From N to Num
3.3 Fine-tuning the position of classifiers
This section refines the structure in (17) and proposes a more fine-grained functional sequence for the lower portion of the DP. The data and analyses in this section emerged in joint work with Aniko Csirmaz as D´ek´any and Csirmaz (2010). This section summarizes our joint work, and I share all credit for it with Aniko. I complement and go beyond the results of our collaboration in Sections 3.4 and 3.5.
3.3.1 Multiple classifier positions
Borer’s functional sequence DP > # >Div> N provides only one position for classifiers. However, it is known that some languages can have more than one classifier in a single DP. This is typical of Mayan languages, for instance. The Akatek Mayan example in (33) contains two classifiers, glossed by Zavala (2000) asnumclf for numeral classifier andsort.numclf for sortal numeral classifier. The Jacaltec Mayan, the Minangkabau (Austronesian), the Persian and the Newar (Tibeto-Burman) examples in (34) through (37) also contain two different classifiers. The Akatek example in (38) contains as many as three classifiers: one glossed asinanim, one glossed as cl and another glossed asnouncl.9,10
(33) kaa-(e)b’
‘two small plums’ (Zavala, 2000, p. 123, ex. 16. b.) Akatek Mayan
(34) caw-a¨n
‘two men’ (Craig, 1977, p. 127, fn. 1.) Jacaltec Mayan
(35) tigo
‘three Toona Sinensies (mahogany species)’
(Marnita, 1996, p. 84., ex. 4.20) Minangkabau
(36) pænj
‘five books’ (Gebhardt, 2009, p. 269., ex. 100.) Persian
(37) ta-gwa-gu
big-cl:2d-cl:inan bal ball
‘a big ball’ (Hyslop, 2008, ex. 64.) Newar
(38) ’ox-eb’
‘three half-folded tortillas’ (Zavala, 2000, pg. 127) Akatek Mayan
Borer’s decomposition cannot accommodate more than one classifier, therefore it needs to be amended. Svenonius (2008a) adapts Borer’s theory just enough to meet the data in (33) through (38), without introducing radical modifications. On the basis of Mayan data, Svenonius proposes that there are altogether three classifier positions in the DP. He suggests that so-called noun classifiers (c.f. Chapter 2, Section 2.2 and ixim in (38)) are harboured in n. So-called sortal classifiers (classifiers that typically sort nouns by shape and size) sit in SortP, which is the equivalent of Borer’s DivP (compareb’ilan in (33) andkupan in (38)). Finally, Svenonius suggests that the head of #P is also a classifier-related position (though he calls it UnitP): it houses so-called unit classifiers (classifiers that name the unit that is counted, like eb’ in (33) and (38) and a¨n in (34)). In essence, Svenonius does not introduce new classifier projections11, rather he suggests that several projections that already exist in Borer’s structure are able to host classifiers. The decompositions of Svenonius (2008a) and Borer (2005) are aligned in (39), the projections that
9Borer (2005) argues that the English plural functions as a classifier. If this is on the right track for the Jacaltec and Akatek Mayan plurals as well, then (33) and (34), too, contain three classifiers.
10For further examples, see Aikhenvald (2000) and Grinevald (2004).
11Modulon, which Borer does not discuss but probably assumes tacitly.
the authors acknowledge as classifier-related are bolded. My own labels for these projections are provided in the last row of (39).
(39) Svenonius (2008a) Dem>Art>Unit Cl>Sort Cl>n>N
Borer (2005) D > # > Div > N
labels used here Dem> D > Num > Cl >n>N
My ClP is thus equivalent to Svenonius’ Sort ClP and Borer’s DivP (as well as Zhang’s 2011 UnitP), and my NumP is equivalent to Svenonius’ UnitP and Borer’s #P. I chose these labels in an attempt to avoid proliferation of terminology and provide maximal transparency with regard to what sort of elements these projections host.
The data (33) through (38) make it necessary to have several functional projections that can host classifiers. Svenonius (2008a) can capture the data quite elegantly, placing the various clas-sifiers into independently motivated functional projections. I will term his proposal for the three classifier-related heads as the ‘Svenonius hierarchy’. The existence of multiple classifier-related functional projections raises the question of which classifier position(s) Hungarian classifiers lex-icalize. To probe this question, D´ek´any and Csirmaz (2010) test the ordering of classifiers with respect to adjectives.12
3.3.2 Adjectives and the Svenonius hierarchy
I repeat the Svenonius hierarchy with my labels in (40). The functional heads that can host classifiers are marked in bold.
(40) Dem > D >Num>Cl>n> N
It is uncontroversial that non-idiomatic, non-focused adjectives are base-generated betweennand Num. There are three logical possibilities as to how they line up with respect to the Cl position.
(41) Adj > Cl (42) Cl > Adj (43) Adj > Cl > Adj
D´ek´any and Csirmaz (2010) point out that Borer (2005) does not discuss adjectives, but her proposal has clear predictions for the ordering of classifiers and adjectives. Specifically, adjectives that can modify undivided‘stuff’ and adjectives that are normally understood to modify divided
‘stuff’ (atoms or units) are predicted to occur in different contexts, along the lines of (43).
Adjectives that can modify‘stuff’ without that ‘stuff’ being packaged into units are predicted to be grammatical independently of the presence or absence of a dividing structure. Adjectives of color, material and nationality, for instance, are such adjectives: they are predicted to occur both in mass and count structures without a change in meaning. Borer makes no predictions for the height of these adjectives. They could, in principle, occur either below or above classifiers.
Adjectives that are normally understood to modify divided‘stuff’ (units, atoms), on the other hand, are predicted to have a more constrained distribution. Adjectives of shape, height, length and other physical dimensions are such adjectives. In the presence of a dividing structure these adjectives are predicted to be grammatical. A further prediction is that when they modify units, these adjectives are merged above classifiers. This follows from compositionality: first the divisor must create units, and only then can these units be modified. In the absence of a dividing struc-ture, however, no units are established and so these adjectives cannot receive an interpretation of modifying units. Depending on how flexible the grammar is, without a divisor these adjectives are either ungrammatical or they are interpreted as modifying types. No prediction is made for the ordering of different adjective classes within the two types.
12I will depart from both Svenonius (2008a) and D´ek´any and Csirmaz (2010) on one point, however: I will assume that adjectives are merged in specifiers of functional projections specialized for adjectives. Svenonius (2008a) assumes that adjectives are in the specifiers ofnand ClP, while D´ek´any and Csirmaz (2010) take no explicit stand on how adjectives are introduced into the syntax.
3.3.3 The position of specific classifiers
To test the ordering of adjectives with respect to classifiers, D´ek´any and Csirmaz (2010) use Scott’s (2002) adjective sequence in (44). D´ek´any and Csirmaz (2010), however, do not commit to the existence of all these separate projections (admittedly, more could be proposed), they merely use it for convenience because it distinguishes a lot of adjective classes.
(44) ordinal > cardinal > size > length > height > speed > width > weight > temperature >
wetness > age > shape > color > origin > material
D´ek´any and Csirmaz test the adjective classes in (44) with classifiers one by one. The complete set of their data is reproduced in Appendix II of this chapter. They find that adjectives chosen from the beginning of (44), specifically between size and weight, neutrally precede classifiers. Two examples are given below.
(45) k´et
The classifier > adjective order, marked here with an asterisk, is not entirely ungrammatical.
Instead, it is ruled out on a unit-modifying interpretation of the adjective, and requires a different context. Specifically, speakers report a type reading for these orders.13 D´ek´any and Csirmaz (2010) note this fact but do not discuss it further. I will return to this issue in Section 3.4.
The pattern changes to the reverse one at temperature adjectives: from here on, classifiers precede the adjective.
‘two cold slices of sausage’
(50) *k´et
‘two cold slices of sausage’
(51) k´et
With these adjectives, the orders marked with the asterisk are very hard to get, even with a change in the interpretation. I can only imagine them with a very strong, contrastive stress, one that could indicate that the adjective in question has moved to a DP-internal focus position (Truswell, 2004;
Scott, 2002; Aboh, 2004b; Giusti, 2005; Svenonius, 2008a). But even with this prominent stress, I highly prefer these adjectives to stay behind the classifier.
To summarize the results, in Hungarian classifiers divide the adjective space into two, yielding an Adj group 1 > classifier > Adj group 2 order. Their neutral surface position is between weight and temperature adjectives.14
(53) ordinal > cardinal > size > length > height > speed > width > weight >classifiers >
temperature > wetness > age > shape > color > origin > material
13This is similar to how certain adjective orders in English are marked with an asterisk on account of not being the neutral order. Althoughblack big car is usually marked as * or *?, it does not mean that the order does not exist at all. Rather, it requires a special context and focus on the first adjective. Hungarian specific classifier >
adjective orders in (45) and (48) are also * in the sense of not being the neutral order.
14I note here that D´ek´any and Csirmaz (2010) check adjectives and classifiers pairwise, but do not test whether the strict ordering among the adjectives in (44) actually holds. I remain agnostic about how many adjective projections there are per adjective zone, as nothing hinges on the exact number. My tentative feeling is that the relative order of adjectives in Hungarian is very similar to that in English, and that Scott’s hierarchy is mostly right for Hungarian but the order may not be as rigid as he claims.
These findings confirm the predictions of Borer’s theory to a large extent: adjectives that modify bounded units are found above classifiers. It also looks like that most ‘mass adjectives’
are found below classifiers. The odd one out might be speed adjectives, but in this case, too, one might argue that speed is generally a property of animates, and thus requires units. The only real puzzle is the position of shape adjectives. Based on compositional semantics, these are predicted to be merged above classifiers, and there is no obvious reason, syntactic or semantic, why they actually surface below classifiers. D´ek´any and Csirmaz (2010) leave this as an open question.15 (54) k´et
two szem cleye
kerek round
rizs rice
‘two round grains of rice’
(55) *k´et two
kerek round
szem cleye
rizs pearl
‘two round grains of rice’
The Svenonius hierarchy, repeated for the reader’s convenience as (56), provides three possible positions for classifiers.
(56) Dem > Art >Num >Cl > n> N
The fact that Hungarian adjectives can be surrounded by adjectives on both sides makes them compatible only with the Cl position. They cannot be in the Num head because numerals sit in the specifier of NumP, and that would leave no position for the adjectives that precede the classifier and follow numerals. They cannot be inn either, as adjectives below this position are idiomatic (e.g. cold war, c.f. Truswell, 2004 and Svenonius, 2008a), and adjectives following classifiers are not idiomatic.16
(57) is a modified version of the Svenonius hierarchy. It incorporates D´ek´any and Csirmaz’s (2010) observations about the ordering of classifiers and adjectives and marks the position of specific classifiers by a frame around the Cl position.
(57) Dem > Art >Num > Adj > Cl > Adj >n > N
3.3.4 Evidence from compositional semantics
D´ek´any and Csirmaz (2010) also support the Adj group 1 > Cl > Adj group 2 decomposition with arguments from compositional semantics. They show that both the kind of stuff that is divided up and the kind of division imposed on the stuff must be known before the semantic contribution of relative dimensional adjectives can be computed. This means that dimensional adjectives that modify units (as opposed to kinds) can only be merged after the classifier. This line of argument thus makes use of the same kind of neat syntax-semantics mapping I am assuming here.
The argument goes like this. Certain classifiers encode dimensionality information. For in-stance, the classifier szem is associated with small spherical objects, while fej is used for big spherical objects. But when we evaluate the dimensionality of a noun modified by such a classifier, it is not enough to know what sort of classifier is involved. The size of theszemunit, for instance, is evaluated differently depending on whether the noun it combines with ismustard seed orcoconut.
15Note that shape adjectives are special among dimensional adjectives not only because of their unexpected position, but also because they are the only intersective dimensional adjectives. This raises the possibility that, as Truswell (2004) and Svenonius (2008a) suggest, Div cuts the adjective sequence into two along the subsective vs. intersective divide. However, if (53) is a base-generated order, then at least age adjectives are misplaced in the sequence. Age is computed relative to a standard of comparison: and old olive tree is exponentially older than an old hamster. The two adjective zones are thus not perfectly aligned along the subsetive/intersective bifurcation either.
16This is not to say that adjectives following classifiers cannot be idiomatic. (i) is an example of an idiomatic adjective following a classifier. The point is that they not have to be (c.f. (ii)), therefore classifiers must be higher thann.
(i) egy one
szem cleye
z¨old-alma green-apple
‘one green apple (characteristically green when ripen, for instance Granny Smith)’
(ii) egy one
szem cleye
z¨old green
alma apple
‘one green apple (an apple that happens to be green but may change color later)’
(58) egy one
nagy big
szem Cleye
must´ar-mag mustard-seed
‘a big mustard seed’
(59) egy one
nagy big
szem Cleye
k´okuszdi´o coconut
‘a big coconut’
Obviously, a size that is big for a mustard seed is not necessarily big for a coconut. Thus adjectives preceding the classifier do not modify the classifier directly: [[ dimensional Adj; classifier ] noun ] is a wrong representation.17
Conversely, some nouns can be partitioned in more than one way. The nounsz˝ol˝o ‘grape’, for instance, can be divided into grains or plants.
(60) a. egy one
szem Cleye
sz˝ol˝o grape
‘a grain of grape’
b. egy one
t˝o Clstem
sz˝ol˝o grape
‘a grape plant’
Ifsz˝ol˝o‘grape’ is modified by a dimensional adjective likebig, it is necessary to know which way the grape is divided. Something that is big for a grain of grape is not necessarily big for a grape plant. Thus in order to evaluate what counts asbig, small, etc. with respect to a particular noun, the partitioning imposed on the stuff must also be known. This means that adjectives preceding the classifier do not modify the noun directly either.
This leaves just one possibility: that adjectives preceding the classifier modify the combination of the noun and the classifier. In the kind of syntax-semantics mapping assumed here, this means that these adjectives can only be merged after a classifier. Thus the correct representation is (61).
(61) [ dimensional Adj [ classifier [ noun ]]]
Of course, the role of the noun in (58) and (59) has long been noted, and it is generally discussed under the rubric of‘standards of comparison’. Mustard seeds have a different standard of comparison than coconuts, therefore the same dimensional adjective will have a different effect on these nouns. This much is acknowledged by everybody. The novelty lies in showing that the partitioning also plays a role in what the standard of comparison is. In other words, the standard of comparison is compositionally computed from the combination of the noun and the classifier. Relative adjectives require a standard of comparison, the standard of comparison for units presupposes a division of‘stuff’, therefore relative adjectives can only come on top of the classifier.
3.3.5 Earlier work on the interaction of adjectives and the count/mass distinction
Similar conclusions about the interactions of a divisor, the count vs. mass distinction and adjectives have been reached independently in Muromatsu (2001); Truswell (2004) and Svenonius (2008a).
Muromatsu (2001) treats adjectives as adjuncts, and suggests that their ordering is a reflection of the mass vs. count distinction. Specifically, she proposes that count structures properly contain mass structures, and adjectives select either a count or a mass structure as their adjunction site.
Thus adjectives selecting for a mass structure can be found both in mass and count phrases, but adjectives selecting for a count structure can only be found with count phrases. This is similar to D´ek´any and Csirmaz (2010) in predicting a different distribution for the two adjective classes.
D´ek´any and Csirmaz (2010), however, do not suggest that adjectives literally select for a projection they want to combine with.
Svenonius (2008a) proposes that adjective ordering restrictions follow from the different mod-ification possibilities ofnand SortP (Borer’s DivP, my ClP). He suggests that modification of n is intersective, and adjectives merged tonP must be, just likenP, non-gradable predicates. Mod-ification of SortP (Borer’s DivP, my ClP), on the other hand, is subsective, and only gradable adjectives are compatible with this level. In short, Svenonius predicts the subsective, gradable >
intersective, non-gradable adjective order. D´ek´any and Csirmaz (2010) are mostly interested in testing the predictions of Borer (2005). Borer’s theory predicts that adjectives cleave into‘count
17See also Hundius and K¨olver (1983), who argue that in Thai adjectives do not modify classifiers.
adjectives’ and‘mass adjectives’ on the basis of whether they presuppose units or not. This is not the same partitioning as the subsective/intersective or the gradable/non-gradable divide. ‘Mass adjectives’, for instance, can be both gradable (green) and non-gradable (wooden), and they can be both intersective (moroccan) and subsective (old).
Closest in spirit to D´ek´any and Csirmaz (2010) is the discussion of adjectives and divisors in Truswell (2004). Truswell proposes that adjectives are adjuncts, and they can occur both betweenn and Div and above Div in Borer’s DP-structure. He suggests that certain adjectives (like size) cannot appear without Div, and so they cannot appear with mass nouns. Like D´ek´any and Csirmaz (2010), he suggests a semantic reason for this, viz. that these adjectives presuppose delimited objects. He further discusses the gradability of adjectives and the subsective/intersective bifurcation.
None of these works, however, actually test their predictions on a classifier language, where the effect of divisors and the mass vs. count distinction is the most visible, and where the different heights proposed for adjectives could translate into a detectable linearization effect (provided that no movement masks the base-generated order). To my knowledge, D´ek´any and Csirmaz (2010) is the first to do this.
3.3.6 The position of the general classifier
Hungarian has a general classifier: darab.18 Darab literally means‘(whole) piece’, but it could also be rendered in English as specimen (c.f. Germanein St¨uck Auto).19 This classifier can combine with any noun that is traditionally categorized as ‘count’. In unpublished work, D´ek´any and Csirmaz observe that darab has a different distribution than specific classifiers. Specifically, all kinds of adjectives followdarab. The following two examples with a relatively high and a relatively low adjective illustrate this point. Their complete set of examples is reproduced in Appendix II of this chapter.
The surface generalization aboutdarab is that it precedes all adjectives and it is right-adjacent to the numeral.20 D´ek´any and Csirmaz conclude that in contrast to the other classifiers,darab sits in the Num head. (66) shows this position on the modified Svenonius hierarchy.
The surface generalization aboutdarab is that it precedes all adjectives and it is right-adjacent to the numeral.20 D´ek´any and Csirmaz conclude that in contrast to the other classifiers,darab sits in the Num head. (66) shows this position on the modified Svenonius hierarchy.