The cycle in generative grammar: a brief history
The transformational cycle was first proposed around 1964 as a way to eliminate generalized transformations from the theory while accounting in a natural way for some empirical observations about the general operation of transformations.
Within the earliest formulations of transformational grammar, phrase structure rules were assumed to be nonrecursive. Recursion in phrase structure was handled instead by generalized transformations, which applied to two or more phrase markers, either embedding one inside another or conjoining them. In addition to these generalized transformations, there were singulary transformations, which operated on a single phrase marker (P-marker). As a result, a derivation was represented by a transformation marker (T-marker), indicating how the system of transformations applied.
Consider the example discussed in Chomsky’s 1964 lectures at the Linguistic Society of America Summer Institute (Chomsky 1966). Sentence (1) would be derived from three base P-markers underlying the (kernel) sentences (2), the derivation being represented by the T-marker (3), where B1, B2, and B3 represent each of the base P-markers.
The T-marker (3) shows that the three base P-markers in (2) are processed as follows. First, (2c) (= B3) undergoes the transformation that turns it into a relative clause. Next, that derived P-marker is embedded in the P-marker for (2b) (= B2), after which the passive transformation converts the expanded P-marker into a passive construction (the man who quit work was fired by someone) and a deletion transformation removes the phrase by someone. Finally, the resulting P-marker is embedded into the P-marker underlying (2a) (= B1) and the embedded sentential structure is converted into an infinitival.
Putting aside the antiquated nature of the particulars of this analysis, the incorporation of T-markers in the theory of grammar created several complications. One concerned the then-evolving semantic theory for generative grammar first proposed by Katz and Fodor (1963). Katz and Fodor had to propose two types of projection rule, whose purpose was to construct readings for constituents. Type I projection rules operated on underlying P markers to construct a reading. Type II projection rules were proposed to account for the effects of generalized transformations that applied to pairs of P-markers. Katz and Postal (1964) demonstrated that the contribution of generalized transformations could be limited to the amalgamation of readings of the P-markers joined together by the operation. Thus, the only function of type II rules was to assign the reading of the embedded structure (a base P-marker) to the P-marker into which it was embedded.
Another complication concerning generalized transformations involved restrictions on the organization of T-markers as discussed by Fillmore (1963). As interpreted by Chomsky (1966), Fillmore’s observations were essentially that (a) generalized transformations need not be ordered (in contrast to singulary transformations), (b) there are no cases where a matrix P-marker must be operated on by a singulary transformation before a constituent P-marker is embedded in it by a generalized transformation (though there are cases where a singulary transformation must apply to the matrix P-marker after the constituent P-marker has been embedded), and (c) the embedding operation should be viewed as a substitution operation that inserts a sentential P-marker in place of a “dummy symbol” A (equivalent to a categorially unspecified position). As Chomsky noted, the earlier theory of T-markers allowed for more complex orderings between singulary and generalized transformations. He concluded, “It is therefore quite natural to generalize from these empirical observations, and to propose as a general condition on T-markers that they must always meet Fillmore’s conditions,” as illustrated in (3) (1966, 62).
However, as Chomsky also noted, this condition as formulated “appears to be quite ad hoc” (p. 62).
Chomsky’s solution was to eliminate generalized transformations in favor of allowing recursion in phrase structure rules so that they could construct generalized P-markers directly. To account for the kind of ordering restrictions Fillmore observed, Chomsky introduced the transformational cycle as a general principle of rule application.1 The linear sequence of singulary transformations applies to the most deeply embedded sentential structure in the generalized P-marker. Then, as Chomsky outlines the process, “Having completed the application of the rules to each such structure, reapply the sequence to the ‘next-higher’ structure dominated by S in the generalized phrase-marker” (1966, 63). This continues until the root S has been so processed.
Chomsky gave the argument for the cycle as follows:
The advantages of this modification are obvious. It provides a more highly structured theory which is weaker in expressive power; in other words, it excludes in principle certain kinds of derivational pattern that were permitted by the earlier version of transformational theory, but never actually found. Since the primary goal of linguistic theory is to account for specific properties of particular languages in terms of hypotheses about language structure in general, any such strengthening of the constraints on a general theory is an important advance. Furthermore, there is good internal motivation for enriching the structure (and hence decreasing the expressive power) of transformational theory in this way, namely, in that this modification permits us to eliminate the notion of “generalized transformation” (and with it, the notion “T-marker”) from the theory of syntax. Hence the theory is conceptually simpler. Finally, the theory of the semantic component can be simplified in that type two projection rules are no longer necessary at all. (1966, 65)
As stated, the argument rests on a mixture of empirical and conceptual/ methodological issues. The reduction in expressive power is motivated by the empirical fact that certain derivational patterns are not found in natural languages. The elimination of generalized transformations and T-markers, as well as of type II projection rules in the semantic component, simplifies the theory conceptually The general methodological goal of accounting for the specific properties of individual languages via general principles of language also favors postulating the cycle.
At this stage in the development of the theory of transformational grammar, the cycle was used as an argument for eliminating generalized transformations. Yet the current theory of bare phrase structure has reintroduced generalized transformation into the theory of grammar, at the same time relying on a version of the cyclic principle. Since this does not lead to any obvious problem, it should be clear that there is no conceptual incompatibility between the cycle and generalized transformations. Actually, generalized transformations became superfluous with the introduction of recursion in phrase structure rules. Without phrase structure rules, generalized transformations are again necessary to embed one piece of phrase structure within another.
Following the enunciation of the cyclic principle, a new empirical argument for the cycle was introduced. The general form of the argument involved identifying a pair of transformations α and β such that the derivation of some sentence of a language required the applications α>β>α, where α had to apply before β in a particular sentential domain and then α had to apply after β in a higher sentential domain. Such arguments tried to establish that the application of β depended on the prior application of α and the second application of α depended on the prior application of β.2 Such arguments turned out to be quite fragile, primarily because they were based on particular formulations of transformational rules. Change the formulation of the rule and the argument collapsed. For example, Ross’s (1976) important analysis of the cyclic nature of pronominalization depended on the existence of a transformational operation (called pronominalization) that converts a referential expression into a coreferential pronoun. The argument vanishes under current analyses where pronouns are inserted in a derivation directly from the lexicon and later interpreted. Several other specific empirical arguments were based on the existence of a There-Insertion transformation and of a Raising transformation that included the extraposition of the infinitival predicate—rules that play no role in the current principles-and-parameters framework. However, even where these rules were assumed, this form of empirical argument for the cycle turned out to be defeasible. Kimball (1972) demonstrated that transformations could apply in a strict linear fashion: transformation T1 applies iteratively to increasingly larger domains and then another transformation T2 applies to the derived P-marker in the same fashion, so that the strict linear ordering T1>T2 holds globally. Kimball showed how such linear derivations generated the same structures that cyclic derivations produced. (See Freidin 1976 for further discussion of the empirical motivation for the cycle based on rule-ordering arguments.)
The formulation of the cycle in Chomsky 1965 and 1966 left open the possibility that a singulary transformation could still apply to a cyclic subdomain of the current cycle, in violation of the spirit though not the letter of the cyclic principle. Thus, the application of a rule to a matrix sentential domain would create a context for the application of a rule to a constituent sentential domain. Chomsky (1973) proposed a sharper formulation of the cyclic principle that eliminated this loophole. This reformulation, called the Strict Cycle Condition (henceforth SCC), is given in (4).
Chomsky (1973) mentioned one example relevant to the SCC, given in (5) (= his (57) and (58)), but did not go into the details of the analysis.
Given an underlying structure (5b), if where moves into the embedded COMP position, then movement of what into the matrix COMP is blocked by other principles—notably the Subjacency Condition, which prohibits movements that extract an element out of more than one bounding node at a time. The derivation of (5a) violating the SCC is given in (6).
In (6a–b), what moves successive-cyclically to each COMP position. Then, in (6c), the movement of where to the empty embedded COMP violates the SCC.3 This derivation is somewhat different from the kind that the original formulation of the cycle was designed to prohibit—that is, derivations where a rule could apply solely to the matrix domain prior to the embedding of the constituent domain in a way that would facilitate the application of a rule applying solely to the constituent domain. In this instance, the rule applying to the matrix domain involves a term in the constituent domain. Nonetheless, examples like (5a) constitute the empirical content of the SCC.
In Freidin 1978, I identified other cases that are ruled out by the SCC, including cases of NP-movement, and showed that, given trace theory, the empirical effects of the SCC can be subsumed under other independently motivated principles—specifically, parts of the θ-Criterion, what eventually became the Case Filter, and the Subjacency Condition construed, crucially, as a condition on representations. Given that the SCC appears to be totally redundant with respect to other necessary parts of the theory, it was argued that the SCC is superfluous and can be dropped as an axiom of Universal Grammar (UG). The elimination of the SCC suggested that the facts of derivation may be epiphenomenal and that what was actually important were the representations generated by the rules of grammar. Thus, in the case of (5a) it didn’t matter whether the derivation of the sentence violated just the SCC or Subjacency (or both together on one derivation) because the output gave the same representation, which violated Subjacency construed as a condition on representations. In short, it didn’t matter whether derivations conformed to the SCC or not because any deviant sentence whose derivation violated it would be excluded by independently motivated conditions.
