Lexical Phonology and its predecessor

Lexical Phonology (LP) is a generative, derivational model: at its core lies a set of underlying representations of morphemes, which are converted to their surface forms by passing through a series of phonological rules. It follows that LP has inherited many of the assumptions and much of the machinery of Standard Generative Phonology (SGP; see Chomsky and Halle 1968). LP therefore does not form part of the current vogue for monostratal, declarative, non-derivational phonologies (see Durand and Katamba 1995, Roca (ed.) 1997a), nor is it strictly a result of the recent move towards non-linear phonological analyses, with their emphasis on representations rather than rules (see Goldsmith 1990, and the papers in Goldsmith (ed.) 1995). Although elements of metrical and autosegmental notation can readily be incorporated into LP (Giegerich 1986, Pulleyblank 1986), its innovations have not primarily been in the area of phonological representation, but rather in the organizational domain.

The main organizational claim of LP is that the phonological rules are split between two components. Some processes, which correspond broadly to SGP morphophonemic rules, operate within the lexicon, where they are interspersed with morphological rules. In its origins, and in the version assumed here, the theory is therefore crucially integrationist (but see Hargus and Kaisse (eds.) 1993 for discussion, and Halle and Vergnaud 1987 for an alternative view). The remainder apply in a postlexical, postsyntactic component incorporating allophonic and phrase-level operations. Lexical and postlexical rules display distinct clusters of properties, and are subject to different sets of constraints.

As a model attempting to integrate phonology and morphology, LP is informed by developments in both these areas. Its major morphological input stems from the introduction of the lexicalist hypothesis by Chomsky (1970), which initiated the re-establishment of morphology as a separate subdiscipline and a general expansion of the lexicon. On the phonological side, the primary input to LP is the abstractness controversy. Since the advent of generative phonology, a certain tension has existed between the desire for maximally elegant analyses capturing the greatest possible number of generalizations, and the often unfounded claims such analyses make concerning the relationships native speakers perceive among words of their language. The immensely powerful machinery of SGP, aiming only to produce the simplest overall phonology, created highly abstract analyses. Numerous attempts at constraining SGP were made (e.g. Kiparsky 1973), but these were never more than partially successful. Combating abstractness provided a second motivation for LP, and is also a major theme here.

The problem is that the SPE model aimed only to provide a maximally simple and general phonological description. If the capturing of as many generalizations as possible is seen as paramount, and if synchronic phonology is an autonomous discipline, then, the argument goes, internal, synchronic data should be accorded primacy in constructing synchronic derivations. And purely internal, synchronic data favor abstract analyses since these apparently capture more generalizations, for instance in the extension of rules like Vowel Shift in English from alternating to non-alternating forms. However, as Lass and Anderson (1975: 232) observe, `it just might be the case that generalizations achieved by extraparadigmatic extension are specious'; free rides, for instance, `may just be a property of the model, rather than of the reality that it purports to be a model of. If this should turn out to be so, then any ``reward'' given by the theory for the discovery of “optimal” grammars in this sense would be vacuous.' In contrast, I assume that if LP is a sound and explanatory theory, its predictions must consistently account for, and be supported by, external evidence, including diachronic data; the facts of related dialects; speech errors; and speaker judgements, either direct or as reflected in the results of psycholinguistic tests. This coheres with Churma’s (1985: 106) view that ```external'' ... data ... must be brought to bear on phonological issues, unless we are willing to adopt a ``hocus pocus'' approach ... to linguistic analyses, whereby the only real basis for choice among analyses is an essentially esthetic one' (and note here Anderson's (1992: 346) stricture that `it is important not to let one's aesthetics interfere with the appreciation of fact'). The over-reliance of SGP on purely internal evidence reduces the scope for its validation, and detracts from its psychological reality, if we accept that `linguistic theory . . . is committed to accounting for evidence from all sources. The greater the range of the evidence types that a theory is capable of handling satisfactorily, the greater the likelihood of its being a ``true'' theory' (Mohanan 1986: 185).

These ideals are unlikely to be achieved until proponents of LP have the courage to reject tenets and mechanisms of SGP which are at odds with the anti-abstractness aims of lexicalism. For instance, although Mohanan (1982, 1986) is keen to stress the relevance of external evidence, he is forced to admit (1986: 185) that his own version of the theory is based almost uniquely on internal data. Elegance, maximal generality and economy are still considered, not as useful initial heuristics, but as paramount in determining the adequacy of phonological analyses (see Kiparsky 1982, Mohanan 1986, and especially Halle and Mohanan 1985). The tension between these relics of the SPE model and the constraints of LP is at its clearest in Halle and Mohanan (1985), the most detailed lexicalist formulation of English segmental phonology currently available. The Halle-Mohanan model, which will be the focus of much criticism below, represents a return to the abstract underlying representations and complex derivations first advocated by Chomsky and Halle. Both the model itself, with its proliferation of lexical levels and random interspersal of cyclic and non-cyclic strata, and the analyses it produces, involving free rides, minor rules and the full apparatus of SPE phonology, are unconstrained.

Despite this setback, I do not believe that we need either reject derivational phonology outright, or accept that any rule-based phonology must inevitably suffer from the theoretical afflictions of SGP. We have a third choice; we can re-examine problems which proved insoluble in SGP, to see whether they may be more tractable in LP. However, the successful application of this strategy requires that we should not simply state the principles and constraints of LP, but must rigorously apply them. And we must be ready to accept the result as the legitimate output of such a constrained phonology, although it may look profoundly different from the phonological ideal bequeathed to us by the expectations of SGP.

Then, I shall examine the performance of LP in three areas of phonological theory which were mishandled in SGP: abstractness; the differentiation of related dialects; and the relationship of synchronic phonological rules and diachronic sound changes. If LP, suitably revised and constrained, cannot cope with these areas adequately, it must be rejected. If, however, insightful solutions can be provided, LP will no longer be open to many of the criticisms levelled at SGP, and will emerge as a partially validated phonological theory and a promising locus for further research.

The three issues are very clearly connected; let us begin with the most general, abstractness. SGP assumes centrally that the native speaker will construct the simplest possible grammar to account for the primary linguistic data he or she receives, and that the linguist's grammar should mirror the speaker's grammar. The generative evaluation measure for grammars therefore concentrates on relative simplicity, where simplicity subsumes notions of economy and generality. Thus, a phonological rule is more highly valued, and contributes less to the overall complexity of the grammar, if it operates in a large number of forms and is exception less.

This drive for simplicity and generality meant exceptions were rarely acknowledged in SGP; instead, they were removed from the scope of the relevant rule, either by altering their underlying representations, or by applying some `lay-by' rule and a later readjustment process. Rules which might be well motivated in alternating forms were also extended to non-alternating words, which again have their underlying forms altered and are given a `free ride' through the rule. By employing strategies like these, a rule like Trisyllabic Laxing in English could be made applicable not only to forms like divinity (~ divine) and declarative (~ declare), but also to camera and enemy; these would have initial tense vowels in their underlying representations, with Trisyllabic Laxing providing the required surface lax vowels. Likewise, an exceptional form like nightingale is not marked [-Trisyllabic Laxing], but is instead stored as /nɪxtVngǣl/; the voiceless velar fricative is later lost, with compensatory lengthening of the preceding vowel, to give the required tense vowel on the surface.

The problem is that the distance of underlying representations from surface forms in SGP is controlled only by the simplicity metric ± which positively encourages abstractness. Furthermore, there is no linguistically significant reference point midway between the underlying and surface levels, due to the SGP rejection of the phonemic level. Consequently, as Kiparsky (1982: 34) says, SGP underlying representations `will be at least as abstract as the classical phonemic level. But they will be more abstract whenever, and to whatever extent, the simplicity of the system requires it.' This potentially excessive distance of underliers from surface forms raises questions of learnability, since it is unclear how a child might acquire the appropriate underlying representation for a non-alternating form.

A further, and related, charge is that of historical recapitulation: Crothers (1971) accepts that maximally general rules reveal patterns in linguistic structure, but argues that these generalizations are non-synchronic. If we rely solely on internal evidence and on vague notions of simplicity and elegance to evaluate proposed descriptions, we are in effect performing internal reconstruction of the type used to infer an earlier, unattested stage of a language from synchronic data. Thus, Lightner (1971) relates heart to cardiac and father to paternal by reconstructing Grimm's Law (albeit perhaps not wholly seriously), while Chomsky and Halle's account of the divine ~ divinity and serene ~ serenity alternations involves the historical Great Vowel Shift (minimally altered and relabelled as the Vowel Shift Rule) and the dubious assertion that native speakers of Modern English internalize the Middle English vowel system. I am advocating that historical factors should be taken into account in the construction and evaluation of phonological models; but the mere equation of historical sound changes and synchronic phono logical rules is not the way to go about it.

Here we confront our second question: how are sound changes integrated into the synchronic grammar to become phonological rules? In historical SGP (Halle 1962, Postal 1968, King 1969), it is assumed that a sound change, once implemented, is inserted as a phonological rule at the end of the native speaker's rule system; it moves gradually higher in the grammar as subsequent changes become the final rule. This process of rule addition, or innovation, is the main mechanism for introducing the results of change into the synchronic grammar: although there are occasional cases of rule loss or rule inversion (Vennemann 1972), SGP is an essentially static model. The assumption is that underlying representations will generally remain the same across time, while a cross-section of the synchronic rule system will approximately match the history of the language: as Halle (1962: 66) says, `the order of rules established by purely synchronic considerations ± i.e., simplicity ± will mirror properly the relative chronology of the rules'. Thus, a sound change and the synchronic rule it is converted to will tend to be identical (or at least very markedly similar), and the `highest' rules in the grammar will usually correspond to the oldest changes. SGP certainly provides no means of incorporating recent discoveries on sound change in progress, such as the division of diffusing from non-diffusing changes (Labov 1981).

It is true that some limited provision is made in SGP for the restructuring of underlying representations, since it is assumed that children will learn the optimal, or simplest, grammar. This may not be identical to the grammar of the previous generation: whereas adults may only add rules, the child may construct a simpler grammar without this rule but with its effects encoded in the underlying representations. However, this facility for restructuring is generally not fully exploited, and the effect on the underliers is in any case felt to be minimal; thus, Chomsky and Halle (1968: 49) can confidently state:

It is a widely confirmed empirical fact that underlying representations are fairly resistant to historical change, which tends, by and large, to involve late phonetic rules. If this is true, then the same system of representation for underlying forms will be found over long stretches of space and time.

This evidence that underlying representations are seen in SGP as diachronically and diatopically static, is highly relevant to our third problem, the differentiation of dialects. The classical SGP approach to dialect relationships therefore rests on an assumption of identity: dialects of one language share the same underlying representations, with the differences resting in the form, ordering and/or inventory of their phonological rules (King 1969, Newton 1972). Different languages will additionally differ with respect to their underlying representations. The main controversy in generative dialectology relates to whether one of the dialects should supply underlying representations for the language as a whole, or whether these representations are intermediate or neutral between the realizations of the dialects. Thomas (1967: 190), in a study of Welsh, claims that `basal forms are dialectologically mixed: their total set is not uniquely associated with any total set of occurring dialect forms'. Brown (1972), however, claims that considerations of simplicity compel her to derive southern dialect forms of Lumasaaba from northern ones.

This requirement of a common set of underlying forms is extremely problematic. Perhaps most importantly, the definition of related dialects as sharing the same underlying forms, but of different languages as differing at this level, prevents us from seeing dialect and language variation as the continuum which sociolinguistic investigation has shown it to be. Furthermore, the family tree model of historical linguistics is based on the premise that dialects may diverge across time and become distinct languages, but this pattern is obscured by the contention that related dialects are not permitted to differ at the underlying level, while related languages characteristically do. It is not at all clear what conditions might sanction the sudden leap from a situation where two varieties share the same underlying forms and differ in their rule systems, to a revised state involving differences at all levels. These theoretical objections are easily swept aside, however, in a model like SGP where the central assumptions require maximal identity in the underlying representations.

The three areas outlined above are all dealt with unsatisfactorily in SGP; moreover, these deficiencies are due in all cases, directly or indirectly, to the insistence of proponents of the SPE model on a maximally simple, exceptionless phonology. The use of an evaluation measure based on simplicity, the lack of a level of representation corresponding to the classical phonemic level, and the dearth of constraints on the distance of underlying from surface representations all encourage abstractness. Changes in the rule system are generally preferred, in such a system, to changes in the underlying forms, which are dialectally and diachronically static. Rules simply build up as sound changes take effect, with no clear way of encoding profound, representational consequences of change, no means of determining when the underliers should be altered, and no link between sound changes and phonological rules save their identity of formulation. This historical recapitulation contributes to further abstractness, and means that, in effect, related dialects must share common underlying forms. King (1969: 102) explicitly states that external evidence, whether historical or from related dialects, may play no part in the evaluation of synchronic grammars; this is presented as a principled exclusion, since speakers have no access to the history of their language or to the facts of related varieties, but is equally likely to be based on the clear inadequacies of SGP when faced with data beyond the synchronic, internal domain.

I hope to show that LP need not share these deficiencies, and that its successes in the above areas are also linked. Working with different varieties of Modern English, I shall demonstrate that the abstractness of the synchronic phonology can be significantly restricted in LP. In general, the strategy to be pursued will involve imposing and strengthening the constraints already existing in LP, most notably the Strict Cyclicity Condition or Derived Environment Condition, and assessing the analyses which are possible, impossible, or required within the constrained model. Because maximally surface-true analyses will be enforced for each variety, we will be unable to consistently derive related dialects from the same underlying representations, and the underliers will also be subject to change across time. Sound changes and related phonological rules will frequently differ in their formulation, and new links between diachrony and synchrony will be revealed.

Of course, this is not the first time that questions have been raised over aspects of SGP: for instance, I have already quoted Lass and Anderson (1975), a Standard Generative analysis of Old English phonology incorporating an extremely eloquent and perceptive account of the difficulties which seemed then to face SGP, a model which had seemed so `stable and unified' (1975: xiii) in 1970, when their account of Old English was first drafted. Lass and Anderson set out to test SGP against a particular set of data. They discover that the theory makes particular predictions; that it permits, or even requires, them to adopt particular solutions. These solutions are sometimes fraught with problems. Lass and Anderson could, of course, have made use of the power of SGP to reformulate the areas where they identify problems and weaknesses; instead, they include a final section explicitly raising doubts about the theory, and the issues they identify have been crucial in remodeling phonological theory ever since.

The conclusion, more than twenty years on, is that these difficulties cannot be solved within SGP: the simplicity metric, the overt preference (without neurological support) for derivation over storage, and the denial of `external' evidence, mean that many of the generalizations captured are simply over-generalizations. The model must be rejected or very radically revised.

LP is one result. But the revisions have so far not been radical enough. I shall show that it is possible to maintain the core of the generative enterprise in phonology (namely, that alternating surface forms may be synchronically derived from a common underlier) without a great deal of the paraphernalia which was once thought to be crucial to the goal of capturing significant generalizations, but in practice encouraged the statement of artefactual and insignificant ones. Thus, we shall reject the SGP identity hypothesis on dialect variation; rule out free rides; prohibit derivation in non-alternating morphemes; revise the feature system; and exclude underspecification, which has recently become an expected ingredient of LP, but is in fact quite independent from it.

In the rest, then, I shall follow much the same route as Lass and Anderson: we shall begin with a phonological model, in this case LP, and assess its performance given a particular set of data, here the vowel phonology, loosely defined, of certain accents of Modern English. The model is characterized by a number of constraints; I shall argue that these should be rigorously applied, and indeed supplemented with certain further restrictions. We can then examine what is possible within the model, and what solutions it forces us to adopt. If we are forced to propose analyses which seem to conflict with internal or external evidence, being perhaps apparently unlearnable, or counter historical, or without phonetic or diachronic motivation, we must conclude that the model is inadequate. Likewise, the model may never make decisions for us: in other words, any analysis may be possible. Such a theory clearly makes no predictions, and is unconstrained, unfalsifiable and uninteresting. On the other hand, we may find that the predictions made are supported by internal and external evidence; that the phonology becomes more concrete, and arguably more learnable than the standard model; that phonetics and phonology can be better integrated, and the relationship between them better understood; and that a more realistic model of variation and change can be proposed.

So far, I have introduced LP only in the broadest terms. A number of outlines of LP are available (Kiparsky 1982, 1985; Mohanan 1982, 1986; Pulleyblank 1986; Halle and Mohanan 1985). However, most aspects of LP, including its central tenets, are still under discussion (see Hargus and Kaisse (eds.) 1993, Wiese (ed.) 1994). Available introductions therefore tend to be restricted to presenting the version of LP used in the paper concerned (Kaisse and Shaw 1985 does provide a broader perspective, but is now, in several crucial respects, out of date). Consequently, it may be difficult for a reader not entirely immersed in the theory to acquire a clear idea of the current controversies, which become apparent only by reading outlines of LP incorporating opposing viewpoints. I shall consequently attempt to provide an overview of LP, considering both its evolution, and current controversies within the theory. First, however, I must justify approaching the problems outlined above in a derivational model at all.