Distant Instantiation and Inheritance (1996 version)

11.1 The Basic Phenomena of Distant Instantiation

The varieties of instantiation we have studied up to now (excluding null instantiation and coinstantiation cases) all have a local character: the instantiations of the elements of a valence set are structurally connected to the relevant valence-bearing constituent as 'sister' in the usual case, as 'niece' in the case of the objects of null prepositions. Non-subject complements show up as sister to the lexical head of their phrase. In the case of subject instantiation, the VP carries the valence of its head verb, so the subject in the Spec-V construction is a sister to the VP, whose valence element it satisfies. In extraposition, the 'extraposed' constituent is a sister to the lexical verb and the it constituent is either a subject, a sister to the verb (in the object case) or a niece to the verb (in the case of prepositional objects). (Students of 220 in 1996 will not have covered those topics in 120 in 1995.) Coinstantiation, too, has a local character: the predicator which provides the coinstantiation is always the lexical head of the phrase immediately containing the complement needing subject instantiation.

There are also, however, instantiation relations which reach across an unlimited number of constituent boundaries. These are sometimes called long distance dependencies. We call the phenomenon distant instantiation, or equivalently, left isolation. (The term distant instantiation is to reflect the idea that the instantiating position can be at any structural distance from its predicator; the term left isolation may be less misleading, since in some cases the structure we have in mind is identical to the position of the subject in a Spec-V construction and therefore not 'distant' from the constituent whose valence it satisfies.) Another traditional name for the phenomenon in generative grammar is extraction, the image being that of 'extracting' a phrase such as what in example (1) from its normal position after bring and moving it to the front of the sentence.

(1) What do you think Pat told Lynn that someone persuaded Max to expect the postman to bring to the house?

In example (1) the word What is the left-isolated, extracted or distantly instantiated element. You should note that what does not take its place at the beginning of the sentence via a chain of coinstantiation links, as in the 'Equi' and 'Raising' sentences we have considered up to now. Rather, in sentence (1), what is linked by instantiation only to the object valence element of bring. This one-word phrase directly satisfies the object requirement of bring, which is structurally far from being a sister to it. Moreover, we can increase the constituent-boundary distance between the left isolated element and the predicator whose valence value it satisfies ad libitum, as illustrated in (2), where the italicized material has been inserted to increase that distance without loss of grammaticality.

(2) What do you think Pat told Lynn that someone said you believe me to have persuaded Max to expect the postman to bring to the house?

In examples (1) and (2) the left-isolated phrase is an interrogative word; but the construction is not limited to wh-words, as can be seen in (3)a, or even to phrases containing wh-words, as illustrated in (3)b.

(3)a Which of Shakespeare's plays do you think she said she would like to see?

b Troilus and Cressida, I don't think he would ever want to see.

In a later section we will deal with the formation of wh-phrases. For the time being we will concern ourselves with the Left Isolation (LI) construction itself.

11.2 The LI construction

11.2.1 Valence Embedding

11.2.2 The LI Construction Displayed and Explained

The Left Isolation construction forms a maximal verbal constituent by locating to the left of a maximal verbal constituent another constituent - the 'left-isolated' or 'extracted' constituent - whose synsem and rel values are unified with those of an element which is valence embedded in the right daughter. In examples (1) and (2) the left daughter of the LI construct consists of the word what and the right daughter consists of everything else. Also, in both of these examples, the left-isolated constituent does not satisfy a valence requirement of the main verb think, but rather of the verb bring, which is embedded several clauses under think.

Left Isolation (LI) Construction

In the diagram of the LI construction given as Figure 1, the notation inside the valence value of the right daughter says that (i) there is a valence element, call it v, which is embedded in the right daughter at an arbitrary depth, (ii) the rel and synsem values of v unify with the corresponding values of the left daughter (= the left isolated constituent) and (iii) that any constituent which is valence embedded in the right daughter and which itself valence embeds v carries the notation 'sealed -'.

The attributes 'sealed' and 'loc' are devoted to the expression of certain limitations of the LI construction. Although extraction can take place over an unlimited number of intervening consitituent of the right kind, there are certain kinds of constructs which, when they constitute what would be intervening constituents in the right daughter of an LI construct, render the resulting structure ungrammatical as an instance of LI. Otherwise put, extraction cannot take place from within certain kinds of structures. One example of a structure from which extraction cannot take place is the LI structure itself, a fact that is expressed by the notation 'sealed +' in its external syntax. We will consider several more examples in a later section. (Outside of the LI construction, it is only necessary that 'sealed +' structures, traditionally called 'islands', be marked for the 'sealed' attribute. Structures unmarked for the 'sealed' attribute will freely unify with the 'sealed -' requirement of the LI right daughter and its descendants, if any. ^{fn. 1})

The 'sealed +' feature has the effect that nothing can be extracted from the constituent that it marks. If a constituent is marked with the attribute 'loc +' , which is to be read 'locally instantiated', that will mean that it is unextractable. We noted in connection with examples (8) and (9) in Chapter 8 that application of the Recipient construction blocks extraction of recipient non-subjects. We repeat those examples as (4) and (5) here.

(4) *Which spy do you think Agent 007 gave a parcel?

(Compare: Do you think Agent 007 gave Goldfinger a parcel?)

(5) *Which spy do you think a parcel was given?

(Compare: Do you think a parcel was given Goldfinger? Recall that A parcel was given Goldfinger is acceptable in `British' English.)

We may now prescribe that the Recipient construction marks such valence elements with the notation 'syn [loc +]'; this marking, together with the requirement that the LI constituent in Figure 1 specifies 'loc -', will prevent these valence elements from being distantly instantiated by the LI construction, and thus will block examples (4) and (5).

Using the 'extraction' terminology again, we can recapitulate what has been said about the two features concerned with restrictions on LI as follows. If a constituent has the feature 'sealed +', nothing can be extraced from it. If a constituent has the feature 'loc +', it cannot be extracted.

An LI construct is a full clause: it is marked 'srs +', which means that its subject requirement is internally satisfied. This subject may, however, either be realized (i) as the left daughter (Who did it?), (ii) within the right daughter (What did they do?), or (iii) within the right daughter by interpretation (...what to put it in). In case (i) the right daughter of the LI construct must be a VP (srs -); in case (ii) the right daughter must be a clause (srs +); in case (iii) the right daughter is marked srs+, in spite of the fact that it does not overtly contain an explicit subject, because it does not need its external context to satisfy its subject requirement.

The notation '[...#1[ ]...]' in the external semantics of the construction indicates that the semantics of the LI construct will contain (or possibly be the same as) the semantics of its right daughter. For example, in a sentence like That story I would never believe, the external semantics contains the proposition 'I would never believe that story' plus the information that 'that story' is topical.

As this chapter progresses we will see that the LI construction, as diagrammed in Figure 1, is really a family of constructions, or better an abstract construction whose properties are inherited by a number of more detailed constructions. What we mean by 'inheritance' of one construction by another construction will be made more precise presently. Our first illustration of this notion will be in connection with the structural pattern that goes by the name of Inversion.

11.3 Inversion

In the structure of English questions, we often find the word-order property known as subject-auxiliary inversion, or just plain inversion. In an inverted clause an auxiliary verb is followed by its subject, and that subject may be followed by additional constituents. In sentences (1) and (2) we found both inversion and the LI pattern, and it is easy to see that both LI and inversion are common in many questions.

Not all LI structures are questions, not all inverted structures are questions, not all LI structures are inverted, and not all inverted structures are LI structures. In short, the two patterns known as Inversion and Left Isolation are independently associated with various different constructions. In our discussion of inheritance, we will introduce a mechanism for describing and accounting for such interrelationships.

Example (6) illustrates both left isolation and inversion.

(6) Where did Lynn go?

In (6) the inverted clause did Lynn go constitutes the right daughter of an LI construct.

Examples (7) illustrate cases of left isolation which do not involve inversion.

(7) a Who stole the tarts?

b Raw oysters, you could never make me eat.

Examples (8) illustrate inversion in constructs which do not incorporate left isolation; the italicized portions represent the inverted (finite) auxiliary verb and subject.

(8) a Never will I leave you.

b So will she.

c Long may you prosper.

d Had I known, ...

The Inverted Clause (INV) construction is given in Figure 2.

Inverted Clause (INV) Construction

An inversion construct is a clause (srs +), headed by an auxiliary verb, fn. 2,

which is followed by its subject, and perhaps by other complements. Since it is a kind of [cat v, max +], it can unify with the right daughter of the LI construction. The Kleene star marking the third interior box indicates that there can be zero or more constituents following the subject. In most ordinary sentences there will be something following the subject, but the 'zero' provision is to cover cases like those illustrated in (9) and (10). In (9) we have non-LI examples, and in (10) we have LI constructs in which the extracted element represents a predicate supported by the copula (= the auxiliary).

(9) a Has she?

b So do you.

(10) a Where is it?

b What kind of linguist is he?

11.4 Inheritance and Main Clause Yes-No Questions

In the case of many constructions, including several to be considered in this chapter, we can capture grammatical generalizations and at the same time achieve economy of presentation by noting that one construction may serve as a 'part' of another construction. At the end of the last section we referred vaguely to 'particular constructions which contain inverted clauses'. When we come to these constructions, we will be able to express them more economically by adopting a notation which refers to the INV construction, rather than by repeating all the information of the INV construction within the inverted clause constituent of the construction being defined.

For example, the syntax of one kind of main clause yes-no question (MCYNQ) is nothing more than that of the inverted clause (INV) construction, as illustrated in

(11) [[Did] [Pat] [remember to put the mayonnaise in the fridge]]?

We recall, however, that we need the INV construction independently of questions of any type for use in licensing sentences such as (6) and

(12) [Seldom] [[did] [Pat] [remember to put the mayonnaise in the fridge]].

In a MCYNQ like (11) the external syntax is simply that of the INV construction, while in the Neg(ative)-Adv(erb) construction, which licenses sentences such as (8)a and (12), the syntax of the right daughter is that of the INV construction. We will say that the relevant MCYNQ construction inherits the INV construction and that in the Neg-Adv construction, its right daughter inherits the INV construction. (We ignore the semantic aspects of these constructions for the moment to concentrate on their syntax.) Thus, semantics aside, the (Inverted) Matrix Y-N Question construction which licenses (11) can be represented as in Figure 3 and the Neg-Adv construction can be represented as in Figure 4.

(Inverted) Main Clause Yes-No Question (MCYNQ) Construction [Syntax only]

Negative Adverb (Neg-Adv) Construction [Syntax only] ^{fn. 3}

In the statement of a construction, when we encounter in some box b the notation 'inherit X', we read that as indicating that all the information expressed in the construction named 'X' is in box b. For instance, all the information of the INV construction is contained in the right daughter of Neg-Adv. Everything else that we can add to a constituent marked 'inherit X' must unify with the construction named 'X' in that place. Thus, constituents (boxes) which contain an inheritance notation may also contain non-inherited information. For example, the construction we have been referring to as the (Inverted) MCYNQ never occurs embedded, as illustrated by the unacceptability of

(13) *I wonder whether did Pat put the mayonnaise in the fridge.

For this reason (among others) we need a way to distinguish main clauses from embedded clauses. Let us add the value for main[clause] to the possible values for the role attribute. So, while complement clauses display the role 'comp', main clauses carry the role 'main '. Then, in order to restrict inversion in questions to main clauses, we need only add the notation 'role main' to the external structure of the (Inverted) MCYNQ construction, thus:

(Inverted) MCYNQ [Syntax only]

Appearances to the contrary notwithstanding, an inheritance notation, such as 'inherit INV', does not represent an attribute value pair and does not denote part of the AVM of the constituent (box) in which it appears. The inheritance notation is, from a mechanical point of view, simply an abbreviatory shorthand because in place of writing, for example, 'inherit INV' we could have written in the appropriate box everything we wrote in the INV construction itself. That is, the revised version of the MCYNQ construction of Figure 5 is equivalent to the following, less perspicuous, representation (including the fact that we have not added below anything to indicate the semantics of interrogation). For present purposes the point that needs to be made is that the following representation of MCYNQ looks exactly like INV except for the addition of 'role main' to the external structure.

From the point of view of a computer program parsing a matrix yes-no question sentence, the inheritance notation (Figure 5) and the long winded notation (Figure 6) are equivalent, since they will accept exactly the same range of candidate sentences and assign them the same structures. But from a linguistic point of view, the inheritance notation says something more. It allows us to express the ways in which certain constructions are alike, and therefore to capture certain generalizations about the language. Thus, the nature of the structural relationship between the MCYNQ and Neg-Adv constructions is captured in the inheritance notation by our having posited the more abstract INV construction and indicating in the two more complex constructions that INV plays a role in each.

If we now take semantics and intonation into account, we will see that there is more than a single kind of MCYNQ and that the concept of inheritance can do much for us in sorting out and making clear the kinds of relationships that can hold among constructions.

Instead of uttering the sentence in (11), one can perform the same (or an extremely similar) illocutionary act by uttering the sentence in (14), which employs Spec-V syntax with a rising intonation contour (in contrast to the INV syntax, with either rising or falling intonation, of 11). ^{fn. 4}

(14) Pat remembered to put the mayonnaise in the fridge? (Amazing!)

We are forced to notice, then, that there are two distinct MCYNQ constructions illustrated by (11) and (14). Obviously we consider these two constructions related, since we find it natural to give them the same name. This expression in fact refers to three common properties of the two constructions: both are main clause constructions (as against clauses that can occur embedded within other clauses), both are questions, and semantically both may be used to present a full proposition and request its acceptance or rejection (yes/no-type questions) versus presenting a partial proposition and requesting its completion (wh-type questions). We have already noted that the two constructions differ in their overall syntactic form (INV) versus Spec-V and in intonation; only the Spec-V type requires rising intonation, the INV type permitting, but not requiring, this contour.

There is a further distinction to be made between the two constructions. This is illustrated in example (15) and (16).

(15) a Did Pat take something from the fridge?

b Did Pat take anything from the fridge?

(16) a Pat took something from the fridge?

b *Pat took anything from the fridge?

The INV type question (15) can create what is called a negative polarity environment, that is an environment in which words like anything, anyone, any, and expressions like a red cent and lift a finger can occur.^{fn. 5}

(i) a *He likes me the least bit
b He doesn't like me the least bit

(ii) a *He has any money
b If he has any money, he'll buy it

(iii) a *He got there until six
b I doubt he got there until six

Although the phenomenon of polarity seems to be fundamentally semantic, it is very difficult to say what the semantics of polarity is. Polarity is a difficult and not well understood subject, both semantically and syntactically. We will have a little more to say about it in a later chapter.

More precisely, the INV question (15) creates an environment which accepts either positive or negative polarity items, while the Spec-V question (16) creates an environment which accepts only positive polarity items.^{fn. 6}

Our strategy is to formulate an abstract MCYNQ construction and then propose two specific MCYNQ constructions, each of which inherits the abstract construction. The former is given in Figure 7. This construction specifies only a finite, main clause with 'yes-or-no' semantics. It does not commit itself on inversion versus Spec-V form.

Figures 8 and 9 represent inverted and non-inverted MCYNQ constructions, respectively. The inverted construction, shown in Figure 8, inherits the abstract MCYNQ construction of Figure 7 and the INV construction of Figure 2. Nothing more needs to be added to these specifications.

We have seen in this section that the mechanism of inheritance can be useful in specifying similarities among constructions and in this way drawing attention to syntactic generalizations. In discussing various related constructions involving left isolation and some additional constructions with which these interact, we will have frequent recourse to inheritance of one construction by another. In particular, we will see that the LI construction itself, like the INV construction and like the abstract MCYNQ constructions, is inherited by several more specific constructions.

11.5 A Partial Classification of Finite Clauses

In order to give reasonable, though not exhaustive, coverage to the phenomenon of LI, it is necesssary to discuss a fairly wide range of structures which do not directly involve left isolation itself. There are two reasons for this. First, a number of LI constructions either include or are included in other (non-LI) constructions. Secondly, there are constraints on the operation of the long distance dependency of the LI construction between the left daughter and the valence element contained somewhere within the right daughter, and to discuss these we must consider additional constructions.

In discussing examples (1) and (2) at the beginning of this chapter, we noted that this dependency can exist across an unlimited number of clause boundaries. But although there are certainly cases that conform to this generalization, there are other cases in which left isolation is not possible. For example, compare examples (17) and (18).

(17) Which rule did you say someone reported to the sergeant that Private Schmallowitz had broken?

(18) *Which rule did the sergeant hear a report that Private Schmallowitz had broken?

Even though the distance - measured in words, phrases or anything else - between the left isolated element which rule and the predicator broken, one of whose valence elements which rule satisfies, is greater in sentence (17) than in (18), only the former is acceptable. It is evidently something about the structure, not the length, of (18) that makes it bad. Operation of the LI construction is constrained in a variety of ways, and to discuss these we will have to consider a number of non-LI constructions which interact with LI.

It will be convenient to classify finite clauses according to a tri-partite division into (i) main clauses, (ii) argument or complement clauses and (iii) clauses serving as modifiers. In the remainder of this section, we present a classification of finite clause types and in section 5 we consider in greater detail the specific constructions which license them.

11.5.1 Main Clauses

English contains a number of types of main clauses, several of which inherit the LI construction. The only one we have considered in detail prior to this chapter does not involve distant instantiation. It is the declarative sentence type, which inherits the Spec-V construction and is illustrated in a sentence like (19).

(19) She won the prize.

There are several types of main clause questions. In section 11.4 two types of main clause yes-no questions were studied. One of these inherited Spec-V and the other INV. Neither inherited LI.

Main-clause non-subject questions inherit both INV and LI, as seen in (20). fn. 8

Main-clause subject questions inherit LI but not INV, as illustrated in (21).

(21) Who won the prize?

There is another type of (non-question) main clause which inherits LI. This is in fact an abstract syntactic type that contains semantically distinct subtypes and thus deserves an abstract construction inherited by the specific constructions licensing its subtypes. We may call this the Topic/Focus construction. The topic version is illustrated in (22)a and the focus version in (22)b.

(22)a The latter, he told me he would never agree to.
b [I like to go to plays and movies.] Operas, I'm unable to work up much enthusiasm for.

In (22)a we have chosen to employ an anaphoric element The latter for the left-isolated constituent to emphasize that this kind of fronting syntax can serve the rhetorical function of presenting an argument as already 'on the floor' in the conversation. Contrastingly, in a sentence like (22)b the left isolation of the element Operas can serve to indicate that it presents the only 'new information' in the sentence, i.e., information not presupposed to be present in the prior discourse. (There is a great deal more to be said about the discoursse functions of various topicalizing and focussing constructions than will be mentioned here.)

There exists a variety of other kinds of main clauses of varying degrees of productivity (versus idiomaticity). Without any claim to exhaustiveness and with no attempt to decribe the special semantic/pragmatic values that attach to their syntactic types, we mention three of these before going on to complement clauses.

(23)a Leave my house!
b Him get into med school?!
c The more syntax you learn the more of a party animal you become.

Example (23)a illustrates an imperative sentence. The construction realized by a sentence like (23)b may be called the Incredulity Construction and is syntactically characterized as having a non-nominative subject and a main verb in the bare stem form. The syntax of a sentence like (23)c - which will be discussed in detail in a later chapter - takes more time to describe than we have room for here. Notice, though, that it incorporates two instances of inheritance of the LI construction: the more syntax is a left-isolated complement of learn and the more of a party animal is a left-isolated complement of become.

11.5.2 Complement Clauses

Verbs, nouns, and adjectives may contain valences which call for clausal complements.

(24)a I think [that she won].
b I am entertaining the thought [that she won].
c I am aware [that she won].

Clausal complements may also be questions.

(25)a I wonder [whether she won].
b I am entertaining the question [whether she won].
c I am uncertain [whether she won].

We can left-isolate a constituent of a non-question complement clause without any problem.

(26) Who do you think (that) she told the news to?

It is generally maintained, however, that English prohibits extraction from question complements. Many examples like (26) support this generalization.

(27) *Who do you wonder whether she told the news to?

It is possible, though, to find examples which seem to violate this generalization.

(28) ?He's the guy who she's been wondering whether she should ask to the prom.

In examples like (28), which many English speakers accept, who is a left-isolated element which unifies with a valence element of ask; the latter is inside the complement question whether she should ask to the prom. Despite the existence of examples like (28), we will go along with the orthodoxy according to which English prohibits extraction from embedded questions.

We have seen that complement clauses can be subclassified as questions or non-questions and that this subclassification has consequences with regard to their behavior in left isolation contexts. In the examples we have considered so far, each complement clause has been introduced with a special word, called a complementizer, which serves just this purpose. In yes-no question complements a complementizer, whether or if, is always necessary (29)a,b, but (finite) non-question complement clauses can come either with or without the complementizer that. fn.9

When a complement clause is introduced by a complementizer, its subject can not be left-isolated.

(30) a Who do you think she likes?
b Who do you think likes her?
c Who do you think that she likes?
d *Who do you think that likes her?

The examples in (30) show that with a non-question complement, extraction of the subject is blocked only when the clause is introduced by a complementizer. With regard to question complements, the situation is apparently similar. However much you accepted or failed to accept (28), we suspect you will find (31) noticeably worse.

(31) *He's the guy who she's been wondering whether will ask her to the prom.

Example (31) appear to suffer because it left-isolates the subject of a complement clause introduced by a complementizer - independently of the degree to which it also suffers from left-isolating a constituent of an embedded question.

In fact, these kinds of cumulative judgments of unacceptability have been used to argue that the impossibility of left isolation of the subject of a complement clause introduced by a complementizer is part of something more general. It is possible that English prohibits extraction of the subject of any clause that is introduced by a 'little word', whether or not the clause is a complement and the 'little word' is a complementizer. It is hard to reach a firm conclusion here because in most such clauses no extractions of any kind are possible, and so one has to have recourse to judgments of cumulative unacceptability of the kind we have just been considering. For example, left isolation of any constituent of an adverbial subordinate clause tends to be unacceptable.

(32) *What was Lynn frying the onions while Pat was chopping?

Many authors judge extraction of a subject from an adverbial clause to sound even worse.

(33) *Who was Lynn frying the onions while was chopping the meat?

It is not implausible to conclude that (33) suffers both from violation of a ban on left-isolation of any consituent of an adverbial clause and from violation of a ban on left isolation of the subject of a clause introduced by a 'little word', including subordinating conjunctions such as while.

In theories that employ a notion of movement, a great deal of effort has been devoted to explaining why the subject of an introduced complement clause cannot be left isolated. In CG, this observation simply follows from the fact that we have no constructions which license structures consisting of a complementizer followed by a VP (i.e., an srs- constituent). You will recall in chapter 6 that we studied constructions like those licensing that-clauses and for-clauses, but none licensing a structure consisting of a complementizer followed by a VP.

11.5.3 Relative Clauses

Most, but not all, relative clauses inherit the LI construction. Relative clauses may first be subdivided into those which serve as modifiers and those which do not. If you are familiar with only one of these types, it is probably the former. Example (34)a contains a modifying relative clause and example (34)b contains a non-modifying relative clause.

(34) a I have visited the town [where Pat used to live].
b [Where Pat used to live] was a one-horse town.

(You should not conclude from this pair of examples that any sequence of words that forms a modifying relative clause also forms a non-modifying relative clause, or conversely: who I know can only be of the modifying type; what the dog ate can only be of the non-modifying type.) Traditionally, what we have called non-modifying relative clauses are called 'headless relative clauses'. The name arises historically from the fact that this kind of clause does not occur within a noun phrase headed by a lexical noun, since such clauses don't modify nominal constituents; rather they function as noun phrases by themselves. We will have more to say about non-modifying, or headless, relative clauses below.

With regard to the semantic function of modifying relative clauses, we must distinguish between restrictive and non-restrictive modification. The distinction came up briefly with regard to modifying adjectives in Chapter 3, where we noted that proper nouns, since they denote unique individuals, cannot be restrictively modified, although they can be non-restrictively modified. (In (35)b the intended 'restrictive' interpretation is signalled by the absence of a comma after Pat.)

(35) a I met a person who likes vodka and tonic.
b *I met Pat who likes vodka and tonic.
c I met Pat, who likes vodka and tonic.

There are three syntactic types of modifying relative clauses, as illustrated in the following.

(36) a He bought the car which she hoped he'd buy.
b He bought the car that she hoped he'd buy.
c He bought the car she hoped he'd buy.

We will call these types 'wh-relatives', 'that-relatives' and 'bare relatives', respectively. All three types may be used for restrictive modification, but only wh-relatives can serve as non-restrictive modifiers.

(37) a The car which my mother gave me is a 1983 Ford.
b *My car, that my mother gave me, is a 1983 Ford.
c *My car, my mother gave me, is a 1983 Ford.

11.5.4 Non-Modifying or Headless Relative Clauses

Headless relatives differ in both external and internal semantics and in external and internal syntax from modifying relatives. Considering external function first, while modifying relatives modify nominal constituents, headless relatives straightforwardly denote entities. They occur in places where we expect a noun phrase. Example (34)b shows a headless relative fulfilling the subject function. Examples (38)a,b show headless relatives fulfilling obj and pobj functions, respectively.

(38) a I found what I was looking for.
b She talks about what she likes.

With respect to their internal syntax, headless relatives can be subclassified into the -ever type and the ordinary type (illustrated in 38). Examples (39)a,b present the -ever versions of examples (38)a,b.

(39) a I found whatever I was looking for.
b She talks about whatever she likes.

11.5.5 Headless Relatives versus Embedded Questions

The syntax of non-ever headless relatives is in many cases identical to that of embedded questions and the two are easy to confuse. Thus, in some contexts a given string of words which can be interpreted as a headless relative may also be interpreted as an embedded question.

The verb wonder takes only an embedded question as a complement, while a verb like cook does not accept any sort of propositional complement, including question complements. Thus the italicized phrase in (40)a is necessarily an embedded question while that in (40)b is necessarily a headless relative.

(40) a I wonder what he cooked.
b I ate what he cooked.

A common rule-of-thumb for distinguishing headless relatives from complement questions pairs the possibility of substituting a wh-ever form for the simple wh-word in the case of headless relatives (only) with the possibility of adding the anaphoric element else only to the embedded question, thus:

(41) a I wonder what else he cooked.
b *I wonder whatever he cooked.
c I ate whatever he cooked.
d *I ate what else he cooked.

Some predicators take either sort of complement. The verb know, for example, allows both NP objects (know the facts ) and embedded question complements (know how to do it ). By using know as both the root and the complement verb, we can get an ambiguity of interpretation of the complement as either headless relative or embedded question. Examples (42) illustrate this with the -ever versus else test.

(42) a I know whatever Syd knows.
b I know what else Syd knows.

So a sentence like

(43) I know what he knows.

can be interpreted along the lines of either (42)a (headless relative) or (42)b (embedded question).

Having briefly subclassified some of the major varieties of finite clauses, we turn now to the constructions which license these clauses. Our attention will be focussed upon those constructions which inherit or which interact with left isolation.

11.6 Main Clause Constructions

An ordinary declarative sentence is, syntactically speaking, simply a Spec-V construct which is marked as a finite main clause.

11.6.1 MCYNQs

MCYNQ constructions were treated in 11.4. We simply recall here that there were two types, one having the syntax of INV and the other the syntax of Spec-V, with differing intonation possibilities and different constraints on polarity items.

11.6.2 Main Clause Wh Questions

Our having named the main clause wh question constructions exemplified in (21) and (22) 'non-subject' and 'subject', respectively, requires some amplification. 'Subject' here refers only to the subject of the right daughter of the LI construct itself, not to the subject of any clause embedded within the right daughter of the LI construct. Thus, repeating (21) and (22) for convenience, we see that an example such as (44)a,b in which either an embedded subject or an embedded non-subject is extracted has, like (21), inversion in the right daughter.

(21) What did she win?

(22) Who won the prize?

(44) a What do you think she won?
b Who do you think won the prize?

Non-subject main clause wh questions are, thus, those in which the left-isolated element unifies with any valence element of any predicator of the right daughter except for the subject requirement of the main verb of the right daughter.

11.6.3 Main Clause Subject Wh Questions

A main clause subject wh question, such as (22), inherits both the LI and Spec-V constructions. The left-isolated, wh phrase and the subject are the same constituent! There is no inversion.

11.6.4 Topicalization and Related Constructions

The Topic/Focus construction raises certain questions regarding how the data should be construed. The question is whether, or rather under what conditions, this construction permits distant instantiation properly so-called. That is, when the left-isolated element does not unify with a valence requirement of the main verb, whether or not the left isolation is possible appears to depend on the meaning of the intervening clause(s). Examples (45)a,b seem to be okay with everyone.

(45)a Braised cockroach I would never eat.
b Braised cockroach I don't believe I would ever eat.

Example (45)a exemplifies extraction from a main clause, while (45)b represents extraction from an embedded clause (distant instantiation). But whereas any sentence like (45)a will be judged fine, many sentences of the (45)b type may be judged questionable.

(46) *Braised cockroach Lynn bet Pat would never eat.

Many would judge sentence (46) ungrammatical, but accept (47).

(47) ?Braised cockroach I'll bet I would never eat.

Sentences (46) and (47) have the same structure; in fact the intervening verb bet is the same in both. What might make the difference is that I'll bet can be considered almost semantically empty in (47). The expression I'll bet often functions as a mild hedge on a prediction: the sentences Pat is going to be late and I'll bet Pat is going to be late probably do not differ in truth conditions (on one reading of I'll bet ). Lynn bet in (46) is, however not subject to this kind of interpretation. Since it is likely that complex aspects of interpretation affect judgments of acceptability in these cases, we tentatively conclude that the grammar itself treats topic/focus constructions as allowing left isolation of all embedded arguments. We propose then, that what makes a sentence like (46) hard to accept is not part of English grammar per se, but rather must ultimately be explained by the mechanisms according to which certain aspects of utterances relating to notions such as topic and focus are processed. The topic/focus constructions are most suited to conversation, and it may simply be that cognitively complex sentences sound unnatural in ordinary conversation.

Although Topic and Focus versions of this construction differ with regard to the differences in the semantics of topics and foci, they all have in common a semantic restriction: the left isolated element must be semantically contentful. Null elements cannot be topicalized as illustrated by the acceptable passivization in (48)a and the unacceptable topicalization in (48)b of the null element and the acceptable topicalization of the non-null element in (48)c.

(48) a It was expected to rain.
b *It they expected to rain.
c Pat they expected to cooperate.

Idiom chunks, similarly, appear more readily in initial position as passive subjects than as topics (or foci).

(49) a Unfair advantage was taken of his customers by Karl Cheating.
b *Unfair advantage Karl Cheating took of his customers.

This constraint on the Topic/Focus construction - that the first element must be semantically contentfull - is reflected by the notation 'rel ~[theta null]' (negation in this text being symbolized by '~' and greek letters being written out) on the left daughter in Figure 13.

Complement clauses can be subclassified according to whether or not they are questions and according to whether or not they are introduced by a complementizer.

11.7.1 Introduced Complement Clauses

We consider first the case of complement clauses introduced by a complementizer. For such clauses we need only a single construction, as we can let the nature of the complementizer determine whether the introduced clause is a question.

The lexical complementizer that carries the notation 'mood declarative', while whether and if carry 'mood interrogative'. This feature is copied into the semantics of the introduced clause and there determines whether or not the introduced clause is interpreted as a question.

This construction will license the italicized parts of examples like

(50) a I believe that Pat ate the pizza.
b I wonder whether/if Pat ate the pizza.
c I am happy that Pat ate the Pizza.
d I am uncertain whether/if Pat ate the pizza.
e The fact that Pat ate the pizza is important.
f The question whether Pat ate the pizza is unimportant.

Since this construction requires that its right daughter be a Spec-V constituent and since the Spec-V construction requires internal instantiation of the subject, left isolation of the subject of the italicized clause in each of the above cases is impossible, as indicated in

(51) a *Who do you believe that ate the pizza?
b *Who do you wonder whether ate the pizza?
c *Who are you happy that ate the pizza?
d * Who are you uncertain whether ate the pizza?
e *Who is the fact that ate the pizza important?
f *Who is the question whether ate the pizza unimportant?

While extraction is possible from the complements of verbs and adjectives, as in (52)a,b, or from the prepositional complements of nouns, as in (52)d,f, extraction is not possible from clausal complements of nouns, as in (53).

(52) a What does Pat believe that Lynn wrote in the notebook?
b What is Pat sorry that Lynn wrote in the notebook?
c The prophet witnessed the destruction of the Roman Empire
d Which empire did the prophet witness the destruction of?
e This is the lid to the pickle jar.
f Which jar is this the lid to?

(53) *What does Pat resent the fact that Lynn wrote in the notebook?

Because of the clear acceptability of (51)a,b, we cannot attribute the unacceptability of a sentence like (52) to sealing of the italicized complement clause. Because of the acceptability of (51)d and (51)f, we cannot say that all complex NPs are sealed. A brute-force solution is to stipulate that noun phrases containing clausal complements are sealed.

11.7.2 Bare Complement Clauses

Complement clauses, either questions or non-questions, can also occur without complementizers. In this case, however, two separate constructions are required. Whether or not the complement clause is interpereted as a question cannot be determined by the complementizer, since there is none. Further, the internal syntax of bare question complements differs from that of bare non-question complements.

Bare question complements take the form of wh-marked LI structures.

(54) a I wonder who gave it to her.
b I wonder what he gave to her .
c I wonder why he gave it to her

We can represent the Bare Complement Question construction, again ignoring semantics, as follows.

(55)a *Which toy do you wonder who gave to the child?
b *Which child do you wonder which toy I gave to?
c *Which toy do you wonder why I gave to the child?

11.7.3 Bare Non-Question Complements

Bare non-question (declarative) complements are exemplified by sentences like

(56) Pat thinks Lynn likes potato chips.

Unlike bare question complements and both types of introduced complements, bare non-question complements allow distant instantiation of their subjects (that is the subject of their main verb). Thus,

(57) Who does Pat think likes potato chips?

These complements are not LI structures, as are bare complement questions, and they are also not specified to be clauses (srs +), as introduced complements are. Bare non-question complements are unspecified with respect to the srs attribute. They are not inverted and will appear either as VPs (57) or as Spec-Vs (56).

We discuss in turn the constructions for the modifying types of relative clause: wh, that, and bare and then for the non-modifying (headless) types: plain and -ever.

11.8.1 Modifying Relative Clause Constructions

A modifier is like a predicator in the respect that both have an unsatisfied valence element. We might say that while a modifier and a predicator are dedicated to doing different jobs (predication versus modification), each requires that it combine with another constituent (the one described in its unsatisfied valence element) in order to do its own job. ^{fn. 10}

(i) A spectator
(ii) A spectator screamed.
(iii) a A screaming spectator
b A spectator who was screaming

We attribute the difference in semantics observed in (ii) versus (iii), not to distinct predicative and modifying lexical entries scream, but rather to a difference between modificational and predicational constructions. Noun modifiers are not the only kind of modifiers that exist. There are also predicate modifiers, which map predicates to predicates (e.g. 'fast' to 'very fast'), and other types as well. A modifier always maps an element of a given semantic type to another element of the same semantic type. A predicate never does this.

(58) a [a man] who I met in Kabul
b [a man] whose sister I met in Kabul

In (58)a the meaning of who will unify with the unsatisfied valence element of the relative clause who I met in Kabul and thence unify with the meaning of the modified noun man. In (58)b the meaning of the left isolated phrase whose sister does not in its entirely unify with the external valence element of the construct whose sister I met in Kabul and thence with the head noun man. Rather a part of the meaning of whose sister, (the 'who' part), does this. Figure 17 depicts the wh modifying relative clause construction.

In Figure 17 we may note first that a new phrasal role, modifier (mdfr), has been assigned to the construction. We may imagine in retrospect that in the Modified Nominal Construction of Chapter 3, and in all other modificational constructions, fn. 11

(59) *a rock who fell on my head

is peculiar because the information 'human' contained in who clashes with what we know about rocks.

11.8.1.1 That -Relatives

The that-relative relative construction is exactly like the wh-relative except that the word that occurs as the sole element of the left daughter.

The construction in Figure 18 specifies the word that to be the whole of the left-isolated constituent. Thus, we can have relative-including structures like (60)a,b,c but not like (60)d.

(60) a the guy that we were talking about.
b the guy whom we were talking about.
c the guy about whom we were talking.
d *the guy about that we were talking.

We think of the word that in this construction as a special kind of relative pronoun. It has the semantics of a wh-pronoun (whatever that may be); it occurs only in this construction; and it necesarily constitutes the whole of the left isolated element in the construction. fn 12

¹² This is not the only possible analysis of that-relatives. Another will be considered later.

11.8.1.2 Some Baffling Mysteries Involving Special Head Nouns

There are a number of puzzling cases in which a head noun accepts a relative clause with a wh-word which seems redundant with part of the meaning of the noun (e.g., 'time' or 'place') and also accepts either a that relative or a bare relative, but does not accept a relative with the wh-word which, unless a preposition is included to carry the temporal- or spatial-location meaning. In examples (61) the word day incorporates the idea of a location in time redundantly with the word when. In examples (62) the word place incorporates the notion of location in space redundantly with the word where. This kind of head noun is happy with semantically redundant relative introducers (when, where) and also with that and bare relatives. But, as noted, a preposition is required with the relative pronoun which.

(61) a the day when Pat and Lynn met.
b the day that Pat and Lynn met.
c the day Pat and Lynn met.
d the day on which Pat and Lynn met.
e the day Pat and Lynn met on.
f *the day which Pat and Lynn met.

(62) a the place where Pat and Lynn met.
b the place that Pat and Lynn met.
c the place Pat and Lynn met.
d the place at which Pat and Lynn met.
e the place Pat and Lynn met at.
f *the place which Pat and Lynn met.

Way and reason as head nouns prohibit stranded prepositions, contrasting in this respect with the f examples in the two preceeding sets.

(63) a the way in which she convinced him.
b *the way she convinced him in.
c the reason for which she prepare the speech so carefully.
d *the reason which she prepared the speech so carefully for.

These special head nouns also accept both that and bare relatives.

(64) a the way that she convinced him.
b the way she convinced him.

(65) a the reason that she prepared the speech so carefully.
b the reason she prepared the speech so carefully.

Examples (61-65) repesent the tip of an iceberg of mysteries surrounding interactions between the members of a small set of head nouns, whose semantics is to some degree redundant with that of a relative pronoun (time, place,...), and certain finicky details of the form of allowable relative clauses.

11.8.1.2 Bare Relative Clause Constructions

Bare relatives behave somehat, but not entirely, like other modifying relatives. In particular, like other modifying relatives they illustrate the phenomenon of long distance dependency, which up to now we have spoken of as restricted to the LI construction.

(66) a the man you thought I said she had spoken about.
b the man that you thought I said she had spoken about.
c the man who you thought I said she had spoken about.

It is important to notice, however, that the long distance dependency in (66)a does not relate two constituents within the relative clause. There is a long distance dependency in (66)a between a valence element of about and the head noun man, but there is no constituent of the relative clause in (66)a which participates in this dependency as do the that and who constituents in (66)b,c. We cannot, therefore say that the bare relative construction inherits LI. There is nothing in bare relatives to fill the left daughter of an LI structure. We will consequently have to devise another means to capture the long distance dependency in (66)a.

Bare relatives behave differently from the other modifying relatives in a second way. In bare relatives the 'extracted' valence element, i.e., the one that ends up dependent upon the head noun to which the relative clause is attached, cannot be the subject of the relative clause. fn. 13

(67)a *The man thought I said we were raising our prices refused to renew his subscription.
b The man who thought I said we were raising our prices refused to renew his subscription.
c The man that thought I said we were raising our prices refused to renew his subscription.

In bare relatives the subject is always instantiated internally, in its normal position. In both wh-relatives and that relatives, the subject is also always instantiated internally, but this comes about differently. In the bare relative case we have an ordinary Spec-V structure with the subject at home. In the wh and that cases we find either (a) the left-isolate daughter unified with a non-subject and the subject instantiated within the right daughter or (b) the subject occupying the left-isolate daughter and marked 'loc -'. Thus, bare relatives have a Spec-V structure, not an LI structure, but they share with the LI relatives the possibility for an embedded argument to be associated with the head noun, creating a long-distance dependency. So, bare relatives are Spec-V structures whose right daughters share with the right daughters of LI relatives (i.e., wh and that relatives) the property that some valence element contained within them - not necessarily one belonging to the main predicate of the relative clause - becomes an external valence element. The following diagram of the bare relative construction incorporates these properties.

The above diagram is like rather like that of the other modifying relative clauses except that the left daughter is the subject and cannot function as the 'loc -' element. The overall structure is that of a Spec-V construct rather than an LI construct. However, the part of the LI construction which makes possible long distance instantiation is written into the bare relative construction (in the form of the unlimited number of optional unsealed constituents dominated by the right daughter and dominating the predicator governing the externally satisfied valence element).

It was mentioned in a note earlier that a second analysis of that-relatives is possible. This analysis treats the that of that-relatives, not as a special kind of relative pronoun, but as a special use of the complementizer that. Under such an analysis the that-relative construction is more directly related to the bare relatives than to the wh relatives. That appears as a left daughter complementizer in a construction whose right daughter is a bare relative and whose external structure inherits everything from its right daughter. A version of such a construction is given in Figure 20.

Ordinary, non-ever, headless relatives have a syntactic shape that often allows them to be confused with complement questions. The meaning of a headless relative is, however, different from that of an embedded question. Also, ordinary headless relatives have an internal syntactic property that distinguishes them as a subclass of the class of structures that can appear in environments where either a headless relative or an embedded question is permitted, as after the verb discover. The following example,

(68) She discovered what he had hidden,.

is ambiguous between an embedded question reading, in which she found out the identity of the hidden object, and a headless relative reading, in which she found the hidden object. However, the examples in (69) are not ambiguous in this (or any other) way. They must each be interpreted as an embedded question, not as a headless relative.

(69) a She discovered which one he had hidden.
b She discovered which he had hidden.
c She discovered whose key he had hidden.
d She discovered whose he had hidden.

Compare the examples in (70), each of which presents an environment which does not accept an embedded question.

(70) a *She picked up which one he had hidden.
b *She picked up which he had hidden.
c *She picked up whose key he had hidden.
d *She picked up whose he had hidden.

The lexical makeup of the left daughter of an ordinary headless relative must consist exclusively of one of the words what, where, when, how or why. Multi-word wh-phrases are not possible, nor are those consisting of the words whose or which, which would require an interpretation in which some nominal element is anaphorically reconstructed ('whose key', 'which key').

The majority of the actual tokens of headless relatives one encounters begin with what, but a few verbs allow as complements headless relatives which begin with the other wh-words listed. For example, the verbs admire and deplore require explicit objects.

(71) a *I admire.
b *I deplore.

From this fact it follows that in (72) the italicized phrases are complements, not adverbial clauses.

(72) a I admire what he did but I deplore why he did it.
b I deplore what he did but I admire how he did it.

Examples (73) show that the clauses italicized in (72) must be headless relatives and not embedded questions (which agrees with our intuitions regarding their meanings).

(73) a *I deplore why else he did it.
b *I admire how else he did it.

Examples (74)a,b illustrate headless relatives in where and when.

(74) a Did you see where it went?.
b I remember when a coke cost a nickel.

(Headless relatives in who, are also marginally possible, but have a tightly restricted distribution: I like who he married is fine but *I wrestled who he married, *Who he married is coming to dinner, and *He gave who he married a lovely ring are bizarre. Who steals my purse steals trash is a remembered line from the earlier English of Shakespeare and Can I help who's next? appears to be a recent idiomatic innovation.)

In semantic type, headless relatives are arguments, not modifiers. They appear in such grammatical functions as subj, obj, pobj, etc. They generally appear where one might expect a NP, for example as the object of a predicator such as pick up (compare examples 70). Briefly, and with several oversimplifications, we can make the following contrast. An expression which is unambiguously a modifying relative because of its syntactic structure, such as

(75) whose book Pat borrowed

cannot be used to refer to an entity. It can only participate in referring to something when it is combined with an expression which already has the potential to refer. The entity that gets referred to by the combination of expressions is one that satisfies the meanings of each expression. (This is an oversimplification.) Thus,

(76) the student

can refer to an entity, and

(77) the student whose book Pat borrowed

refers to an entity x such that (i) x is a student and (ii) Pat borrowed x's book.

A headless relative, on the other hand, can refer to an entity and can perform most of the other semantic functions that garden variety NPs do. Hence we are not surprised to find that headless relatives occur in environments where ordinary NPs are welcome. Example (78) represents a construct which can't be a modifying relative, because what cannot introduce a modifying relative.

(78) what Pat had hidden behind the couch

An expression like (78) can be used directly to refer to an entity x such that Pat hid x behind the couch.

(79) Lynn stepped on what Pat had hidden behind the couch.

In formulating the Headless Relative Clause construction, we must account for the fact that headless relatives behave externally as noun phrases despite the fact that they are internally organized by LI, which licenses a verbal constituent according to the formulation of Figure 1. We might do this in either of two ways. The first, and probably preferable, way involves slightly redoing the LI construction of Figure 1 in such a way that syntactic category is unspecified, rather than specified as 'v', in the external structure. Then all the kinds of LI structures that we have treated up to now would inherit a second 'Verbal LI' construction, which would itself inherit LI and simply add 'cat v' and 'srs +'. That would leave the more abstract LI unspecified for category and thus available to be inherited by the headless relative construction of Figure 21. (In our diagrammatic representation of the headless relative construction (Figure 21), the external semantics value is intended to be no more than suggestive of a serious treatment of the semantics of the construction.)

We will not show anything in our representation of the -ever headless relative construction to deal with its (sometime) semantic differences from the ordinary headless relative construction beyond the purely mnemonic and unexplanatory notation 'ever +'.

Syntactically, -ever relatives differ from ordinary headless relatives in that they permit the full range of possible wh-phrases as left-isolate daughter.

(80) a Whatever it was made of must have been expensive.
b *Which material it was made of must have been expensive.
c Whoever's dogs have been in my garden had better watch out.
d *Whose dogs have been in my garden had better watch out.

Consequently, in the representation of the -ever relative construction, there is no 'lex +' notation in the syn value of the left daughter. Essentially the same alternative analyses represented in Figures 21 and 21a for ordinary headless relatives are possible for -ever headless relatives.

Since all headless relative constructions inherit LI, we have accounted for the fact that all headless relatives are islands.

11.9 'What the ... !'

Frequently, but by no means always, a wh-phrase can contain, directly following the wh-word, an exclamatory expression like the heck, the daylights, the devil, in heaven's name, in (the) blue blazes, in tarnation, and so on. But what on earth are we to make of the distribution of wh-exclamations exemplified in (81)?

(81) a Who the devil cares? [main clause question]
b I can't imagine who on earth would care [complement question]
c Whoever the heck cares about such a thing ought to have their head examined [-ever headless relative]
d *I met someone who the devil cares about it [modifying relative]
e *You found what in blue blazes? [ordinary headless relative]
f Who the hell's car did you get here in? J[preposition stranding(main clause question)]
g *In who the hell's car did you get here? [pied-piping (main clause question)]

The contrast in acceptability between f and g may perhaps be explained as a register clash in g. Wh-exclamations appear to be characteristic of casual speech, while pied-piping characterizes writing, or speech which emulates written style. In a sentence which reaches for a humorous effect through a mixture of written and casual registers (such as the final sentence of the preceding paragraph), wh-exclamations with pied-piping seem more at home.

(82) To what on earth may we attribute the distribution illustrated in (81)?

Assuming we have accounted for the unacceptability of g, we are left with the observation that wh-exclamations are acceptable in questions, either root (81)a or embedded (b,f) and in ever-relatives (c), but not elsewhere. If we could find some factor common to questions and -ever-relatives, but lacking elsewhere, that would provide the desired explanation of the distribution.

11.10 Constraints on Left Isolation

... to be continued...