

rule, but a general account of how we understand facts to be knitted


together by efficient cause.

3.3.2. English causatives. In support of these highly speculative


remarks. I would like to draw attention to a very different motivation for
bonding relations. Whatever we say about what can be semantically
packaged into a clause ought to also provide an upper bound on what can
be semantically packaged into a lexical item. It thus ought to contribute
to our characterization of what a possible lexical item is. The notion of
cause has, of course, quite a notorious role in the lexicon, both in
morphologically rich languages with causativization affixes, and in
English, where an extraordinary number of verbs can be both transitive
and intransitive, with the transitive interpretable as a causative version of
the intransitive:
(9) The vase broke.

(10) The hammer broke the vase.


If what I have suggested about efficient-causality as a bonding relation is
true, then the relation between the two entries for break is very simply
expressible:

[breakr: V: {(LOC $LOCO


REL break
SUBJ $IND0
POL $POLO)}]

[breakT:V: {(LOC $LOC0


REL break
OBJ $IND0
POL $POL0)
(LOC $LOC1
REL ACT
SUBJ $IND 1
POL $POLO)}]

I assume, for parsimony, that the predicate ACT here is general enough
to allow for both instigating and non-instigating actors; that is, the
transitive entry above should do for both John broke the vase and the
hammer broke the vase. This, then, raises an interesting question when we
consider the interaction of the transitive version with co-predication:
(lla) Jeanette broke the vase against the wall.
356 JEAN MARK GAWRON

(llb) Jeanette broke the vase.

(llc) T h e vase broke against the wall.

( l l a ) does not cover a scenario in which Jeanette knocks over the vase
and it hits the wall and breaks, or one in which she rolls the ball down an
incline and it hits the wall and breaks, or one in which she throws the
vase at the wall and it breaks (the judgement is perhaps less secure here).
Yet all three scenario's are consistent with ( l l b ) and (11c). This
difference in " t r u t h " or "description" conditions would at first blush
appear to be an argument against using the same verb content for all the
sentences in (11).
Before mounting a dress rehearsal of how bondedness relations may
help account for the extra specificity of ( l l a ) , let me try and articulate an
intuition about what is going on. First, the three scenarios for which (1 la)
is inappropriate all involve Jeanette becoming physically separated from
the impingement fact itself. It seems that once the extra impingement
fact enters the scene, the sort of effect Jeanette can efficiently exert
becomes constrained. She can move-guide the vase towards destruction,
but she cannot simply give it a shove. T h e vase must be manipulated. I
want to try and capture this contrast by distinguishing between con-
tiguous causality and non-contiguous causality. In the usual case, in-
stigating actors have a great deal of leeway about how they b e c o m e
efficient causes. It is quite difficult to say, as one constructs more and
more elaborate causal chains linking Jeanette to the vase's breaking, just
when we can no longer use ( l l b ) to summarize matters (though we
clearly c a n go too far). In the case when the event being described
becomes more complex, however - when it becomes a chain of events
such as a ,vase impinging against a wall and then breaking - then an
instigator's action must be contiguous to the chain of events: she must be
physically connected to the impingement.
Let me try and sketch how an account of these facts based on
"contiguous" causality might go. Consider the semantics that results for
( l l a ) if we use the transitive entry for b r e a k given above:
(llT) {($LOC1, A C T , $A)(Jeanette, $A)
($LOC2, break, $B)(the-vase, $B)
($LOC3, I M P I N G E M E N T , $B, $C)(the-wall, $C)}.

Let us suppose that bondedness relations must be able to connect all the
located facts in a clause, and that when there are more than two such
facts, the connections are made in a " n e s t e d " fashion; thus, in the case of
three located facts, two facts are first bonded together into what we will
SITUATIONS AND PREPOSITIONS 357

tentatively call a complex fact, and then that complex fact must be linked
by some bondedness relation to the third. This gives us three possible
"bondedness structures" for (llT), which we may write, schematically:
(i) R~((IMPINGEMENT $B, $C), R2((ACT $A), (break $B)))
(ii) R~(ACT $A), R2((IMPINGEMENT SB, $C). (break $B)))

(iii) Rl((break $B), R2((ACT, $A)(IMPINGEMENT $B, $C)))


For simplicity, let us assume that the bondedness relation that comes into
the issue with these examples is efficient causality. This eliminates struc-
ture (i), because the impingement is "mediating", it comes between the
action and the breaking. Let us call a complex fact knitted together with
efficient causality a causal chain. Then the following assumptions explain
the peculiar "narrowed" sense of (1 la).
(I) Only a fact involving the relation A C T can be the efficient
cause of a causal chain.
(II) When an A C T fact is the efficient cause of a complex fact
then the kind of causality linking it to the complex fact must
be contiguous causality.
Assumption (I) immediately rules out possibilities (i) and (iii) for our
bondedness structure. Assumption II guarantees that the kind of action
that causes the complex impingement-break fact is spatio-temporally
contiguous with it. This allows a move-guide action, but rules out a
simple push. Compare this with Jack rolled the tire against the wall. Here
again, roll is a causative verb, a causal chain will be involved, but since
the mediating relation will be a roll rather than an impingement, it will be
contiguous even with the quickest push. Or, again, compare Joan pushed
the ball against the wall. Since push is not causative, no complex fact is
involved, and a quick shove is allowed. The prediction here is that no
causative verb will allow non-contiguous causality with a co-predicator.
Similar assumptions may help explain the peculiar:
(12) The hammer broke the vase with the nail.
On any interpretation in which the hammer isn't a causal instigator, this
sentence seems out. How do we explain this with our proposed lexical
entries, under the assumption that A C T allows both instigating and
non-instigating actors? The direct approach may here be the best:
(Ill) An A C T fact that is the efficient cause of a complex fact must
involve an instigating actor.
358 JEAN MARK GAWRON

One general account of the unacceptability of sentences like (12) that has
been proposed is that the "instrument" is a core participant of the verb
and that in (12) the "instrument" role has been filled twice. That
something more general may be going is shown by the equal unac-
ceptability of the hammer broke the vase against the wall and the bat hit
the ball against the fence. Again, for three participant sentences, an
animate instigator seems to be required. An account based on bonded-
ness relations would make the same requirement of the subject in both
sentences.
Given our "causative" entry for break, Principle (III) automatically
accounts for the break sentence. In order for it to extend to the hit case,
we need to analyze hit as a causative verb. Some such extension of our
analysis appears to be called for in any case: since we have analyzed the
"instrument" as an explicit argument of hit, our current three-place entry
for hit can not account for examples like the stick hit the fence. We noted
in 1.1 that all the other impingement verbs come in both three- and
two-place versions (the a and /3 verbs of our I M P I N G E M E N T con-
straints); if we assume that all the/3 versions are two-place specifications
of I M P I N G E M E N T , then the two-place sentences are accounted for, and
correctly related to the three-place sentences. The three-place valence of
hit used in (1) would now be given by the following entry:
[hit: V: {(LOC $LOC0
REL hit2
OBL $IND1
OBJ $IND2
POL $POL0)
(LOC $LOC1
REL ACT
SUBJ $IND0
POL $POLO)}]

against would be eligible to mark the argument realized as O B L because


hit2 has I M P I N G E M E N T as a component.
Features like instigation, animacy, and control seem to be intrinsically
fuzzy and the critical facts here grow murky quite quickly. Contrast (12)
with something like the water wheel broke a paddle against the dock. My
intuition is that the water wheel can be considered instigating here, since
it is sufficiently under its own control. But even if, as is quite likely, I
have missed the mark here, the real question is whether something more
general than an idiosyncrasy of a particular verb is going on. If so, then a
general principle of clause interpretation may be the right place to state
it.
SITUATIONS AND PREPOSITIONS 359

3.3.3. Co-predication as a Lexical Rule. Bondedness relations are


obviously not mainstream stuff; their function is to fill the semantic
cracks. A real worry in having truck with such creatures of darkness is
that they may implicate us in violations of compositionality. (They won't,
if we can maintain that bondedness relations apply to actual situations.)
In that case, it would be good practice to try to limit their domain
of application. Someone particularly concerned to preserve com-
positionality in their combinatorial semantics might use a sentence like
(12) to argue that there must be a new verb break which differs from
simple transitive break in being linked with a three-place relation, in
obligatorily subcategorizing for against, and, more subtly, in the seman-
tic relations among its participants. And in so far as we can find other
meaning shifts in other cases of what I have called co-predication (as we
will in the next section), such an advocate has further support for moving
co-predication into the lexicon. The troubling thing about this move is
that lexicalization is generally the recourse when confronted by "in-
explicable idiosyncrasy". And if what I have been saying about bonded-
hess relations is anywhere near correct, then there are regularities to be
accounted for in the way that semantic components like break and
against combine, whether that combination is licensed directly in the
combinatorial semantics, or in the lexicon by valence-changing rule. In
principle, I see no objection to moving co-predication into the lexicon as
long as some account of the kind of semantic relations we have been
discussing is preserved.
Thus, if the co-predication rule were re-stated as a lexical rule, it might
output a lexical entry for break that looked like:

[breakr: V: {(LOC $LOC0


REL break
OBJ $IND0
POL $POL0)
(LOC $LOC1
REL ACT
SUBJ $IND1
POL $POL1)
(LOC $LOC2
REL IMPINGEMENT
OBJ $IND0
OBL $IND1
POL $POL2)}]
A slightly modified version of the Argument Principle for Prepositions
would correctly allow this "verb" to subcategorize for against, and using
360 JEAN MARK GAWRON

the combination rule for obliques given in section 1 would give some-
what redundant, but correct, semantics for sentences like (11). Then, for
those so inclined theoretically, all talk of "bondedness" relations can be
confined to the department of intra-lexical affairs. Indeed, one could
argue, a lexical rule might be the only way to handle co-predication in a
language that covered the sort of semantic ground we have been con-
sidering by valence-changing morphological operations like prefixation
(for example, Polish or Hungarian). Yet even with these amendments, all
the semantic points discussed here go through. The real thrust of lexical
semantic parsimony is not to keep the number of lexical entries down; it
is to seek some guarantees for the integrity of lexical semantic structure,
to require, where the language allows it, that the semantically distinct
entries we have can be related in semantically principled ways. The most
trivial such relation is identity. A morphology-poor language like English
encourages us to build morphologically poor lexicons, and since this is an
unabashedly English-bound paper, that has been the strategy I followed.

4. CONTROLLED COMPLEMENTS AND CONTROLLED


ADJUNCTS

I want briefly to consider still another account of sentences like (4), one
which assimilates them to the phenomenon of control. Bresnan (1982)
distinguished two different kinds of control, exemplified in:
(16) Virginia considered Jack foolish.
(17) Virginia visited the Louvre drunk.
In both cases one of the NP arguments of the verb can be called the
controller of the adjectival complement, in (16) the direct object, in (17)
the subject. But in (16) the adjectival complement is obligatory and
alternates with NP-complements; Bresnan calls it an XCOMP. Analyses
of sentences like (16) have a long history in generative grammar dating
back to Rosenbaum (1967). (17) is an example of something which has
generated relatively little grammatical literature, a constituent that does
not appear to be a genuine complement of the verb, but is controlled by
one of its arguments; Bresnan calls it an XADJ.
I will consider X C O M P and an X A D J analyses of (4) in turn, starting
with the X C O M P analysis.

4.1. The Role of PCOMPs


First, consider a classic example of a raising verb, expect, with a PP
SITUATIONS AND PREPOSITIONS 361

complement:
(18) I expected John in Rome
(18T) {($LOCI, exp, $A, {($LOC2, in, $B, $C)})
(Jack, $A)(the-man, $B)(Rome, $C)}.
(18T) gives a likely content for (18), along the lines of a classic raising
analysis. The chief difference between (18) and a translation using the
Co-Predication Rule is that in (18), the situation-type formed by com-
bining the PP with the direct object is an argument on the verb; in
Co-predication it is not. It is simply fused with the described situation by
union. We could thus assimilate the account of co-predicating PP's to
complement control simply by always adding an extra argument to the
verbs in question. Since the result will exhibit a rather varied cast of
characters it will facilitate discussion somewhat if we label argument
positions in lexical entries with mnemonic role names; this is done purely
as a convenience; it is not a covert introduction of deepcase/
thematic/theta/karaka roles:

[hit: V: {(LOC $LOC0


REL hit
SUBJ $Agent
OBJ $Patient
OBL $Inst
XCOMP $Result
POL $POLO)}]
[break: V: {(LOC $LOC0
REL break
SUBJ $Agent
OBJ $Patient
OBL $Goal
POL $POLO)}].

One argument against the control account is parsimony, but the


argument needs to be made carefully.
Consider (19):
(19a) Jack hit the ball against the fence with the aluminum bat.
(19b) Jack hit the wall.
An entry like the above one for hit will handle cases like (19a), but it will
not do for (19b). A defender of a single-entry account might try to claim
362 JEAN MARK GAWRON

that if results aren't explicit in the sentence, described situations will


simply randomly instantiate them, parallel to our account for instruments
with hit. But in fact hit takes only a very special kind of result involving
movement. T h e result-PP must describe the trajectory or ultimate
destination of the patient after being struck. But, as (19b) suggests, it is
simply not the case that every Patient of a hitting situation has such a
trajectory or destination. In short, (19a) entails m o v e m e n t of the patient;
(19b) does not.
Thus, a controlled-complement account of hit will involve a new verb
content with a four-place lexical relation; the new argument position will
be filled by an obligatory trajectory/goal. Parsimony argues against this
move, but for a theorist, parsimony may not always be the prime
consideration.
As it happens, we have already seen an example that is problematic
for a control analysis of the "result" slot in hit, namely sentences with
more than one co-predicator specifying the path and goal: Joan hit the
ball through the narrow alleyway into Mrs. Magillacuddy's window.
Similar, stacked co-predicators can be adduced for other verbs we have
looked at:

(20) A d a m broke the hammer into pieces against the vase.

Exactly the same considerations that led us to argue for a co-predicating


account of against the vase in (4) can argue for a co-predicating account
of into pieces in (20). T h e r e is a class of semantically-related transitive
verbs such as turn, transform, and transmuted which can plausibly be
argued to be three-place verbs, and use to establish a transforms-into
sense of into. T h e imposition of a transform-into situation on break also
involves imposing the role of transformed-object on the direct object,
and only a predicational structure involving control can do this. Yet with
two complements controlled at once a controlled-complement account of
(20) becomes highly problematic.
Presumably, control entries for verbs like hit and break would make
use of the available syntactic structures in the grammar. But no current
framework that I know of allows more than a single controlled com-
plement of a verb. In LFG, for example, a control account of (20) would
require a new grammatical function (an X C O M P 2 or some such), which,
as far as I know, would be needed solely for the co-predication cases.
This high theoretical price is the best argument against a controlled-
complement account of the co-predication cases.
Lest readers wonder whether I am calling into question the very
existence of controlled PP's, I note here that there are some cases where
SITUATIONS AND PREPOSITIONS 363

a controlled-complement analysis of a PP will not only be possible; it will


be essential. Consider an obvious candidate, the verb thrust.

(21) She thrust the money into their hands.

(22) *She thrust the money.

In contrast to hit, thrust always requires that the Patient move, a


requirement reflected by the syntactic obligatoriness of the PP comple-
ment. Note also that a number of different locative/goal prepositions can
fill the same obligatory slot, resulting in sentences with different mean-
ings and different truth conditions. Given the Argument Principle, this
fact alone rules out an argument PP account. The Argument Principle
says that the argument PP's are redundant. That means that if the same
verb content shows different valences with different prepositions, and
minimal pairs are not paraphrases, then the prepositions do not mark
arguments of the verb. But thrusting something into some tiring is
different from thrusting it onto some thing, or under some thing. So the
PP's with thrust do not mark its arguments. The only account left that is
compatible with a syntactically obligatory complement, an accoun~ that,
for Bresnan, predicts the syntactic obligatoriness, is a control account.
I want to now consider a verb in some ways similar to thrust, insert:

(23) Jack inserted the key into the lock.

(24) Jack inserted the key.

(23) and (24) suggest that insert, too, requires a three-argument account:
although the PP-complement of insert is not syntactically obligatory, (24)
still implies the existence of a particular containing space; in fact, it
presupposes it, since the negation of (24) has the same entailment.
In the last section, we exploited the fact that a verbal lexical entry has
an entire situation-restriction by including more than one fact in certain
lexical entries (the causatives). We can deal with lexical presuppositions
like those of insert in a similar way:

[insert: V: {(REL insert


SUBJ $IND0
OBJ $IND1
OBL ($IND2, {(REL the-container
ARG $IND2
POL 1)})
POL $POLO)}]
364 JEAN MARK GAWRON

insert is here treated as a three-argument non-control verb; and an extra


fact, embedded in a role, records the presupposition about the OBL
indeterminate using the same kind of semantics we have been using for
definite determiners. Presumably, when a semantics for resource situa-
tion has been worked into our system, this treatment could be modified
to accord with it. Since the entire role that is the value of the OBJ
attribute is an indeterminate, it can be substituted for by a conventional
syntactic filler if one is available. In that case the entire role would
disappear, and there would be no definite-reference involved with the
clause; this insures that a sentence like He inserted the key into a lock,
isn't required to presuppose a lock. 6
insert and thrust are different in that the prepositional complement of
insert is not syntactically obligatory. They are alike in that the oblique
object must somehow be available or be made available to the discourse;
it is, so to speak, semantically obligatory. Charles Fillmore (personal
communication) has pointed out to me a number of other examples in
which a verbal complement, although zeroed, must be interpreted as
having definite reference. They include sentences like He won, John
accused Mary, and Sonya applied. In each case, the sentence can only
be felicitously used in a discourse if a unique referent for the missing
complement can be supplied: there must be a relevant conflict/contest,
offense, or position; the same seems to be true of the negations of these
sentences.
Despite the logistical similarity between inserting and thrusting, I have
analyzed insert as a non-control verb. This is possible because insert is
different from thrust in another way: it is much more selective about its
prepositional complements:

(25) John inserted the key under the door into the crack.
(26) John thrust the key under/past/behind the door.

In (25) only the crack is the container; under the door is a circumstantial
path PP, another co-predicator. There can be any number of such path
PP's, but the only preposition that seems to be able to mark the container
itself with insert is into. Thus, the Argument Principle allows us to
analyze it as a three-argument non-control verb. Furthermore, analyzing
it that way facilitates the statement of the lexical presupposition in the
above lexical entry. If insert were analyzed as a raising verb like thrust,
there would be no obvious way of stating presuppositional information
about the object of the preposition, because the whole PP complement
would correspond to a single $SIT indeterminate in the lexical entry.
SITUATIONS AND PREPOSITIONS 365

Similar minimal choices sometimes arise in deciding between a co-


predication analysis and a control analysis. Compare He wheeled the
gurney down the hall and He pushed the gurney down the hall. T h o u g h
push and wheel are logistically quite similar, like insert and thrust, the
sentence with push must be analyzed as co-predication, the sentence with
wheel as complement control. Two considerations enter in here: first,
only with wheel is a path or goal complement obligatory; second, only
push is like hit in not entailing the m o v e m e n t of its direct object. Thus,
only wheel has m o v e m e n t as a necessary component. We will say a bit
more about the role of "necessity" in section 4.3.
A right-thinking realist might well be troubled by these decisions:
logistically similar verbs should have similar argument structures. With
insert, the choice was difficult precisely because we had hit a kind of
limiting case. PP X C O M P ' s have a peculiar semantic status. T h e y take
no right recursive modifiers, so that they can drag in no new parti-
cipants apart from the prepositional object. As the kind of situational
information that can be a c c o m m o d a t e d in an X C O M P PP becomes more
and more constrained, the P - O B J will look more and more like an
argument, because the constrained X C O M P situation will simply be an
extension - or rather a c o m p o n e n t - of the verbal situation. At that point
the choice between an argument account and an P C O M P account has no
" r e a l " (i.e., logistical) consequences, and it is made according to
theoretical convenience. A n o t h e r way of making the same point: we
could have analyzed all argument-PP's as X C O M P s , as far as the
logistical "facts" of the matter went, because we claimed that in the
contents of clauses with argument-PP's, the prepositional fact was
redundant with the verb fact. But that would have lost us a neat account
of valence alternation; it would also, incidentally, have conflicted with
the predictions made by most syntactic theories about reflexives in
controlled complements. So, general theoretical considerations dictated
an argument analysis. Similar considerations can help us make our choice
with insert and wheel.
But how do we reconcile such cavalier talk of theoretical convenience
with our realist commitments? Surely, if the insert relation is real, its
"third" argument is either a situation or a container, and that is a fact
about the facts of the matter, and not a theoretical choice. The truth is
that I don't think being realists is going to help us decide things about
argument-structure. In general, we shall have to assume that whatever
can be c o m p r e h e n d e d as a totality can be used as an object in our
ontology; the world is chock full of relations, only some of which we
carve out for special use. Both candidates for the insert relation are real,
366 JEAN MARK GAWRON

in the sense that both can be comprehended as regularities that other


objects can "stand in" across situations; But only one will be associated
with the English regularity written "insert", and the decision about which
must be made by consulting the general system of regularities constitut-
ing a grammar.

4.2. The X A D J Account


We turn now to the question of whether co-predicators can be analyzed
as XADJ's like that in (17). In one sense, a co-predication account IS an
XADJ account. Consider the following rule for composing the semantics
of (17), Virginia arrived drunk.

(S---~NP VP (XP): QI([[NP]]SUBJ U([[XP]][[VP]]))}


{($LOC0, drunk, $A)(Virginia, SA)($LOC1, arrive, $A)}.

The XADJ rule here assumes all XADJ's exhibit the function SUBJ;
given this the NP will be QI'ed in for all instances of the SUBJ function
and will automatically "control" two positions. Except for this point the
rule quite resembles our semantics for a co-predicating PP with an
intransitive verb. A similar, if slightly more elaborate, analogy could be
drawn for the transitive cases like Jack ate the fish raw.
But before simply concluding that what I have called co-predicators
are syntactically all XADJ's, there are two important issues to deal with,
one syntactic, one semantic.
To begin with, there is an important syntactic difference between
examples like Virginia arrived drunk (call these predicative XADJ's)
and the co-predicators discussed in sections 2 and 3. Predicative XADJ's
seem to select their controller freely from among the nuclear terms; thus
we have pairs like Jack ate his dinner drunk and Jack ate his dinner raw
with the same verb. The controller of a co-predicator seems to be syntac-
tically fixed for a given verb; thus with transitive break the controller is
the direct object, with intransitive break, the subject. At the moment I
have no theoretical commitments to offer that make subcategorization
the natural way to express this difference, but ! would argue that some
syntactic difference must exist, despite the semantic likeness between
XADJ's and co-predicators.
For all that semantic likeness, there is also a semantic difference
between co-predicators and predicative XADJ's: the truth conditions
associated with the verbal participants may change with co-predicators in
the clause. Showing this was the burden of (11), reproduced here:
(27) John broke the vase against the wall.
SITUATIONS AND PREPOSITIONS 367

Where (27) cannot be used to describe a situation in which John knocks


the vase against the wall, John broke the vase can. Furthermore the
occurrence of a co-predicator can have selectional effects on participants
other than its controller; we noted that instrumental subjects do not seem
to be possible when a causative verb co-occurs with a co-predicator. I
know of no parallel phenomena with predicative XADJ's.
One possible explanation for these semantic contrasts is this: bonded-
ness relations are constraints on what located facts can be packaged into
a single clause: perhaps XADJ's don't contain located facts. In support
of this conjecture, note that verbal XADJ's must be non-finite, either
past or present participles; it could be argued that the nature of such
modifiers was fundamentally that of unlocated predication.
But this move runs up against some immediate problems. (17) shows
that a class of predicative XADJ's are adjectives; calling all adjectival
complements unlocated would prevent extending the analysis of resul-
tative PP's offered here to resultative adjectives:

(28) He hammered the nail flat/into the wood.


(29) He mailed the diploma fiat.
(28) with the adjective is as resultative as it is with the PP, as a resultative
as any sentence like example (5); (29), with the same adjective, is a
typical XADJ. One might try claiming that (28) involved located
flatness, while (29) did not, but this would miss an important facet of the
semantics of (29). The diploma is flat precisely while being mailed, not
necessarily before, and not necessarily after. What seems to be called for
in (29) (and in (17), as well) is that the location of the adjective fact be
unified with that of the verb fact. The following modified semantics for
XADJ's would do that:

X A D J (version 2):
QI([[NP]] SUBJ M([[XP]] LOC [[VP]] LOC)).
If we adopted version 2 of the XADJ rule, we could leave what we said
before about bondedness relations unmodified except in one particular:
bondedness relations applied to distinctly located facts.
Yet another possibility is to abandon any attempt to distinguish struc-
turally between predicative XADJ's and co-predicators, and simply say
that the differences between can all be explained in terms of the different
bondedness relations inferred. Bondedness relations involving causality
may involve constraints on all the participants in the clause, resulting in
"narrowed" contents like that of (27). The bondedness relation "depic-
368 JEAN MARK GAWRON

tion,, to borrow the term used in Halliday (1967) for cases like (29), will
only involve the P-OBJ and its controller; it will also force interpreting
the verb location and the adjunct location as the same. This account
gives us a new bondedness relation whose effects must somehow be
explained. One could often rule out a depiction reading on the basis of
the relation in the XADJ fact, since it would presumably have to be a
state relation; thus, relations like into and against would be non-depic-
ters. This would still seem to allow a depiction reading for sentences like
the adjectival version of (28), but perhaps that reading is just barely
possible. A good deal more would have to get said about the pragmatic
function of the depiction relation before such an account could be taken
seriously. For example, what factor governs the contrast in (30)?
(30) Lanya bought/?washed the car new.

In sum, it is quite possible that XADJ's could subsume co-predicators, at


least semantically, but the semantic half of the issue can only be resolved
in the context of a fully fleshed out theory of bondedness relations.

4.3. Should Co-predicated PP's be Subcategorized-for?

Suppose that the question of whether all co-predicators were XADJ's


were resolved in the negative, and that, at the very least, a syntactic
difference was assumed. Then we would still not have established that
co-predicated PP's must be subcategorized-for. Although I have called
them subcategorized-for throughout this paper, I have offered no direct
argument for that choice, and I am not at all sure how to go about
making such an argument. For some theories, say one which directly
adopts Chomsky's Projection Principle, there is no choice at all: sub-
categorization is equivalent to argumenthood. One can only argue for or
against both simultaneously. For others, like GPSG, in which, for
example, a syntactic VP can, via control, correspond to a semantic
proposition, there is no direct commitment to argumenthood for all
subcategorized-for complements. There is, however, a general commit-
ment to a small inventory of ways in which subcategorized-for comple-
ments can be semantically combined with their heads (see, for example
Klein and Sag, 1985); a new subcategorized-for complement-type with a
new combination operation does not, therefore, come theoretically cost-
free.
An alternative to complicating one's theory of subcategorized-for
complements, which still makes the distinctions I've claimed must be
made, is to say that there is a third domain, neither adjunct nor sub-
SITUATIONS AND PREPOSITIONS 369

categorized-for, into which the co-predicators fall. Co-predicators would


be distinct from subcategorized-for complements in that they would
never be mentioned in a verb's lexical entry; they would be distinct from
all adjuncts in that they would have grammatically defined control
properties, and enter into bondedness relations.
Still another possibility is the lexical rule route: one could sub-
categorize for all co-predicators and still call the oblique participants
"arguments" by using "promotion" rules which turned out the kind of
complex verbal contents we discussed in section 3.3.
By one theoretical device or another, co-predicators will have to have
a different status for run-of-the-mill argument PP's. Both co-predicated
PP's and argument PP's can be said to semantically select for certain
verbs; but the Argument Principle draws a sharp line between the types
of selection. The content of against is the situation-type for impinge-
ment; thus against can mark the arguments of precisely those verbs
whose contents involve impingement. That gives us one "sub-'category
of verbs. Given what we've said about bondedness relations and the
meaning of against we would expect against to appear as a co-predicator
with various change-of-state verbs in which the change of state can be
effected by forceful movement: crack, bend, and cut, for example. Have
we now identified another "sub-'category of verbs? Argument PP's
D I R E C T L Y partition verbs into situation-types according to their con-
tents; but against PP's used as co-predicators "sub-'categorize verbs
like cut and crack according to the role of their contents in the causal
order of things. Those roles are not necessary roles. Thus, a hitting is
always and everywhere an IMPINGEMENT (just as a kissing is always a
touching); but a breaking does not necessarily involve an impact (the
soprano's high C broke the glass).
The interesting question then is how this difference between necessary
and non-necessary connection should be reflected by constraints. The
situation semantics account of non-necessary constraints seems to still be
developing but the general idea, judging from the account of meaning
constraints given in Barwise (1984b), is that non-necessary constraints
are constraints read against a background of other constraints. Part of
the background for a meaning constraint, for example, is David Lewis's
truth-telling convention, the convention that language users use
language truthfully. That "convention" is, presumably, itself a back-
grounded constraint. But the question of how in general such back-
grounds should be formulated, and how in particular they would be
formulated for the logistics of cutting and breaking is, as they say,
another story, one that will not be told in this paper.
370 JEAN MARK GAWRON

5. ADJUNCTS

I assume that iteration is a useful diagnostic, but neither a sufficient nor a


necessary condition for adjuncthood. Both co-predicators and adjuncts
can iterate. My chief criterion for calling something an adjunct PP rather
than a co-predicator will thus be semantic intuitions about the sort of
predication a preposition's lexical relation makes. In this preliminary
treatment, I will discuss only two relatively clear cases, benefactives and
locatives.

5.1. Locatives: Extending the Formalism


W e begin with what appears to be the simplest case, a kind of paradigm
example of prepositional modification: the locative. Barwise (1984) treats
all adjuncts as located-relations between objects (either individuals or
situations, in particular). While I will argue below that this is unsatis-
factory for all adjuncts, I believe it is the right course for locatives. T h e
following rendering of "Miles walks in R o m e " should make the intuition
clear:

(1) in e: at I; in, el, Rome; 1


in ea: at 11; walk, Miles; 1.

T h e hitch in writing the semantic rule for this is that we must construct a
situation type (corresponding to e above) which has something true in it
(Miles's walking) and something true of it (its being in Rome).
We now see a serious limitation on the contents allowed thus far. So
far, we can only constrain a described situation via its constituent facts,
constrain it, thus, only in terms of its internal structure. But now it
appears that in treating certain adverbial adjuncts we want to constrain
the described situation in terms of predications on it, in terms, thus, of
external constraints. We can use roles to express external constraints on a
situation-type, and situation-restrictions to express internal constraints,
but to use both at once means using more than one indeterminate at once
(since any roles a situation bears will not be roles inside itself).

5.1.1. On why external predication is necessary. One might try to get


around constructing contents with external predications on them by
having prepositions like in be internal one-place predicates. On this
account, the indeterminate for " J o h n walked in R o m e " would be:

{($LOC1, walk, John)($LOC2, in, Rome)}.


SITUATIONS AND PREPOSITIONS 371

This may seem a little strange at first, but there is no a priori objection to
treating in this way. First, in our new ontologicaUy promiscuous frame of
mind, just about any regularity we can consistently identify can be found
a real relation. Second, there is going to be a general correspondence
definable between internal predications in a situation and external predi-
cations on it. For example, we can define a predicate H I T T I N G - S I T true
of a situation just in case it contains a fact involving the relation hit. Why
not, then, go in the other direction, and say that those actual situations of
which one can truthfully predicate external-in are just those containing
facts with another relation, of one less argument, internal-in?
One argument against this approach seems to be the observation that
in some cases adjuncts do not seem to commute without meaning
change. Thus "Bill made a sweater for Mary for Miles" has a different
interpretation than "Bill made a sweater for Miles for Mary." For such
adjuncts, the internal-predication account would allow no way to dis-
tinguish between the readings.
Another consideration is that treating adjuncts as two-place relations,
and thus external predicates, will allow us to use the same preposition
contents in handling subcategorized-for prepositions and adjuncts. Thus,
the same preposition will be able to occur in both functions.
Finally, treating a locative as a two-place external predicate allows us
to use the same preposition content to handle locatives in noun phrases
like the box in the corner. Since ordinary individuals have no internal
structure, there is no possibility of handling this example as internal
predication on an individual that is also a box. But we might try to
reduce it to a predication in the situation that defines the box role. In
T C & S O the lexical indeterminate for a common noun is a role:

($IND0, {(box, $IND0)}).


This raises the possibility that the nominal box in the corner could have
the content picked out by:

($IND0{(box, $1ND0)($LOC0, in, $IND1)


(the-corner, $IND 1)})
That is, th~ role of being a box is an in-the-corner situation (again, on the
analogy of a hitting situation). But now consider what would happen if
we had a relative clause. Presumably the content of box that John bought
would be given by something like:

($IND0, {(box, $IND0)($LOC0, buy, $IND1, $IND0)


(John, $IND 1)}).
372 JEAN MARK GAWRON

But now how would we capture the contrast between nominals like
box in the corner that John bought and box that John bought in Rome? In
one case the PP modifies the relative clause, in the other, the noun. To
make this distinction we need the PP to introduce an external predication
on an individual in one case, and a situation in the other.

5.1.2. Described situations as roles. What we need is to generalize our


means of representing contents, so that we can represent more kinds of
information about them. In particular, we need distinct places to record
their external and internal information. The following possibility seems
relatively straightforward. For every situation-restriction, there is an
absolutely equivalent role. Thus, equivalent to:
{($LOC, walk, John)}

we have:
($SITO, {(equal, $SITO, {($LOCO, walk, John)})}).
To make external predications on the $SIT0-role, we simply add facts
predicating things of $SIT0 to the situation-restriction in the role. So the
indeterminate for the content of John walks in Rome would be:

($SIT0, ((equal, $SIT0, {$LOC0, walk, John)})


($LOC1, in, $SIT0, Rome)}).

The only question now is to fit this sort of representation in with our
system of labeled indeterminates and our current composition operations.
We begin with verbs. Where before the content for walk was given by:

{(LOC $LOC0
REL walk
SUBJ $IND0
POL $POL)}.

Now it will be given by the equivalent:


($SIT0, {(HEADREL equal
ROLE-SIT $SITO
HEADSIT {(LOC $LOCO
REL walk
SUBJ $INDO
POL $POLO)})}).
What now has to happen is that composition operations must be rela-
tivized to the value of the HEADSIT function when applied to V and its
SITUATIONS AND PREPOSITIONS 373

projections. Call:

($SIT0, {(HEADREL equal


ROLE-SIT $SIT0
HEADSIT o-)})
a head-frame, where o- is a situation-restriction. By convention, the
content of a verb or its projections is always given by a head-frame with
the appropriate situation-restriction filling the HEADSIT slot. Where we
wrote:
VP---~ V NP
[[VP]] = QI([[NP]] OBJ [[V]])
we will now write:
[[VP]]:HEADSIT = QI([[NP]] OBJ [[V]]:HEADSIT).
By a semantic convention analogous to the syntactic head-feature con-
vention in GPSG (discussed, for example, in Gazdar and Pullum, 1982), a
verb projection will always inherit the content indeterminate of its head,
except in so far as it is explicitly stipulated to have changed. In the above
rule, the only thing changed was the value of HEADS1T. Rules that
change the HEADSIT of a verb projection will be just those that
constrain a sentence content by adding internal facts to the verb content.
This is in contrast to the adjunction rule we give now:
[in: P: {(LOC $LOC0
REL in
P-SUBJ $IND0
P-OBJ $IND1
POL $POLO)}]
Adjunct Rule
VP ~ VP PP
[[VP]]srr=M([[PP]] P-SUBJ [[VP]]stT ROLE-SIT).
Here, only [[VP]]s~T is changed. We write [[VP]]stv to mean the situa-
tion-restriction of the role associated with the VP. We can illustrate the
effect by returning to example (1):

[[walks]] = (SSIT0, {(HEADREL equal


ROLE-SIT $SITO
HEADSIT {(LOC SLOCO
REL walk
SUBJ $IND0
POL $POLO)})})
374 JEAN MARK GAWRON

[[walks]]s,w = {(HEADREL equal


ROLE-SIT $SIT0
HEADSIT {(LOC SLOE0
REL walk
SUBJ $IND0
POL $POL0)})}
[[in Rome]] = {(LOC $LOC1
REL in
P-SUBJ $IND1
P-OBJ $IND2)
(REL Rome
ARG $IND2)}
[[walks in Rome]]srv =
{(HEADREL equal
ROLE-SIT $SIT0
HEADSIT {(LOC $LOCO
REL walk
SUBJ $INDO
POL $POL0)})
(LOC $LOC1
REL in
P-SUBJ $SIT0
P-OBJ $IND2)
(REL Rome
ARG $IND2)}
Several points ought to be made about this semantics. First, note that, the
same sort of preposition entries are being used here as were used for the
semantics of subcategorized-for prepositions. In particular, the same
contents can belong to both subcategorized-for and adjunct prepositions.
Second, the above composition operation will work for noun adjuncts as
well. Thus, if we assume the following entry for box:
[box: N: ($IND0, {(REL box.
ROLE-SIT $IND0)})]
we could, borrowing the feature notation introduced in Gazdar and
Pullum (1982), re-state our verb adjunct-rule as an adjunction rule for
N-bar and V-bar:
XI[(Maj N V)]--> X1 P2
[[Xl]]siw=M([[e2]] P-SUBJ [[X1]]S~T ROLE-SIT).
SITUATIONS AND PREPOSITIONS 375

If this is the right generalization for the major categories, then neither
adjectives nor prepositions take right recursive modifiers. As far as I
know, this is correct.
This collapse of noun and verb rules raises the further question of how
generally useful elaborate headframes and HEADSIT-changing rules
are. Since nouns have ROLE-SIT, should they also have HEADSIT?
Nothing really stands in the way, in the case of those nouns that can
sensibly be said to have arguments. Thus we might imagine the following
entry for that hoary old example, destruction: 7

[[destruction]] = ($SIT0, {(HEADREL equal


ROLE-SIT $SIT0
HEADSIT {(LOC $LOC0
REL destroy
SUBJ $IND0
OBJ $IND 1
POL $POLO)})}).

Other nominalizations suggest that relations other than equal might serve
as the H E A D R E L , and situation variables might not be needed for the
ROLE-SIT. Thus, to designate the product of an action of criticism, we
might use the following entry for criticism:

[[criticism]] = ($SIT0, {(HEADREL product-of


ROLE-SIT $IND2
HEADSIT {(LOC $LOC0
REL criticize
SUBJ $IND0
OBJ $IND 1
POL $POLO)})}).

Similar headframes could be argued for for adjectives (although I am not


at all certain that adjectives take prepositional adjuncts, or any right
adjuncts). Interestingly enough, among the major categories, the hardest
case to make seems the case for prepositions. Prepositions neither seem
to take right adjuncts of their own nor to have a genuine variety of
argument-binding complements; analogously, the case for prepositions
as genuine heads in the X-bar system has always been the weakest.
Nevertheless, in the next section I will propose that one type of prep-
osition, at least, should have a verb-like entry with a situation-in-
determinate of its own.
376 JEAN MARK GAWRON

5.2. Benefactives

Let us return to the benefactives, whose chief peculiarity seems to be that


they do not allow commutation. T o capture the facts, I want to propose a
different rule of semantic combination. Until now, the H E A D SIT , the
internal content of the syntactic head, has always been the same as the
ROLE-SIT, the described situation. The facts about benefactives sug-
gest that this is not always the case; informally, in Bob made a sweater for
Sue the sweater-making situation is in the benefactive relation with Sue.
In Bob made a sweater for Sue for Mary, the benefactive situation with
Sue as beneficiary is in the benefactive relation with Mary. The general
picture, then, is that the VP content becomes the argument of the
benefactive relation, and that the content of the new VP formed is a
benefactive situation. Semantically, the described situation shifts from
being the H E A D S I T descended from the main verb to being the
" A D J S I T " descended from the benefactive preposition. Two in-
novations will be used in what follows: first, for will be given a lexical
entry more like a verb's, because it needs to offer a described situation
indeterminate, which can be the argument of further adjuncts. Second,
the composition rule given will drive a wedge between the syntactic
notion of a H E A D S I T , and the semantic notion of a ROLE-SIT:

[for: P: ($SIT1, {(ADJREL equal


DESC-ADJSIT $SIT0
ADJSIT {(LOC $LOC1
RE L for
P-SUBJ $SIT1
P-OBJ $IND 1)})})]

VP ~ VP PP

[[VP PP]] =
<[[PP]]ROLZ, U([tVP]]s~T
B(B({[VP]]ROLE P-SUBJ {[PP]]:ADJSIT)
ADJSIT
[{PP]]srr))).

The complexity of the semantics is due to the fact that this rule makes no
use of our semantic head convention; the entire content of the mother
VP is constructed from scratch. The new VP content has a new in-
determinate as its role indeterminate and a new situation-restriction as its
situation-restriction. Here, [[X]]ROLE is the complement to [[X]]sIT; one
SITUATIONS AND PREPOSITIONS 377

picks out the role indeterminate of a role, the other the situation
restriction. We will illustrate the working of the rule with an example:
sing for Joan for Miles.
[[sing]]ve
($SIT0, {(HEADREL equal
ROLE-SIT $SIT0
HEADSIT {(LOt $LOC0
REL sing
SUBJ $INDI)})})

[[for Joan]]pp
($SIT 1, {(ADJREL equal
DESC-ADJSIT $SIT1
ADJSIT {(LOC $LOC1
REL for
P-SUBJ $SIT4
P-OBJ $IND2)
(REL Joan
ARG $IND2)})})

[[sing for Joan]]vr,


($SIT 1, {(REL equal
ROLE-SIT $SIT0
HEADSIT {(LOC $LOC0
REL sing
SUBJ $IND1)})
(ADJREL equal
DESC-ADJSIT $SIT1
ADJSIT {(LOC SLOe1
REL for
P-SUBJ $SIT0
P-OBJ $IND2)
(REL Joan
ARG $IND2)})}).
Note that the labeled indeterminate for the full VP still has the in-
determinate descended from the main verb as the value of HEADSIT;
when the rule for combining the subject changes the HEADS1T by
altering its SUB J, the new subject will be fit into the right slot. At the
same time, the new VP has a new role indeterminate. Thus, a second
378 JEAN MARK GAWRON

benefactive will give:


[[for Miles]]pp
($SIT2, {(ADJREL equal
DESC-ADJSIT $SIT2
ADJSIT {(LOC $LOC2
REL for
P-SUBJ $SIT5
P-OBJ $IND1)
(REL Miles
ARG $IND3)})})
[[sing for Joan for Miles]]vp
($SIT2, {(HEADREL equal
ROLE-SIT $SIT0
HEADSIT {(LOC $LOC0
REL sing
SUBJ $IND1)})
(ADJREL equal
DESC-ADJSIT $SIT1
ADJSIT {(LOC $LOC1
REL for
P-SUBJ $SITO
P-OBJ $IND2)
(REL Joan
ARG $IND2}})
(ADJREL equal
DESC-ADJSIT $SIT2
ADJSIT
{(LOC $LOC2
REL for
P-SUBJ $SIT1
P-OBJ $IND3}
(REL Miles
ARG $IND3)})}).
The first sentence described a benefactive situation with Joan as
beneficiary, the second a benefactive situation with Miles as beneficiary.

6. CONCLUSION

In this paper I have argued for a unified lexical semantics of prepositions,


and for a combinatorial division among PP occurrences into three
groups: arguments, co-predicators, and adjuncts. The Argument Prin-
SITUATIONS AND PREPOSITIONS 379

ciple makes a strong claim about how the selection of argument prep-
ositions is semantically constrained, and gives that class of preposition
occurrences a natural characterization. It also grounds a motivated
account of what preposition.contents are, whether the prepositions o c c u r
as arguments or co-predicators. I have outlined a semantics which admits
something called bondedness relations into the interpretation of clauses,
remaining neutral on whether bondedness relations constrain only lin-
guistic connections or some significant ontological class such as actual
situations. In either case the interpretations of clauses are constrained by
requiring that the located facts in a situation-restriction enter into one or
more among a very small number of relations, for example constituent-
hood or efficient causality.
Two main sorts of evidence force us to admit V P adjuncts as distinct
from co-predicators and arguments: (1) genuine ambiguities that
cannot be chalked up to the co-predicator/argument distinction; (2) cases
where the content of a clause, C, is most simply explained by assuming
that a preposition has the same content we have seen elsewhere, and that
in C it is taking the entire described situation as its P-SUBJ. Particularly
important for evidence of type (2) is the analogy between verb and noun
adjuncts. NP's like the book in the room suggest a very straightforward
analysis on which in is a two-place relation between individuals. T h e
same relation then extends naturally to c o v e r the sentences with V P
adjuncts, where one of the individuals is a described situation. One of the
appealing aspects of the analysis proposed here has been the uniform
treatment of verb and noun adjuncts. T h e same preposition content can
be both a noun and a verb adjunct, and a single rule handles both types
of combination.
A central theoretical assumption inherited from T C & S O was that of a
level of representation of grammatical function. One problem with the
treatment proposed here is that in isolation labeled indeterminates that
make reference to grammatical function serve no particular theoretical
purpose beyond interpretation; and it is not clear how they interact with
other theoretical apparatus to do the work usually expected of gram-
matical functions. Yet it is not essential that interpretation happens
.t

relative to trees. One likely alternative is the sort of functional structure


used in LFG.
I do not propose to present an indeterminate-based semantics for an
L F G here; that is a task for another paper. But I conjecture that
composition rules stated on f-structures, or something very like f-
structures, could be a good deal simpler than those stated here, since
most of the requisite pointing structure has already been set up. 8
W h a t e v e r the virtues and faults of the particular semantics offered
380 JEAN MARK GAWRON

h e r e , t h e d e s i d e r a t a for a s a t i s f a c t o r y t r e a t m e n t of p r e p o s i t i o n s d e r i v e
f r o m s o u n d t h e o r e t i c a l goals. M y h i g h e s t t h e o r e t i c a l p r i o r i t y h a s b e e n n o t
so m u c h l e x i c a l s e m a n t i c p a r s i m o n y as a d e s c r i p t i v e c o m m i t m e n t to
l e x i c a l s e m a n t i c s t r u c t u r e : w h e r e o u r c o n c e p t i o n of s e m a n t i c i n f o r -
m a t i o n r e q u i r e s it, l e x i c a l e n t r i e s s h o u l d b e r e l a t e d in s e m a n t i c a l l y
p r i n c i p l e d ways. T h u s , a g u i d i n g p r i n c i p l e in c o n s t r u c t i n g this a c c o u n t is
t h a t t h e t h r e e - w a y d i s t i n c t i o n a m o n g P P o c c u r r e n c e s is n o t s i m p l y t h e
p r o j e c t i o n of a d i s t i n c t i o n a m o n g t h e s e m a n t i c s o r s e m a n t i c t y p e s of
p r e p o s i t i o n s . T h e s a m e p r e p o s i t i o n c o n t e n t c a n in p r i n c i p l e f u n c t i o n in
a n y of t h e t h r e e ways. T h e s e m a n t i c d i f f e r e n c e s t h a t r e s u l t all r e f l e c t
d i f f e r e n t w a y s of c o m b i n i n g v e r b c o n t e n t s w i t h ( p o t e n t i a l l y ) t h e s a m e
preposition content.

ACKNOWLEDGEMENTS

I d i d m u c h of t h e w o r k o n this p a p e r as L e v e r h u l m e F e l l o w at the
U n i v e r s i t y of E d i n b u r g h a n d l a t e r as a R e s e a r c h S c i e n t i s t at t h e C o u r a n t
I n s t i t u t e of M a t h e m a t i c a l S c i e n c e at N Y U . T h a n k s a r e d u e b o t h t h e s e
institutions for their nurturing, both intellectual and otherwise. Special
t h a n k s a r e d u e to t h e m e m b e r s of t h e s i t u a t i o n s e m a n t i c s s e m i n a r I
a t t e n d e d at E d i n b u r g h : I v a n B l a i r , E w a n K l e i n , H a n R e i c h g e l d , a n d
M a r y T a t e . I a m also g r a t e f u l f o r s o m e v e r y h e l p f u l c o m m e n t s f r o m an
anonymous referee.

NOTES

This paper is about semantics. The arguments made here go through whether or not both
co-predicators and argument PP's are syntactically subcategorized for. The important point
is that there is a kind of PP which needs to be semantically distinguished both from the
argument PP's, presumably the prototype of the subcategorized-for PP, and both types of
adjuncts discussed in section 5. I will discuss the question of whether co-predicators need to
be subcategorized-for in section 4.3.
2 I have rather cavalierly assumed here that there is a meaning of against uniquely
associated with impingement, distinguishing it from, for example, a locative against, (the
broom leaned against the wall), and an against of contention, (she fought against disarma-
ment), and doubtless others as well. The chief motivation for this move is simply the
compositional burden of the preposition. It introduces movement into the "scenes" (in
FiUmore's sense) associated with verbs that normally do not imply movement, or else imply it
of different participants than they do with against. This can be explained neither by a
meaningless against, nor one with an ultra-vague locative-impingement-contention Grund-
bedeutung.
3 Gawron (1985) discusses an important case of valence change which does require a lexical
redundancy rule to be correctly described: the "load the truck" examples of Partee (1971).
4 Non-located facts within the situation-restriction tell us things about the participants in the
located facts; so we might have another bonding relation describes-participant-of, which
relates unlocated facts to located facts.
SITUATIONS AND PREPOSITIONS 381

5 It's only fair to point that this conjecture seems to be at odds with the view of situations
emerging in some of Barwise's recent work relating situations to sets (Barwise, 1985), in
which situations are, roughly speaking, interpreted as any chunk of reality that can be
apprehended as a whole. In its strongest form this view adopts the axiom that any set of facts
determines a situation. Obviously, my way of talking, which limits actual situations to those
containing "bonded" facts, is incompatible. But reconciliation might still be possible if I
would give up using the term "actual situation" for whatever it is clause situation-types can
be instantiated by. At the moment, I haven't got anything better.
6 The same kind of treatment could also be extended to handle the sort of missing argument
facts we noted in section 1 for verbs like rub. What we would have would be:
[rub: V: {(REL rub
SUBJ $IND0
OBJ $IND1
OBL ($IND2, {(REL hand-of
ARG1 $IND2
ARG2 $IND0
POL 1)})
POL $POL0)}].
Here, there is a role in which the added fact contains an indefinite, rather than a definite
description. The "instrument" that does the rubbing is, by default, a hand belonging to the
entity denoted by the SUBJECT.
7 But see Thomason (1985) for a compelling barrage of arguments against this sort of
approach to nominalizafion. It is not clear to me, however, that any of Thomason's
arguments apply to a mass noun like destruction.
8 For a full development of a model-theoretic semantics for an LFG, see Halvorsen (1983).
The semantics for an LFG, see Halvorsen (1983). The semantics there is presented in a
Montagovian framework.

REFERENCES

Anderson, John: 1970, "On Case Grammar', Croom Helm: London.


Barwise, Jon and John Perry: 1983, 'Situations and Attitudes', MIT Press, Cambridge,
Mass.
Barwise, Jon: 1984, '5 Lectures on Situation Semantics', Center for the Study of Language
and Information, Stanford typescript.
Barwise, Jon: 1985, 'Situations, Sets, and the Axiom of Foundation', Center for the Study
of Language and Information, Stanford, typescript.
Bresnan, Joan: 1983, (ed.), 'The Mental Representation of Grammatical Relations', MIT
Press, Cambridge, Mass., 874 pages.
Cooper, Robin: 1984a, 'A Situation Semantics for GB', Center for the Study of Language
and Information, Stanford, lecture notes.
Cooper, Robin: 1984b, 'Conjunction in Situation Semantics', in Proceedings of the 20th
Annual Chicago Linguistics Society Conference.
Dowry, David: 1978, 'Lexically Governed Transfomations as Lexical Rules in a Montague
Grammar', Linguistic Inquiry 9, 393-426.
Dowty, David: 1982, 'Grammatical Relation~ and Relational Grammar', in Pauline Jacob-
son and Geoffrey K. Pullum (eds.), The Nature of Syntactic Representation, Reidel,
Dordrecht, pp, 79-130.
Gawron, Jean Mark: 1983, 'Prepositions and the Semantics of Complementation', Uni-
versity of California at Berkeley, doctoral dissertation.
Gawron, Jean Mark: 1985, ' A Parsimonious Account of Prepositions and CAUSE', CLS
21: Parasession on Causatives and Agentivity.
382 JEAN MARK GAWRON

Gawron, Jean Mark: 1986, 'Types, Contents and Semantic Objects', Linguistics and
Philosophy 9.
Gazdar, G.: 1983, 'Phrase Structure Grammars', in G. K. Pullum and P. Jacobson (eds.),
The Nature of Syntactic Representation, Reidel, Dordrecht.
Gazdar, G. and G. Pullum: 1982, 'Generalized Phrase Structure Grammar: A Theoretical
Synopsis', Indiana Unuversity Linguistics Club, August.
Halliday, M. A. K.: 1967, 'Notes on Transitivity and Theme in English', Part I, Journal of
Linguistics 3, 37-81.
Halvorsen, Per-Kristian: 1983, 'Semantics for Lexical-Functional Grammar', Linguistic
Inquiry 14, 567-615.
Klein, Ewan and Ivan Sag: 1985, 'Type-Driven Translation', Linguistics and Philosophy 8.
Montague, Richard: 1970, 'The Proper Treatment of Quantification in English', Yale
University Press, New Haven.
Partee, Barbara: 1964, Subjects and Objects. MIT dissertation, Cambridge, Mass.
Parsons, Terry: 1984, 'Underlying Events in the Logical Analysis of English', Paper given
at the Rutgers Conference on the Philosophy of Donald Davidson.
Simpson, Jane: 1983, 'Discontinuous Verbs and the Interaction of Morphology and Syntax',
in Proceedings of the Second West Coast Conference on Formal Linguistics, University
of Southern California, Los Angeles.
Thomason, Richmond H.: 1985, 'Derived and Gerundive Nominals', Linguistics and
Philosophy 8, 73-80.

Center for the Study of Language and Information


Ventura Hall
Stanford University
Stanford CA 94305
U.S.A.

You might also like