1 Introduction

One of the most significant open debates about the constituency of NP centers around the question of whether the classifier forms a constituent with the noun (1) or the numeral (2).

The problem of choosing between the two views has been around for decades, and it has proven to be difficult to solve on empirical grounds. In a recent paper Bale et al. (2019) offer four arguments for the [[Num Cl] N] structure from Ch’ol, a head-marking V-initial Mayan language (Chiapas, Mexico). These are briefly previewed in (3).

Importantly, Bale et al. suggest that the way they test NP-internal constituency has applicability beyond Ch’ol as well: their “diagnostics can be used as a template to test the constituency structure in other classifier languages” (p. 1). The goal of this paper is to demonstrate that Bale et al.’s diagnostics are insufficient for distinguishing between [[Num Cl] N] and [Num [Cl N]] even internally to Ch’ol: while all of their data are indeed compatible with [[Num Cl] N], these data can also be captured on [Num [Cl N]] with equal ease. In prior work (Dékány 2021; 2022a; b; to appear) I have supported the universal [Num [Cl N]] hypothesis, so readers are referred to those works for empirical and conceptual arguments for [Num [Cl N]]. In this paper I will not show that the structure of the Ch’ol NP must be [Num [Cl N]], or that [[Num Cl] N] is not viable in this language. My aim is to show that in order to determine constituency, cross-linguistically as well as in Ch’ol, we have to go beyond the tests in Bale et al. (2019).

Before we delve into the discussion, a caveat is in order regarding the structural representation of nominal modifiers. N-modifiers, including the numeral and the classifier, have been approached in two ways. On the one hand, they have been viewed as occupying positions on the main projection line of the extended NP in the form of functional heads or specifiers, taking a projection of the noun as their compelment. On the other hand, they have also been analyzed as adjuncts to a projection of the noun. This paper is primarily concerned with the issue of constituency, which is logically independent of the adjunct/non-adjunct status of N-modifiers. In principle, in (1) the numeral can be either a (specifier of a) functional head or an adjunct, and both options are available for the classifier as well. Similarly, in (2) the constituent that comprises the numeral and the classifier can, in principle, be either an adjunct (as in Bale et al. 2019) or a specifier of a functional head. Since a commitment to the adjunct/non-adjunct status of N-modifies is not necessary for my present purposes, I will remain neutral on this issue. Where labels will be useful in trees, I shall use xNP for ‘extended nominal projection’. This is a cover-term for NP, NumP and DP, with no commitment as to the (non)adjunct status of the left-hand daughter. The abbreviations Num and Cl stand for the descriptive terms ‘numeral’ and ‘classifier’, respectively, and are not meant to be understood as ‘the functional head number’ or ‘the functional head classifier’. (When I will have to talk about ‘the functional head classifier’, I will write this out in full as ‘the functional head Cl⁰’.)

This paper is organized as follows. In Sections 2 through 5 I go through Bale et al.’s (2019) tests one by one and demonstrate why they are not suitable adjudicators of NP-internal constituency. In Section 6 I tackle some further arguments for [[Num Cl] N] made on the basis of Ch’ol, and show that these arguments do not carry force either. I conclude in Sect. 7.

2 Phonological constituency

Bale et al. (2019: 17) argue that “the first piece of evidence” for the [[Num Cl] N] structure is the fact that “the numeral and the classifier form a single phonological word in Ch’ol”. This is illustrated in (4), where the hyphen indicates the phonological dependency between Num and Cl.¹

Bale et al. suggest that this is easily captured if the numeral and the classifier form a constituent. At the same time, they acknowledge that the alternative [Num [Cl N]] structure can also derive this fact, mentioning that earlier literature, specifically, Borer (2005), proposed that the classifier undergoes head movement to Num. Therefore the phonological facts do not, in and of themselves, constitute an argument for or against either structure.

While this much is acknowledged by Bale et al. (2019), phonological constituency still plays a major – albeit indirect role – in their analysis. This is because the logic of their paper centers around the fact that on Borer’s approach the numeral and the classifier form a complex head, while on the [[Num Cl] N] structure the numeral and the classifier form a phrasal constituent (possibly having a complex head inside). Taking this as their starting point, Bale et al. (2019) present three arguments to the effect that the numeral+classifier unit has phrasal status (these will be discussed below), and, in turn, this is taken to directly support [[Num Cl] N] and rule out [Num [Cl N]]. A major problem in this argumentation is Bale et al.’s tacit assumption that the [Num [Cl N]] structure can only capture the phonological facts of Ch’ol via Cl⁰-to-Num⁰ head movement, and so [Num [Cl N]] is inextricably linked to the complex head analysis. Bale et al. (2019) operate under the assumption that if in a particular environment the complex head analysis can be shown to be incorrect, then we automatically have an argument against [Num [Cl N]]. This is not the case, however. To be sure, syntactic head movement is one of the ways in which morphologically complex words can be derived. However, there are at least two other (non-lexicalist) approaches to such data as well.

On a well-known alternative to syntactic head movement as the (only) purveyor of morphologically complex words, nodes exponed by affixes can remain syntactically independent from their hosts throughout syntax, and may unite with the host, forming a complex head with it, only in the morphological component: via Lowering (Halle & Marantz 1993 et seq.) or post-syntactic Raising (Harizanov & Gribanova 2019). Similarly to syntactic head movmement, these operations are subject to strict locality (they can only target the next lower or next higher head, respectively), but they take place post-syntactically. On this approach parts of a morphologically complex word do form a complex head at some point, but not necessarily in narrow syntax. Therefore the syntactic unit containing the relevant parts might externally look like a constituent without underlyingly being a constituent.

Furthermore, it has been known since at least Marantz (1984; 1988) that mismatches between syntactic and phonological structures do occur, and parts of certain phonological words do not form a complex head either in narrow syntax or in the post-syntactic component. Among others, Marantz mentions the English auxiliary clitic ’ll (I’ll go now) and the Saxon Genitive ’s (the peacock in the park’s feathers) as examples where a morpheme forms a syntactic constituent with (and in the case of ’ll scopes over) material to its right, but it forms a derived word with material to its left (without being in a complex head or other constituent with it either in or after sytax). Such data have inspirted Marantz’s Morphological Merger. Marantz suggests that phonological adjacency is associative: a phonological structure such as (5a) can be rebracketed into (5b), if the (morpho-phonological) properties of the lexical items involved motivate this. More recent work in the framework of Distributed Morphology calls (5b) leaning (Embick & Noyer 2001; Lipták & Saab 2016).

(5b) means that in certain cases a syntactic structure will be linearized in such a way that the parts of a complex word end up in just the right configuration at PF for affixation to occur between them. The syntactic representation that will be mapped to such a linear order may be either derived or base-generated; this is immaterial. In the English examples above, the sentential subject and the possessor have arguably ended up in their surface position in front of the clitic via movement from the verbal and nominal thematic domain, respectively. Koopman & Szabolcsi (2000) et seq. have generalized this derivation to essentially all morphologically complex words, and eschewed syntactic head movement. In their approach a morphologically complex word Y-X is formed when a constituent containing Y as its last element moves to a specifier above X, such that there is no overt material between them (7). Y and X never from a syntactic constituent, but linearization makes them string-adjacent, and this can provide the bound morpheme with the host that it requires at PF.

We should emphasize, however, that leaning has also been been profitably applied to nodes that arguably stay in situ. This is the case, for instance, with the verbal complex in Chichewa and other Bantu languages, where the left-to-right order of prefixes directly corresponds to the order of functional heads in the clause (Cinque 1999; Mchombo 2004).

As argued in Cinque (1999), if examples such as (8) were derived by syntacic head movement (or post-syntactic Raising or Lowering), then we would expect the bound morphemes to be post-verbal and in the mirror order, contrary to fact. The attested order is therefore best derived via in-situ leaning.²

In addition to the English and Bantu examples mentioned above, cases of so-called suffix-stranding NP ellipsis also provide strong evidence that parts of phonological words do not have to form a syntactic (or post-syntactic) constituent. Suffix-stranding NP ellipsis can be observed in some agglutinating languages. It consists in ellipsis which deletes the noun but leaves some or all of the nominal suffixes unaffected. These stranded suffixes find a phonological host in the nearest overt NP-modifier (e.g., adjective, numeral or possessor) and get affixed to it at PF. There is no constituent at any part of the derivation that contains all and only the overt host and the suffix.

Consider possessive structures in Huallaga Quechua. The possessed noun bears agreement for the ϕ-features of the possessor and case appropriate for its role in the clause, while the possessor is marked with genitive case (9).

NP-ellipsis from this structure deletes the possessed noun and the agreement marker. The case affix is retained, however, and at PF the possessor serves as is host. This is an effect of linear adjacency rather than a matter of any particular narrow syntactic relationship.

Further cases of suffix-stranding NP ellipsis, with adjectival and numeral hosts, are illustrated in Appendix A. That section also demonstrates that the stranded suffix becomes part of the vowel harmony domain of its new adjectival, numeral or other host, and thus indeed forms a phonological word with it.

To summarize, parts of a morphologically complex word do not necessarily form a constituent either in syntax or in the morphological component; what is required is a syntax that will be linearized in such a way that these parts end up adjacent at PF. Applying Morphological Merger (5) to the case of the numeral and the classifier in Ch’ol (and other languages where the classifier is a suffix to the numeral), in a structure like (11) these two elements are exactly in a configuration that will lead to their adjacency at PF, where affixation can take place.

The preceding discussion has shown that in principle, phonological constituency between morphemes can arise in three different ways:

It is now prudent to ask if we can determine which of these theoretical possibilities apply to Ch’ol numerals and classifiers. Most previous work takes classifiers to correspond to a functional head in syntax, cf. Tang (1990); Cheng & Sybesma (2005); Simpson (2005); Jenks (2011); Ott (2011); Alexiadou & Gengel (2012); Mathieu (2012), to mention just a few papers covering various language families. In the majority of cases examined in the previous literature, this appears to be well-motivated (e.g., classifiers generally cannot be modified). In Ch’ol, however, this approach would be difficult to maintain. As discussed in Arcos López (2009: 41–52) and Bale et al. (2019), most Ch’ol sortal classifiers are derived forms: they have the shape C₁VjC₂, which is derived from corresponding C₁VC₂ transitive verbal or positional roots. (Due to a phonological restriction on adjacent fricatives, fricative-final C₁VC₂ roots maintain their C₁VC₂ from as classifiers, see Coon & Preminger 2009: n. 6.) The following examples illustrate the positional root koty ‘standing on four legs’ as a basis for the main sentential predicate (14) and for a sortal classifier (15).

Orthographic <j> is IPA [h], and there is disagreement in the literature as to whether it involves infixation of a consonant (Vázquez Álvarez 2011: 65–66), or it reflects a special vowel quality (a lengthened aspirated vowel) whose value is supplied by templatic morphology (Coon 2017b). In either case, a C₁VjC₂ form is the product of a structure that is bigger than just a root/head. This has important repercussions for how to represent Ch’ol classifiers in syntax, which was not discussed in previous literature.

There are two ways in which the complexity of classifiers could be modeled. A base-generated head-adjunction approach (16) would offer the possibility that either the transitive/positional root or <j> corresponds to the classifier head Cl⁰ projecting a ClP, with the other morpheme being a head-level adjunct to Cl⁰ and not projecting its own phrase.

Similar so-called conflation structures, with a non-projecting N being head-adjoined to a projecting V, have been proposed in the verbal domain by Haugen (2009) to account for denominal verbs such as ‘(to) hammer’ and by Mateu (2012) to analyze resultatives.

This cannot work in Ch’ol because neither transitive/positional roots, nor <j> is consistently associated to classifier structures. Transitive roots can give rise to transitive verbs and positional roots can give rise to stative predicates (14). <J> goes back to a Proto-Mayan nominalizer (Law 2020) and is used in contemporary Ch’ol as one of the strategies to form passives (Vázquez Álvarez 2011: Chap. 10.2, see (18)).

If neither the transitive/positional root nor <j> corresponds to the functional head Cl⁰, then (16) is unviable, and we must model the internal complexity of classifiers in a different way. This leads us to the second logical possibility, namely that Ch’ol classifiers are phrasal categories. A specific implementation of this idea can be found in Clausen (2023). Clausen suggests that the <j> of classifiers and passives is a nominalizer, and thus both classifiers and passive predicates are externally nominal phrases of sorts. I will adopt this line of thinking, and for expository purposes, I will use the label nP for structures with <j>.³ I would like to emphasize, however, that what is important for me is that Ch’ol sortal classifiers correspond to phrases; whether these phrases are nominal and what their exact label is is immaterial for the present discussion. I suggest that the phrase comprising the transitive/positional root and <j> merges in the specifier of ClP, and Cl⁰ itself is morphologically unrealized.

The phrasal nature of Ch’ol classifiers is supported by the fact that up until the middle of the 20th century they could be modified by PPs (see (95) in Sect. 6.2).⁴

This proposal means that cross-linguistically, languages vary as to whether their classifiers spell out the functional head Cl⁰ itself or they correspond to a phrase in the specifier of Cl⁰. That this variation exists should not surprise us: Dékány (2021) argues that cross-linguistically and even internally to Hungarian, demonstratives show the same two-faced behaviour. That is, some demonstratives spell out the functional head Deix⁰, while other demonstratives are phrasal categories Merged as specifiers of a phonologically null Deix⁰.

Let us circle back to the issue of phonological constituency between Num and Cl. We now have a simple explanation for the fact that the classifier and the numeral together do not have the distribution of a complex head: the classifier itself is phrasal to begin with. Notice that this does not correlate with NP-internal constituency in any way: (19) could, in principle, either first-Merge with NP ([Num [Cl N]]) or with the numeral ([[Num Cl] N]). We can now also answer the question of which option in (13) best characerizes Num-Cl phonological constituency in Ch’ol. Both syntactic head movement and post-syntactic head movement (Raising or Lowering) are excluded: these are operations on heads, and Ch’ol classifiers are phrasal constituents. This leaves Morphological Merger (leaning) as a model for the phonological constituency, and as we have seen, this is compatible with both [[Num Cl] N] and [Num [Cl N]] as the underlying syntax.

Summarizing this section, [Num [Cl N]] does not force us to assume that Num and Cl form a complex head in syntax. As a result, ruling out the complex head analysis does not provide an argument in favour of [[Num Cl] N]. In the following sections I will go through Bale et al.’s remaining three diagnostics in detail and show that a) they are unsuitable as adjudicators of the [Num [Cl N]] vs. [[Num Cl] N] debate, and b) the data they are built on are even compatible with the view that in languages where the classifier spells out the functional head Cl⁰, the numeral and the classifier form a complex head in syntax via head movement.

3 Coordination

Bale et al. (2019) observe that a syntactic unit containing the numeral and the classifier can undergo asyndetic disjunctive coordination.⁵

They suggest that this would be difficult to derive on an approach where Num and Cl form a complex head via head movement, and they conclude from this that the Num+Cl unit has phrasal status. This, in turn, is taken as evidence for a structure in which the numeral and the classifier form a constituent to the exclusion of the noun: [[Num Cl] N].

The underlying assumption behind this line of argument is that the relevant examples have a WYSIWYG representation: the coordinated constituent contains only the numeral and the classifier.

This is indeed a logically possible source for examples like (20) and (21). However, there are two alternative structures, not at all considered in Bale et al. (2019), which can derive the data equally well. Specifically, the data also fall out from a structure in which entire extended NPs (complete with a numeral, a classifier and a noun) are coordinated, and the noun is either elided from the first conjunct (23) or it undergoes extraposition from both conjuncts via Right Node Raising (24).⁶

On both of these analyses, the coordinated units are indeed phrases rather than complex heads, but critically, both are neutral as to whether the classifier forms a constituent with the numeral or the noun. The two possible structures on the first, elliptical approach are shown below. Notice that on both representations, it would even be possible to assume that in languages where the classifier spells out the functional head Cl⁰, the classifier undergoes head movement to and forms a complex head with Num, though this is not necessary to capture the data.

Importantly, NP/nP ellipsis in classifier expressions in independently attested in Ch’ol.

Moreover, NP/nP ellipsis is also possible specifically in the disjunctive coordination under consideration: in (29) there is no overt noun at all. That reference is made to people is recoverable from the human classifier tyikil.⁷

The structures in (25) and (26) predict that examples in which low NP-modifiers other than numerals and classifiers appear in either conjunct should be grammatical. (30) shows that this is indeed the case. Here both Num+Cl units are followed by a relative clause marked by the relativizing clitic -bä, but the crucial position is the one after the first conjunct.

Examples like (30) must definitely be derived from (25) or (26), with ellipsis or Right Node Raising, as there is no approach on which numerals, classifiers and relative clauses sit together in a left branch.⁸ But if (30) is not indicative of constituency between numerals, classifiers and relative clauses, then (20)–(21) do not supply an argument for [[Num Cl] N] either. This just reinforces the well-known thesis that since conjuncts can contain silenced material, coordination is a very unreliable test for constituency.

The examples above show that coordinating extended NPs/NumPs with ellipsis in the first conjunct or Right Node Raising is a possibility offered by Universal Grammar, moreover, this possibility is part of Ch’ol grammar as well. With these derivations in the toolbox, it becomes very difficult to exclude (23) or (24) as possible sources of (20)–(21). This, then, obviates the disjunction facts as an argument for the [[Num Cl] N] structure.

4 Complex numerals

Bale et al. (2019) suggest that complex numerals above 20 also provide an argument for [[Num Cl] N]. In the case of Ch’ol simplex numerals, the numeral is immediately followed by a classifier.

Ch’ol is a vigesimal language. 20 and its powers (i.e., 400, 8000) have a distinguished role in that they a) form the basis of complex numerals and b) occupy the classifier slot (with their multipliers sitting in the regular pre-classifier numeral slot.)

Since 20 and its powers are in the classifier position themselves, they block the occurrence of a ‘regular’ classifier. Compare (33), where ‘cow’ appears with kojty, the classifier for animals, and (35)–(36), where k’al, the root expressing 20 occupies the post-numeral slot and the appearance of kojty is ruled out.

This means that with ‘round’ numerals (i.e., multiples and powers of 20) the information about the shape, size or disposition of the noun that is normally coded in a garden variety classifier (33) is not present: the lexeme that is in the Cl slot is neutral as to whether the counted noun is human, animate, round and flat, long and slender, etc. In this respect, 20 and its powers resemble so-called generic or general classifiers (see Grinevald 2004). These “effectively suspend the classificatory function” while still filling the obligatory classifier slot (Law 2024: around 5’42”) and can be found both within the Mayan family and outside of it. Generic classifiers perform the function of connecting numerals and nouns the same way as more specific classifiers do (in the sense of Borer 2005, they portion out ‘stuff’); they are just more vague/generic in their lexical semantics than others. In many languages, generic classifies can replace more specific classifiers in NPs, similarly to what happens in (35).

Complex numerals that fall between two multiples of 20 employ a so-called overcounting pattern: they comprise a numeral in the 1–19 range and the next higher multiple of 20. For instance, 22 is expressed with the juxtaposition of 2 and 40, 36 involves the juxtaposition of 16 and 40, and 90 is coded by juxtaposing 10 and 100, etc.⁹

In these complexes the numeral falling between 1 and 19 is followed by a ‘regular’ classifier, one that matches the animacy, shape or disposition of the counted noun. Compare (38) and (40): the former is appropriate for nouns that are counted with p’ej, the general classifier, while the latter is used with nouns that take kojty(/koht), the classifier for animals.

The second numeral of the complex, on the other hand, is always a multiple of 20 (or one of its powers), so it has 20 (or one of its powers) in the post-numeral classifier slot regardless of what the counted noun is. This second numeral is preceded by a 3sg prefix (belonging to Set A, which is used to cross-reference ergative arguments on verbs, possessors on possessed nouns and the Ground on relational nouns, Coon 2013).

Bale et al. (2019) argue that since these complex numerals are multi-word expressions, they correspond to XPs rather than complex heads. This is then taken to directly support [[Num Cl] N] over [Num [Cl N]]. As shown in detail in Sect. 2, this conclusion does not follow: the complex head analysis is just one of the ways in which [Num [Cl N]] can model phonological constituency between numerals and classifiers (and it cannot be applied to Ch’ol, where classifiers are syntactic phrases). Therefore ruling out the complex head analysis does not mean that [Num [Cl N]] is also ruled out.

The internal structure of Ch’ol complex numerals is not investigated in Bale et al. (2019), as the Authors consider this topic to fall outside the scope of their paper. However, as the XP status of these numerals does not adjudicate the constituency in and of itself, a more detailed look at their internal structure is required. As shown above, Ch’ol overcounting numerals are built on the following abstract template:

On Bale et al.’s (2019) [[Num Cl] N] approach this simply involves a complex phrase whose internal make-up is indeed immaterial.

On [Num [Cl N]], the structure of these numerals can be approached in the same way as the coordinated Num+Cl sequences in Sect. 3: the two numerals and two classifiers on the surface are indicative of two Ns at deep structure. That is, overcounting numerals involve a syntactic relation (whose nature will be made more precise below) between two xNPs, both of which contain a numeral, a classifier and an noun. The two xNPs share the same listeme in the N slot, which allows one of the Ns to be deleted under identity (or to undergo Right Node Raising).

With both the first and the second classifier in a local relationship to a token of N, there is no argument for constituency either way: the data are compatible with both [[Num Cl] N] and [Num [Cl N]]. Moreover, in languages where the classifier corresponds to a functional head, both (43) and (44) are even compatible with the head movement approach to the phonological dependency between the numeral and the classifier.

The analysis in (43)/(44) receives support from Ionin & Matushansky (2006; 2018), who argue that (non-multiplicative) complex numerals, in fact, cross-linguistically involve operations between xNPs (such that each xNP contains an instance of the lexical noun), and components of the complex numeral do not form a constituent to the exclusion of the noun. Ionin & Matushansky (2006; 2018) suggest that additive numerals have the following, invariable structure across languages:

Cross-linguistic variation among additive complex numerals stems from the following parameters: i) whether the relationship of the xNPs is mediated by the functional head J(unction) or by an adposition (P), ii) how the mediating head is exponed (J: ø or and; P: locative (on) or comitative (with)), and iii) which of the identical Ns, if any, is deleted on the surface.

Evidence for (45) is furnished by languages which allow (or require) multiple instances of the lexical noun to be spelled out overtly. Such languages include Biblical Welsh, Luvale (Bantu), Archaic Russian (Ionin & Matushansky 2018: Chap. 5.2.1) and Ainu (Tamura 1988/2000). A Biblical Welsh example is provided below.

Further evidence for an instance of the lexical noun inside complex cardinals is provided by languages in which the counted noun appears after a conjunct other than the last one (Ionin & Matushansky 2018: Chap. 5.2.2). Examples of this pattern can be found in Biblical Welsh, Scottish Gaelic and archaic English (cf. threescore years and ten from Psalm 90:10 of the King James Version of the Bible). Compare (46) and (47):

Ionin & Matushansky (2018) argue that languages like (modern) English also build their complex numerals as in (45); where they differ from Biblical Welsh and similar languages is that the ellipsis of the noun in the first conjunct is obligatory.

Ionin & Matushansky (2006) argue that complex subtractive numerals (e.g., Latin 19: un-de-viginti lit. ‘one-from-twenty’) likewise involve a relationship between two full xNPs (48).

Subtractive numerals involve a P as a connector, and cross-linguistic variation again comes down to i) the exponence of P (caritive or ablative) and ii) whether the lexical N is ellipted from the subtrahend or the minuend (see Ionin & Matushansky 2006 for refinements). The noun can again be observed inside the complex numeral in some cases:

Ionin & Matushansky (2006) do not investigate the Ch’ol-type overcounting numerals. I suggest that similarly to additives and subtractives, these also involve a syntactic relationship between full xNPs, as in (43)–(44). As in the case of additives and subtractives, the components of the complex numeral do not form a constituent to the exclusion of the noun. Support for this position comes from theoretical and empirical considerations. On the theoretical side, this provides a unified representation for additives, subtractives and overcounting numerals: it would be difficult, if not impossible, to motivate why the first two involve relationships between xNPs while the last one has a different structure, with a relationship between just numerals. As for the empirical side, if the noun was represented just once, in its surface position after the whole complex numeral, then it would remain a mystery how the first classifier, deeply embedded inside a complex phrase (42b), could co-vary with the noun: see (38) and (40). This fact is straightforwardly captured, however, if the first classifier is also in a local relationship to an instance of the noun. Finally, overcounting numerals in Ainu (a Northeast Asian isolate with a vigesimal numeral system) provide explicit evidence for a nominal after the smaller numeral.

I conclude from the foregoing discussion that it is indeed the schema in (43)/(44) that underlies complex numerals in Ch’ol. This means that it is not possible to place any theoretical burden of constituency on these numerals: they can be easily accounted for on [Num [Cl N]] as well.

In order to give a full-fledged account of the syntax of Ch’ol overcounting numerals, both (42) and (51) need to address the questions of i) how exactly the two components of the complex numeral are connected in syntax, and ii) why the a3 morphology appears on the second numeral of the complex. Since these issues are, strictly speaking, independent of constituency, I will not take them up here. In Dékány (to appear) I provide a cross-linguistic survey of overcounting numerals and conclude that the most frequent overt connector they feature is a lative (‘toward’) adposition. In the same work I suggest that Ch’ol overcounting numerals are also connected by a silent, latively interpreted P (which undergoes P-drop), thus the Ch’ol numeral ‘twenty two’ is essentially ‘two towards forty’.

The plausibility of this analysis is supported by the fact that an overt P connector is also attested in overcounting numerals in the Mayan family: in 18th century Yucatec Maya t, a reduced from of the all-purpose preposition links the two parts of the complex numeral.

I account for the a3 morphology with a possessive structure, the particulars of which need not concern us here. I refer the interested reader to Dékány (to appear) for detailed discussion.

5 A-bar extraction

The final, and on the face of it, perhaps strongest argument for the [[Num Cl] N] structure in Bale et al. (2019) is that in Ch’ol it is possible to front the numeral and the classifier together to the focus position, leaving the noun in the postverbal domain.

Bale et al. (2019) make a strong case that examples such as (54) involve extraction from the Noun Phrase, as the fronting of the numeral and the classifier is subject to the same restrictions as the fronting of interrogative possessors, the latter being a well-known case of A-bar extraction. In the Tila dialect of Ch’ol, possessor subextraction is possible only out of internal arguments: the subjects of unaccusatives and the objects of transitive verbs (Coon 2009; 2013; Bale et al. 2019).

Extraction out of the subjects of transitive verbs (and, as expected, out of adjuncts), on the other hand, is banned.¹⁰

Num+Cl fronting obeys the same constraints: it is felicitous out of unaccusative subjects (60)–(62) and transitive objects (63)–(65).

Extraction out of transitive subjects in ungrammatical; here the only option is to move the entire Noun Phrase.

Based on these parallels, Bale et al. (2019) conclude that discontinuous NPs such as (54) involve extraction out of NP, and since Num and Cl undergo this movement together, they must form a constituent. This is then taken to support [[Num Cl] N] as the structure of NP over [Num [Cl N]].

Unassailable as this line of argument may seem, these data are far from being conclusive for the constituency of classifier expressions. Discontinuous Noun Phrases of the type in (54) constitute a case of Left Branch Extraction (Little 2020a), a phenomenon thoroughly investigated in Slavic languages. The similarities between Ch’ol and Slavic LBE are extensive: i) various types of NP modifiers such as possessors, numerals and adjectives may be extracted (though the extraction of possessors and adjectives is more restricted in Ch’ol than in Slavic and Ch’ol demonstratives don’t extract at all, see Little 2020a for discussion), ii) the fronted constituent receives a contrastive interpretation (focus in Ch’ol and Serbo-Croatian, contrastive topic or focus in Russian), and iii) the fronted constituent may contain more than one NP modifier. These properties are illustrated below for Russian.

Given the close correspondence between LBE in Slavic and Ch’ol, it is worth considering what could be profitably used from the extensive literature on the former in the analysis of the latter.

LBE in Slavic has received three different analyses: direct subextraction of the modifier from the NP (Ross 1967; Borsley & Jaworska 1988; Corver 1990; Bošković 2005), remnant movement of the NP after evacuation of the N head (Franks & Progovac 1994; Abels 2003; 2012; Bašić 2008) and scattered deletion (Fanselow & Ćavar 2002; Pereltsvaig 2008; Fanselow & Féry 2013; Bondarenko & Davis accepted). For Tila Ch’ol, scattered deletion can be safely excluded: there is nothing in the theory that could restrict its application to all and only complements. This restriction is straightforward to capture on a movement-based account, however. As pointed out in Little (2020a), the fact that Tila Ch’ol bans discontinuous NPs which correspond to specifiers can be easily understood in light of the fact that specifiers are islands in a variety of languages.¹¹ This thus leaves two options on the table for Tila Ch’ol: direct subextraction and remnant movement.

Direct subextraction of the left branch from [[Num Cl] N] yields (54) in a simple and straightforward manner. Bale et al. (2019) put aside the possibility of deriving (54) from [Num [Cl N]] with remnant movement in a footnote, citing two reasons why they do not find it a plausible account of their data. Firstly, they do not see independent motivation for the first step of the derivation which moves NP/nP out of the nominal phrase. Secondly, they suggest that the remnant movement approach cannot capture the parallels between Num+Cl movement and other types of subextraction (see (55)–(67)). I will take up these concerns in turn, and show that they can, in fact, be easily eliminated.

Let us start with the issue of what might motivate the evacuation of the noun from the extended NP (73).

As discussed in Bale et al. (2019) and Little (2020a), examples with full NP movement and with LBE are discourse-pragmatically different. While pied-piping the noun results in focus on the entire Noun Phrase, LBE focuses the quantity alone. Thus (75) is appropriate as a reply to ‘who arrived?’, while (76) can be used as a reply to ‘how many women arrived?’ or to correct a previous claim that one woman arrived.

The evacuation step in (73) is thus motivated by information-structural considerations. Specifically, it serves to background the noun and thus allows focus to be on the quantity only. (LBE of numerals in Russian and Serbo-Croatian has the same effect.)

Turning to the parallels between separated possessors ((55)–(58)) and separated numeral modifiers ((60)–(67)), these fall out from the remnant movement account without further ado. If the first movement that creates the remnant (73) takes the noun outside of the extended noun phrase, then the evacuating step is a case of subextraction from xNP, therefore the same constraints that apply to subextraction of other material (e.g., possessors) from xNP are expected to apply here. Since Tila Ch’ol disallows subextraction from specifiers, the evacuating step from the subject of transitive verbs will be blocked, and so the remnant cannot be created. On the other hand, if the landing site of the noun is somewhere high within the extended Noun Phrase, then it is the movement of the remnant to the focus position (74) that constitutes a case of subextraction from xNP, and the restrictions on subextraction are predicted to apply at this point. In both scenarios, it is fully predicted that specifiers and adjuncts require the whole xNP to be pied-piped to the left periphery.¹²

It is worth pointing out here that remnant movement has been profitably applied in the analysis of the Ch’ol extended VP as well in Coon (2010b) and Little (2020a).

To summarize, Bale et al. (2019) make a successful argument for subextraction being involved in the derivation of (54). They do not, however, make a convincing case for [[Num Cl] N] being the base-generated structure in the extended NP, as NP/nP subextraction from [Num [Cl N]] followed by remnant movement achieves exactly the same empirical coverage. Note again that for languages in which the classifier corresponds to a functional head, the derivation in (73)/(74) is fully compatible with Cl⁰-to-Num⁰ head movement in syntax.

Future work will no doubt uncover compelling arguments for one analysis over the other. But even if it turns out to be the case that the direct subextraction approach is on the right track, we should be careful with drawing conclusions for the underlying structure of the NP from the grammaticality of (54). This is because as shown by data across a variety of Slavic languages, it is possible for the LBE-fronted constituent to contain multiple elements in such a way that they certainly do not form a deep-structure constituent. This phenomenon has been termed as ‘extraordinary LBE’ (Bošković 2005). Illustrative examples are provided below.

This pattern is, of course, straightforwardly captured (in fact, predicted) under the remnant movement analysis, but proponents of the direct subextraction analysis of LBE have also found ways to address such data. These involve some sort of re-arrangement or restructuring of the constituency within the NP in one way or another before the extraction takes place (see Borsley & Jaworska 1988; Corver 1990; Radkevich 2010; Talić 2018; Martinović 2019 for various implementations of this idea). In either case, the constituent that undergoes A-bar fronting arises only at an intermediate step of the derivation.¹³ As the constituents targeted by LBE can be derived, LBE is not a reliable tool for probing into deep structure constituency in the NP.¹⁴ This, in turn, disarms a key argument for [[Num Cl] N] in Bale et al. (2019).

Before closing this section, it should also be noted that the general logic of my argument does not depend on the fronting movement being specifically a case of LBE. What for split in Germanic provides independent evidence that the fronted part of Discontinuous Noun Phrases is not necessarily directly extracted from the NP. Consider the following examples from German.

As Germanic languages do not allow LBE, the fronting of the bracketed constituent does not involve LBE. But as discussed in Abels (2003: Chap. 4); Leu (2008) and Leu (2015: Chap. 6), it cannot involve direct extraction in general, because was does not form a constituent with the adposition in the underlying representation. (See the referenced papers for a remnant movement analysis of these data.) This further strengthens the general point made in this section: there is more than one way to arrive at a split NP, and the fronted phrase may be a derived constituent.

6 Further arguments

In this section I turn to two additional arguments for [[Num Cl] N] that have been made in the literature on the basis of Ch’ol, and I show that these arguments do not carry force either.

7 Conclusions

This paper considered four arguments for [[Num Cl] N] from Bale et al. (2019). The argument based on phonological constituency was recognized as inconclusive already in Bale et al. (2019). Two further arguments, namely those based on disjunctive coordination and A-bar extraction, have no probative value. These tests are known to produce false positives cross-linguistically, thus they cannot be pressed into service to adjudicate controversial cases. As for the argument based on overcounting numerals, here everything depends on what kind of structure one adopts for complex cardinals in general. Overcounting numerals can be captured on [Num [Cl N]] with Ionin & Matushansky’s (2018) theory of complex numerals, developed independently of the issue of classifiers. The plausibility of this approach was supported with the shape of overcounting numerals in Ainu. An analysis along these lines is worked out in more detail for Ch’ol in Dékány (to appear).

At the end of the paper I also addressed a further constituency test, based on the mutual dependency between Num and Cl, put forth in Little et al. (2022a). I showed that the dependency between numerals and classifiers is (or at least until the 1950’s was) unidirectional rather than bidirectional: classifiers can occur without numerals, but numerals cannot occur without classifiers. I argued that numerals are prosodically dependent and need classifiers (or some other suitable host) for phonological reasons. This alone can explain the occurrence of classifiers in the counting sequence; no conclusions can be drawn for syntactic constituency.

In my previous work I supported a universal [Num [Cl N]] structure for classifier expressions. In Dékány (2021) I showed that the standard of comparison is determined compositionally from the semantic contribution of the noun and the classifier. This semantic constituency aligns with a [Num [Cl N]] syntax but produces a syntax-semantics mismatch on [[Num Cl] N]. In Dékány (2022a) I argued that so-called autoclassifier (aka repeater) constructions support the [Num [Cl N]] constituency, and in Dékány (to appear) I suggested that so does the semantic selection between the classifier and the noun.

In this paper, I did not show that the structure of xNPs in Ch’ol is [Num [Cl N]]. I also did not show that [[Num Cl] N] is definitely excluded for Ch’ol. The structure of classifier expressions, in UG as well as in specific languages, remains subject to an ongoing debate. What I hope to have shown is that going forward, we should not rely on the diagnostics proposed in Bale et al. (2019).

Appendix A: Suffix stranding NP-ellipsis

This appendix illustrates suffix stranding NP-ellipsis in more detail with data from Hungarian, and demonstrates that via this process numerals, too, can end up forming a morphological word with another morpheme with which they do not form a syntactic constituent. As shown in (97), numeral modifiers in Hungarian are prenominal and occur in the Dem > Num > Adj > N order. Participles can appear at multiple places within the extended NP; (97) features one in the immediately post-numeral position.

NP ellipsis may scope over all of part of the post-numeral domain, but it leaves the case suffix intact. If only the nominal head is elided, then the color adjective bears the case suffix that would be appropriate for the noun if it was overt (98). We can be sure that the case suffix belongs to the elided noun rather than the adjective because adjectives, numerals and participles do not show concord (97), and the case suffix can only appear on the NP-modifier which is linearly closest to the stranded suffix.

If NP ellipsis affects the noun and the color adjective together, then the case suffix is hosted by the quality adjective (99). If both adjectives are elided with the noun, then the case affix will appear on the participle (100), and if the entire post-numeral domain is elided, then it will be affixed to the numeral (101). All examples below can be understood to be elliptical versions of the full DP in (97), with progressively bigger and bigger structure undergoing ellipsis.

Note that the appearance and quality of the linking vowel on the stranded suffix is determined by its newly found host: the adjective in (98) and (99), the participle in (100) and the numeral in (101). That the stranded suffix is in the vowel harmony domain of its new host is unambiguous indication that they are part of the same phonological world. But there is no doubt that in (98)–(101) the suffix does not form either a complex head or any other type of syntactic constituent with its host: it is merely adjacent to the host at PF.

The same phenomenon can be observed with the plural marker: it is left behind by NP ellipsis, and it leans onto the last remnant preceding the elliptical gap. This is illustrated below with an adjective and a participle. (The plural marker is in complementary distribution with numerals, so numerals will never act as a host for this type of affix.)

Kester (1996) discusses similar data from the related North Sámi language. Outside of the Uralic and Quechua families (see Sect. 2), suffix stranding NP-ellipsis has been observed in Persian, Turkish and Basque (Lipták & Saab 2010) as well as the Northeast Caucasian languages Archi and Kubachi (Kibrik 1995).

Abbreviations

a: set A person marking (∼ ergative, genitive), acc: accusative, adn: adnumerative, b: set B person marking (∼ absolutive), cl: classifier, cop: copula, dat: dative, det: determiner, dir: directional, dtv: derived transitive, ep: epenthetic insertion, ext: existential, f: feminine, fut: future, gen: genitive, hon: honorific, imfv: imperfective aspect, ins: instrumental, itv: intransitive verb suffix, iv: status marker for intransitive verb in imperfective, loc: locative, nom: nominative, pass: passive, pfv: perfective, pl: plural, poss: possessive, prep: preposition, prfv: perfective, prog: progressive, prt: particle, pst: past, ptcp: participle, rel: relative, sg: singular, sp: Spanish borrowing, stat: stative suffix, th: theme vowel, tv: transitive verb suffix

Funding information

This work was funded by the János Bolyai Scholarship of the Hungarian Academy of Sciences, the ÚNKP-23-5 New National Excellence Program of the Ministry for Culture and Innovation as well as NKFIH grants KKP 129921 and FK 145985.

Acknowledgements

I am grateful to Jessica Coon for discussion regarding the distribution of C₁VjC₂ forms and to Lena Borise for providing the Russian examples. I also wish to thank Carol Rose Little, the reviewers, the handling editor and the audience of SinFonIJA 16 for their constructive feedback. All errors are mine.

Competing interests

The author has no competing interests to declare.