Theoretical implications

I have presented a model of phonological well-formedness conditions on English -ize verbs in the framework of OT. It has been shown that earlier approaches by E. Schneider, Gussmann, Kettemann, and Raffelsiefen are not satisfactory. A number of violable output constraints were proposed that can both account for the apparent phonological variability of recently formed -ize derivatives and predict the phonological make-up of possible derivatives.

The proposed analysis raises some important and controversial theoretical issues. The first of these concerns the controversy between rule-based and constraint-based theories in linguistics, the second the status of the proposed constraints within the prosodic system of English, and the third relates to the nature of the morphophonological mechanisms at work.

Let us start with the problem of rules vs. constraints. The analysis presented above does not only show the versatility of OT for the description of morphophonological phenomena in English but also makes a strong point against rule-based approaches to allomorphy, or word-formation in general. It was pointed out that rule-based approaches cannot cope ade-general. It was pointed out that rule-based approaches cannot cope adequately with the obvious variability in the application of a putative rule. The non-uniformity of certain effects might be captured by imposing complex special restrictions on the rules, e.g. of the type "apply the rule except when...", similar to the generalization in (Haplology and stress (1), but such mechanisms are ad hoc and require an explanation themselves. As we have seen, violable constraints provide an empirically superior approach to morphophonological alternations.

Most recently, Neef (1996) has argued for non-violable well-formedness conditions in morphology instead of violable constraints of the OT type. At this stage it is unclear how his theory could handle the infrequent, but nevertheless attested, violations of some almost exceptionless well-formedness conditions of -ize, for example the near-non-occcurrence of stress shift or the rarity of iambic bases. In the present model, violation of constraints can be expected but it was argued that the high rank of certain constraints makes their violators highly marked - and therefore rare - structures. Although the exact relationship between type frequency and markedness remains to be studied (but see Golston and Wiese 1998 on German roots), the present model is preferable to Neefs, because in his approach such forms would have to be treated as truly idiosyncratic. Furthermore, generalizations (or 'word design conditions' in Neefs terminology) such as the ones put forward. are not fine-tuned enough to predict the different kinds of phonological alternations that occur, which means that the implementation of violable constraints yields the empirically more adequate results. But even if better word design conditions could be found, the same problems arise as with rule-based frameworks: why do we find these kinds of design conditions and not others? How can these conditions be independently motivated? In the model I have proposed, the violable constraints play a significant role in other areas of English phonology as well as in the phonological systems of other languages. Thus a growing number of studies in OT have demonstrated that constraints such as MAX, NONFINALITY, IDENT-HEAD etc. are universally relevant. Neefs design conditions, to the contrary, appear accidental and are necessarily language-specific.

The second theoretical problem raised by my analysis is that it entails a number of assumptions that are still under debate in OT. Thus I have argued that output-output constraints are needed to regulate stress preservation effects, that conjoined constraints are needed to account for some destressing effects, that crucial non-ranking of constraints is required to account for the observed OCP-effects and that morpheme-specific constraints can be integrated into the general hierarchy of constraints which is responsible for the prosodic structure of nouns in English. None of these points are original to this study but have been proposed by other scholars before in order to account for different phenomena in various languages (see the references above). An in-depth discussion of the theoretical problems involved would go beyond the scope of this investigation, but future research will show whether the proposed model is indeed tenable. More detailed evidence from a wider range of affixes is needed to find out more about possible and impossible constraints and their rankings and about the interaction of purely phonological constraints with those that only apply to morphologically complex words (e.g. output-output constraints).

The third problem concerns the - so far implicit - assumption that the stem allomorphy of -ize derivatives can be accounted for without making reference to the morphological structure of the base word. Thus, the shape of the optimal derivatives was solely determined on the basis of purely phonological constraints. It could, however, be argued that the constraints do not trigger phonological alternations in the putative base word, but that they either trigger the truncation of base-final morphemes, or that they only ensure the selection of the optimal stem allomorph. I will argue in the following that it is neither necessary nor possible to explain the stem allomorphy of -ize derivatives by making reference to morphological structure.

The strongest argument against phonologically governed allomorph selection or morphological truncation is of an empirical nature. Morphological truncation rules have been proposed, for example, in Aronoff (1976) or Booij (1977). The classic example of such a rule is the truncation of the verbal ending -ate when -able is attached (e.g. demonstrate - demonstrable). According to Aronoff (1976), this rule must make reference to the morphemic status of -ate since non-morphological -ate (as in debate) does not truncate (*debable). As pointed out by Anderson (1992:280), these facts can equally well be expressed as a purely phonological generalization: only secondarily stressed -ate can be deleted, irrespective of the morphological status of the string [eɪt].

In general, it can be observed that truncation does often not involve morphological constituents. Corbin (1987:345), for example, provides a whole range of French data which demonstrate that the deleted sequence at the end of the base cannot be a morpheme, and the same is true for -ize derivatives. For example, the deletion of base-final [ɪ] in many of the neologisms under discussion can hardly be analyzed as the truncation of a putative morpheme -y, since [ɪ] does not seem to have morphemic status in words like anthropólogy, (Madame) Bo̒vary, fántasy, mediócrity, Nórmandy or vaséctomy. In particular, I am unaware of any claims that -logy or -ity are bi-morphemic suffixes, consisting of -log and -y, or -it and -y, respectively. What the six words really have in common is not the morphology of putative -y, but their prosody: they are dactyls that end in a vowel. As we have seen above, such a phonological structure leads to the truncation of the final vowel across the board, ignoring morphological structure: anthropologize, bovarize, fantasize, mediocritize, Normandize, vasectomize. This point is corroborated by the form patinize, where the final schwa of the base word patina does not represent a suffix either, but is nevertheles truncated because the base conforms to the kind of prosodic structure that necessitates the deletion of the final vowel when -ize is attached.

Similar arguments hold for the haplology cases, where -ite in appetite or -is in metathesis can hardly be regarded as English morphemes (if morphemes are units of sound and meaning). However, Kiparsky (personal communication, August 1997) pointed out to me that a form like parallelize may serve as evidence that perhaps only morphemes can be deleted, since, according to the proposed OCP constraints this derivative would be less optimal than the truncated *parallize. Under an account that only allows the deletion of [əl] if it represents a morpheme, the existence of parallelize could be predicted.

This argument is not entirely convincing, however. First, parallelize is a very old form, first attested 1610, which means that it does not necessarily reflect the constraints of present-day English morphology. This is corroborated by the fact that native speakers generally prefer the synonymous converted verb parallel to parallelize (see also the respective entries in the OED). Second, it seems that suffixal -al is equally never truncated (cf. federalize and many more examples), so that the argument in favor of morphological truncation collapses. Third, even if we allow parallelize to be a possible derivative (and not only an actual one), it is still conceivable that OCP-ONSET needs to be further specified for phonological features. For example, [I] might be allowed to appear in identical onsets whereas other consonants might not. Unfortunately, there are no data that could show this, because all stems which can take adjectival -al and which end in [I] take -ar instead of -al as an adjectival suffix (cf. *polal vs. polar).¹ In principle, however, the featural specification of OCP constraints is not unusual² and could perhaps also solve the problem at hand. In view of the arguments just presented it is however preferable to assign idiosyncratic status to the word parallelize.

In summary, the idea of morphological truncation has two main flaws. It necessitates the postulation of otherwise unmotivated morphological structure, and it cannot explain the robust phonological generalizations that hold across morphologically diverse derivatives.

The only remaining alternative to a purely phonological model is therefore to assume that -ize selects certain stem allomorphs in order to satisfy the constraints. Under this approach, the deleted material need not be a morphological constituent, but only the base would have to be a morphological unit, namely a bound stem. This idea is preferable to the one just dismissed because it makes it unnecessary to assign questionable morphological status to certain sound sequences at the end of words. The problem is, however, to provide in each case independent evidence for the existence of the respective bound stem. While this may still justifiable with a stem like anthropolog-, which also surfaces in words like anthropolog-ical and anthropolog-ist, there are words where the putative bound stem only occurs in combination with -ize. Consider Normandize, whose stem Normand.· is not attested outside the base word Normandy, whereas all other complex words feature Norman(n)- as their base (.Norman-esque, Normann-ic, Norman-ish, Norman-ism, etc.). Similar empirical problems emerge with words like metathesize, where there is no independent evidence for a bound stem metathes- (unless one assumes that -is is a nominal suffix of English, which would be a truly innovational but nevertheless unconvincing claim). To complicate the situation, the only bound stem allomorph that does exist ends in -t- (as in metathetical), and it is mysterious why -ize would not select this stem allomorph if it prefers bound stems with these kinds of base lexemes.

Apart from these problems, the allomorph selection approach faces similar difficulties as the morphological truncation approach in accounting for the observed phonological generalizations across the many different derivatives. These generalizations would be accidental by-products of stem allomorphy because the stem allomorphs to be selected must be available a priori. But where and how do they originate? Under the present account, stem allomorphy is not (necessarily) a peculiar, given fact of the language, but can at least partially be explained as the result of independently motivated prosodic constraints that define the phonological make-up of a derivational category. In functional terms, stem allomorphy ensures the phonological coherence of the morphological category of -ize derivatives at the potential cost of reduced recognizability of the bases of individual derivatives. Sometimes these allomorphs are lexicalized or may become so, sometimes not. The important point is that under the selection model they must be lexicalized whereas under the present model they need not. The latter is clearly the more desirable solution.

On theoretical grounds, it seems that one should prefer an account that is most conservative in its assumptions. In the model proposed here, only one type of very general mechanism is needed to account for the data. Additional machinery like allomorphy selection should only be allowed if there is good independent evidence for it. The only derivatives where this is undoubtedly the case are those on the Greek bases discussed above. Note again that the selection of the Greek bound stem is not triggered by the phonological constraints, but by a purely lexical mechanism. The only other case where allomorphy selection might be involved are -ize derivatives on the basis of adjectives in -ous. It seems that these adjectives generally undergo truncation inspite of the fact that the non-truncated forms would be chosen as phonologically optimal in the above system. Thus, the only form which features the sequence -ous-ize listed in the whole OED is graciousize (obsolete/rare), with only one citation in 1701. Among the neologisms there is also only one form with a putative base in -ous, indígenìze. According to the model, this derivative is less optimal than the unattested ! indígenousìze. However, this analysis presupposes that indígenìze is indeed based on the adjective and not on the bound stem indigen-. Under the assumption that there is indeed a bound stem indigen which can serve as the basis for derivation, the attested form is also optimal. Due to the extremely low number of attested derivatives involving bases in -ous (truncated or not), I do not attempt to draw any firm conclusions with respect to bases ending in -ous. What the example of indigenize shows again, however, is that it is sometimes problematic to determine the base from which a complex word is actually derived. Thus it may well be the case that speakers coin words on the basis of bound stems, if available. Notably, this is not precluded by the model advocated above. What I have argued for is that the reliance on stem allomorphy selection as the general or the only mechanism can neither explain the kinds of stem allomorphy we observe nor the phonological generalizations that hold across morphologically diverse derivatives.

Although the foregoing discussion of the theoretical problems only concerned -ize derivatives, parallel arguments hold for the other complex derived by the affixation of -ate, -ify, eN- and -en and by conversion.

¹ This is an interesting fact by itself, because it shows the operation of another OCP constraint proposed in Plag (1998) which does not allow identical onset and coda in a single syllable. This constraint is obviously ranked lower with agentive -er since this suffix can be attached to stems ending in /r/, as in murderer. Note that the alternation between -al and -ar cannot be explained in terms of phonological alternation but should be regarded as an instance of allomorph selection, with the allomorphs being inherited from Latin.

² For example, the occurrence of epenthetic schwa in English verbs and nouns involving the inflectional suffix -s has been explained by a constraint which prohibits adjacent sibilants (see, for example, Yip 1996, Russel 1997).