206 Louisa Sadler and Andrew Spencer

10 Morphology and
Argument Structure
LOUISA SADLER AND ANDREW SPENCER

1 Introduction

In English we can say (1a) or (1b):

(1) (a) The peasants loaded hay onto the wagon.

(b) The peasants loaded the wagon with hay.

However, although we can say (2a), we can’t say (2b):

(2) (a) The peasants poured water into the tank.

(b) *The peasants poured the tank with water.

And while we can say (3b), we can’t say (3a):

(3) (a) *The peasants filled water into the tank.

(b) The peasants filled the tank with water.

Further, we can say (4a) or (4b):

(4) (a) Ira broke the vase.

(b) The vase broke.

Yet, although we can say (5a), we can’t say (5b):

(5) (a) Ira cut the bread.

(b) *The bread cut.

In addition to the contrasts in examples (4) and (5), we have those exhibited
by examples such as (6) and (7):
 Morphology and Argument Structure 207

(6) (a) Children easily break such vases.

(b) Such vases are easily broken (by children).
(c) Such vases break easily (*by children).

(7) (a) An ordinary knife will easily cut such bread.

(b) Such bread is cut easily (with an ordinary knife).
(c) Such bread cuts easily (with an ordinary knife).

Examples such as these raise the question of how participants which are
entailed by the lexical meaning of predicates are made explicit in the morpho-
syntactic representation, and whether and under what conditions they may
remain implicit: that is, issues of valency. In addition, they raise the question
of alternations: that is, where two morphologically related (or even identical)
predicates differ in their lexical semantics and in the way participants are
realized in the morphosyntax and, in particular, in morphology. This facet
of the morphology–syntax interface has come to be referred to as ‘argument
structure’. This term means different things to different authors, and one of
our aims will be to make explicit a number of distinctions between some of
the different types of realization and different types of alternation that have
fallen under this term.
We begin our discussion with a distinction between two sorts of operation
which affect valency. Section 3 goes on to examine the question of linking, the
way that grammars describe the associations between semantic and/or argu-
ment structure representations and surface grammatical relations. We describe
how this is accomplished in two major frameworks: Lexical Functional Gram-
mar and Principles and Parameters Theory. We also discuss the vexed question
of unaccusativity and its representation, and conclude the section with brief
mention of other frameworks. In the fourth section we provide three case-
studies illustrating the architecture we propose: passives/middles in English,
reflexives/reciprocals in Bantu, and causatives in Japanese, then close with a
brief discussion of noun incorporation and synthetic compounding.
There are several important topics we cannot address for reasons of space.
First, we are restricting our attention to verbs, and ignoring the argument
structure properties of adjectives, nouns and prepositions. More specifically,
we will side-step a number of difficult but important questions about par-
ticiples and nominalizations, especially the inheritance of argument structure
in deverbal nominalizations (cf. Grimshaw 1990: chs 3, 4). We are also unable
to delve into the question of light verb constructions (and the related problem
of serial verbs). Finally, there are interesting questions surrounding argu-
ment structure and derivational morphology (e.g. the argument structure of
deverbal adjectives such as readable from read) which we leave untouched.
However, we hope to have provided sufficient foundation in this chapter for
the interested reader to investigate the primary literature in those areas with
some confidence.
208 Louisa Sadler and Andrew Spencer

2 Two types of operations

Between them, lexical semantics and morphosyntactic theory must enable us
to relate, at some level, the two uses of load in (1) and break in (4), and the
passive voice to the active. There is by now a broad consensus that lexical
semantics plays a large role in determining such morphosyntactic realizations,
but disagreement as to whether it can all be reduced to semantics. For those
who believe that semantics is insufficient on its own, it is necessary to stipu-
late some other level of representation to account for these phenomena. In
addition, we would wish to see evidence that such extra structure is empiric-
ally or conceptually necessary.
Any verb will have some number of (optional or obligatory) syntactic
dependents, and the lexicon and grammar of a language must therefore
include information about these valency requirements. This information may
be expressed in a variety of ways, appealing directly to grammatical functions
such as subject and object, as in Lexical Functional Grammar (Bresnan 1996)
or Relational Grammar (see Blake 1990), or to syntactic configurations, as in
Principles and Parameters Theory (Chomsky 1981), or to some combination
of grammatical functions and category labels, as in Head-Driven Phrase Struc-
ture Grammar (HPSG, Pollard and Sag 1994). In addition, there must be some
representation of the linguistic aspects of word meaning: that is, a semantic
level of representation characterizing the necessary properties of the semantic
arguments of predicates. A number of different proposals have been put for-
ward in the literature concerning the nature, structure and vocabulary of the
level of lexico-semantic representation, summarized by Levin and Rappaport
Hovav, Morphology and Lexical Semantics.
Given this background, two fundamentally important, and clearly related,
issues arise:

(i) To what extent is syntactic valency idiosyncratic, and to what extent

can it be said to be predictable from the lexico-semantic representa-
tions associated with individual predicates?
(ii) How can the relationships between different ‘uses’ of the same word
form and between related word forms be captured/predicted in a
non-redundant manner?

Our first response to these questions will be to follow a number of authors in

drawing a distinction between two sorts of operation ‘in’ the lexico-semantic/
syntax interface (cf. Levin and Rappaport Hovav, Morphology and Lexical
Semantics, for a succinct summary of this claimed distinction). The first,
‘meaning-changing’ operation alters the semantic content of predicates, and we
refer to such operations as ‘morpholexical operations’. The second, ‘meaning-
preserving’ operation alters the syntactic manifestation of a given semantic rep-
resentation, particularly the way that it is mapped on to grammatical relations.
 Morphology and Argument Structure 209

We will refer to these as ‘morphosyntactic operations’.1 In a sense, this division

corresponds to the traditional distinction between derivation (lexeme-creating)
and inflection (creation of distinct forms of a given lexeme), though it may not
always be helpful to push this analogy. The distinction is reminiscent of that
found in morphophonology, in which phrasal phonology becomes morpho-
logized by giving rise to morphophonological alternations and ultimately
suppletive allomorphy.

2.1 Morpholexical operations

To makes things more concrete, consider (8):

(8) Resultative construction

(a) The blacksmith hammered the metal.
(b) The blacksmith hammered the metal flat.

This example illustrates an operation which is appropriate for verbs in cer-

tain semantic classes (roughly, the meaning of the verb must be compatible
with an eventual change of state), and adds a semantic argument to a predic-
ate. This argument expresses the resultant state, flatness, of the object, metal.
Evidently, the resultative construction increases the syntactic valency of the
predicate – in (8b), hammer in the resultative complex hammer flat has a sur-
face syntactic valency of three. The claim that result predication is a semantic
or morpholexical operation is based on the assumption that the syntactically
bivalent predicate illustrated in (8a) expresses a relation between just two
semantic arguments, without entailing an end result. That is, (8b) crucially
means that the blacksmith flattened the metal by means of hammering activ-
ity. Example (9) is a particularly clear case in which the main verb is atelic:

(9) They drank the teapot dry.

Since one cannot drink a teapot, (9) must be interpreted as ‘they rendered the
teapot dry by drinking (from it)’.

2.2 Morphosyntactic operations

Two constructions in English, dative shift and passive, are often taken to be
examples of morphosyntactic operations. These are illustrated in (10) and (11):

(10) Dative shift

(a) Tom gave a bone to his dog.
(b) Tom gave his dog a bone.
210 Louisa Sadler and Andrew Spencer

(11) Passive
(a) Tom broke the vase.
(b) The vase was broken (by Tom).

Each operation brings about an alteration in the morphosyntactic manifesta-

tion of the semantic dependents of a predicate, but they do not alter the basic
semantics of the predicate itself (though see section 3.3 for discussion of dis-
senting views).
The first of these alternations, dative shift, appears to involve a simple altern-
ation between two different syntactic manifestations of the same semantic
roles. In (10a) the direct object realizes the Theme role, and in (10b) it realizes
the Recipient. Now, other things being equal, we might expect morphosyntactic
operations to be unconstrained by the semantics of the predicate. This is largely
true of the passive in English, for instance. On the other hand, dative shift is
restricted in applicability to verbs of transfer respecting rather subtle semantic
constraints (see Pinker 1989 for detailed discussion and for examination of
some the consequences of this for learnability).
Turning now to the second alternation, it is common to treat passivization
as a morphosyntactic operation involving the suppression of the external argu-
ment, or most prominent argument. If passive is a morphosyntactic operation,
we would expect that the semantics of the predicate would remain constant
across the voice alternation. A consequence of this in English and many other
languages is that the Agent is available semantically, and enjoys a certain pres-
ence syntactically without necessarily being syntactically expressed. In many
languages, this suppressed argument may be expressed as an oblique or an
adjunct of some sort, as in the English optional by phrase illustrated in (11b).
If the passivization process is simply one of syntactic suppression (as opposed
to downright deletion), we would expect the first argument to be available for
processes which are semantically rather than syntactically governed, and indeed
this seems to be the case. This is discussed in more detail in section 4.1.
A distinction something like that between our morpholexical and morpho-
syntactic operations is widely assumed, and it is very often taken to motivate
a further (third) notion of dependent and a third level of information. This con-
ceptual level is often known as ‘argument structure’ or ‘predicate–argument
structure’ (pas), and it occupies the interface between the two sorts of opera-
tions. The morpholexical operations alter (add, delete, identify) semantic com-
ponents of predicates and create new semantic representations, lcss. Each of
these is associated with its own argument structure, pas. The morphosyn-
tactic operations intervene between pas and syntactic structures, resulting in a
multiplicity of syntactic realizations for one and the same argument structure.
Argument structure is essentially a syntactic representation: in fact, it is the
syntactic reflex of certain semantic properties. These properties determine the
arity (adicity) of the predicate and the relative prominence of the dependence.
Both these properties will determine the way the arguments project into the
syntax. In a two-place predicate, if no morphosyntactic operations intervene,
 Morphology and Argument Structure 211

the most prominent argument (the more ‘agent-like’) will map to the subject
position, and the less prominent (the more ‘theme-like’) will map to the object
position. In many accounts, the parallel between pas and constituent syntactic
structure proper is increased by distinguishing a special argument position,
that of external argument (E. Williams 1980) or most prominent argument
(Grimshaw 1990), which always surfaces as the subject. This is the argument
of which the entire VP is predicated (hence, it is in a sense external to the verb
as such). In a two-place predicate, the remaining argument is often called
an internal argument, while a three-place predicate may distinguish a direct
internal argument, generally associated with the direct object, and an indirect
internal argument, associated with an indirect object. For many purposes it is
convenient to assume a further pas position, denoting events (cf. Higginbotham
1985). This is a position which can be bound by tense operators in the syntax,
and to which certain sorts of adverbial may have access.
We give a simplified sketch of this architecture in an essentially theory-
neutral fashion. We take examples (11) for illustration, ignoring various com-
plexities such as the precise surface representation of tense elements, participles
and so on:

(12) (a) Active form: Tom broke the vase.

[[x ACT] CAUSE [BECOME [BROKEN(y)]]] lcs
break: <x <y>> pas
Tom broke the vase. syntax
SUBJECT OBJECT
(b) Passive form: The vase was broken by Tom.
[[x ACT] CAUSE [BECOME [BROKEN(y)]]] lcs
broken: <(x) <y>> pas
The vase was broken (by Tom). syntax
SUBJECT OBLIQUE

In the pas representations, the external argument is leftmost, the direct internal
argument is written in its own set of angle brackets, . . . <y>. . . . In the passive
representation, the suppression of the external argument is notated by means
of parentheses <(x) . . . >. The reader must expect to see a number of notational
variants on this theme. If we wish to include an event position, we would
write the pass as break <e <x <y>>> and broken <e <(x) <y>>> (assuming
passive participles are ‘eventive’ in the appropriate sense).
In this section, we have motivated and illustrated Levin and Rappaport’s
(Morphology and Lexical Semantics) distinction between two conceptu-
ally different sorts of relations between lexemes or word forms. The morpho-
syntactic operations are meaning-preserving, but alter the syntactic realization
212 Louisa Sadler and Andrew Spencer

of the predicate. The morpholexical operations are meaning-altering, and

add, delete or identify certain components of meaning. They therefore create
slightly different lexemes with syntactic realizations different from those of
the base predicate. Either type of operation may be, but is not necessarily,
morphologically mediated. Morphosyntactic operations regularly arise when
fully fledged syntactic processes become morphologized (as when a verb
becomes an affix). Not infrequently a morphosyntactic operation becomes a
morpholexical operation in historical change (lexicalization). As a result of
this, one and the same piece of morphology may realize a morphosyntactic
operation in one language/dialect and a morpholexical operation in a closely
related language/dialect.
In most of the rest of this paper, we will be turning our attention to
morphologically mediated operations of these two sorts. Before doing so,
however, we will turn in the next section to the question of how the surface
syntactic form (possibly through the mediation of pas) is related to the lcs
representation.

3 Linking

3.1 Standard cases in LFG and PPT/GB

In this section we provide a brief sketch of linking theories, within two frame-
works, Lexical Mapping Theory (LMT), within Lexical Functional Grammar
(LFG) (Bresnan and Kanerva 1989, Bresnan and Moshi 1990), and the Principles
and Parameters Theory (PPT, also called Government Binding Theory, GB) of
Chomsky 1981.
The lexical semantic representation in LMT uses a set of thematic roles,
including Agent, Patient, Theme, Experiencer, Beneficiary, Goal, Instrument,
Location.2 For break, this lexical representation, known as the argument struc-
ture in LFG, would then be as in (13):

(13) break: <Agent, Patient>

The roles are ordered by the hierarchy given in (14) (Bresnan and Kanerva
1989):

(14) Agent < Benefactive < Goal/Experiencer < Instrumental < Patient/
Theme < Locative . . .

The argument structure given in (13) is then mapped to a set of sub-

categorized grammatical functions, which are syntactic primitives in LFG (SUBJ,
OBJ, etc.). These are decomposed into binary distinctive features as shown
in (15):
 Morphology and Argument Structure 213

(15) SUBJ [–r, –o]

OBJ [–r, +o]
OBJ2 [+r, +o]
OBL [+r, –o]

Subjects and ordinary objects can express any thematic role, so they are
unrestricted, [–r]. Secondary objects and obliques are associated with some
specific thematic role, and are hence restricted [+r].3 Genuine objects are able
to complement transitive verbs, but not, say, nouns or adjectives, and these
are marked [+o]. Subjects and obliques lack this property. This leads to the
classification given in (15).
Mapping is achieved as follows: first, a set of Intrinsic Classification (IC)
principles associate thematic roles with particularly syntactic feature values,
as exemplified in (16):

(16) Agent Patient/Theme Locative

[–o] [–r] [–o]

The association of Patient/Theme with [–r] in (16) reflects the fact that Patients
or Themes alternate between subject and object functions. Next, a set of Default
Rules (DR) apply, filling in redundant values in accordance with (17):4

(17) (a) 0ˆ (b) 0

[–r] [+r]

The symbol 0ˆ stands for the highest thematic role. The effects of these can be
seen in (18):

(18) break <Agent, Patient>

–o –r IC
–r DR

Notice that the system is monotonic, so it would be impossible for DR (17b)

to override the intrinsic specification [–r] on the Patient role. Thus far, (18)
guarantees that the Agent is mapped to the SUBJ position. To complete the
derivation, we appeal to a principle of Function-Argument Biuniqueness which
states that every lexical argument position must be uniquely associated with
a grammatical function (and vice versa). This limits the specification for [o]
which may be given to the Patient in (18). It cannot be [–o], as that would
mean that the clause would have two subjects, thus violating bi-uniqueness.
Hence, the Patient must be marked [+o], and therefore maps to OBJ.
214 Louisa Sadler and Andrew Spencer

Of course, we may find that the lexical form in (18) is altered by a valency-
affecting (for us, morphosyntactic) operation such as Passive. In LFG passive
is an operation which suppresses the highest thematic role:

(19) Passive 0ˆ

The morphosyntactic operations apply before the Default Rules, a consequence

of the Elsewhere Condition, under which the more specific of two rules in com-
petition applies in preference to the more general (see Stump, Inflection, for
general discussion of the Elsewhere Condition). The derivation for the passive
of break (as in The vase was broken) would therefore be (20):

(20) break <Agent, Patient>

–o –r IC
∅ Passive

The Default Rule cannot apply (even vacuously). We now appeal to a further
well-formedness condition which states that every clause must have exactly
one subject. Thus, the Patient is mapped to the SUBJ function, as required.
These represent very simple cases, of course (the minimal requirement on a
successful linking theory). Rather complex problems emerge when we look at
more tricky types of predicate, especially those with double objects in some
languages (cf. Alsina and Mchombo 1993, Bresnan and Moshi 1990) and, per-
haps most notoriously, psychological predicates. The problem with the latter
is that there are languages (such as English) which have two types of predicate
with roughly the same meaning, as illustrated in (21):

(21) (a) Tom fears enclosed spaces.

(b) Enclosed spaces frighten Tom.

Suppose we say that these are exactly synonymous, so that the thematic rela-
tions borne by Tom and enclosed spaces are identical in each case: how do we
then construct a linking theory which will obtain both mappings? This conun-
drum has been the subject of much recent debate (see e.g. Belletti and Rizzi
1988, Dowty 1991, Grimshaw 1990, Pesetsky 1995).
We now turn briefly to the second model, in which grammatical relations
are not primitive labels or feature bundles, but are positions in a constituent
structure. The pas representation for break will be that of (22):

(22) break: <x <y>>

This representation conveys information about arity, just as representation (13)

did, but it also contains limited information about prominence. In addition
 Morphology and Argument Structure 215

to indicating that there are two syntactically realizable arguments, (22) also
specifies the x argument as the external argument and y as the internal
argument. This is obtained by means of rules mapping the lcs to the pas.
The semantic argument of an ACT (or CAUSE) predicate is more prominent
than, say, the semantic argument of a stative predicate such as BROKEN (or
BECOME[BROKEN]). (This is another way of saying that Agents are more
prominent, or higher on a hierarchy, than Themes.) Most of what has been
said about the use of a thematic hierarchy can be automatically translated into
this framework (including the problems with psychological predicates!), so we
will not rehearse this (see Jackendoff 1990 for detailed discussion).
The next question is how (22) is related to a syntactic representation. In the
standard Principles and Parameters Theory of Chomsky (1981) there is a D-
structure representation, which reflects argument structure very directly, and
this is mapped into S-structure by the general rule of Move-α (constrained
in various ways). We therefore map the pas (22) for sentence (11a) into the
D-structure shown in (23):

(23) S

NP VP

V NP

Tom break vase

Representation (23) corresponds very closely to the S-structure representation.

In PPT, nominals have to be licensed in syntactic form by receiving abstract
Case. The subject position receives Nominative Case from I, the position asso-
ciated with tense marking and subject agreement, while the direct object gets
Accusative Case from the verb. These conditions are satisfied in (26), the S-
structure derived from (23):

(24) IP

NP I′

I VP

V NP

Tom broke e vase

216 Louisa Sadler and Andrew Spencer

The passive operation is viewed as suppression of the external argument,5

notated by putting it in parentheses. As in all other theories, there is then a
rider permitting this argument to be expressed as an oblique of some kind in
many languages. The passivized verb form, which in English is a periphrastic
construction involving a participle and auxiliary verb be or get, is inserted into
the D-structure and the internal argument, vase, is linked to the object posi-
tion (technically, the nominal complement governed by the V head of the VP).
However, the external argument has been suppressed, and cannot therefore be
linked to the subject position. The best it can hope for is to be realized as an
optional PP adjunct, as in (25):

(25) S

NP VP

V NP PP

be broken vase (by Tom)

There is a general principle known as Burzio’s Generalization, which states

that a predicate lacking an external argument cannot assign Accusative Case.
This means that passive participles cannot assign Accusative Case and hence
cannot license their own objects. This means that vase in (25) cannot remain
where it is. It is therefore moved to the only landing site where it will receive
a legitimate Case: namely, the (currently unoccupied) subject position. Hence,
we obtain (26):

(26) IP

NP I′

I VP

V NP PP

vase j was brokeni ei ej (by Tom)

PPT, like LMT, is governed by ancillary assumptions: in particular, that all

clauses must have a subject (the Extended Projection Principle) and all lexical
 Morphology and Argument Structure 217

argument positions must map on to a structurally defined argument position

in the syntax, and vice versa (the Theta Criterion).
It will be seen that although the details of the architecture differ, and a num-
ber of theoretical positions contrast starkly (e.g. whether grammatical relations
are primitive or not), the two models manipulate very much the same ideas,
especially with respect to argument-structure representations.

3.2 The mapping of intransitive verbs

As is well known, many languages distinguish morphosyntactically between

two types of intransitive verb. In the first type, the unergative, the subject
fulfils an active semantic role (such as the traditional Agent), while in the
second, the unaccusative, the subject is more passive semantically, and corres-
ponds to a Theme or Patient. The important point here is that there are some-
times morphosyntactic processes which treat the subject of an unaccusative
and the direct object of a transitive verb as a single class, distinct from the
subject of an unergative verb.6 An example of this is found in the English
resultative construction (seen in (8) and (9) above). We can say They hammered
the metal flat, and also The river froze solid, but we can’t say She ran tired with
a resultative meaning ‘she tired herself by running’. This corresponds to the
fact that freeze has a Theme subject and is thus unaccusative, while run has an
Agent subject and is thus unergative. In other languages the distinction is said
to manifest itself in terms of the auxiliaries selected for certain tense/aspect
forms (Italian, French, Dutch, Danish), whether an impersonal passive is
permitted (Dutch, German and many other languages), whether a genitive
subject is possible under negation (Russian), which argument certain quanti-
ficational or aspectual prefixes apply to (Slavic generally), or whether the
argument can undergo noun incorporation (Mohawk, Chukchee and many
other languages).
In LMT this distinction is coded in terms of the specification of grammatical
function features. The Intrinsic Classification presented in (16) above will give
us the representations (27) for the unergative and unaccusative verbs run,
freeze:

(27) (a) run<Agent> (b) arrive<Theme>

–o –r
---------------- ---------------------
SUBJ SUBJ

We can now make those processes diagnostic of unaccusatives sensitive to the

presence of a [–r] argument.
218 Louisa Sadler and Andrew Spencer

In the PPT/GB framework an unaccusative predicate can be character-

ized in two ways. First, we could say that it is a one-place predicate with no
external argument; second, we could say that it is a one-place predicate whose
argument occupies a D-structure object position. In principle, one might even
expect to find different processes sensitive to these different characterizations,
though such evidence is hard to come by.
The pas representations for run and freeze are shown in (28), and the corres-
ponding D-structures are given in (29):

(28) (a) run<x> (b) freeze<<x>>

Here we follow Grimshaw (1990) in representing the internal argument in

double angled brackets.

(29) (a) run – unergative (b) freeze – unaccusative

S S

NP VP NP VP

V V NP

x run freeze y

By Burzio’s Generalization, the unaccusative predicate cannot assign Accusat-

ive Case (because it has no external argument), so the ‘y’ argument has to
move to subject position to receive Nominative Case.

3.3 Alternative approaches to argument structure

This chapter is investigating the idea that argument structure is an independ-

ently definable level of representation, and for that reason we have not delved
into those approaches under which alternations are the result of head move-
ment in the syntax (Baker 1988a, D. G. Miller 1993, Ackema 1995; for review see
Carstairs-McCarthy 1992, Spencer 1991). It is not clear how this relates to pas/
lcs representations, especially in a theory such as Baker’s which countenances
an autonomous morphology module. It is worth mentioning, though, that most
approaches to morphological causatives treat them at some level as a kind of
 Morphology and Argument Structure 219

complex predicate (see section 4.3), which is reminiscent of Baker’s view that
they have the morphosyntax of V–V compounds.
An interesting offshoot of Baker’s work which links it to the notion of lcs
is that of Hale and Keyser (1992, 1993). They propose that argument structure
be described in terms of lexical argument structures or lexical relational struc-
tures (LRS), essentially a sparse form of lcs built up out of binary-branching
syntactic structures and obeying syntactic principles from PPT such as the
Empty Category Principle. The idea is that the argument positions are repres-
ented by NP or PP complements, and the differences in argument structure
associated with various verb types are coded as occurrences of (generally
empty) V, A or sometimes P slots in the LRS. Thus, a causative verb is one
which has a lexical VP structure headed by a V slot (corresponding roughly
to a causative predicate in other frameworks).
Hale and Keyser consider location verbs such as shelve, in which the incorp-
orated noun corresponds to a locative prepositional phrase:

(30) She shelved the books.

Cf.
(31) She put the books on the shelf.

They argue that the LRS for shelve is (32), akin to the syntactic structure cor-
responding to (31):

(32) V′

V VP

NP V′

(the books) V PP

P NP

shelf
220 Louisa Sadler and Andrew Spencer

The final verb form is derived by multiple application of head movement of

shelf, successively through P, V and V:

(33) V′

V VP

P V NP V′

N P (the books) V PP

shelve e P NP

e N

Hale and Keyser argue that similar derivations account for cases such as
saddle (Harriet saddled the horse; cf. Harriet provided the horse with a saddle). They
claim that syntactic principles explain why this is not possible when the con-
verted noun corresponds to an indirect object. Thus, although we can say
Harriet donated a fortune to the church, we can’t say *Harriet churched a fortune.
They also analyse causative/inchoative pairs in terms of movement of nouns
or adjectives through empty V positions (The gravy thinned, The cook thinned the
gravy). In this theory there appears to be no room for a distinction between
morpholexical and morphosyntactic operations. Most of the alternations Hale
and Keyser discuss are lexically restricted, and are clearly morpholexical (e.g.
conversion from shelf to shelve).
One very interesting alternative to the two-level architecture presented here
is the theory of argument-structure alternations developed by Goldberg (1995)
within the framework of Construction Grammar. For Goldberg, an alternation
such as dative shift or the locative alternation is the result of fusing the lexical
structure of an individual lexical item with a more general frame, the dative
shift construction. A construction modulates the original verb entry, and is even
capable of adding extra arguments or underlying predicates such as CAUSE.
The framework is a particularly attractive way of handling alternations in which
a semantic component is added, such as causatives and resultatives. It is less
obvious how it handles cases of argument identification (e.g. reflexivization)
or suppression (passives, antipassives). There is little scope in such a frame-
work for the lcs/pas distinction, because all alternations are regarded as on
a par.
 Morphology and Argument Structure 221

4 Morpholexical and morphosyntactic operations

Having discussed background notions and general theoretical approaches to
argument structure, we will now illustrate our overall typology of operations
with specific examples. We will discuss three sets of alternations which are
each similar except that one can be regarded as a morpholexical relationship
while the other can be viewed as a morphosyntactic operation. In the first
morphosyntactic operation, passive, an argument position is suppressed. A
similar morpholexical operation (which is not expressed morphologically) is
English middle formation. In Chichewa reflexivization we see a process in
which one argument position is linked with another referentially. The mor-
phological reciprocal turns out to be morpholexical, however. In productive
causatives an argument position is added. This has a drastic effect on the lcs,
but in regular cases this is a predictable side-effect of the morphosyntactic
change brought about by fusing the argument structure of the base verb with
that of the causative. Finally, we look briefly at noun incorporation and at
some of the theoretical responses to this phenomenon.

4.1 Passives and middles

In both the passive and the middle alternations in English an argument is
lost and fails to be projected in the syntax. Consider (34, 35) (cf. Ackema and
Schoorlemmer 1995: 175):

(34) (a) Tom painted the walls.

(b) The walls were painted (by Tom). (Passive)

(35) These kinds of walls paint easily. (Middle)

The suppressed external argument of the passive is still syntactically ‘active’

to some extent. It can appear as a by phrase more or less irrespective of the
semantics of the verb (34b), and it can license agent-oriented adverbials (36),
and can control the null subject of purposive clauses (37):

(36) The walls were painted on purpose.

(37) The walls were painted to protect them against the rain.

As is well known, the middle construction imposes semantic constraints: the

resulting sentence is interpreted as a stative, and stylistically is preferred, there-
fore, with a generic subject. In addition, the construction is generally difficult
or impossible without adverbial support (in the form of adjuncts referring to
ease or difficulty).
222 Louisa Sadler and Andrew Spencer

An important contrast between passives and middles is that the lost subject
is not syntactically available, even where the above felicity conditions are met:

(38) (a) *These kinds of walls paint easily by professional painters.

(b) *These kinds of walls paint easily on purpose.
(c) *These kinds of walls paint easily to protect them against the rain.

English passives are possible with a great variety of verb construction types,
including raising verbs (39), double object constructions (40) and with idiom
chunks, as in (41):

(39) (a) The coach expected Tom to win.

(b) Tom was expected to win (by the coach).

(40) (a) Tom gave Dick a book.

(b) Dick was given a book (by Tom).

(41) (a) The used-car salesman took advantage of Tom.

(b) Tom was taken advantage of by the used-car salesman.
(c) Advantage was taken of Tom by the used-car salesman.

Middles fail in such cases:

(42) Such committed athletes readily expect to win.

(This is ungrammatical on a reading synonymous with are readily expected to

win.)

(43) (a) *Such well-educated children give books easily.

(Also *Such books give easily to young children.)

(44) (a) *Tom takes advantage of easily.

(b) *Advantage takes easily of Tom.

We will follow Ackema and Schoorlemmer (1995) in assuming that middles

are formed by a ‘pre-syntactic’ – for us a morpholexical – process, whereas
passives are formed by a morphosyntactic process. The middle, then, has a
single syntactically projectable argument (an external argument, according to
Ackema and Schoorlemmer, hence the middle is an unergative form) and no
syntactically available implicit argument. This lone argument corresponds to
a patient argument in semantic structure (and, of course, corresponds to the
patient argument which is associated with the direct internal argument of
the ordinary active form of the verb). The passive, however, has an internal
argument which is linked to subject position in the syntax, and a suppressed,
 Morphology and Argument Structure 223

but implicit, argument corresponding to the external argument of the active

form. The existence of such an implicit argument is, for us, a sufficient con-
dition for a morphosyntactic valency-reducing operation.7
Given the logic of the overall architecture of grammar which we have taken
as our descriptive starting-point, the active and passive alternants of a verb
are forms of one and the same lexeme, while the middle form is effectively
a closely related lexeme, with slightly different semantics. The difference in
syntactic projection in the middle follows from the difference. The fact that
there are fairly strict lexical and semantic restrictions on the formation of
middles should come as no surprise. However, this is not to say that all verb
forms in a given language which behave like the English middle with respect
to argument structure necessarily exhibit such restrictions.
In Bantu languages – for instance, Swahili – verbs accept a wide range
of suffixes to form new voices or new lexemes. Swahili has a form which is
generally referred to as a passive, formed by a suffix -w-. Hence, from pika
‘cook’ we have pikwa ‘ be cooked’, funga ‘close’, fungwa ‘ be closed’. The Swahili
passive permits expression of the suppressed external argument, as in (45):

(45) Chakul kili-pik-wa na mwanamke yule.

food PAST-cook-PASS by woman that
‘The food was cooked by that woman.’

In addition, we find a class of derivates known as ‘stative verbs’, formed

regularly by suffixation of -k. Thus, we have verb stems such as the following
(Wilson 1985: 63; Ashton 1944: 226–8):

(46) (a) vunja ‘break’ vunjika ‘be broken’

(b) pasua ‘crack’ pasuka ‘be cracked’
(c) funga ‘close’ fungika ‘be closed’
(d) fungua ‘open’ funguka ‘be opened’

Stative verbs refer to a resultant state without any indication of an agent.

Thus, we have examples such as (47) and (48) (Ashton 1944: 229, 361):

(47) Sikuvunja kikombe hiki, kimevunj-ika tu

NEG.I.broke cup this broke-STAT just
‘I didn’t break this cup; it merely broke.’

(48) Sikufunga mlango, umefungika tu.

NEG.I.open door open-STAT just
‘I didn’t shut the door, it shut of itself.’

The difference between mlango ulifungwa ‘The door was closed (passive)’ and
mlango umefungika ‘The door is shut’ is essentially the same as in the English
translations: the passive refers to an event, the stative to a state (cf. Mchombo’s
discussion of similar facts, in Chichewa (Bantu).
224 Louisa Sadler and Andrew Spencer

Very intriguingly, the stative form is associated with a potential meaning in

addition to the simple intransitive meaning illustrated so far. Thus, the stems
fungika and funguka can also mean ‘be closeable/openable’ respectively. This
is reminiscent of the meaning of the English middle (This book reads easily ⇔
It is easy to read this book).
Wilson (1985: 65) claims that ‘any verb, provided its meaning allows it,
can be made into a stative form’, by which we take it that the verb’s lexical
semantics must be such as to imply the possibility of a resultant state. Thus,
stative formation is very productive, and arguably part of the paradigm of the
verb. However, from the brief descriptions provided here, it would seem to have
the same argument structure as the English middle. Here, then, is a case in
which we have something akin to English passive and middle constructions,
but both are realized by regular and productive suffixation. The difference is
that the passive merely suppresses the external argument, leaving it syntactic-
ally available, while the stative disposes of that argument altogether.8

4.2 Reflexives and reciprocals

Many languages have within their inventory of morphological operations a
class of processes which may be viewed as deriving reflexive or reciprocal verb
forms from transitive verb forms (see e.g. the brief discussion and exemplifica-
tion in Levin and Rappaport Hovav, Morphology and Lexical Semantics).
Syntactically, these operations are valency-reducing, resulting in predicates
which do not permit a direct function to be assigned to an NP corresponding
to the reflexive or reciprocal affix or clitic, as in the French examples (49):

(49) (a) Jean voit l’homme dans le miroir.

Jean sees the.man in the mirror.
‘Jean sees the man in the mirror.’
(b) Jean se voit dans le miroir (*l’homme).
Jean REFL sees in the mirror (the.man).
‘Jean sees himself in the mirror.’

Levin and Rappaport Hovav (Morphology and Lexical Semantics) suggest

that such processes result in a predicate with the same semantic representation
as the input predicate, and on this basis we might be tempted to classify all
such processes as morphosyntactic. There are a number of issues here. First,
although intuitively speaking it is clear that the basic verb and its reflexive
or reciprocal form have the same semantics, these operations clearly have a
semantic effect: namely, that of identifying the semantics of the fillers of two
role slots (and, in the case of reciprocal, placing a constraint on plurality). The
question is whether this identification is brought about syntactically (by syn-
tactic binding) or whether it falls purely within the lexical domain. Second,
 Morphology and Argument Structure 225

we have argued that the prototypically morphosyntactic processes are simply

relation-changing alternations. These operations, of which the best examples
are voice alternations, may be viewed as doing nothing more than providing
different sets of syntactic prominence arrays for sets of roles, or as mapping
between alternative syntactic realizations for the arguments of predicates. In
this context, then, we may ask whether reflexivization and reciprocalization
are morphosyntactic in the appropriate sense, specifically:

(i) Do they provide different grammatical function arrays/surface real-

izations for arguments?
(ii) Does the reflexive or reciprocal ‘role’ remain accessible in the syn-
tax (in the way that the suppressed argument of a passive remains
accessible)?

Both these questions are addressed in Mchombo’s (1993a) discussion of reflex-

ives and reciprocals in the Bantu language Chichewa. In Chichewa the object
marker (OM) is optional, and Mchombo treats it as an agreement marker (unlike
the subject marker, SM, which is ambiguous in status between an agreement
marker and an incorporated argument/function). Reflexives are realized by a
prefix -dzi- occupying the OM slot (FV = ‘final vowel’):

(50) Mkângo u-na-dzí-súpul-a

3-lion 3SM-past-REFL-bruise-FV
‘The lion bruised itself.’

Mchombo offers a most interesting argument for the syntactic ‘presence’ of

the reflexive marker on the basis of ambiguities in comparative clauses. In (51)
the reflexive gives rise to strict as well as sloppy identity readings, and also
to a comparative object (rather than subject) reading:

(51) Alenje á-ma-dzi-nyóz-á kupósá asodzi

2-hunters 2SM-hab.-reflex.-despise-FV exceeding 2-fishermen
(i) The huntersi despise themselvesi more than the fishermen i (despise
themselvesi) – sloppy identity reading.
(ii) The huntersi despise themselvesi more than the fishermen j (despise
themi) – strict identity reading.
(iii) The hunters despise themselves more than (the hunters despise)
the fishermen – comparative object reading.

The existence of strict identity and comparative object deletion readings points
to the presence of a syntactic argument (for Mchombo, in fact, an object)
corresponding to the reflexive in the two clauses.
Given Mchombo’s reasoning, it follows that at least at pas, reflexivized pre-
dicates have two arguments, so they remain bivalent even if they are not
transitive in surface syntax. This indicates that reflexivization in Chichewa is
226 Louisa Sadler and Andrew Spencer

a morphosyntactic operation. If, furthermore, Mchombo is correct that the

reflexive really is an object, then we have no difference in surface grammatical
function, although we do have a difference in surface expression (as an affix
rather than an independent NP).
Mchombo presents further data to support this position. In (52) we see that
under gapping, the reflexive verb patterns in a way parallel to a normal trans-
itive verb:

(52) Alenje á-ma-dzi-nyóz-á kupósá asodzi alimi.

2-hunters 2SM-hab.-reflex.-despise-FV exceeding 2-fishermen 2-farmers
‘The hunters despise themselves more than the fishermen the farmers.’

The essence of Mchombo’s claim concerning the Chichewa (and by extension,

the Bantu) reflexive is that it is present in the syntax, as an anaphoric syntactic
element, subject to syntactic binding. Since the domain of binding is syntax,
and this domain is separate from that of the morpholexical rules, reflexivization
cannot be a morpholexical rule.
The behaviour of the reflexive morpheme contrasts sharply with that of the
reciprocal. The reciprocal marker is a suffix to the verb root, hence part of the
verb stem. This means that in contrast to the reflexive marker, the reciprocal
participates in the process of vowel harmony, reduplication, nominalization and
imperative formation. This puts the reciprocal in the same position as expon-
ents of morphosyntactic operations such as passive, applicative and causative;
but this does not mean that it is a morphosyntactic operation itself. To see this,
note its behaviour in comparative clauses. As seen in (53), the reciprocal gives
rise only to the sloppy identity reading:

(53) Alenje á-ma-nyoz-án-á kupósá asodzi

2-hunters 2SM-hab.-despise-recip.-FV exceeding 2-fishermen
‘The huntersi despise themselvesi more than the fishermen j (despise
themselvesj).’

This strongly suggests that the process involved identifies the fillers of the two
semantic role slots lexically and not syntactically; that is, we have a (productive)
semantic derivation providing a predicate with a slightly altered semantic
representation. In parallel fashion, the counterpart of (52) is impossible with
reciprocal verbs.9

4.3 Causatives
In a causative construction a bare verb, V, an adjective, A, or sometimes a
noun stem, N, alternates with a verb meaning ‘cause/allow/persuade/help
. . . to V’ or ‘cause/allow/persuade/help . . . to become A/N’. In many genetic-
ally and typologically varied languages this is realized morphologically in a
 Morphology and Argument Structure 227

completely regular fashion. The causative alternation has been the subject of
considerable research, to which we can hardly do justice here. At first blush
this would seem to be an instance of the creation of an entirely new lexeme,
as in English They darkened the room (from adjective dark) or They enslaved the
populace (from noun slave). This is because there is an additional semantic
component of causation. Moreover, this component often receives subtly dif-
ferent interpretations (as indicated in our glosses above), such as persuasion,
instruction (‘tell someone to do something’) or permission (‘allow someone to
do something’). However, many researchers regard the morphological causat-
ive as an instance of an argument-structure alternation, rather than lexemic
derivation proper. This is particularly attractive when the causative is com-
pletely productive and lacking in lexical idiosyncrasies.
Japanese has an interesting and well-studied causative morphology. For
convenience we will follow the discussion in Tsujimura’s (1996) overview of
Japanese grammar (cf. also Shibatani 1976). The examples in (54) are adapted
from Tsujimura (1996: 247):

(54) (a) Hanako-ga arui-ta.

Hanako-NOM walk-PAST
‘Hanako walked.’
(b) Taroo-ga Hanako-o aruk-ase-ta.
Taroo-NOM Hanako-ACC walk-CAUSE-PAST
‘Taroo made Hanako walk.’
(c) Taroo-ga Hanako-ni aruk-ase-ta.
Taroo-NOM Hanako-DAT walk-CAUSE-PAST
‘Taroo had Hanako walk.’

The causative morpheme is a suffix taking the form -(s)ase (-sase occurs after
vowel-final stems). As is typical cross-linguistically, the subject of the basic
verb walk in (54a), Hanako, is expressed as a direct object marked by -o in
the causative version in (54b). This sentence means something like ‘Taroo
forced Hanako to walk’. In (54c), however, Hanako is marked with a dative
case marker -ni, and the interpretation is closer to ‘Taroo persuaded Hanako
to walk’.
When a transitive verb is causativized, the embedded subject is always
marked with dative case, as in (55):

(55) (a) Taroo-ga hon-o yon-da.

Taroo-NOM book-ACC read-PAST
‘Taroo read a book.’
(b) Hahaoya-ga Taroo-ni hon-o yom-ase-ta.
mother-NOM Taroo-DAT book-ACC read-CAUSE-PAST
‘His mother made/had Taroo read a book.’
228 Louisa Sadler and Andrew Spencer

This has been linked to a general prohibition in Japanese against two nominals
marked with -o in one clause. In other languages it is quite common for the
embedded subject of a transitive verb to be marked as an (optional) oblique
rather than as a direct object, but Japanese is only accidentally similar to this
type.
Morphological causatives in Japanese are very productive, and are relatively
free of lexical restrictions or idiosyncrasies. Moreover, there is good evidence
(cf. Shibatani (ed.) 1976) that at some level a causativized transitive sentence
such as (56) behaves like two clauses, much like the bi-clausal English transla-
tion (again, adapted from Tsujimura 1996: 255):

(56) Taroo-ga Ziroo-o/ni zibun-no heya-de benkyoo-sase-ta.

Tarooi-NOM Zirooj-ACC REFLi/j-GEN room-in study-CAUSE-PAST
‘Taroo made Ziroo study in his own room.’

Reflexivization in Japanese is subject-oriented, so (56) suggests that both Taroo

and Ziroo correspond to subjects at some level. We can say that, in a certain
sense, Ziroo is the subject of study, while Taroo is the subject of CAUSE. Evid-
ence from ambiguities with temporal and subject-oriented adverbs points the
same way. We can call a causative construction of this sort ‘bi-clausal’.
In what sense is a productive morphological causative a morphosyntactic
operation? One viewpoint, which in some ways incorporates the insights of
Baker (1988a), is to say that the causative operation comprises the fusion or
union of two argument structures. Thus, suppose we take the argument struc-
ture of predicates such as walk or hit as something like (57):

(57) (a) walk: <arg. 1>

(b) hit: <arg. 1, arg. 2>

Then we can say that the causative will be (58):

(58) (a) cause-walk: <arg. 0 <arg. 1>>

(b) cause-hit: <arg. 0 <arg. 1, arg. 2>>

To some extent it is immaterial for our typology whether structures such

as (58a,b) are the result of an operation ‘in the lexicon’ or the result of V-
movement ‘in the syntax’. The point is that it makes sense to say that the opera-
tion is defined over an essentially unrestricted set of predicates and results in
an argument structure representation along the lines of (58), such that all three
of the arguments are syntactically visible in some sense. The operation thus
qualifies as morphosyntactic in our sense.10
Languages differ in how these three arguments are realized (see Baker 1988a
for a detailed study of this). Cross-linguistically, arg. 1 in (58a) is uniformly real-
ized as a canonical direct object. However, the (58b) structure (where it exists)
will sometimes realize arg. 1 as a direct object, and sometimes arg. 2 retains the
 Morphology and Argument Structure 229

direct object status, depending on the language. When arg. 1 becomes the
derived direct object, arg. 2 is generally marked as a second object, though this
will often be inert to object-oriented processes like passive. Thus, despite the
fact that Taroo-ni in (55b) is an obliquely marked phrase, it is in fact the direct
object of the causative. This is clear when we try to passivize a causative.
Consider the data in (59) (cf. Tsujimura 1996: 259):

(59) (a) Ziroo-ga Mitiko-ni kodomo-o home-sase-ta.

Ziroo-NOM Mitiko-DAT child-ACC praise-CAUSE-PAST
‘Ziroo made Mitiko praise the child.’
(b) Mitiko-ga Ziroo-ni kodomo-o home-sase-rare-ta.
Mitiko-NOM Ziroo-DAT child-ACC praise-CAUSE-PASS-PAST
‘Mitiko was made to praise the child by Ziroo.’

Here, Mitiko-ni in (59a) has been promoted to subject when (59a) is passivized,
while kodomo-o ‘child-ACC’ remains marked as an accusative. However, if we
try to passivize on the accusative marked object in (59a), kodomo-o, we get an
ungrammatical sentence:

(60) *Kodomo-ga Ziroo-ni Mitiko-ni home-sase-rare-ta

child-NOM Ziroo-DAT Mitiko-DAT praise-CAUSE-PASS-PAST
‘The child was made praised by Mitiko by Ziroo.’

Thus, kodomo-o behaves like an inert, or frozen, second object. Such ‘freezing’
of the second object is typical, though in some languages both objects retain
full object properties, including passivization, object agreement marking and
so on (see Baker 1988a and Bresnan and Moshi 1990 for discussion of variation
in Bantu in this respect).
In other languages arg. 2 in (58b) is treated as the object of the causative, with
arg. 1 becoming an oblique marked adjunct, generally optional. This is how
causatives work in Turkish, for instance. Consider (61) (Comrie 1976: 268):

(61) Di/çi mektub-u müdür-e imzala-t-tı.

dentist letter-ACC director-DAT sign-CAUSE-PAST
‘The dentist made the director sign the letter.’

Here the object of the basic verb, letter, remains the object of the derived
causative verb, while the subject of the basic verb, director, appears as a Dative
marked oblique.
In general, when arg. 2 is treated as the derived (as well as basic) object,
the causative construction behaves syntactically like a single clause. Thus, if
the object is a reflexive pronoun, it can only refer back to the surface subject,
arg. 0, and not to arg. 1. In other words, arg. 1 cannot be treated as a kind of
subject for the purposes of reflexivization, and we have a monoclausal causat-
ive construction. This is in contrast to the biclausal causative of Japanese.11
230 Louisa Sadler and Andrew Spencer

The semantic effects of the causative operation will differ somewhat from
language to language, and, importantly, from construction to construction, so
Japanese causatives with -o marked objects have the coercive or the adversity
reading, while those with the -ni object have the permissive reading. However,
the basic semantic relationships are the same cross-linguistically.
Japanese also has verb pairs which we can interpret as lexical causatives,
often related by non-productive ablaut. Some examples are given in (62)
(Tsujimura 1996: 260):

(62) causative intransitive

tomeru tomaru ‘stop’
ageru agaru ‘rise/raise’
sageru sagaru ‘lower’
okosu okiru ‘wake up’
nekasu neru ‘sleep’

Despite the meanings of these pairs, the causative member does not behave
like a morphological causative derived with -(s)ase. Thus, with reflexives we
get a monoclausal patterning. Compare (63) with (56):

(63) Taroo-ga Ziroo-o zibun-no heya-no mae-de tome-ta.

Tarooi -NOM Zirooj -ACC REFLi /*j room-GEN front-at stop-PAST
‘Taroo stopped Ziroo in front of his room.’

In (63) Ziroo is a ‘pure’ object, and hence cannot be the antecedent to the
reflexive. Contrast this with (64), in which the intransitive toma- is causativized
morphologically, not lexically:

(64) Taroo-ga Ziroo-o zibun heya-no mae-de tomar-ase-ta.

Tarooi-NOM Zirooj -ACC REFLi/j room-GEN front-AT stop-CAUSE-PAST
‘Taroo made Ziroo stop in front of his room.’

Here, his room can refer to Taroo or to Ziroo.

Clearly, we want to relate the intransitives to their lexical causatives by
means of a morpholexical operation, while we will argue that the morpholo-
gical causatives (at least in Japanese) are the result of morphosyntactic opera-
tions. A further piece of semantic evidence in favour of this is that the lexical
causatives signify direct causation, in which the agent must come into direct
contact with the patient. Morphological causatives, however, denote indirect
causation. Thus, (64) could refer to a situation in which Ziroo is brought to a
stop outside Taroo’s room by a large obstacle which Taroo has left there.
Taroo could be hundreds of miles away when Ziroo is thus halted. Sentence
(63) cannot have such an interpretation.
Discussion of how exactly the linking to grammatical functions is achieved
would require a separate chapter in itself, and would require a detailed analysis
 Morphology and Argument Structure 231

of the morphosyntax of various subject- and object-oriented processes in the

languages which have such causatives. However, we must address one question
which arises with a morphosyntactic analysis of causatives. Since there is a
sharp shift in meaning (the addition of a CAUSE predicate), in what sense can
any causative be morphosyntactic: that is, an operation over argument struc-
tures? We assume that what is actually happening here is that the causative
operation involves addition of an argument structure, consisting minimally of
an external argument position and a further argument position corresponding
to the embedded proposition. This then ‘fuses’ or ‘merges’ with the argument
structure of the basic predicate. Language-particular principles dictate exactly
what happens to the elements of the embedded pas. We thus obtain a repres-
entation such as (65):

(65) CAUSE(x, Ev) [evSTOP(y)]

<x <y>>

This is a complex predicate, whose overall argument structure is a function

of two independent argument structures. However, the fusion takes place at
the level of pas, not lcs. The lcs portion of (65) is what we would see in a
syntactic causative, so the causative predicate and clausal semantic argument
can behave, to some extent, independently.
In the lexical causatives we simply have a causative lcs which includes the
embedded predicate:

(66) [CAUSE(x)[STOP(y)]]

<x <y>>

This does not differ significantly from the representation for any (mono-
morphemic) transitive verb with a causative component.

4.4 Noun incorporation

Levin and Rappaport (Morphology and Lexical Semantics) illustrate
English verbs derived by conversion from nouns, such as butter in to butter
toast (with margarine). Here the converted noun corresponds to an object in a
syntactic construction such as to spread butter on the toast. In some languages
the conversion is signalled morphologically, by affixation: for example, Dutch
ver-botter-en from the noun botter, Hungarian meg-vaj-az from the noun vaj. In
other languages, we can create a lexical unit akin to the VP to spread butter by
means of noun incorporation. An example from Chukchee is cited by Gerdts
(Incorporation), from Polinskaja and Nedjalkov 1987: 240:
232 Louisa Sadler and Andrew Spencer

(67) ftlfg-e kawkaw mftqf=rkele-nin.

father-ERG bread.ABS butter=spread.on=AOR.3SG:3SG
‘Father spread the bread with butter.’

This corresponds to (68) without incorporation:

(68) ftlfg-e kawkaw-fk mftqfmft kele-nin.

father-ERG bread-LOC butter.ABS spread.on=AOR.3SG:3SG
‘Father spread the butter on the bread.’

In the model proposed by Baker (1988a) such an alternation would be the

result of movement applying to the head noun of the object NP butter, forming
a compound verb as in (69):

(69) (a) S

NP VP

V NP NP

father spread butter on-bread

(b) S

NP VP

V NP NP

N V

father butteri-spread ei on-bread

The movement of butter leaves a trace, ei , which has to be properly gov-

erned by the verbal complex. Given this analysis, Baker is able to account
for a number of important features of NI, such as the stranding of modifiers
in some languages. In this type of approach, the fact that the noun butter is
interpreted as the direct object of the verb is a consequence of the fact that
it is, indeed, the verb’s object in the syntactic representation. Further discus-
sion of some of the implications of this type of approach is given in Borer
(Morphology and Syntax).
A similar kind of analysis is available in principle for an English synthetic
compound such as butter-spreader (e.g. Roeper 1988, Lieber 1992), though here
 Morphology and Argument Structure 233

there are important questions surrounding the fact that the verb is nominal-
ized and cannot occur in finite forms, as witnessed by the ungrammaticality
of *Tom butter-spreads his toast every morning (with margarine). An alternative
analysis, which is also open to noun incorporation proper, would say that the
incorporated element discharges an argument position, but not in the same way
as a syntactic direct object. Sproat (1985) and DiSciullo and Williams (1987)
propose rather different solutions, which, however, are alike in not appealing
to syntactic movement. This is especially attractive for synthetic compound-
ing, in which compounding renders the verb syntactically intransitive: there is
no stranding of modifiers, and no ‘doubling’ is possible: *butter-spreader of bread
with rancid ei , *butter-spreader of bread with margarine.
What is the argument-structure status of butter-spreader? Here, again, we
have a combination of lexemes, which produces what is morphologically
more akin to a single word than to a genuine phrase. We can then say that the
argument structure of the noun butter is somehow ‘fused’ with that of the verb
stem spread in such a way that butter is interpreted as the direct internal argu-
ment of spread. If this is a direct operation over the verb argument structure,
then presumably we have to say that the nominalizing suffix -er is attached
after the internal argument is discharged. This means we must assume that the
grammar creates an otherwise non-existent verb stem [butter-spread] which
then triggers discharge of the internal role. Only then does suffixation take
place to give [[butter-spread]er] (cf. Sproat 1985). On the other hand, we might
argue that the argument structure of the verb is in some way inherited by the
nominalization. Thus, [[spread]er] retains at least the direct internal argument
of the verb stem. This argument can then be projected either as an of phrase
(a spreader of butter (on bread)), or as part of the compound to give [bread
[spread-er]] (cf. DiSciullo and Williams 1987).
Approaches to synthetic compounding which appeal to operations over
argument structures leave a number of questions unresolved, of course (see
Carstairs-McCarthy 1992 for discussion). One of these is the status of the notion
of ‘inheritance’ of argument structure (see Lieber 1992 for discussion of this).
Another concerns the generality of the approach. As stressed by Roeper and
Siegel (1978) in an early generative treatment of synthetic compounds, English
permits compounds in which an adverbial modifier is incorporated, as in quick-
drying (paint), sun-dried (tomatoes), home-made (cakes) and many others. Recent
discussion has tended to ignore these cases. An approach which appeals solely
to the discharge of argument-structure positions has little to say about them.
The situation is rendered more interesting by the fact that incorporation
of adverbials is observed in some noun-incorporating languages. Thus, in
Chukchee it is possible to say things like Tom quick-ran or The mother tent-sewed
the shirt (i.e. the mother sewed the shirt in the tent) (cf. Muravyova, Chukchee
(Paleo-Siberian). The incorporation of adverbials is a distinct embarrassment
to the framework of Baker (1988a), which is so constructed as to explicitly
exclude such constructions (for extensive discussion of this point see Spencer
1995). It seems to us that there is merit in exploring the idea that incorporative
234 Louisa Sadler and Andrew Spencer

structures of this sort can, in part at least, involve something akin to the com-
plex predicate formation we proposed for morphological causatives. Thus, the
Chukchee example in (67) might involve a representation along the lines of
(70), in which the verb and its object have distinct lcs representations but at
pas that of the object butter is indexed with the verb’s internal-argument posi-
tion, thereby saturating it and preventing it from being realized syntactically:

(70) [[x ACT] CAUSE [y BECOME-ON z [by x SPREAD y]] [BUTTER(w)]

spread <x <w, Ploc z>>

This can be thought of as the pas equivalent of syntactic incorporation in

such theories as those of Baker (1988a) or Sadock (1991) (for the latter, cf.
Sproat, Morphology as Component or Module). In the relatively rare cases
like Chukchee where adverbials can be incorporated, we can adopt a similar
analysis, in which the adverbial’s argument structure is fused with an event
position at pas.
Noun incorporation in many languages is lexically restricted, non-productive
and idiosyncratic, much like noun-to-verb conversion of the type butter the toast
in English (see Mithun 1984 for extensive discussion of this). For such languages,
we would argue that the incorporation takes place at the lcs level, despite
being realized morphologically by compounding, giving a representation such
as (71):

(71) [[x ACT] CAUSE [BUTTER BECOME-ON z [by x SPREAD BUTTER]]

This could then correspond to any of the pas representations in (72), depend-
ing on the language, corresponding to syntactic structures (73):

(72) (a) spread <x <Ploc z>>

(b) spread <x <z>>
(c) spread <x <y, Ploc z>>

(73) (a) butter=spread on to the toast

(b) butter=spread the toast
(c) butter=spread margarine on to the toast

5 Summary

We have argued that valency alternations can be of two distinct types: morpho-
lexical operations at a semantic level and morphosyntactic operations at a level
of argument structure. The morpholexical operations are likely to be semantic-
ally or lexically restricted, and to bring with them semantic changes which
 Morphology and Argument Structure 235

cannot always be predicted from the valency shift as such. Morphosyntactic

operations are more often semantically unrestricted, and are thus often defined
solely in terms of input/output conditions on argument-structure representa-
tions, independently of the semantic representation. They generally do not give
rise to additional semantic affects (modulo other aspects of the construction).
The result of a morpholexical operation tends to behave syntactically in the same
way as a corresponding monomorphemic predicate, whereas syntactic pro-
cesses may have access to the individual parts of the result of a morphosyntactic
operation (cf. the difference between morphological and lexical causatives).
We illustrated these distinctions by contrasting passives/middles in English
(and Bantu), reflexives/reciprocals in Bantu, and morphological/lexical causat-
ives in Japanese. We finally discussed noun incorporation and synthetic com-
pounding as possible instances of complex predicate formation, in which pas
positions (rather than lcs positions) are saturated morphologically.

NOTES

1 The terminology, with essentially 5 There are languages in which this

this interpretation, is due to restriction does not hold: e.g.
Ackerman (1992). We acknowledge Turkish, in which unaccusative
that the nomenclature is potentially predicates, including passivized
rather misleading, particularly given verbs, can be passivized. See
that ‘morpholexical’, which already Spencer 1991 for brief discussion.
has a number of unrelated uses in 6 This explains in part the alternative
linguistics, is used to refer to both term used for unaccusative
our morpholexical operations and predicates, ‘ergative’. However, this
our morphosyntactic operations in is a rather misleading term, since
the LFG literature. ergative patterning would lead one
2 In practice, thematic role labels are to expect objects to pattern with
used for convenience, not out of all intransitive subjects. The more
theoretical commitment to these appropriate alternative to ‘ergative’
labels. It is generally understood would therefore be ‘inactive’,
that they stand for more complex though as far as we know, no one
lcs representations, perhaps of the has ever made this terminological
kind argued for by Jackendoff 1990. proposal.
Alternatively, some theorists take 7 The reader should bear in mind
Dowty’s (1991) Proto-roles as their that this conclusion is meant to
starting-point (cf. Ackerman 1992). follow for English passives and
3 Though this is a rather complex English middles. Constructions
matter: see Ackerman 1992, Alsina which are called ‘passive’ or
and Mchombo 1993, Bresnan and especially ‘middle’ in other
Moshi 1990, for discussion. languages, e.g. French (Levin and
4 There are various formulations of Rappaport Hovav, Morphology
these operations in the literature. and Lexical Semantics), may well
236 Louisa Sadler and Andrew Spencer

have different properties from the 10 Alsina (1992), to whom the

ones described here (French middles approach here offers a certain
can be eventive, for instance). debt, provides a detailed analysis
In addition, many factors govern of causatives in Chichewa along
whether a semantic argument is similar lines, except that for him
accessible to syntactic processes, the arg. 1 position is an argument
so absence of an implicit argument of the causative itself. It is a matter
cannot be taken as criterial for a of considerable debate whether the
morpholexical valency-reducing ‘real’ argument of causation is a
operation. patient (‘causee’) and an event, or
8 We haven’t investigated the full set just an event, or whether these
of properties of the Bantu passive represent direct and indirect
and stative constructions. Given causation respectively. The matter
the level of disagreement over the is tangential to our main concerns
status of Germanic middles in the here.
recent syntactic literature, a detailed 11 In some languages of the Japanese
cross-linguistic study of that sort on type, it is only arg. 1 which can be
Bantu might be rather timely. the antecedent of a reflexive in a
9 A terminological warning: causative. Japanese allows arg. 0
Mchombo uses the term to serve as antecedent, because
‘morpholexical’, but does not its reflexive allows ‘long-distance’
explicitly make our terminological binding, by the subject of a higher
distinction between ‘morpholexical’ clause.
and ‘morphosyntactic’.