Paradigm Function Morphology

1 Introduction
In realization-based morphologies inectional afxes are not classical morphemes, that is, lexical entries which contribute a meaning of their own. Realizational models take as their starting point the full set of features or morphosyntactic properties which characterize a cell in the paradigm of a lexeme, and then provide a set of instructions for constructing or accepting the word form which occupies that cell, and which hence realizes those morphosyntactic properties. In Paradigm Function Morphology this is achieved by means of the paradigm function for the language. Stump (2001:32) characterizes the paradigm function as ... a function which, when applied to the root of a lexeme L paired with a set of morphosyntactic properties appropriate to L, determines the word form occupying the corresponding cell in Ls paradigm. The paradigm function itself is dened by a set of realization rules, which determine the way that specic morphosyntactic properties are realized by stem selection, afxation and so on.

2 Brief overview of PFM

2.1 Realization rules
In PFM inected forms are derived by the successive application of blocks of realization rules. IN the most straightforward instances, each rule introduces an afx, triggered by some set of feature specications. The order in which the rule blocks apply corresponds to the linear order of afxation. A simple example will illustrate. The Finnish noun form talo-i-ssa-ni in my houses consists of the root talo, followed by plural number, inessive case and 1 SG possessor afxes. These can be introduced by the rules shown in (1)1 : (1) a. RRI, {NUM:plural}, N (< X,>) = < X,> b. RRII, {CASE:iness}, N (< X,>) = < Xssa,> c. RRIII, {POSS:1sg}, N (< X,>) = < Xni,> Formally speaking these rules are functions which map an ordered pair consisting of set of features, , and a form to another ordered pair consisting of a (possibly distinct) form and the same set of features. The feature set is a complete characterization of the feature content of the cell in the paradigm, in this case the plural inessive 1sg possessed form. The X represents the output of the previous block of rules. For the rst block it represents the stem dictated by that particular feature set. By default, X denotes the root of the lexeme, in this case talo. The subscripts on the rule name RR denote successively (i) the rule block in which the rule applies, (ii) the set of morphosyntactic properties (feature values) the rule realizes (which we can refer to as ) and (iii) the class of lexemes to which the rule applies, for example verbs or Class IIIa nouns. The convention for determining whether a given realization rule can apply is this: we inspect the subscripted feature set . If is a subset of the complete feature set and
RR for realization rule, noting that he used RR in earlier works to mean rule of referral. A slightly more perspicuous notation for PFM is developed in Ackerman and Stump (2004).
1 Stump uses the abbreviation

the lexeme is of the appropriate class then the rule can apply in that block. The complete feature set is repeated on the right hand side of the realization for essentially technical reasons. For most purposes, however, the repeated feature set in the output can be ignored. The nominative singular unpossessed form of house is just the root, talo. In a strict morpheme based theory this would require a string of zero morphemes meaning respectively singular, nominative and unpossessed. This, however, completely fails to capture the intuition that in Finnish nouns the root form is the default form, used to express default meanings. In a paradigm-based theory the grammar contains a set of features and their permissible values. This will include the specications [NUM:sg], [CASE:nom] and [POSSESSOR:none]. Hence, any Finnish noun is necessarily associated with a cell labelled nominative singular unpossessed, which must be lled by some appropriate expression (provided the paradigm is not defective). The PFM model appeals to an important principle, that of the Identity Function Default. An identify function is a function, f , which applies to some value, x, and delivers the same element: f ( x) = x. In PFM it is assumed that in any block of realization rules there is one realization rule which has exactly the format of the identity function. This is a realization rule which takes any feature set for any category and maps X to X, as shown in (2): (2) Identity Function Default RRn,{any},any (<X,>) = <X,>

For Stump, this is a universal default which applies in any rule block where no explicit rule has applied. In the case of the word form talo house, we would nd that Finnish grammar lacks specic realization rules in Blocks I, II, II for [NUM:sg], [CASE:nom] and [POSSESSOR:none] respectively, and so the Identity Function Default would apply. Since no special form of the root is required the form talo serves as the X in the rule set and this is the output of Block I. This form then serves as the input to the Block II realization rules, where again there is no Finnish-specic rule to apply and the Identity Function Default is invoked. Similarly, talo is the output of Block III. In the specication of the form talo-ssa-ni in my house, where the property singular receives no overt realization, the Identity Function Default applies in Block I and the two overt afxation rules apply in Blocks II, III. In the specication of the unpossessed form talo-i-ssa in (the) houses, the Identity Function Default would apply solely in Block III. The Identity Function Default is formalized as the most general rule imaginable: for any set of features, X = X. Other things being equal, the Identity Function Default would guarantee that no word would ever be overtly inected. Where we have an explicit realization rule, such as that for [NUM:pl], however, the default is pre-empted by the more specic or narrow rule. This illustrates another important feature of PFM. Rule block organization is governed by a principle known under various names but referred to by Stump as P ninian Determinism:2 a. where two rules are in competition the narrowest applicable rule applies. For instance, imagine that OX is the only noun in English with an irregular plural. Then, we can state plural formation by saying (i) for OX the plural is oxen, (ii) for any noun with root X the plural is X-z. Without P ninian Determinism the regular plural rule (ii) would derive *oxes but this application is a. pre-empted by the more specic rule (i). The fact that rules (i) and (ii) are in paradigmatic opposition within the same block is captured by P ninian Determinism. a. In models such as PFM, which are based on the logic of default inheritance and in which the whole grammar can be thought of as a set of nested defaults, P ninian Determinism is a.
2 The reference, of course, is to the great Sanskrit grammarian, P nin, whose grammatical description of Vedic Sanskrit, a. the Astadhy y or Eight Books is widely regarded as the rst, as well as being one of the greatest works in formal linguistics. a .. P nins grammar makes crucial use of the default logic deployed in PFM and other inferential models of linguistics. a.

the principal organizing factor. For example, consider the Finnish nominative plural form talot houses. This has an unexpected (and unique) desinence -t, signalling simultaneously the properties nominative and plural, what is sometimes known as cumulation. Classical morpheme theory has problems with cumulation and other deviations from a strict one-one relationship between form and content. Is -t the nominative case ending used with the plural or is it the plural ending used in conjunction with the nominative case ending? In other words, is the form really (3a) or (3b)? (3) a. talo t house pl nom b. talo t house pl nom In the case of (3a) we would have to explain how the grammar knows that the exceptional zero plural marker has to be selected, while in the case of (3b) it would have to know how the exceptional -t plural marker was selected. In each case we would have sensitivity to an outer afx. In point of fact, a strict morphemic approach would probably have to opt for the analysis in (3b), because if (a) were adopted it is unclear what would prevent the regular plural afx i from being inserted to give the ungrammatical *talo-i-t. In addition, the analysis in (3b) would be regarded as a little more elegant because it would only require one exceptional allomorph, for the plural. Nonetheless, the technical problems with such an example remain: assuming the analysis in (3b) some way has to be found to ensure that the regular plural sufx is not accepted: *talo-i- and some way has to be found to ensure that case sufxes other than the nominative sufx are not sufxed to the -t form: *talo-t-ssa etc.3 Even more problematic are cases in which the two afx positions are separated by other afxes. Stump (2001: 162f) exemplies precisely that situation with the Swahili past negative. The normal past tense prex in Swahili is li and the normal negative prex is ha. In simple cases the order of prexes is ha-SUBJECT. MARKER- TENSE- VERB . ROOT. However, just in the past tense there is a special prex found when the verb is in the negative form: ku. Thus, a negative past tense verb has the form ha-SUBJECT. MARKER-ku-VERB . ROOT . Clearly, measures have to be taken on a strictly morpheme-based theory to ensure that the incorrect string *ha-SUBJECT. MARKER-li-VERB . ROOT isnt accepted as a grammatical word form. In PFM this Finnish example could be handled very simply by adding a rule to the Block I rule set to accommodate the unexpected nominative plural form, as in (4): (4) a. RRI, {NUM:pl, CASE:Nom} ,N(<X,>) = <Xt,> b. RRI, {NUM:pl},N (<X,>) = <Xi,>

Rule (4a) expresses cumulation by allowing one afx to be the realization of two feature specications. (Extended exponence occurs when a single feature specication is mentioned
and Sproat (2007), referring to class notes of mine taken from my website, are surprised by my arguments here. They claim that analysis (3b) is clearly correct and, moreover, that this causes no problems to a morpheme-based account. All we need to say is that the zero nominative morpheme is marked to select the t-allomorph of the plural morpheme |i| (in other words, we need to add a statement of the form if the left adjacent ax is the num:plural morpheme then it must be the t allomorph. However, this presupposes some kind of theory of (inwards) allomorph selection. Specically, some way must still be found in the classical approach of preventing the default forms from being wrongly selected, here, *taloi for the nominative plural, and to ensure that the non-default plural allomorph is incompatible with the other cases. In the classical approach, before the advent of formal or generative ways of thinking, these matters were not made explicit. As Roark and Sproat would have learnt had they attended the relevant lectures, making such assumptions explicit will often lead to precisely the kind of realizational model proposed by Stump and others. This discussion illustrates the importance of the advice we give to our undergraduate students about the dangers of surng websites and quoting their contents without checking the precise context.
3 Roark

in rules from two distinct rule blocks). Note that both rules (4a) and (4b) could in principle apply to a form bearing the feature set [NUM:Pl], [CASE:Nom]. However, rule (4a) pre-empts rule (4b) because it is the more specic. Another way of putting this is to say that rule (4b) subsumes rule (4b). Moreover, this fact can be readily computed by simple feature counting: a rule A is more specic than rule B if its feature set properly contains that of rule B. Thus, we see how P ninian Determinism captures the idea of disjunctive sets of afxes within a single a. position class (in structuralist terms, paradigmatic organization as opposed to syntagmatic organization). An alternative solution to the Finnish nominative plural problem would be to say that there was a single portmanteau sufx which simultaneously realizes the properties plural and nominative, as shown in (3c):

(3)

c. talo t house pl.nom

This, indeed, is the kind of solution often proposed for such data in the classical approach. However, its important to realize that a solution of this sort is represents an important weakening of the principles of the classical morphemic approach. The portmanteau type of solution often has much to recommend it and we will see below in 3.2 how such an analysis is coded in PFM. One could imagine a language in which every afx was distinct from every other and every feature specication was found just once, associated with a single rule in a single rule block. A language which exhibited such rule block coherence would be a perfectly agglutinating language. However, there are no known languages with this property and all the logically possible deviations from the canonical one-one mapping are common. A particularly widespread deviation is underdetermination, when a set of morphosyntactic feature values receives no overt expression in the inected word form. In classical morphemics this is handled by postulating zero morphemes. Thus, on a strict morpheme-based analysis the Finnish word form talo house-nominative-singular-unpossessed would receive the analysis talo---. However, in the great majority of cases where such zero morphemes appear to be needed, the zero expresses the default feature specication for that block and is therefore handled in PFM by the Identity Function Default. Occasionally, we nd genuinely signicative zeros. An instance is provided by the aorist tense forms of Bulgarian verbs. This is signalled by a sufx o: krad-o-x I stole, krad-o-xme we stole, igra-o-x, I played, igra-o-xte you (pl.) played, dava-o-xme we gave dava-o-xa they gave. However, in the 3sg forms the o is ellided before the 3sg agreement sufx e: krade s/he stole, igra-e s/he played, dava-e s/he gave. Stump (2001: 45) overrides the o-sufxation rule in the rst sufx block with a rule of the form shown in (5):

(5)

RRI, {TNS:Aor, AGR:3sg},V (<X,>) =def <X,>

Again, it is not that we add a zero here, rather rule (5) stipulates that no afxation occurs even though this is not the default situation for this property. Finally, none of the afxes is a lexical entry with its own meaning. An inectional afx is simply a marker providing (partial) information about the set of feature specications associated with the cell in the paradigm occupied by the completed word form. 4

2.2 Paradigm functions

Stump (2001: 43) gives (6) as the general characterization of the paradigm function: (6) PF(<X,>) = <Y,>

Thus, the paradigm function for Finnish applied to the lexeme HOUSE, root talo to specify the inessive singular non-possessed form talossa in the house would be (7): (7) PF(<talo, { NUM:sg, CASE:iness, POSS:no}>) = <talossa, {NUM:sg, CASE:iness, POSS:no}>

What this function does is to take a lexemes root paired with a complete set of features required to fully specify a cell in the paradigm and it delivers a pairing of that cells features together with the inected word form that occupies that cell. In other words, the right hand side of the function enumerates a paradigm in the sense of form-property paradigm as dened in Chapter ??. Where a word form is dened by the successive application of rule blocks the order of application is part of the denition of the paradigm function. Assuming three rule blocks for Finnish the paradigm function in (7) can therefore be thought of as a concrete instantiation of the paradigm function shown in (8), where X is the root of any (nominal) lexeme and = {NUM:, CASE:, POSS:}: (8) PF(<X, >) = RRIII (RRII (RRI (<X, >)

The paradigm function is dened over the root of the lexeme, and instructs us how to form completed words from that root. In effect, the root form of the word is being used as an index for the lexeme of which it is a root. This leads to immediate problems, of course, wherever we have root homophony. Stump (2001: 43) allows for this by assuming that each root carries a lexemic index (L-index). In order to reect the fact that an inected form remains a form of one and the same lexeme, Stump adds principle (9), the persistence of L-indexing: (9) For any realization rule, RR(<X,>) = <Y,>, L-index(Y) = L-index(X)

3 Ax ordering
3.1 Three types of deviation
Afxes do not always line up in the way expected and there are several sets of deviations from the agglutinative ideal presupposed in classical structuralist morphotactics. There are three main types of deviation. First, a given afx may appear to straddle a sequence of slots, that is, it may appear to belong simultaneously to a sequence of two or more consecutive slots (portmanteau position classes). Second, we may nd that one and the same set of afxes may appear in distinct positions depending on their grammatical function. A common situation is for a single set of pronominal afxes to be found in different positions when realizing subject, respectively object agreement features. This is the problem of parallel position classes. Third, we may nd that the relative ordering of afxes changes depending on the exact set of features associated with a word form. Again, subject/object markers provide a common scenario, and we may nd that a set of markers (not necessarily parallel) may occur in one order when realizing one feature set and the opposite order when realizing a different feature set (reversible position classes). 5

3.2 Portmanteau position classes

Portmanteau position classes are handled in PFM by portmanteau rule blocks. To describe such phenomena we suspend the assumption that the blocks of rules are necessarily applied sequentially. We can illustrate this situation by returning to the problem of the Finnish nominative plural. An alternative analysis of the Finnish case can be given if we suppose that the t nominative plural sufx occupies not Slot I, but Slots I, II simultaneously. In PFM we would then posit a realization rule dened over the sequence of slots, notated by the composed block index [II,I] as shown in (10): (10) RR[II,I] {NUM:pl, CASE:nom},N (<X,>) = <Xt,>

In effect, the rule (10) is outside the rule block system of Finnish inection, in that rule (10) is in paradigmatic opposition simultaneously with the rules of Block I and of Block II. In order to ensure that rule (10) can apply, we have to ensure that the paradigm function for Finnish nouns can be realized by such a rule. This means that we must revise the paradigm function shown in (8), as in (11): (11) Revised paradigm function for Finnish nouns Where = {NUM:, CASE:, POSS:} PF(<X,>) =def RR[II,I] (RRI (<X, >))

But now we must ensure that the standard Block I, II rules can realize the paradigm function. Therefore, Stump (2001: 142) introduces the Function Composition Default (FCD): (12) RR[n,m] (<X,>) =def RRn (RRm (<X,>))

The composed rule (10) will always be more specic than any individual rule in Block I, II, so that (10) will pre-empt the default plural rule and any other case rule. Where is other than nominative plural, the FCD, (12), applies and the normal realization rule sequence is called into play.

3.3 Parallel position classes

Parallel blocks occur when the same set of afxes occupy different position classes in order to express slightly different functions, as when identical subject and object markers occupy different positions depending on the function. For the sake of exposition let us consider an idealized Bantu language whose verbs inect according to the schema in (13) (where SM/OM mean Subject/Object Marker respectively and TAM means Tense/Aspect/Mood Marker): (13) Bantu prexes SM III TAM II OM I Verb stem 0

Suppose further that the SM, OM cross-reference person and number and are identical in form. If we write independent realization rules for Slots I, III then we would effectively be stating six identical rules twice, but with different rule block indices, thus missing a clear generalization about the language. Stump therefore modies the rule block model by permitting realization 6

rules which effectively have no block index (Stump 2001: 147). The rules of Block I, III are conated as Block Agr. When the paradigm function comes to evaluate a feature set in Block I, III it is referred to the corresponding realization rule in Block Agr.4 If there is an incomplete overlap between SM and OM (as in Lingala) then Blocks I, III will contain their own specic rules which will pre-empt the Block Agr rules because they will make reference specically to subject or object agreement.

3.4 Reversible position classes

Reversible rule blocks are found when we have cases of afxal metathesis. Stump (2001: 149f) discusses cases from Fula subject/object agreement and Stump (1993) discusses similar phenomena in Swahili. In Fula in certain tense forms the default afx order is Verb stem SM OM. However, for certain person/number combinations we nd the opposite order. This can be seen from the forms shown in (14) (cf Stump 2001: 151): (14) a. mall-u-Pon-mo help-past-2pl-3sg you(pl) helped him V-SM-OM b. mall-u-mi-Pe help-past-1sg-3pl I helped them V-SM-OM c. mall-u-moo-mi help-past-3sg-1sg I helped him V-OM-SM The order of application of realization rules is dened as part of the paradigm function (generally using iconically labelled block indices such as I, II, III, . . . or A, B, C, . . . ). In Stumps analysis of the Fula verb the SM, OM slots are labelled III, IV. He therefore sets up realization rules for SM in Block III and rules for OM in Block IV, but adds a portmanteau rule block [IV,III]. By the FCD, (12), this is normally realized by the sequence of realization rules RRIV (RRIII (. . . ))) to give the order V-SM-OM. However, for certain feature combinations the FCD is pre-empted by rule (15) (adapted from Stump 2001: 154): (15) Where = {certain feature combinations} RR[IV,III],,V (<X,>) =def RRIII (RRIV (<X,>))

4 Rules of referral
We often nd that forms occupying cells in one part of a paradigm are systematically identical to those of other cells. This is inectional homonymy or syncretism. This arises for a variety of reasons, as detailed minutely, and with a wealth of examples in Baerman et al. (2005).
4 Technically

the way this is achieved is by means of a rule of referral, discussed below in 4.

In some cases its not actually appropriate to think in terms of homonymy at all. For instance, in Russian there are three genders, masculine, feminine, neuter, but this distinction is completely neutralized for all nominals in the plural. The simplest (i.e. best) way of describing this is to dene a dependency between gender and number, such that the plural subparadigm simply doesnt make any gender distinctions, i.e. the property gender is undened for [NUMBER:plural]. Alternatively, and perhaps equivalently, we can think of plural number as a kind of fourth gender. In other cases the homonymy can be explained in terms of an underspecication of properties For instance, in Latin it is clearly necessary to distinguish ve case forms, including the dative and the ablative. However, there are no nominals (nouns, adjectives, pronouns or whatever) that distinguish the dative and ablative case in the plural. One possible way of describing this situation is to set up a special dative/ablative case which splits into dative and ablative in the singular but remains one case in the plural, fullling all the morphosyntactic functions of both the dative and ablative. A number of other situations are discussed by Baerman et al., together with the various logically possible ways of analysing them. There is one set of syncretisms, however, that are of importance for paradigm-based realizational models. These are instances in which two featurally distinct cells are associated systematically with exactly the same forms but in which one of those formfeature pairings can (or must) be taken as prior, and the other formfeature pairing dened as being identical to the rst. A rule which encapsulates such an equivalence statement is a rule of referral (Zwicky, 1985). Stump provides an interesting case from Bulgarian (and Macedonian) conjugation. Bulgarian has one present and two past tenses, aorist and imperfect. Verb endings consist of a theme vowel, a tense marker and an agreement marker, however there is no simple correspondence between form and function: the 3sg aorist of RABOTJA work is rabot-i while the 3sg imperfect of IMAM have is im-a-e. Nonetheless, there is an exceptionless generalization in s the conjugation system: for any verb, including even the highly irregular verb BE, the 2sg aorist/imperfect forms are identical to the corresponding 3sg aorist/imperfective forms. Stump argues that this syncretism should be expressed by means of a rule of referral, taking the 3sg forms as the basic ones and referring the 2sg forms to the 3sg in either of the two past tenses. In PFM a rule of referral for such a syncretism is dened by referring each realization rule for 2sg forms to the corresponding realization rule for 3sg forms on a block-by-block (effectively, afx-by-afx) basis. In principle, it would be possible to refer the entire word form expressing 2sg features to the entire word form expressing 3sg features, but Stump argues (p. 217f) that there are instances in which only some of the afxes of a form need to be subject to a referral. He therefore treats all referrals as block-by-block referrals, with whole-word syncretism being simply the limiting case in which all rule blocks are subject to the syncretism. For the Bulgarian example, Stump (2001:55f) proposes a rule which in prose states the following: let be a feature set containing the features {TENSE:past,AGR:2sg}; let n be any of the rule blocks needed for describing Bulgarian conjugation. Then, for any realization rule RRn,,V (<X,>) = <Y,> serving to (partially) realize the features {TENSE:past, AGR:2sg}, apply instead the corresponding rule with AGR:3sg substituted for AGR:2sg.

5 Allomorphy in PFM: morphological metageneralizations

The formulation of realization rules given thus far would be sufcient for purely concatenative morphology. However, there are instances of inection in which some kind of morphophonological process applies to a stem in addition to or instead of afxation. An afxation process itself may involve more than just concatenation and may trigger such processes as haplology or other types of truncation. In addition, we may nd that the afx is an inx. Other processes 8

include ablaut, stress/tone/accent shift, consonant mutation, truncation and, very commonly, reduplication. Many of these effects would normally be subsumed under the operation of the morphological metageneralizations argued for by Stump (2001: 47f). These metageneralizations capture the traditional intuition that a morphophonological (non-automatic) alternation might be a part of the morphological process itself. Stump generalizes the format of the realization rule to accommodate morphophonological alternations. Suppose a realization rule introduces a sufx, Z, so that we have RRn,,C (< X,>) = < XZ,>. If we designate the output XZ as Y, then the generalized realization rule takes the form RRn,,C (< X,>) = < Y ,>. The expression Y defaults to Y (i.e. XZ), but for certain rules in certain rule blocks Y might be some phonological alternant of Y. For instance, Y might be X Z, where X is a palatalized variant of X. This alternation would be captured by means of a morphophonological redundancy rule with roughly the form of (16): (16) Where RRn,,C (< X,>) = < Y ,>, if X = WC, where C is a palatalizable consonant whose palatalized alternant is C , and Y = WCZ, then Y = WC Z

In prose, (16) states that result of afxing Z to X is to palatalize C, the last consonant of X, C . Stump (2001: 47) proposes that . . . for each realization rule R, there is an unordered set R of morphophonological rules constraining the evaluation of R in any instance of its application. If the palatalization rule is the only morphophonological rule relevant to applications of R, then Stump would assume a morphological metageneralization of the form Morphophonological rule (16) applies to realization rule R. In some cases a morphophonological rule would be applicable to just one realization rule, capturing the idea that a single afx triggers that alternation, while in other cases every rule in a given block might be subject to the metageneralization. Since a single block often realizes values of a single morphosyntactic category (such as non-past tense or subjunctive mood), this would capture the idea that the morphophonological alternation is a partial exponent of that property. To varying degrees it is possible for the morphophonological rules to capture phonological invariants. For instance, if all and only front vowel initial sufxes trigger a class of palatalizations this fact can be written into the rule. On the other hand, if a language has completely regular [ATR] harmony, such that an afx with a [+ATR] vowel causes all [ATR] vowels elsewhere in the word to become [+ATR], this can in principle be handled by underspecifying the [ATR] vowels for the harmonic feature and by assuming a default rule which species all vowels underspecied for [ATR] after the application of all realization rules and morphophonological rules as [ATR]5 .

6 Stems in PFM
In this section I briey summarize the role of stems in PFM. The stem notion will be discussed in more detail in Chapter ??. Following Aronoff (1992, 1994) Stump argues for the importance of stems and stem selection. Stems may be listed lexically for a lexeme or they may be the result of a completely general word formation rule. For instance, in Romance languages we nd that the verb in the default conjugation has a root but most inected forms are built on a stem formed from the root together with a theme vowel (e.g. Latin amo love root am-, stem ama:-). There is no need to assume that the theme vowel has any meaning or featural content. Indeed, Aronoff argues
5 For arguments that the morphological metageneralizations model is superior to the rather similar system of co-phonologies

proposed by Inkelas and Orgun (1998), see Stump (1998).

that there are cases in which it would be entirely wrong to assume that a stem form expressed a meaning. He discusses the Latin third stem, illustrated by the form ama:t(um). The stem form ama:t- is the basis for the passive perfective participle ama:tus (having been) loved, as well as the supine form ama:tum. However, the third stem is itself the basis for the active voice future participle ama:tu:r(us) about to love. Aronoff outlines a variety of other formations from inection and derivation which also appeal to this stem form. He points out that we know that we are dealing with a specic stem form because many verbs have an irregular third stem but this behaves in exactly the same way as the regular forms. For instance, the verb FERO I carry has a suppletive third stem la:t- which forms both the passive perfective and the active future participles: la:tus carried and la:tu:rus about to carry. Aronoff concludes from such cases that the Latin third stem is just a form devoid of any feature content or meaning over and beyond the lexical meaning associated with the lexeme itself. He argues that morphology has to be couched in terms of processes and representations that appeal just to forms and not to meanings. Such asemantic processes and representations he calls morphomes. Thus, the Latin third stem is an example of a morphomic stem.

6.1 The nature of stems

Stump (2001: Chapter 6) provides a very detailed demonstration of the need for morphomic stems, and an exhaustive illustration of the way that a complex set of stem alternations may operate, basing himself on Sanskrit. We often nd that stems come in groups or sets (the stem sets of Anderson, 1992), which may be associated with each other by means of regular rules of stem formation. A particularly common instantiation of this is the phenomenon of the theme extension to a verb root. This is found throughout Indo-European languages, for instance. In Latin, verb fall into four traditional conjugations dened by theme vowels: 1st conjugation in -a/a: (am-a:-re to love), 2nd conjugation in -e: (mon-e:-re to advise) 3rd conjugation in -e/i (reg-e-re/reg-i-t to rule, s/he rules), and 4th conjugation in -i: (aud-i:-re to hear), but similar phenomena are found through the worlds languages. In other cases, stems may be associated with each other by means of morphophonological generalizations which dont, nevertheless, have the status of morphological rules proper, because they are idiosyncratic and lexically conditioned. In many languages verb stems undergo vowel alternations (ablaut, apophony, largyngealization and so on), or alternations in tone, accent, length. We frequently nd that consonants at the edge of verb stems systematically undergo changes such as (de)voicing, palatalization, spirantization and so on. Stump shows that in Sanskrit there is a class on lexemes for which it is possible to dene stem sets in terms of their distribution as Strong or Middle and there is another set of lexemes for which the Middle class can be divided into Middle and Weakest. For example, the possessive adjective BHAGAVANT fortunate belongs to the class which has two stem forms, Strong and Middle, while the perfecti active participle TASTHIVANShaving stood has the three-way three stem alternation, Strong, Middle and Weakest: (17) a. Strong(BHAGAVANT) = bh gavanta Middle(BHAGAVANT) = bh gavat- (morphophonologically bh gavnt) a a " a b. Strong(TASTHIVANS) = tasthivnsMiddle(TASTHIVANS) = tasthiv ta Weakest(TASTHIVANS) = tasthivus. In the declension of BHAGAVANT in the masculine the Strong stem is used for the nominative/accusative forms, singular and plural together with the nominative plural form. In the 10

neuter gender declension on the nominative plural uses the Strong stem. Elsewhere both genders use the Middle stem. This is an instance of paradigmatic stem selection, which is dealt with by reference to an abstract morphomic stem furnished with an arbitrary index. In the declension of TASTHIVANSthe Middle stem is divided into a Middle and Weakest stem. The Middle/Weakest distinction is effectively allomorphic in the sense that the Weakest stem is selected with the following sufx is vowel-initial and the Middle stem is selected elsewhere (Stump 2001: 174f), i.e we have phonologically conditioned suppletion. This is syntagmatic stem selection, handled by reference to the morphophonological form of the stems and afxes as well as by reference to morphosyntactic and morpholexical properties. We will briey consider both types, starting with the paradigmatic type of stem selection. The three-way characterization of stems into Strong, Middle, Weakest corresponds in general to their morphophonological characterizations, the traditional Zero Grade, Guna Grade . and and Vrddhi Grades (Stump, 2001: 186). We might imagine that there is only need for one . of the two sets of labels: by knowing that a stem is, say, Zero Grade we can usually predict that it has the distribution of a Weak stem and vice versa, while the Strong stem generally appears in the Vrddhi Grade. However, Stump demonstrates at great length that the correspondence . is not perfect: the default mapping can be overridden. For instance, there are occasions when the Strong stem is expressed by the Guna Grade rather than the Vrddhi Grade. An example . . is the vocative singular of a certain class of nouns including R AJAN king. Similarly, there are occasions when the Middle stem is expressed by the Zero Grade rather than the Guna Grade, . among other mismatches. At the same time, however we identify a stem, whether by default from its form, or by stipulated labelling, we cannot in general predict how that stem will be used in the paradigm. In fact, the distribution of the three stem types is very complex and has to be stated as part of the morphology of Sanskrit, sometimes on a lexeme-by-lexeme basis. Moreover, if we examine the actual forms of lexemes such as BHAGAVANT we nd that one and the same sufx may be associated with different stems. For instance, in the declension of lexemes such as BHAGAVANT the masc accusative plural is based on the Middle stem, while the masc nominative plural is based on the Strong stem. Yet the sufx is -as in both cases: masc nominative plural bh gavant-as masc accusative plural bh gavat-as. In other words, Sanskrit a a stems are paradigm examples of morphomic stems in Aronoffs sense. In some cases the selection of stems can be written into the realization rule which introduces the afxes attached to those stems. Stump provides three compelling reasons why this is not sufcient in the general case to account for stem distribution. He provides two such reasons on the basis of the Breton partial paradigm shown in (18): (18) Breton conjugation SKRIVA N write Present Indicative 1sg 2sg 3sg 1pl 2pl 3pl skriv-a-n skriv-e-z skriv skriv-o-m skriv-i-t skriv-e-r Imperfect skriv-e-n skriv-e-s skriv-e skriv-e-m skriv-e-ch skriv-e-d Future skriv-i-n skriv-i skriv-o skriv-i-m skriv-int skriv-o-r Imperative skriv skriv-e-t skriv-o-m skriv-e-nt

Stump assumes that the forms skriv, skriv-e, skriv-i and skriv-o are all stem forms. The rst point is illustrated by the Future forms skriv-i and skriv-o. Both of these are uninected stem forms (on Stumps analysis) and so they dont involve any (non-trivial) application of a realization rule. Therefore, we cant link the selection of the two distinct stems to such a realization rule. The second point is illustrated by the 1pl forms skriv-o-m, skriv-e-m and skriv-i-m. In each case 11

these are the result of the (default) realization rule for the 1pl agreement inection, but in each case the stem is distinct (the imperative form is syncretic with the present indicative form so there are, in fact, only three stem forms here). The third argument comes from Bulgarian conjugation. The verb JAM eat has the stem form jad- for most of the present tense paradigm, and this reects the fact that it belongs to the so-called non-truncating consonantal inectional class, whose 1sg form is regularly -@ as in krad I steal. However, but the 1sg form of JAM @ is not *jad, as would be expected but ja-m. This form is regular on the assumption that the @ 1sg form is built on a stem belonging to the non-truncating, non-consonantal class. But by this reasoning we have to assume that it is the stem selection which determines the nature of the sufx and not the other way around. In other words, stem selection is prior to the operation of any realization rules effecting inection. Phonologically conditioned stem allomorphy is illustrated by the selection of the Middle/Weakest stems of the TASTHIVANS-class of nominals. Consider the masculine accusative and instrumental plural forms of the relevant classes of adjectives. The forms for BHAGAVANT are respectively {bh gavat-as, bh gavad-bhis} (where the t/d alternation is purely phonological) a a but the forms for TASTHIVANS are {tasthus-as, tasthiv d-bhis}. The two case forms of BHAGAV. a ANT are built on the same stem form (modulo regular phonological voicing assimilation), but the accusative plural form of TASTHIVANS differs from the instrumental plural form in that the accusative form is built on the Weakest stem while the instrumental form is built on the Middle stem. This is because the accusative ending is vowel-initial while the instrumental ending is consonant-initial. Stump appeals to a morphological metageneralization which states, in effect select the Weakest stem form when the stem comes before a vowel (see his rules (8, 9), p.181). Interestingly, the stem selection rule for tasthus- cant be attributed solely to the phonology of . the resulting word form. In particular, the Weakest stem is not found prevocalically in compound formation (rather, the default Middle stem is selected in such cases). For this reason, Stump argues that the morphological metageneralization is relativized to Block I sufxation. In this sense the stem selection is governed both by phonological form and by morphological structure. Stump (2001: 183) provides the set of rules shown in (19) for determining the relevant forms of the three stems of an alternating lexeme: (19) Stem-formation rules for L = perfect active participle Where L is a perfect active participle, properties (a) and (b) imply each other and both imply (c): a a. Ls Strong stem is X ivns b. Ls Middle stem is X ivta c. Ls Weakest stem is X ivsu

Parallel to the determination of morphologically distributed Strong/Middle/Weakest stems students of Sanskrit provide a phonological characterization of stem types or grades. The rst two are generally referred to by their Sanskrit names, Vrddhi, Guna and the third is the Zero . . grade. These grades are essentially a type of moraically-dened ablaut root vowels, showing respectively the alternations long- , short-a and no vowel. The basic denition is provided by a Stump (2001: 186) in (20): (20) Sanskrit grades For any gradational nominal L, each of (a)-(c) implies the other two: a. The Vrdhi-grade stem of L has the form X (R)C0 a . b. The Guna-grade stem of L has the form Xa(R)C0 . 12

c. The Zero-grade stem of L has the form X(R)C0 " In an ideal world these phonological denitions would correspond exactly to the morphologically dened stem types, but this is not the case for Sanskrit. We nd, for instance, the following correspondence between grade type and stem type: (21) Strong stem = Guna grade, Middle stem = Zero-grade . R AJAN king: Strong stem = Vrddhi . PAD foot: Strong stem = Vrddhi, Middle stem = Guna . . ATMAN soul: Strong stem = Vrddhi, Middle stem = Zero, Weakest = Guna . . BH AGAVANT ,

Moreover, what correspondences there are only apply to gradating nominals. Many words fail to show gradation at all, so that their morphological stem classes differ from their phonologically dened stem classes by denition. To be sure we can identify a default mapping, as shown in (22) but even this rather complex default statement is frequently overridden: (22) Sanskrit stem defaults Where L is a gradational nominal a. by default, Ls Strong stem is its Guna-grade stem. . b. by default, Ls Middle stem is its Zero-grade stem. c. if L belongs to the Weakest class, then by default, Ls Weakest stem is its Zero-grade stem d. if L {n-stem nominals, perfective active participles, comparative adjectives in -y ms, MAH ANT, . . . }, then Ls Strong stem is its Vrddhi-grade stem. a. .

Such deviations from default mappings between the phonological and morphological characterizations of stems lead Stump (2001: 199f) to conclude that morphological theory requires us to distinguish stem forms and stem indices. The stem index is an arbitrary integer or other label which uniquely identies a given stem for a given class of lexemes, for instanceStrong, Middle Weak. Stump (2001:184, 188f) summarizes these conclusions in (23), as his Indexing Autonomy Hypothesis: (23) The Indexing Autonomy Hypothesis The determination of a stems index is in principle independent of the determination of its form.

Finally, we illustrate stem selection by seeing how it interacts with the notion of portmanteau position class introduced in 3.2. Consider the situation in which an irregular stem serves not as the basis for afxation but as the complete inected word form itself. For instance, the comparative form worse of the lexeme BAD is an irregular stem which doesnt accept the normal comparative sufx -er: *wors-er. In other words, we must ensure that in such cases the word form-cum-stem doesnt feed into the normal afxation rules. By contrast, one might wish to analyse the irregular comparative better as consisting of an irregular (suppletive) stem bett- followed by the regular -er sufx. We shall assume that the comparative is an inectional morphosyntactic category governed by the feature DEG:{positive, comparative, superlative}. Here I shall simplify Stumps rule schemata in obvious ways to make the discussion more perspicious. The analysis of worse and better appeals to the notion of portmanteau position class and denes worse as the result of applying a stem selection rule to the class [I,0], which in English denes an entire inected word form. Thus, we dene rule (24) where BAD is the label for the lexical class containing BAD as its sole member (Stump 2001: 208f): 13

(24)

RR[I,0],{comparative}.{BAD} , (X) = worse

The Function Composition Default (12) will apply unless pre-empted by rules such as (24), for example to dene taller as the comparative of TALL: (25) PF(tall, comparative) = RR[I,0],{comparative},ADJ (tall) = RRI (RR0 (tall) = RRI (tall) = taller

(by FCD)

The Block 0 realization rule selects the root tall as the stem and the Block I realization rule is the regular -er sufxation rule. Now consider better, which, for the sake of argument we are assuming consists of the suppletive stem bett- with the regular -er afx. The stem selection rule for better takes the form (26): (26) RR0,{comparative}, {GOOD} (X) = bett

This differs from the stem selection rule for worse in that it refers only to Block 0. It thus denes an irregular stem to which the regular sufx is attached in Block 1. The irregular stem selection rule (26) pre-empts the default stem selection rule which would have selected the root form good. 6.1.1 Stems - summary

In sum, we require the following system of rule-types and principles: Stem-formation rules, which specify the phonological shape of a stem or set of stems. In default cases the stem-formation rule will also permit us to predict the stem index and vice versa, but not always. Stem-indexing rules, overriding the default specications given by the stem-formation rules or other default principles. Stem-selection rules, which specify which stem is to be used in a given cell of the paradigm. Such rules are the rst rules in the paradigm function of the lexeme and the rst block of inectional realizational rules proper takes this stem as its input. Where there is no stem-selection rule, stem-selection proceeds by default (in which case the input to the realization rules is typically the lexical root).

6.2 Paradigm linkage in PFM

6.2.1 Form-paradigms and content-paradigms

In a number of works (Stump, 2002, 2005, 2006; Stewart and Stump, 2007) Stump argues for an enriched conception of the notion morphosyntactic feature, which does justice to the syntactic functions of features as well as to their morphological functions. He argues that each lexeme is associated not only with a set of form-function pairings in the morphology, but also with a set of pairings between the lexemic index of the lexeme and those properties which 14

govern the behaviour of that lexeme in the syntax. For instance, an English count noun such as cat has singular and plural forms, but these forms correspond to syntactic terminals where appropriate singular, resp. plural forms can appear. This means that, in addition to the form paradigm of the lexeme cat, given in (27) we also have a content paradigm, as shown in (28): (27) Form paradigm for cat a. <|kat|, sg> b. <|kat|, pl> (28) realized as /kat/ realized as /katz/

Content paradigm for cat a. <cat, SG> b. <cat, PL>

An example such as this is trivial, of course, because in the default case the morphologically relevant form paradigm is isomorphic to the content paradigm. Needless to say, however, there are interesting cases of mismatch in which the default isomorphism is subverted. There are several types of mismatch between morphological form and syntactic function. A classical example is deponency (Stewart and Stump, 2007:393f) such as that found in Latin conjugation. A whole host of verbs in Latin have passive morphology but are active in meaning. For instance, the verb rego rule has an active form regit rules (3sg, present) and a passive form regitur is ruled (3sg, present), while the verb loquor speak has just the passive form loquitur which, however, has active meaning speaks. There is no form *loquit. Therefore, we have to say that their form paradigms are dened over whatever features distinguish active verb forms from passive forms, but their content paradigms contain only the active voice feature, as illustrated in (29): (29) Deponency as an instance of form-content mismatch rego rego loquor (loquor (loquor Form cells <reg, 3sg, pres, act> <reg, 3sg, pres, pass> <lokw, 3sg, pres, pass> <lokw, 3sg, pres, act> <lokw, 3sg, pres, pass> Content cells <rego, 3SG, PRES, ACT> <rego, 3SG, PRES, PASS> <loquor, 3SG, PRES, ACT> <loquor, 3SG, PRES, ACT> <loquor, 3SG, PRES, PASS> Realization regit regitur loquitur ) loquitur)

The mappings for loquor in brackets in (29) are undened for this lexeme. Other types of mismatch include syncretisms, heteroclisis (Stump, 2002, 2006), periphrasis (Ackerman and Stump, 2004) and principal parts phenomena of various kinds (Stewart and Stump, 2007). The (default and non-default) relations between form and content paradigms are specied by rules of paradigm linkage. The default case is represented by the universal rule of paradigm linkage given in (30) (Stewart and Stump, 2007: 392; see also Stump, 2006: 286): (30) The universal default rule of paradigm linkage Given a lexeme L, where R is Ls root, <L, > <R, >

The binary connective is to be interpreted as expressing the relation between corresponding feature specications in the obvious way. 15

6.2.2

Heteroclisis

One particularly clear instance of the need for a distinction between form and content paradigms is the phenomenon of heteroclisis, in which a lexeme inects according to one inectional class for one part of its paradigm but according to a distinct inectional class for the other part Stump (2002, 2006). Heteroclisis also illustrates the need for an articulated theory of stem selection. We will consider the example of the Czech noun pramen spring, source. This declines as a soft (palatalized) noun in the singular, like pokoj room and as a hard (non-palatalized) noun in the plural, like most bridge, as shown in Table 1 (adapted from Stump 2006:xxx): Declension Singular Nom Voc Acc Gen Dat Instr Loc Plural Nom Voc Acc Gen Dat Instr Loc soft pokoj room pokoj pokoji pokoj pokoje pokoji pokojem pokoji pokoje pokoje pokoje pokoj u pokoj m u pokoji pokojch heteroclite pramen source pramen prameni pramen pramene prameni pramenem prameni prameny prameny prameny pramen u pramen m u prameny pramenech hard most bridge most moste most mostu mostu mostem most e mosty mosty mosty most u most m u mosty mostech

Table 1: Heteroclite declension of Czech pramen source

However, although the soft/hard declension pattern in Czech has its historical origins in phonologically dened stem types, this phonological motivation has been obscured, so that there is no way of telling from the form of the root that PRAMEN takes anything other than the hard set of endings. What this means is that the lexical entry of PRAMEN has to be furnished with a root with two distinct labels, say, PRAMENSOFT and PRAMENHARD. The lexeme is then given a class feature indicating that the hard root is to be used with plural forms. The paradigm function for a heteroclitic word now has to made sensitive to the distinction between the two roots. The way that Stump (2002, 2006) achieves this is to distinguish two sorts of paradigm, a syntactic/content paradigm and a morphological/form paradigm. The form paradigm is dened over a pair <root, features>, while the content paradigm is dened over a pair <L, features>, where L is the lexemic index of the lexeme. In the normal state of affairs, the two paradigms are homomorphic, and there is a trivial mapping between them. This is illustrated schematically for the non-heteroclite nouns POKOJ room and MOST bridge (with obvious abbreviations):

(31)

Form-content mappings for POKOJ, MOST 16

a.

< pokoj, gen sg> < pokoj, gen pl> ... < most, gen sg> < most, gen pl> ...

< POKOJ, GEN SG> < POKOJ, GEN PL> < MOST, GEN SG> < MOST, GEN PL>

= = = =

pokoje pokoju mostu mostu

b.

However, for heteroclite nouns such as PRAMEN this default linkage is subverted. Taking the soft root to be the default, the hard root form is then a co-radical. Stump then denes a special linkage rule, (32), stating that the hard co-radical is selected for the plural number part of the paradigm. (32) Czech rule of stem licensing: If lexeme L belongs to the PRAMEN class, then {NUM:pl} licenses Ls coradical stem.

6.3 Stems and the English verb

English provides a simple example of how stem formation and indexing works6 . A regular transitive verb can express the morphosyntactic features past tense, perfect participle and passive participle. However, the form is the same for all three features, e.g. walked. We can call this Stem1. This stem is derived by regular afxation from the root and on its own can realize each of the three features. However, some verbs have an irregular Stem1, such as BUY: root buy, Stem1 bough- (/bO:/) (or perhaps Stem1 = bought). This listed Stem1 form overrides the default stem formation rule, but it behaves in other respects exactly like Stem1. In other verbs, the past tense and perfect/passive participle features are realized by different stem forms, as in WRITE, root write, past wrote, participle stem writt- (/rIt/). What we can say here is that there is a Stem3 which is by default identical to Stem1 but which in some verbs has a form distinct from Stem1. In all verbs the perfect and passive participles are realized by the same stem form (Stem2). This can be pictured as in (35), where the denition of the different stems is made explicit in (34): (33) English verb stem system Stem0 (=root) X X X X X etc. (34) walk buy write take put Stem1 X-ed Y Y Y X walk-ed bough-t wrote took put Stem2 = Stem1 Y Z X X example
WALK BUY WRITE, SING TAKE PUT

(regular)

walk-ed bough-t writt-en take-n put

(35)
6 For

Stem-to-feature mapping for English

a detailed discussion of stems in English and West Germanic generally see Blevins (2003).

17

By default Stem2 = Stem1 Stem2 [VFORM:perfpart] Stem2 [VFORM:passpart] Stem1 [TENSE:past] Rules of morphology dene which stem is used for realizing which set of features. For instance, the realization of [VFORM:perfpart] appeals to Stem2 (by default set equal to Stem1). In the case of verbs of the class[WRITE] this stem serves as the basis for -en sufxation. In the case of verbs of the class[SING] Stem2 on its own realizes that property set, and hence the stem form occupies a portmanteau position class (Slot[1,0]; cf the discussion of portmanteau stem-selection rules, Stump 2001:208f).

7 Derivational morphology in PFM

7.1 Derivational paradigms
A number of authors have argued that derivational relationships can in some cases exhibit paradigmatic structure, though in some instances different authors may mean slightly different things by paradigm (or indeed derivation) (Bauer, 1997, Booij, 1997, 2002, (Spencer, 1988)). Consider the set of words in (36): (36) Subject nominalizations in English drive walk act reside study claim apply chair guide cook y driver walker actor resident student claimant applicant chairman/chairperson guide chef pilot

Clearly, the English lexicon is structured in such a way as to warrant us dening a notion of subject nominalization, under which for any given verb there is (usually) some noun which denotes the subject argument of that verb. The morphological means used to realize this correspondence are varied, though its also clear that the sufxation of -er (which may also include -or as a graphemic variant) is in some sense the default morphology. To see these correspondences as a paradigm we only need to compare the list in (36) with a similar list of regular and irregular plurals. In each case we have a systematic relationship, expressed by means of generally afxal morphology, but with exceptions, including conversion (guide), stem suppletion (applicant) and whole word suppletion (chef, pilot). The only difference in the formal means of realization between the list of plurals and the list of subject nominals is illustrated by chairperson, which is a compound (and which many morphologists, no doubt, would omit from the list of subject nominals for that reason). To describe singular/plural inection we set up a feature [NUMBER] with two values {sg, pl}. To describe subject nominalization we cannot deploy a binary feature in this way, however, 18

because the base verb and its subject nominalization are not in an equipollent relation to each other. Rather, the subject nominalization process is determined by a privative (single-valued) feature subject nominalization. On this understanding, the subject nominalization paradigm is a very simple paradigm, consisting of exactly two cells per verb lexeme, one occupied by the base verb and the other by its nominalization. The individual morphological means for realizing the nominalization are in paradigmatic opposition to each other, hence, the application of the paradigm metaphor.7 There is a second notion of paradigm that is of some interest to models of morphology. In Spencer (1988) I discuss so-called bracketing paradoxes of the type generative grammarian or bass guitarist. These are instances of morphosemantic mismatches. Morphologically the expressions are clearly segmented as written: [[generative] grammarian], [[bass] guitarist]. However, semantically they deserve the segmentation [[generative grammar]-ian], [[bass guitar]ist]. There are a great many expressions in English which exhibit this kind of morphosemantic mismatch, some of them systematically. Of particular interest to paradigmatic approaches to word formation are specically the personal nouns of the kind generative grammarian. The reason is that these represent a highly productive pattern, though one which can only be dened in terms of a structured lexicon. Spencer (1988) discusses cases of the form (37) in some detail: (37) generative grammarian criminal lawyer baroque autist moral philosopher modern linguist plastic surgeon electrical engineer organic chemist theoretical linguist monumental mason North American East German Southern Dane What these examples have in common is that they are all personal nouns derived from nominal expressions of the form Adjective + Noun. In each case, the base expression has to be lexicalized. Thus, for afcionados of pre-classical music, the term baroque ute denotes a specic type of (pre-Bohm) instrument, almost certainly made out of wood and with at most one key. It is a xed expression with a xed denotation, and for that reason can serve as the base for derivation.
7 There is in the literature another interpretation of the notion of paradigm. Bauer (1997) has argued for the need to include derivational paradigms in morphological theory, where by paradigm he means something more akin to a syntagmatic notion. Thus, from, say, the noun nation he would derive the paradigm consisting of the words {nation, national, nationalize, re-nationalize, nationalization, . . . }. This notion corresonds, if anything, to the notion of word nest sometimes used by Soviet morphologists (@Uluxanov 1996?? other refs - check). Although such a notion is potentially of lexicographic and psycholinguistic interest, I dont really see how it has any role to play in a theory of morphology, so I shall ignore this interpretation.

19

The fact that baroque autist is the personal noun form can be deduced from the fact that autist is the personal noun form of the head noun ute. In other words, the relationship between the base expression baroque ute and the personal noun baroque autist is crucially dened with respect to the head of the expression, ute. A derivational way of picturing this is to deploy what Hoeksema (1985, 1989) has called a head operation. We rst form the personal noun from the head, to get autist. Then we take the phrasal base baroque ute, circumscribe out the modier, baroque, apply the personal noun forming process to the remaining head, and replace the modier: (38) Derivation of baroque autist baroque ute circumscription of head noun ute personal noun formation autist re-instatement of modier baroque autist

Another way of picturing this is as in (39): (39) baroque ute (baroque) ute = (baroque) autist
ist

and in Spencer (1988) I present the process as a kind of Latin Square: (40) Latin Square representation of baroque autist ute baroque ute autist baroque autist

However we choose to picture it, the crucial facts are that (i) the base has to be a lexicalized expression and (iii) the process is entirely productive. Point (i) is evident when we look at minimal pairs which involve ordinary phrases rather than lexicalized phrases. In (41) we see some examples of failure of the personal noun formation process: (41) wooden ute modern linguistics wooden autist modern linguist

Wooden autist can only mean autist who is wooden and modern linguist, if the output of personal noun formation, can only mean specialist in/student of Modern Languages, and thus has to be derived from the (xed) expression modern languages, not the syntactically formed modern linguistics. A further salient feature of personal noun formation is that it often involves suppletive or subtractive morphology. For this reason it cant sensibly be treated in terms of ordinary morphemic analysis8 . As I point out in Spencer (1988), personal noun formation is paradigmatic, though not in quite the sense that inection is paradigmatic. The grammar of English has to include some feature, label or whatever which encodes the fact that personal noun formation from nouns and lexicalized nominal phrases is (in principle always) possible. It must also specify that the
and Neeleman (2004) develop a very interesting analysis of these constructions in terms of an abstract type of morpheme, their AFFIX, which circumvents some of the problems raised for a classical morphemic approach.
8 Ackema

20

morphological form of a phrase-based derivation is computed from the morphological form obtained by applying that process to the noun head of the phrase. In this respect, we have a derivational paradigm, just as we do for deverbal subject nominalizations, though not one which can easily be dened in terms of a default afxation process.

7.2 Derivational paradigms in PFM

The notion of derivational paradigm illustrated by 37 is the second of the two interpretations mentioned by Stump (2001:255). We can think of it as a meaning-, feature-, or content-driven notion of paradigm, in which formal relationships are secondary. The rst is that of Stump (1991), devoted to morphosemantic mismatches of the baroque autist variety, in which he treats a derivational paradigm as being dened by the (usually afxational) process which derives the new lexeme. On this interpretation the examples in (??) and (??) would all represent distinct paradigms, that is, there would be an -er/or paradigm, an -ent/ant paradigm and so on. This is a form-driven notion of paradigm. One consequence of Stumps (1991) way interpreting this type of paradigm is that the semantic relationship between the derived lexeme and the base lexeme can be very (in principle, completely) different. In fact, these two notions of paradigm have been current in the East European structuralist tradition for some time. The meaning-driven notion corresponds to the notion of derivational category, while the form-driven notion corresponds to the notion of derivational type (see Szymanek, 1988, 1989 for detailed discussion of these distinctions). I will take the crucial notion here to be that of derivational category, i.e. that which appeals to the notion of content-paradigm. Stump (2001:257) illustrates the application of PFM to derivation by considering the case of friendless from friend. Since we are assuming a derivational paradigm there must be a paradigm function which delivers the derived lexeme. This means that the paradigm function (in its most general sense) for English must include an application of the form (42), in which is a syntacticosemantic (i.e. derivational) category: (42) PF(<X,>) =def <Y,> PF(<friend, privative adjective>) = <friendless, privative adjective>

Notice that the paradigm function is dened as usual over a pairing of root and property to derive a new root.

8 Head Marking and the Head Application principle

In morpheme-based models of morphology, on one interpretation of the notion of morpheme concatenation, it is natural to suppose that morphology gives rise to branching structures. Indeed, this will be the case if we regard morphemes as lexical entries in their own right, so that morpheme concatenation is homologous with compound formation. This is the essence of so-called Word Syntax (see Selkirk, 1982, for an early model and Toman, 1998 for an overview of the issues). In clear cases of endocentric compounding the head is that element which determines the overall syntactic category of the compound, which determines the meaning of the compound (by dening a denotation which is then delimited in some way by the non-head) and which receives the inections of the compound as a whole. Thus, mousetrap is headed by trap because it is a kind of trap and because more than one of them are called mousetraps rather than *micetrap. Similarly, a blackbird is a bird and the word blackbird is a noun, just like bird (but unlike black). Given this background, morphologists have asked whether afxed words 21

are headed in the same way. For instance, are the word forms walked, re-write or walker headed and if so what is the head? The most extreme answer to the question is that given by Williams (1981), who argues that for all languages the head is the right-most element. I and many other observers have been mystied by this claim so I will leave it to one side. Another possibility is that the most recently added afx is the head, mimicking compounding. This would make -ed, re- and er respectively the heads in our examples. For this to respect the putative homology with compounding this would mean that the afxes would have to have a syntactic category which is the category of the whole word, and a meaning which is modied by the non-head (the lexical root in this case). While this may make some sense for walker, its less obvious that it can be applied to examples other than highly canonical derivational morphology. For re-write we would have to say that the basic meaning was that of a repeated event, further delimited by the concept write. For walked we would have to say that the meaning is that of past event, modied by the concept walk. Not surprisingly, some theorists have balked at this way of analysing inection. Selkirk (1982) proposed a notion of relativized head, under which different parts of a complex word can contribute head properties. For instance, we could argue that the wordform walked has two heads. The lexical head walk provides the lexical semantics (and syntactic class, perhaps) to the word as a whole while the -ed afx provides the feature [TENSE:past]. In a sense, this is equivalent to treating an inected word form as though it were a periphrastic expression such as has walked. In that expression the auxiliary verb is the inectional head (in that it takes tense and agreement inections) while the verb form walked is the lexical head. The inected word form as a whole inherits the properties of its parts by a process of feature percolation under which morphosyntactic and morphosemantic properties pass from a lower node to a higher node. In languages which permit several inections per word form, each inection is a head in its own right. The natural critique of the relativized head notion is that it evacuates the notion of head of its crucial meaning. In realizational models the question of headedness doesnt arise in this form, because there are no afxal morphemes as such to serve as pseudo-syntactic heads and morphology is not identical to (endocentric) compounding. However, there do arise situations in which the notion of head seems to be warranted even in realizational theories. Consider English prexed verbs of the type undertake, understand, withstand, withhold, uphold. These verbs inect in the same way as their unprexed bases, whether the prexes are unproductive and noncompositional or productive and semantically transparent (such as stressed re-). This means that whatever lexical property it is that verbs such as take, stand, hold may have which determines their conjugation that property is preserved under prexation. A popular way of thinking of such situations is to say that the prexed words are headed by their verb bases and that it is the heads that determine inection. This way of looking at things is particularly attractive when the inections are also prexes and they appear closer to the root than the derivational prexes. This is a common situation, being found in classical Indo-European languages, as well as Modern Greek, German and a whole host of other languages. For instance, in German we have the verb nehmen to take, past participle ge-nommen. A prexed form of this verb, such as mit-nehmen to take with (one) has the past participle mit-ge-nommen, in which the prexal part of the lexical root is added to the form inected for past participle. Stump (1991, 2001:Chapter 4) argues that a necessary (though not, of course, sufcient) condition for analysing complex words as headed structures of this type is that they are the result of derivation or compounding which preserves word category, as in the examples cited above. If in addition such a process is transparent with respect to some property of the base then we can speak of a headed structure. Stump offers a number of examples of such transparency from diminutives in various languages (more generally, evaluative morphology, encompassing perjoratives, augmentatives and so on as well as diminutives sensu stricto). Perhaps the com22

monest use of evaluative morphology is found with diminutives of nouns. In some languages the diminutive morphology is no different from straightforward derivation. For instance, in German the diminutive sufx -chen creates a neuter gender noun, irrespective of the grammatical (or natural) gender of the base (masculine, feminine or neuter). However, a property of diminutive processes in many languages is that they preserve the gender of the base noun, even where this is at variance with the default gender assignment for the inectional class of the output. For example, Russian is rich in diminutive sufxes, all of which preserve the gender of their bases. Thus, the affectionate/diminutive sufx -ulja creates a feminine gender noun mamulja from mama Mummy (feminine) and a masculine gender noun, papulja, from papa Daddy (masculine). In this case the afx determines the inectional class of the output, but the gender is transparent. In other cases the gender and also the inectional class are transparent. Thus, the sufx -e ek-e ka-e ko is added to consonant-nal (masc.), -a nal (fem.) c c c and -o nal (neuter) lexemes, as in celovek elov-e ek person (masc.), doskadoc-e ka board c c s c (fem.), slov-oslov-e ko word (neut.) preserving both gender and inectional class (though c it can be added to Class III nouns ending in a palatalized consonant, such as do do -e ka c c c daughter, in which case it imposes Class II membership while preserving gender). Stump (p. 99f) cites further examples from Southern Barasano and Breton, in which the diminutive formation rule preserves syntactic category (noun, adjective, adverb). The Russian diminutives, while headed in Stumps technical sense, inect in a way that targets the sufx, not the base noun. In this respect the Russian examples are not head marking, and Stump refers to this type of (standard) inection as external marking (EM). However, the Southern Barasano diminutives do exhibit head marking. The sufx is -aka. When attached to a noun such as wi house it gives wiaka little house but the plural of the diminutive is formed by sufxing -aka to the plural inected form of the base, wi-ri, to give wiriaka. This is head-marking (HM). Diminutive formation in Breton is, if anything, even less canonical: the diminutive sufx -ig gives rise to double plural marking when attached to a noun: bag boat, bag-ig little boat, bag-o` boats, bag-o` -ig-o` little boats. Thus, Breton exhibits double u u u marking (2M). The property of being head-marking or not is a property of the derivational or compounding construction itself: either all the outputs of the given construction (coderivatives) exhibit head-marking or none do (the Coderivative Uniformity Generalization, p. 98, 108). Moreover, if an output exhibits head-marking in some part of its inectional paradigm it exhibits head-marking throughout the whole of the paradigm (the Paradigm Uniformity Generalization p. 98, 109). This means that there are three types of category-preserving derivational or compounding construction. Those constructions that exhibit EM are called root-to-root rules, those constructions that exhibit HM are called word-to-word rules and those that exhibit 2M are called word-to-stem rules. The logic of the terminology is this. In inecting a Russian diminutive the realization rules apply to the derived word in exactly the same way they would apply to a simplex word. Thus, the base form of the derived word behaves itself exactly like a root with respect to the operation of the realization rules of the inectional component. However, when a HM word such as the Southern Barasano diminutivized noun is inected, the effect is to take an inected word (the plural form wi-ri houses) and apply the diminutive process to deliver another inected word (perhaps word form-to-word form would be a more accurate description). In the case of a doubly marked Breton diminutive, the -ig sufxation applies to an already inected word form, as in Southern Barasano:[bag-o` ] + -ig. However, the output, u bago` ig is not a completed word form, rather it is the stem to which a second round of plural u afxation applies, to give bago` igo` . Hence, the -ig sufxation creates a stem from a word. u u The behaviour of headed constructions produced by word-to-word rules is governed by the Head Application Principle (HAP) (p. 115). In simple terms the HAP states that for any word Y that is headed by Z, every inected word form of Y is headed by the corresponding 23

word form of Z. For instance, the prexed verb lexeme UNDERTAKE is headed by TAKE. This means that each inected form of UNDERTAKE (i.e. undertake, undertakes, undertaking, undertook, undertaken) is headed by the corresponding forms of TAKE.

References
Ackema, Peter and Ad Neeleman. 2004. Beyond Morphology: Interface Conditions on Word Formation. Oxford: Oxford University Press. Ackerman, Farrell and Gregory T. Stump. 2004. Paradigms and periphrastic expression: a study in realization-based lexicalism. In L. Sadler and A. Spencer, eds., Projecting Morphology, pages 111158. Stanford, CA: Center for the Study of Language and Information. Anderson, Stephen R. 1992. A-Morphous Morphology. Cambridge: Cambridge University Press. Aronoff, Mark. 1992. Stems in Latin verbal morphology. In M. Aronoff, ed., Morphology Now, pages 532. Albany, NY: SUNY Press. Aronoff, Mark. 1994. Morphology by Itself: Stems and Inectional Classes. Cambridge, MA: MIT Press. Baerman, Matthew, Dunstan Brown, and Greville G. Corbett, eds. 2005. The syntax-morphology interface: A study of syncretism. Cambridge: Cambridge University Press. Bauer, Laurie. 1997. Derivational paradigms. In G. Booij and J. van Marle, eds., Yearbook of Morphology 1996, pages 243256. Dordrecht: Kluwer Academic Publishers. Blevins, James P. 2003. Stems and paradigms. Language 79(4):737767. Booij, Geert. 1997. Autonomous morphology and paradigmatic relations. In G. Booij and J. van Marle, eds., Yearbook of Morphology 1996, pages 3554. Dordrecht: Kluwer Academic Publishers. Booij, Geert. 2002. Constructional idioms, morphology, and the Dutch lexicon. Journal of Germanic Linguistics 14:301327. Hoeksema, Jack. 1985. Categorial Morphology. New York: Garland Publishing. Hoeksema, Jack. 1989. Head-types in morpho-syntax. In G. Booij and J. van Marle, eds., Yearbook of Morphology 1988, pages 123137. Dordrecht: Kluwer Academic Publishers. Inkelas, Sharon and Orhan C. Orgun. 1998. Level (non)ordering in recursive morphology: evidence from Turkish. In S. G. Lapointe, D. K. Brentari, and P. M. Farrell, eds., Morphology and Its Relation to Phonology and Syntax, pages 360392. Stanford, CA: Center for the Study of Language and Information. Roark, Brian and Richard Sproat. 2007. Computational Approaches to Morphology and Syntax. Oxford Surveys in Syntax and Morphology. Oxford: Oxford University Press. Selkirk, Elizabeth. 1982. The Syntax of Words. Cambridge, MA: MIT Press. Spencer, Andrew. 1988. Bracketing paradoxes and the English lexicon. Language 64:663682. Stump, Gregory T. 1998. Comments of the paper by inkelas and orgun. In S. G. Lapointe, D. Brentari, and P. Farrell, eds., Morphology and Its Relation to Phonology and Syntax, pages 393405. Stanford, CA: CSLI Publications. 24

Stump, Gregory T. 2001. Inectional Morphology: A Theory of Paradigm Structure. Cambridge: Cambridge University Press. Stump, Gregory T. 2002. Morphological and syntactic paradigms: arguments for a theory of paradigm linkage. In G. Booij and J. van Marle, eds., Yearbook of Morphology 2001, pages 147180. Dordrecht: Kluwer Academic. Stump, Gregory T. 2006. A theory of heteroclite inectional paradigms. Language 82:279322. Szymanek, Bogdan. 1988. Categories and Categorization in Morphology. Lublin: Wydawnictwo Katolickiego Uniwersytetu Lubelskiego. Szymanek, Bogdan. 1989. Introduction to Morphological Analysis. Wydawnictwo Naukowe. Warsaw: Panstwowe

Toman, Jindch. 1998. Word syntax. In Andrew Spencer and Arnold Zwicky, ed., Handbook of r Morphology, pages 306321. Oxford: Blackwell Publishers. Williams, Edwin. 1981. Argument structure and morphology. The Linguistic Review 12:18114. Zwicky, Arnold M. 1985. How to describe inection. In M. Niepokuj, M. Van Clay, V. Nikiforidou, and D. Feder, eds., Proceedings of the Eleventh Annual Meeting of the Berkeley Linguistic Society, pages 372386.

25