Stratal constraints:

Level ordering, Lexical Phonology, and the Latinate Constraint

Bloomfield already stated that "Normal roots combine with normal affixes, learned roots with learned affixes" (1933:252). A distinction between two classes of affixes in English has been widely acknowledged and often reformulated since then, finding its way eventually into the theory of Lexical Phonology. The attractiveness and importance of the distinction and the restrictions that go along with it lie in its simplicity: if correct, a whole range of difficult data could naturally be accounted for by positing a single theoretical device. Unfortunately, it is not as simple as that.

The differences between two classes of affixes in English observed by Bloomfield concern not only their etymology but also their phonological and morphological behavior. An example of this is the well-known stress-shift that some affixes impose on their bases, and others do not. In SPE this distinction is associated formally with two kinds of boundaries, the word or strong boundary '#' and the morpheme or weak boundary '+'. #-affixes attach to words, i.e. they are outside the domain of cyclic phonological rules like stress assignment, whereas +-affixes do not block the application of such rules. An oft-cited example illustrating this difference is

Siegel (1974) argues that each suffix is associated with one (and only one) boundary and establishes what later became known as the Affix Ordering Generalization (Selkirk 1982a), which says that +-affixes, or 'class I’ affixes in Siegel's terminology, are always attached before, and class II affixes (i.e. #-affixes) after stress assignment. This generalization predicts that certain combinations of affixes are ruled out on principled grounds. A derivative like atom#less+ity violates the Affix Ordering Generalization and is therefore ungrammatical.

Later on the ideas of Siegel and other researchers (e.g. Allen 1979, Selkirk 1982a) have led to the development of various more refined stratificational models in the framework of Lexical Phonology (e.g. Kiparsky 1982b, Halle & Mohanan 1985, Mohanan 1986) which distinguish between two levels or strata of affixation in English. In (4), a number of affixes and their stratal affiliation are listed.

Note that all stratificational models (except the most recent one by Giegerich 1998) are based on the assumption that the information that crucially decides on class-membership, and thus on the application of phonological and morphological processes, rests exclusively with the individual affix.

The advantages of a stratal model of morphology are obvious. As mentioned above, only one mechanism is needed to explain a whole range of restrictions on the combinability of stems and affixes or affixes amongst each other. However, there has been abundant criticism of the stratificational model, involving both empirical and theoretical arguments.¹

The empirical weakness of level-ordered affixation is that the generalizations are both too weak and too strong. Thus it has been pointed out by Strauss (1982) that class I suffixes can be attached outside class II prefixes as in the notorious example ungrammaticality. From this and similar facts he concludes that level ordering is not applicable to affixes on opposite sides of the stem, which considerably weakens the power of the generalizations. But even if we limit our attention to only one side of the stem, there appear to be numerous counterexamples to the Affix Ordering Generalization where class II suffixes attach inside class I suffixes. Forms in -ist-ic or -ize-ation illustrate this phenomenon, since the stress-neutral suffixes -ist and -ize appear inside the stress-shifting -ic and -ation, respectively. Such facts suggest that many suffixes seem to behave as if they were associated with both strata. For instance, -able is sometimes stress-shifting, sometimes not (cómpar+able, compár#able). Although the association of one suffix with more than one level can explain away some undesired facts, this is achieved by further weakening the predictive power of the model considerably. Yet another problem for level-ordered morphology is that sometimes affixes may attach to phrases as in out-thereness and spreadoutness (see also note 10 above), which is rather unexpected for a well-behaved suffix, since affixes in general should not have access to processes that apply later in the derivation.

The counterexamples mentioned so far reveal that the model rules out many combinations that are actually attested. To make things worse, the model also permits combinations that never occur. Thus, in addition to what stratum-models predict, we need further, perhaps more idiosyncratic, restrictions on stacking. The existing stratum models (again with the notable exception of Giegerich 1998) are not able to account for stacking restrictions or precedence rules within one level. Consider, for example, the failure of -(u)ous to be followed by -ize, both of them level 1 affixes, as in *sens-uous-ize. Finally, it is far from clear whether ungrammatical forms like *home-less-ity are ungrammatical because of a level ordering violation or because of other reasons.

In essence, a level-ordering theory of English affixation is too strong because it rules out attested combinations of affixes, and it is too weak because it fails to predict a number of apparent restrictions on affix stacking. The most pronounced systematic attack on level ordering was launched by Fabb (1988) who presents impressive evidence against standard arguments for a stratificational approach to English suffixation. He investigates the combinability of a broad range of English suffixes and comes to the conclusion that the attested and non-attested combinations cannot be explained satisfactorily by level ordering. He therefore claims that English suffixation is constrained only by the selectional restrictions of the suffixes involved. This is not a very attractive solution to begin with, because it assigns all kinds of restrictions to stipulated idiosyncrasies of the suffixes. Although this is a viable solution in principle, one would prefer an account under which at least some of the restrictions observed in the data might be due to more general principles. I will demonstrate that many of the apparent selectional restrictions proposed by Fabb are the natural consequence of independent morphological mechanisms that have to be stated anyway to account for other phenomena in the realm of English derivational morphology.

The rejection of level ordering is however not equivalent to a complete rejection of Lexical Phonology. Linguists like Booij (1994) have pointed out that the basic insight of Lexical Phonology is that morphological and phonological rules apply in tandem. This assumption is logically independent of level ordering and level ordering can therefore be disposed of without any harm. This is shown by Inkelas and Orgun (1995), who demonstrate that certain (classes of) affixes can trigger certain phonological mechanisms. Although they still call the classes of processes "levels", their notion of level has hardly any resemblance to the one used in earlier models. The standard view of level ordering expressed by Kiparsky (1982b) or Mohanan (1986) is that every form is subject to every level of the phonology. Mohanan conceptualizes Lexical Phonology as a factory with the levels as individual rooms: "There is a conveyor belt that runs from the entry gate to the exit gate passing through each of these rooms. This means that every word that leaves the factory came in through the entry gate and passed through every one of these rooms" (1986:47). Inkelas and Orgun challenge this claim by showing "that in Turkish forms are not subject to the phonological levels at which they do not undergo morphology" (1995:764, original emphasis)

An alternative to the idea of level ordering is the (re-)introduction of an etymological-morphological restriction known as the Latinate Constraint (cf. e.g. Aronoff 1976, Booij 1977, 1994, Anshen et al. 1986, Fabb 1988). In general this constraint has taken two forms. Bloomfield (for English) and Booij (for Dutch) postulate a general constraint, whereas the other authors mentioned advance rule-specific restrictions. In the following I will propose an intermediate position.

Refining the aforementioned authors' observations somewhat, it seems necessary to distinguish between three types of suffixes in English. The first type only combines with native stems, the second type only (or primarily) attaches to Latinate stems, and the third type ignores this distinction and combines with both kinds of stems. The only real example of the first type is probably verbal -en which almost exclusively attaches to Germanic stems (e.g. Anshen et al. 1986:2, 5), whereas most native affixes are immune to this distinction and attach to any kind of base. Affixes of the second type are all of Latinate etymology themselves, as for example adjectival -ive and -al, nominal -ity and -cy, or verbal -ize. The third category of affixes consists of the majority of Germanic affixes and some Latinate affixes such as -ment, and -able.

Affixes and stems can now be characterized as belonging to one of the three categories, expressed by the features [+ Latinate], [- Latinate] or [± Latinate]. The distributional facts mentioned can now be formally accounted for by the following morphological constraint, which operates on stems and affixes:

The constraint in (5) rules out the only possibility that indeed should be ruled out in principle, namely that a [+ Latinate] element combines with a [- Latinate] element. Thus, a [+ Latinate] base may only combine with affixes marked [+ Latinate] or [± Latinate], a base marked [- Latinate] may only combine with affixes that are [- Latinate] or [± Latinate].

Alternative general formulations of the Latinate Constraint have been suggested, but these fail to capture the facts correctly. If Bloomfield's term "normal" in the quotation above is interpreted as being equivalent to our feature [- Latinate], his constraint fails to account for the many attested combinations of [- Latinate] suffixes with [+ Latinate] stems. If the term "normal" is interpreted as equivalent to our feature [± Latinate], Bloomfield's constraint fails to account for the combination of [± Latinate] affixes with [+ Latinate] stems. A stronger version of the constraint would say that [+ Latinate] affixes do not combine with [± Latinate] stems, which is, however, too strong a prediction. Consider, for example, the suffix combination -able-ity. -able attaches to both [+ Latinate] and [- Latinate] stems and must therefore be characterized as [± Latinate]. The suffix -ity on the other hand is, by all accounts, [+ Latinate] because it attaches only to [+ Latinate] adjectives. However, there is one systematic exception, namely adjectives in -able, which are [± Latinate]. The only possible generalization is therefore the one suggested in (5), which is in line with the observations made by Plank (1981:132) in regard to Dutch and German. More far-reaching restrictions must therefore be rule-specific.

The crucial question for the researcher, and the native speaker, is of course how the stratal feature of a given suffix can be determined. Unfortunately, one cannot simply consult an etymological dictionary, because even if a certain suffix is borrowed from, say, Old French, this does not entail that it behaves in the expected way. Besides -able, an example of such a suffix is -ment, which etymologically is Latinate, but morphologically is not, since it attaches equally to native and non-native bases (cf. e.g. Marchand 1969:331-332). Solely relying on the etymology is inadequate in any case, as it presupposes that native speakers need to have etymological knowledge in order to master English derivational morphology. Although etymological knowledge may influence the individual's morphological competence (especially what concerns morphological creativity), it does not make sense to make the native speaker's morphological competence dependent on meta-linguistic knowledge. Thus the Latinate Constraint is, strictly speaking, a morphological constraint, and not an etymological one.² The presence of a particular stratal feature may therefore only be detected structurally, by way of observing the combinatorial properties of a given affix. Thus, it is often said (e.g. Burzio 1994:334) that [+ Latinate] affixes attach to bound stems, whereas [- Latinate] only attach to words. As a consequence, [+ Latinate] affixes are always closer to the base than [- Latinate] ones. Another, related difference between Latinate and Germanic affixes is of a phonological nature. The former often affect the phonological make-up of the base through segmental or prosodic changes, while the latter do not. Furthermore, native and latinate words often differ in their phonological structure. Thus it has been frequently pointed out that Latinate stems tend to be polysyllabic with stress patterns different from that of Germanic stems which tend to be either monosyllabic, or, if poly-syllabic, to bear primary stress on the first syllable.

It seems, however, that none of the criteria can be applied in a strict fashion, since counterexamples can be found quite easily. Furthermore, it is not clear whether phonological properties (instead of a morphological feature related to etymology may be the real cause of the observed combi nations of elements. Following suggestions by Inkelas (1989, 1993), Booij (1994:20) has proposed to eliminate the Latinate constraint altogether in favor of a purely phonological distinction between affixes that attach to a phonological stem and those that attach to a phonological word. This ap pears to be a promising move, since such prosodic constituents are independently needed to account for a number of morphological and morpho phonological facts (see e.g. Nespor and Vogel 1986, McCarthy and Prince 1993b). No matter which kinds of features will ultimately turn out to be the most adequate, it seems that affixes form clusters or strata which function as organizational entities in the derivation of complex words.

¹ See, for example, Aronoff and Sridhar (1983, 1987) for a summary of arguments.

² Consequently, I continue to use square brackets to refer to the morphological property, and no brackets where I refer to the etymological characteristic proper.