Argument Structure and Morphology

The Linguistic Review l :81-114

Argument Structure and Morphology

EDWIN WILLIAMS

That there is some regulär relation between the argument structures ofmorphologically related words is clear and well-known, and there areseveral proposals about how such relations can be characterized. Thispaper presents another such proposal, one that differs from previousproposals mainly in that it represents an attempt to be highly restrictive,perhaps erring in that direction.

By argument structure of a lexical item I mean simply a labelled listingof the arguments that a lexical item can have. I will adopt the labellingproposed by Gruber (1976) (Actor, Theme, Goal, Source), though theactual labels themselves are not important. These are called thematicrelations.

Now suppose that a lexical item has an argument structure, and weapply some morphological rule to that lexical item to derive a new lexicalitem, with a new argument structure. What we want to characterize inthis paper is the function that will relate the old argument structure tothe new argument structure. One way to do this would be to dive right inand write a notation for writing argument structure transformations in.This is essentially what past proposals have done. What I intend to doinstead is to specify two functions ("internalize X" and "externalize X"),each parameterized by the thematic relations. Since there are finitelymany thematic relations, there are finitely many functions, and the claimis that these exhaust the possibilities.

Before these functions can be specified, there are a number of topicsthat must be discussed which are independent of these functions. Forexample, we must have a theory of exactly how many argument types(thematic relations) there are. Furthermore, we must determine whetherthere is any further structure to the argument structure of a verb than

* This paper is a revision of a paper with the same title written in February, 1980.The author wishes to thank the Max Planck Institute for Psycholinguistics for theopportunity to complete the work on this paper äs a member of its research staff inthe summer of 1980.

Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)Authenticated | 172.16.1.226

Download Date | 3/4/12 2:05 PM

82

that embodied in the notion "unordered list of arguments". In Section l,I will propose that there is one minimal additional element of struc-ture to the <4unordered list of arguments": the distinction between "inter-nal" and "external" arguments. Such things would need to be studied evenif there were no morphology or no need for any function from one argu-ment structure to another.

We also would need some specification, independent of morphologicalfunctions, indicating how the various arguments, or thematic relations, are"realized" in syntax; that is, for a given verb in a given sentence, how doyou know which phrases in that sentence correspond to which argumentsof that verb. There are innumerable ways to specify the realization of thearguments, and there are several important issues. One issue concerns theneed for grammatical, in addition to thematic, relations. Without gram-matical relations, it is necessary to specify directly the realization of eachthematic argument; for example, that the theme is realized äs the NPdirectly dominated by the VP that dominates the verb, etc. But if thereare grammatical relations, this specification can be indirect; for example,one might specify that the theme is the direct object and that the directobject is the NP immediately dominated by the VP that dominates theverb. I will explore the possibility in this paper that there are no gram-matical relations; that is, that the realization of the thematic relations isdirectly specified. Some of the issues involved in this decision are thefamiliär ones: if there are no grammatical relations, then of course therecan be no rules, either syntactic or lexical, that refer to them, etc.

Another issue that arises in the "realization" problem is the level ofgenerality that holds for the rules governing the syntactic realizationof arguments, and similar problems arise in theories with and withoutgrammatical relations. In the worst case, every verb of every languagewould need to specify the realization of every one of its arguments - andin the best case, realization would be completely fixed, admitting of noVariation. Now reality clearly lies between these: given that some languagesrealize arguments positionally and others according to a case System, andothers with some mixture of these, then the best case is not possible. Onthe other hand, it seems possible to avoid specification of realizations forparticular lexical items in particular languages, so the worst case does nothold either. Now it appears that some parameters of realization can varyfrom language to language, and others from category to category within alanguage, though hopefully the latter can be kept to a minimum. In Sec-tion 2 we will specify a solution to the realization problem for English andmake some general remarks about the universal characterization of realiza-tion rules.

Now, the material in the first two sections, äs I mentioned earlier, isindependently required; even in a language with no morphology, we will



83

want to know what the argument structure of lexical items is, and howarguments are syntactically realized, and we will want universal theoriesof these things. Clearly, how these theories are formulated will determineto some extent what kind of relation can hold between the argumentstructures of different morphologically related words. It is to this thatSection 3 is devoted.

1. ARGUMENT STRUCTURE

As mentioned above, we will take the argument structure of a lexical itemto be an unordered list of its arguments, with one further complication.We will use the labels proposed by Gruber (1976), whichare Actor, Theme,Goal, Source, etc. Thus we might represent the argument structure ofhit in the following way:

(1) hit: (Actor, Theme)

Now, in the theory we will develop here, we will want to distinguish oneof the arguments in an argument list, which we will call the "external"argument, and we will adopt the notation of underlining to mark thisargument; thus the argument structure of hit might be:

(2) hit: (Actor. Theme)

The notion external argument corresponds in some respects to subject, butwe avoid that term because of its ambiguity. The external argument issimply a distinguished argument, not a syntactic position, a case, or some-thing eise. Another reason we avoid the notion subject is that this termimplies other terms äs well, such äs direct object, indirect object, etc.,which will play no role in our svstem.

The obvious question at this point is: why is there one and only oneexternal argument in an argument structure? We can answer this questionwith reference to the theory outlined in Williams (1980), where the termexternal argument was introduced. In that paper, it was proposed thatthe subject predicate relation was to be indicated by coindexing subjectand predicate; one reason for this was that a c-command restriction oncoindexing then would automatically apply to the subject predicate re-lation, requiring that a subject c-command its predicate, and this seemscorrect. Thus for example, the subject of a sentence and its verb phrasewill be coindexed by a rule of predication:

(3) [



84

The use of indices to mark the subject predicate relation has anotherimportant consequence. If we take the X-bar theory seriously, then everyphrase, and in particular, every predicative phrase, apart from S, is themaximal projection of some lexical item. The use of indices to indicatethe subject predicate relation entails that there can be only one subjectfor any maximal projection used predicatively, since any maximal pro-jection can bear only one index. In essence then, all predicative phrasesthat are maximal X-bar projections are necessarily monadic. Given thatlexical items are not monadic in general, but that maximal projections ofthem are, we may ask what the relation is between a head and its maximalprojection. Clearly, all of the arguments of that head must be specifiedinternal to the maximal projection, except for the one external argument— that is, the item that is located external to the maximal projection butwith which the maximal projection is coindexed. Thus, for example, theverb A/Y, äs we mentioned, has two arguments: an Actor, which is external,and a Theme, which is internal. This means that in a sentence with hit, thetheme must be specified within the verb phrase of which hit is the head,and the Actor of hit must be specified in a position external to the verbphrase of which hit is the head, but with which that verb phrase is coin-dexed.

Thus, the external argument of a lexical item is located outside of themaximal projection of that item, and since it must be coindexed with thatmaximal projection, and only one index is allowed, there can only be oneexternal argument. There is certainly no necessity forthings tohave workedout this way, and in fact, if one adopts an unconstrained use of the lambdacalculus, äs was done in Williams (1977), then things will not work outthis way. If a maximal syntactic projection receives a logical translationwhich lookslike:

(4) [XxXy. . . . ]^

then the thesis is denied: this phrase has two external arguments, onecorresponding to each lambda. In essence then, the use of coindexing toindicate the subject predicate relation amounts to a limited use of thelambda calculus: you only get one lambda per X-bar maximal projection,and the index on the predicative phrase serves the role of this single ab-straction operator.

So, the external argument of a lexical item is the one that correspondsto the NP in a sentence of which a phrase with that item äs its head is pre-dicated. The use of indices to indicate the relation between a predicativephrase and its subject sets an upper bound (of one) on the number ofexternal arguments that that phrase can have, but no lower bound. Thelowest conceivable bound is, of course, zero, and this is the analysis given



85

in Williams (1979) to the VP with head seems. Bresnan has shown thatseems cannot have a "subject" (Bresnan 1972):

(5) *That John is here seems*John's presence seems

In the terms outlined here, this amounts to the claim that there can beno external argument. We can indicate this by giving seems an argumentstructure in which nothing is underlined:

(6) seems: (Theme, Goal)

This means that when seems heads a VP, all of its arguments must belocated internal to that VP; hence, seems has no external argument.

In order to maintain this, we must distinguish between occwrence ofanargument and binding of an occurrence of an argument. An item mightoccur internal to a phrase but be bound external to the phrase; it is never-theless an internal argument of the head of that phrase. This is exactlywhat happens with raising seems:

(7) Johnj [seems [tj to be happy]§ to me.Jyp.

Here, both of the arguments of seems (the Theme, t to be happy, and theGoal, to me) occur within the VP. However, part of one of the argumentsof seems is bound from without: John binds the trace. In this case,/0A/zis the subject of the predicative phrase "seems to be happy to me" but notby virtue of corresponding to the external argument of the head of thatphrase, which does not have one, but rather by virtue of binding a trace inthat VP.

Of course, phrases other than VP are used predicatively. APs for ex-ample, äs in John is proud of Mary. Here, proud has two arguments, oneinternal, Mary, and one external, John. There are also adjectives likecertain, which have no external argument; hence we have raising for thiskindof adjective:

(8) Joluij is [certain [t. to win] ]yp(Theme) i

The case of Noun Phrase is a bit more complicated, but the result will bethe same: that nouns have one and only one external argument. In the caseof verb, when there is an Actor and there is an external argument, then theActor is the external argument - this is a fairly clear universal. Now,nouns have Actors äs well; witness the famous NP: thearmy's destruction



86

of the city, or its passive: the city 's destruction by the army. But note thatin both of these, the Actor is internal, that is, it is located inside themaximal projection of N (NP, or N").

So, if Ns have external arguments, they are not Actors. But do Ns haveexternal arguments? I believe that we want to say that they do, at leastwhen they occur äs the head of a NP that is used predicatively, äs in "Johnis a fool". We might say that fool had an external Theme (that is, it hasthe argument structure (Theme). Then we could say that when foolappeared äs the head of a NP used predicatively, the subject of that pre-dication corresponded to the theme of fool. But we can show that thiswill not do generally, because we can show that this would require somenouns to have both internal and external Themes. Consider the following,forexample:

(9) I consider that [destruction of a city by evil forces]

Here, "destruction of a city by evil forces" is predicated of "that". Butthe predicative NP has internal Theme and Actor - thus to what argumentof destruction does that correspond; i.e., what is the external argument ofdestruction, if not Theme or Actor? Clearly the external argument ofsuch a noun has no counterpart in the verbal System; suppose we inventthe label R to name that argument of the noun which is external. Then wewould assign destruction the argument structure (R, Actor, Theme). Thelabel R is meant to suggest "referential", since it is this argument positionR that is involved in referential uses of NPs äs well. Thus, for example, iffool has the argument structure (E), we might assign the following logicalrepresentations for the predicative and referential uses respectively: Johnand the variable is in the position of R:

(10) John is a fool fool (John)The fool left E!x(fool (x) & left(x) )

To sum up this section, we have assigned a minimum amount of structureto the noüon argument structure. The argument structure of a lexical itemis simply a list of its arguments. In addition, one of those arguments isdistinguished from the rest äs being the external (underlined) argument.The external argument of a lexical item corresponds to the NP of whichthe major projection ofthat lexical item is predicated.

It is worth emphasizing, that for our purposes the argument labels(Actor, etc) have no intrinsic content or structure; the only property thateach one has is that it is different from the others. No doubt "Actor" hassome cognitive content, but we claim that this is linguistically irrelevant;i.e., that the rules determining the argument structure of derived words



87

which we will discuss in Section 3 are not concerned with whatever thatcontent is.

2. THE SYNTACTIC DEPLOYMENT OF ARGUMENTS

We now tum to the characterization of the rules of "realization". Thebasic question, for a lexical item in a sentence, is how is it determinedwhich NPs in that sentence function äs realizations of which argumentsof that lexical item?

A partial answer has already been given: the internal arguments must berealized internal to the maximal projection of the lexical item, and theexternal argument (if there is one) must be coindexed with the maximalprojection of the lexical item. The coindexation of the external argumentis governed by the rules of predication, äs outlined in Williams (1980),and will not be the subject of further discussion here. In the previous sec-tion t wo universals concenüng the external argument were given:

(11) l. If there is an Actor, it must be external for V2. R must be external for N

The first of these will have important consequences for the analysis ofthe passive constructioii, which we wfll consider in Section 3.

We now turn t o the deployment of the internal arguments. To describethe deployment of the Theme of the verb Hit, for example, we might say,"the Theme of hü is the NP that is immediately dominated by the VP ofwhich hit is the head" - or, to use a familiär notation:

(12) Theme: (NP,VPhit)

Similarly, to describe the location of the Goal of a verb like give wemight write:

(13) Goal:(NP,<PPto,VP,*e))

In fact, though, the reference to VPhit and VP^ is entirely superfluous,since by definition these arguments must be internal to the maximalprojection of the lexical item (in this case, V) of which they are arguments,so we may drop the reference to VP. We might suppose äs well that theseparticular realizations are not particular to the verbs in question, but infact are valid for verbs that have the argument types Theme and Goal,and thus we may drop the reference to particular verbs. We are left withgeneral realizations for Theme and Goal:



88

(14) Theme: NPGoal: (NP,PPto)

The subscripting notation X will be useful in case languages, where, forexample, we will refer to realizations like the following:

(15)

which is to be read: the Goal is the NP headed by the feature "dative"(or, whose head bears the feature "dative").

Notice that the realization for Goal given above is not category specific,since we have dropped the reference to VP. This is a welcome result, sinceGoal is realized äs PPto in all English categories:

(16) the trip to Houstonapparent to Mary

In general, a realization rule wül have the form:

(17) A:(Xy,Zw)

where X and Z are major categories, and y and w are either prepositions orcase features, and A is one of the argument types.

We will assume that the realization rules are verb-independent; that is,if a verb has an argument of type X, then any realization rule for typeX in the language will be applicable.

Below, we list some of the realization rules for English:

(18) Actor:(NP,PPby)

Actor:(NP,PPof)(limited to NP)Goal:(NP,PPto)

Goal: (NP2)Theme: (NP)Theme :(NP,PPof)

Source:(NP,PPfrom)

X: (NPposs)

John was seen by BillThe destruction by the armyThe shooting of the hunters

to give to Johnthe gift to Johngive Bill the bookhitBillthe destruction of the citydeprive Bill of moneyaware of the accidentspeak of somethingarrive from Houstonthe arrival from Houston



89

Most of these are self-explanatory. The last one says simply that a possess-ive NP may bear any relation whatever to the head noun; this is a greatexaggeration, but it is a first approximation that is difficult to improveupon. The subscript "2" on NP in the realization rule for Goal is simply ashorthand for whatever the appropriate distinction is between the firstand second NP of the double object construction in English; perhaps theydiffer in terms of a case distinction that is not visible in English.

It is important to emphasize that these rules are specified once and forall for a language (English, in this case) and that no particular lexical itemcan set up its own realization rules. Thus this Information is not listed withparticular lexical items.

It might be objected here that our claim to have avoided the use ofgrammatical relations is radier empty now, since the realization rules wehave offered are written in the notation that was originally invented (inChomsky (1965) ) for grammatical relations. But this objection misses thepoint. In a theory with grammatical relations, there will be two sets ofentities, the thematic relations and the grammatical relations, and therewill be two sets of realization rules - one set which says which thematicrelations are realized äs which grammatical relations, and another setwhich says which grammatical relations are realized in which syntacticconfigurations. In the theory here, there is one set of realization rules andone set of entities.

There is more at stake than these numbers, however. Given that the setof thematic relations and the set of grammatical relations is not isomorphic(under the realization rules), then in a theory with grammatical relations,one might imagine there being rules, say morphological or lexical rules,which operate on representations, say of words, purely in terms of gram-matical relations, äs well äs rules that operate on thematic relations. Ofcourse this will be impossible if there are no grammatical relations. In thesystem outlined in this paper, the closest thing to a grammatical relationis the distinction between internal argument and external argument,where external argument corresponds to the notion subject. Even here,though, there is a great difference. In the Systems with grammaticalrelations, the subject argument has the same Status äs the others; itsrealization, for example, is written in the same notation (e.g., in the Sys-tem of Chomsky (1965), the subject is (NP, S) whereas the object is (NP,VP)). In the System of this paper, however, the two receive entirelydifferent treatments. The internal arguments are subject to the realizationrules; the subject, or external argument, is handled by an entirely differentmechanism, the predication coindexing mechanism outlined in Williams(1980). Thus, although some notation has been borrowed from the de-finition of grammatical relations in Chomsky (1965), the system in thispaper differs in a number of important respects from Systems based ongrammatical relations. Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)

Authenticated | 172.16.1.226Download Date | 3/4/12 2:05 PM

90

There are some problems — to which we will return for detailed dis-cussion in Section 6 — with the notion that the realization schemes are notverb-particular. This predicts, for example, that if there are two ways torealize a certain argument type, then a verb that has an argument of thattype will be able to use both realizations. And there are many examplesin which this is true; for example, many verbs with Goal can use the dou-ble object construction, äs well äs the PPto construction. But what aboutthe verbs for which only one of these is possible (donate, for example)?For these we can use the mechanism of subcategorization: donate is NPPP, whfle give is NP NP and _NP PP.

3. HOW MORPHOLOGY CAN AFFECT ARGUMENT STRUCTURE

It wfll be our contention that a morphological rule can affect only theexternal argument of its input, and that it can affect this external argu-ment in only one of two ways: it can make one of the internal argumentsinto the external argument, or it can add a new external argument. Ob-viously, morphologically related words differ in more ways than this willbe able to account for. For example, the subcategorizations of morpho-logically related words differ, and when the words are in different cate-gories, the realizations might be different. We will first review some ofthese differences, suggesting that they in fact need not be specified in anyparticular morphological rule, but will follow from general considerations.Afterwards, in Section 3.2, we will examine the actual morphologicalrules themselves.

3.L Predictable Side-Effects of Morphological Rules

Morphological rules do not specify a number of observable differencesbetween their inputs and their Outputs. Consider, for example, the rule ofnominalization which derives hatred from hate. Both of these words haveinternal themes: the hatred ofBül, I hate Bill. In the case of the verb, thetheme is realized äs a direct object, but in the case of the noun, in a PPof.Why this difference? Now, a verb can have its theme in the direct object,or in a PPof (äs in to speak of x), but a noun can have only the latter,never *hatred Bill. But surely this is because the base rules of English donot generate NPs in the complement structure of NP (or perhaps that thecase rules do not assign case there); and since there is only one otherrealization of Theme, PPof, NP must use that one. Adjectives also used PPoffor Theme, and presumably for the same reason äs in the case of NP. Thepoint is that none of this need be specified by the rule which takes hateinto hatred - the difference wfll follow automatically from other differ-ences between NP and VP.



91

In the above analysis of the nominal of, we make no recourse to an ofinsertion transformation, so popul r in studies of nominalizations. We haveseveral reasons for not doing so. First, in nonthematic objects of V, we donot get of in the corresponding nominal:

(19) John's promise r of º Pete to leave.Uo J

We know that the direct object of promise is not a theme because it doesnot control the complement clause (see Williams (1980) ) and because itdoes not Heavy NP Shift:

(20) *John promised to leave all of the people that were there.(cf. * John promised the books all of the people who were there.)

A theory with of insertion does not predict this; the scheme we have out-lined earlier does though—only Themes can be realized in PPof.

In sum then, the fact that nominalizations of verbs with direct objectthemes always have PPof themes will follow from features of the base (orof case theory, which might account for this difference in the base) andneed not be specified in morphological rule.

Subcategorizations of moöhologßcally related words differ s well. Forexample, readable differs from read, from which it is derived:

(21) read: _NPreadable: .

Again we can ask whether this difference is specified in the morphologicalrule which relates these two. Hopefully, it will turn out that it is notnecessary to specify it for each rule; hopefully, it will follow from anexamination of those differences which are assigned to related lexicalitems by moöhological rule. We will return to this question in Section 6.

3.2. TheRules

Postponing until later further questions about what morphological rulesneed not specify, we turn now to a delineation of that part of the rulesof morphology which is specified in each rule and which concerns themapping of the input argument structure into the output argument struc-ture. Our claim is that morphological rules can alter argument structurein only two ways. To preview, a rule can efther (a) externalize an internalargument, or (b) internalize the external argument. We will symbolizethese s:



92

(22) E(X),I(X)

where X is the name of one of the thematic relations. We wfll define theseoperations in 3.2.1 and 3.2.2 respectively.

These two ftmctions specify only the change of argument structurebetween the input and output of a morphological rule. Obvioudy, mor-phological rules specify other changes; for example, the phonological formof the output (in terms of the input), the category of the output, etc.These other aspects of morphology will not concern us here. We mayassume any of a number of different views on these matters (see Aronoff(1976); WiUiams (1981); or Allen (1978) ); the choice will not matter forour present purposes.

Whether a morphological rule is a rule of affixation or whatever, it hasan input stem or word and it defines an output word. The input word orstem has an argument structure, and the morphological rule specifies theOperation on that argument structure that will derive the argument struc-ture of the output word. We are claiming that the Operation must be oneof the two mentioned above. Clearly, the argument structure of the outputof a morphological rule must conform to whatever well-formednessconditions apply to the argument structures of words in general, or ofwords in that language, or of words in the category of the output word,etc. For example, if the output of the rule is Vs, and if there is an Actorin the output, then the Actor must be the external argument if there isone; this, because there is a universal law to this effect, and this law, likeall laws, holds for both mo hologically complex and morphologicallysimple words. Similarly, if the output is a noun, then the external argu-ment will be R, unless the rule specifies otherwise. All this is äs expected,if these conditions apply to words.

3.2.1. E(X) - "Externalize an Argument": E(X) is an Operation (or rathera set of operations parameterized by the set of thematic relations) onargument structures, which yields new argument structures. The Operationis defined äs follöws:

(23) E(X): erase the underline on the external argument, if there is one,and underline X. If X=0, then underline nothing.

This Operation takes an internal argument and makes it external, where Xis the intemal argument that is to be made external. If X is not specifiedin some particular instance of this rule, then nothing is made external, andso we derive words with zero extemal arguments. If X is specified in someinstance of this rule, and if some word does not have an X, then the ruleis not defined for that word.



93

We will now look at several examples of E(X) rules: One instantiationof E(X) is E(theme), or "Extemalize the theme". Two examples of thisare the rule which adds -able to form adjectives, and the rule which createsthe adjectival past participles. Given a form like read with argumentstructure (A, Th), and a rule of morphology which (a) adds -able to thestem, and (b) applies E(Th) to the argument structure, we will derivereadable, which has argument structure (A, Th). This can be symbolizedäs:

(24) E(Th): read (A, Th) —>readable (A, Th)

This correctly represents the meaning ofreadable, and its relation to read:the subject of readable corresponds to the direct object of read. Further-more, äs Wasow (1977) has pointed out, -able can be added only when theTheme is externalized (to use my terminology, not his):

(25) Those things are | promisable l (theme externalized).jperishable J

*Those people are runnable (Actor externalized).Those people are promisable (Goal externalized).

In Williams (1979), Aronoff (1972) and Wasow (1977), it is shown thatthere are two passives in English - roughly, an adjectival passive and averbal passive. These two are most clearly distinguished in Wasow (1978).The adjectival passive permits the attachment of -un (which attaches onlyto adjectives), it can be fronted by WH-movement (äs all adjectives, and noverbs, can) and, most importantly for our purposes, the adjectival passivealways has a Theme for its subject, whereas the verbal passive can haveTheme, Goal, or Source for its subject. The two passives are illustratedbelow:

(26) Adjectival:The rules are ungiven (Theme of give is subject of given).

*We are ungiven (the rules) (Goal ofgive is subject of given).How firmly promised are these things (Theme ofpromise is

subject of promised).*How firmly promised were those people (Goal ofpromise is

subject of promised).Verbal:

Those people were promised these books.These books were promised to those people.

/e will give an analysis of these two passives in terms of the theory ofBrought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


94

morphology developed here. We will assume what is argued for in Williams(1979); namely, that the verbal passive involves NP movement, the classical,or EST assumption concerning the verbal passive, and we will assume thatthe adjectival passive does not involve trace hinding. The goal, of course,will be to try to explain how this is so, and why it must be the way it is.This goal cannot be achieved solely in terms of the theory of this paper,but in conjunction with the theory of case, we can go fairly far towardsthis goal.

To begin with the adjectival passive, note its similarity to -able affixa-tion. In both cases we derive an adjective, and in both cases we havethematic constancy: the Theme must be the subject of the output word.From this we conclude that the adjectival past participle is derived byE(Th), just äs adjectives in -able are. The adjectival passive, then, has nointeresting syntactic component to its description - no NP movement, forexample — beyond the fact that the adjectival past participle is an adjec-tive; to syntax, it is just another adjective.

The verbal passives, on the other hand, cannot be derived by anymorphological rule (alone), however, since the verbal passives do notexhibit thematic constancy; that is, the subjects of verbal passives are notuniformly of one thematic type (we have seen that the verbal passive canhave at least Theme and Goal subjects). Thus the verbal passive cannot beanalyzed äs E(X), for any particular X.

On the other hand, though, the argument structure of the verbalpassive cannot be simply identical to the active either, for we would thenhave no account of the difference between the active and the passive. Inthe following, we will entertain the idea that the verbal passive participleis formed by the Operation of E(0) (or just E( ) ), which we may read äs:

(27) E(0): erase the underlining in the argument structure giving anargument structure with no external argument.

In other words, the verbal passive has no external argument. There is al-ready a class of underived words with this property, äs we noted earlier:the verbs seem, appear, etc. So E(0) simply maps verbs into this class. (See(5) ). Unfortunately, this proposal makes quite peculiar predictions for theverbal passive; it predicts, for example, the grammaticality of the following:

(28) *It was seen Bill by Fred.

E(0): see > seen(A,Th) (A, 1h)

Here, all of the arguments of the verbal past participle are internal to theBrought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


95

VP, and the dummy fills the subject, with which no argument is associated.Now, in syntax, there is a rule of NP movement, and if we apply it to thisexample, we can derive the grammatical:

(29) Billj was seen t{ by Fred.(A,Th)

Here, all of the arguments ofseen (t{ and Fred) are again internal; äs in thecase of seem raising, we must distinguish between external occurrence ofan argument and the binding of an internal argument by an external NP.1

The problem posed by the last two examples is this: while applyingE(0) to derive the verbal past participle will permit movement to thesubject position because it thematically "frees up" the subject position,it does not force movement to take place, but it is only when movementtakes place that we get grammatical verbal passives. The problem is not sobad when one notices that movement (of any kind) is not an essentialfeature of passive in any case; we have the following passives in which nomovement has taken place:

(30) (a) Es wurde getanzt."It was danced"

(b) ete discute de beaucoup de choses."It was discussed about many things".

(c) It was [reasoned that Bill had left.Jyp

In all of these cases we have a passive with no movement and a dummysubject. What they have in common is that there is no NP to move. Recentproposals in the theory of case suggest that movement is forced to takeplace when an NP is in a non-case-marking position - the NP mustmove to a case-marking position in order to get case. Thus if the verbalpassive participle does not assign case, direct objects NPs will be forcedto move. We may, therefore, view the verbal passive äs having two sig-nificant features: it is the Output of E(0), which leaves it with no externalargument (thus permitting movement to subject position) and it does notassign case, thus forcing movement, when there is an NP object. In thecases just given, we see that the subject position can be freed up thematic-ally (by E(0) ) and no movement takes place, since there is no NP object.Thus these two features of passive are really independent of one another.

The examples (30a) and (30c) require some further discussion. The(30a) case shows that English and German are decidedly different: theEnglish translation is ungrammatical, and we have no explanation for thisdifference. We might suppose that the difference is simply that the Englishpassive participle can be added only to verbs that are transitive - that is,



96

verbs which have either an NP or S complement. When it has an NP com-plement, movement will always be forced, for the reasons just discussed(here and elsewhere); when it has an S complement, movement is notforced, äs in (30c). The example (30c) cannot be analyzed äs the resultof applying extraposition to a structure to which NP movement hasapplied, since the intermediate structure is not grammatical:

(31) *That Bill had left was reasoned.

NP movement cannot take place in such examples, because the verb reason(like certain other verbs, such äs hint and/ee/) do not take NP objects, andthe taking of NP objects is prerequisite for this construction:

(32) *We reasoned it (*That Bill left was reasoned).*We hinted it (*That Bill left was hinted).We hinted at it (That Bill left was hinted at).We believed it (That Bill left was believed).

See Williams (1980) and Koster (1978) for further discussion.We may summarize the difference between the two passives in the

following way: The adjectival passive participle is formed from verbs byE(Th), and the output from is an adjective. The verbal past participle isformed from verbs by E(0), and the output form is a verb with the samekind of argument structure of the seems type verb. Thus the two formsdiffer lexically. They also differ syntactically — the adjectival past partici-ple has simply the distribution of an adjective, and there is nothing moreto say about it. The verbal past participle has the further property that itfails to assign case, and this has predictable consequences for the kinds ofsyntactic environments in which it is found. The difference between thesetwo passives can be illustrated by a rather nice comparison with the verbbelieve (from Williams (1979) ). We find both adjectival and verbal passiveof the verb believe. The passive of the verb believe that is "raising" canonly be verbal, because the raised NP bears no thematic relation to believe(much less a constant one):

(33) John believes there to have been a riot.Therej is believed [t{ to have been a riot.]

Since it must be verbal, due to the failure of thematic constancy, thefollowing are predicted:

(34) *How widely believed is John to have left.



97

Here, the passive is not an instance of E(Th), so the participle must beverbal, but if it is verbal and not adjectival, it cannot be fronted by WHmovement. On the other band, there are passives of believe which arethematically constant, and therefore these can be, and in fact are, adjec-tival:

(35) That story is widely believed (story is the theme ei believed).It is widely believed that Bill was here (that Bill was here is the

theme of believed)

If these are adjectival, then we predict the grammaticality of:

(36) How widely believed is that story.How widely believed is it that Bill was here.

These contrasts ftirther confirm the correlation of E(Th) with the adjec-tival past participle and of E(0) and NP movement with the verbal pastparticiple.

We might consider the question of whether this correlation isaccidental.These facts with believe are rather subtle (they have never been reported inthe literature, to my knowledge) and one would thus hope that knowledgeof them would not be required to settle questions of English morphology.To see that this correlation is not accidental, consider what would happenif we reversed the correlation - imagine E', just like E(nglish) except thatin E' the thematically constant passive is verbal, and the thematicallyinconstant passive (the Output of E(0) ) is adjectival. Is E' grossly differ-ent from E, or is it subtly different from E?

Certain considerations lead one to suppose that it might be grosslydifferent from E. Suppose that adjectives do not have NP positions inthem but only PP positions, and suppose further that every P is a caseassigner. These are certainly reasonable assumptions. It would follow thatAP would have no bindable positions, if NP movement trace must occupycaseless positions. Thus the adjectival passive could involve no NP move-ment, except out of embedded Ss. Thus we would have adjectival passiveslike the following:

(37) ItwashitA of John.Joluij was believedA t{ to have left.

The last possibility exists in E, äs well äs E', with such adjectives äs certain,äs in:

(38) John is certainA t{ to have left.Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


98

E' will not have, however, such passives äs:

(39) John was hitA (by Bill).

Furthermore, there would be no verbal passives, because it is impossible tohave anything external in the category V except Actor, when Actor is amember of the argument structure. This principle was needed, it will berecalled, to explain why Actor is always external in active verbs (and topermit such cases äs seem (where nothing is external, and there is noActor). Thus, the passives in E' would be quite an odd assortment - therewould be practically none of the normal passives that are found in E. E',then, even if it is an allowable language, which it is according to ourtheory, would be extromely different in gross surface form from E andeasily distinghuished from E.

Given the government theory, we predict further differences between Eand E'. According to the government theory (see, for example, Chomsky(1981) and Rouveret and Vergnaud (1980), empty nodes must be"governed"; that is, "minimally c-commanded " by lexical nodes. Giventhat the English passives are of two kinds, an adjective with one externalargument, and a V with no external argument (and NP movement), wepredict the following surface structures:

(40) (a) John is [knowny t^ have left]yp(b) It is [known^]^y> that John left.

In (a) the complement is inside the VP, äs it must be, in order for the traceto be governed by the verb; in (b), the complement is outside of the AP,äs it must be in order to be the external argument of the head of the AP.Given this difference in surface structure, which can be inferred from acombination of universal principles and rather gross properties of E, wepredict the following differential wrt to the action of VP deletion:

(41) (a) *John is known to have left, but he isn't to have gone to themovies.

(b) It is known that John left, but it isn't that he went to themovies.

In the (a) sentence, only a subpart of a VP has been deleted, whereas inthe (b) sentence, a whole phrase, the AP, has been deleted. As was shownin Williams (1977) it is ungrammatical to delete a subpart of a VP. Again,it seems that a rather obscure fact follows from a combination of universalprinciples and gross properties of the English passive.



99

3.2.2 É (×) - "Internalize the external argument": The rules E(X) kept thenumber of arguments in the input and the output the same. The rulesI(X), on the other hand, add one argument to the input argument struc-ture. Since these two rules exhaust the possibilities, we predict that no ruleof morphology can shorten argument structure. In Section 4 we willdefend this.

I(X) is for causativization and for nominalization. Although morecomplex than E(X) to specify, it admits of no more descriptive freedom.Its action on an argument structure is s follows:

(42) I(X):(a) Set the external argument of the input word "equal to"

X in the output word;(b) Add a new external argument, A for verbs, R for nouns.

As an example, consider the rule which adds -ize to words. This is anI(X) rule - in particular, I(Th). Suppose we apply this nile to random,which has argument structure (Th):

(43) I(Th): randomA—> randomy (A, Th=Th)

The argument structure for randomize says that it has a new externalargument, A, and that its theme, also new, "corresponds" to the themeof the input word. This last specification serves to capture the fact thatthe theme of randomize is the thing that becomes random.

The rule which adds causative en is also I(Th), and derives words withargument structures just like -ize. Also, the rule which relates intransitivemelt to transitive melt (The ice melted; John melted the ice) if this is arule of morphology, is another instance of I(Th), and derives transitivesin the same way that -ize does.

We may now turn to a brief sketch of the Japanese causative System.In Japanese, causative verbs are formed lexically from simple verbs byadding the suffix -sase onto simple verbs. We shall see that the resultingverb has an argument structure determined by applying I(Th) or I(G)to the original verb. (The Situation is slightly more complicated in fact- in certain circumstances I(G) forms permissives, rather than causatives,but for our present purposes, this distinction is irrelevant.)

As a background, let us assume that Japanese has the following realiza-tion rules:

(44) Theme: PP0Goal: PPni



100

That is, Themes are located in "o" postpositional phrases, and Goals in"ËÉ" postpositional phrases.

When -sase is added to an intransitive verb, either I(Th) or I(G) mayapply (with the difference in meaning noted above):

(45) I(Th): V(A) - > Vsase (A, Th=A)

If I(Th) applies, we will get sentences like:

(46) John ga Bill o Vsasemasu."John made BU1 V"

If I(G) applies, we will get sentences like:

(47) John ga Bill ni Vsasemasu ."John let Bill V"

When -sase is added to a transitive verb, that is, a verb which takes an "o"postpositional phrase, then only I(G) can apply, not I(Th). There aretwo ways we could account for this. First, we might point to the fact thatthe result of applying I(Th) to an argument structure which has a Themealready would result in two Themes:

(48) I(Th): (A, Th) -» (A, Th, Th=A)

Now, we may ask, is "Th" the same thematic relation s "Th=A"? We havebeen quiet on this point thus far. Suppose we say yes — then this will notbe a well-formed argument structure, since we have two instances of thesame thematic relation, and since the items in an argument structure aresimply labels, we cannot have distinct occurrences of the same labels.

But in fact, it appears that causativization by I(G) of a verb which hasa Goal phrase already is permitted, despite the fact that this results in twoGoal phrases. So perhaps after all, we would like to consider "X" and"X=Y" distinct and prohibit the application of I(Th) to transitives by arestriction on the number of "o" phrases that may appear in a surfacestructure clause. This latter type of explanation is similar to that set forthin Aissen (1974), but without her notion of "clause union".

Nominalizations and causatives may seem to be an odd natural class,but it seems that nominalizations are also instances of I(X) rules. When averb or an adjective is nominalized, its external argument is made an inter-nal argument of the noun. We will examine two cases of this in English,I(A)andI(Th).



101

In the formation of -ing nominals, either I(Th) or I(G) may apply.Recall that the (at least unmarked) external argument for nouns is R:

(49) I(Th): growly (A) -> growlingN (R, Th=A) "the growling of thelions"I(A): growly (A) -+ growlingj^ (R, A=A) "the growling by thelions"

As in the case of the Japanese causative, if we have a transitive verb withtheme object, then we cannot apply I(Th), but only I(A):

(50) the shooting of the lionsj by the hunters ll *of the hunters J

This example perhaps suggests the answer to the question raised earlierabout the Japanese causative. If we assume that shooting is always transi-tive, then we can derive the two argument structures:

(51) (A, Th)^ (R, Th, Th=A) (by I(Tli) )( , Th)— (R, Th, A=A) (by I(A) )

In fact both of these can be found for the nominalization ofshoot (shoot-ing\.

(52) the shooting of the huntersthe shooting of the lions by the hunters

U seems that we do want to prohibit the structure (R, Th, Th=A), wesimply want to prevent two o/phrases, each linked to a role of theme, toappear in surface structure. The suffix -ion works in much the same way:either I(A) or I(Th) may apply:

(53) I(Th): the pontification of the Pope (R, Th=A)I(Th): the pontification by the Pope (K, A=A)

(from pontificate (A) )

One prediction unfortunately made by the System thus far represented isthat intransitive verbs with external themes (like the intransitive melf)should be able to undergo I(A), which they cannot:

(54) I(A): melt (Jh) -> melting (R, A=Th) "the melting by the snow"

I have no particular insight äs to why this should be so, but I do have aBrought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


102

couple of suggestions. Suppose that we have a hierarchy of thematicrelations, äs proposed in Jackendoff (1972), which runs: Agent, Theme,Goal, ... and suppose that we stipulate that in the "X=Y" Statementsderived by I(X) that X cannot be higher than Z on the hierarchy. This willprohibit the above derivation of melting, but it will permit all otherderivations that we have considered.

We might take a somewhat different tack. Suppose that the normal casefor nominalizations is 1(0), where this is interpreted to be:

(55) 1(0): (a) add a new external argument (the appropriate one forthe category created; R for nominals, etc.)

Since the new argument is external, the external argument of the input ismade internal, by automatic convention (äs in the case of E(X)). Aproblem could arise if the new external argument was the same äs one ofthe old arguments, since we would then wind up with identical arguments.This could not arise for deverbal nouns, since the external argument addedto make a noun is R, and R is never a verbal argument. But the problemwould arise for denominal nouns (e.g., tableness).

Let us assume that the normal rule for nominalizations is 1(0) and thatthe problems mentioned with it above are solved. Then, the expected caseswfllbe:

(56) the melting of the snowthe shooting by the hunters

since in these cases the argument types are preserved: snow is the Themeof melting, just äs it was the Theme of melt; hunters is the Actor ofshooting just äs it was the Actor ofshoot. From this point of view, one ofcourse would not expect to find such cases äs "the melting by the snow"since in such cases the Theme of the input (melt) has become the Actorof the Output (melting). That is a good result, but one would also notexpect the form, the shooting of the hunters, which is well-formed, sincehere äs well the thematic roles differ in the input and Output. For suchcases one could either apply I(Th) äs a special "marked" alternative, or,one could say that PPof was a special "marked" realization of Actor in NP.I will leave the problem äs it Stands.

1(0) äs the unmarked case for nominalizations does well by adjectives,though, if Theme is the external argument for adjectives. Consider thesuffix -ness, for example, äs an example of 1(0) (or perhaps I(Th) ):

(57) 1(0): complete (Th) —*> completeness (R, Th)The completeness of the results.



103

3.3. Concluding Remarks

To summarize the point of this section: once you divorce the rules whichdetermine the syntactic realizations of various argument types from therules which relate the argument structures of morphologically relatedwords, then this latter class of rules can be stated in a quite restrictiveway. What these rules seem to boil down to is that you can affect theexternal argument of the input word in one of two ways: you can makeone of the internal arguments into the external argument, with a con-sequent demoting of the old external argument to internal Status (E(X) ),or you can add on a new external argument, again with consequent demot-ing of the old external argument to internal Status (I(X) ).

It is important to emphasize that grammatical relations play no role inthis System, or, to the extent that they do, there are only two: subjectand nonsubject (where I use internal and external instead). In this Systemoutlined here, thematic roles are tied directly to syntactic realizations, bythe realization rules. In a theory with grammatical relations, this relationis mediated by the set of entities called grammatical relations, and thereare two sets of realization rules: one set which says which thematic rolesare realized äs which grammatical relations and another which says whichgrammatical relations are realized in which syntactic positions. Since thereare no grammatical relations in the System presented here, there can ofcourse be n o rules, syntactic or lexical, which operate on representationswhich contain or are composed in terms of grammatical relations. Thereinlies part of the restrictiveness of the System presented here.

Another portion of the restrictiveness of the theory presented here liesin the Substantive claim that morphological rules can affect argumentstructure only by affecting the external argument (I(X) or E(X) ). This isfar from obviously true; in the next section, we will defend this view fromone kind of attack on it, the case of lexical "detransitivization" rules.

In fact, any putative lexical rule which does not affect the externalargument of a lexical item cannot be a lexical rule, according to the theorypresented in this paper. Proposals for such rules are quite common. A goodexample is the rule of dative movement in English, which, it is argued bysome, cannot be a transformation - and thus must be a lexical rule (seeOehrle (1976) ). We may agree with Oehrle that this rule is not a trans-formational rule, but we need not conclude that it is a lexical rule - infact there is no need for it to be a rule at all. First of all, it cannot be alexical rule, because it cannot be an instance of either E(X) or I(X), sincethe external argument of the input (no matter which way the rule runs)is not affected. Second, it need not be a lexical rule, if we consider thefoUowing line of thought: Suppose English has two ways of realizing Goal,äs outlined in Section 3.2.1 — äs NP2 and äs PPto. Any verb with a Goal



104

internal argument must subcategorize for one of these two possibilities, orfor both. In the case of verbs that use both realizations, there is no need tosay anything other than that they use both — there is no need, in partic-ular, to consider such verbs to be two verbs and to relate these two verbsby lexicalrule.

It is telling that there is an affix (-en) for passive, but no affix for dativemovement. Why is this? According to the theory presented here, therecould be no such affix äs a "dative movement" affix, because there can beno such lexical rule, and this is because the relation that would hold be-tween the argument structures of the forms related by the rule could notbe stated (is neither I(X) nor E(X) ).

A rather strong (embarrassingly, I am sure) prediction is made: languagescan have all sorts of passives, with affixes - for example, a passive whichmoves locatives and themes and adds a different affix for each of these -and these would be cases of E(X). But no language could have a lexicalrule like English dative movement, nor, of course, an affix associated withsuch a rule, since such rules do not affect external arguments.

4. LEXICAL DETRANSITIVIZATION VS OPTIONAL SUBCATEGORIZATION

There are two ways one might represent the relation between transitiveeat and intransitive eat. One might use the notation of optional subcate-gorization and write "eat, (NP)." Or one might have two eats onetransitive and one intransitive, related by a lexical rule of detransitivization:

(58) V[+obj] *V[-obj]

For simple cases, it is difficult to distinguish these empirically. But forsome more complicated cases, a difference in prediction can be derived.

The transitive-intransitive relation is sporadic — not all transitives havecorresponding intransitives. Thus put and attempt do not, while eat andhope do. In the subcategorization theory, this means that some frames willhave parentheses in them, and some will not. In the lexical detransitiviza-tion theory, this means that the rule of detransitivization is not completelygeneral. This consideration is certainly not decisive, since lexical rules arein general not completely general.

Put is a case where two things are obligatorily present, an NP and a PP.Position, on the other hand, whfle it has two arguments in the VP, hasonly one obligatory argument. And promise, which has two arguments äswell, has both of them optional. The subcategorization theory predictsthat four different cases will be found for verbs with two arguments in theVP:



105

(59) XY one expansion: X Õ put(×)Õ two expansions: X Õ, Õ make (see below)× (Õ) two expansions: × Õ, × position(×)(Õ)four expansions: × Õ, ×, Õ, 0 promise

All of these are attested, except for the second. As a case of the second,we might consider the dative that occurs in such cases s:

(60) Make Bill a cake.

This dative is always optional and precedes a direct object.Now the subcategorization theory predicts that a verb with two argu-

ments in the VP will exhibit one of these four "expansion sets" but noother. There are other expansion sets, however, besides the ones predictedto occur by the subcategorization theory. For example, the expansion sets

(61) X,Y,0X Y,X,0

are conceivable but are predicted not to exist by the subcategorizationtheory. In the case of the second, for example, in order for 0 to belongto the expansion set, both X and Õ must be made optional; but if bothX and Õ are optional, this implies that Õ will belong to the expansion set.The lexical detransitivization rule, however, makes no such prediction.Consider the hypothetical lexical item M with two arguments in the VP,X and Y. First we derive M' by applying lexical detransitivization toeliminate the Õ argument ( s we might to derive "position NP" from"position NP PP"). Then, we apply lexical detransitivization again, toderive M", just like M', except that now the X argument is removed:

(62) M >M' >M"× Õ × 0

We see that this triple (M, M', M") exhibits an expansion set which wesaw was predicted not to occur by the subcategorization theory. Ingeneral,the lexical detransitivization theory predicts a wider r nge of possibilitiesthan the subcategorization theory. To the extent that exactly the r nge ofcases predicted by the subcategorization theory is found, and no others,that theory is to be preferred.

Note that it will not do to object that the diagram above is "incomplete"in the sence that lexical detransitivization has not applied to M to deriveM'", which has only Õ s an argument. This objection is pointless becausewe know from the Start that the transitive-intransitive relation is sporadic,and the only real question is what is the r nge of possibilities.Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


106

More is at stake than whether the transitive-intransitive relation shouldbe done by one mechanism or another. The theory of lexical rule is atstake. I have proposed in this paper a theory of lexical rule that will notcountenance such a rule äs lexical detransitivJzation, for the simple reasonthat such a rule does not affect the external argument of words (i.e., it isnot an instance of either I(X) or E(X) ). Such a role would affect only thearguments internal to a VP. Thus it is encouraging to find that such a rulehas other faults äs well when compared with reasonable alternatives. Tosummarize, the rule of lexical detransitivization has two significant faults:it predicts that a broader ränge of "expansion sets" will be attested thanthe subcategorization theory, and it forces the abandonment of anattrac-tively restrictive theory of lexical rule.

5. THE FRENCH MIDDLE CONSTRUCTION

The theory of morphology outlined here, in conjunction with the theoryof predication in Williams (1980) affords some explanation of aspects ofthe French middle construction, with some help from recent ideas of casetheory. The French middle is a construction in which (a) the normalobject appears in subject position, (b) the verb is in the active form, andagrees with the subject; and (c) the reflexive pronoun se is attached to theverb. The following is an example:

(63) Ce livre se vendThis book sells itself

We will proceed with our discussion of the middle, first by assuming acertain representation for the middle, and then considering why it has tohave that representation.

Suppose that there is a lexical rule for forming middle verbs. The ruledoes not alter the verb morphologically; it does, however, alter a verb 'sargument structure by the function E(0). Now suppose that we have averb like vend with argument structure (A, Th). When we apply this rule,we will get vend (A, Th). Next, suppose that we generate this verb with adirect object reflexive clitic and a subject, äs follows:

(64)0

The "0" under vend is simply to indicate that the verb is a 0-externalargument verb.

When the rule of clitic movement and the rule of predication (seeWilliams (1980) ) have applied to this form, we will derive:



107

(65) NPitsej0

This representation seems to capture the basic facts about the middleconstruction. The surface subject is interpreted äs the direct object (here,it winds up coindexed with it) and the verb bears a reflexive clitic. Fur-thermore, the superficial subject has the thematic relation of the object,since it is coindexed with it, and the subject position itself receives nothematic relation, since the verb form was derived by E(0), and thus hasno external arguments.

Given the deep structure described above, we can derive a reasonableaccount of the middle. But why this deep structure? In other words, whycan't the middle verb (derived by E(0) ) appear in other kinds of deepstructures? If we can answer this question, then we can be said to haveexplained some properties of the middle construction.

First, why is a reflexive clitic involved? And why does it wind up coin-dexed with both the trace and the NP subject? To answer the second ques-tion first : the reflexive is coindexed with the trace by movement - all cHticsmust move, by definition; the clitic must be coindexed with the W thatcontains it by the requirement stated in Williams (1980) (äs a variant ofthe SSC) that a reflexive cannot be free in a VP (and where being coindex-ed with the VP counts äs being bound within the VP); the VP and thesubject are coindexed by the rule of predication; therefore, by the transiti-vity of coindexing, the subject wfll be coindexed with the clitic and thetrace of clitic movement.

Now for the first question: why is a reflexive clitic involved? Thisquestion divides into two parts. First, why can't there be a nonreflexiveclitic, or ordinary lexical NP, and second, why can't there be no clitic -i.e., why can't NP movement move a direct object of a middle verb direct-ly into subject position.

The reason a nonreflexive clitic cannot be used is this: again, a non-reflexive clitic must move (since it is a clitic); furthermore, a nonreflexiveclitic must not be coindexed with the VP that immediately contains it(again, a variant of the SSC, äs given in Williams (1980)); hence, we willget the structure:

(66)0

Here, the subject NP is not coindexed with the object trace. Since it isnot coindexed with anything that is an argument of vend, and since itcannot be interpreted äs the external argument of vend, since that verbis in the middle and thus has no external argument, then the subject NP



108

can receive no Interpretation with respect to vend, and this is why thesentence is bad. Thus no nonreflexive clitic can be used. A similar ex-planation can be given of why a füll NP cannot be used in place of thereflexive clitic.

Now, why must there be any clitic at all? Why not derive it by NPmovement, from subject to object position, giving the representation:

(67) N0

Here, the NP subject will receive a thematic Interpretation, by virtueof bin ding the trace, which occupies an internal argument position. Sothe theory of argument structure is not really going to explain this.

But this is where the theory of case Steps in. By current theorizing oncase (see, for example, Chomsky (1981) ), case must not be assigned to thetrace of NP movement. If the active form of the verb assigns case, thenthe trace above will receive case, thus ruling out this example. The pastparticiple of the passive construction, on the other hand, does not assigncase, and so NP movement is permitted in the derivation of passives —the trace left behind will not get case in that construction.

A bit more needs to be said, because the stipulation that traces cannotget case, or be in case-marking positions, will also rule out the actualmiddle construction, since it has a trace in it. The crucial property of thetrace in the middle construction is that it is a clitic trace and not an NPtrace. Clearly, we want to say that the clitic trace is exempt from thecase-marking restriction that the NP movement trace is subject to. Wemight do this in a number of ways, but, however it is done, the solutionwill not apply simply to the middle, but to all uses of clitics in all con-structions, and so the solution will not be an ad hoc answer to our currentconcern; that is, the same question arises for the use of non-reflexiveclitics in non-middle constructions:

(68) Jean [le{ mange t{]

One way to do this is to say that clitic movement takes place after casemarking. The exact mechanism does not concern us here, though, oncewe have noted that the problem immediately generalizes to all clitics inall uses.

From two assumptions - that middle verbs are formed by E(0) andthat active verbs are case assigners - we have explained a number of factsabout the middle construction: why it has a clitic, why that clitic is re-flexive, why the subject is interpreted äs the object argument.2 I regardthis äs a fair attempt at explanation, but there are some problems, and



109

some further speculations. Some of these arise with the English middleconstruction.

English has two middle constructions. One of them is just like that inFrench, except of course English does not have clitics. Suppose that thereis a middle formation rule in English äs well, and that this rule is an E(0)rule. Then in English, we can derive middles like:

(69) This book sells itself

which will have the representation:

(70) This boolq [sells itselfj yp0 !

This is exactly like the French middle, and its properties can be explainedin exactly the same way äs the French middle: the reason a reflexive isinvolved is that otherwise, the subject NP could not be coindexed withanything in the W, and if it were not, it could not receive any thematicInterpretation (the verb being a/) external argument verb).3

This is an interesting construction from the point of view of the "thetacriterion". In this construction, an NP gets a thematic relation by virtueof being coindexed with a reflexive pronoun which occupies an internalargument position. In the more ordinary use of reflexives, the reflexiveand its antecedent are independently assigned thematic roles, and it isusually thought that it is only through mövement that an NP will get itsthematic role by virtue of what it is coindexed with. These cases refutesuch a view. Furthermore, the facts of such cases fall out automatically —in fact the same äs in French, modulo the clitic/nonclitic distinction whichdivides these languages independently. One concludes from such examplesthat the faflure of case assignment to object position does not always attendthe application of E(0). These are independent. Problems arise whenwe consider another English middle construction, which differs from theEnglish reflexive middle construction and the French middle constructionin not having a direct object reflexive:

(71) These books seil well.

There is no problem in saying that these cases are the result of applyingE(0) to verbs; the problem is that we will have to use NP mövement toderive the appropriate surface forms:

(72) These books. [seil t{ well.] yp0 '



110

The problem with this representation has to do with case assignment: ifactive verbs must assign case, then we have no explanation of why thisform is permitted to have a NP movement trace after the active verb. Wemight suppose that there is a special rule which removes the case-assigningproperty from these verbs, despite the fact that they remain active verbs.This would weaken our explanation of why the French middle used areflexive clitic instead of NP movement, but in fact we now need someway to distinguish this middle from both the French middle and theEnglish reflexive middle, and case assignment could be the way to do that.

In sum, then, the description of the middle involves two differenttheories. First, there is the theory of argument structure. As far äs thattheory goes, the middles are all E(0). Second, there is the theory of case.If the middle verb is a case assigner, then we will get a reflexive object;if not, NP movement. In either of the latter two cases, the surface subjectgets its thematic relation by virtue of being coindexed with somethinginside of the VP. The one further difference between the French middleand the English reflexive middle — namely, that the reflexive moves inFrench, but not in English — follows from the theory of clitics, and thefact that French has them, but English does not. One virtue of the analysispresented here is that the reflexive clitic in the middle construction hasthe properties of any clitic in any other construction, and there is no needto replicate the basic property of clitics ("they move in front of verbs")inside the theory of the middle. Nor does the reflexiveness of the re-flexives in the two constructions require any special treatment: in the caseof the middle, äs in the case of any other use of the reflexive, we needonly say that bound anaphors are subject to certain opacity conditions.Thus a good deal of the description of the middle reduces to nothingspecial: the theory of bound anaphora and the theory of clitics. In notevery imaginable treatment of the middle, however, is this reduction pos-sible.

6. SUBCATEGOR1ZATION

We have ignored the question of subcategorization so far, except to rely onit for explanations that were not supplied by other parts of the theory. Wehave assumed more or less theAspects theory of subcategorization, wherea head of a phrase can determine the category of the phrases that appear inthat phrase with that head. We will now look more closely at those pro-posals about subcategorization which will be consonant with the theo-rizing we have done thus far.

Of course, the application of a morphological rule changes subcategori-zation, but the question is, do we need a set of rules for specifying thechanges that can be effected in subcategorization? The best answer is no -



111

that any changes in subcategorization will follow from other changesinduced by the application of a rule, most likely the changes in argumentstructure and in case-assigning properties. In this section, we will strive forthis best answer; in particular, we will outline a theory by which theargument structure and the category of the derived word will determinethe subcategorization of the word, thus avoiding the need to specify thesubcategorization of the Output of a morphological rule.

The theory assumes a "markedness" analysis, or ranking, of the reali-zation rules. As can be seen by inspecting (18), there is potentially morethan one realization rule for each argument type. Thus, for example,Goal can be realized äs a NP2 or äs the NP in a PPto. If a particular verbhas a Goal argument, it must subcategorize for one or the other or both ofthese options.

Now suppose that we say that the realizations for any particular argu-ment type are ranked, and that for Goal in English, PPto is first, andNP2 is second. One use we might make of this is to predict that if a newmonomorphemic verb were invented, and that verb had a goal, then itwould have the PPto subcategorization, and not the NP2 subcategoriza-tion. We might also expect that a new morphologically complex wordwould have an unmarked realization of Goal, if it had a goal; thus, wemight automatically supply new words with unmarked realization possibil-ities, regardless of the source.

Other factors can prevent the manifestation of the unmarked case. Forexample, of the two ways to realize Theme, äs NP or äs the NP of PPof (äsin think of) surely the bare NP is the unmarked case. But in the comple-ment structure of NP, bare NP is not allowed at all, presumably becausecase cannot be assigned to it there (N does not assign case) and so wefind only the marked realization of NP.

There remain several further questions to which we can give only tenta-tive answers. For example, we know that subcategorization is sometimesobligatory ( NP) and sometimes optional ( (NP) ). Do morphologicalrules affect the obligatory/optional Status of Output words? The suffix-ize, for example, seems to create obligatorily transitive verbs: one says,John rubberized the raft, but not *John rubberized. Again, we mightsuppose that " NP" is the unmarked choice for internal theme, and" (NP)" is the marked choice.

A further question is, what Information is subcategorization sensitiveto? For example, is it sensitive to case? At first glance it appears to be; inGerman, for example, certain arguments are sometimes instantiated äsgenetive NPs for certain verbs, and this is idiosyncratic; presumably, verbsmust be subcategorized for such cases. And there is some indication thatin English, subcategorization must refer to heads of PPs. A perhaps crucialpair of cases is give and present, both of which have goals, and both ofwhich have " NP PP" subcategorization:Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)


112

(73) I presented it to him.I presented him with it.I gave it to him.

But, if all we say about give and present is that, we predict:

(74) *I gave him with it.

There are two Solutions: we can either distinguish "NP" and(which we earlier called "NP2") or we can distinguish PPto and IIn the former case, we will give the subcategorizations in (a) below, andin the latter, the ones in (b):

(75) a. Present: _NP2 PP Give: _NPPP—NPPP _NP2NP

b.Present: _NPPPwith Give: _NPPPto_NPPPto _NP2NP

It is unclear whether such Solutions will be available generally, or whetherthey should be allowed generally.

A possible objection to the above Solutions, which allows the subcate-gorization restrictions to "look" at the heads of phrases, is that this iscontrary to the claim made earlier in the paper "that words do not haverealization schemes, languages do" since by careful subcategorization averb can determine uniquely which realizations it will be subject to.While the ability to subcategorize for heads of complements may in factallow a verb to uniquely determine which realizations it will be subject to,this does not reduce the original claim to vacuity, since there will stillonly be a few fixed realizations for any argument type, and only thesewill be available. Thus, no English verb, for example, can determine thatits Goal wfll be realized äs PPfrom-

We have a theory now where subcategorization can refer not only tophrases, but also to heads of phrases. Perhaps we could make the theoryat least sound more restrictive by allowing subcategorization for headsonly, but heads under different descriptions. Thus, instead of PP, wehave P, and by the same token, instead of PPto we have 70, and subcate-gorizations for -bar äs well (or to-double-bar). And S could be referredto äs tense-bar, if tense is the head of S.

NOTESl. Even though Bill is not the subject (or external argument) ofseen in this example,

(i) Billj was [seen tj]



113

(rather, it binds the intemal argument position occupied by trace), we still would liketo say that Bill is the subject of the VP "seen tj" in the sense of Williams (1980). Billis the subject of "seen t" not by virtue of being an argument of the head of "seen t"(i.e. of seen), but rather by virtue of binding the trace. This the predication involvedin the verbal passive is not of the type called "simple headed" in Williams (1980)but rather of what was called there "complex," since the predicate "seen t" is a pre-dicate only because it contains a syntactic variable (t). The füll representation of thissentence in the notation of Williams (1980) would thus be

(ii) Billj was [seen tj] vp.

Raising would work similarly; although Bill is not the subject of seems, since, äsmentioned before, seems has no subject (or external argument), still, Bill is the sub-ject of "seems t to be sick," and using the notation of Williams (1980) we have:

(iii) Billj [seems [tj to be sick] ] yp.

Again, the VP "seems t to be sick" is a complex predicate, that is, a phrase that is apredicate by virtue of containing a syntactic variable.2. It is worth notin g that the only difference between the middle and the ordinaryuse of the reflexive is that in the middle, the verb is a 0-external argument form, butin the ordinary use, the verb is an ordinary l-external argument form:

(i) Middle: Le livre{ [sq vend tj] 4The book sells itself"0

"Ordinary": Jean^ [se^ lave tj] "John washes himself"

In the middle, there is only one thematic position, the object position, and the sub-ject gets its thematic role by virtue of being coindexed with the trace in object Posi-tion. But in the "ordinary" case, there are two thematic positions, the subject andthe object position, and each receives a different thematic role.3. Many such cases are listed in Fiengo (1974). The construction is also quitecommon in Dutch, formed on zichzelf (Riny Huybregts, personal comm.).

REFERENCES

Aissen, J. 1974. "Verb Raising." Linguistic Inquiry, 5. 325-366.Allen, M. 1978. "Morphological Investigations." University of Connecticut disserta-

tion.Aronoff, M. 1972. "Studies in Analogical Pseudosyntax," no. 1. MS.Aronoff, M. 1976. Word Formation in Generative Grammar. MIT Press: Cambridge,

MA.Bresnan, J. 1972. "The Theory of Complementation in English Syntax." MIT dis-

sertation.Bresnan, J. 1978. "A Realistic Transformational Grammar." In Bresnan, Halle, and

Miller, eds., Linguistic Theory and Psychological Reality. MIT Press: Cambridge,MA.

Chomsky, N. l965.Aspects ofthe Theory of Syntax. MIT Press: Cambridge, MA.Chomsky, N. 1980. "On Binding." Linguistic Inquiry, 11. lChomsky, N. 1981. Lectures on Government and Binding. Foris Publications: Dor-

drecht.Fiengo, R. 1974. "Semantic Constraints on Surface Structure." MIT dissertation.Gruber, G. 1976. Lexical Structure in Syntax and Semantics. North Holland: Am-

sterdam. Brought to you by | Chinese University of Hong Kong (Chinese University of Hong Kong)Authenticated | 172.16.1.226


114

Jackendoff, R. 1972. Semantic Interpretation in Generative Grammar. MIT Press:Cambridge, MA.

Kayne, R. 1975. French Syntax. MIT Press: Cambridge, MA.Koster, J. 1978. "Why Subject Sentences Don't Exist." In S.J. Keyser, ed., Recent

Transformational Studies in European Languages, MIT Press: Cambridge, M A.Oehrle, R.T. 1976. "The Grammatical Status of the English Dative Alternation."

MIT dissertation.Rouveret, A. and J.-R. Vergnaud. 1980. "On Specifying Reference to Subject."

Linguistic Inquiry, 11.1.Wasow, T. 1977. "Transformations and the Lexicon." In Culicover, P., T. Wasow,and

A. Akmajian, eds., Formal Syntax. Academic Press: New York.Wasow, T. 1978. "Remarks on Processing, Constraints, and the Lexicon," In Pro-

ceedingsofTINLAP2.Williams, E. 1977. "Discourse and Logical Form." Linguistic Inquiry, 8. 3.Williams, E. 1979. "Passive." MS.Williams, E. 1980. "Predication." Linguistic Inquiry, 11.1.Williams, E. 1981. "On the Notions 'Lexically Related' and 'Head of a WordY' Lin-

guistic Inquiry, 12.2.

Max Planck Institute for PsycholinguisticsBerg en Dalseweg 79Nijmegen, HollandandDept. of LinguisticsSouth CollegeUniversity of MassachusettsAmherst, MA 01003 U.SA.

