The architecture of the english lexicon

1.5 Exceptions and the lexicon

Regardless of how well the linguist designs a traditional generative grammar for a given language (whether derivational or OT-based), there will always be some words which fail to behave as predicted. Chapter seven investigates the nature of such exceptions, and how they can be accounted for in OT. Some words are regarded as exceptional because of assumptions about the nature of their underlying forms. For example, because the surface forms of underived words is assumed to be identical to its underlying form, words like van’lla are understood as showing exceptional stress (cf. —rigin). The best solution for this type of problem involves what Inkelas, Orgun & Zoll (1994) refer to as prespecification. In this case, the lexical entry is understood as containing a geminate /l/, whose effects surface only in the surface stress of the word. Other cases of long segments (i.e., long vowels) result in a melodic as well as a stress contrast (e.g., hor’zon vs. —rigin).

Inkelas, Orgun & Zoll (1994: 28-29) regard prespecification as the best explanation of this type of exception within OT because:
On this theory, exceptions differ from "non-exceptions" É by possessing some underlying phonological structure causing them to resist or trigger a different class of phonological alternations than morphemes not similarly specified undergo. Crucially, alternations are governed by phonological structure, not by rule or grammar features. The contrast between a "regular " and "exceptional " morpheme is formally indistinguishable from that between any two phonologically distinct morphemes. Membership in a static subregularity and lexical exceptionality are formally identical, a species of simple lexical varietyÉ The second advantage of prespecification theory is that it offers the only workable approach to morphologically complex, phonologically heterogeneous wordsÉ A third significant advantage of prespecification is that it can locate and identify the exceptional area of the morpheme. (pp. 28-29)
Following this approach, regular and exceptional morphemes of this kind differ only in respect to structure, and are otherwise treated identically in the Gen and Eval processes. The exceptional structure can be identified and located in relation to the segments of the morpheme, and should be expected to appear consistently wherever that morpheme appears. Limiting the structures involved to members of the constituent hierarchies limits the types of exceptionality and helps to predict what kind of exceptionality is operating. Following the proposals made in chapters four, five and six, most exceptions to stress placement and vowel shortening can be accounted for by the presence of a lexical mora in the exceptional words, attached either to consonants (unexpected stress on an apparently light syllable) or vowels (unexpected failure of "vowel shortening" environments).

Another, less frequent, type of exception involves unexpected subcategorization for affixation. While most instances of affixation bring predictable prosodic effects attributable to subcategorization, some words display properties which are not expected on the basis of their presumed morphological structures. However, these properties are always identifiable with some other existing structure in the language. For example, the noun rŽcord does not show the expected prefixed noun structures seen in e.g., rep—rt or dŽfect, which foot both the prefix and the stem, but rather a single foot with unlengthened prefix syllable. Such a structure would be expected for a unitary stem like lŽgend or br’gand, but not for a prefixed stem. The combination of /re/ and /cord/ to produce a noun has arbitrarily (lexically) been governed by an unexpected subcategorization. Other exceptions of this kind include words like imprŽgnate, which shows an unexpected stress on the penult, normal in words suffixed by, e.g., /-al/, but almost absent from words in /-ate/, which normally show penultimate stress regardless of syllable weight. This is again accounted for by an unusual subcategorization, the suffix /-ate/ in this case arbitrarily subcategorizing for the stem rather than the morphological word (¤ 6.1).

A third type of exception involves words whose "input" morphemes themselves present an unexpected segmental form. For example, the pairs destr—y Ü destrœction, ret‡in Ü retŽntive cannot be linked to common underlying forms without proposing constraints which will yield erroneous results when applied generally to the lexicon. Furthermore, it should be clear that the selection of English morphemes in combination is in general arbitrary. For example, the verbs comp—se and prop—se are both built on the root /pos/, and their derived nouns c˜mpos’tion, pr˜pos’tion show similar morphology, but the further nouns prop—sal and comp—nent, the latter of which shows unusual allomorphy, are not predictable either in their structure or meaning. While words showing a common morphological root are often relatable semantically, their specific denotations and their specific morphological structures are not predictable. Other morphological entities, such as some prefixes (e.g., ab-) and roots (e.g., -cŽive), have no consistent meanings associated with them, although they might show a predictable behavior morphologically (e.g., concŽive Ü concŽption, decŽive Ü decŽption, etc.).

To account for these different types of arbitrariness, subcategorization is again applied. Just as Inkelas (1989) proposed that all morphemes have subcategorization frames, in an OT context all morphemes require at least one constraint associating their melodic segments with some morphological structures. McCarthy & Prince’s (1993a) subcategorization constraints for affixes are just a special case of the subcategorization constraints necessary for all morphemes to combine. Thus, while most roots will be governed by a predictable constraint of the type Align( /limit/, R; Root, R), linking the appropriate segment string to the root category, some roots will require an additional constraint which demands a prefix:

(1.20) Align( /ceive/, R; Root, R)

Align( /ceive/, L; Prefix, R)

Furthermore, the fact that semantically related words with identifiably related but nonidentical "input" strings exist (e.g., destr—y, destrœction) suggests that there is some semantic element involved in selecting morphemes and their subcategorizations. For example, if a semantic constituent, or lexeme {destroy} is proposed to represent the concept behind these two words, the following complex subcategorization constraints might be used to account for them:
(1.21) Link( {destroy}{Verb}, /de-stroy/)

Link( {destroy}{Noun}, /de-struct-ion/)

The constraint Link here simply refers to an unspecified associative relation, or simply morphological selection. Such a constraint chooses the root morpheme, its content-less prefix, and the nominal suffix in question, based on the semantically defined identity of the word in question. Following the same approach, a "regular" form would require less complex subcategorization constraints:
(1.22a) Link( {act}, /act/)

(1.22b) Link( {Noun}, /-ion/)

The constraint given in (1.22a) is a minimal association between a lexeme and the segmental string which consistently carries its meaning, while (1.22b) is a more generally applicable constraint for creating nominals, which can account for many more forms. Such constraints will do for "regular" words, but irregulars will require more complex and specific constraints such as those given above in (1.21). These constraints will have to be ranked above the more generally applicable constraints to have an effect upon the grammar.

Accepting such a proposal leads to a single conclusion: that all material traditionally ascribed to the lexicon in generative approaches must be encoded, in one way or another, into a series of subcategorization constraints of this type. In other words, the lexicon is entirely contained within the constraint hierarchy itself. This follows a proposal by Russell (1995) that constraints should be used to represent the lexicon itself as well as the processes which govern word-formation. Besides accounting for the various types of subcategorization required above, this approach also makes it easier to treat underspecification effects (¤ 7.3.3). Deriving the grammar exclusively from constraints within the hierarchy allows for all grammatical relationships to be expressed through the properties of the constraint hierarchy, such as ranking.

Constraints which refer to complex structures, like (1.21), or subcategorization constraints which refer to certain affixes, are referred to here as specific constraints. These can be contrasted with general constraints, which refer only to grammatical constituents (¤ 7.3.2). Constraints like (1.22b) fall into this category, as do most traditional OT constraints. Exceptions arise when specific constraints outrank their corresponding general constraints; the converse relation means that the specific constraints in question have no effect and do not participate in the grammar. Since the number of potential specific constraints is astronomically high, while the number of general constraints is limited by the number of constituents in the grammatical system, it is assumed that only high-ranking specific constraints, which would have an effect on the grammar, are understood to be present in the grammar. The fall of a specific constraint below its corresponding general constraint in the hierarchy effectively eliminates it from the grammar, and this is how regularization would be represented in this framework.

In ¤¤ 7.2-7.3, such a conception of the lexicon is contrasted with previous thinking. The traditional generative lexicon, which treats exceptions and regular forms through two distinct types of processing, is contrasted with Bybee’s (1995) proposal of a lexical network to model the relations between words, which takes into account type (pattern) and token (word) frequencies while rejecting discrete morphemes and process-oriented word formation. Each of these competing models is seen to focus on different sets of grammatical relationships, the generative model (like early OT) focusing on the prosody/morphology relationship, while Bybee’s network model focuses on the morphological and semantic connections between words. It is suggested herein that both models can be superseded by an OT grammar as outlined above, with regular correspondences captured by general constraints, which are indicative of high type frequencies, while exceptions are maintained by highly-ranked specific constraints supported by their high token frequency. The only true "input" forms are lexemes from a third semantic constituent hierarchy, which account for both the selection of morphemes and their subcategorization. While such selection is often consistent per lexeme, there are instances in which high-ranking selection constraints will arbitrarily choose allomorphs like /de-stroy/ and /de-struc/ for the same lexeme. It is these selected morphemic structures which comprise what has been until now treated as the OT "input string". Using the constraint hierarchy to model such relationships makes it logically equivalent to Bybee’s lexical network, but more effective since a wider range of relationships (including purely prosodic ones) can be captured, and all relationships can be described formally through constraints and their constituent arguments.

A further advantage of this approach is the presence of a built-in evaluation metric, which allows, unlike the previous approaches, for the complexity of a proposed grammar to be adequately quantified. The number of general constraints, which define the high-frequency relationships between members of the constituent hierarchies, should remain more or less constant between grammars (both cross-linguistically, and between competing grammars for a single language), as these relationships must always be specified. This follows the idea of McCarthy & Prince (1994) that all constraints are present in all languages, with the added proviso that only general constraints are meant. It is then the specific constraints which capture the idiosyncrasies and complexities of a real language; it is in fact preferable that such constraints are not thought to be present in all grammars. Specific constraints correspond to the unpredictable elements in the grammar, such as subcategorization, morpheme selection, and allomorphy. General constraints, driven by regularity (type frequency), are selected by the process of Lexicon Optimization to maximize the number of forms which can be accounted for by the grammar. In the least complex grammar, only those aspects of the language which are truly unpredictable via the general constraints will require specific constraints, enforced by the high token frequency of the irregular words which require them. The number of specific constraints necessary to produce the grammar is thus an indication of its complexity, and the grammar which minimizes the number of these specific constraints will be the least complex for a given language.

The final sections of chapter six deal with the further implications of the model outlined above for the theory. This proposed OT allows for the representation of the entire grammar through the constraint hierarchy, dispensing with the heterogeneous system of traditional generative theory, with its independent lexicon and morphological and phonological rules. This promises a more constrained system, but also then depends heavily on the constraints themselves. The need for a formal system of constraint definitions is discussed, although an exhaustive treatment is left to future work. Proposals for dealing with what is traditionally described as the post-lexical level are also introduced, although segmental phonology will not be elaborated upon in this work. Since a single constraint hierarchy must account for all relationships in this approach, the different correspondences captured previously in Lexical Phonology by level ordering must be accounted for instead by the use of hierarchically arranged constituents. A distinction between morphemic phonemes and prosodic segments (and their corresponding features) is proposed to account for various levels of segmental relationships.

The goal of the theory pursued here is the representation of the maximum number of grammatical relationships through the use of the minimum number of theoretical constructs, in this case the members of the three grammatical hierarchies and the mechanisms of the constraint hierarchy within Optimality Theory. By proposing a system in which the majority patterns can be produced by one ranking of the general constraints, while the minority patterns can be generated by lexically (through "lexical" subcategorization constraints) specifying some minimal element of structure (prosodic or morphological) for the underlying morphological forms involved, an argument is made both for the validity, productivity and elegance of the proposed model, and against the disconnected and arbitrary mechanisms of previous approaches.

2. Alternations in English

2.1 Patterns and regularity

As in derivational theory, regular morphology in OT involves those words whose proposed lexical "input" forms yield the correct "output" surface forms, while irregular morphology denotes the set of words for which the expected underlying forms would not produce the correct optimal candidates. Unlike the derivational approach, however, OT does not provide any expression for exception marking.¹⁴ Instead, irregular words will have to provide their surface forms on the basis of some lexical input which will produce the optimal candidate. Under this definition, only regular stems will be mappable to a single "underlying" morpheme, and a word’s irregularity may be based on the fact that its apparent lexical form must necessarily differ from those seen in related words. However, since the constraint hierarchy must produce all surface forms of the language, a proposed hierarchy that would necessarily prevent an attested surface word from appearing cannot be the correct hierarchy for that language. The hierarchy of general prosodic constraints must be by definition minimally violable for all surface forms.

To describe the interactions of constraints that take both morphological and prosodic categories as their arguments, forms showing irregular morphology must be removed from the set of data. Defining a complete constraint hierarchy (or grammar) for English thus involves determining which patterns are general and the result of the interactions of prosodic and morphological constraints, and which forms rather have idiosyncratic morphological structures. Once the irregular forms are removed from consideration, the patterns and sub-patterns found in the regular cases must be categorized and accounted for by known constraint interactions, and a hierarchy must be developed which yields all but the irregular cases.

The issue of phonological alternation in English will be explored in this chapter by looking at a class of moraically-characterized vowel alternations found in the English data. These vowel alternations, which appear to interact with prosodification processes expressible via the constraint hierarchy, will be presented and previous treatments of the data will be discussed. The computational corpus investigation undertaken for this study will be introduced and described, and data from that study will be used to illustrate the problems inherent in previous approaches, including solutions couched in Optimality Theory. The question of vowel length alternation will be shown to be dependent upon stress placement, which in turn suggests that the prosodic category "foot" and the constraints which take it as an argument will ultimately provide the explanation for these alternations. Evidence for patterns in the English stress system and the majority and minority stress patterns seen in the data will be investigated in the following chapter.

2.2 Investigating regular patterns in English

Despite the importance of alternations as evidence for particular constraint hierarchy configurations, there is actually not a great deal of stem alternation in English. The vast majority of words in the sample corpus do not provide examples for vowel alternation or stem allomorphy. Most suffixation in English, viewed in terms of surface phonetic realizations, appears to be simply concatenative. It might seem possible then, that we could simply follow Vennemann (1974) and just list the cases of apparent alternation, writing them all off as due to multiple stems. However, a prosodic analysis of English is necessary to address the issue of English stress patterns, and as will be illustrated below, generalizations seen across the categories required to describe English prosodic structure will provide the means for explaining many apparent phonemic alternations seen across morphemes. Particular alternation will be focused in the sections below upon vowel alternation, of interest because it involves both prosodic (moraic) and morphological components and is intimately connected with the question of stress and the structure of the prosodic system in English. Previous attempts to account for the alternation in a generative context will be discussed; in the course of the discussion some other relevant phonological alternations involving consonant loss and palatalization will be discussed in the context of the theories to be described.

2.2.1 Vowel alternation within the derivational paradigm

Myers’ (1987) analysis is useful as a starting point for a description of the problem, as he presents a summary of the traditional environments for vowel alternation in English, as well as an account of the processes involved, from a derivational perspective. The terms used below in the discussion of the traditional derivational approach will be, of course, couched in the language of derivational theory. The surface vowel alternation in question, between diphthongs and simple lax vowels, has been concisely described, in abstract phonological terms, as an alternation between long and short vowels (¤ 1.1), that is, vowels distinguished by the presence of phonemic vowel length, which would be described in prosodic phonology in terms of the presence or absence of a prosodic constituent, the mora. Myers identifies the following types of vowel alternation found between morphologically related words. The sample pairs given below are taken from his much longer lists. The first three sets involve regular alternations based on suffix-type:

(2.1a) Consonant-initial suffixes (Myers 1987: 488-9):

( /-t/, /-th/, /-tive/, /-tion/, /-ture/)

keep / kept wide / width

describe / descriptive convene / convention

discreet / discretion revise / revision

scribe / scripture five / fifth

(2.1b) Multisyllabic vowel-initial suffixes (Myers 1987: 494-5):

( /-ity/, /-ative/, /-itive/, /-atory/, /-ible/, /-able/, /-ify/, /-itude/, / ual/, /-acy/ )

sincere / sincerity sane / sanity

derive / derivative appeal / apellative

prime / primitive define / definitive

exclaim / exclamatory oblige / obligatory

creed / credible placate / placable

mode / modify vile / vilify

sole / solitude grateful / gratitude

grade / gradual rite / ritual

supreme / supremacy conspire / conspiracy
(2.1c) Monosyllabic vowel initial suffixes (Myers 1987: 499-500):

( /-ic/, /-id/, /-ish/, /-ule/)

cone / conic state / static

pale / pallid vapor / vapid

final / finish Spain / Spanish

globe / globule node / nodule

In all these cases, the addition of a suffix to a "base" (unsuffixed) form showing a long vowel apparently causes the vowel of the base form to shorten. These suffixes are, for the most part, those described in Lexical Phonological work as level I, the stress-affecting suffixes (Kiparsky 1982a, b, Halle & Mohanan 1985). These contrast with the level II affixes, such as / ship/, / ness/, /-ing/, which do not affect the location of main stress.

The shortening environments described above are familiar from Chomsky & Halle (1968), who proposed three separate shortening (in their terms, "laxing") rules to cover these cases, namely CC Shortening for set (2.1a), -ic/ id/-ish Shortening for set (2.1c), and Trisyllabic Shortening for set (2.1b). CC Shortening shortened underlying vowels appearing before two consonants, -ic/ id/-ish Shortening shortened vowels before that set of suffixes, and Trisyllabic Shortening shortened stressed vowels that were followed by two unstressed syllables. These rule environments reflected generalizations that Chomsky & Halle drew from their own English data.

Additionally, Myers notes the shortening seen in what have been called the "Latinate" prefixes, in suffixed words of the following kind¹⁵ (p. 500):
(2.2) re-fœte / rŽ-fut-ation de-f—rm / dŽ-form-ation

pre-sŽnt / prŽ-sent-ation pro-v’de / pr—-vid-ent

In these words, addition of a level I suffix to the base form apparently causes a long vowel in the prefix to shorten; furthermore, the main stress of the word shifts off the base to the prefixal syllable.

The preceding cases may be compared with a series of vowel-initial suffixes of level I which, although they may affect stress assignment, do not produce such vocalic alternations; the suffixed forms contain the same vowel as the base forms:

(2.3) Suffixes producing no length alternations:

( /-al/, /-or/, /-ory/, /-ary/, /-ous/, /-ive/)

scribe / scribal tone / tonal

advise / advisor propel/ propellor

advise / advisory supervise / supervisory

fame / famous desire / desirous

These suffixes have a similar level of productivity and distribution in the corpus of English to that of the "shortening" suffixes noted above.
2.2.2 Stress assignment and vowel quantity

According to Myers (1987: 486), "vowel quantity in English depends to a great extent on stress." This statement encompasses a range of notions, from the fact that Chomsky & Halle’s shortening rules referred explicitly to stress, to later observations, described in chapter one, that sound change occurs in the context of prosodic constituents. As was described above, the earliest metrical accounts of English stress, such as Liberman & Prince (1977), Hayes (1982), and Selkirk (1980), moved toward doing away with the segment-based stress rules of Chomsky & Halle (1968), replacing them with prosodic or metrical analyses. Vowel alternation was clearly being related to stress in these analyses, and in the relevant literature that touches upon this problem (Chomsky & Halle 1968, Myers 1987, Halle & Vergnaud 1987, Prince 1990, Sainz 1992, Burzio 1993), stress is always a factor in the formulation of vowel shortening rules. Consequently, such accounts are necessarily dependent upon their theoretical framework for stress assignment in their explanation of vowel alternation.

However, the early metrical accounts of English stress conspicuously did not address the issue of vowel alternation. For example, the influential account of stress given in Hayes (1982) does not treat any cases involving vowel alternation. Since Chomsky & Halle's vowel alternation rules were dependent upon their stress rules, improving upon their stress rules invited similar improvement for the vowel alternation rules. For the same reasons, discarding Chomsky & Halle's segment-based approach forced later studies to discard their formulation of the shortening rules; Myers (1987) attempted to use a prosodic category, the syllable, to help achieve the same results. Myers followed the prosodic theory of Selkirk (1984), which did not recognize the foot as a category, and based its approach to stress upon syllables and a metrical grid. Thus, the only relevant prosodic category available to the analysis is the syllable, and any rules which affect segments are bound to the syllable domain.