Trouble with transformations

This chapter discusses the problems which arose with transformations. Attempts to limit their power proved impossible to specify. As a result, Chomsky started to look at general constraints, ways of preventing grammars from being able to do anything and everything.

Transformational grammar (TG) transformed linguistics, so it seemed. All linguists now had to do was to agree on the form which deep structures took, which were assumed to vary relatively little from language to language. They also had to produce a final and definitive list of possible transformations. At least, that was the general hope in the 1960s. However, little by little, problems crept in. Let us consider why.

[]() Waving a magic wand

The most obvious trouble with transformations in the Standard Theory (1965) was that they appeared to be a kind of magic wand, something which could change a deep structure into any kind of surface structure by any means whatsoever. But this would clearly be absurd. We would not want a device which altered []()a deep structure something like:

[]()Bill kept the dodo in the bath.

into, say,

[]()My goldfish eats bumble-bees.

There must obviously be some limits on the operations which transformations can perform. The search for the limits or constraints which must be placed on them ultimately led to a fundamental reorganization of TG. In this chapter we will look at how this has happened, showing how cracks appeared in what at first looked like a magnificent theory.

[]()Insight

It was recognized that transformations should not randomly change sentences around; they needed to have constraints (limits) put on them.

[]() Preserving the meaning

In the Standard Model of TG, the strongest constraint placed on transformations (T-rules) was that they should not be allowed to change meaning. One could therefore alter:

[]()Bill kept a dodo in the bath.

into

[]()A dodo was kept in the bath by Bill.
[]()What Bill kept in the bath was a dodo.
[]()In the bath Bill kept a dodo.

These alterations were simply stylistic nuances. They did not alter the basic proposition that Bill kept a dodo in the bath. There was no change in the lexical items, or in who did what to []()whom. Bill remained the ‘agent’, the person doing the keeping, and the dodo was still the ‘patient’, the recipient of Bill’s action.

But this was where the problem started. Certain basic transformations changed the meaning of the deep structure, as in the following examples.

First of all, consider T-passive, the T-rule which related active and passive sentences. Look at the active and passive pair:


active:	Many cowboys do not ride horses.
passive:	Horses are not ridden by many cowboys.

This pair, according to the Standard Theory, shared a common deep structure, so they should mean the same thing. Yet the reaction of many English speakers was that the two sentences had different interpretations. The active sentence implied that, although many cowboys do not ride horses, many still do. The passive sentence, however, suggested that hardly any cowboys ride horses.

This problem was not necessarily insoluble. One way out of the dilemma was to claim that both sentences had two meanings:

[]()Many cowboys do not ride horses, although many still do.
[]()Many cowboys do not ride horses, and hardly any still do.

The position of the word many in the sentence biased the interpretation of the active towards 1, and of the passive towards 2 – though in theory either meaning was possible for each sentence.

But when different transformations were involved, the problems could not be explained away so easily, as with a pair of sentences which shared a common deep structure involving a transformation known as T-conjunction reduction. This optionally eliminated repeated elements in sentences joined by and:

[]()Few women are rich and few women are famous.
[]()Few women are rich and famous.

In the second []()sentence, T-conjunction reduction has optionally been applied, and produced a sentence with a different meaning. If transformations did not change meaning, this should not happen.

How could this dilemma be solved? One way out was to suggest that there was something odd about the sentences above: they both involved quantifiers – words such as many, few, which express a quantity. In this case, we could assume that transformations changed meaning in certain circumstances, one of them being the presence of a quantifier.

This was the solution adopted by a number of linguists, who called their revised grammar the Extended Standard Theory (EST), since it represented an extension of the Standard Theory, in which (as noted above) transformations could not change meaning.

A second possible response, made by another group of linguists, was to maintain that transformations preserved meaning, but to assume that the deep structure had been wrongly formulated in the first place (). This viewpoint became known as Generative Semantics (GS), for reasons which will become apparent below.

[]()

Figure 17.1.

[]() Generative semantics

The group of people who refused to accept[]() the Extended Standard Theory (EST) claimed that the problem of T-rules which seemed to change meaning lay not so much with the T-rules themselves, as with Chomsky’s conception of deep structure which, they asserted, was insufficiently subtle. It should be elaborated so that each member of the pairs of sentences discussed above had a different deep structure. For example, they denied that the pairs:

[]()Few women are rich and few are famous.
[]()Few women are rich and famous.

shared a common deep structure. If these sentences had different meanings, then they must have different underlying structures (as the generative semanticists preferred to call deep structures). It was crucial for the underlying structures to deal in more detail with the ‘scope’ of quantifiers, the parts of the structure affected by words such as few, many. The main problem was to decide what these complicated and subtle underlying structures were like.

Generative semanticists, however, did not only argue that if two sentences had different meanings, they must have different underlying structures. They also argued (contrary to the Standard Theory) that if two sentences had the same meaning, they must also have the same underlying structure. Consider the sentences:

[]()Henry stopped Drusilla.
[]()Henry caused Drusilla to stop.

According to the Standard Theory, the deep structures would look fairly different. The first consists of only one underlying sentence, the second of two ():

[]()

Figure 17.2.

According to the generative semanticists, []()the difference between the two sentences was purely superficial. The similarity between them could be represented if the words were decomposed into their component parts. That is, generative semanticists abandoned the assumption (made in the Standard Theory) that lexical items in the deep structure were unanalysable units. Instead, they analysed a word such as stop in Henry stopped Drusilla into cause stop, and a word such as kill into CAUSE DIE (or, more accurately, into CAUSE BECOME NOT ALIVE), and a word such as remind into STRIKE AS SIMILAR TO. Then a special type of transformation packaged up the various components into a single word.

The result was that generative semanticists elaborated their underlying structures to such an extent that eventually they became indistinguishable from semantic structures. This had crucial consequences for the grammar. If underlying structures were the same as semantic structures, then clearly there was no need for them to be separate components within a grammar. The base would initiate or generate a set of underlying structures which was the semantic structure. For this reason, those who upheld such a theory were known as generative semanticists.

The general idea behind generative semantics was superficially appealing. It seemed like common sense to many people that meanings should come first, and syntactic means of expressing them follow. Remember, however, these linguists were not talking about the processes involved in producing speech. In writing a grammar, their primary aim was to specify what was, or was not, a well-formed sentence of English. And this became increasingly difficult within the generative semantics framework.

The main problem was that of specifying []()the underlying structures, which became more and more unwieldy. Nobody could agree on what they should be like, and they seemed to reflect above all the intuitions of the linguists writing them, rather than any objective reality. Furthermore, they required extraordinarily complicated rules for showing how the varying sections of the underlying structure should be combined. In the end, most supporters of this approach gradually gave up on the impossible task of specifying the details.

[]() Trace theory

Eventually, the majority of TG adherents turned away from generative semantics, and admitted that at least some surface structures were important for the interpretation of meaning. Building this possibility into the grammar resulted, as mentioned above, in the Extended Standard Theory (EST). As research continued, many linguists came eventually to the conclusion that the surface structure alone was responsible for meaning, and the resulting amended grammar became known as the Revised Extended Standard Theory (REST).

However, the assumption that surface structures alone were responsible for meaning had several repercussions on the rest of the grammar. Above all, it became important to know where items had been moved from in the deep structure. This was necessary in order for the grammar to be able to deal with the meaning of sentences in which an NP had been shifted away from its original position, as when what had been switched to the front. When an NP was moved, therefore, it was assumed to leave behind a faint trace of its previous location, marked conventionally by the []()letter t for ‘trace’. So, a deep structure something like:

[]()Q Drusilla find what in the cave. ()

[]()

Figure 17.3.

would have a surface structure something like:

[]()What (did) Drusilla find t in the cave? ()

[]()

Figure 17.4.

(Both the deep structure and the surface structure have been considerably simplified here, as elsewhere in this book: only the features relevant to the point under discussion are included.)

[]() Limiting the power of transformations

Remember, the main constraint on T-rules in the Standard Theory was that they could not change meaning. All the meaning resided in the deep structure, and this []()had to be retained in the surface structure. But in this later version of TG (REST), the surface structures alone provided the semantic interpretation. Therefore linguists were left with the problem they started out with. How could they limit the power of transformations, and prevent them creating a hopeless morass of randomly moved items? It therefore became important to provide firm guidelines as to what could move where. One proposal was that transformations could only move items around within the structure already set up by the phrase structure rules: they could not create a totally new set of structures. In linguistic terminology, they had to be ‘structure-preserving’ – and a subsidiary effect of this was to alter some of the phrase structure rules.

Another proposal was to limit the distance which items could travel, so that, except in exceptional circumstances, they were unable to hop too far away from their own clause. For example, take the sentence:

[]()The discovery that the picture of the aardvark had been stolen was quite upsetting.

The phrase of the aardvark was forbidden to hop outside the whole structure:

[]()\*The discovery that the picture had been stolen was quite upsetting of the aardvark.

A major preoccupation of linguists working on REST, then, was working out ways of constraining the power of transformations, in particular, ways of preventing items from moving uncontrollably in all directions. But this lack of constraints was not the only problem.

[]() Sharing out the work

A major advantage of Standard TG was that it appeared to simplify the superficial confusion of language. Instead of listing numerous sentence types, it specified []()a few basic patterns, and the remainder were treated as variants of these basic few.

However, in order to cope with these variants, there were literally dozens of different transformations. This large number raised its own set of problems. Above all, listing umpteen transformations which specified how to produce variants of the basic patterns was not necessarily any more economical than listing different patterns in the first place.

A further problem was that different T-rules sometimes had the same effect, yet this was not recognized in the grammar. Take the sentences:

[]()Marigold was impossible to please.
[]()Bill seems to be ill.

The surface structure subjects (Marigold, Bill) had been brought to the front by a transformation, the deep structure of each sentence being something like:

[]()It was impossible to please **Marigold**.
[]()It seems that **Bill** is ill.

Yet the transformations were quite separate, because they applied to different structures: in the first, Marigold is the ‘object’ of the verb please, in the second, Bill is the ‘subject’ of is. There was no indication in the grammar that the transformations might be linked, even though they performed similar operations.

These two problems: the large number of different transformations, and the fact that some of them appeared to have the same effect, led people to re-examine the transformations one by one. They came to two general conclusions: first, some transformations were not ‘proper’ transformations, and the operations they performed could be dealt with better in some other component of the grammar. Second, some of the remaining transformations could be combined. Let us briefly look at the offloading and combining which took place.

[]() Offloading

The lexicon and the semantic component were the two components which were seemingly underworked []()in the existing grammar, and onto which some of the existing transformations were offloaded. Consider the following sentences:

[]()Arabella gave the champagne to Charlie.
[]()Arabella gave Charlie the champagne.

In a Standard TG, the deep structure was somewhat like the first sentence. In order to arrive at the second, a transformation switched the words champagne and Charlie, and deleted the intervening to. But specifying this transformation turned out to be rather difficult. It certainly didn’t seem to be a general rule applying to the structure V NP PP. After all, you could say:

[]()Arabella took the picnic to the wood.
[]()Jim donated the book to the library.
[]()The TV station transmitted the programme to Japan.

but not:

[]()\*Arabella took the wood the picnic.
[]()\*Jim donated the library the book.
[]()\*The TV station transmitted Japan the programme.

In brief, this proposed transformation applied just to a few lexical items, such as give, tell, offer. It seemed somewhat strange to have a T-rule, which was meant to be a general syntactic rule, narrowed down in this way. The lexicon might be a more obvious place for information about the structures following a few verbs. A number of ‘transformations’, therefore, were reassigned to the lexicon.

Other transformations were reassigned to the semantic component. For example, a Standard TG assumed that the sentence:

[]()Antonio claimed that he was ill.

had a deep structure something like:

[]()Antonio claimed that Antonio was ill.

Then a transformation changed the []()second Antonio to he. But this was a fairly unnecessary complication. A simpler alternative would be to have he in the deep structure to begin with, but to keep an index of NPs as they occurred, noting which ones were ‘co-indexed’, that is, referred to the same person or thing. This allowed the semantic component to make the correct interpretation at a later stage, without any extra transformational complexities.

In general, then, the transformational component was gradually whittled down as tasks previously dealt with by transformations were offloaded onto other components in the grammar, especially the lexicon and the semantic component.

[]() Combining

As the various transformations were peeled off, only two major processes remained: transformations which moved wh- (what, which, etc.) around, and transformations which moved NPs around as in:

[]()**What** did Arabella buy?
[]()**Arabella** was difficult to please.

which had deep structures something like:

[]()Q Arabella bought **what**.
[]()It was difficult to please **Arabella**.

Yet even these two types of transformation had certain features in common, in that they both moved items to the front. Perhaps, therefore, it was suggested, there was just one basic transformation, which said: ‘Items can be moved,’ instead of lots of different transformations, each of which had to be specified separately. This general transformation then would have to be combined with some clearly stated principles about what could []()be moved where. This marked a distinct change of emphasis in the grammar. It now relied on general principles almost more than on individual rules.

[]() Slimmed-down transformations

Let us now summarize how transformations became a mere shadow of their former opulent variety.

People started with the general problem that transformations appeared to be able to do anything. Moreover, there seemed to be dozens of these powerful devices.

At first, it was hoped to constrain their power by not allowing them to change meaning. But this proved to be impossible. They clearly did change meaning in a number of cases. This therefore triggered a search for new constraints, principles which would prevent them from altering sentences around in random ways.

As people sought to specify constraints, they realized that the transformations themselves were something of a ragbag, and that a number of them should be removed from the transformational component, since the tasks they did would be better achieved within another component of the grammar. In this way both the lexicon and the semantic component became more important.

Transformations, meanwhile, dropped off one after the other. In the long run, only one major T-rule remained, which said in effect ‘anything can be moved’, but which was combined with strong constraints on what could be moved where.

These alterations paved the way for a fundamentally new version of grammar. This will be outlined in the next chapter, and so will some alternative proposals from people who want to ditch Chomskyan linguistics altogether.

[]()[]()THINGS TO REMEMBER

[]()Transformations were too powerful; they seemed to be able to do far too much.
[]()Linguists wanted to constrain the power of transformations.
[]()Originally, transformational scholars had tried to constrain transformations by not allowing them to change meaning.
[]()But several transformations did apparently change meaning.
[]()Some linguists argued that one could allow some exceptions, as when a sentence contained a quantifier, a word which specified a quantity, such as ‘few’, ‘many’.
[]()Other linguists suggested that maybe one should maintain the idea that transformations did not change meaning, but refigure the deep structures involved; however, their deep structures became so complex that this idea was eventually abandoned.
[]()Another problem was that there were too many different T-rules, and some of them seemed to have the same effect.
[]()A further problem was that some T-rules seemed to only be needed for particular lexical items.
[]()Linguists started to move the tasks of some of the T-rules to different components in the grammar.
[]()Eventually, only one T-rule remained.