Review of The Minimalist Program by Noam Chomsky
This collection of Chomsky’s recent technical papers (henceforth MP) documents what may well herald a third major breakthrough in the study of language and mind. At the very least, it provides a radical reformulation of certain parts of linguistic theory based on some broad speculations about language design and how it might accommodate the architecture of the mind/brain.
The first three chapters of MP are reprinted with minor revisions from previous publications. Ch. 4 ‘Categories and transformations’ is a substantial elaboration of Chomsky 1994, which was published under a different title. The eleven-page introduction, most of it taken from §1 of Chomsky 1994, puts the following four chapters in historical/conceptual perspective. Ch. 1 ‘The theory of principles and parameters’ (co authored with Howard Lasnik) presents a general overview of the PRINCIPLES AND PARAMETERS framework circa 1991, one which is transitional between a government based and an economy-based theory because it incorporates notions of economy discussed in the second chapter (written almost three years before, in September 1988). Ch. 2, ‘Some notes on economy of derivation and representation’, elucidates proposals for economy conditions (e.g., FULL INTERPRETATION (Chomsky 1986a, henceforth FI) and LAST RESORT), which Chomsky cautions should be viewed as guidelines rather than principles of Universal Grammar (UG) because their formulations are too vague. These guidelines are intended to ensure that representations contain no superfluous elements and derivations no superfluous steps. The last two chapters deal specifically with a minimalist view of grammatical theory. The third chapter, ‘A minimalist program for linguistic theory’, sketches goals and directions for research under minimalist assumptions—in particular, the elimination of D-structure and S-structure as levels of linguistic representation—and further develops the economy-based view by adding new conditions, PROCRASTINATE and GREED. Ch. 4 subjects the basic framework for a theory of UG under the principles and parameters model to a searching critical analysis ‘in an effort to approach as closely as possible the goals of the Minimalist Program outlined in the introduction’ (219). It presents a radical reformulation of phrase structure theory and a significant revision of the theory of transformations. As Chomsky notes, ‘the end result is a substantially different conception of the mechanisms of language’ (219).
MP is unique among Chomsky’s published works because of the theoretical distance it covers between the first chapter and the last. For example, while Ch. 1 contains a substantial section on government theory and Ch. 2 refers to the EMPTY CATEGORY PRINCIPLE (ECP) throughout, Ch. 3 suggests that government is not a legitimate grammatical concept (its effects derivable from other more basic notions), and Ch. 4 essentially assumes that the notion is not available for grammatical description. Not only does the ECP disappear from the landscape by the time we reach Ch. 4, but so do the Case filter, the projection principle, the head movement constraint, the subjacency condition, most of X-bar theory, and parts (potentially all) of the θ-criterion. This alone makes MP a highly challenging book. Moreover, central concepts and mechanisms undergo quite significant reformulation as the discussion proceeds (e.g., the formulations of checking domain in Chs. 3 and 4, a minor readjustment compared to the discussions of phrase structure and transformations in Ch. 4).
Without the fourth chapter, MP would still be Chomsky’s richest and most complex collection of essays. Ch. 4, arguably his most difficult paper to date, raises MP to another level entirely. At 175 pages, this chapter makes up almost half of MP and is just about equal to the length of Chomsky’s two Linguistic Inquiry monographs combined. Unlike the first three chapters or the monographs, Ch. 4 focuses on very broad and abstract conjectures about language design and the theoretical constructs that account for it. Where most of the numbered examples in Chs. 1–3 contain linguistic expressions from some particular language, those of Ch. 4 deal mostly with definitions, questions, conjectures, assumptions, abstract structures, lists of relevant topics or properties of constructions, and of course, various principles, conditions, guidelines, and just plain stipulations—less than half cite linguistic expressions. As a result, the empirical basis for much of the discussion is either assumed or passed over. Chomsky acknowledges that the ‘empirical questions that arise are varied and complex, and it is easy enough to come up with apparent counterevidence’, but states clearly that he is putting these problems aside and assuming ‘the best outcome, namely, that UG settles the matter’… ‘hardly an innocuous step’ he cautions (266). Thus Ch. 4, unlike much of Chomsky’s previous writing, focuses on conceptual as opposed to empirical considerations and eschews detailed analyses of linguistic data.
From the way Chomsky talks about conceptual arguments in contrast to empirical arguments it is pretty clear that he prefers the former, though ideally conceptual and empirical considerations should eventually converge (as they did in much of the work that developed the principles and parameters framework, where the replacement of language-specific grammatical rules by mechanisms and principles of UG led to a new model of comparative syntax). With the exception of Ch. 2, the book contains many references to ‘conceptual necessity’, ‘conceptual grounds’, ‘conceptual arguments’, and ‘conceptual naturalness’ (most frequently in Ch. 4). Consider for example the following comment about a model of grammar that includes levels of D-structure and S-structure:
The empirical justification for this approach, with its departures from conceptual necessity, is substantial. Nevertheless, we may ask whether the evidence will bear the weight, or whether it is possible to move towards a minimalist program. (187)
As part of the minimalist program, Chomsky wants to eliminate the levels of D-structure and S-structure simply because they fail to meet the criterion of conceptual necessity in spite of substantial empirical justification.
This methodology is central to the minimalist program, which seeks to limit the theory of grammar to what is conceptually necessary. In the case of levels of representation, this includes the interface levels of PF and LF and excludes all others. Empirical motivation for such a move may be unknown or even nonexistent. If so, then the conceptual argument suffices. The empirical motivation for the contrary conclusion has to be reanalyzed. This is of course quite reasonable, given that empirical arguments themselves are often based on assumptions whose justification is less than solid. Conceptual advances often result in the reevaluation of empirical evidence and arguments. Thus the criterion of conceptual necessity operates in the minimalist program as a type of Ockham’s razor. Exactly what constitutes conceptual necessity in the case of levels of representation is particularly clear, but in other domains may be open to interpretation.
This style of reasoning, which has to some extent always been in the background of Chomsky’s work, is likely to strike most readers as rather alien and difficult to deal with, primarily because before the advent of the minimalist program most research in the field has been data-driven. This does not mean that Chomsky is no longer concerned with constructing an explanatory account of the properties of natural language. Rather, it seems that he has become increasingly skeptical that such an account can be constructed primarily on empirical arguments—which perhaps indicates an intensification of his critique of empiricism, now applied to the methodology of generative grammar. The result is that Chomsky has moved the principles and parameters model to unfamiliar ground. Not only is it difficult to evaluate conceptual arguments, but it is also unclear exactly how to proceed, if for the present empirical arguments are not strong enough to move the program forward—in marked contrast to the situation when Lectures on government and binding appeared in 1981.
Not surprisingly, the motivation for proposing a minimalist program is essentially conceptual, though of course based on considerable empirical research within the principles and parameters framework. It comes from two related conjectures, stated in MP as questions.
a. To what extent is language a ‘perfect system’?
b. Is the computational system for human language (henceforth CHL) optimal (‘in some interesting sense’) and therefore unique among cognitive systems?
Putting aside expected ‘imperfections’ in morphology and phonology Chomsky is optimistic that both questions will receive positive answers. He mentions two criteria for establishing the perfection of the system. One is that language should meet a condition of INCLUSIVENESS: the computation that generates structures (especially PF and LF) applies only to the elements of lexical items initially selected for the computation (the NUMERATION in the new terminology), and is restricted to rearrangements of lexical properties—thus the computation cannot create new objects (e.g., indices or bar-levels).4 The other criterion concerns the extent to which the language faculty is determined by general conditions, that is, without postulating additional special structure.5 Such conditions in turn depend on (1) how the language faculty interfaces with other cognitive systems and (2) ‘general considerations of conceptual naturalness that have some independent plausibility: simplicity, economy, symmetry, non-redundancy’, and so forth. At the moment we know almost nothing about (1) and considerably more about (2), which provides some basis for believing that (a) can be answered in the affirmative. The optimality of CHL will also be determined by the extent to which it conforms to some of the same considerations of conceptual naturalness, especially, economy. With these considerations in mind, let us turn to the set of assumptions that constitute Chomsky’s minimalist program.
The minimalist program shares at least three assumptions with its predecessors. First and foremost, there exists a language faculty in the mind/brain, a component that interacts with other cognitive systems. Next, the cognitive system of language connects, via levels of linguistic representation, with performance systems. Finally, performance systems do not differ across languages. Even at this fundamental level Chomsky remains cautious, noting that these assumptions are ‘not at all obvious’.
As the minimalist program is grounded in the principles and parameters model,6 the two share further specific assumptions [see 17]:
regarding CHL, the initial state S0 contains invariant principles and parameters (i.e., options restricted to functional elements)
selection Σ of parameters determines a language
a language determines an infinite set of linguistic expressions (SDs), each pair (π, λ) obtained from the interface levels, PF and LF
language acquisition involves fixing Σ
the grammar of language states just Σ, lexical arbitrariness and the PF component aside
Because these assumptions have been standard for over a decade, we can dispense with further discussion here.
There is one further assumption, articulated for the first time in Ch. 3, which could have been considered prior to the minimalist program. Chomsky puts forward as a narrow conjecture the claim that there is no variation in the overt syntax or the LF component. Ignoring PF options and lexical arbitrariness, Chomsky wants to show that Variation is limited to nonsubstantive parts of the lexicon and general properties of lexical items’ (170) (but see n. 7 in this review). If this is correct, then ‘there is only one computational system and one lexicon’, apart from the variation mentioned. Whether this can be convincingly demonstrated may be a distant prospect, but it comes close to saying that human languages share a central core so that at the appropriate level of abstraction they are the same language—a bold and fascinating conjecture.
We come at last to those assumptions that are unique to the minimalist program. The fundamental assumption from which the rest more or less follow is that the theory of grammar must meet a criterion of conceptual necessity (as mentioned above). In fact it is so basic that it is never explicitly mentioned as an assumption. It leads directly to the assumption that the interface levels LF and PF are the only relevant linguistic levels, in spite of apparent empirical evidence to the contrary (169). Chomsky further assumes that (i) all conditions are interface conditions and (ii) a linguistic expression is the optimal realization of these conditions (194).
The discussion of interface conditions in MP is rather vague. Only the Case filter, now interpreted as a general condition on morphological feature checking, is explicitly mentioned (197). On p. 189 Chomsky refers to ‘(external) interface conditions’, presumably to leave open the possibility that there are internal interface conditions as well (i.e., output conditions that belong to CHL—for example, the various conditions of the binding theory perhaps). In Ch. 4, however, he discounts this possibility: ‘I will continue to assume that the computational system CHL is strictly derivational and that the only output conditions are the bare output conditions determined “from the outside,” at the interface’ (224), specifically the perceptual-articulatory interface for PF8 and the conceptual-intentional interface for LF. Thus CHL contains only conditions on the application of rules, whereas cognitive systems that use information from CHL impose conditions on representations (filters), BARE OUTPUT CONDITIONS (henceforth BOCs) in the new terminology.9 (For further elaboration of the derivational nature of CHL, see Epstein 1994 which proposes that ‘the only relations that exist for CHL are those established by the derivational process itself (254).) A related assumption of the minimalist program is that BOCs determine which items are ‘visible’ for computations (242). The example cited is that only maximal and minimal projections are accessible to CHL given BOCs.
Although Chomsky offers no specific proposals for BOCs, the principle of FULL INTERPRETATION (FI) might be considered a candidate because it operates as an output condition on PF and LF which is imposed by the cognitive systems that use interface representations. From the generally informal discussion of this principle in MP, we might construe FI to require that every element in the interface representation must be interpretable to the cognitive system that uses it. This assumes that the external cognitive system will mark an interface representation as deviant if it contains an uninterpretable element, which requires that external cognitive systems can recognize uninterpretable elements even though they cannot assign them a legitimate interpretation.
Under this conceptualization, the notion of interpretability plays a crucial role. This is developed in Ch. 4 in terms of a binary distinction ± interpretable for features. Categorial features and of nouns are+interpretable, whereas Case features generally and of verbs are -interpretable. When a [-interpretable] feature occurs at LF, for example, FI is violated and the derivation is said to CRASH. If only [+ interpretable] features occur at LF, the derivation CONVERGES. A convergent derivation must converge at both interface levels and thus must eliminate any [-interpretable] feature introduced via the selection of lexical items. Consider how this works for the following examples.
The Case feature of John is licensed in 2 but not in 1. In MP licensing is achieved via a checking procedure (see §4.5.2 for details) which results in the elimination of the— interpretable feature. Thus the derivation of 1 crashes because the uninterpretable Case feature remains at LF violating FI, whereas the derivation of 2 converges. The effects of the Case filter for lexical NPs are subsumed under FI, now interpreted as a BOC, so the former condition can be dropped from the inventory of UG. In this way, BOCs do seem to determine the structure and function of CHL itself, as Chomsky suggests.
Another fundamental assumption of the minimalist program is that phrase structure representation is bare—meaning that it is composed of lexical features and objects constructed from them. Thus phrase structure is assumed to meet the condition of inclusiveness mentioned above. As a result, a large part of X-bar theory is discarded, in particular, bar levels and rule schema.
It is worth recalling that as the principles and parameters framework developed, it became increasingly clear that phrase structure rules were largely redundant language particular stipulations that basically followed from principles of UG in conjunction with parameter settings. However, for well over a decade Chomsky made no specific proposal for how phrase structure was to be generated if there were no phrase structure rules. The proposal in Ch. 4 is that phrase structure is constructed via a transformational operation MERGE, which concatenates two elements (lexical items and/or concatenations of lexical items) and projects the category label of one on the newly formed constituent— basically classical adjunction (which does not produce a two-segment category, as in the analysis of Chomsky 1986b, which adapts a proposal in May 1985). Interestingly, this analysis establishes transformations as fundamental (and phrase structure rules as unnecessary) in contrast to earlier attempts to show the converse (e.g., Harman 1963 and Freidin 1970).
Merge interfaces with the lexicon via a numeration N, a set of lexical items (including functional elements) selected from the lexicon for a derivation. Merge must apply to each item in N. If some item is selected more than once (e.g., he in he thinks he is clever), then merge must apply to that item the same number of times it was selected. If N contains any unprocessed elements, what has happened does not count as a derivation.
Merge constructs phrase structure bottom-up, in contrast to the top-down nature of derivations under some versions of phrase structure grammar. Merge cannot construct a VP containing a verb and a sentential complement until the sentential complement has been constructed. Furthermore, merge cannot construct classical D-structures where all lexical elements are inserted into phrase structure prior to the application of movement transformations. For example, the classical D-structure representation of we expect John to be arrested soon would have John in complement object position and we in matrix subject position. With bare phrase structure, although John is inserted into complement object position by merge, it must be moved to complement subject position by an operation MOVE to construct the sentential complement before the matrix predicate can be formed. The analysis rests on an assumption that the lexicon does not contain empty categories that have no interpretation (in contrast to PRO). Thus there is no strict ordering between lexical insertion and movement operations, as there is in models which include D-structure.
While classical D-structure disappears under the new analysis of phrase structure, its central property is maintained, namely, that arguments are inserted by merge into their canonical D-structure positions (i.e. where they are assigned θ-roles). Presumably this is forced by selectional properties of predicates—otherwise it would be mysterious under minimalist assumptions why the position of arguments would be relevant anywhere except at LF.
Bare phrase structure theory raises the question as to which category label of the two elements concatenated by merge or move projects. Section 4.4.2 (projection of target) of Ch. 4 is devoted to showing how for movement it is always the target that projects. In essence the minimalist program raises such questions as why a VP is a ‘VP’ and not something else. To account for the order of constituents within phrases, another basic issue, Chomsky adapts the LINEAR CORRESPONDENCE AXIOM (Kayne 1994), which is based on the idea that ‘order reflects structural hierarchy universally’ (335). These are basic questions whose answers were simply assumed as stipulated in earlier analyses. In raising such questions about fundamental points of analysis, the minimalist program focuses on the root and beginning of syntactic inquiry, not the ‘end of syntax’ as has been claimed (e.g., Marantz 1995).
Given BOCs, perhaps just FI, there must be a point in the derivation where one part splits off in the direction of PF and the rest proceeds to LF. Chomsky assumes that the presence of phonetic features at LF will cause a derivation to crash. This entails that an operation of spell-out, which splits a derivation into the two parts, extracts those aspects of SDs relevant to PF (e.g., phonetic features). Given FI, no element of a numeration having phonetic features can be merged after spell-out. Similarly, an element bearing semantic features cannot be merged after spell-out in the derivation to PF. Presumably the point at which spell-out applies in a derivation is determined by BOCs, so no stipulations are necessary. This accords with what Chomsky calls a ‘guiding intuition’ of the minimalist program: ‘that operations apply anywhere, without special stipulation, the derivation crashing if a “wrong choice” is made’ (231).
Operations that apply pre-spell-out are overt because their consequences show up in PF as well as LF—thus post-spell-out operations are covert, their consequences showing up only in LF. In this way CHL can be divided into an overt and a covert component. The distinction would be a potential problem for a minimalist approach if a difference between the two components had to be stipulated, which would thus specify when spell out applied. Instead, Chomsky takes the uniformity of CHL from the numeration to LF to be an assumption of the minimalist program—in other words, the same kind of operations apply in both the overt and covert components. If this is correct, then any distinction between pre- and post-spell-out should be a reflection of other factors.
There are two ways of meeting the uniformity assumption. Either the operation move applies to categories in both components (the analysis of Ch. 3) or move applies only to features (the analysis of Ch. 4). Since only feature movement is required for checking in the covert component, the minimal operation would be MOVE FEATURE, not MOVE CATEGORY.17 Chomsky notes that the latter is ‘an unnatural interpretation of the operation’ on general minimalist assumptions (262).
The move-feature analysis creates a major problem for an account of overt movement, which involves categories, not just features. If only features are required to move in order to check and thereby eliminate -interpretable features from a derivation, then why does overt movement involve more than just features? Moreover, why does it involve single constituents (rather than strings of categories that do not form constituents)? The move category analysis answers such questions by mere stipulation. Chomsky’s analysis in Ch. 4 is, however, somewhat less than satisfying. He proposes ‘a natural economy condition’: a feature which moves ‘carries along just enough material for convergence’, ‘a kind of “generalized pied-piping’”, determined by BOCs. This displacement property of overt movement, the fact that constituents in sentences can occur in a position different from the one in which they are interpreted at LF, might actually constitute ‘a striking departure from minimalist assumptions in language design’, Chomsky cautions at the outset of Ch. 4 (221). If so, then one of the most interesting properties of human language falls outside the domain of the core theory. In spite of this apparent problem with overt movement, the morphologically driven feature-movement analysis is conceptually attractive, in part because it raises fundamental questions that were previously buried under a stipulation. Furthermore, it suggests a natural explanation for the economy condition PROCRASTINATE, which prefers covert movement to overt movement. Since overt movement involves an additional operation (pied piping), pure feature movement in the covert component is always preferable on general economy grounds.
Given this incomplete sketch of Chomsky’s minimalist proposals, let us consider the analysis of some familiar examples of NP movement.
In 3 the Case feature of John will raise to the matrix Infl to be checked and to check the Case feature on Infl itself. In addition, the NP John will be pied piped to Spec IP for reasons of convergence that remain to be specified. The analysis is basically straightforward. In contrast, the analysis of 4 is open to a couple of interpretations. We know that the Case feature of John can only be checked once, since being—interpretable it becomes invisible to CHL (deletes) after it is checked. If John checks the Case feature of the complement Infl, the last resort condition, which allows a feature to move only if it enters into a checking relation with its target, might block any further movement. However, given the VP-internal subject hypothesis, John would have been inserted in the complement VP and therefore could have moved directly to the matrix position. In this case John will be able to check the Case feature of the matrix Infl; the movement however would violate the minimal link condition by skipping the complement Infl. But then both the D-feature and the Case feature of the complement Infl will not be checked. In either case, we are left with -interpretable features at LF, in violation of FI.
The analysis of 4 shows that both last resort and the minimal link condition appear to overlap with FI. If elimination of redundancy is a methodological prerequisite of the minimalist program as Chomsky suggests, then it seems that we still have a fair amount of sorting out to do. The overall character of the minimalist program is highly speculative, as Chomsky notes throughout MP. In a recent paper (Chomsky 1998) he is virtually categorical on this point: ‘There are minimalist questions, but no specific minimalist answers’ (recall n. 6). Whatever answers can be discovered will result from following the research agenda of the program. Unfortunately, how this is to be done is rather unclear as compared to the research agenda outlined in Chomsky 1981.
With this in mind, consider the issue raised at the outset of this review: does the minimalist program as outlined in MP constitute a major breakthrough in the study of language and mind? As we have seen, the minimalist program raises fundamental new questions that did not arise in previous work. It also leads to a radical reformulation of syntactic theory within the principles and parameters framework. The phrase structure component is reduced to a fundamental part of the transformational component. Furthermore, the modular subcomponents of UG in standard GB theory are essentially eliminated or substantially reduced in favor of economy conditions and BOCs. The subtheory of government is disallowed for conceptual reasons. Case assignment and licensing, and also Case conflicts are handled indirectly by FI and other conditions on economy, so that there seems to be no need for a special Case theory module. The empirical effects of the θ-criterion are subsumed under FI (the fact that an argument must be assigned a θ-role (but see n. 10)) and last resort, which generally prevents overt movement between two θ-positions, with perhaps the exception of overt control structures (see Hornstein 1996), so there may be no special θ-theory apart from some mechanism for assigning θ-roles to arguments. Under minimalist assumptions, binding theory too must undergo some major reformulation (as suggested in Chs. 1 and 3)— especially if the use of indices is excluded under an inclusiveness condition (see Freidin 1997b for discussion). While these are major changes, the extent to which they will lead to new and deeper insights into the nature of language and mind largely remains to be determined. In terms of empirical payoff, the minimalist program in its present state does not yet constitute a breakthrough of the same sort as the discovery of transformational generative grammar or the major conceptual shifts of the principles and parameters framework (see Freidin 1994b).
Nonetheless, there is much about the minimalist program that is promising. Bare phrase structure theory in and of itself is intuitively very appealing. So too is the general idea that the nature of CHL, its function and output, is partly determined by the cognitive systems that access its interface representations. This new perspective requires a broader and more integrated approach to the cognitive system of language by raising as a central question how it interacts with other cognitive systems. Furthermore, the minimalist program has a special therapeutic value beyond the question of its validity in that by focusing on conceptual justification of assumptions it discourages ‘the temptation to offer a purported explanation of some phenomenon on the basis of assumptions that are of roughly the order of complexity of what is to be explained’ (233). To be sure, this is just good scientific practice, but this is the first time Chomsky has made it a serious issue, indicating perhaps a belief that linguistic theory might have developed to the point where it makes sense to invoke such rigorous strictures. So even though it may take a while before we have a strong sense that it is on the right track, the minimalist program seems worth pursuing.
In the first paragraph of the first chapter of Syntactic structures, Chomsky enunciated one of the goals of the study of syntax as follows: ‘The ultimate outcome of these investigations should be a theory of linguistic structure in which the descriptive devices utilized in particular grammars are presented and studied abstractly, with no specific reference to particular languages.’ In MP, especially the fourth chapter, Chomsky has taken a giant step forward towards realizing this goal. From this perspective, the minimalist program constitutes a major breakthrough. In raising theoretical discussion of the principles and parameters model to a new level of abstraction, it offers us a glimpse of a potentially simpler and deeper theory of human language (with of course the important caveat that it is actually on the right track). However, in doing this, Chomsky has introduced a kind of discontinuity between our previous understanding and the new one he is trying to achieve. The situation is somewhat reminiscent of the one in mathematics noted in Courant 1937.
The point of view of school mathematics tempts one to linger over details and to lose one’s grasp of general relationships and systematic methods. On the other hand, in the ‘higher’ point of view there lurks the opposite danger of getting out of touch with concrete details, so that one is left helpless when faced with the simplest cases of individual difficulty, because in the world of general ideas one has forgotten how to come to grips with the concrete. The reader must find his own way of meeting this dilemma. In this he can only succeed by repeatedly thinking out particular cases for himself and acquiring a firm grasp of the application of general principles in particular cases; herein lies the chief task of anyone who wishes to pursue the study of Science.
