present walk is opposed by the dedicated third singular form walks and accordingly ...... will subcategorize for [VFORM
J A M E S P. BLEVINS
SYNCRETISM
AND PARADIGMATIC
OPPOSITION*
This paper aims to consolidate the insights of two feature-theoretic treatments of inflectional syncretism within an account that recognizes morphological paradigms as a locus of distinctive grammatical oppositions. The proposed account in large part adopts the underspecification analysis of syncretism presented in Andrews (1990), in which 'non-third person singular' forms like walk in English are simply unmarked for person and number features. It also follows Andrews in attributing the illformedness of *he walk to a morphological blocking condition that prohibits the use of an unmarked form when a more specific alternative - in this case, walks exists. However, I will argue that the operative notion of 'alternative' is more revealingly and efficiently characterized in terms of paradigmatic relatedness than in terms of the synonymy relation invoked by Andrews. In this respect, the present account develops the more 'logistic' approach to syncretism represented by Pollard and Sag (1987), in which the specifications assigned to walk are defined in direct opposition to the features associated with the marked form (again walks) of the same lexeme. The principal innovation in the synthesis outlined below involves organization of the lexicon into hierarchical paradigms, consisting of lexical items that either share a lexeme-identifying feature value or are defined from a common base specification. The feature type hierarchies employed in Pollard and Sag (1987) complement the underspecified paradigms in this account by providing a global representation of feature neutralization. Thus the coalescence of nominative and accusative case or masculine and neuter gender in German declensional paradigms is succinctly represented in a hierarchy that includes 'nonoblique' and 'nonferninine' macrotypes. The inclusion relation between such feature macrotypes and their partially neutralized subtypes supplies a restricted alternative to the disjunction of feature specifications. Disjointness relations among the feature types in a hierarchy likewise provide a constrained alternative to negative specifications. These alterna-
* I am grateful to Shelly Harrison, Roger Higgins and Edwin Williams for helpful discussion, and to two anonymous reviewers for comments and suggestions that have led to improvements in the present version.
Linguistics and Philosophy 18: 113-152, 1995. © 1995 Kluwer Academic Publishers. Printed in the Netherlands.
114
JAMES P. BLEVINS
tives moreover achieve the economy of disjunction and negation while incurring few of the attendant complications. The use of feature underspecification and structured feature inventories primarily identifies their distinctive contribution to the reduction of inflectional syncretism. However this account also highlights a couple of points of general interest for current feature-based approaches. In the first place, it suggests the utility, both theoretical and practical, of incorporating information about paradigmatic relations in the description of syntactic systems. Moreover, it illustrates a useful division of logical labour, as the structure assigned to the general feature inventory permits radical simplification of individual entries. This simplification or partial underspecification in turn facilitates a morphological blocking analysis of the sort proposed in Andrews (1990). The body of the paper is divided into four sections. The first identifies more precisely the sort of formal redundancy that the present account seeks to eliminate. The second section reviews the approaches to this problem proposed in Karttunen (1984) and Andrews (1990). The third presents a paradigm-based alternative and provides illustrative fragments of English verbal morphology and German declensional morphology. The final section concludes by elucidating the basic distinction between the alternatives considered with reference to the GPSG feature system, which represents a simple inheritance hierarchy in terms of a propositional language. 1.
PARADIGMATIC SYNCRETISM
It is useful at the outset to distinguish 'artifactual' syncretism, which reflects an overarticulated grammatical description, from 'fortuitous' or 'accidental' syncretism, which involves simple homophony within a paradigm. The present section discusses examples of artifactual and accidental syncretism, and outlines a general strategy for reducing - and, in some cases, eliminating - the former type. While the classification of individual instances is not always straightforward, these examples illustrate how artifactual syncretism arises through distinguishing properties that are in fact fully or partially neutralized within a grammatical system. The strategy for collapsing such hyperarticulated descriptions involves designating one form as the unmarked member of a lexical paradigm, and recognizing only as many marked alternatives as there are contrastive oppositions in the paradigm. By representing only distinctive oppositions, the resulting description prunes artifactual syncretism. However, this procedure will not collapse arbitrary homophones, as entries that cannot be amalgamated
SYNCRETISM
AND PARADIGMATIC
OPPOSITION
115
through the neutralization of noncontrastive features will remain distinct. This arguably reflects a useful distinction, as the recognition of nondistinctive features is a correctable defect of a linguistic description, whereas homophony is an intrinsic, albeit contingent, property of a grammatical system. Traditional classifications of English verbal paradigms typically involve considerable syncretism. This is clearly illustrated in the proliferation of analyses for walk in the present indicative paradigm in (1), from Curme (1947). (1) 1ST 2ND 3RD
SINGULAR
PLURAL
walk walk walks
walk walk walk
The evident redundancy in this description is due to the fact that it does not accurately represent the distinctive oppositions within the English present indicative paradigm or, for that matter, within any verbal paradigm in English. Rather, the artificial distinction between the non-third person singular forms reflects extrinsic considerations, often, as Huddleston (1984:77) suggests, the fact that these positions contrast in the verbal paradigms of classical languages. As Huddleston (1984) and Quirk et al. (1985), among others, have observed, the distinctive opposition in such regular paradigms is much more perspicuously described in terms of a contrast between a marked third person singular form, and an unmarked, or general, form. 1 Thus the six-way contrast in (1) can be reduced to the more revealing binary paradigm in (2). (2)
3rd Singular Form
walks
General Form
walk
This economical classification identifies the syncretism in (1) as an artifact of a defective description that fails to represent the complete neutralization of person and number distinctions in non-third person singula; forms. Unmarked general forms likewise permit the simplification of systems that contain more than one marked alternative. This can be illustrated 1 Qu irk et al. (1985) describe the m a r k e d form as the -s form.
116
JAMES P. BLEVINS
with reference to the basic three-way contrast in the present indicative paradigm of the English copula. The traditional description in (3) again incorporates various noncontrastive person and number distinctions.
(3) 1ST 2ND 3RD
SINGULAR
PLURAL
am are is
are are are
The artifactual syncretism in (3) can accordingly be eliminated by representing only distinctive oppositions, as in (4). (4)
1st Singular Form
am
3rd Singular Form
is
General Form
are
The concise descriptions in (2) and (4) reveal a dynamic complementarity that is obscured by the proliferation of entries in (1) and (3). In each case, the combinatorial options of marked forms are determined by their agreement specifications, while the distribution of the unmarked form has an 'elsewhere' character. As a consequence of this dynamic opposition, the distribution of general forms may vary across paradigms. 2 The regular present walk is opposed by the dedicated third singular form walks and accordingly combines with the non-natural class of person and number combinations that conflict with walks. The general present form of the copula, are, is likewise compatible with the person and number combinations that conflict with third singular is or first singular am. The preterite walked, in contrast, is unopposed by marked alternatives and hence compatible with all person and number combinations. The simple paradigms in (2) and (4) distinguish fully specified from completely unspecified forms. However, more complex paradigms typically exhibit more intricate patterns involving the partial neutralization of
2 Such variation in the distribution of general forms is what ultimately permits a uniform description of paradigms. This contrasts with the account in Huddleston (1985:85fn8), where the uniform distribution, assigned to general forms demands a correspondingly overarticulated description of the copula, which, Huddleston suggests, "does not have a general present tense form, for in the present tense we have a three-way person-number contrast is (3rd person singular) vs a m (1st person singular) vs a r e (2nd person or plural)".
117
SYNCRETISM AND PARADIGMATIC OPPOSITION
features that are contrastive elsewhere in the paradigm. The nominal declensions of modern German clearly illustrate the rampant syncretism that may arise as a result of such 'local' neutralization. For example, the matrix description of the definite article in (5), from Lederer (1969), assigns a total of sixteen analyses to the six distinct forms of the article.
(5) SINGULAR
NOMINATIVE ACCUSATIVE DATIVE GENITIVE
PLURAL
MASCULINE
NEUTER
FEMININE
MASC/N-EUT/FEM
der den dem des
das alas dem des
die die der der
die die den tier
This description again obscures significant patterns, as not all combinations of person, number and gender specifications define contrastive entries. Thus the distinction between masculine and neuter genders is entirely neutralized in the dative and genitive cases, while feminine and plural forms are identical except in the dative. Nominative and accusative are likewise collapsed in all but the masculine gender, while the feminine dative and genitive also coincide. Yet these limited neutralizations cannot be expressed in the four-way case and three-way gender system in (5). Forms like dem or des are neither uniquely masculine or neuter, nor wholly unmarked for gender, but rather non-feminine. Neuter das is likewise neither distinctively nominative nor accusative, nor unmarked for case, but rather non-oblique. Conversely, feminine der is better described as oblique than as unmarked or as ambiguously dative and genitive. To capture these local neutralizations, we require some means of imposing additional structure on the feature system in (5). The form that this structure takes is to some degree dependent on how the contrast between marked and unmarked entries in (2) and (4) is represented. One strategy, proposed in Karttunen (1984) and developed in much subsequent work, accommodates unmarked entries and local feature neutralization by extending lexical entries to include negative and disjunctive feature specifications. The alternative proposed in Andrews (1990) preserves the unmarked character of general forms by leaving them unspecified for agreement features, and attributing the distributional complementarity with marked forms to a morphological blocking condition. The present account extends this basic strategy to deal with cases of partial neutraliz-
118
JAMES P. BLEVINS
ation by encoding the locus of neutralization in a global feature type hierarchy. 3 The following section contrasts these alternatives after reviewing their use of syntactic features.
2. FEATURE-BASED
A N A L Y S E S OF S Y N C R E T I S M
Nearly all current syntactic frameworks make some use of syntactic features and articulated feature structures. In the simplest case, features are atomic attribute-value pairs, such as [PER or [NUMSg], and feature structures are collections of such pairs. These collections are typically represented as attribute-value matrices (ArMs), as in (6). 4
3Fd]
(6)
3rd] LNUMsg 3 [PER
The array of feature-value pairs in such representations is interpreted conjunctively, so that (6) represents a linguistic object that is both third person and singular in number. A distinguishing property of the family of 'constraint-based' or 'unification-based' formalisms is the use of features with complex values. The treatment of agreement within these formalisms provides a clear illustration of the descriptive utility of complex features. Generalized Phrase Structure Grammar (GPSG; Gazdar et al. 1985) expresses subject agreement requirements in terms of a feature AGR that takes, as in (7), a feature structure as its value. (7)
[AGR[PER 3rd]] LNUM s g
33
Functional Unification Grammar (FUG; Kay 1979) and Lexical Functional Grammar (LFG; Kaplan and Bresnan 1982) likewise incorporate agreement specifications in the feature structure that occurs as the value of a functional SUBJ(ECT) attribute.
sg
LNUM
j•
s As suggested in the introduction, the analysis of German proposed in Section 3.2.1 recognizes a primary distinction between feminine and nonfeminine gender, and a subsidiary distinction between the nonfeminine subcases. Case features are analogously divided into primary oblique and nonoblique (or 'direct') macrotypes which subsume the four basic values in (5). 4 Attributes are uniformly represented in SMALLCAeS and attribute values in italics in AVMs.
SYNCRETISM AND P A R A D I G M A T I C OPPOSITION
119
Current models of Head-Driven Phrase Structure Grammar (HPSG; Pollard and Sag 1994) similarly introduce complex agreement specifications as values of an embedded INDEX feature.
(9)
[SYNSEM ] LOCAL I CONTENT [ INDEX IP?LR ~gd 1 ]
However, since the present paper is concerned more with the logical structure of categories than with the precise inventory of agreement features or their location in the feature geometry of signs, the substantive differences in analysis reflected in (7)-(9) are of less importance than their formal similarities. The illustrative paradigms below accordingly suppress the inessential variation in (7)-(9), and amalgamate agreement features with the categorial features in the SUBCAT lists that represent valence in HPSG. This generic alternative, illustrated in (10), again integrates agreement specifications with subcategorization demands, as in FUG and LFG. 5 (10)
While these simple examples clearly do not determine a theory of agreement, they do show how complex features provide a straightforward analysis of the agreement properties of marked verb forms. The partial feature structures in (7)-(10) each represent the subject agreement specifications of a third person singular verb like walks. The requisite match between these specifications and the agreement features of a subject is then regulated by a principle that UNIFIES the subject's features with the verb's agreement features. 6 For example, the agreement specifications in (7)-(10) all unify with the feature structure in (6), but do not unify with any feature structure that contains alternative values for PER or NUM. Associating these specifications with the entry for walks thus expresses the fact that walks requires a third person singular subject. Complex features also provide a transparent representation of the highly syncretistic paradigms in (1) and (3), as each member of these paradigms can be associated with distinct person and number specifications. How5 The SIJBCATattribute in (10) suppresses the path prefix SYNSEM[ L O C A L I CATEGORY,while the familiar symbol 'NP' is viewed as a mnemonic alias for appropriate feature specifications. 6 Unification is essentially the feature structure analogue of set union. The unification of two feature structures A and B is the smallest feature structure that contains atl of the specifications from A and B, provided that A and B are consistent. If A and B assign distinct values to any feature, unification fails. See Shieber (1986a) for discussion of unification and its role in a variety of constraint-based formalisms.
120
JAMES
V. B L E V I N S
ever, a clear descriptive p r o b l e m arises in connection with the 'disjunctive' agreement requirements of u n m a r k e d forms in the collapsed paradigms in (2) and (4). The various person and n u m b e r combinations associated with walk and are cannot be expressed as the conjunction of atomic person and n u m b e r specifications. The following subsections review two approaches to this problem. The proposals summarized in Section 2.1 extend feature structures (or corresponding constraint languages) to include negation and disjunction operators. In contrast, the account in Section 2.2 maintains the basic conjunctive format for feature structures, while adding a general wellformedness condition on lexical preterminals that allows partial underspecification of general forms. Let us now consider each alternative in turn.
2.1. Logical Extensions of Feature Structures Karttunen (1984) captures the disjunctive specifications associated with the general forms in (2) and (4) by introducing disjunctive and negative feature structures. Karttunen argues that the use of single entries with negative and disjunctive specifications yields m o r e concise and revealing descriptions than the corresponding disjunction of nondistinctive lexical entries. There are many cases, especially in morphology, in which the most natural feature specifications are negative or disjunctive... Although the features "number" and "person" are both clearly needed in English verb morphology, most verbs are very incompletely specified for them. In fact, the present tense paradigm of all regular verbs has just two forms of which one represents the 3rd person singular ("walks") and the other ("walk") is used for all other persons. Thus the most natural characterization for "walk" is that it is not 3rd person singular. The alternative is to say, in effect, that "walk" in the present tense has five different interpretations... The system of articles in German provides many examples that call for disjunctive feature specifications. The article "die", for example, is used in the nominative and accusative cases of feminine singular nouns and all plural nouns. The entry given [in (12)] succinctly encodes exactly this fact. (pp. 24-5) The specific analyses that Karttunen proposes are repeated in (11) and (12) below. The structure in (11) represents the negative agreement specifications assigned to non-third person singular forms like walk.
(11)
[AGR [ PER3rail /NUM sg J]
Negative specifications are interpreted as constraints on unification; a specification of the form -qA blocks unification with the feature structure
SYNCRETISM AND PARADIGMATIC
OPPOSITION
121
A (and any feature structure subsumed by A). 7 The agreement specifications in (11) thus bar unification with any third person singular feature structure, and permit unification with any other person-number combination. As desired, this expresses the disjunctive subject agreement requirements associated with walk. The disjunctive inflectional specifications associated with the German article die are likewise amalgamated in (12).
(12)
]
L[Nu pq
J
LCASE {nom acc} Disjunctive specifications of the form {A B} unify with a structure C just in case C unifies (ultimately) either with A or with B. 8 Hence the INFL specifications in (12) unify with structures that are compatible with one of the disjunctive values of the complex AGR feature and one of the disjunctive values of the atomic CASE feature. In the passage cited above, Karttunen identifies the reduction of syncretism within morphological paradigms as the primary empirical motivation for these extensions. The agreement specifications of the general form walk in (11) are the negation of the specifications assigned to the marked paradigmatic opponent walks, while the disjunctive entry for die collapses the four elements in (5) that contrast with the specifications for the competing articles der, das, dem, des and den. This implicit morphological opposition is somewhat more explicitly represented in the hierarchical lexicon proposed in Pollard and Sag (1987:209ff) as a contrast between lexical rule-mediated relations. The fragment in (13) illustrates the general structure in which "regular inflected forms are computed (deduced) from a single base form".
7 Subsumption imposes a partial 'informativeness' order on feature structures, corresponding to the inclusion relation on sets. See again Shieber (1986a) for discussion. s The qualification here is required by the fact that disjunctions, as Karttunen notes, often cannot be directly unified, but must be maintained until one of the disjuncts is excluded as a consequence of later unification. Karttunen (1984:28-9) proposes that the unification of a structure C with a disjunction {A B} introduces a 'positive constraint' on A and B requiring that "[a]ll later unifications involving them must keep at least one of the two pairs (A,C), (B,C) unifiable. If at some later point one of the two tuples becomes inconsistent, the members of the sole remaining tuple finally can and should be unified. When that has happened, the positive constraint on A and B can be discarded".
122
JAMES e . BLEVINS
ialki 1tl 3rd
SUBCAT NP
LNUM sg • I ]
-PHON walk] 'b~~/ ~ SUBCAT (NP}/ VFORM base J
