2.1 Filters or patches?

A notable example concerns constraints on extraction, which are sometimes regarded as a towering achievement of generative syntax. It is indisputable that previous traditions did not formulate intricate constraints to bar the questioning of deeply embedded oblique dependents, the relativization of individual conjuncts from larger coordinate structures, the clefting of subconstituents within clausal dependents of nouns, and other similarly aberrant patterns. However, as C&J note, it is also true that few if any antecedent traditions assumed that learners would massively overgeneralize the questions, relative clauses and cleft constructions that they do encounter, and thereby run the risk of producing the types of patterns barred by explicit constraints on extraction.

The basic issue that C&J are raising is whether close to a half-century of research into island constraints has yielded genuine insights into extraction constructions or whether this work has merely clarified descriptive shortcomings of unrestricted movement operations. Although C&J refer to the later operations Move wh and Move α, non-local movement devices are present from the earliest transformational accounts. In Syntactic Structures, the movement transformation T_w applies to a string of the form X−NP−Y, ‘where X and Y stands for any string’ (Chomsky Reference Chomsky1957: 69), and derives a string of the form NP−X−Y. The body of proposals that have come to be known as ‘island constraints’ represent attempts to restrict variables such as X. The first general constraint of this type was the ‘A-over-A’ principle of Chomsky (Reference Chomsky1964), which barred extraction of an element of a given category from within a constituent of the same category. To overcome problems with this excessively general condition, Ross (1967) proposed a set of more specific constraints that referred to particular categories, such as the Complex NP Constraint and the Sentential Subject Constraint, or to particular constructions, as in the case of the Coordinate Structure Constraint. The ‘Ross constraints’ led in turn to attempts to generalize the restrictions on extraction, beginning with the programmatic conditions on transformations in Chomsky (Reference Chomsky, Anderson and Kiparsky1973). The goal of unifying island constraints even played a role in the development of some short-lived frameworks, such as Chomsky (Reference Chomsky1982), which explored the ‘functional determination’ of null elements, or Chomsky (Reference Chomsky1986), which developed a notion of ‘barriers’ to extraction. But a ‘grand unification’ of the Ross constraints remained elusive, as each of the initiatives explored a different strategy for expressing the disjunctions that held together disparate sets of corrective conditions.

Those working within the Chomskyan tradition may discern progress in these developments, and argue – at the very least – that this work has succeeded in closing off avenues of research that might have seemed to have some initial plausibility. But those who are not already convinced of the central importance of constraining constituent structure displacements may find little evidence of incremental progress. C&J invite this audience to entertain the possibility that the entire enterprise has been based on the wrong intuition. The problem raised by extraction phenomena takes on an entirely different character if one starts from the assumption that word order dislocation is local in the general case, restricted to domains such as finite clauses, and that non-local dislocations represent extensions of local patterns. From this perspective, one might expect different non-local extensions to become established in different languages, much as different clause-internal word order patterns do. There may be syntactic – alongside functional or communicative – motivation for particular extensions, just as there is for clause-internal patterns. So the challenge of characterizing extraction phenomena would not go away, though the challenge could be met by a more language- or construction-specific account. The task of describing (or acquiring) extraction patterns would not involve the ‘parameterization’ of general constraints on variables or domains but would instead involve the learning of non-local word-order alternations, in essentially the same way that local alternations or other constructional patterns are learned (Culicover 1999).

A constructional treatment of extraction phenomena, one which focuses on overt, observable patterns that are present in a learner's input, is particularly compatible with the conservative learning strategies that learners appear to employ (Pullum & Scholtz Reference Pullum and Barbara2002, Tomasello Reference Tomasello2003). Hence the shift in perspective to an input-driven or usage-based perspective would avoid the need to prevent speakers from ‘generalizing wildly beyond their experience’. From this standpoint, the key insight expressed by generative constraints on extraction would just be that any account that assumes unconstrained generalization requires corrective and ultimately heterogeneous conditions to rein in this process.

The same shift undermines the motivation for ‘post-syntactic’ counterparts of extraction in languages such as Chinese and Japanese, which exhibit no ‘overt’ dislocation of question phrases. A distinctive feature of Chomskyan accounts from the early 1980s is the idea that ‘wh-phrases in these languages move covertly, to occupy the same position in LF as they do in English’ (71f.). As C&J note, covert movement ‘was motivated by the claim that the distribution of such in situ wh-phrases is constrained by the same principles as overt movement’. However, they also observe that the empirical support for post-syntactic movement is greatly weakened by discrepancies between overt and LF movement.

Efforts to generalize constraints on extraction also had a more subtle effect on the development of Chomskyan approaches, as they fostered a division of labour between unrestricted movement rules and conditions that ‘filtered’ rule output. As the task of constraining movement became increasingly central to the transformational agenda, there was an accompanying narrowing of focus to syntactic phenomena perceived as relevant to that task. Correspondingly, the traditional goal of broad descriptive coverage receded in importance, and the formal architecture of the language faculty reflected a progressively narrower focus. By Chomsky (Reference Chomsky1981), transformational accounts dedicated an entire grammatical ‘component’ to locality (Bounding Theory), devoted a second component to anaphoric dependencies (Binding Theory), and expropriated the terms ‘Case’ and ‘government’ for theory-internal notions that bore little formal or functional relation to their traditional namesakes. The enhanced prominence of pronominal items, which had scarcely figured in the models of Chomsky (Reference Chomsky1957) and Chomsky (Reference Chomsky1965), was particularly striking, especially given that these items form small inventories, with fairly circumscribed – hence, presumably, learnable – patterns of distribution. But the importance of anaphoric dependencies derived ultimately from more theory-internal factors.

A critical step was the idea that ‘filler-gap’ dependencies could be assimilated to cases of binding such that the ‘filler’ functions as an antecedent, and the ‘gap’ as an anaphor or pronoun (Chomsky Reference Chomsky1976) By treating extraction as a type of anaphoric dependency between a dislocated ‘operator’ and a ‘variable’ in an extraction site, Chomskyan accounts arrived at a solution to the longstanding problem of ‘derived constituent structure’. As had been noted by Less (Reference Lees1957) and Stockwell (Reference Stockwell and Hill1962), early Chomskyan models defined transformations as mappings from phrase markers to phrase markers. Yet the rules formulated in these models specified only the order of constituents in the output, with no information about the structure of these constituents. The notion that extraction constructions contained a dislocated filler ‘binding’ a gap provided the first general solution to the problem of assigning derived structure to transforms, though this solution is by no means empirically unproblematic (Blevins Reference Blevins1994).

Working backwards through these developments, one can see the formulation of constraints on extraction, and the subsequent efforts to isolate those constraints, as decisive steps in the evolution of Chomskyan approaches. But instead of yielding new insights about language, the focus on movement led, via a succession of models with steadily decreasing empirical coverage, to a programme in which it seems appropriate to explore the question ‘How “perfect” is language’ (Chomsky Reference Chomsky1995: 221). The opposing view that C&J present is that extraction constructions are learned, like other constructions (Tomasello Reference Tomasello2003, Goldberg Reference Goldberg2006), and that the constraints on extraction in a given language reflect a process of conservative generalization from the patterns that are established in the language. From this perspective, the study of a priori notions of ‘perfection’ and deviations from this ideal have as little to do with the study of language as they do with biology. C&J do not deny that their alternative may be regarded as ‘uninteresting’ by syntacticians who have imprinted on the Chomskyan aesthetic. But they maintain that the issue is an empirical one and that the available evidence – developmental as well as syntactic – supports their position.

2.2 ‘Hidden’ levels and honest accounting

More generally, as the other case studies in SSyn show, the quest for universal principles linked by a rich deductive structure has led to steadily declining empirical coverage, as a small set of ‘core’ phenomena have assumed a progressively larger role, culminating in the aptly termed ‘Minimalist’ accounts. A key role in the discrimination of core from ‘peripheral’ phenomena is played by what C&J term ‘hidden’ levels of analysis. The postulation of a ‘pre-syntactic’ level of ‘deep structure’ (Chomsky Reference Chomsky1965) or ‘D-structure’ (Chomsky Reference Chomsky1981) may initially have reflected a genuine belief that ‘the structures for which universality is claimed may be quite distinct from the surface structures of sentences as they actually appear’ (Chomsky Reference Chomsky1965: 118). Yet the hope of using transformational analysis to discover an underlying unity remained almost completely unfulfilled. The earliest transformational accounts were acknowledged to be far too construction-specific: a descriptive apparatus designed to classify strings and tree structures (technically string sets) turned out to be ill-suited to the task of capturing common properties across constructions in different languages, as argued with particular cogency in the early Relational Grammar literature (Perlmutter Reference Perlmutter1983). By the late transformational period, the search for a deep commonality had effectively been called off, and, as SSyn points out in chapter 3, the universal properties of underlying structures came to be imposed by fiat, with only the most cursory empirical justification. The level of ‘logical form’ or ‘LF’ (Chomsky Reference Chomsky1976) was never suited to the task of discovering common patterns, given that its structure was designed to reflect specific assumptions about extraction constructions. Furthermore, as C&J show, the arguments advanced for LF were undermined not only by discrepancies between ‘overt’ and ‘covert’ movement but also by the fact that the use of ‘covert movement … to syntacticize quantifier scope … cannot account for the full range of phenomena’ (533).

As C&J acknowledge, their own model incorporates a hidden level of sorts, in the form of a ‘grammatical function (GF) tier’ that associates dependents with a ranked set of grammatical functions. The elements of the GF tier are more like the argument structure (arg-st) terms of HPSG accounts (Pollard & Sag Reference Pollard and Sag1994). than like the grammatical relations of RG accounts (Perlmutter Reference Perlmutter1983). or the grammatical functions of LFG accounts (Kaplan & Bresnan Reference Kaplan, Bresnan and Bresnan1982), in that they ‘are not explicitly labeled subject, object, and indirect object’ but ‘are just ranked positions’ (191). However, the distinctions expressed by the grammatical function tier are broadly similar to those expressed in the relational networks of RG, in the functional structures of LFG, or, indeed, in traditional grammars (Jespersen Reference Jespersen1937). Moreover, because the grammatical function tier classifies the functions of surface structures, it does not shift the locus of grammatical analysis away from observable structures onto more abstract, idealized representations, such as underlying D-structures or post-syntactic LF-structures.

The sustained argument in SSyn against idealized ‘hidden’ levels that diverge significantly from observable structures thus echoes the verdict reached by Hockett (Reference Hockett1987) in his discussion of ‘the great agglutinative fraud’. Although Hockett had been instrumental in developing ‘item and arrangement’ analysis, he came to regard it as an empty formal exercise in which an agglutinative structure was foisted onto a language by ‘devising an “agglutinative analog” of the language and formulating rules that would convert expressions in that language into the shapes in which they are actually uttered’(Hockett Reference Hockett1987: 83). In the continuation of this passage, Hockett notes the success of the agglutinative ideal in jumping hosts and establishing itself within the generative paradigm, where it fostered a more abstract and theory-internal style of analysis. Theoretical analyses ceased to address the facts of a given language directly, but instead constructed idealized analogues, which could then be brought into correspondence with the actual patterns in the language. Hence, the focus of linguistic analysis shifted from the study of patterns of cross-linguistic variation to the study of the properties of abstract analogues and the nature of the devices needed to mediate between these idealized forms and attested surface patterns.

The contemporary relevance of Hockett's remarks is confirmed by the many cases discussed in SSyn in which theoretical artifacts divert attention from issues of real substance. The treatment of constituent structure offers an instructive range of cases and issues. C&J note that the late transformational preference for binary branching structures rests on the extremely slender evidence offered in Kayne (Reference Kayne1983). They also stress that arguments based on the putative ‘restrictiveness’ of binary branching must be assessed in the context of a ‘cost-benefits analysis’ over the grammar as a whole. A ‘restriction’ in one part of the grammar plainly achieves little if it imposes limitations that must be overcome by greater complications elsewhere. As C&J observe, constraints on branching present a particularly direct trade-off, given that ‘simplifying the branching possibilities requires the tree to have more nodes, and reducing the number of nodes requires more branching possibilities’ (113). This point brings out a recurring theme of SSyn, namely that the evaluation and comparison of analyses should be guided by a principle of ‘honest accounting’ that counts global as well as local consequences of analytical choices. Although C&J are surely right about the virtues of honest accounting, the fact that they feel compelled to propose such a self-evident principle at this stage goes some way toward explaining why the generative paradigm has failed to live up to its early promise.

Significant as it was, the adoption of binary branching was only one factor in the development of increasingly abstract constituent analyses in ‘Principles and Parameters’ (P&P) models (Chomsky Reference Chomsky and Freiden1991). Binary branching structures pushed the syntactic analyses in P&P models away from the relatively flat, ‘bushy’ trees which were characteristic of earlier transformational models and are retained in traditional grammars and constraint-based frameworks. By reducing branching options, P&P models incorporated a larger number of nodes into increasingly ‘spindly’, bamboo-like trees. These structural developments were reinforced by shifting assumptions about the content of individual nodes.

Reflecting their origins in the distributional model of Harris (Reference Harris1951), the models of X-bar theory derived from Chomsky (Reference Chomsky, R. A. and P. S.1970) regulate only information about word class within a phrase. So the initial Chomskyan models of X-bar theory enforced a match between the category (N, V, A, or P) of an endocentric phrase and its head, but did not regulate any other grammatical features, such as tense or agreement. This restriction on the type of information governed by the X-bar conventions fed the use of constituent-structure displacements to distribute other types of feature information, notably via the advancement of heads to higher head positions or movements linking heads to ‘specifier’ slots. However, the unbundling of the former ‘INFL’ node proposed in Pollock (Reference Pollock1989) offered a different way around the restrictions imposed on X-bar models. By treating ‘functional categories’ as abstract heads, P&P models invoked the X-bar conventions to regulate the distribution of tense, aspect or agreement features. Reifying morphosyntactic features as heads is consistent with the distributional approach to syntax that underlies X-bar theory, as these heads rehabilitate the ‘abstract morphemes’ that played a significant role in the distributional model of Harris (Reference Harris1951). At a more technical level, recasting inflectional features as functional heads allows the X-bar conventions to regulate a larger class of features. But this coercion is only necessary because the class of X-bar features was arbitrarily restricted in the first place. As the X-bar models developed in GPSG (Gazdar et al. Reference Gazdar, Klein, Pullum and Sag1985) and HPSG (Pollard & Sag Reference Pollard and Sag1994) show, traditional notions of endocentricity can be captured by generalizing the X-bar conventions to regulate a broader class of ‘head’ features. Classifying person, number and gender as nominal head features permits generalized X-bar conventions to accommodate constructions in which agreement features of a noun are preserved on the noun phrase that it heads. Classifying tense, aspect and mood as verbal head features likewise allows the inflectional features of a verb to be shared with the verb phrase or clause that it heads.

A principle of honest accounting would dictate that any benefits obtained by restricting the X-bar conventions to word-class features should be balanced against the cost of reclassifying entire inventories of morphosyntactic properties as word-class features. Constraining the X-bar conventions so that they only regulate word-class features has no meaningful effect if there are no restrictions on which properties can be classified as word-class features. Similar cost-benefit calculations apply to other division of labour choices. There is plainly a trade-off between the density of feature information present at individual nodes and the number of nodes required to represent a fixed quantity of information. Any analysis must face the substantive challenge of identifying the features that are distinctive in a system and assigning appropriate features to particular items or constructions. But the way that these features are distributed over a representation will tend to reflect more general considerations. Constituent structures can be simplified by increasing the density of feature information associated with individual nodes. Or the complexity of individual feature structures can be reduced by distributing simpler feature ‘bundles’ over a larger number of nodes. Given that comparable grammatical information can be distributed over spindly or bushy trees, one might have expected some discussion in the P&P literature regarding the formal or empirical benefits of the spindly variety. Yet, as becomes clear from C&J's attempts to impose a coherent narrative on the developments within late transformational models, there is almost no discussion of this kind.

C&J bring out even more striking cases of accounting lapses in their summary of ‘VP-shell’ analyses (in which verbs are evacuated from their initial positions and advanced into higher ‘shells’ in a layered VP). From their origins in Larson (1988) as arboreal reimplementations of Montague grammar accounts (Bach Reference Bach1979), VP-shells became progressively more influential in transformational models, extending to patterns traditionally regarded as derivational. C&J review published criticisms of the initial proposals, including the key objection that the use of shells serves mainly to preserve the view that anaphoric dependencies are subject to hierarchical, not linear, constraints. They then consider the remarkable extension of VP-shells in Hale & Keyser (Reference Hale, Keyser, Hale and Keyser1993), in which a sentence with a denominal verb such as shelves is treated as containing an underlying noun shelf that becomes a verb in the course of a syntactic derivation. C&J run through some of the most glaring deficiencies of this account, a number of which appear to be freely conceded by its proponents. Specifically, the account does not attempt to explain which nouns will participate in the alternation, what meaning changes may be effected by the derivation or even whether there will be any phonological effects (as in the voicing of shelf to shelve). The even-handed tone that C&J maintain in their review only makes their criticisms more devastating: there is no escaping the fact that the Hale & Keyser account fails to address basic questions that it would be expected to answer. Just as extraordinary is the fact that the account does not set out to solve any outstanding problems whose resolution might that would balance the costs associated with the substantial problems that it itself creates.

SSyn attributes the acceptance of the Hale & Keyser account to its use of an established technology (VP-shells) and to the fact that it ‘purports to explain semantic parallelism without having actually to invoke a structured theory of semantics’(56). The first of these factors is of additional interest for the way that it highlights the pivotal role that ‘exemplary analyses’ have played in Chomskyan accounts. As C&J note in section 2.2, the treatment of active-passive alternations provided a template for early transformational accounts and subsequent exemplary analyses served a similar function in later versions. Transformational analyses of passive constructions exemplified the notion of ‘significant linguistic generalization’ in the relation between active and passive clauses, and clarified the role of underlying structures and derivational strategies. But even in early (and comparatively explicit) models, the use of a ‘canonical’ account to exemplify principles of analysis was not unproblematic, and contributed to what are sometimes known as the ‘linguistic wars’ (Harris Reference Harris1993). At least part of the rancour generated by what ought to have been an intellectual dispute derived from the fact that proponents of Generative Semantics regarded their proposals as legitimate extensions of exemplary transformational analyses. Starting from the assumption – widely accepted at the time – that underlying structures were the primary locus of semantic interpretation, the Generative Semantic programme explored the hypothesis that underlying structures were themselves semantic representations. By expanding the scope of transformations (which mapped semantic representations onto surface structures), this hypothesis seemed to extend earlier exemplary analyses, which had shifted the division of grammatical labour away from Immediate Constituent structures to transformational analyses. The ‘lexicalist’ response in Chomsky (Reference Chomsky, R. A. and P. S.1970) did not so much refute the Generative Semantics positions (Huck & Goldsmith Reference Huck and John1995) as establish a different exemplary model, based on analyses of derived and gerundive nominals.

Moreover, as transformational models became more programmatic, the task of inducing principles from ‘canonical’ individual analyses became substantially harder. By Chomsky (1993), the reader is faced with the daunting challenge of trying to ‘reverse engineer’ constitutive principles from exemplary analyses that consist largely of implementation details, key terms in scare quotes and little actual linguistic data. Although C&J offer a characteristically charitable assessment of the ‘complex and impressive’ (56) achievements of Chomskyan approaches, their account of the evolution of ‘mainstream’ models makes sobering reading. From an intellectual movement that once seemed destined to revolutionize the study of language in much the way that plate tectonics revolutionized geology, Chomskyan theories have become increasingly detached from the concerns of traditional descriptive and theoretical approaches as well as from those of allied disciplines. A central question raised by this volume is whether the flaw lies in the execution of the Chomskyan programme, or whether the basic conception is misguided.