Deep Structure, Surface Structure and their Demise


(Note: the following is very rough, informal; many distinctions have been elided for the sake of clarity and space.)


The enterprise begun by Chomsky in 1949 has gone through many changes, some superficial, some profound. In recent years, Chomsky has highlighted what he calls “virtual conceptual necessities”; these are properties any theory of linguistic competence must reflect or accommodate:


(i) A lexicon - a ‘list’ of exceptions.

(ii) Combinatorial structure - lexical items must be combined.

(iii) Interfaces - the combined structures must interface with ‘external’ components which put the structures to ‘use’, minimally, ‘sound’ and ‘meaning’.


These characteristics are the bare minimum required to account for the structure of the language faculty - an aspect of the mind/brain.


From the earliest theories to the latest, three principal phenomena have animated the generative enterprise:


(i) Dissociation of sound and meaning

Roughly, because sound is linear - along one dimension - the structure required to encode it must be likewise linear, i.e., the items of the structure must be well-ordered in the mathematical sense. On the other hand, the structure required to encode meaning must be at least two-dimensional, perhaps three dimensional. Compare:

(i) Bill likes himself

(ii) Which pictures of himself did Bill like?


(ii) Dislocation

Where an item in a structure is phonologically spelt out, is not necessarily where it is construed. Passives offer a clear example; also see the ‘tough movement’ case below.


(iii) Empty categories

Some elements of structures are not spelt out, e.g., The boat was sunk to collect the insurance. The boat didn’t collect the insurance, the sinkers did.


The transformational map from Deep Structure to Surface Structure was an attempt to make sense of these phenomena.


The Early Years

The theory of The Logical Structure of Linguistic Theory (1955-56) and Syntactic Structures (1957) did not appeal to either Deep Structure or Surface Structure. The architecture proposed consisted of


(i) a context-free rewrite base (essentially a Post system, demonstrably Turing equivalent), inclusive of the lexicon; and

(ii) a transformational component that takes P-markers - structures generated by the base - as argument, and produces P-markers as value. Transformations are either optional, or obligatory. A kernel sentence is one whose structure is generated just by obligatory transformations. All further transformations take such structures as argument.


The architecture as a whole constitutes an algebra:

L = [L , f , R1 ,…, Rm , l , F , w1 ,…,  wn]


(1) L = set of primes (primitive elements)

(2) f = concatenation

(3) R1 ,…, Rm = set of classes/relations defined over (1)-(2), where R1 is =

(4) l = set of constructed objects by (1)-(3)

(5) F = map to descriptive conditions

(6) w1 ,…,  wn = set of classes/relations defined over L and L¢, L¢,…



(i)     S t NP + VP

(ii)    VP t Verb + NP

(iii)   NP t Npsing, NPp

(iv)   Npsing t (T) + N + Ā

(v)    NPpt T + N + S

(vi)   T t the, few, every,…

(vii)  N t Bill, Mary, man,…

(viii) V t kiss, love, kill,…

(ix)   Aux t C(M) (have + en) (be + ing)

(x)    M t may, shall, will, can, must


(PM)                                                 S





       NP                                                                                            VP



                                                                           Verb                                                 NP


 NPsing                                                 AUX                              V                              NPsing       





     Bill                             C may have en be ing                      kiss                              Mary



A transformation takes an abstract phrase marker (not a sentence), as specified by a structural analysis, and maps it onto another phrase marker. Thus, each transformation is fully specified by a structural analysis (SA) of the input and a specification of the structural change (SC).


A base generated structure, such as the terminal string of PM, is not a sentence, for it has no number/tense. There must, therefore, be an obligatory transformation:


Number transformation:

SA: X - C - Y

SC: C t s, if X = Npsing

               Ā in other contexts

               past in any context


This generates the sentences:


Bill kisses Mary

Bill kissed Mary

Bill may kiss Mary

Bill may have kissed Mary

Bill may have been kissing Mary

Bill will kiss Mary, etc.

Bill has kissed Mary, etc.

Bill is kissing Mary, etc. 


Any sentence represented by a structure which results from obligatory transformations is a kernel sentence: a mono-clausal sentence in the active.


There are optional transformations. These take the structure of kernels as input, and produce a non-base generated structure. Passive is optional in this way.


Passive transformation:

SA: NP - AUX - V - NP

SC: X1 - X2 - X3 - X4 t X4 - X2 + be + en -X3 - by - X1


Why Transformations?

The basic reason is that re-write grammars are inadequate. They merely ‘list’ the structures available without explaining them. We can see this in numerous ways.


Take the AUX rule:

Aux t C(M) (have + en) (be + ing)


This tells us that HAVE introduces the perfect morphology en, and BE introduces the progressive morphology ing. So far so good, but it simply lists this; it does not explain its curious distribution. In particular, while HAVE introduces en, it is affixed to the following ‘verblike’ element; similarly, ing is introduced by BE, but is affixed to the following ‘verblike’ thing. In other words, the dependency of the morphology is cross serial, but re-write grammars can only represent serial relations, such as nesting:





                                            have     be      en      kiss     ing


         Further examples are abundant in the ‘felt relations’ between, say, indicatives and interrogatives, actives and passives, topicalisation, structural ambiguities, etc. Transformations can explain these relations by taking the linear morphological order as generated from an underlying structure.


Deep Structure and Surface Structure    


Where does meaning fit in? A not uncommon thought is that Chomsky, at least at the time of Syntactic Structures, wholly ignored the issue of meaning. This is false. The confusion arose from a misunderstanding of the so-called autonomy thesis


(AT) Semantic information is not required for the selection of a grammar, i.e., the base and transformations are not derivable from semantic information.


As far as we can tell, AT was as true then, as it is now. But it does not follow that the syntax tells us nothing about semantics. The following claim suggests itself:


(S-S) The meaning of a sentence is determined by its underlying kernel and its transformational history from the kernel. That is, a theory of meaning need only target the kernels and the transformations.


Chomsky said this was no more than a suggestion, which, if proposed in generality, would face a number of problems. Still, the idea was developed into the Katz-Postal hypothesis (An Integrated Theory of Linguistic Descriptions, 1964), which Chomsky accepted in Aspects (1965):


(K-P) Transformations are semantically irrelevant, i.e., interpretation targets the underlying kernel alone.


Assume K-P is true. It follows that the meaning of a sentence must be wholly determined by its kernel. Chomsky re-christened kernels deep structures, and derived phrase markers surface structures. This provides the following model:


Deep Structure Ž Semantics


                  r Transformations


Surface Structure Ž Sound


This is the model assumed in Cartesian Linguistics.


An Example

Take the sentence The Invisible God created the visible world. Let its deep structure be as follows


(DS)                                                          S




        NP                                                                                                     VP




the         God      S                                                          created




      NP              VP                                                                                    NP



    God        is       invisible                                                         the         world          S




                                                                                                                       NP              VP 



                                                                                                              world      is     visible


The structure captures the inferential relations from the single sentence to (i) God is invisible; (ii) the world is visible; and (iii) God created the world.





(i) Substitute the relative pronouns who and which for the respective nouns of the embedded Ss (the incident propositions).


SA: X - Y - Z, where X = Z

SC: X1 - X2 - X3 - X4 t X1 - wh - X3 - X4  


This gives us: The god, who is invisible, created the world, which is visible.


(ii) Delete the wh word and the copula in both embedded Ss.


(iii) Invert the remainder of the reduced clauses with the heads of their respective dominating NPs.


This gives us: The invisible God created the visible world.


Properties of deep structure (circa ‘66)

(i) The interface with semantics.

a. Subcategorization

b. Selection restriction

(ii) Universal.

(iii) The output of a now recursive base.

(iv) The site of lexical insertion, form a now independent lexicon.


Properties of surface structure (circa ‘66)

(i) The interface with sound.

(ii) Variable.

(iii) The output of the transformational component.



Changes Since 1966

In 1966 was approximately the last time there was a consensus on the architecture of the language faculty. Soon were to begin the so-called linguistic wars, which were to do with the status of deep structure. Chomsky (Studies on Semantics in Generative Grammar, 1972) was to reject the Katz-postal hypothesis and contend that surface structure essentially contributed to meaning, in the form of, e.g., scope, presupposition, focus, etc. Ironically, these were precisely the problems which had stopped Chomsky proposing K-P in 1957. At the same time, the base was finally dropped, to be replaced by X-bar geometry, which amounted to a new theory of deep structure:


(i) XP t ZP X'

(ii) X'* t X' YP

(iii) X' t X0 YP


Where ZP and YP are phrases, X0 is a lexical item, and XP is the projection from the item. In terms of a sentence, ZP would be the subject, YP the VP, and XP the inflectional projection which heads the whole structure. The basic geometry fits any phrase whatsoever, permitting a radical minimization of the base.







                                                  ZP                    X'




                                                                          X0                   YP



X-bar theory was the corner stone of the so-called ‘government and binding’ approach (Chomsky, Lectures on Government and Binding, 1981), which differed from the Aspects model in various ways, principally, deep structure (now called D-Structure) is no longer the interface with semantics; a new level was introduced - LF - for that purpose. Similarly, surface structure (= S-Structure) doesn’t interface with sound; instead the level of PF is now the interface:


D-Structure [X-bar schemata, theta assignment]




                     S-Structure [BT, Case, Bounding theory]




     A-P Ü PF                               LF Ž I-C


The minimalist program (initiated in Chomsky, The Minimalist Program, 1995) rejects D-Structure and S-Structure, to leave just the two interface levels.







Minimalist Architecture (Circa 1995)















                           A-P Ü PF                     * [SPELL-OUT]








                                                     LF Ž I-C  


The levels of D-structure and S-Structure are dispensed within on two grounds.

(i) Firstly, neither is an interface level. Effectively, then, both were posited for internal reasons, to satisfy descriptive demands, not on the basis of virtual conceptual necessity. On the assumption that the language faculty is perfect, all properties of the faculty should follow from what is demanded of the interfaces. In this light, the levels are departures from perfection.

(ii) The information the levels are intended to encode can be encoded at the interface or in the lexicon. Indeed, D-structure only problematically meets its demands of theta-realisation. Consider ‘tough movement’ (Chomsky, 1981, p.309) cases:


(1) Bill is easy (for anyone) to please.

(cf. It is easy (for anyone) to please Bill.)


The matrix Bill cannot be inserted at D-structure, for the position is not theta-governed, i.e., it doesn’t realise an agent position. But nor can Bill be inserted as object of please and then move - via transformation - to matrix position. But D-structure is where lexical insertion is supposed to take place! Things get worse. Consider


(2) A man who is easy (for anyone) to please is easy (for anyone) to convince.


One might be persuaded to insert Bill anywhere in the derivation, but it is certainly much

harder to accept the lexical insertion of a man who is easy (for anyone) to please, in the very same context where Bill creates a problem in (1). The very same transformation that raises difficulties for the analysis has taken place inside the subject in this instance.


When Howard Lasnik raised this difficulty in Chomsky's class - so the anecdote goes - Noam paused for an unusually long time, to then concede what Lasnik was asking him to accept: the need for generalized transformations. Of course, if the latter are part of the grammar, we can (in minimalist terms) merge not just Bill, but in fact also a phrase of arbitrary complexity, such as the (previously assembled) a man who is easy (for anyone) to please. But to concede this point, is to reject D-structure, for the whole purpose of that level of representation was to create a unified object to express configurational relations prior to the occurrence of any transformation.


In the last few years, proposals have been advanced against the interface levels. This, again, is a drive for perfection: levels, with their sui generis features, appear to be a departure from perfection. As Chomsky writes:



The principled elements of S0 are the conditions imposed on FL by the systems with which it interacts. If language is to be useable at all, its design must satisfy on “interface condition” IC… The goal is to determine just what aspects of the structure and use of language are specific to the language faculty, hence lacking principled explanation at this level.

      [I]nitial conditions…


(i)   unexplained elements of S0

(ii)  IC (the principled part of S0)

(iii) general properties


Principled explanation, going beyond explanatory adequacy, keeps to (ii) and (iii). An extremely strong minimalist thesis: (i) is empty.

¾ N. Chomsky, ‘Beyond Explanatory Adequacy’, unpublished.








Level-Free Architecture













                                         A-PÜ · Ž I-C








                                         A-PÜ · Ž I-C








                                         A-PÜ · Ž I-C