wordandparadigmmorphology - university of...

30
Word and Paradigm Morphology James P. Blevins University of Cambridge Farrell Ackerman UC San Diego Robert Malouf San Diego State University To appear in J. Audring & F. Masini (eds.), e Oxford Handbook of Morphological eory. Oxford: OUP, Abstract e -odd years since the publication of Matthews’ () Inflectional Mor- phology have witnessed a broad resurgence of interest in word-based approaches and a particular rehabilitation of classical ‘word and paradigm’ (WP) approaches as general models of analysis. WP models have shown themselves well adapted to the description and analysis of inflectional patterns, particularly those in- volving inflection class morphology. Contemporary WP approaches have clarified the formal structure of classical models, while isolating assumptions that reflect their longstanding use in pedagogical or reference materials. is tradition has helped to clarify the degree to which the conception of morpho- logical analysis in WP models is fundamentally implicational: the central role of words (and paradigms) reflects their predictive value in a morpho- logical system. An understanding of the nature of morphological organiza- tion, within and across languages, requires exploration of the fundamental el- ements of implicational relations. From a descriptive perspective this involves identifying the internal structure of words and the ways that this structure fa- cilitates an external organization into patterns of relatedness. From a theoret- ical perspective, it is necessary to identify analytic tools that are appropriate for specifying and quantifying word-internal and word-external organization. To validate these perspectives, it is also important to investigate the types of learning theories that may play a role in determining the types of organization that are found to occur and thereby help to explain their learnability. Introduction e common opposition between ‘word’ and ‘morpheme’ tends to obscure, rather than elucidate, what appear to be two distinct approaches to the study of morphol-

Upload: truongkhue

Post on 17-Sep-2018

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Word and ParadigmMorphology

James P. BlevinsUniversity of Cambridge

Farrell AckermanUC San Diego

Robert MaloufSan Diego State University

To appear in J. Audring & F. Masini (eds.), The Oxford Handbook ofMorphological Theory. Oxford: OUP,

AbstractThe -odd years since the publication of Matthews’ () Inflectional Mor-phologyhavewitnessed a broad resurgence of interest inword-based approachesand a particular rehabilitation of classical ‘word andparadigm’ (WP) approachesas generalmodels of analysis. WPmodels have shown themselveswell adaptedto the description and analysis of inflectional patterns, particularly those in-volving inflection class morphology. Contemporary WP approaches haveclarified the formal structure of classical models, while isolating assumptionsthat reflect their longstanding use in pedagogical or reference materials. Thistradition has helped to clarify the degree to which the conception of morpho-logical analysis in WP models is fundamentally implicational: the centralrole of words (and paradigms) reflects their predictive value in a morpho-logical system. An understanding of the nature of morphological organiza-tion, within and across languages, requires exploration of the fundamental el-ements of implicational relations. From a descriptive perspective this involvesidentifying the internal structure of words and the ways that this structure fa-cilitates an external organization into patterns of relatedness. From a theoret-ical perspective, it is necessary to identify analytic tools that are appropriatefor specifying and quantifying word-internal andword-external organization.To validate these perspectives, it is also important to investigate the types oflearning theories that may play a role in determining the types of organizationthat are found to occur and thereby help to explain their learnability.

Introduction

The common opposition between ‘word’ and ‘morpheme’ tends to obscure, ratherthan elucidate, what appear to be two distinct approaches to the study of morphol-

Page 2: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

ogy. While both examine the ‘structure’ of words, they do so from very differentperspectives, assuming very different objects of analysis and arriving at very dif-ferent results. The ‘word-based’ tradition regards morphology as the “branch oflinguistics which is concerned with the ‘forms of words’ in different uses and con-structions” (Matthews : ): the emphasis is on the ways that the internal struc-tures provide whole words with distinctive shapes and how these faciliate the de-velopment of patterns of relations between them. The ‘sub-word’ tradition treatsmorphology in terms of the internal structure of words and the nature of the pro-cesses responsible for their disassembly and reassembly. The two basic approachesfall within parallel lineages that derive from different branches of ancient Indo-European grammatical traditions, as schematized in Figure .

Sanskrit Grammarians

Altindische Grammatik

‘American Structuralism’

Morphemic analysis inlexicalist and non-lexicalist

frameworks

Graeco-Roman Tradition

Neoclassical WP

Realizational WP

Neogrammarians

‘European Structuralism’

Information-theoretic andSimulation Based

Modelling

Figure : Morphological lineages

Morphology from a Graeco-Roman, Word and Pattern, perspective can be con-strued as the study of part-whole organization, with respect simultaneously to theinternal and external organization of words. What elements occur as parts of in-dividual words, and how are words themselves organized into larger patterns inmorphological systems? This suggests a division of morphology into the study ofsyntagmatics, which explores the internal organization of words, and paradigmat-ics, which explores the external organization of relations between words. The rolethat syntagmatic variation plays in determining a higher-level paradigmatic classi-fication within a classical WP model is cogently expressed in Matthews (: ):

Page 3: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

In the ancient model the primary insight is not that words can be splitinto roots and formatives, but that they can be located in paradigms.They are not wholes composed of simple parts, but are themselves theparts within a complex whole. In that way, we discover different kindsof relation, and, perhaps, a different kind of simplicity.

. The locus of morphological variation

Much of the variation exhibited by morphological systems is manifested by seg-mental or suprasegmental contrasts within individual words. Yet detailed analysesof morphological systems, inflectional and derivational, reveal principles of organi-zation that apply to more than just linear (or hierarchical) arrangements of ‘recur-rent partials’. The analysis of a system also requires the recognition of word-sized(or larger) units as basic objects of morphological inquiry. In some cases, theseunits are synthetic formations, comprised of single contiguous sequences. In othercases, morphological units are periphrastic constructions consisting ofmultiple freeforms. It is only by expanding the scope of morphology to include these kinds oflarger units that the principles underlying patterns of relatedness can be discovered.

Given the general applicability of a part-whole approach to internal and exter-nal dimensions of morphological organization, it is reasonable to ask why descrip-tivist approaches initially neglected the paradigmatic dimension, and what featuresofmorphological systemsmight have favoured the persistence of this bias in the sub-sequent generative tradition.¹ Intriguingly, the challenges faced by a purely syntag-matic approach would have been familiar to the descriptivists from their schoolingin Latin and Greek. However, the Post-Bloomfieldian goal of developing a ‘scienceof linguistics’ came to involve not only a rejection of classical models but also a ne-glect of the patterns that had motivated these models. By the time that Matthews(: ) remindsmodern readers of the empirical andmethodological challengesthat arise in decomposing words in classical Greek and Latin, it is the languages thatappear anomalous, not the a priori assumption of a basic agglutinative structure.

Many linguists tend to boggle at such systems. They seem complicated,while agglutinating systems seem so simple. They may even seem per-verse. Why should a language have rules which obscure the identityand functions of its minimal elements?

From the perspective of WP models, the examination of part-whole relations inthese “perverse” systems is particularly instructive, since they reveal the underlying

¹The historical origins of the syntagmatic model are discussed in Hockett (), Matthews ()and Blevins ().

Page 4: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

commonality between “complicated” and “simple” systems. These common fea-tures provide a stable basis for models that combine broad empirical coverage withcognitive plausibility. As illustrated below, fundamental principles of part-wholeorganization not only scale to apply to languages of varying complexity, but alsoclarify that the appearance of ‘obscurity’ is a symptom of a limiting idealization.

. Models and classification

It is customary, following the tradition initiated by Hockett (), to regardmodelsas ‘word-based’ or ‘morpheme-based’. Yet this unit-based classification is mislead-ing: it is ultimately the role and status of words within competing types of analysisthat is essential. For present purposes, it is important to distinguish two dimensionsof wordhood. Themorphosyntactic dimension concerns the ‘grammatical mean-ing’ or ‘features’ associated with words. The morphotactic dimension concernsthe formal constitution of words. One class of approaches is ‘word-based’ in thesense that they treat the features associated with paradigm cells or ‘morphosyntac-tic representations’ associated with word-sized units as basic ‘units of meaning’ ina grammatical system. A second class of word-based approaches treats wordformsas morphotactically basic and regards roots, stems and exponents as abstractionsthat serve to discriminate the distinctive shapes and patterns of word types. Thediscriminative function of form variation represents a central difference betweenmorphotactically word-based approaches and accounts that treat sub-word units asbasic. In models that are word-based, morphotactically and morphosyntactically,words function as a primary locus of grammatical meaning, as a maximally stableunit of form and as basic units of larger paradigmatic structures. In less consistentlyword-based accounts, wordforms and paradigmatic structures tend to be viewed asderivative or ‘epiphenomenal’. However, this assessment largely makes a virtue ofnecessity, given the difficulties that arise in reconstructing wordforms and paradig-matic systems from inventories of genuinely independent sub-word units.

Although a distinction between word-based and ‘morpheme-based’ models im-poses a fairly crude split, it suffices to classify the main contemporary approachesto morphological analysis. A first cut can contrast a class of word-based approacheswith accounts that only recognize morphemes as morphological units of analysis.The most consistently word-based approaches are classical ‘word and paradigm’(WP) models of the type familiar from the description of classical languages. Inthese models, the word is the smallest grammatically meaningful unit and surfacewords are also the basic elements of form, since there is no unit recognized be-tween words and sounds. The realizational WP tradition that grew out of the workof Matthews (, ) is somewhat less uniformly word-based. This traditionincludes, among other approaches, the Extended WP model (Anderson )/A-

Page 5: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Morphous Morphology (Anderson ), Paradigm Function Morphology (Stump; Bonami and Stump ), Network Morphology (Corbett and Fraser ;Brown and Hippisley ) and the family of allied realization-based and lexeme-based approaches (Zwicky ; Aronoff ; Beard ). These approaches allpreserve the morphosyntactic assumption that (grammatical) words are the pri-mary locus of grammatical meaning. But their morphotactic assumptions alignmore with those of morphemic models, in which surface wordforms are assembledfrom more basic elements. Hence, they can be classified as morphosyntacticallyword-based but morphotactically constructive (Blevins ), in that they as-sume a model of the lexicon in which open-class items are represented by stems(and/or roots), and surface wordforms are constructed by applying rules that suc-cessively modify lexical stems. The focus of these models is on building individ-ual words, not identifying patterns that distinguish or relate words to one another.In modern realizational approaches, analysis is an interpretive process, in which‘bundles’ of distinctive features are ‘spelled out’ by realization rules that map fea-tures onto forms. The organization of these approaches is elaborated in detail in theworks cited above as well as in Montermini (this volume) and Stump (this volume).In these models, the paradigm cells or morphosyntactic representations that repre-sent the properties of word are basic, but the wordforms that realize those propertiesare not only derived but also non-persistent. In no realizational model is the outputof all possible realization processed cached in any kind of full-word lexicon.

The differentmorphotactic assumptions adopted in classical andmodern word-basedmodels yield alternative strategies for expressing content-form relations. Clas-sical models represent these relations by sets of exemplary patterns, together with(proportional) analogies that extend patterns to new items. Internal structure isimportant not because it provides inventories of pieces useful for deriving words,but because it exhibits the dimensions of variation that define the organization ofmorphological systems. The remainder of this chapter clarifies how word-internalvariation determines the paradigmatic structure of WP models, in their classicaland pedagogical formulations, as well as in more recently formalized variants.

Words and paradigms

Reflecting the pedagogical origins of classical models, the pivotal role of words andparadigms in these models derives from considerations of morphosyntactic andmorphotactic stability and predictive reliability. As units of form and exponentsof grammatical properties, word-sized units are more stable and consistent than the

Page 6: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

sub-word units that are isolated in models of morphemic analysis.² The closed anduniform feature space defined by inflectional paradigms likewise sanctions highlyreliable deductions about the existence and shape of other words in a paradigm. Thestable grammatical information associated with a word not only serves to identifyits own meaning and function, but also locates it within an inflectional paradigm orderivational ‘morphological family’ as well as within the larger morphological sys-tem. In this way, the information associated with a wordform facilitates deductionsabout other forms, based on systematic interdependencies in a language.

Aspects of form that sanction implications express a type of information thatcan be modelled by notions developed within Information Theory (Shannon ).In the case of an inflectional paradigm, entropy or uncertainty corresponds to theformvariation exhibited by paradigm cells. The informativeness of a given form cor-relates with the degree to which knowledge of that form reduces uncertainty aboutthe shape of other words within the same paradigm. A notion of uncertainty reduc-tion is clearly implicit in the traditional use of ‘diagnostic’ principal parts, wherewordforms or sets of wordforms are identified as reliable guides to the shapes ofother wordforms. The utility of principle parts rests on their value in reducing re-ducing uncertainty about other forms. In the idealized pedagogical case, principalparts would reduce all uncertainty, though the complete elimination of uncertaintyis neither feasible nor necessary in actual language learning or use.

The split between ‘predictive’ principal parts and ‘predicted’ forms also reflectsthe pedagogical goals of a classical grammar. Nearly all wordforms are informa-tive to some degree about the shape of other wordforms. The more closely related,the more mutually informative forms tend to be, ranging from members of an in-flectional paradigm, through inflection classes and derivational paradigms, outwardtomorphological systems and language families. Measures of interpredictability arelargely restricted to word-basedmodels, given that predictability cannot be reducedto ‘derivability’ or to any other relation defined over sub-word units by morphemicor other types constructive approaches. These measures receive a measure of exter-nal validation from their contribution, in conjunction with discriminative learningmodels, to explanations of the learnability of complex morphological systems.

. Two models of grammatical description

In what, following Hockett (), have been come to be known as Item and Ar-rangement (IA) models, the task of morphological analysis involves identifying theconstitutive pieces of complex words and the nature of the reassembly processes thatreconstitute the whole from its parts. In Hockett’s terms “morphology includes the

²Traditional philological arguments for the stability and informativeness of words are reinforcedby compression-based measures of the locus of regularities within a language (Geertzen et al. ).

Page 7: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

stock of segmental morphemes, and the ways in which words are built out of them”(Hockett : ). Words, accordingly, are concatenations of individually mean-ingful sub-word units, whose distribution is regulated by rules specifying their lin-ear and hierarchical arrangement (with ‘morphophonemic’ contextual adjustmentsas needed). Yet as observed in Jackendoff () andBochner (), any gains in de-scriptive economy achieved by morphological decomposition are undermined by,among other factors, the redundancy between the rules proposed to build wordsand the limitations of these rules in constructing various types of complex words.³

In a second variant of morphemic analysis, termed the ‘Item and Process’ (IP)model by Hockett (), the meanings of words are not associated directly withunits of form, but instead with the processes that apply to build complex wordsfrom lexical stems. As with IA approaches, IP approaches are centrally concernedwith building words from simpler components – in this case stems and processes –guided by the goal of reducing or avoiding redundancy in linguistic representations.Surface words and their organization into paradigms again have no direct status ina morphology system and questions concerning their relevance do not even arise.

Both models are concerned exclusively with the assembly and disassembly ofindividual wordforms in isolation, so relations between words fall outside theirdescriptive scope. The word is not regarded as a primary object of morphologicalanalysis but functions instead as the ‘output’ of processes that assemble forms fromtheir constitutive pieces or as the ‘input’ to processes which disassemble forms intotheir ultimate constituents. From this perspective, words are only of interest to theextent that they provide clues for identifyingminimal elements and operations. Theidealization that individual words can be analyzed wholly in isolation precludes theinvestigation of relations between words or principles that organize words into net-works of related forms. Given these severe limitations, words and paradigms canonly occupy an ‘epiphenomenal’ status within models of morphemic analysis.

On these ‘morpheme-based’ conceptions, morphological analysis encompassesa limited stock of research questions. Analysis involves the search for minimum in-dividually meaningful units of form, along with the batteries of rules that determinetheir shape and distribution in larger expressions. Following Bloomfield (: ),this enterprise is guided by a priori notions of “scientific compactness” and avoid-ance of segmental redundancy. Among the questions that fall outside the scopeof this perspective are those that concern relations between surface words and theorganization of words into larger collections within a grammatical system.

³See Jackendoff and Audring (this volume) for further discussion of these limitations, and Jack-endoff (), Aronoff () and Bochner () for metrics that measure the cost of morphologicalpatterns rather than the degree of (presumed) segmental redundancy. This perspective will be brieflyexplored in Section below in relation to discriminative Information-Theoretic WP models.

Page 8: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

. In defense of WP

The limitations ofmorphemicmodels were already clear by the publication ofHock-ett (), and Hockett () provides a first-hand account of the origin, develop-ment and retrenchment of these models. Although Hockett was familiar with theclassical tradition (and responsible for the moniker ‘Word and Paradigm’), Robins()was the first to take up the challenge implicit inHockett’s (: ) assertionthat “WP deserves the same consideration here given to IP and IA”. Recognizingthat “[t]he discussion can be carried no further until the word-and-paradigm ap-proach has been characterized at least as clearly as current versions of morphemics”,Matthews (: ) adddressed the task of formalizing the classicalWPmodel. Al-though many of the technical details of this initial model have since been revised,the conception of analysis outlined in Matthews () remains largely intact:

The word is its central unit, and the grammatical words (the VocativeSingular of brutus, for example) are minimal elements in the study ofsyntax. At the same time, the intersecting categories form the frame-work ormatrix withinwhich the paradigm of a lexeme can be set out. Ifa schoolboy is asked to recite the paradigm of mensa or the paradigmof amo he will deliver the sets of word-forms (mensa, mensa, mensam…; amo, amas, amat …) in an order which explicitly or implicitly ad-dresses their assignment to the individual Cases, Persons, and so on.

Although developed to describe the complex patterns of exponence in classicallanguages, the motivation for a paradigmatic treatment of inflectional patterns isoften compelling in much simpler systems. In an early generative sketch of Germanmorphology, roughly contemporary with Matthews (), Chomsky (: )briefly summarizes the complications that arise in representing declensional pat-terns in terms of discrete case, number and gender morphemes, and concludes:

I know of no compensating advantage for the modern descriptive re-analysis of traditional paradigmatic formulations in terms ofmorphemesequences. This seems, therefore to be an ill-advised theoretical inno-vation.

Two of the problems that Chomsky identifies with morpheme-based models con-cern (i) the reliance on unrealized (‘null’) morphological expressions, and (ii) theneed to impose a fixed order on ‘sequences’ of realized and unrealized elements.

These issues are illustrated by the contrasts exhibited by the forms of the Mord-vin noun kal ‘fish’ in Table . There are no overt markers for nominative or singu-lar in the Mordvin indefinite declension. Consequently, within a strict morphemicmodel, one zero marker would be required for nominative case and another for

Page 9: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Indef Sg Indef Plu Def Sg Def Plu

Nominative kal kal-t kal-os kal-t-neElative kal-sto kal-sto-nt kal-t-ne-ste

Table : Partial paradigm of Mordvin kal ‘fish’ (Rueter : ff.)

singular number, and an order would need to be imposed on these elements. Inprinciple, the two zero markers could occur in either order. The nominative pluraldefinite form kal-t-ne identifies only one order of plural markers, with the basic plu-ral marker -t preceding the definite plural marker -ne. However, the elative formsindicate that this order is not invariant, and that there is also no invariant order be-tween case and number markers. Both of the possible orders are possible, with vari-ation conditioned by number. In the singular, the definite marker -nt follows thecase marker -sto; in the plural, both the basic plural marker t and the definite pluralmarker ne precede the case marker ste. Hence even in the case of overt markers,it is not possible to establish an invariant order. Introducing ‘unrealized’ markerscan only make this recalcitrant problem even more intractable. The postulation ofabstract hierarchical structure likewise expands the space of unobservable variation.

As reflected in Chomsky’s characterization of morphemic analysis as “an ill-advised theoretical innovation”, the challenges presented by morpheme orderingare self-inflicted problems. Like the other classes of problems discussed from theoutset of the morphemic tradition Harris (); Hockett (), these problemslack analogues in other morphological models, and have no clear relevance to mor-phological acquisition or use. Speakers must be able to discriminate the forms of asystem and determine their function, within the system. But there is no establishedneed to impose a consistent ordering on the contrasts abstracted from those forms,let alone extend an ordering to ‘unrealized’ markers. WPmodels do not offer a solu-tion to the problem of morpheme ordering; instead they impose analyses that avoidcreating this kind of problem in the first place.

In the ErzyaMordvin patterns in Table , which obtain across all cases, the al-ternation does not depend on specific case values. An instructive contrast is foundin Udmurt (Kel’makov and Hännikäinen ), in which order is sensitive to thechoice of markers. Figure sets out the basic sets of markers and associated mor-photactic patterns. Examples of each pattern are presented in (). In each pattern astem precedes a possessive marker and a member of a specified set of case markers.

This type of data, together with recent research on variable affix orders (Luuto-nen ; Bickel et al. ; Plag and Baayen ), casts substantial doubt on thegeneral applicability of approaches that describe the organization of morphological

Page 10: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

systems in terms of arrangements of minimum meaningful elements. Insofar as or-der is salient, it will tend to serve a discriminative function in distinguishing wordsor larger units that play a communicative role in a language. However, this functiondoes not depend on basic (let alone universal) orderings of inflectional markers orother formatives. As with the challenges that confound attempts to impose orderon ‘unrealized’ elements, the difficulties that confront an order-based model ariseimmediately and point to a basic deficiency in the underlying conception of mor-phological analysis.

C1: Abessive/Caritive, Ablative, Adverbial, Approximative, Dative, GenitiveC2: Egressive, Elative, Inesssive, Illative, Instrumental, Prolative, TranslativeC3 Terminative

P1: Stem–Person/Number Marker–CP2: Stem–C–Person/Number MarkerP3: Stem–Person/Number Marker–C or Stem–C–Person/Number Marker

Table : Case sets and patterns in Udmurt

() Variable marker order in Udmurta. P

pi-ed-lịboy-2sg.px-dat

‘to your boy’b. P

pi-en-idboy-inst-2sg.px

‘with your boy’c. P (where the Person/Number Marker is Sg)

busi-jed-ozfield-2sg.px-term

∼∼

busi-ioz-adfield-term-2sg.px

‘up to your field’

To some extent, discussions of element ordering underestimate the scale of thechallenge, by presupposing the availability of general solutions to the problem ofisolating minimum ‘units of form’ and specifying minimum ‘units of meaning’. Thesame is true of debates about the status of ‘biunique’ form-meaning relations. In

Page 11: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

both cases, the Segmentation Problem (Lounsbury ; Spencer ) must be ad-dressed in order for the arrangement and interpretation of segments can be deter-mined. The atomization of ‘meaning’ raises analogous challenges, though these tendto be obscured by the use of disparate classes of ‘features’ as proxies for meaning. Inmany cases, these challenges can be met or at least attenuated by shifting the focusto larger units. To rephrase Robins () and Matthews () in more contem-porary terms, identifying the shape and specifying the grammatical meaning of aninflected word involves less uncertainty in general than identifying the shape andspecifying the meaning of the component formatives. As has been observed sinceat least Bloomfield (), an analysis in terms of word-sized units does not elimi-nate uncertainty. Nevertheless, a variety of factors support the traditional view thatwords enjoy a stronger claim to psychological reality than their component parts.⁴

. Periphrastic expression

Word-based models exhibit two principal divergences between basic morphosyn-tactic and morphotactic categories. As mentioned in Section ., a model may re-gard words (corresponding to paradigm cells or morphosyntactic representations)as grammatically basic but treat stems and formatives as morphotactically basic.Conversely, a model may again treat grammatical words as basic but recognizemor-photactic units that consist of multiple free forms. The ‘compound tense’ analysisof periphrastic verb constructions provides one familiar case (Curme , ).Nominal case and number inflection in Tundra Nenets illustrates a further pattern.

Sg Dual PluNominative ti texɘh tiqAccusative tim texɘh tiGenitive tih texɘh tiqDative tenɘh texɘh n’ah texɘqLocative texɘna texɘh n’ana texɘqnaAblative texɘd texɘh n’ad texɘtProlative tewna texɘh n’amna teqm

Figure : Synthetic and periphrastic exponence in Tundra Nenets

In Tundra Nenets, the three ‘grammatical’ cases: nominative, accusative and

⁴See Blevins (: §) and Geertzen et al. () for recent summaries of some relevant evidence,Dixon and Aikhenvald () and Haspelmath () for more skeptical perspectives, and Arkadievand Klamer (this volume) for a typological overview of the notion of the word.

Page 12: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

genitive, are normally distinguished from the remaining four ‘local’ cases. A prop-erty common to grammatical cases is that they are realized synthetically in all threenumbers, whereas local cases exhibit number-sensitive variation. Singular and plu-ral forms of local cases are realized synthetically. However, local dual forms consistof an invariant dual form and an appropriately case-inflected postposition. As Ack-erman and Stump () observe, the variation between synthetic and periphrasticrealizations in Tundra parallels a pattern observed in languages with only syntheticrealizations. Specifically, themorphosyntacticmarkedness of case andnumber com-binations tends to be reflected by the formal markedness of their synthetic encod-ings. A similar correspondence obtains inTundraNenets nominal paradigms, wherethemostmarked feature combination, i.e., local case and dual number, is realized bythe most marked realization, i.e., periphrasis. Insofar as these types of generaliza-tion are not reducible to independent factors (such as time of grammaticalization),they reinforce the functional parallels between synthetic and periphrastic forma-tions in morphological systems, and support a view of morphological analysis thatincludes synthetic and periphrastic modes of expression. This conception leads inturn to a reconsideration of the demarcation of morphology and syntax, echoingthe deliberations of Matthews (: ) at the outset of the WP revival.

The history of morphology since the s has led to a progressivelycomplex and non-patent relationship between the elements of gram-mar, on the one hand, and their phonological realization on the other:Is there any reason why the domain in which an element may be real-ized should be kept within traditional limits?

Word and Paradigm models are usefully agnostic on this point. WP models donot require that paradigm cells (or morphosyntactic representations) must be ex-pressed by single synthetic expressions. As in traditional descriptions, multi-wordmorphological expressions can be associated with paradigm cells and integratedwithin a patternedmorphological organization that includes synthetic expressions.⁵

⁵Compelling evidence of the centrality of paradigmatic relations to an understanding themorpho-logical character of periphrastic expressions can be found in theAlgonquian studies ofGoddard ()and LeSourd () and in the cross-linguistic account of Bonami (). A considerable body of re-cent work identifies criteria for classifying particular multi-word expressions as morphological, anddistinguishing them from syntactic phrases. See Ackerman (); Ackerman and LeSourd ();Börjars et al. (); Ackerman and Webelhuth (); Ackerman and Stump (); Brown et al.(); Bonami (), among others. Parallel arguments for morphological periphrasis in the deriva-tional domain are given in Ackerman (); Booij (); Masini (); Masini and Benigni ().

Page 13: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

. Parts and wholes

From a Word and Paradigm perspective, there is no reason to expect morphosyn-tactic and derivational properties to be expressed by units of the same granularity inall languages. The functional load of a morphological system may, in principle, bedistributed across units of varying sizes in different languages, depending on otherproperties of the languages. In the case of the familiar (mainly Indo-European)languages that have had the greatest influence on contemporary models, it is gen-erally acknowledged that words are more useful than sub-word units for practicaldescriptions, including teaching grammars, dictionaries and reference grammars.Even Bloomfield (: ) concedes that “[f]or the purposes of ordinary life, theword is the smallest unit of speech”. For proponents of traditional WP models, thepractical benefits of word-based descriptions also carry over to the use of words for“the systematic study of language”(Bloomfield : ). This reflects the belief thatthe same properties that make words useful for practical descriptions, notably a sta-ble relation between forms and morphological classes and grammatical meaning,are of equal value to theoretical accounts. The word is, as Robins (: ) em-phasizes, “a grammatical abstraction” from the speech stream, but it is a maximallyuseful abstraction, whether for abstract analysis or for more practical purposes.

The word is a more stable and solid focus of grammatical relationsthan the component morpheme by itself. Put another way, grammat-ical statements are abstractions, but they aremore profitably abstractedfromwords aswholes than from individualmorphemes. (Robins : )

From the perspective of aWPmodel, nothing precludes the possibility that sim-ple units of grammatical meaning might stand in correspondence to minimal unitsof form in some languages. An agglutinating pattern could arise in a languagewhosemorphotactic organization preserved the structure determined by successive wavesof morphologization. However, there is no motivation within the WP traditionto regard this pattern as normative, given that tradition was initially developed todescribe languages in which “categories and formatives are in nothing like a one-to-one relation” (Matthews : ). For a familiar example, we can turn to theexemplary Classical Greek verb elelýkete ‘you had unfastened’ in Figure below.The full wordform elelýkete stands in a biunique relation to the second person pastperfective indicative active cell in the paradigm of lyo ‘unfasten’. But as Matthews(: ) observes, the realization of aspect and voice in Classical Greek verbs con-founds any attempt to establish biunique property-formative relations.

Page 14: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

e le lý k e teperf past ind 2nd plu

past active

Figure : Morphological analysis of Greek elelýkete (Matthews : )

. Gestalt exponence

Even more acute difficulties arise in associating properties with formatives in casesof what has been termed ‘gestalt exponence’ in Ackerman et al. (). The obser-vations that the properties of words are not in general reducible to the properties oftheir parts applies not only to the properties of individual words but also to relationsbetween words. Just as the grammatical meaning of a word cannot always be bro-ken down into discrete units of meaning that are assigned to sub-word formatives,relations between the elements of a paradigm cannot invariably be reconstructedfrom relations between their parts. Irreducibly word-level properties bring out abasic asymmetry between wholes and parts. The formatives that make up a wordmay uniquely identify the place that the word occupies in its inflectional paradigmor, more broadly, within the morphological system of a language. But if the sameformatives occur in different combinations in other forms, as is often the case, it isnot possible to associate discrete meanings with individual formatives. Much thesame is true of implicational relations. In most cases where a stem or exponent is ofpredictive value, the value is preserved by a word containing the stem or exponent.But in cases where the predictive value of a word is keyed to the absence of an ele-ment or to distinctive combinations of elements, the predictive value of the whole islost when it is disassembled into parts. This is because combinations are distinctiveand the meaning of a whole word is often more than the meanings of its parts.

A simple example will help to clarify how combinations of elements can havedistinctive meanings and predictive values within a language. The first four datacolumns in Table contain the singular grammatical case forms of a group of nounsthat exhibit productive ‘weakening’ gradation in Estonian (in which nominativeforms are based on a ‘strong’ stem, marked here by the double consonant -kk, con-trastingwith genitive forms based on a ‘weak’ stem in -k). Each rowof forms exhibitstwo dimensions of variation: the choice of a strong or weak stem and the presenceor absence of one of the lexically-specified ‘theme vowels’ a, e, i and u.

Now consider the locus of the property ‘partitive singular’. The partitive singu-lars of this class contain two ‘recurrent partials’: a strong stem and a theme vowel.Thus sukka can be analyzed as sukk + a, kukke as kukk + e, pukki as pukk + i andlukku as lukk + u. But partitive case cannot be associated either with strong stems or

Page 15: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Nominative sukk kukk pukk lukk luguGenitive suka kuke puki luku looPartitive sukka kukke pukki lukku luguIllative 2 sukka kukke pukki lukku lukku

‘stocking’ ‘rooster’ ‘trestle’ ‘lock’ ‘tale’

Table : Singular nouns in Estonian (Erelt et al. ; Blevins )

with theme vowels in isolation. The strong stems sukk, kukk, pukk and lukk cannotbe analyzed as partitive, because these same stems realize the nominative singularwhen they occur without a theme vowel, and also underlie the second ‘short’ illa-tive singular forms. Partitive case also cannot be associated with the theme vowels,because the same vowels occur in the genitive and illative singular forms.

Hence partitive case is an irreducibly word-level feature that is realized by thecombination of a strong stem and a theme vowel. This type of ‘gestalt’ or ‘construc-tional’ exponence (Booij ) is difficult to describe if stems and theme vowels arerepresented separately. Because the grammatical meanings associated with strongstems and theme vowels are context-dependent, these elements cannot be assigneddiscrete meanings that ‘add up’ to partitive singular when they are combined. Froma traditional perspective, this context-dependence underscores the difference be-tween ‘analyzability’ with respect to word internal structure and claims about themorphemic status of word internal structure. An individual wordform is often an-alyzable into parts that recur elsewhere in its inflectional paradigm or in the mor-phological system at large. But these parts may function solely to differentiate largerforms, so that the minimal parts that distinguish a pair of wordforms cannot be as-sociated with the difference in grammatical meaning between these wordforms. Toreturn to the patterns in Table , the theme vowel -u distinguishes the partitive sin-gular lukku ‘lock’ from the nominative singular lukk. In isolation, however, thevowel -u neither realizes a specific case value nor expresses ‘the grammatical dif-ference’ between nominative and partitive. Exactly the same is true of the gradecontrast between partitive singular lukku and genitive singular luku.

Furthermore, the implicational value of a form often depends essentially on theparadigm cell – or, more generally, the grammatical properties – that it realizes. Astrong partitive singular such as lukku identifies lukk ‘lock’ as a first declensionnoun and permits the deduction of the other forms in its paradigm (Blevins ).However, in isolation the strong vowel-final lukku provides limited information,because in the paradigm of a noun such as lugu ‘tale’ in Table it may only realizethe short illative singular and thus be dissociated from the other forms of the noun.

Page 16: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

These examples illustrate some of the fundamental limitations of stems quaforms. While it is often possible to identify stems from the wordforms that theyrealize or underlie, the separation of stems from exponents raises recalcitrant prob-lems inmany languages. As Spencer () notes, this is a familiar issue in Romancelanguages, where motivating a particular segmentation among several alternativesis a perennial problem. From a WP perspective, this type of challenge is an artifactof a flawed method of analysis. Determinate segmentations can arise and remainstable within specific languages. However, from the standpoint of an implicationalWP model, it is unsurprising that different, and possibly overlapping, sequencesmay be of value in predicting different patterns. A stem and theme vowel may beuseful for identifying the lexical class of an item, whereas the vowel and inflection –or even a stem consonant, vowel and inflection – may predict patterns of inflection.

This gestalt-based conception of word structure is implicit in the rules and cor-respondences ofclassical grammars, and ismore explicit in the (proportional) analo-gies developed by Neogrammarians such as Paul (). As remarked by Mor-purgo Davies (: f.), much of the initial appeal of analogy derived preciselyfrom the fact that “it offered an algorithm for a structurally based form of morpho-logical segmentation, without making any claims about the segments in question”.

The ‘item and pattern’ model

As themoniker ‘word and paradigm’ suggests,WP approaches assign a special statusto words, and attach grammatical significance to inflectional paradigms and othercollections of words. However, this designation is somewhat misleading, modelshould in fact have been ‘item and pattern’, where comparison of the item againstthe pattern sanctions the deduction of new forms. Reclassifying traditional WP ap-proaches as specific instantiations of a general ‘item and pattern’ model is of morethan purely historical interest. This characterization highlights the fact that themodel is defined less by the units it recognizes than by the relations it establishesbetween units. Instead of disassembling a language into inventories of ‘atoms’ thatcan be combined to build larger units,WP analyses focus on the implicational struc-ture defined over networks of interrelated elements. The privileged status of wordsin these models does not rest on claims about their epistemological priority, or theirplace within procedures of classification or methods of analysis. Instead, the statusof words is due to their relative informativeness, as reflected in Robins’ claim abovethat “the word is a more stable and solid focus of grammatical relations than thecomponent morpheme by itself ” (Robins : ). A second dimension of infor-mativeness is expressed in “the general insight that one inflection tends to predictanother”(Matthews : ). In this domain, the primary locus of form-based im-

Page 17: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

plication is again “words as wholes, arranged according to grammatical categories… distinguished by their endings”(Matthews : ). The role of paradigms like-wise follows from the fact that implications are most reliable within the essentiallyclosed and uniform feature space of an inflectional paradigm.

. Morphological organization

The networks of interdependencies within an inflectional system allows it to be fac-tored into exemplary paradigms and sets of principal parts. This traditional fac-torization rests on a genuine insight about the structure of morphological systems.Paradigms conceived as consisting of fully independent forms cannot be factored inthis way, because such a ‘paradigm’ cannot be identified by any subset of its forms.It is this essential interdependency, rather than numerical bounds or extrinsic con-straints on paradigms, that accounts for the ‘paradigm economy’ effects discussedby Carstairs (); Carstairs-McCarthy () and Ackerman and Malouf ().

For themost part, classicalWPmodels do not go beyond this basic insight aboutthe relational nature of morphological organization. Classical formulations of WPmodels also incorporate a range of extrinsic assumptions that reflect the practicaluses to which these models have been put. While the prominence of words andparadigms is one obvious assumption, there are various other less productive formalassumptions as well. Inflectional systems are almost always factored into a discretenumber of inflection classes, usually with provisions for macroclasses or subclassesin cases where there is considerable overlap. Principal part inventories are also nor-mally ‘static’ in the sense of Finkel and Stump (), in that the same forms or setsof forms, e.g., the neuter nominative singular or first person singular present, aretaken to represent non-exemplary items. The deduction of new forms by match-ing principal parts against exemplary paradigms is likewise attributed to a symbolicprocess of the kind represented by (four-part) proportional analogies.

Each of these assumptions creates problems that are just as artifactual as thoseraised by form segmentation or meaning atomization. The difficulty of motivatingthe choice of principal parts (or ‘leading forms’) is the best known of these problems.

One objection to the Priscianic model … was that the choice of leadingform was inherently arbitrary: the theory creates a problem which it isthen unable, or only partly able, to resolve. (Matthews : )

The other assumptions are equally problematic. Although pedagogical grammarsoften converge on a similar number of classes for a given language, this consen-sus tends to reflect practical considerations of utility. In the absence of criteria forclass identification, the number of classes associated with a language can vary enor-

Page 18: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

mously, as in the celebrated case of Estonian declensions. This generates pseudo-problems and competing solutions to them.

Aswith the problemof stem segmentation, the best solution involves developingmodels in which the artifactuality of these problems becomes plain. For pedagog-ical purposes, it is useful to draw the most informative cells of a paradigm to theattention of language learners. However, there is no reason to assume that a singleform will always suffice to identify the inflectional pattern of an item. Conversely,there is no reason to ignore the partial information supplied by other forms. Thereis also no reason to assume that a language can be organized into some fixed setof classes. Instead, different sets of interdependent forms will, like segments, bedefined by their predictive value. For pedagogical purposes, it will again be usefulto bundle these sets of interdependent forms into larger collections that specify theshape of each form of an item, irrespective of how loosely the forms of different setsare connected to each other. The number of such larger connections will depend onthe uses that they are meant to perform and, accordingly, the level of detail at whichthey are defined and individuated. It is also important to acknowledge that there isno principled reason to assume that the analogical processes that deduce new formsof an item should be representable symbolically, rather than sub-symbolically by areasoning system such as TiMBL (Daelemans and Van den Bosch ).

. Morphological information

Formulating WP models in information-theoretic terms provides strategies avoidseach of these problematic commitments. Each paradigm cell can be associated witha measure of variability or uncertainty that correlates with the number of waysthat the cell can be realized (and the frequency of those alternatives). One cell isof diagnostic value in identifying the realization of another cell (or set of cells) ifknowing the realization of the first cell reduces uncertainty about the realization ofthe second cell (or set). To formalize these intuitions, it is useful to regard paradigmcells as random variables that take realization ‘outcomes’ as their values. The un-certainty associated with the realization of a cell C can then be defined in terms ofthe entropy (Shannon ) of the cell, H(C).

()

H(C) = −∑x∈RC

p(x) log2 p(x)

In this definition, RC represents the set of realization outcomes for C , x representsoutcomes in RC , and p(x) represents the probability that C is realized by x. As in

Page 19: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Shannon’s original definition, entropy is measured in bits.⁶The entropy of a cell is determined both by the number of outcomes and by the

uniformity of their distribution. The greatest uncertainty arises in a system witha large number of equiprobable outcomes. Uncertainty is reduced in a system thathas fewer ‘choices’, either few outcomes in total or else outcomes with highly skeweddistributions. The cumulative uncertainty associated with a paradigmP depends inturn on the uncertainty of its cells C1, C2 . . . , Cn. On a traditional model, cells aregenerally assumed to be interdependent, so that the entropy of a paradigm, H(P),will correspond to the joint entropy of its cells, H(C1, C2, . . . , Cn).

Given a measure of uncertainty, the diagnostic value of an individual cell canbe defined in terms of uncertainty reduction. The relevant notion can be based onconditional entropy,H(C2|C1), whichmeasures the amount of uncertainty thatremains about C2 given knowledge of C1. The more information that C1 providesabout C2, the lower H(C2|C1) will be. If C1 is fully diagnostic, then H(C2|C1)will approach . If C1 is completely uninformative about C2, then H(C2|C1) willpreserve the uncertainty of H(C2). Yet the more uncertain C2 is to start with, thehigherH(C2|C1)will also tend to be. Hence in order to determine relative informa-tiveness, the original entropy values as well as the conditional entropy values mustbe known. Both values are incorporated into the general measure in (), which de-fines morphological information M(C2|C1) as a value between and that isobtained by subtracting from the proportion of the original uncertainty in C2.

()

M(C2|C1) = 1− H(C2|C1)

H(C2)

Thebasic notion of uncertainty reduction expressed in () generalizes directly tocollections of cells. Given that the uncertainty of a paradigm,H(P), can be definedin terms of joint entropy, the morphological information that a cell C expressesabout a paradigm P is expressed in ().

()

M(P|C) = 1− H(P|C)

H(P)

For a concrete example, consider the realizations of the partitive cells in the par-tial Estonian paradigms shown in Table . There are five realizations of the partitive

⁶Information-theoretic WP models assume standard definitions of entropy and related measures.For discussion of some of the formal issues that arise in using Shannon entropy to measure uncer-tainty (including the use of a logarithmic scale, the choice of a binary base, the applicable notion ofprobability, etc.) see Shannon (), Moscoso del Prado Martín et al. () and Milin et al. ().

Page 20: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

singular: str+a, str+e, str+i, str+u, wk+u. If, for the sake of illustration, we assumethat each is equally likely, the entropy of the partitive realization will be:

()

H(part) = −(5× 1

5log2

1

5

)= 2.32 bits

In contrast, there are only four distinct realizations of the ‘short illative’ (str+a, str+e,str+i, str+u) in Table , so the entropy of this cell is slightly lower:

()

H(ill) = −(3× 1

5log2

1

5+

2

5log2

2

5

)= 1.92 bits

Again considering only the forms in Table , there are five equally likely pairsof partitive and second illative realizations (both str+a; both str+e; both str+i; bothstr+u; one wk+u and the other str+a), so the joint entropy H(part, ill) is also .bits. So the conditional entropy H(part|ill) of the partitive given the illative is:

()

H(part|ill) = H(part, ill)−H(ill) = 0.4 bits

It should be intuitively clear at this point how the selection of principal parts isimplicitly guided by entropy reduction. A fully diagnostic cell, such as the parti-tive singular in the Estonian paradigms in table , has a morphological informationvalue approaching because it all but eliminates uncertainty. Fully non-diagnosticcells (such as the dative, locative and instrumental plurals in Russian), have a valueapproaching because they preserve uncertainty. If diagnostic value were an all ornothing affair, then principal parts could be defined as cells that eliminated the un-certainty associated with the paradigms to which they belong. A cell C would be aprincipal part for a paradigm P whenever the value of M(P|C) approached . Ofcourse nothing guarantees that every class system will contain such fully diagnosticforms. However, any system that can be decomposed into exemplary paradigmsand principal parts will consist of partially informative, forms which, in variouscombinations, eliminate the uncertainty associated with a paradigm.

The diagnosticity of a set of cells cannot be determined by summing their in-dividual morphological information values, since multiple forms may reduce theuncertainty of the same (or overlapping) cells in a paradigm. Instead, the diagnos-tic value of a set of forms is measured by their collective morphological informationvalue. Since conditional entropy is also defined for sets of given forms, the diagnos-tic value of cellsC1, C2, . . . , Cn can be determined by generalizing the single cellCin () to the set C1, C2, . . . , Cn.

Page 21: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

The availability of a range of different diagnostic combinations clearly enhancesthe robustness of class identification and form deduction, since speakers are not de-pendent on encountering one uniquely informative form of a paradigm. A classifi-cation based on morphological information thus offers a principled solution to thetraditional problem of identifying principal parts (or ‘leading forms’). The idea thatdiagnosticity correlates with variability across inflection classes can be expressedmore precisely in terms of the uncertainty reduction measured by the morphologi-cal information of a cell or set of cells. From the present perspective, one can see thatthe choice of leading forms is to a large degree arbitrary. A pedagogical or referencegrammar might use seemingly arbitrary criteria to select a particular cell or cell set.A description might select the smallest set of cells, the set with the most highly fre-quent members, or, more capriciously, the cell or cell set with the morphologicallysimplest members, etc. Since any fully diagnostic set of cells will do, all are equallysuitable and the arbitrariness involved in selecting one is harmless.⁷

Moreover, given that cells are informative about the realization of other individ-ual cells or sets of cells, there is no need to mediate the deduction of new forms viaa reified class structure. Instead of being part of the linguistic system, classes canbe regarded as being imposed within a description of the system. Much the sameis true of proportional analogies, which merely provide a symbolic representationof the deductions sanctioned by the morphological information that cells expressabout other cells. In this connection, the present appeal to information-theory de-velops the advantages identified by Hockett (: ) for WP approaches:

One of the most dangerous traps in any of the more complex branchesof science (by no means absent even in the simplest branch, physics)is that of confusing one’s machinery of analysis with one’s object ofanalysis. One version of this is pandemic in linguistic theory today:almost all theorists take morphophonemes (by one or another name)to be things in a language rather than merely part of our equipmentfor the analysis and description of the language A correct principal-parts-and-paradigms statement and a correct morphophoneme-and-rule statement subsume the same actual facts of alternation, the formermore directly, the latter more succinctly.

Although formalized in information-theoretic terms, the notion of morpho-logical information invoked here captures the traditional intuition that Matthews

⁷If the goal is class identification rather than form deduction, there can be also a trade-off betweenthe number of cells required to identify class in a system andwhether one uses the same cells to identifyclass, as Finkel and Stump (, ); Stump and Finkel () show. Recent information-theoreticstudies based on large databases with frequency information have also established the relevance ofjoint entropy measures (Bonami and Luís ; Bonami and Beniamine ).

Page 22: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

(: ) expresses as “the general insight … that one inflection tends to predictanother”. This notion is largely absent from morphological traditions that repre-sent only grammatical information, such as case, number and gender. Part of theproblem lies in the fact that predictive value is not a property of a form in the sameway that, say, case is and hence cannot readily be expressed as a ‘feature’, even if oneaccepts the use of diacritic features for expressing notions like class affiliation.

Concluding observations

The usefulness of information theory for formalizing traditional WP models pro-vides confirms the linguistic relevance of information theory, as initially suggestedbyDescriptivists such asHockett (, ) andHarris (, ).⁸ There is also anumber of more general consequences of recognizing implicational relations as thecornerstone of WP models. This immediately avoids the need to impose a uniformanalysis on all languages at the level of units. From the outset of the rejuvenatedWP tradition, it was clear that many languages do not conform to the ‘agglutina-tive ideal’ of a morpheme-based model and that at least some depart quite radicallyfrom that ideal. Following a classic demonstration of the non-biunique patterningof ‘form’ and ‘meaning’ in Latin verbs, Matthews (: ) observes:

One motive for the post-Bloomfieldian model consisted, that is to say,in a genuinely factual assertion about language: namely, that there issome sort ofmatching betweenminimal ‘sames’ of ‘form’ (morphs) and‘meaning’ (morphemes). Qua factual assertion this has subsequentlyproved false: for certain languages, such as Latin, the correspondencewhich was envisaged apparently does not exist … One is bound to sus-pect, in the light of such a conclusion, that the model is in some sensewrong.

An implicationalmodel also provides a framework for incorporatingmorphemesalongside the patterns that Aronoff () terms ‘morphomic’. A morphome is sim-ply a unit of predictive value. The patterns that Matthews () terms ‘Priscianic’or ‘parasitic’ sanction predictions about the shape of one form based on the shapeof another (as do the morphomic patterns discussed in Maiden ()).

For any Verb, however irregular it may be in other respects, the PresentInfinitive always predicts the Imperfect Subjunctive. For the Verb ‘toflower’, florere → florerem; for the irregular Verb ‘to be’, esse → essem,and so forth without exception. (Matthews : )

⁸Harris’ perspective has had amore direct influence onworks such asGoldsmith () andPereira() and on statistical approaches to segmentation.

Page 23: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

From an implicational perspective, morphemes can be regarded as special cases ofmorphomes that encapsulate a biunique implication between properties and forms.

Finally, an implicational interpretation of classical WP models offers a novelperspective on morphological variation. One of the most striking things aboutmorphology is how much it appears to vary across languages. Languages may havemany, few or even no inflection classes, paradigms may have many cells or few, theforms that realize individual cells may vary widely or be relatively invariant, and soon. Languages may even lack morphology altogether, which has encouraged a viewwhich, in its most provocative form, holds that morphology is somehow “unnatu-ral” or even a “pathology of language” (Aronoff : ). However, much of theproblematic variation involves aspects of morphological systems that are largely ir-relevant to the acquisition or use of language by human speakers. A speaker doesn’tneed to determine the exact number of inflection classes in a language, providedthat the patterns within inflectional paradigms provide a reasonably secure ana-logical base for deducing new forms. The fact that certain formatives may be un-ambiguous while others have a wider range of functions and meanings also posesno great difficulties if these occur in larger wordforms or constructions with sta-ble grammatical properties. Hence there is no reason why the pressures imposedby language acquisition and use should mold languages in ways that produce clearanswers to questions concerning the number of inflection classes in a language, themeaning of specific formatives, or which determine an unambiguous segmentationof forms into stems and exponents and associate discrete meanings with these parts.

From a contemporary WP perspective, these questions, and many of the otherkinds of questions that tend to vex theoretical studies, are either ofmainly pedagogi-cal interest (numbers of classes or sizes of principal part inventories) or else are arti-facts of methods of analysis or schemes of classification (meanings of formatives or‘correct’ segmentations). Consequently, these issues can be seen to fall within whatone might call ‘theoretical lexicography’ rather than within the study of linguisticmorphology per se. Since languages do not develop in response to the demandsimposed by lexicography, there is no reason to expect that these properties will bebroadly similar across languages. In contrast, speakers of all languages do face whatAckerman et al. () call the ‘Paradigm Cell Filling Problem’, the task of deducingor interpreting new forms of an item, based on exposure to other forms of the item.Hence, a traditional model would lead one to expect that the difficulty of this taskwould fall within a fairly circumscribed range. Information-theoretic notions pro-vide the tools tomeasure the difficulty of predicting paradigms from subsets of theirforms, and a research question that is being actively pursued in the current litera-ture explores what Ackerman andMalouf () term ‘the LowConditional EntropyConjecture’, namely that the difficulty of this task is in fact relatively low and largelyindependent of the properties that give rise to apparently extreme morphological

Page 24: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

variation. In effect, this different perspective on the organization and learnabilityof morphological systems has begun to explore what Hockett (: ) anticipatedwould be a benefit of developing WP approaches, namely, that

… there would be a net gain in realism, for the student of the languagewould now be required to produce new forms in exactly the way thenative user of the language produces or recognizes them- by analogy

In line with this, the revival of old ideas and insights associated with earlierexemplification-based WP approaches to morphology are animating and alteringmodern perspectives on “realism” with respect to morphological systems and theirorganization. This arises from the use of quantitative techniques, including information-theoretic measures as well as Bayesian and discriminative learning models (Chateret al. ; Seyfarth et al. ; Blevins et al. ; Ramscar et al. ; Sims ) toinvestigate phenomena and patterns that were less well delineated and understoodwithout these tools. The adoption of this perspective and these methodologies en-folds the study of linguistic morphology within the fertile research program in thedevelopmental sciences that explores and explains phenomena in terms of the dy-namics of interdependencies within complex adaptive systems.

References

Ackerman, Farrell (). Miscreant Morphemes: Phrasal Predicates in Ugric. Ph.D.thesis, UC Berkeley.

Ackerman, Farrell, Blevins, James P., and Malouf, Robert (). Parts and wholes:Implicative patterns in inflectional paradigms. In Analogy in Grammar: Formand Acquisition (ed. J. P. Blevins and J. Blevins), Chapter , pp. –. OxfordUniversity Press.

Ackerman, Farrell and LeSourd, Phil (). Toward a lexical representation ofphrasal predicates. In Complex Predicates, pp. –. CSLI, Stanford.

Ackerman, Farrell and Malouf, Robert (). Morphological organization: TheLow Conditional Entropy Conjecture. Language, , –.

Ackerman, Farrell and Stump, Gregory T. (). Paradigms and periphrastic ex-pression: a study in realization-based lexicalism. In Projecting Syntax (ed. A. J.Spencer and L. Sadler), pp. –. CSLI, Stanford.

Ackerman, Farrell and Webelhuth, Gert (). ATheory of Predicates. CSLI, Stan-ford.

Page 25: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Anderson, Stephen R. (). Where’s morphology? Linguistic Inquiry, , –.

Anderson, Stephen R. (). A-Morphous Morphology. Cambridge UniversityPress, Cambridge.

Aronoff, Mark (). Word Formation in Generative Grammar. MIT Press, Cam-bridge.

Aronoff, Mark (). Morphology by Itself: Stems and Inflectional Classes. MITPress, Cambridge, MA.

Beard, Robert (). Lexeme-Morpheme Base Morphology: A General Theory ofInflection and Word Formation. SUNY Press, Cambridge.

Bickel, Balthasar, Banjade, Goma, Gaenszle, Martin, Lieven, Elena, Prasad Paudyal,Netra, Purna Rai, Ichchha, Rai, Manoj, Kishore Rai, Novel, and Stoll, Sabine(). Free prefix ordering in Chintang. Language, (), –.

Blevins, James P (). Word-based morphology. Journal of Linguistics, (),–.

Blevins, James P (). The post-transformational enterprise. Journal of Linguis-tics, (), –.

Blevins, James P (). Word and ParadigmMorphology. Oxford University Press,Oxford.

Blevins, James P., Milin, Petar, and Ramscar, Michael (). The Zipfian ParadigmCell Filling Problem. In Perspectives on Morphological Structure: Data and Anal-yses (ed. F. Kiefer, J. P. Blevins, and H. Bartos). Brill, Leiden.

Bloomfield, Leonard (). Sentence andword. Transactions of the American Philo-logical Society, , –. Reprinted in Hockett (), –.

Bloomfield, Leonard (). Language. University of Chicago Press, Chicago.

Bochner, Harry (). Simplicity in Generative Morphology. Mouton de Gruyter,New York.

Bonami, Olivier (). Periphrasis as collocation. Morphology, , –.

Bonami, Olivier and Beniamine, Sarah (). Implicative structure and joint pre-dictiveness. InWord Structure andWordUsage. Proceedings of the NetWordS FinalConference (ed. V. Pirelli, C. Marzi, and M. Ferro), pp. –.

Page 26: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Bonami, Olivier and Luís, Ana R. (). Sur la morphologie implicative dans laconjugaison du portugais : une étude quantitative. Mémoires de la Société de Lin-guistique de Paris, , –.

Bonami, Olivier and Stump, Gregory T. (). Paradigm Function Morphology.In The Cambridge Handbook of Morphology (ed. A. Hippisley and G. T. Stump),Chapter , pp. –. Cambridge University Press, Cambridge.

Booij, Geert (). Compounding and derivation: Evidence for constructionalmorphology. In Morphology and its Demarcations (ed. W. U. Dressler, P. Kas-tovsky, and F. Rainer), pp. –. John Benjamins, Amsterdam.

Booij, Geert (). Construction Morphology. Oxford University Press, Oxford.

Börjars, Kersti, Vincent, Nigel, and Chapman, Carol (). Paradigms, periphrasesand pronominal inflection: A feature-based account. In Yearbook of Morphology (ed. G. Booij and J. van Marle), pp. –. Springer.

Brown, Dunstan, Chumakina, Maria, Corbett, Greville G., Popova, Gergana, andSpencer, Andrew (). Defining ‘periphrasis’: Key notions. Morphology, (),–.

Brown, Dunstan and Hippisley, Andrew. (). Network Morphology: A Defaults-based Theory of Word Structure. Cambridge University Press, Cambridge.

Carstairs, Andrew (). Paradigm economy. Journal of Linguistics, , –.

Carstairs-McCarthy, Andrew (). Inflection classes, gender, and the principle ofcontrast. Language, , –.

Chater, Nick, Clark, Alexander, Goldsmith, John A., and Perfors, Amy (). Em-piricism and Language Learnability. Oxford University Press, Oxford.

Chomsky, Noam (). Aspects of the Theory of Syntax. MIT Press, Cambridge,MA.

Corbett, Greville G. and Fraser, Norman (). Network Morphology: A datraccount of Russian nominal inflection. Journal of Linguistics, , –.

Curme, George O. (). A Grammar of the German Language. Macmillan, Lon-don.

Curme, George O. (). A Grammar of the English Language. Heath, Boston.

Page 27: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Daelemans, Walter and Van den Bosch, Antal (). Memory-Based LanguageProcessing. Cambridge University Press, Cambridge.

Dixon, R. M. W. and Aikhenvald, Alexandra Y (ed.) (). Word: A Cross-Linguistic Typology. Cambridge University Press, Cambridge.

Erelt, Mati, Erelt, Tiiu, and Ross, Kristiina (). Eesti keele käsiraamat. Eesti KeeleSihtasutus, Tallinn.

Finkel, Raphael and Stump, Gregory (). Principal parts and degrees of paradig-matic transparency. In Analogy in Grammar (ed. J. P. Blevins and J. Blevins), pp.–. Oxford University Press, Oxford.

Finkel, Raphael and Stump, Gregory T. (). Principal parts and morphologicaltypology. Morphology, , –.

Geertzen, Jeroen, Blevins, James P., and Milin, Petar (). The informativeness oflinguistic unit boundaries. Italian Journal of Linguistics, (), –.

Goddard, Ives (). Primary and secondary stem derivation in Algonquian. In-ternational Journal of American Linguistics, (), –.

Goldsmith, John A (). On information theory, entropy and phonology in theth century. Folia Linguistica, , –.

Harris, Zellig S. (). Morpheme alternants in linguistic analysis. Language, ,–. Reprinted in Joos (), –.

Harris, Zelig S. ().Methods in Structural Linguistics. University of Chicago Press,Chicago.

Harris, Zellig S (). A Theory of Language and Information. Clarendon Press,Oxford.

Haspelmath, Martin (). The indeterminacy of word segmentation and the na-ture of morphology and syntax. Folia Linguistica, (), –.

Hockett, Charles F. (). Problems ofmorphemic analysis. Language, , –.Reprinted in Joos (), –.

Hockett, Charles F. (). Review of The Mathematical Theory of Communicationby Claude L. Shannon and Warren Weaver. Language, , –.

Hockett, Charles F (). Two models of grammatical description. Word, , –. Reprinted in Joos (), –.

Page 28: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Hockett, Charles F (). AManual of Phonology. Indiana University Publicationsin Anthropology and Linguistics, Memoir , Bloomington.

Hockett, Charles F (). A Course in Modern Linguistics. MacMillan, New York.

Hockett, Charles F. (). The Yawelmani basic verb. Language, , –.

Hockett, Charles F (ed.) (). A Leonard Bloomfield Anthology. University ofChicago Press, Chicago.

Hockett, Charles F (). Refurbishing our Foundations: Elementary Linguisticsfrom an Advanced Point of View. John Benjamins, Amsterdam.

Jackendoff, Ray (). Morphological and semantic regularities in the lexicon. Lan-guage, (), –.

Joos, Martin (ed.) (). Readings in Linguistics I. University of Chicago Press,Chicago.

Kel’makov, Valentin and Hännikäinen, Sara (). Udmurtin kielioppia ja har-joituksia. Suomalais-Ugrilainen Seura, Helsinki.

LeSourd, Phil (). On the analytic expression of predicates in Meskwaki. InHy-pothesis A / Hypothesis B: Linguistic Explorations in Honor of DavidM. Perlmutter(ed. D. B. Gerdts, J. Moore, andM. Polinsky), pp. –.MIT Press, CambridgeMA.

Lounsbury, Floyd (). Oneida Verb Morphology. Yale University Publications inAnthropology . Yale University Press, New Haven. Chapter reprinted in Joos(), –.

Luutonen, Jorma (). The variation of morpheme order in Mari declension.Suomalais-Ugrilainen Seura, Helsinki.

Maiden, Martin (). Morphological autonomy and diachrony. In Yearbook ofMorphology (ed.G. Booij and J. vanMarle), pp. –. Springer, Dordrecht.

Masini, Francesca (). Phrasal lexemes, compounds and phrases: A construc-tionist perspective. Word Structure, (), –.

Masini, Francesca and Benigni, Valentina (). Phrasal lexemes and shorteningstrategies in Russian: the case for constructions. Morphology, (), –.

Matthews, Peter H. (). The inflectional component of a word-and-paradigmgrammar. Journal of Linguistics, , –.

Page 29: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Matthews, Peter H. (). Recent developments in morphology. In New Horizonsin Linguistics (ed. J. Lyons), pp. –. Penguin, Harmondsworth.

Matthews, Peter H. (). Inflectional Morphology: A Theoretical Study based onAspects of Latin Verb Conjugation. Cambridge University Press, Cambridge.

Matthews, Peter H. (). Morphology. Cambridge University Press, Cambridge.

Matthews, Peter H (). Grammatical theory in the United States from Bloomfieldto Chomsky. Cambridge University Press, Cambridge.

Milin, Petar, Kuperman, Victor, Kostić, Aleksandar, and Baayen, R. Harald ().Words and paradigms bit by bit: An information-theoretic approach to the pro-cessing of inflection and derivation. In Analogy in Grammar: Form and Acquisi-tion (ed. J. P. Blevins and J. Blevins), pp. –. Oxford University Press, Ox-ford.

Morpurgo Davies, Anna (). Nineteenth-century Linguistics. Volume IV. Long-man.

Moscoso del Prado Martín, Fermín, Kostić, Aleksandar, and Baayen, R. Harald(). Putting the bits together: An information-theoretical perspective onmor-phological processing. Cognition, , –.

Paul, Hermann (). Prinzipien der Sprachgeschichte. Max Niemayer Verlag,Tübingen.

Pereira, Fernando (). Formal grammar and information theory: togetheragain? In The Legacy of Zellig Harris: Language and Information into the thCentury (ed. B. E. Nevin and S. B. Johnson), Volume : Mathematics and com-putability of language, pp. –. John Benjamins, Amsterdam.

Plag, Ingo and Baayen, R. Harald (). Suffix ordering and morphological pro-cessing. Language, , –.

Ramscar, Michael, Dye, Melody, Blevins, James P., and Baayen, R. Harald ().Morphological development. In Handbook of Communications Disorders: The-oretical, Empirical, and Applied Linguistic Perspectives (ed. A. Bar-On andD. Ravid). De Gruyter, Boston/Berlin.

Robins, Robert H. (). In defence of WP. Transactions of the Philological So-ciety, , –. Reprinted in Transactions of the Philological Society , ,–.

Page 30: WordandParadigmMorphology - University of Cambridgepeople.ds.cam.ac.uk/jpb39/pdf/BlevinsAckermanMalouf2017.pdf · Intheancientmodeltheprimaryinsightisnotthatwordscanbesplit into roots

Rueter, Jack (). Adnominal Person in the Morphological System of Erzya.Suomalais-Ugrilaisen Seuran Toimituksia . Suomalais-Ugrilainen Seura,Helsinki.

Seyfarth, Scott, Ackerman, Farrell, and Malouf, Robert (). Implicative orga-nization facilitates morphological learning. In Proceedings of the th AnnualMeeting of the Berkeley Linguistics Society, pp. –.

Shannon, Claude (). Amathematical theory of communication. TheBell SystemTechnical Journal, , –, –.

Sims, Andrea D. (). Inflectional defectiveness. Cambridge University Press,Cambridge.

Spencer, Andrew J. (). Identifying stems. Word Structure, , –.

Stump, Gregory T. (). Inflectional Morphology: ATheory of Paradigm Structure.Cambridge University Press, Cambridge.

Stump, Gregory T. and Finkel, Raphael (). Morphological Typology: FromWordto Paradigm. Cambridge University Press.

Zwicky, ArnoldM. (). How to describe inflection. In Proceedings of the EleventhAnnual Meeting of the Berkeley Linguistics Society (ed. M. Niepokuj, M. Van Clay,V. Nikiforidou, andD. Feder), pp. –. Berkeley Linguistics Society, Berkeley.