2.1 Theory-internal arguments

2.1.1 Number of specifiers

The first argument for the DP-hypothesis comes from examples like (3), where in addition to the possessor there is also a prenominal determiner present (such structures are limited in English but frequent in other languages, e.g., Hungarian, cf. Abney 1987: 270–276 and Section 2.2.2 below):

(3)	John’s every secret wish

Such examples posed a problem under the X′-theoretic assumptions of the Government-and-Binding era because it was assumed that heads only project one specifier (Abney 1987: 288–297). Since both the quantifier and the possessor have to occupy specifier positions of N, analyzing (3) as in (4a) was not an option.² No problems arise under the DP-hypothesis where the possessor occurs in the specifier, while the determiner occupies the D-position, cf. (4b):³

(4)	a.
	b.

Importantly, the restriction is no longer adhered to in more recent work within the Minimalist Program (though see antisymmetric approaches where for reasons of linearization, heads also take just one specifier, which is treated as an adjunct, cf. Kayne 1994). Even in languages that do not have overt multiple wh-movement to the same functional projection it is standardly assumed that at least v can take several specifiers: Next to external Merge of the subject, an additional specifier is projected when objects undergo successive-cyclic movement and make an intermediate movement step via SpecvP:

(5)

Several specifiers of v are also often posited for languages with free word order; for German, for instance, it has been argued that multiple scrambling can target different specifiers of v, see Heck & Himmelreich (2017).

One may object that under the DP-hypothesis, the word order restrictions, i.e., the fact that the possessor precedes the determiner and the determiner the adjective, follow more straightforwardly than under the NP-hypothesis: The specifier of D precedes the head D and D precedes adjuncts adjoined to NP. However, since there are also languages where the possessor can follow the determiner (see below on Hungarian), this does not strike me as a very strong point. In that case, one has to postulate another functional projection below D which can host the possessor. Under the NP-hypothesis, one will have to treat determiners and possessors as optional arguments that have to be introduced in a certain order. Thus, while under the DP-hypothesis the order results from the functional sequence (and possible movement operations of the possessors), under the NP-hypothesis, the order follows from ordered selectional features on N. The fact that D-elements precede adjectives need not be stipulated under the NP-hypothesis but follows from the semantics (see also Bošković 2009): D-elements close off the noun phrase by mapping a predicate onto an individual. The reverse combination is simply impossible. Note also that semantic considerations are inevitable even under the DP-hypothesis given that adjectives also occur in a preferred order according to a semantic hierarchy, see Scott (2002). Thus, unless one adopts a cartographic approach, the DP-hypothesis will also have to resort to semantics to account for some ordering restrictions within the noun phrase.

2.1.2 Distribution of D-elements and consistent selection

The next theory-internal argument comes from selection and the distribution of D-elements. First, verbs can be combined with both nouns/noun phrases and pronouns. Since pronouns belong to category D, consistent selection by the verb favors treating noun phrases as DPs (Abney 1987: 74, 281–284). Importantly, pronouns are sometimes reanalyzed as determiners taking a silent NP-complement (Elbourne 2005), cf. (6a). However, once parts of the noun phrase can be silent, one can reanalyze pronouns as NPs headed by a silent N that take a D-specifier (or D-complement under Bare Phrase Structure), cf. (6b) (unless there is an independent restriction that only maximal projections can be silent, see Section 2.3.1 below; cf. also Abney 1987: 278–281):

(6)	a.
	b.

Essentially the same issue arises with demonstrative pronouns, which have the same distribution as noun phrases with demonstrative determiners. Capturing the identical distribution is straightforward under the DP-hypothesis; however, once we allow the head N to be silent, the distribution also follows under the NP-hypothesis, cf. (7):⁴

(7)	a.
	b.

The argument in this subsection crucially rests on the assumption that noun phrases and pronouns in the same language are of the same syntactic category. Therefore, it loses much of its force once it is possible for a verb to occur with nominal complements that can be shown to be different from DP. This is, for instance, the case if different types of pronouns correspond to different categories, as argued in Déchaine & Wiltschko (2002), who distinguish ϕP, nP and DP. Thus, if a predicate is compatible with different pronouns and thus different XPs, one might as well assume that in addition, it is also compatible with NPs. The issue also arises with noun phrases, where there is a debate about whether all noun phrases are necessarily of the same size within the same language. The case is particularly pressing with bare nouns, where the postulation of silent structure is contested. Once verbs can also occur with noun phrases of smaller size, e.g., NumP or even NP, the argument from consistent selection is further weakened. See Section 4.2.1.1 below for more discussion on selection of potentially different types of noun phrases.

2.2 Parallelism arguments

This subsection deals with several arguments that are based on parallelism between the verbal and the nominal domain. The argument usually goes like this: Given that we find X in the verbal domain/X is a functional head in the verbal domain, its counterpart in the nominal domain should also be a functional head. As we will see, while these arguments are not per se implausible, there is always an analytical alternative under the NP-hypothesis.

2.2.1 Agreement – on functional heads only

The first argument is based on the observation that in some languages we find possessor agreement on nouns (Abney 1987: 37–52), which in some languages additionally looks morphologically identical to agreement at the clausal level, viz. on the T-head. From this it is concluded that there must also be a corresponding functional head in the nominal domain. Consider the following data from Yupik Eskimo discussed in Abney (1987: 39–42):

1. (8)

1. Yupik

1.  
1. a.

1. angute-t
2. man-ERG.PL
1. kiputa-a-t
2. buy-OM.SG-SM.PL
1. ‘The men bought it.’

1.  
1. b.

1. angute-k
2. man-ERG.DU
1. kiputa-a-k
2. buy-OM.SG-SM.DU
1. ‘The men (du) bought it.’

1. (9)

1. Yupik

1.  
1. a.

1. angute-t
2. man-ERG.PL
1. kuig-a-t
2. river-SG-AGR.PL
1. ‘the men’s river’

1.  
1. b.

1. angute-k
2. men-ERG.DU
1. kuig-a-k
2. river-SG-AGR.DU
1. ‘the men’s (du) river’

The markers that reference agreement with the ergative argument in (8) are identical to those that reference agreement with the possessor in (9). Upon closer inspection, however, it becomes clear that the parallelism is not as perfect as it may seem at first sight: Crucially, the object marker in the verbal paradigm in (8) occupies the same slot as the marker which encodes the noun’s inherent ϕ-features in (9). If we were to interpret the parallelism in a very strict way, we would be led to the rather questionable conclusion that the possessee occupies the equivalent of an object position in the nominal domain (see Bruening 2009: 32 for the same argument based on Passamaquoddy). It is also important to stress that the perfect morphological parallelism shown in the examples above is by no means found in all languages with possessor agreement; quite often, the agreement markers in the verbal and the nominal domain are different, cf., e.g., Turkish.

Possessor agreement is usually handled by positing a DP with the D-head equipped with a ϕ-probe that targets the possessor in its specifier (the possessor is usually assumed to be extracted from a lower position, e.g., from Spec,nP or Spec,PossP, cf. Adger 2003). Movement of the noun/possessee via Poss to D then ensures that the agreement affix ends up on the noun:

(10)

However, these facts can also be captured under the NP-hypothesis once we locate the ϕ-probe on N:

(11)

Note that under current assumptions agreement no longer requires a spec-head relationship (as was presupposed at the time of Abney 1987). Rather, c-command between probe and goal at some point of the derivation is sufficient (for an application to case assignment in the nominal domain, see Georgi & Salzmann 2011). Consequently, the correct result obtains as long as the possessor is c-commanded by the noun at some point of the derivation; this requires possessors to originate within the projection of the possessee and movement of the possessor can no longer be linked to the Agree operation with D. The trigger for movement of the possessor and an analytical alternative will be discussed in the next subsection. As for projecting the possessor within NP, and even as a complement as in (11), this may seem unusual given that nowadays, possessors are usually projected in a functional layer above NP such as nP or PossP. But such layers are usually also largely motivated by parallelism (the possessor is treated as the equivalent of the external argument of the verb). From a purely empirical point of view, though, no problems arise if the possessor originates lower. I am not aware of any evidence, analogous to those at the clausal level (cf. external vs. internal arguments of verbs), that would show that the possessor must be generated outside the projection of the noun. As discussed in fn. 14, when several arguments and/or a possessor of N are present, they are projected in a fixed hierarchy. But nothing, to my knowledge, requires the distinction between projecting an argument within NP and outside of NP. Since restricting phi-probes to functional heads is a purely theory-internal assumption, nothing seems to preclude a reanalysis within the NP-hypothesis.⁵

Related to the agreement argument is the argument based on silent pronouns. It has been observed that possessor agreement, like agreement on T, licenses a silent pro. Consider the following pair from Hungarian, cf. Coene & d’Hulst (2003: 5):

1. (12)

1. Hungarian

1.  
1. a.

1. a
2. the
1. pro
2.  
1. kalap-unk
2. hat-POSS.1PL
1. ‘our hats’

1.  
1. b.

1. a
2. the
1. mi
2. we.NOM
1. kalap-unk
2. hat-POSS.1PL
1. ‘OUR hats’

In the absence of an overt possessor the inflection is interpreted as pronominal, cf. (12a). In the presence of an overt pronoun an emphatic interpretation obtains, cf. (12b). This is parallel to what is found in pro-drop languages at the clausal level, cf. the following pair from Italian:

1. (13)

1. Italian

1.  
1. a.

1. pro
2.  
1. Canta.
2. sing.PRS.3SG
1. ‘He/she sings.’

1.  
1. b.

1. Lui
2. he
1. canta.
2. sing.PRS.3SG
1. ‘HE sings.’

While there is indeed an obvious parallelism, the assumption that pro is only licensed by agreement on functional categories is a purely theory-internal assumption.

2.2.2 Possessors in different positions

At the clausal level there is solid evidence for two subject positions in many languages (cf., e.g., expletive there constructions in English; for more evidence, see, e.g., Hornstein et al. 2005: 80–90). Concretely, it is generally assumed that the subject is introduced in SpecvP and moves from there to SpecTP:

(14)

The following pair from Hungarian shows that possessors can occur both below and above the definite article; the alternation correlates with differences in case-marking: the higher possessor bears dative, while the lower occurs in nominative case (cf. also Abney 1987: 275):⁶

1. (15)

1. Hungarian

1.  
1. a.

1. a
2. the
1. Mari
2. Mari.NOM
1. kalap-ja
2. hat-POSS
1. ‘Mari’s hat’

1.  
1. b.

1. Mari-nak
2. Mari-DAT
1. a
2. the
1. kalap-ja
2. hat-POSS
1. ‘Mari’s hat’

We thus have evidence for two different surface positions for possessors in the noun phrase.

Given the parallelism with the clausal domain, it seems straightforward to argue in favor of movement of the possessor from a lower position closer to the possessee to a higher specifier position associated with a functional head, e.g., SpecDP. In fact, things are more complex in that the possessor in the lower position precedes other noun phrase internal material such as numerals, quantifiers and adjectives in Hungarian. Thus, both surface positions of the possessor are derived. Assuming that the possessor is merged in Spec,PossP, it will then either move to the lower Spec, e.g., Spec,AgrP, or the higher one, viz., Spec,DP, as in (16):

(16)

However, strictly speaking, the Hungarian data above only show that the dative possessor is structurally higher than the definite article which in turn is higher than the nominative possessor. They do not show that the possessors have to occupy the specifiers of a higher head. Moving the dative possessor from a lower complement position within NP, e.g., the complement of N, to the highest specifier of N is not a priori ruled out:

(17)

Admittedly, triggering such a movement may be somewhat difficult, but one could adopt a rule that assigns to dative-assigning nouns an EPP-/edge feature that triggers movement to the highest specifier of N.⁷

Alternatively, it is conceivable that possessors can be base-generated in different positions within NP depending on their case-feature. This will be sufficient for the possessor to receive a theta-role from N. One may regard an approach where the order is exclusively handled by means of ordered selectional features of N as ad hoc. However, note that much of the ordering could follow from general principles such as a hierarchy of features and semantics. One should bear in mind that under the DP-hypothesis, the order of functional heads also has to be stated somewhere, e.g., by means of selectional features on the relevant functional heads. As far as I can tell, the major difference would be that the different surface orders of the possessor are derived by movement under the DP-hypothesis and by base-generation under the NP-hypothesis (in the version that does not employ the movement in (17)). Note that I am not aware of any empirical evidence that would clearly show that a possessor in a high position has to have been merged in a low position and moved to its surface position. In the clausal domain, however, floating quantifiers or scope ambiguities (e.g., w.r.t. negation) can provide evidence that the two subject positions are indeed related by movement. Thus, while perhaps unusual, it is not clear to me that an account that largely involves base-generation is necessarily inferior.

2.3 Constituency arguments

The arguments in this subsection (from Adger 2003: 247) show that the noun forms a constituent together with certain modifiers, to the exclusion of D-related elements. This is crucially different from showing that D is the head.

2.4 Head-movement evidence

The data discussed in this section show that there has to be another head-position above the base position of the noun. As we will see, this does not necessarily imply that this head is to be identified with D or some other functional head.

The literature contains quite a number of arguments for head-movement within the noun phrase, one of the first being put forward by Ritter (1988), who posited N-to-D-movement to account for the NSO-order in Hebrew construct state nominals; see, e.g., Carstens (2017) for a recent overview. Since the facts are quite well-known, I limit myself to just one example:

The following Italian data show that the determiner precedes the possessive adjective, cf. (25a/b). While the possessive adjective normally precedes the noun, cf. (25b), the order is crucially reversed if there is no overt D-element, cf. (26a vs. b).

The alternation between (25b) and (26b) is consequently argued to result from movement of N to D:⁹

(27)

The data indeed suggest that there must be a head-position above the projection of the noun. However, even if we accept this conclusion,¹⁰ this need not necessarily imply that there is a DP-layer on top of NP. Such noun phrases could be smaller than DP, e.g., corresponding to nP or NumP, see, e.g., Danon (2006) for arguments to this effect for Hebrew construct state nominals. A more radical alternative involves reanalyzing (certain kinds of) head-movement in terms of reprojection where the moving head projects rather than the target, see Georgi & Müller (2010) and references cited there. The basic idea is that heads that undergo movement have a probe feature linked to a selectional feature that cannot be checked in-situ because the relevant constituent is a specifier and thus not c-commanded. In (28), N has the probe feature [*D*] requiring Agree with a DP under c-command. Reprojection is a last resort mechanism that moves the head to a position where it c-commands the relevant constituent, viz., D and thus can discharge its probe feature. Under this assumption, a DP-layer is no longer needed, NP is sufficient (see Georgi & Müller 2010: 19 for discussion of how to capture the dependency of N-reprojection on D being null):¹¹¹²¹³

(28)

2.5 Two remaining arguments

Before concluding this section, I will briefly mention two further arguments for the DP-hypothesis that do not fall into the categories above, viz., gerunds and asymmetries between languages with and those without articles. Both are interesting arguments but are either based on a poorly understood phenomenon (gerunds) or only constitute an indirect argument.

As for gerunds, Abney (1987: Chapter 3) proposes that gerunds involve a DP embedding a VP as in (29):

(29)

This captures both the verbal properties (accusative case, modification by adverbials) as well as the gerund’s external distribution of a noun phrase. Furthermore, the absence of an N-layer accounts for the impossibility of adjectival modification. This would seem like an interesting argument for the DP-hypothesis since replacing D by a silent nominalizing N-head would probably predict – incorrectly – that modification by adjectives should be possible. However, many aspects of the phenomenon remain puzzling. First, Abney (1987: 197) presents examples from the older literature where gerunds do co-occur with adjectival modification. Ackema & Neeleman (2004: 176, fn. 27) suggest that parsing factors could be responsible for this change: once the gerund is encountered after an adjective, it is preferentially parsed as a noun; a following DP-object will require reanalysis. Second, if D is the head of gerunds, it is somewhat surprising that determiners are impossible, at least in English gerunds. Third, mixed nominalizations in other languages, e.g., in Dutch or German, are not subject to this modification restriction. Both an adjective or an adverb are possible, suggesting that nominalization can take place at different levels; note that Ackema & Neeleman (2004: 172–182) explicitly argue in favor of a silent nominalization affix in such constructions. See also Alexiadou et al. (2007: 477–546) for detailed discussion of the various proposals that have been advanced. Given these uncertainties, I prefer not to draw any conclusions from the gerund facts.

One of the most interesting though indirect argument for the DP-hypothesis comes from the work of Bošković (2009) on syntactic asymmetries between languages with articles and those without. He shows that the two types of languages differ in systematic ways from each other. For instance (the list is considerably longer), languages without articles allow left-branch extraction and extraction of adjuncts from NP, while languages with articles do not. The basic idea is that these asymmetries can be best accounted for if the languages have a significantly different noun phrase structure: Languages without overt articles have no D-layer whatsoever, i.e., noun phrases are NPs, while those in languages with articles are DPs. I find the basic logic of the argument convincing, even though it is not fully trivial to relate all the properties Bošković lists to D. However, in principle, it is also conceivable that the mere presence vs. absence of an element (viz., the article) has drastic syntactic consequences; this may hold irrespective of whether it is a head (DP-hypothesis) or a specifier (NP-hypothesis). To make a concrete example, the ban against left-branch extraction in languages like English is related to anti-locality in Bošković (2005): An adjective that attempts to extract from DP has to move via SpecDP; however, this movement step is too local because it fails to cross a maximal projection (by assumption, the adjective starts out as an adjunct to NP and thus only crosses a segment of NP and D′). A reformulation under the NP-hypothesis does not seem impossible: Assuming, with Bošković (2016), that only the highest specifier of N constitutes an escape hatch, in a language like English, an adjective would have to move from a lower specifier of N to the highest one, above D-elements. This movement step could plausibly also be ruled out under a certain definition of anti-locality. In Serbo-Croatian, given the absence of definite articles, adjectives can be directly merged in the highest specifier of N and therefore are extractable. The crucial difference between English and Serbo-Croatian would thus not be the headedness of the DP but the presence vs. absence of D-related elements in the left periphery of the noun phrase. Needless to say, this should not be considered an attempt to reanalyze all the asymmetries discussed in Bošković (2009). But if they are to be reanalyzed within the NP-theory, the differences would have to follow from the presence of D-related elements rather than a difference in headedness. Because of such reservations and the fact that the approach does not directly tell us anything about the headedness of languages with D-elements, I will not discuss this perspective any further. See also Bruening et al. (2018: 37–38) for critical discussion; in Section 4.3 below, I will provide an argument that noun phrases in Serbo-Croatian are in fact DPs.¹⁴

2.6 Intermediate summary

It should have become clear that the arguments in the literature in favor of the DP-hypothesis are not as strong as one would expect given its success in the last thirty years. All of the phenomena reviewed can arguably be handled under the NP-hypothesis as well. Since this will sometimes require somewhat non-standard assumptions that some syntacticians might reject, the DP-hypothesis may still be at a certain advantage, but as far as I can tell, none of the phenomena are such that they necessarily require the DP-hypothesis. Thus, the phenomena reviewed so far cannot in my view be used to decide between the NP- or the DP-hypothesis. Note also that none of the arguments surveyed so far actually refers to standard headedness criteria. I will address arguments of this type in Section four below and show that some of them do provide more insight into the headedness of the noun phrase.¹⁵

3.1 Noun incorporation

In noun incorporation D-related constituents like demonstratives, numerals and possessors can be stranded, see Baker (1988: 93–97). (30a) is the baseline example, (30b) illustrates stranding of a demonstrative, (31a–c) illustrate stranding of a relative clauses, a numeral and a possessor, respectively (the verb rakv ‘be white’ is unaccusative):

The structure of a noun incorporation example like (30b), which is usually assumed to involve head movement, looks as follows under the NP- and the DP-hypothesis, respectively (see Assmann et al. 2012 for a recent account based on the NP-hypothesis):

(32)	a.
	b.

Under the DP-hypothesis in (32b), the head-movement constraint (Travis 1984) is violated as the D-head has been skipped. No problems obtain under the NP-hypothesis, where the demonstrative is located in a specifier of N. One possibility to uphold the DP-hypothesis is to assume that all overt D-related elements are specifiers of D and that noun incorporation involves N + a silent D (alternatively, N-to-D movement could be followed by excorporation of N, but excorporation is often ruled out on principled grounds, cf. Baker 1988):

(33)

Such a reanalysis under the DP-hypothesis is not impossible, but at the same time it seems nearly impossible to falsify. What makes the stranding examples independently problematic is the fact that in the relevant languages, the putatively stranded elements are actually well-formed independent constituents that can occur outside noun incorporation contexts and can be roughly translated by means of NP-ellipsis, see Rosen (1989). Consequently, much more careful investigation of such ‘stranding’ cases is necessary before any conclusions can be drawn, since the stranding argument is probably the strongest one for a syntactic approach to noun incorporation and against a lexicalist one (i.e., one involving compounding).

Note also that it has been argued that noun incorporation may involve incorporation of more than just the root. Barrie & Mathieu (2016: 27–31) observe that incorporated nouns often bear categorizer morphology and argue that noun incorporation therefore must involve phrasal movement rather than head-movement. I think it is fair to say that the proper analysis of noun incorporation is still very much a matter of debate.¹⁶

3.2 Possessive determiners

Müller (2007: 85–87) provides an argument for the NP-hypothesis on the basis of possessive determiners in German as in (34):

If the possessive determiner were the head of the noun phrase, Müller argues, we would expect the DP to carry its referential index and not that of the noun, contrary to fact. Since the possessive pronoun is visible for binding (cf. his picture of himself), one cannot simply identify its index with that of the noun; the problem is thus real. Müller therefore concludes that the possessive pronoun must be projected as a specifier of N. An alternative may consist in projecting the possessive determiner in SpecDP with D remaining empty. This would avoid projecting the wrong index up to DP.

The solution with a silent D will, however, no longer work once the possessor doubling construction is taken into account, which can be found in all German dialects and in Colloquial German. Here, the possessive pronoun is preceded by a dative possessor:

In this construction the referential index of the possessor is arguably not borne by the possessive determiner but by the dative possessor, viz. dem Hans. Consequently, the possessive determiner can be analyzed as an agreement element as in languages like, e.g., Turkish which have systematic possessor agreement (in addition to dependent marking on the possessor; cf. also the examples from Yupik in (9) above):

For the version without an overt possessor in (34) above, one can assume that, as in languages like Turkish, there is a small pro in SpecDP. However, this is still not sufficient to ensure that the referential index of the noun is present on the maximal projection of the noun phrase if this is mediated via D. One possibility is to decompose the possessive determiner into a possessive and a D-part (cf. sein-e ‘POSS.3SG.M-F.SG’). The possessive part would be merged as a lower head within DP and then moved to D, which carries the referential index. The possessor could be initially merged in the specifier of the possessive determiner and then moved to Spec,DP.

Given these assumptions, possessive determiners are no longer problematic under the DP-hypothesis. Let me add, though, that the possessor doubling construction is, of course, certainly compatible with the NP-hypothesis. For more detailed discussion of these issues, see Georgi & Salzmann (2011).

4.1 Head-complement asymmetry

The head-complement asymmetry is manifested in two areas: Heads select their complements rather than vice versa and heads often determine properties of their complements. As we will see, while the prediction as such is clear, the result will be somewhat inconclusive for our debate.

4.1.2 Form determination

It is well-established that heads determine the form of their complements. This can be seen in case marking and morphological selection. With respect to morphological selection between verbs (also referred to as status government), it is generally the case that V_n determines the form of V_n+1, i.e., the governing verb determines the form of the immediately dependent element, as, e.g., in (40) (from Bruening 2009: 30):

(40)	I might have been being handed some cocaine (when the police caught me).

Furthermore, it is usually the functional (auxiliary, modal, restructuring) verb that determines the lexical verb rather than the other way around (a possible exception being auxiliary selection, but the choice of auxiliary is often determined by several elements, including, e.g., directional PPs, not just the lexical verb):

(41)	I broke/was breaking/have broken/want to break the vase.

Bruening (2009) argues that there are asymmetries within the noun phrase suggesting that N is the head rather than D. Indeed, when we look at nominals the reverse situation seems to obtain: It is the noun that determines the form of other constituents inside the noun phrase. For instance, the count/mass distinction affects the choice of modifiers, cf. (42):

(42)	a.	too many/*much people
	b.	too much/*many rice

While this may not constitute proper agreement but rather represent some semantic compatibility requirement based on the count/mass distinction, concord within the noun phrase is a clear case where the form of determiners and adjectives depends on features of the noun. This holds at least for those features that are inherent to the noun; this can be number as in the case of pluralia tantum, cf. (43a) or gender, cf. (43b):¹⁸¹⁹

1. (43)
1. a.

1. these/*this scissors

1.  
1. b.

1. Latin
1. puer
2. boy.M
1. bon-us/
2. good-M
1. puella
2. girl.F
1. bon-a
2. good-F
1. ‘good boy/good girl’

The argument from concord is interesting, but it must be pointed out that it is crucially different from morphological selection or Case assignment: It involves the copying/sharing of (inherent) features rather than the assignment of features to other constituents that the head/governor itself does not bear. Note, for instance, that a verb that assigns accusative is not accusative itself, i.e., there is no morpheme that would identify it as an accusative assigning element. If there were feature sharing, one might expect verbs to have special morphology depending on the case they assign, but that is not what one finds. Similarly, verbs that select a certain type of infinitive do not have any morphological property that would signal this (they may belong to a different class, e.g., restructuring vs. modal, but that is orthogonal). In concord, the noun will usually have a morphological exponent that expresses the feature, e.g., the plural-s ending above.

I conclude from this that the relationship between D and N is crucially different from that between the verbs in the sentence in (40) above. Therefore, in my view the concord data do not provide any conclusive evidence that N is the head rather than D.²⁰

Another possible argument on the basis of form determination can be made by means of adjectival inflection, which in Germanic languages is affected by the choice of D. I will illustrate the argument on the basis of German, where attributive adjectives display so-called strong and weak inflection. Synchronically, this inflection is not related to definiteness (unlike in some of the other Germanic languages). Rather, the inflection is determined by the form of the determiner. Simplifying somewhat, a zero determiner or a determiner without an ending selects the strong inflection, while a determiner with an ending selects the weak inflection:

1. (44)

1. German

1.  
1. a.

1. (ein)
2. a
1. rot-er
2. red-STR
1. Wein
2. wine
1. ‘a red wine’

1.  
1. b.

1. ein-em
2. a-DAT
1. rot-en
2. red-WK
1. Wein
2. wine
1. ‘to a red wine’

1.  
1. c.

1. rot-em
2. red-STR
1. Wein
2. wine
1. ‘to red wine’

This clearly suggests that D is the head relative to material in its c-command domain. The dependency can be captured straightforwardly under the DP-hypothesis, where the determiner will be a head that c-commands the adjective. This represents a standard configuration for Agree. Under the NP-hypothesis, where the determiner occupies Spec,NP, one will have to assume that phrasal elements can as probes (under Bare Phrase Structure, the determiner would be [+max,+min]). We thus obtain the structures in (45a) and (45b) (of course, under Bare Phrase Structure, the adjectival modifier would be merged as a complement under the NP-hypothesis):

(45)	a.
	b.

The force of this argument to some extent depends on how concord on adjectives is handled. If they are adjoined as in (45a) and attributive adjectives probe for features of N, the probe will be on a non-head. The configuration would thus be very similar to that in (45b), where the probe determining adjectival inflection is also on a non-head, viz., the determiner. Thus, the argument from adjectival inflection only works if we assume that form determination is crucially different from concord, viz., that only projecting heads can determine the form of other elements (as is the case in the verbal domain, recall example (40)). Finally, since D does not interact with N, the phenomenon strictly speaking does not tell us anything directly about the relationship between D and N; but given that D displays head-like behavior w.r.t. form determination, it may be considered an indirect argument in favor of its being the head of the noun phrase.²¹

4.2 The features of the head are present on the maximal projection

This fundamental property of heads has the consequence that it is the features of the head that are visible to the syntactic context outside the noun phrase. Features of other noun phrase-internal constituents, on the other hand, are much less accessible.

This can have the following (well-known) syntactic effects: First, the head determines the distribution of the phrase because its category label is visible on the projection. Second, the head is the preferred element for higher heads to interact with. This is seen in the fact that the head/features of the head are the preferred target for selection and agreement. Furthermore, given the A-over-A principle, the features of the head will be closer to probes outside the noun phrase and can therefore potentially block interaction of outside probes with constituents inside the noun phrase.

We have already discussed the distribution argument in Section 2.1.2 above and came to the conclusion that no compelling argument could be made for either the NP- or the DP-hypothesis. In the following (sub)sections, I will discuss possible evidence based on selection and agreement that capitalizes on the fact that the preferred target for noun-phrase external probes is the (features of the) head.

4.2.2 Agreement asymmetries and intervention

Given that the features of the head of the noun phrase are present on the maximal projection, we expect the (features of the) head to be the preferred goal for probes outside the noun phrase.

Suppose that a functional head such as v/T targets the noun phrase for subject/object agreement. One expects it to preferably target (features of) the head because that constitutes the closest goal. Furthermore, the head may block access to features of lower heads; this may be related to phasehood if the highest head of the noun phrase is a phase head or more generally to Relativized Minimality since the head of XP c-commands/the XP dominates (A-over-A-principle, cf. Chomsky 1973) lower heads/projections. The two different analyses of the noun phrase make crucially different predictions in this respect: Under the DP-hypothesis features of N may be inaccessible, while under the NP-hypothesis, things would be expected to be the other way around: Features of D may be inaccessible (material that may be inaccessible is set in gray):

(50)	a.
	b.

In the argument to be presented below, noun phrase-internal concord will play an important role. For this we need to make certain assumptions about the locations of features within the noun phrase. What follows is a certain simplification because for each of the relevant features within the noun phrase it has been proposed that they reside in separate functional heads (see, e.g., Danon 2011 for discussion). For my purposes, the following simplification will not cause any problems: I will assume that both gender and number are inherent features of N, while person and definiteness are features of D (recall that a comparison of the DP-hypothesis with the NP-hypothesis is only possible if at least some features originate on N). When a verb shows agreement with features of the noun phrase that originate from different noun-phrase-internal heads, there are two main analytical possibilities: Either the head (v/T) can probe several times and target different heads of the noun phrase or the head (N or D) ‘collects’ the features; consequently, it is sufficient for v/T to target only the head of the noun phrase and copy everything from there. In the derivations below, I will assume that determiners, demonstratives, adjectives (and possibly other noun-phrase internal constituents) bear probe features and that the head collects the relevant features.²⁸

4.3 Hybrid agreement

We now turn to the empirical argument in favor of the DP-hypothesis. It is based on agreement with hybrid nouns in BCS (Bosnian-Croatian-Serbian), more specifically, on the possible switches between grammatical and semantic agreement within the same clause. The facts will show that the features of demonstratives are more accessible to outside probes than the features of N. This strongly suggests that N is not the head.

In many languages agreement can target both grammatical as well as semantic features of nouns. For instance, in English, the verb can show plural agreement with collective nouns like committee that denote a plurality. In the data to be discussed below the relevant feature is gender. Nouns denoting human entities often have grammatical gender features that diverge from their natural/biological/semantic gender. Most interesting are so-called hybrid nouns that can in principle agree both in semantic and in grammatical gender. Even more interesting is the fact that in some languages, different probes within the same clause (e.g., D vs. v/T) can access different types of features of the noun, i.e., some may target the grammatical and some the semantic feature. Interestingly, both semantic agreement as such cross-linguistically as well as hybrid agreement within the same clause (i.e., different probes targeting either grammatical or semantic gender features of the noun) are governed by a hierarchy:

Relevant in the present context are the effects of the hierarchy in the same clause: It allows for agreement switches from grammatical to semantic agreement. Languages can have different cut-off points where they switch from grammatical to semantic agreement, and the cut-off point may be flexible even within a given language. For instance, languages can also switch within the attributive domain, viz., between A and D or even between different adjectives. Importantly, once there has been a switch to semantic agreement in a sentence, one cannot switch back to grammatical agreement, as schematically shown in (52) (A2 is a lower, A1 a higher adjective, V is the participle used in compound tenses; ‖ indicates an agreement switch):

The following example from BCS involves the noun vladika ‘bishop’, which is grammatically feminine, but its semantic gender is masculine. We observe a (optional) switch between D and V as in (52a) (Puškar 2018: 304):

An example from BCS with a switch between A and D (52b/d) is given in (54) (Zorica Puškar, p.c; see Corbett 2006: 206 for a similar example).

It should be pointed out that there appears to be some speaker variation w.r.t. the possible switches and cut-off points, especially within the noun phrase; for instance, BCS does not seem to readily allow a switch between different adjectives, while Russian does, see Pesetsky (2013: 37–38). What is sufficient for our purposes, though, is that D can agree in either feature.

Note that the effects of the agreement hierarchy are not restricted to gender. They are also found with number, as discussed for Lebanese Arabic in Pesetsky (2013: 45–49), and for Hebrew in Landau (2016). In what follows, I will base the argument on gender, but nothing really hinges on this. I will adapt the analysis proposed in Puškar (2017; 2018) to my present concerns (this will be a significant simplification in many details and cannot do justice to the many aspects addressed in Puškar’s work). Let me stress at this point that there are, of course, different ways of distributing phi-features in the noun phrase and capturing the relationship between them (e.g., by means of feature hierarchies). The assumptions and the analysis based on them presented below have largely been chosen for reasons of concreteness and, more importantly, because they allow a clear comparison between the predictions of the NP-hypothesis and those of the DP-hypothesis. I thus do not mean to imply that the following analysis is necessarily the only and best way to tackle agreement and concord within the noun phrase. To simplify the discussion, I will in what follows ignore number agreement and focus exclusively on gender agreement, even though there is an interesting interaction between the two in BCS.²⁹

The analysis is based on the following assumptions: First, N has features for both grammatical and natural gender (Wechsler & Zlatić 2000). Second, natural gender is featurally more complex than grammatical gender as it also includes animacy. The two features will be represented as follows:

A probe (A, D, V) that is searching for a gender feature can in principle target either gender. For this to be possible, the probe can be of variable complexity (so-called relativized probing, see, e.g., Nevins 2007; Béjar & Řezáč 2009; Preminger 2011; probes are rendered as star-features, cf. Heck & Müller 2007).

Furthermore, the following assumptions hold for Relativized Minimality in the interaction of probes and goals:

• a complex probe can probe past a simple feature to find a complex feature below it (Relativized Minimality: a subset does not intervene)

• a simple probe cannot probe past a complex feature to find a simple feature (Relativized Minimality: a superset intervenes)

• a simple probe cannot be valued by a complex feature, which leads to a crash (feature copying requires matching of all features)

Importantly, the choice of probe is free, i.e., all heads that probe for phi-features (A, D, V) can enter the derivation with either a simple or a complex gender probe. As we will see, this does not lead to converging derivations in all cases.

Note that many of the assumptions above are familiar from relativized-probing based approaches to person hierarchy effects. First, the preference for agreement in a certain value is captured by both the complexity of the feature and the complexity of the probe (cf. local person, which is often analyzed as consisting of [+participant] and [+person]). Second, a goal that has only a subset of the probe’s features cannot satisfy it (cf. 3rd person, which cannot value a probe specified for [participant]). What is different from classical cases of relativized probing for person as in Béjar & Řezáč (2009) is that (i) a goal that has a superset of the features of the probe cannot value the probe (unlike a first person argument that can value a probe specified for participant) and that (ii) the gender-probe can differ in complexity. These differences are a consequence of the different workings of gender agreement in languages with hybrid agreement. The resulting system in fact also bears a certain similarity to the version of Relativized Minimality put forward in Starke (2001), where only supersets but not subsets intervene in movement operations.

The relevance of the complexity of probe and goal and their interaction is illustrated in the following two tree diagrams (simple probes and simple gender features are set in gray):

(57)

(58)

In both derivations, the adjective first agrees with the noun and copies a gender feature from it, the grammatical one in (57) and the biological one in (58). Then, the D-head probes. (57) illustrates that a complex probe can search past a simple feature. In this case we get a switch from agreement in grammatical gender to agreement in biological gender. (58) illustrates that a simple probe cannot search past a complex feature. In this case, the derivation crashes because D’s feature cannot be valued (recall that since the biological gender feature contains more structure than the simple probe could accommodate, probing fails and the derivation crashes). A consequence of this is that once there has been a switch to biological agreement, the derivation will only converge if the next higher probe is complex. This captures the generalization that once there has been a switch to biological gender, there is no way back. (59) illustrates how a complex gender feature on D probes and finds the corresponding feature on A (after A has agreed with N):³⁰³¹

(59)

The tree diagram in (60) shows the successful derivation (52b) of (54) where the verb agrees in biological gender features:

(60)

The probe on V is complex and thus targets the biological gender feature on D. Probing past D is blocked by Relativized Minimality, thus ensuring that a switch back to grammatical gender is blocked. In the alternative derivation (52d) for (54) illustrated in (61), the probe on V is simple:

(61)

The simple probe on V cannot target the complex feature on D as it cannot accommodate both features of D. Because of Relativized Minimality, however, it cannot probe past D either, with the consequence that the probe on V remains unvalued and the derivation crashes. The only converging derivation in this constellation involves a complex probe on V as illustrated in (60), resulting in biological agreement. A switch back to grammatical agreement is correctly ruled out.

Importantly, this result only follows under the DP-hypothesis because under the DP-hypothesis, D is closer to V than both A and N. Under the NP-hypothesis one would expect the reverse result: Because the features of N are present on the maximal projection, they would always be closest to outside probes like V. Consequently, one expects switching back and forth between grammatical and biological gender to be possible, contrary to fact. Since N is the closest goal for all probes in the tree, the value that obtains after Agree between A/D and N has no influence on Agree involving V – unlike in the derivation based on the DP-structure, where the earlier Agree operation affects the possibilities of the later. The tree in (62) illustrates the derivation of the ungrammatical example (52d) above with a switch to biological gender on D and a switch back to grammatical gender on V:

(62)

Hybrid agreement is thus a phenomenon that can be used to tease apart the NP- vs. the DP-hypothesis. Note that as discussed above, it primarily shows that N is not the head of the noun phrase but some other head such as the demonstrative. Whether one wants to call this DP is an orthogonal issue. Note that the workings of hybrid agreement thus also provide evidence against approaches like that in Bošković (2005) which treat the noun phrase in article-less languages like BCS as an NP.

4.4 A reanalysis within the NP-hypothesis?

Obviously, the argument in favor of the DP-hypothesis proposed in the previous subsection is only relevant as long as the facts cannot be reanalyzed within the NP-hypothesis. If it turns out that a plausible alternative is available, the argument will have the same status as the (putative) arguments for the DP-hypothesis discussed in Section two.

Bruening (2019; 2020) argues that it is indeed easily possible to reanalyze the agreement switches under the NP-hypothesis (I am discussing both the published version and a previous draft version since they nicely show what is at stake). In Bruening (2019), previous work on agreement switches like Matushansky (2013); Pesetsky (2013); Landau (2016) is adapted. The basic assumptions are the following: The hybrid noun starts out with a grammatical gender feature only. A biological gender feature can be introduced at any point within NP. This can be done in various ways, either by directly adding it to the projecting head (it is not shown how this can be done), by introducing it via a feature on the higher probes (i.e., adjectives, determiners, verbs etc.) as in Matushansky (2013) or by a designated functional head that bears only this feature as in Pesetsky (2013: 35–50) or by a Num-head as in Landau (2016), who discusses semantic number agreement. It is further assumed that the biological gender will be the only accessible gender feature. In Matushansky’s approach, this follows from the assumption that the biological gender feature is interpretable, while the grammatical one is uninterpretable (and by assumption a higher grammatical gender feature cannot combine with a lower biological gender feature). In the other approaches it follows under the assumption that the biological gender feature is structurally higher than the grammatical gender feature. Under these approaches, the position of the switch will thus depend on the height of the biological gender feature: If it is introduced by the determiner (Matushansky (2013)/between adjective and determiner (Pesetsky 2013, Landau 2016), probes below the determiner agree in grammatical gender, while the probes above it agree in semantic gender.

The tree diagram in (63) presents the reanalysis of the agreement involving the noun vladika ‘bishop’ within the NP-hypothesis using the mechanism proposed in Pesetsky (2013). The functional head introducing the biological gender feature is labeled ‘m’.

(63)

This leads to agreement in grammatical gender on the adjective and to biological gender on D and V, viz., the grammatical version (52b) of (54).

Let me discuss this alternative. It should first be pointed out that the approaches Bruening bases himself on adopt the DP-hypothesis potentially containing further functional projections. In my view this makes a non-trivial difference. At least in Landau, where Num projects to NumP and takes NP as its complement, this implies that the biological gender feature will automatically be a feature of NumP and thus of a higher projection than NP. The fact that no switch back follows naturally. Under Bruening’s reanalysis, however, this is not the case, it must crucially be the non-head that imposes the biological gender feature on the head.³² In other words, under the NP-hypothesis, the agreement switch and the effects of the agreement hierarchy can only be captured if a well-established notion of headedness is given up; the approach is certainly incompatible with Bare Phrase Structure (Chomsky 1995), where the head and its projections share all features and their values. There are further arguments one can raise against these approaches and their implementation irrespective of the NP/DP debate: Under Landau’s approach, to capture a switch between D and T, the gender feature would have to be introduced in a functional head above D, implying that noun phrases would not be DPs in this case but FPs/GenPs. A consequence of Matushansky’s approach is the curious fact that the semantic gender feature is only found on elements where it is arguably not interpretable; this is especially clear for a semantic gender feature introduced by a verb. Finally, in all approaches, although natural masculine gender is an inherent property of these hybrid nouns (at least of nouns like ‘bishop’), it is introduced by a separate head, which implies that some extra selectional relation (or at least some semantic restriction) must be established to ensure that this feature is only added to the right type of noun. Thus, it seems there are quite a number of reasons why one may want to prefer not to pursue this approach to agreement with hybrid nouns, not the least as it effectively seems to require a different theory of syntax, viz., one that gives up endocentricity.

Arguably for such reasons, Bruening (2020) pursues a different analysis: Phi-features on nouns are structured. Crucially, there is a higher level where phi-features are unvalued. This level dominates a level of valued features. In the case of gender, this will be grammatical and semantic gender features as in (64) for a noun like vladika (both features are thus on N):

(64)

He further assumes that in principle, a probe on some higher head (A, D, V) can access either feature. However, once there has been agreement in semantic features, the gender feature on N is valued as semantic, and further Agree operations cannot probe past the valued feature; this derives the effects of the agreement hierarchy. This proposal is certainly less problematic than the previous one in that the semantic gender feature is introduced on the head it is an inherent feature of. Since the valued gender feature is present on higher projections of N, the analysis is also more endocentric. However, at the same time, it is also clear that an extra level of feature structure has to be introduced for this to work, and the valued gender feature essentially plays the role of a separate functional head above N whose function is to ensure the effects of the agreement hierarchy. None of this seems to be needed anywhere else in the grammar. Thus, eventually, it seems to me that this kind of proposal primarily highlights the difficulties that the phenomenon presents for the NP-hypothesis.

Still, while perhaps unorthodox, it is certainly a possible theory of syntax. Does this mean that the argument in favor of the DP-hypothesis introduced above is void? I think not. Theories can only be compared as long as certain assumptions are held constant, and this is what I have attempted in this paper. Once in principle anything is possible, comparison is no longer possible and no interesting arguments for any theory can be made.