1. THE CHARACTER OF THE STANDARD THEORY

The central innovation of the theory of Chomsky and Halle (1968) over pre-generative discussions is undoubtedly the extent to which it attempted to make explicit the principles relating phonological to phonetic representation. The existence and association of two levels of sound structure had of course been generally accepted at least since the works of de Saussure and Baudouin de Courtenay (with some idiosyncratic exceptions, such as Bloomfield’s rejection of the significance of phonetic representations). Phonological (or ‘phonemic’) structure was generally accepted as a representation of the sound properties of an utterance that distinguish it from others and establish its position in a network of interrelations within a particular language as opposed to phonetic structure, which identifies an utterance in the range of possible language-independent human vocalizations. While the degree of abstractness of the relation between these two was subject to considerable variation across theories, Chomsky and Halle’s position on this issue was certainly within the range of those accepted by previous writers.

The unique contribution of the standard theory of generative phonology, then, was not in the character of the phonological representations it proposed (though these did indeed represent a break with the ideas of their immediate predecessors in the structuralist tradition), or even in the extent to which the character of these representations was made explicit and a goal of scientific inquiry (which was also a characteristic of the structuralist tradition, especially in America), but rather in the theory’s concentration on making explicit the principles governing the association of phonological with phonetic representation, conceived as a system of rules or algorithm for converting one into the other in a series of steps. The theory required the individual rules to be specified with a formal precision quite unprecedented in discussions of phonological structure (which had concentrated on the character of the representations themselves). Further, since the rules of such a description are in principle interdependent (as distinct from those in, e.g., American structuralist or ‘autonomous phonemic’ descriptions), their potential interactions have to be specified with comparable precision.

In this conception, then, a phonological theory is first and foremost an explicit, notational system of description specifying a) the formal character of phonological and phonetic representations; b) the formal character of the individual components of the operation mapping one representation onto the other (the ‘rules’); and c) an effective procedure for applying the set of rules so as to effect the required mapping, involving both a definition of what it means to ‘apply’ a rule to a representation and a specification of the ways in which the rules can interact.

The standard theory of generative phonology also aspired to another goal, however: in addition to providing an explicit descriptive framework for the formulation of phonological descriptions, the theory also wished to address the problem of providing an account of how one out of a range of possible descriptions is to be chosen and justified as ‘correct’ for a given range of data. The notational system, that is to say, ought to facilitate the direct, formal comparison of alternative accounts of the same set of facts.

It is important to note that this latter project, the achieving of ‘explanatory adequacy,’ is logically quite distinct from the project of description; but in the proposals of Chomsky and Halle (1968) they often seem to have become inextricably merged. Since in the standard theory a phonological theory is a notational system, over expressions in which an evaluation metric can be defined, questions of notation come quickly and apparently inevitably to involve descriptive and explanatory concerns.

Much theoretical discussion since the publication of SPE, including a number of the positions represented in this volume, has been based on the explanatory concerns of that theory, under the assumption that the purely descriptive problems have already been (or at least could trivially be) solved. The strategy of workers in the field has been to propose radical limitations on the class of possible phonological rules or on the internal complexity of grammars which will have the effect of insuring that at worst only a very few accounts will be available for any given set of data. In the course of this development of the concerns of phonological theory, many interesting proposals have been made and much has been learned about the limits of linguistic structure. Considerations of the phonetic content and motivation of rules especially make it clear that a formal notational system in which all possible expressions are directly comparable is at the very least far in the future.

Indeed, the famous last chapter of Chomsky and Halle (1968) recognizes and addresses this problem, to which we will return below in section 5. There is little doubt that phonetic substance plays an important role in determining and justifying the choice of grammars, because it is basic to the nature of language. This is, in an important sense, however, completely foreign to the essential thrust of the standard theory, which aims to construct a formal descriptive system in which any expression is at least potentially a rule of a natural language.

Given the observed tendency of languages to incorporate at least some ‘crazy rules’ (rules whose synchronic phonetic motivation is apparently lacking), and of restructuring and other forms of language change to produce such rules, it would probably be Utopian to expect phonetic substance to serve directly as the basis for a system of descriptive phonology. In constructing a descriptive system, then, the formal approach of the standard theory undoubtedly still makes sense, so long as we realize that issues of explanation still remain to be considered.

At any rate, while the descriptive issues have been less emphasized in recent discussion than the more ‘glamorous’ ones of explanation, there are still problems that arise within and proceed from these formal and mundane aspects of the standard theory. The character of these problems suggests that this is an appropriate place to begin in the search for a Revised Standard Theory.

2. NOTATIONAL CONVENTIONS

The set of abbreviatory conventions proposed in SPE form one of the central problem areas for investigation within the standard theory of phonology. A consideration of the rather interesting history of these conventions and their role in the theory will serve to indicate both some of the character of the ‘notationalist’ program and some of the directions of subsequent work.

In early work in generative phonology (as summarized, e.g., in Chomsky and Halle 1965), abbreviatory conventions such as curly brackets, parentheses, Greek-letter variables, etc. were justified as aspects of the evaluation metric for descriptions, rather than as a part of the descriptive system itself. The fully expanded set of primitive rules was taken to be the grammar, and the notational conventions were assigned the task of assessing certain aspects of the internal coherence of the system of these primitive rules. Each notational convention, that is, was justified by the claim that certain (purely formal, in terms of the rest of the notational system) resemblances between rules are highly valued, and contribute to ameliorating the complexity of a set of primitive rules within which they obtain.

The notational conventions soon came to be employed for another purpose, however. Chomsky (1967) suggested that, in addition to their evaluatory function, notational conventions such as those of Greek-letter variables and parentheses also defined subclasses of rules within which systematic deviations from otherwise valid organizing principles of grammars could be localized. By associating simultaneous and disjunctive (as opposed to sequential, conjunctive) ordering with rules displaying these formal resemblances, such a move considerably enriches the content of the claim of significance for these abbreviatory conventions. In the process, however, purely descriptive issues (what grammars are possible?) came to be merged with issues of explanation (what grammars are preferred?) in a way that has been taken much further in later work, much of which is represented in the other contributions to this conference.

Along with this shift in the content of the notational conventions of phonological theory there took place a corresponding shift in the sort of justification offered for a given proposed notational device. For example, the first discussion of Greek-letter variable rules (in Halle 1962) proposed this device as a “descriptive device for the treatment of assimilation”, in order to remedy the obvious loss of generality in a formulation which required, e.g., nasal assimilation to labials, dentals, and velars to be performed by separate rules. Arguing that sets of rules of this type are in fact descriptively common, but would be counterintuitively clumsy to express in the primitive notation, Halle suggested that the device of feature counting (the fundamental mechanism by which formal expressions were to be assessed for complexity) be supplemented by a special provision for those rule-sets meeting the formal conditions of the proposed notation. Essentially the same justification as Halle’s is offered by Bach (1968) for the introduction of a notational device for ‘neighborhood’ or ‘mirror-image’ rules. The basic support offered for a proposed notation as part of the evaluation function was to demonstrate that a given formal resemblance between rules recurs in the grammars of a number of diverse languages, together with an appeal to the intuition (of other linguists) that the resemblance in question is not an accidental one.

The discussion of Chomsky (1967), however, introduced a new basis for justifying notational proposals: besides showing that some formal resemblance is a recurrent one in the languages of the world, one could also demonstrate that some aspect of the internal organization of grammars is associated with all (and only?) sets of rules displaying the resemblance in question. Chomsky thus argues that Greek-letter variables, besides defining a relationship of significance for the evaluation of descriptions, also define the class of ‘exchange rules’ for which simultaneous rather than sequence application is required. Similarly, the discussion in Anderson (1974) argues that (at least in the domain of phonological, as opposed to phonetic rules) the mirror-image notation defines a class of rule-sets which must be applied disjunctively rather than conjunctively. Such a descriptive justification for a convention is quite different in kind, and subject to rather different sorts of confrontation with the facts of natural languages, than the fundamentally evaluative justifications offered in most early work.

Another sort of justification for notational proposals also has a fundamentally descriptive basis, and was first suggested by Kiparsky (1968b). Kiparsky argued that rules abbreviable into a single schema could be shown to function as a unit in historical change, as when the entire schema is simplified, re-ordered, lost, etc. If rules function as a unit in this way, of course, it is not unreasonable to conclude that there is an important sense in which they are a unit in a synchronic description of the language in question, and hence that a theory which provided no way of expressing this unity would fail to achieve descriptive adequacy independent of explanatory concerns.

An example of the interaction of these concerns in relation to a specific notational device is furnished by a consideration of the ‘curly brackets’ notation. This notation, which allows shared material in two adjacent rules to be ‘factored out’ and counted only once as contributing to the complexity of the total grammar, was originally justified by its contribution to the evaluation function: it allowed the coherence of a set of partially identical rules to be taken into account. It was considered self-evident in early works on generative phonology that sets of rules demonstrating such partial overlap were more natural or expected, and hence to be evaluated more highly, than arbitrary sets of primitive rules of the same basic complexity.

An attack was made on the significance of the curly brackets notation by McCawley (1970), who observed that the purely formal sense of shared material embodied in this device allows for a wide range of surely spurious simplifications. McCawley argued that the use of the notation generally allowed the concealment of flaws in an analysis, and that in every case where it appeared to be applicable, either the analysis required further refinement (so as to result in a formulation in which the supposedly collapsible rules were in fact treated as the same rule) or there was in fact no significant generalization to be captured.

It is not, in fact, possible to defend the use of curly brackets by an appeal to the internal organization of sets of rules to which they can be applied. This is because, unlike the cases of parentheses and Greek-letter variables, which are associated with exceptional disjunctive application of rules, curly brackets are associated with conjunctive application: precisely the mode of interaction of rules having nothing in common at all. Thus, one might take this fact as a sort of argument ex silentio for McCawley’s position.

In some cases, however, evidence from historical change is available to confirm the significance of the collapsing performed by curly brackets. Consider the discussion of consonant gradation in Finnish in Anderson (1974). There it is shown that consonant gradation in fact involves two distinct sub-rules, one applying to geminate stops and one applying to single stops preceded by a sonorant (consonant or vowel). The separateness of these two rules is confirmed by, among other things, the distribution of exceptions in the language; and hence McCawley’s position (given the evidence against a refinement of the anlysis which reduces the two to a single case) implies that there must be no generalization involved in the rule pair. Aside from the clearly counter-intuitive nature of this conclusion (linguists’ intuitions being, as shown in Anderson 1977, more reliable as descriptions of linguists than of language), there is clear historical evidence against it. According to Collinder (1969), the gradation rule originally applied only in the position between a stressed and an unstressed syllable. In Modern Finnish, however, there is no such restriction; and what is significant for our purposes, the limitation has been lost equally from both cases of gradation. Similarly, in related languages such as Estonian and, more distantly, Lappish (as well as Finnish itself, to a small extent) subsequent phonological change has rendered the gradation environment opaque and the rule has been partially or completely morphologized. The crucial point, again, is that the replacement of phonological by morphological conditioning has proceeded identically for the two subcases of gradation. In considering the historical and comparative data concerning gradation, it is difficult to avoid the conclusion that changes have taken place in a single collapsed schema (which still abbreviates two distinct rules, however) rather than separately in two unrelated rules.

Such an argument would seem to establish the significance of a device for collapsing related but non-identical rules in a grammar, even in the absence of a special type of interaction among the rules so collapsed. On the other hand, a purely formal approach to the question of which rules potentially fall under this collapsing is bound to lead to problems. Consider a language containing a limited intervocalic voicing process, affecting only the dental stop /t/; and also a process of assibilation of /t/ before high front vowels:

By the definition of the curly brackets notation, the two rules in (1) would have to be collapsed to the single schema in (2):

Now imagine that a change takes place, by which intervocalic voicing is extended to the other stops (/p,k/). If the collapsing of (1) into (2) truly represents a synchronic unity between the two rules, we would expect this extension of intervocalic voicing to entail a similar extension of spirantization, causing /p,k/ to become /f,x/ before high front vowels. In fact, we can probably agree that such an extension of spirantization as a consequence of the extension of intervocalic voicing is exceedingly unlikely, and (subject to the reservation noted above about the validity of linguists’s intuitions about what languages could possibly be like) would probably wish to deny that the collapsing in (2) is valid or significant.

We must then ask ourselves what it is that makes collapsing valid for the rules of Finnish gradation, but invalid for the hypothetical case just considered. We might conclude that collapsibility, rather than being a mechanical effect captured by formal similarities in an explicit notation, is rather an idiosyncratic and variable property of particular grammars: the Finnish rules collapse, the others don’t, tout court.

It is probably possible to improve on that conclusion, however, but only by going beyond the notational program of the standard theory. In that theory the form of the rules themselves, in an appropriate expression, is all that is relevant to determining the applicability of the collapsing operations: this is precisely what is meant by the name ‘notational conventions’. If we consider the function of the rules, however (in what are at present the intuitive, pre-systematic terms of traditional phonetics), we can distinguish between the two cases. The two rules involved in Finnish gradation are both weakening processes (degemination of double stops, voicing or sonorization of single stops) while the intervocalic voicing and assibilation of (1) belong to quite unrelated categories (in the sense of having unrelated motivations or phonetic explanations) in traditional terms. We might suggest, then, that collapsing be limited to adjacent rules which are related in terms of their substance, function, phonetic motivation, etc. Of course, in order to make such a notion other than arbitrary, we must examine the traditional categories of phonetic explanation and attempt to make them (or whatever should replace them) as precise as possible. As we will note again below in section 5, however, it may well be only through such a return to the traditional concerns of the study of phonetic explanation that the goals of explanation espoused by phonologists can be approached.

The conclusion that the formal approach to phonological description which is at the heart of the standard theory must be, at the least, supplemented by an appeal to functional considerations is further confirmed by the discussion of the principle of disjunctive ordering which has taken place since the publication of SPE. A number of cases have been discussed in the literature which indicate the need for disjunctive application of rules under circumstances in which no principled notational resemblance such as that captured by the parentheses notation, etc., obtains. A principle which was probably first employed by Pānini has been argued for by several writers (Anderson 1969, 1974; Kiparsky 1972; Koutsoudas, Sanders, and Noll 1974). Roughly, this could be stated as, “When two rules conflict, apply the more specific to the exclusion of the less specific.” That is to say, when two rules would yield different results, if one of the rules applies to a proper subset of the forms to which the other applies, the more specific rule takes precedence over the more general, ‘elsewhere’ rule.

Not only does there seem to be a need for such a principle to govern the internal organization of grammars, but it also appears that this principle subsumes the cases in which the parentheses notation can be unambiguously shown to be associated with disjunctive ordering. If true, this conclusion would not of course establish the non-significance of the parentheses notation: this notation could still be shown to be of importance by an argument similar to that given above for the curly brackets notation. The argument for the significance of this notation which originates with Chomsky (1967), however, would have been shown to be incorrect. Disjunctive ordering appears to be associated with an independent principle, rather than with schematization by any particular notation.

The interest for our concerns at this point, however, lies in the fact that, in order to state the principle on which disjunctive ordering appears to depend, it is necessary to give some substance to the clause “when two rules conflict.” Practically any pair of rules in a grammar can be seen to do different things, and it is not of course simply this that is meant. Rather, we are concerned with cases in which the rules involved are somehow comparable. It would be strange to speak of, say, a rule of penultimate stress assignment and a rule of velar softening as ‘conflicting,’ although they do of course yield different modifications of the input string. Since a general stress assignment rule can apply to any form in the language, while velar softening only applies to forms meeting a specific condition, the most mindless form of the principle we have discussed would seem to predict that stress is not assigned to forms that undergo velar softening in such a language! Obviously, the sort of case which interests us is one in which, say, a rule of antepenultimate stress assignment applies to some specific class of forms and precludes the application of the general rule of penultimate stress applicable to the rest of the language. Stress assignment in distinct positions obviously constitutes a significant sort of ‘conflict’ in a way the hypothetical case of stress and velar softening does not. Similarly, in an example discussed in Anderson (1969, 1974), lengthening (with concomitant lowering) of vowels in open syllables conflicts with shortening of vowels in a certain subset of open syllables in Middle English.

The formulation of a principle of disjunctive ordering, then, requires that we give precise sense to the notion of ‘conflicting rules,’ and this in turn appears to require an appeal to the same substantive, functional considerations that we argued above were important to a proper understanding of schematization by the curly brackets convention (or some suitable alternative). From this we can conclude that the purely formal program of notational conventions of Chomsky and Halle (1968) leads naturally to such issues of phonological substance and function, which take us beyond the standard theory and constitute an important revision of it.

Another area in which the problems originally posed in the standard theory of notational conventions lead to an extension of that theory is the recognition of a distinction between morphological rules and phonological rules. In the program of SPE, essentially all operations on sound structures are performed by phonological rules, which in turn are freely allowed to contain morphological conditions without entailing a significant difference between rules that do and rules that do not involve such non-phonological information.

Subsequent work, however, suggests that the distinction between morphological and other rules is of significance, and must be made in order to carry out the program of defining notational conventions. In connection with the analysis of the synchronic reflex of the Great Vowel Shift in Modern English, Chomsky and Halle had to deal with the question of whether ‘exchange rules,’ performing an interchange of two phonological elements in some uniform environment, are possible operations in a grammar. They conclude that such rules do exist, but that they can be precisely localized as rules abbreviable by the device of Greek-letter variables. Accordingly, Chomsky (1967) proposed that part of the definition of this notation ought to be a principle that rules so abbreviable apply simultaneously rather than sequentially.

More extensive examination of real (and purported) cases of exchange rules by Anderson and Browne (1973) confirms the existence of such examples, but suggests an important limitation on them: apparently, all genuine exchange rules contain significant morphological conditions. This is in fact rather intuitive: rules which are purely phonological in character are generally fairly close to their phonetic motivation, and it is hard to imagine that some environmental conditions which motivate the replacement of A by B could simultaneously motivate the replacement of B by A. Rules involving morphological environments, however, are to just that extent removed from a phonetic motivation. Accordingly, there is no reason to expect that an exchange arising under such circumstances should be especially unlikely.

The principle of simultaneous application for exchange rules, then, is apparently limited to morphologically conditioned rules, and accordingly suggests the significance of a typological distinction between such rules (which class must, of course be properly defined in a motivated way) and purely phonological rules. A similar conclusion is suggested by a consideration of rules involving variables (or some other device, such as the X₀ notation of Chomsky and Halle 1968) which can range over arbitrarily long strings. In Anderson (1974) it is argued (on the basis of a stress assignment rule from Tahitian) that where several distinct but overlapping interpretations of such a variable are possible in conformance with the rest of the rule’s structural description, only the longest of these should be taken. This principle is quite close to the spirit of the other convention(s) for disjunctive ordering, and seems to be supported for a variety of cases (cf. also Howard 1972).

This conclusion regarding variable terms in phonological rules, however, does not seem to extend to morphologically conditioned rules. In Acoma, for example (cf. Miller 1965), there is a rule of accent ablaut, which applies when certain suffixes are added to a word. There is apparently no phonological characterization of the relevant class of affixes, and the rule must thus be considered to be conditioned by an arbitrary morphological property (at least from a synchronic standpoint). The effect of the rule is to replace the tonal accents of all vowels in the affected word by a ‘high’ accent:

The formulation of this rule is no doubt imprecise, but the important point to note is that, if the above interpretation of variables were correct, it would apply only to the leftmost vowel of the word, rather than to all vowels as desired. If we assume that the interpretation given above for phonological variables does not apply in morphologically conditioned rules, however, the correct result can be obtained.

Another example supporting the same conclusion comes from the languages of the Abkhaz-Abaza group of Northwest Caucasian languages. In these languages, we find a dissimilation rule of roughly the following type: when a verb prefix with the shape /r/ (marking 3rd person plural) precedes a stem which begins with /r/, or one which is preceded by the prefix /r/ marking causative formations, it is replaced by /d/.

Again, the formulation of this rule could doubtless be refined (we assume that the causative prefix forms part of a ‘stem’ when added to a ‘root’, and that some boundary delimits this stem from the other prefix material). What is important however is the fact that any amount of other material may intervene between the prefix affected and the beginning of the stem (the Abkhaz-Abaza verb allows a large number of positions for prefixes marking various functions), and also the fact that the rule is obviously morphologically (rather than purely phonologically) conditioned.

We can now ask what happens when more than one /r/ prefix appears in the same verb. Given the interpretation of phonological variables which seems valid for phonological rules per se, it would seem that only the leftmost prefix should dissimilate. From the underlying form /y+r+r+r+ba+d/ “it-them-they-cause-see-aorist, they showed them it” we get (ignoring predictable inserted vowels) [yddrbad], showing that in fact both prefixes are affected. Note, incidentally, that the presence of two /r/ prefixes alone is not sufficient to provoke dissimilation: from /y-r-a-r-h°-d/ “it-them-to-they-tell-aorist, they told it to them” we get simply [yrarh°d], with no dissimilation of one /r/ by the other. From this evidence we conclude that this rule, too, must not involve the interpretation of variables proper to phonological rules in the pure sense.

In addition to the properties surveyed above, other formal distinctions seem to mark off the morphologically conditioned rules of a grammar from the purely phonological rules (for some discussion of this issue, cf. Anderson, 1975). The recognition of a typological distinction between morphological and phonological rules, while it is not a part of the standard theory, clearly follows directly from the range of problems which that theory poses as central; and indeed the greatly increased attention paid to the properties of morphology in recent discussion is a major area in which the standard theory has been revised since the publication of SPE.

3. THE INTERNAL ORGANIZATION OF GRAMMARS

To the questions of how rules interact with one another and of how they apply to a form, the standard theory proposes a fairly straightforward set of answers. A rule is ‘applied’ by scanning the form for all places where its structural description is met, and performing all of the indicated changes simultaneously. The rule does not then re-apply to the result of this process. On the other hand, it can easily be seen that in some cases rules do have to apply to the results of applying other rules; and for this reason the main mode of rule interaction in the standard theory is application in conjunctive sequence. The rules are assumed to be organized in a (single) linear list, and applied by performing whatever changes result from the first rule on the list, then applying the second rule to the result of this process, and so on until the end of the list is reached.

One complication in this clear picture is introduced by the device of the transformational cycle in phonology proposed in SPE. Some of the rules are assumed to be applied cyclically, in that this sub-list of rules is applied in sequence to the innermost layer of constituent structure, then re-applied to the next layer of structure, etc. (For details, cf. Chomsky and Halle 1968.)

All of these aspects of the internal organization of grammars have been seriously questioned in recent years, and few of the standard theory’s positions on these issues enjoy general acceptance currently. The device of the transformational cycle has seemed particularly dubious, especially since the cases in which it was apparently motivated seemed to be confined to rules involving the assignment of stress (cf. Brame 1974 for some discussion). No clear cases of segmental rules which must be applied cyclically have apparently been found, though some candidates have been proposed (particularly by Kisseberth, 1972). It is particularly plausible that stress rules should have this character: the cycle is after all a way of making a rule’s application dependent on internal constituent structure, and the demarcative function of stress in many languages makes its association with constituent structure quite intuitive. It is not clear that the cycle is the correct way to capture this relationship, however, since it does not seem to extend to other sorts of rules in the way we might expect of a fundamental property of the organization of grammars.

In fact, the problems surrounding the cycle have resulted in proposals by Liberman and Prince (1977) for a different representation of stress than has generally been assumed. Liberman and Prince’s approach provides a particularly elegant way of integrating the dependence of stress on syntactic structure with other properties of stress in a way that obviates the use of cyclic rules of stress assignment. For this reason, it appears at present that the device of the cycle can be completely dispensed with and we will therefore ignore it in the remainder of this discussion.

Turning to the question of whether the rules of a grammar should be represented as organized into a linear list, we can distinguish two aspects of the problem which have received a great deal of discussion in the literature: a) the question of whether the interactions of rules can in general be represented as a linear ordering; and b) the question of whether the ordering (whether linear or not) is a language-particular property of grammars (“extrinsic ordering”) or can rather be predicted by a set of general principles of linguistic theory (“intrinsic ordering,” though this expression also has a narrower and more precise sense for which it should perhaps be reserved). These two questions are, to varying degrees, subject to empirical examination, in the sense of presenting analyses of particular languages whose individual rules in isolation can be determined and then observing how they must be assumed to interact in order to yield the correct set of forms. We can also ask whether the ordering relations that are observed are consistent with the hypothesis of overall linear ordering, and whether it is possible to relate all (significant) types of rule interaction to a single consistent set of predictive principles.

It will perhaps come as no surprise to some readers of these remarks that my own opinion on the issue of linearity is quite clear. It seems undeniable that there exist cases in the phonologies of natural languages in which one or more of the defining properties of a linear order is violated (antisymmetry, irreflexivity, or transitivity). A range of such cases was discussed in Anderson (1974), and others have appeared (sometimes in the literature) since then. It is of course possible to argue about the details of specific examples, and there is no doubt that some of the cases discussed in this connection require refinements or revisions of the original analysis, sometimes so as to remove the particular example from the class of those supporting non-linear orderings. But other examples remain, and indeed there is reason to believe, based on our knowledge of the processes involved in language history, that synchronic grammars ought to contain violations of linear ordering (cf. Anderson, 1977).

The example of u-umlaut in Icelandic, and its interaction with other rules of the grammar, forms a central part of the argument in Anderson (1974) and will suffice for our discussion here. Essentially, the violation of linearity involved (which has been left unimpugned by subsequent discussion, as far as I can tell) arises from the fact that in order to derive [jökli] from underlying /jak+ül+e/ it is necessary to apply umlaut before syncope; but in order to derive [kötlüm] from underlying /katil+üm/ it is necessary to apply syncope before umlaut. An account of these orderings within a synchronic grammar, based on a synchronic analog of the principles of feeding and bleeding orders first invoked by Kiparsky in discussing historical change (cf. Kiparsky 1968b) was suggested, but regardless of the correctness of that account (in which I still believe, although logically that account forms a quite separate issue, related to predictability of ordering rather than to non-linearity) the violation of linear ordering seems firmly established. Some authors have suggested that this conclusion could be avoided by establishing underlying /jök+ül/ as the stem for ‘glacier’ rather than /jak+ül/, and thus abandoning the derivation of [ö] from /a/ by umlaut in these forms; but this possibility was raised and (to my mind) conclusively eliminated on the basis of evidence from the highly structured metrical system of Old Icelandic skaldic verse (cf. Anderson 1972). The violation of linearity, which has apparently been part of the grammar of Icelandic for more than a millenium, seems well established; and in light of other examples from a variety of languages, not isolated. We can only conclude, it seems, that the claim of the standard theory that ordering relationships are linear must be replaced with a richer set of possibilities.

Some linguists have seen the possibility of non-linear ordering relationships as completely absurd and ‘counter-intuitive.’ The intuition involved here is again that of linguists about formal systems, and there are good reasons to expect these intuitions (as opposed to the sorts of cognitive processes that underlie language) to favor linear orderings. A sober consideration of the facts of natural languages, however, seems to require the conclusion that these intuitions are mistaken (cf. Anderson 1977).

With respect to the question of whether all orderings can be predicted from a single coherent set of universal principles, the question is somewhat less clear. In principle, it is probably not possible to disprove the claims of those who assert the existence of such a set of principles. After all, someone may come forward with a correct proposal tomorrow. Such as we know about ordering relationships, however, does not appear to give much hope to those who await this. One of the best established predictive principles, for example, is that of the preference of languages for ‘feeding’ orders. Despite the overwhelming predominance of feeding orders over counter-feeding orders, however, it is difficult to deny that at least a small number of cases of the counter-feeding sort exist. To return to Icelandic u-umlaut, there is another rule in the language which inserts certain epenthetic u vowels; this rule could feed u-umlaut, but fails to do so. There is no other obvious principle to explain this fact; it is presumably due to the fact that the epenthesis rule appears comparatively late in the history of Icelandic, well after the umlaut rule was well established, and its appearance at ‘the end of the grammar’ after (and in counter-feeding relationship to) umlaut thus has a historical but not a synchronic explanation. Some discussion of this issue will be found in the exchange between Iverson (1976) and Anderson (1976a, 1976b); the result of that interchange would appear to be agreement that in at least some cases (despite some disagreement on just which stages of what language are the ones concerned) epenthesis has failed to feed umlaut when such a relationship was possible; and this would appear to go some ways toward establishing that some cases of “extrinsic” ordering are unavoidable.

Another generally valid principle of rule interaction suggests that morphological rules precede rather than follow phonologically conditioned ones, at least in the general case. In Anderson (1975) evidence was presented, however, that natural languages sometimes contain violations of this principle as well, though many fewer examples of this sort can be found than there are cases which support its general validity. Aronoff (1976) attempted to refine the hypothesis by suggesting that a given morphological rule either precedes all of the phonological rules or follows all of them; but examples discussed by Chung and Anderson (to appear) show that indeed morphological rules and phonological rules can be fully intermixed in principle (though of course the situation in most languages is much simpler than that). Further examples which appear to demonstrate the difficulties of the position that all ordering relationships are predictable are given by Vago (1977a).

What are we to make of the apparent fact that we know of a number of principles (preference for feeding orders, minimizing opacity, the precedence of morphological over phonological rules, etc.) which have a high degree of generality, but which fail in specific instances? In Anderson (1974) a hierarchical arrangement of such principles was suggested, but as the number and complexity of principles with some apparent explanatory value increases, this seems less and less likely to be worked out in practice. Yet we surely do not wish to discard such principles entirely on the basis of isolated counterexamples, any more than we wish to discard from the grammar of a given language every rule which has even a single exception.

It seems to me that the situation is in some ways parallel to that in which the neogrammarian movement found itself at the beginning of this century. The close and detailed studies of sound change which had been carried out in those terms had suggested some broadly explanatory principles of language change (especially the pair ‘Lautgesetz’ and ‘Analogie’), and it is undeniable that the understanding of linguistic change was immensely increased thereby. Nonetheless, the conflict among these principles was such that it was quite impossible to achieve a predictive sort of explanation, in the sense of providing principles that would determine whether one or the other would prevail in a given change.

If we distinguish the ex post facto understanding of a linguistic fact which a given principle can provide from the ability of that principle to predict beforehand what the facts will be, we can I think find a proper value to set on principles such as those that have been suggested for rule interaction. When we find a particular sort of interaction in a given language, we can suggest that it is motivated by, or implements, or whatever, a tendency to reduce opacity, e.g., or to maximize feeding orders, etc. When we say that, we have, it seems to me, made some progress in understanding the facts as they are, though not in the sense of showing that they could not be otherwise. We might perhaps describe a theory which provided a set of principles meeting this requirement as achieving a level of ‘Exegetic adequacy’ (as opposed to explanatory adequacy in the predictive sense). The work on ordering relationships and their bases which has proceeded out of the standard theory in various directions gives good grounds for hoping that such a level of exegetic adequacy is attainable in this domain, but does not encourage the further hope for a predictive account, which would eliminate the need for ‘extrinsic orderings’ in the grammars of natural languages.

The other issue raised by the standard theory in the domain of the internal organization of grammars is that of how to apply a rule to a form. In the general case, of course, this is trivially resolved; the difficulties arise in cases in which the same rule is potentially applicable at a number of places in the same form, in overlapping environments. The literature dealing with this question has become quite extensive in recent years (including proposals by Johnson 1970; Howard 1972; Anderson 1974; Vago 1977b; and an excellent summary of issues in Kenstowicz and Kisseberth 1977). By now it has come to be generally accepted that the simple algorithm of simultaneous application proposed (for want of any reason at that time to think otherwise, it should be emphasized) by Chomsky and Halle (1968) is inadequate to deal with all of the cases that have been discussed. There are primarily two contenders for an effective algorithm to incorporate all of the examples to date: a revised form of the simultaneous theory, as proposed by Anderson (1974), or a ‘directional’ theory as proposed originally by Johnson (1970) and developed further by others.

The revised simultaneous theory of Anderson (1974) is based on an algorithm which is rather difficult to summarize, given its complexity. This is, of course, no objection to it in principle, since it is to be assigned the status of a metatheoretic construct, rather than as a rule to be stated in the grammars of particular languages. Its principal virtue is probably the fact that it allows the same principles which appear to govern the interaction of different rules to be invoked to give an account of the interaction of multiple applications of the same rule. There is also at least one example, an accentual rule from Acoma, for which it appears to provide the only viable account; though the isolation and complexity of this example make it rather insecure as evidence in the absence of further support.

This algorithm is capable of generating a variety of patterns of surface interactions of multiple applications of a single rule; there are, however, some patterns which it cannot describe, and which it must thus claim are non-occurring in natural languages. One of these is the case which is in fact simplest for the standard theory algorithm: the case in which every potential application of the rule is realized. For reasons which are too complex to go into here, an example in which three or more overlapping potential applications of a rule all undergo it would serve to falsify this theory (under some modestly restrictive assumptions about the type of rule involved).

It is not in fact at all easy to find examples which falsify this theory in this way. In part, this is because the examples with which we are concerned are of considerable complexity, and it is hard to find any evidence which bears on these questions at all. Some examples which appear to have the necessary character turn out on further analysis to require a revision of the rules involved which removes them from the class of counterexamples (some cases of this sort are discussed in Anderson 1974); and of course a number of examples exist which are perfectly consistent with the claims of the theory. It now appears, however, that there are some genuine counterexamples to the revised simultaneous application algorithm, and that it must be abandoned.

An example of the required sort is furnished by a rule of vowel shortening in Fula, a West Atlantic language. In this language, a suffix containing a basic long vowel shortens that vowel before another (basic) long vowel. Thus the suffix /-i:-/ appears as long in njog-ii-mi “I grasped, will grasp” but as short in njog-i-maa-mi “I grasped, will grasp you.” The rule involved has the right character to provide a criterial example, and thus we can ask what happens when a form contains a sequence of four (or more) basic long vowel suffixes. If all of these (except the last, of course, which is not followed by a long vowel) shorten, the revised simultaneous theory is unable to describe this fact; and unfortunately that is precisely the case. In the form /suuɗ+etee+noo+moo+ɗaa/ we have the required conditions; and the surface form of this word cuuɗetenomocɗaa “he was hidden from you/you were hidden from him” shows that all but the last vowel in the sequence do indeed shorten (this form was provided me by Mr. Yero Syllah, a native speaker of the Futa Toro dialect who is currently preparing a phonological description of Fula).

The Fula shortening rule, then, forces us to abandon the revised simultaneous application theory, at least as a general and universal claim. The same conclusion is supported by Kenstowicz and Kisseberth (1977), in their discussion of the Slovak Rhythmic law, which appears to be another counterexample contrary to the discussion of this case given in Anderson (1974). Furthermore, Vago (1977b) has suggested the possibility of an account of the Acoma accent loss rule which eliminates this as the single case apparently supporting the revised simultaneous theory.

If this theory cannot be maintained, then, and the claim of the standard theory cannot either, does this mean that the theory of directional rules is confirmed? This theory is certainly capable of accounting for the great bulk of the data, including essentially all the cases that have been discussed in the literature. There is one case, however, which would appear to pose problems for this theory as well and thus to suggest that the answer to the problem of rule application is not yet at hand in a general form.

Many languages have patterns of stress assignment or, conversely, syllable reduction which affect every other syllable of a word or of some subsequence within a word. In general, such processes can be described by means of a variety of application algorithms, but there is one instance in which this is not clear: the case of a rule which deletes every other vowel within the relevant sequence. Potawatomi, for example (cf. Hockett, 1948) has a rule which applies to a sequence of syllables containing the vowel u and deletes the first, third, fifth, etc., of these (this account is slightly simplified, but not enough to alter the bearing of the example on our problem). Now let us imagine a directional account of this fact. Suppose the rule applies from left to right: in that case it will first examine the leftmost occurrence of u and delete that, and then move on to the next. But now the (originally second) u it encounters is the first of a sequence, and thus the rule must apparently delete it too. This mode of application will clearly delete all u’s, contrary to what is desired.

Suppose, on the other hand, the rule applies from right to left. But now the first vowel that is considered will be the rightmost of a sequence, and in order to determine whether this is an odd- or an even-numbered vowel, it must look all the way to the left end of the sequence, contrary to the spirit at least of the directional theory. There does not seem to be a satisfactory account of these facts within that theory, without resorting to ad hoc devices such as some diacritic alternating stress or other feature (not realized phonetically) which is assigned to all syllables to be (or not to be) deleted before the deletion actually takes place, obviously an undesirable alternative.

The structure of this example appears to be such that the determination of which vowels are odd-numbered and which even should be made all at once, since once some are deleted the relations within the sequence have been changed. This example can of course be accommodated quite well by the revised simultaneous theory, but we have seen above that that theory is not able to account for other sets of facts.

It may well be that there is an adequate account of examples such as the Potawatomi case, which seems to indicate a simultaneous theory of some sort, within the theory of directional rules. We should also allow for the possibility, however, that current discussion has simply misconceived the terms of this problem raised by the standard theory, and that a totally new approach to the facts that have entered into discussions of it thus far is called for. It may also be the case that there simply is no interesting (or at any rate completely general) solution: there may be some elements of language-particular specification in the interpretation of how to apply rules of this type to a given form. At any rate, discussion of this problem has been a major area in which phonologists have attempted to revise and extend the standard theory.

4. THE STRUCTURE OF REPRESENTATIONS IN PHONOLOGY

The standard theory’s view of the internal structure of phonological and phonetic representations is an appealingly straightforward and homogeneous one. At all levels of representation, these have the structure basically of a simple two dimensional array traditionally treated as a matrix. This array is divided into sequential elements of uniform size (segments) with no additional hierarchical structure; each segment is a simultaneous specification for each of the features provided by universal phonetic theory. It is assumed that the matrix is provided with a simultaneous proper bracketing as a result of syntactic and lexical structure, though this bracketing is really of interest only insofar as it controls the operation of the transformational cycle, which as we have seen above may well be unnecessary. Additional structure is provided by the assumption that morphemes, words, and phonological phrases are delimited by boundary elements; these boundaries, however, are treated as quasi-segmental entities with their own internal feature structure. They do not so much provide a (hierarchical) structure to the string of phonological/phonetic segments per se as simply partition it: they are intercalated among the segments as entities of the same basic sort.

There are three aspects of this sort of structure that have been extensively discussed in the recent literature: a) the question of what boundary elements should be posited, what sort of internal structure a feature system for these should have, etc.; b) the question of whether some additional hierarchical structure, especially an organization of segments into syllables, is necessary and appropriate; and c) the question of whether (in one way of posing the problem) the segments themselves have internal structure, and whether a given feature specification always takes exactly one segment as its domain.

With regard to the first two of these questions, I have very little to say. Stanley (1972) provided an important re-study of the nature and inventory of boundaries in phonology, and other papers since have addressed this issue at least tangentially, but there do not presently appear to be distinct, substantive positions concerning boundaries that divide phonologists in the way other problems discussed above do.

The question of the role of syllabic structure in phonology also must be neglected here. This is, of course, not an issue arising particularly within the standard theory of generative phonology: it is rather a standard problem of traditional phonetics, though posed here in different formal terms. The problem of the syllable has been mostly examined from the perspective of ‘non-standard’ theories, especially that of Natural Generative Phonology (cf., for example, Hooper 1972 and Vennemann 1972b). Recent work by Kahn (1976) involves proposals for motivating and incorporating syllabic structure within the standard theory, but this discussion is still in an early stage as concerns its ultimate import for revisions in the theory. Perhaps the upcoming Symposium on Segment Organization and the Syllable, to be held in October, will clarify some of these issues.

A great deal of interest has appeared in recent years, however, in the problem of the integrity of the segment. The first basic questions in this area arose from the study of tone, where the pioneering work of Leben (1973) made it clear that the traditional assumptions concerning segmental structure were in need of revision. A basic point arises in the treatment of contour tones, where it can be shown clearly that a single tone-bearing segment (usually a vowel) must be associated with more than one tone feature (for a summary of this and related issues, cf. Anderson, forthcoming). On the other hand, the study of other areas of phonological structure, particularly the phonology of nasality, argues that a single specification for a single feature may also have as its domain more than a single segment (though perhaps not an integral number of segments): cf. Anderson (1976c). The general conclusion seems clear, then, that the assumption of uniform-sized segments (uniform from the point of view of all features, that is) must be abandoned.

Two more or less independent approaches have been made to this conclusion: the program of Autosegmental phonology, represented at this conference by Goldsmith’s contribution, and the line taken by Anderson (1976c) which is quite close to the sort of phonetic description appearing in work by Pike (1943) and some American Structuralist accounts. There are obvious superficial differences between these two approaches, but there are also obvious similarities of principle, and it is not clear at present that there are empirical grounds for distinguishing between the two. Some discussion of this issue will be found in Goldsmith’s paper at this conference and we will not go into it further.

Discussion of these problems has reached general agreement on the proposition that feature specifications need not take as their domain precisely one segment: single segments (in the usual phonetic sense; of course precisely the problem we are dealing with here raises fundamental difficulties with the notion of “segment”) may have more than one specification for the same feature, sequenced in some way, and a single specification may take more than one segment as its domain. This substantial enrichment of the notion of phonological/phonetic representation opens up a wide range of problems and possibilities which are only beginning to be explored.

The study of tonal patterns, for example, has suggested that in many cases such patterns extending over substantial stretches of an utterance may be made up of only a very few elements. This, in turn, raises the question of whether other phonological processes of similar appearance may best be treated in a similar way. An obvious candidate for such a status is the process of vowel harmony: one might well propose that, instead of specifying each vowel of the word for the harmonic features, there is only one specification, taking as its domain the entire word (or whatever stretch constitutes an appropriate scope for the harmony). Just this sort of analysis has been proposed by Clements (1976b, 1976c). It appears at present that the arguments for treating harmony in this way rather than as a segmental assimilation do not carry conviction (cf. Anderson, forthcoming, for discussion of this controversy), but the issue can hardly be considered resolved. In any event, there are undoubtedly other cases, outside the domain of tone, for which such a ‘prosodie’ (as opposed to assimilatory) approach to intersegmental identities is appropriate. Some facts concerning nasals appear to have this character (cf. Anderson 1976), and the search for others is a particularly exciting area of current research.

Another particularly promising line of research follows from the fact that, if the sort of structure we envision here is appropriate for phonological representations, it should be the case that rules of the phonology of particular languages can manipulate it; and the properties of such rules remain to be examined. To turn again to the phonology of tone, we can consider the proposals of Hyman and Schuh (1974). They distinguish between rules of tonal assimilation and rules of tone spreading; rules of the latter sort are precisely the sort whose formal properties require discussion, since they involve not simple changes in feature values, like garden variety rules, but rather extensions and/or contractions of the scope of a feature value.

Rules of this sort seem particularly suggestive in the context of the problem of diphthongization, as discussed in an important paper by Andersen (1972). He distinguishes two stages in diphthongization: the phonetic diphthongization, in which a single segment becomes internally complex by a polarization of one or more of its features and a phonemic diphthongization, by which such a complex segment is reinterpreted as a sequence of two (or more) segments. In the development of diphthongs from tense vowels in English, Icelandic, and many other languages, for example, we could distinguish stages a) in which, e.g., /o:/ is a single, homogenous segment; b) in which /o:/ remains a single segment, but comes to be realized phonetically with a rise in tongue height at its offset; and c) in which /o:/ has been reinterpreted as the sequence /ow/ (at some level of structure, not necessarily under-lyingly):

The two processes of phonetic diphthongization and phonemic reinterpretation of complex segments seem to yield a more interesting account of the process of diphthongization than do the hitherto accepted processes of glide formation, etc. involved in previous generative discussion. Once we admit such processes (whose essence is the manipulation they perform on the domains of feature specifications, not their values) other areas of potential application are opened up.

It seems possible, for instance, that by recognizing such rules we can propose a constraint that phonological rules never insert epenthetic segments. Epenthesis would then be treated as a sequence of internal diphthongization, followed by phonemic re-interpretation, followed perhaps by some segmental modifications. This is clearly a reasonable approach to a large number of segment-insertion processes that have been suggested in the literature (consider, for example, the appearance of epenthetic stop consonants between nasals and following obstruents in English, discussed in these terms in Anderson, 1976c). Other cases that might not seem to be of the same sort are probably amenable to such a treatment as well, however. Consider the development of epenthetic e before clusters of s plus consonant in Spanish. We can suggest, following Hooper (1976), that this process has its origin in a revision of the syllable structure canon of Spanish, by which such clusters were prohibited. As a result of this change, the s of such clusters was expelled from its original syllable and came to form a syllable of its own. Thus (marking syllable boundaries with.), sta.re becomes s.ta.re. We can now assume that such ‘syllabic’ s’s undergo diphthongization with respect to the feature [±syllabic], becoming [ss]; subsequently, syllabic s (or perhaps more generally syllabic obstruents) are replaced with a more natural vowel [e], yielding modern estar. Such a history is of course largely conjectural, but if validated in some interesting cases it would suggest the existence of a constraint forbidding ‘structure building’ rules in phonology other than through the processes of internal phonetic diphthongization followed by subsequent re-interpretation.

We should note at this point that such a constraint would only apply to phonological rules, and not to rules with morphologically conditioned environments. Most of the modern Algonquian languages, for example, have a rule which inserts a t between a possessive prefix and a following vowel-initial stem. Such a rule obviously involves morphological conditioning; since there is no reason to believe it was ever phonetically motivated, there is no reason to expect a constraint of the form we are considering to have applied to it. This is of course another example of the utility of the distinction between phonological and morphological rules discussed above in making precise the domain of theoretical principles in phonology.

Processes of the sort we have been discussing can also be invoked, it appears, to exclude another type of formal operation in phonology: compensatory lengthening. While it is not certain that all putative cases of this classic category of phonological change can be so analyzed, preliminary investigation (cf. DeChene and Anderson, 1977) suggests that compensatory lengthening can generally be treated as a combination of sonorization to a glide of the element to be lost, followed by monophthongization of the resultant vowel plus glide sequence—the inverse of the diphthongization process dealt with above.

We conclude, then, that the enrichment of phonological structure involved in feature specifications having either more or less than a whole ‘segment’ as their domain raises a number of interesting possibilities for phonology. A significant number of traditional phonological process types may well turn out to be more insightfully (and restrictedly) formulated in these terms than in the terms of previous purely segmental operations.

5. THE RELATION OF DESCRIPTION TO EXPLANATION

In the last chapter of Chomsky and Halle (1968), a number of problems are cited with the strictly formal approach of the standard theory. These all center around the fact that such an approach completely neglects the phonetic content and function of phonological representations and rules. An approach is suggested, based loosely on ideas of Trubetzkoy and Jakobson, which will allow some of these notions to be brought within the notational approach to phonology that characterizes the rest of their work.

It is more than a little ironic that Chomsky and Halle characterize the earlier Prague approach to markedness by stating that “[A]fter a promising start, the exploration of these problems was not continued, largely because it seemed impossible to surmount the conceptual difficulties that stemmed from the taxonomie view of linguistics, which was all but universally accepted at that time. The attempts to break out of the confines of this view [. . .] elicited little positive response and almost no interest among contemporary workers, and the notion of markedness is hardly mentioned in the phonological literature of the 1940’s and 1950’s.” (Chomsky and Halle 1968:402). While it would certainly be difficult to attribute a ‘taxonomie view of linguistics’ to Chomsky and Halle’s work, it is striking that their attempts to surmount the limitations inherent in a purely formal approach to phonology by developing a theory of phonological substance elicited virtually the same lack of response as did the earlier proposals of Trubetzkoy and Jakobson. To the best of my knowledge, not a single phonological description published since 1968 has attempted to carry out the program of incorporating marking conventions, and (while occasional favorable references to the idea of having such a theory appear in the literature) virtually no discussion and development of their ideas can be found.

The basic idea of the markedness proposal involves the observation that some configurations (of features or of segments) are more to be expected (perhaps in a contextually determined way) than others. These (putatively universal) facts are expressed in a set of marking conventions, considered to be part of linguistic theory and thus stated once and for all rather than being part of the grammar of a particular language.

Now just as the notational conventions which capture the degree of coherence of a description are treated as fundamentally part of the procedure for evaluating grammars, so the marking conventions are treated as an aspect of the evaluation metric, and their function is to render forms conforming to the universally expected conditions less ‘costly’ than forms which violate them.

This notion of natural configuration is then to be extended to capture the notion of natural rule in a fairly straightforward way: rules producing expected configurations are to be rendered less costly, while those that produce configurations violating these conditions are to become more costly. This is to be achieved by assigning each marking convention an added function: it is also to serve as a ‘linking rule’, applying to the output of every phonological rule in a derivation to perform corrective therapy on the output of the rule unless explicitly prevented (at some added cost) from so altering the otherwise expected result. Thus, those aspects of a rule’s effects which are ‘natural’ or ‘expected’ need not be stated in the rule at all (and are thus rendered cost-free) since they will come about by the operation of the (universal) marking conventions; while any aspect of a rule’s operation which violates expectation, or indeed which simply fails to enforce it, requires overt statement and a consequent increase in cost.

It is clear then that such an approach is completely consistent with the rest of the program of the standard theory, as it attempts to bring aspects of phonological substance under the scope of a formal treatment in terms of the notation for the expression of phonological processes.

Unfortunately, there are a number of problems with this approach which become apparent upon examination. Some of these are largely mechanical (though not insignificant), and relate to the fact that even highly natural rules are still rules of the grammar, with their own idiosyncratic statements, limitations, exceptions, interactions with other rules, etc. This point is developed to some extent by Anderson (1974) and by Stampe (1973a). It would appear that it is simply not possible to extract all of the naturalness from a phonological description and state it once and for all in linguistic metatheory, simply because each language’s implementation of such naturalness is potentially different, and the line between natural and unnatural processes is a continuum rather than a dichotomy.

A more serious limitation of the SPE approach to markedness comes from the fact that it is based on a theory of natural configurations, rather than on natural processes. Thus, it might be possible to state that clusters of obstruents heterogeneous in voicing are unexpected (‘marked’), but whether a given language eliminates them by progressive assimilation (as in English inflectional endings), regressive assimilation (as in, e.g., Russian), bidirectional assimilation to one preferred value (as in Swedish), or by breaking them up or simplifying them (as in some other languages) cannot be predicted from this statement alone. Thus, all of these ways of achieving the state in which all obstruent clusters are homogeneous in voicing must be treated as completely language-particular. Aside from the fact that this leaves the unmarkedness of voicing assimilation (or whatever replaces it) completely unstated, it also fails to capture the fact that regressive assimilation is far and away the most expected way of achieving the desired configurations, apparently, although all of the other possibilities are in a sense rules of ‘voicing assimilation’ too.

This limitation is an essential feature of the SPE proposal, since that proposal is simply an attempt to extend the notion of natural configuration (which has a straightforward interpretation in terms of a notation and an evaluation applied to that notation) to encompass the notion of natural rule as well. On such an approach, it is inevitable that anything having to do with the preferred or natural way of achieving a given natural configuration, or any phonetically motivated natural process which creates configurations that are (in isolation) unnatural (such as syncope rules which increase cluster complexity) must stand outside the theory. Thus, at best, even with a theory of marking conventions and linking rules a great deal of the original program of taking account of phonological substance still remains to be carried out elsewhere in the theory.

Partially in recognition of such problems, Chomsky and Halle suggest that the theory of markedness would have to be supplemented by a theory of ‘plausible’ phonological processes. Such a notion in fact comes quite close to the sort of substantive typology of phonological processes which we have suggested at various points above is necessary elsewhere in phonology (to define the possibilities of rules schematization, disjunctive order, and other properties of the organization of a grammar). Such a theory of plausible processes might well distinguish, say, weakening processes such as degemination, voicing of obstruents, spirantization of stops, replacements of fricatives by h, etc. as a unitary class, associating particular environments favoring such weakenings with this class, and imposing substantive constraints on (precisely) rules which fall into this category. Such a theory might well give us a way to bridge the gap between formal phonological statements and constraints and phonetic substance, by providing a way of incorporating aspects of phonetic explanation into phonological theory in an appropriately limited way (limited, that is, to the class of processes to which the explanation in question is applicable).

It would appear that phonological theory must also recognize a systematic statement of the distinctions among (sub-phonemic or) phonetic rules, phonological rules, morphological rules, and quasi-systematic relations among forms within the lexicon. Such a distinction (or at least the distinction between phonological and morphological rules) has already been argued for above (and more extensively in Anderson 1975). It corresponds closely to the ideas of Baudouin de Courtenay (1895), who gives a similar classification of processes.

It is Baudouin de Courtenay’s discussion, in fact, that points up what is probably the most important use for such a classification of process types, when he treats the conditions under which phonetic rules become phonologized, phonological rules become opaque and subsequently morphologized or lexicalized. Given a) a substantive typology of phonetically motivated rule types, each narrowly and specifically constrained; and b) an account of the conditions under which rules classified along another dimension (that of phonetic, phonological, morphological, etc.) can change their status along that dimension, we might well come quite close to a theory of grammar that is as explanatory as is intrinsically possible, in terms of our present notion of the nature of language. Concrete application of this approach to the specific problem of understanding “compensatory lengthening” can be seen in DeChene and Anderson (1977). Such a theory should not, of course, be expected to meet the condition of complete predictability, but rather the more limited notion of ‘exegetic adequacy’ suggested above. The very fact that a single linguistic system can apparently evolve in any one of several ways (as shown by the existence of a variety of descendents of a single ancestral language) would seem to argue that a perfectly predictive theory of language is impossible in principle, in any event. Such an exegetically adequate theory as suggested above, then, may well constitute both a more reasonable and an achievable approach to the reconciliation of problems of explanation while giving proper attention to those of description, the basic paradox approached through the notion of markedness in the standard theory.