Rules and constraints

Let us begin with the issue of rules versus constraints (see Goldsmith (ed.) 1993a, and Roca (ed.) 1997a). There seems to be a prevailing opinion in current phonology that it is somehow more respectable to work with constraints only, than to propose rules and then constrain their application, however heavily. For instance, Government Phonology (Kaye, Lowenstamm and Vergnaud 1985, Kaye 1988) includes principles and parameters, but no destructive operations, while Optimality Theory (Prince and Smolensky 1993) incorporates only constraints.

Wemight assume that positing constraints per se is uncontroversial, as they are part of all the phonological models surveyed here: but they are still criticized when they are part of theories which also contain rules, like LP. For instance, Carr (1993: 190±1) accepts that LP may in principle be highly constrained and therefore relatively non-abstract, but argues that `The crucial issue here is whether such constraints (if they are desirable) come from within the theory or have to be imposed from outside. If the latter is the case, then the LP theory itself is, for those seeking a non-abstract phonology, in need of revision.' How are we to assess whether constraints are `imposed from outside'? Is the condition against destructive operations in Government Phonology not `imposed from outside'? Why should the specification of the number of vowel or consonant elements, or the assumption that reference should be made to universal, innate principles, have the status of internally determined, intrinsic aspects of the theory, while the constraints of LP should not? For example, I shall argue below that the main constraint on LP is the Strict Cyclicity Condition (SCC), which does follow from the architecture of the model, insofar as it is restricted to the (universally cyclic) first lexical level. Moreover, it is quite possibly derivable from the arguably innate Elsewhere Condition, and may not therefore require to be independently stated. Even so, why should this be seen as such a conclusive advantage? If we consider language change, we see that purely formal attempts to explain developments have rarely been very successful. For instance, in the domain of word order change, scholars like Lehmann (1973) and Vennemann (1974) attempted to account for the correlations of certain logically independent word order properties, and the fact that the change of one often seemed to have repercussions for others, in terms of the principle of natural serialization; this would probably be interpreted today as a principle or a parameter (see Smith 1989). However, this principle is not, on its own, explanatory (Matthews 1981): it is only when issues of parsing and learnability (see Kuno 1974) are invoked that we begin to understand why change should proceed so regularly in a particular direction. It seems highly likely that the same should be true of phonology: synchronically or diachronically, we need external evidence to explain why certain patterns occur and recur. Thus, the SCC is not purely a formal constraint. Instead, like Kiparsky's Alternation Condition (Kiparsky 1973), which it is partially intended to formalize, it is a learnability constraint: grammars violating either condition will be harder to learn. This means that, for instance, a grammar ordering rules on Level 1, within the domain of SCC, should be easier to acquire than a similar grammar with the same rules permitted to apply on Level 2, where they will not be controlled by SCC.

However, there is one crucial difference between the constraints of LP and those of Optimality Theory, for instance: the former restrict rule applications, whereas the latter replace rules. The next question, then, is whether rules are required at all. There are two considerations here, which relate in turn to the question of transparency in the synchronic grammar, and to the importance accorded to universality.

Anderson (1981), in a study of `Why phonology isn't ``natural''', argues that the effects of sound changes may build up in a language over time so that ultimately extremely opaque phonological processes may be operating synchronically. For instance, in Icelandic, Velar Fronting operates in a synchronically highly peculiar environment, giving back velars before the front vowels [y] and [ø], and front velars before the diphthong [ai], with a back first element. However, once we know that historically, the problematic front vowels are from back [u] and [ɔ], while the difficult diphthong was earlier front [æ:], we can see that Velar Fronting applies in the context of historically front vowels. Anderson points out that a synchronic grammar must nonetheless contain a description of these facts, and that this synchronic rule will not be phonetically motivated, or universal. The synchronic state is simply the result of language-specific history, and the fact that we have a historical explanation means the synchronic rule need provide no more than a description.

Everyday, work-horse descriptive work of this language-particular kind is what phonological rules are for, and it is my contention that phonological theories need them, whether their proponents are happy to admit it or not. For instance, Goldsmith's introduction to his (1993) collection of papers, entitled The Last Phonological Rule, argues that rules and derivations should not be part of a theory of phonology. However, Hyman's (1993) paper, despite setting out to find cases where extrinsic rule ordering will not work, comes to the conclusion that it is, in fact, a viable approach, while other papers (notably Goldsmith's and Lakoff's) involve language-specific constraints, such as Lakoff's (1993: 121) statement that `When C precedes ?# at level W, an /e/ absent at level W intervenes at level P', which is surely an epenthesis rule by any other name. As Padgett (1995) notes, these papers also include sequential, extrinsic level-ordering of constraints, and are therefore scarcely free of the apparatus of derivational phonology.

Similarly, Coleman (1995: 344) argues that `Far from being a rule free theory completely unlike the SPE model, as its proponents claim, Government-based phonological analyses employ various derivational devices which are transformational rules in all but name ... Government Phonology is therefore as unconstrained as the models it seeks to replace.' For instance, Coleman points out that, to model the ostensibly prohibited deletion of segments, Government Phonology can first delete each marked element in turn, which the theory will permit; this will ultimately leave only the single `cold' element which can be removed by the Obligatory Contour Principle (OCP). Furthermore, many of the principles invoked in Government Phonology seem language-specific; for instance, Harris (1994) argues that the loss of [r] in non-rhotic English dialects results from the innovation of the Non-Rhoticity Condition, which allows the R element to be licensed only in onsets. This condition allows an accurate description of the synchronic situation: the question is why such a constraint should become operative in the grammar of a particular dialect or set of dialects at a particular time. We might be dealing with a parameter resetting; but then, of course, we would have to ask why the resetting happened. Principles and parameters theory is faced with similar difficulties in historical syntax; thus, Lightfoot (1991: 160) remarks that, at the point when a parameter is reset, `an abrupt change takes place, but it was preceded by gradual changes affecting triggering experiences but not grammars'. So, Lightfoot recognizes `piecemeal, gradual and chaotic changes' in the linguistic environment; these can affect, for instance, the frequency of a construction, and may be introduced for reasons of contact, or for stylistic effect. These changes are not amenable to systematic explanation; but they are important in creating the conditions for parameter resetting, which is intended to be explicable in terms of Lightfoot's theory of grammar. It is quite unclear where the language change actually begins, and what the status of these preparatory changes is. Of course, a rule-based theory has no particular advantages here; a rule of [r]-deletion would simply be written as a response to the loss of a segment which was present before, and we would seek out reasons for the loss in, for instance, phonetics or sociolinguistics. But we would not be taking the portentous step of labelling this variety-specific behavior as a condition or a constraint, or falsely implying universality.

Finally, and most controversially, we turn to Optimality Theory (OT). In this theory, Universal Grammar for phonology consists of two components, a function Gen, and a set of universal constraints on representational well-formedness. Gen (for `generate') takes a particular input, which will be a lexical entry, and generates all possible outputs - an infinite set of possible candidate analyses, which is then evaluated by the list of constraints. These constraints are universal, but crucially ordered differently for each language, to give the different attested surface results. Most theories of constraints in phonology have held that constraints are exceptionless. In OT, every constraint is potentially violable. This means that the `winning', or maximally harmonic representation will not necessarily be the one which satisfies every constraint. It will be the one which violates fewest. More accurately, since constraints are ranked, it will be the candidate parse which violates fewest high-ranking constraints.

Prince and Smolensky (1993: 101) accept that `Any theory must allow latitude for incursions of the idiosyncratic into grammar.' However, they argue that idiosyncratic behavior is not modellable using rules, but rather by `(slightly) modified versions of the universal conditions on phonological form out of which core grammar is constructed ... [which] interact with other constraints in the manner prescribed by the general theory' (ibid.). This assumption has various consequences. First, constraints may be too low-ranked in particular languages to have any discernible effect. This is not taken to affect learnability adversely, since the strong assumption of universality means the constraints do not have to be learned, only their ranking; note, however, that acquisition is non trivial given the explosion of constraints to be ranked in recent versions of the theory: Sherrard (1997) points out that only five constraints will give 120 possible grammars, while ten will allow 36 million. Contrast this with a rule-based approach, where a rule is written only where it captures phonological behavior in the language concerned; we would not write, for instance, a universal version of the Vowel Shift Rule with effects tangible only in English and concealed elsewhere. To do so would be against every requirement of learnability, and would also unacceptably blur the distinction between the universal and the language-specific.

However, the question also arises of quite how different a constraint based theory like OT is from a rule-based one. Prince and Smolensky's contention that constraints can be language-specifically modified leads to formulations like the now notorious Lardil FREE-V (1993: 101), which states that `word-final vowels must not be parsed (in the nominative)', and again seems a static recasting of a very language-specific deletion rule. In similar vein, Prince and Smolensky (1993: 43), in considering the constraint NONFINALITY, note that `It remains to formulate a satisfactory version of NONFINALITY for Latin.' What this means is that, logically, the issue is not solely one of determining the place of constraint C in the hierarchy of Language X. The formulation of C may also differ, and it is not clear how appreciably, between Languages X and Y. More generally, there is an issue of extrinsic ordering here, since while many constraints must be ranked language-specifically, there are others which are never violated, and which must therefore be placed universally at the top of the hierarchy. Prince and Smolensky (1993: 46) argue that this is acceptable since `we can expect to find principles of universal ranking that deal with whole classes of constraints'. If ordering is acceptable when it refers to classes of ordered items, a rule-based model should be equally highly valued provided that it involved level-ordering, or ordering all lexical before all postlexical rules, for instance.

Even closer to the core of OT, the definition of the function Gen is itself controversial. Although Prince and Smolensky (1993: 79) advocate a parallel interpretation, they concede that Gen can also be understood serially, in which case its operation is much closer to a conventional derivation:

some general procedure (DO- α) is allowed to make a certain single modification to the input, producing the candidate set of all possible outcomes of such modification. This is then evaluated; and the process continues with the output so determined. In this serial version of grammar, the theory of rules is narrowly circumscribed, but it is inaccurate to think of it as trivial.

However, this serial interpretation of Gen may be necessary; Blevins (1997) argues strongly that, without it, there is no way of verifying constraint tableaux, as each tableau will contain the allegedly maximally harmonic parse plus a random set of other candidates, but will not contain all possible parses, and therefore crucially does not contain all the evidence necessary to permit evaluation.

The perceived advantage of an OT account is the absence of specific processes; but it is unclear why such a theory, with vast overgeneration courtesy of Gen, should be seen as more parsimonious than a derivational theory with a finite number of non-overgenerating language specific rules. Of the papers in Roca (ed.) (1997a), which focus on the rules-constraints debate, a surprising number contend that rules and derivations are still necessary, while Roca himself notes that `OT is stretching its original formal fabric in ways that closer scrutiny may reveal are nothing but covert rules, and perhaps even derivations' (1997b: 39). Indeed, some work in OT is entirely open about the addition of rules: McCarthy (1993: 190) includes an epenthesis rule to account for the distribution of English /r/, and states quite explicitly that `By a ``rule'' here I mean a phonologically arbitrary stipulation: one that is outside the system of Optimality.' As Halle and Idsardi (1997: 337±8) argue, `Conceptually, reliance on an arbitrary stipulation that is outside the system of Optimality is equivalent to giving up on the enterprise. Data that cannot be dealt with by OT without recourse to rules are fatal counter examples to the OT research programme.' At the very least, this introduction of rules alongside constraints removes the alleged formal superiority of OT, making it just as theoretically heterogeneous as LP, for instance, in containing both categories of statement.