1 The data: Agreement and Interpretation

In this section I introduce the two basic dimensions in which languages differ with respect to D[N1&N2] constructions. I start by laying out the core concern of this article, i.e., the two morpho-syntactic agreement patterns. Subsequently, I introduce Heycock and Zamparelli’s notion of joint- and split-readings, showing that this concept and its parametrization account for the distribution of number mismatched cases. While the morpho-syntactic phenomenon is completely accounted for by the language family, the semantic parametrization is independent of it.

2 The proposal in a nutshell

My proposal aims to derive both the two illustrated morpho-syntactic patterns as well as the reported differences in interpretation from one common, ‘What you see is what you get’ structure (Heycock and Zamparelli 2000; 2005) using only the featural specification of the functional head Num° (no gender in Romance, gender in Germanic, number ([Pl]) in both language groups) to model the cross-linguistic picture. The analysis will consist of the following “ingredients”:

1. In order to model the three-part dependency between the determiner and the phi features of the noun in the morphologically resolving languages of the sample I adopt an unrestrained version of multiple Agree (Hiraiwa 2001) that is syntactically solely constrained in terms of closeness (agree is always with the closest goal Chomsky (2000)) and relativized phi completeness Danon (2011) (a probe can only agree with a goal that carries a subset of the features it needs). The actual “work” in this approach is done by the interplay of the interfaces.

   Multiple Agree is free to apply in any of the languages but requires a syncretic form for spell out and / or must be interpretable for the semantics.

2. An independently needed modification of the semantic underpinnings of Heycock and Zamparelli (2000; 2005) in the spirit of Lamoure (2020) to derive readings of number-mismatched cases will yield a derivation in which Num° is required to join up with both nouns in a feature sharing dependency (Pesetsky and Torrego 2007; Brody 1998) for reasons of interpretability. This can only be achieved by the application of multiple Agree as one probe faces two goals. The revamping of Num° is the basis for deriving i.) the cross-linguistic distribution of number mismatched cases and ii.) the cross-linguistic distribution of the observed morpho-syntactic patterns:

   1. As pertains to the distribution of number-mismatched cases, the idea is the following: Languages that disallow number-mismatched cases will receive two contradictory filtering instructions for Num° ([Plur:–] from the singular noun and [Plur:+] from the plural one), which will yield a crash. Num° in languages which allow number mismatches will receive the same features but due to them allowing for split readings in singular D[N1&N2] constructions, the singular (“-“) feature value does not trigger any filtering and will thus not yield a crash.

   2. Due to the assumed lack of gender on Num°, Romance D° will not probe Num° but either of the two nouns (or both cf. above). Germanic Num° which is specified for gender is however a suitable goal for D° and can’t be ignored following closeness. Hence, Germanic Num° will inevitably receive multiple feature bundles from the nouns via Num°, while in the Romance languages left-conjunct agreement is always a second viable option.

3. Regarding spellout, I will follow Bjorkman (2016), adopting her Vocabulary Insertion rule. This rule predicts that in cases where a single head has multiple features of the same type, both need to be associated with the same form. Non-identity of the two outputs will yield a crash since it will require two lexical items to be realized in the same position of exponence. This will make sure that morphologically resolving languages will produce a crash if no suitable syncretic form is available.

In the next section (section 3), I will introduce Heycock and Zamparelli (2000; 2005) who provide the framework and semantic underpinning of this analysis, before I introduce necessary semantic and syntactic ramifications (in section 4)

3 The basis of the analysis: The structure of DPs Heycock and Zamparelli (2000; 2005)

In this section, I will outline the starting point, or rather the framework of the analysis to be developed here. I will modify the seminal work works by Heycock and Zamparelli (2000; 2005) in order to accommodate the data just presented.

Although coordination and the nominal domain in general are not exactly key concerns of Chomskyan generativism, D[N1&N2] constructions have been scrutinized by some authors. Since most of these approaches are tailor-made to derive left-conjunct agreement (Demonte, Fernandez-Alcalde and Pérez-Jiménez 2011; Demonte and Pérez-Jiménez 2012; Le Bruyn and de Swart 2014), they cannot handle the Germanic pattern.⁸ Heycock and Zamparelli (2000; 2005) is attractive as its semantics is fruitful with respect to the data (cf. the discussion above) and at the same time offers enough ‘syntactic wiggle room’ to integrate the two agreement patterns and, simultaneously, allows one to develop a more complete image of the syntax-semantics interface.

As already mentioned above, Heycock and Zamparelli’s (2000; 2005) approaches are primarily concerned with modelling the difference between Italian (-type languages) and English (-type languages).

While English D_SG[N_SG&N_SG] constructions allow for both a joint interpretation (reference to one individual (29)a)) and a split-interpretation (reference to more than one individual, depending on the number of conjuncts (29)b)), Italian only allows for a joint interpretation:

In D_PL[N_PL&N_PL] constructions both languages allow for split readings:

Heycock and Zamparelli (2005) assume the structure in Illustration 1:

Illustration 1

Structure Heycock and Zamparelli.

The system works as follows: Merge begins by creating the coordinate complex consisting of the two nouns. For Heycock and Zamparelli count nouns enter the derivation, denoting sets of singletons. They bear a valued [Plur] feature, which corresponds to their morpho-syntactic number and an unvalued [Latt] feature, which reflects their semantic number feature. This system allows for a fine-grained distinction between the different types of nouns: regular plurals are [Plur:+] [Latt:+], while singular counts are [Plur:–] and [Latt:–] and mass nouns are [Plur:–] and [Latt:+].

Coordination is envisioned as an operation they dub Set Product (SP) basically an unordered variant of the Cartesian product, which unifies every element of the denotation of one noun with every element of the denotation of the other noun.

For example, in a world where two philosophers (C(homsky) and K(ant)) and two linguists (C(homsky), M(arcel Den Dikken)) exist, SP yields a set that contains couples as well as singletons (if there is an intersection between the denotation of the two nouns). This yields the basis for deriving the two readings – split readings will refer to the couples (triples in cases where there are three conjuncts, etc.) and joint readings to the singletons.

Linguist = {{C}, {M}}                                             Philosopher = {{C}, {K}}

SP (Linguist, Philosopher) = {{C}, {C,K}, {C,M}, {K,M}}

Next, Pl° merges. Depending on whether Pl° carries [Latt:+] or [Latt:–], the following scenario unfolds: In case it is valued for [Latt:+] and [Plur:+] count noun-pluralization obtains. Heycock and Zamparelli (2000) implement pluralization via Link’s *-operator, while Heycock and Zamparelli (2005) suggest that recursive application of SP is responsible for pluralization. The outcome is the same, however – pluralization and conservation of the input (input marked in bold below):

*Linguist and Philosopher = {{C}, {C,K}, {C,M}, {K,M}, {C,K,M}}

In the case of D_SG[N_sg&N_sg] constructions, [Latt] and [Plur] on Pl° are, however, valued for ‘–’, which will result in the application of the identity function (alternatively, Heycock and Zamparelli (2005) assume that Pl° is absent in these cases).

Finally, Num° – the core device of this proposal – is merged. It is implemented as a filter, regulating ‘[…] the cardinality of the PIP denotation, filtering from it all the elements with the wrong number of atoms’ (Heycock and Zamparelli 2000: 347). If in Italian-type languages Num°’s [Latt] feature is valued for ‘–’ (i.e., we are dealing with a D_SG[N_SG & N_SG] construction), filtering will apply, removing all non-singleton elements (i.e., couples in our case) from the denotation, which leads to the prediction that only joint readings are possible in these languages:⁹

Num (*Linguisti e filosofi) = {{C}, {C,K}, {C,M}, {K,M}}

If it is valued for ‘+’ (i.e., we are dealing with a semantically plural D[N1&N2] construction), no filtering will apply.¹⁰

In English-type languages no filtering whatsoever applies. According to Heycock and Zamparelli (2005) this is so because English-type Num° never carries an unvalued [Latt] feature.

NumP-filtering applies in all languages if Num° is realized by an overt numeral cardinal, in which case filtering will apply according to the cardinality of the inserted element.

Whether or not [Latt:–] triggers Num°-filtering in a given language is envisioned as a parametric choice. This system straightforwardly models the absence of split readings in Italian D_SG[N_SG&N_SG] constructions.

4 Modifying Heycock and Zamparelli

To model the present data, Heycock and Zamparelli’s approach needs to be modified in four dimensions:

• i.)   Heycock and Zamparelli (2005), pace Heycock and Zamparelli (2000), hold that the feature [Latt] is triggering the filtering on Num° not [Plur]. However, the behavior of mass nouns in Italian-type languages casts doubt on this assumption.

• ii.)   Their system is not able to model a fine-grained semantic difference in the interpretation of the linguists and philosophers and the men and women.

• iii.)   Their approach makes no concrete predictions concerning the interpretation of number-mismatch cases; I claim that ii.) and iii.) are tightly related and can be fixed in one go.

• iv.)   Syntactically, their approach leaves the question open as to how exactly agree applies in D[N1&N2]constructions. Agreement dependencies that are set up with coordinate structures multiply either the amount of goals or the amount of probes. Given that according to the canonical implementations of agree, all unvalued and all uninterpretable features have to be eliminated upon transfer to the interfaces, it is not at all clear how this can be modelled with a binary operation.

Issues i.)-iii.) will be dealt with in 4.1 while issue iv.) will be solved in 4.2.

4.2 Syntactic ramifications

Heycock and Zamparelli remain vague about exactly how the derivation converges with respect to agree(ment) in D[N1&N2] constructions. In this section, I will thus introduce general principles I need to adopt in order to account for agreement in the context of DPs and coordination, as well as specify the individual heads for features.

4.2.1 Agreement with coordinate structures: Multiple Agree

In its canonical definition, agree is conceptualized as a binary relation between one probe and one goal. Obviously when dealing with dependencies such as in Germanic D[N1&N2] constructions, such a conception is inherently unfit to handle not only the Germanic data but also to fire the semantic mechanism outlined above. I therefore suggest allowing agree relations to span over one probe and multiple goals (remember that coordination is recursive). This is akin to Hiraiwa (2001). However, the goal is to do without a dedicated trigger or parametrization across lexical elements or languages.

In essence, I allow multiple agree to be a viable option at any point the derivational system applies agree. In this sense, agree is merely a special case of multiple agree. I follow Chomsky in the definition of agree (Chomsky 2000: 122):

1.  

1. The probe P agrees with the closest matching goal in D

1.  
1. a)

1. Matching is feature identity.

1.  
1. b)

1. D[omain](P) is the sister of P.

1.  
1. c)

1. Locality reduces to ‘closest c-command’¹⁴

Since agree is with the closest goal, this implies for multiple agree (conceptualized as one simultaneous operation (cf. Hiraiwa 2001)), that in configuration (38), four options are available (P stands for Probe, G for goal and ‘…’ for c-command).

1. (38)

1. P… G1…G2…G3

1.  
1. a)

1.   P agrees with G1

1.  
1. b)

1.   P agrees with G1 & G2

1.  
1. c)

1.   P agrees with G1 & G3

1.  
1. d)

1.   P agrees with G1 & G2 & G3

1.  
1. e)

1. *P agrees with G2

1.  
1. f)

1. *P agrees with G2 & G3

1.  
1. g)

1. *P agrees with G3

All invalid combinations have in common that they do not include agree with the closest goal. One may wonder why (38)c) is allowed in this system, given that G2 is skipped over. Strictly speaking, there is no second agree operation, but only one. Combining this with Chomsky’s canonical assumption about locality cited above, agree follows to be free to agree with any selection of goals in its c-command domain, so long as the closest goal is part of that selection.¹⁵

Whether or not a given combination will finally lead to a converging derivation further depends on the interface conditions, i.e., i.) is the result interpretable? ii.) can the outcome of multiple agree be spelt out?

Pertaining to ii.), I assume that multiple agree by an unvalued feature always yields multiple features:

1. (39)

1. P [F:___] agrees with G1 [F:val1] & G2 [F:val2]
2. yields P [F: val1] [F: val2]

For spell-out relevant features this means that spellout via a syncretic form, (pending further explanations in section 8, page 37) will be necessary. Following Bjorkman (2016) I assume that PF will have to run Vocabulary Insertion once per feature bundle. Should the resulting forms differ, a crash will obtain, should they coincide (i.e., PF picks out the same form twice) the derivation morphologically converges.

For features relevant to the interpretation, this means that the triggered processes at LF may not exclude one another (e.g., removing all singletons and removing all pluralities) and that interpretability (see further below) is generally obeyed. Otherwise, a crash will result (cf. above).

Before moving on to a brief discussion of agreement in the context of the DP, I want to stress that multiple agree is needed in Heycock and Zamparelli’s system, independently of the agreement patterns scrutinized in this paper. Apart from agreement patterns there are two points at which multiple agree is required:

• 1.)  Under a canonical, ‘monogamous’ view of agree, it would be impossible for Pl° – the only source of valued [Latt] – to value both unvalued instances of unvalued [Latt] on N° (cf. Illustration 1). Since valuation of unvalued features is a well-formedness condition upon transfer, according to Heycock and Zamparelli (2005), such structures could never converge.¹⁶

• 2.)  In order to account for the cross-linguistic distribution of number mismatched cases, I suggest Num°’s [Plur] to be simultaneously valued for ‘+’ and ‘–’ thus triggering both filters and eliminating the entire denotation in Italian-type languages. Since under monogamous agree such a valuation is beyond the reach of the derivation of number mismatched cases, in Italian languages it becomes impossible: Configurations such as D[N_plur & N_sing] Num would agree with the closest goal that is found in Pl [Plur:+] in the left-conjunct. The resulting prediction would be that in Italian-type languages, number mismatched cases should be fine if the left-conjunct is occupied by a plural noun. This is of course counter to the fact.¹⁷

4.2.2 Agreement in the DP: Relativized phi-completeness

Agreement within the nominal domain (or concord) differs from sentential agreement phenomena in that there are no phi-complete goals. Most Chomskyan-generative work on the nominal domain agrees that individual phi-features may very well be hosted on individual heads. Therefore, I will abandon phi-completeness for the present purposes and replace it with Danon’s (2011) relativized phi-completeness:

Relativized φ-completeness: An Agree operation leads to feature sharing if the goal matches all the unvalued φ-features of the probe.

4.2.3 Agree applied to Heycock and Zamparelli (2005)

The feature specification in Heycock and Zamparelli (2005) differs substantially from the one assumed in the canonical minimalist architecture in that it makes no reference to interpretability, which is standardly viewed to be the motor of syntax. Therefore, an implementation in terms of agree requires some further ramifications, which will be elaborated in this section. More concretely, I adopt a refined version of Pesetsky and Torrego’s feature sharing, which in turn has been strongly influenced by Michael Brody’s (1998) Elegant Syntax.

Heycock and Zamparelli’s system seems to break with the ‘standard theory’ in two respects:

1. At least in standard accounts, it is assumed that unvalued features are also always uninterpretable. For Heycock and Zamparelli this is not necessarily true. For example, [Plur] on N° is valued but uninterpretable, while [Plur] on Pl° and Num° is unvalued and interpretable. Outside of the canonical implementations, such conceptions are actually not unheard-of (cf. Pesetsky and Torrego 2007).

2. Usually it is assumed in the minimalist literature that for every feature there exist two subtypes, i.e., uninterpretable and interpretable features. For a derivation to converge, all uninterpretable instances must have been checked off by agreeing with an interpretable instance. Typically, uninterpretable and interpretable instances are distributed according to a one-to-one ratio. Some frameworks, e.g., Pesetsky and Torrego (but also Hornstein, Nunes and Grohmann 2006, c2005), embrace the possibility that there are multiple uninterpretable occurrences of one feature. However, in Heycock and Zamparelli (2005) there are multiple interpretable features, although the authors do not refer to them as such:¹⁸ that is, different occurrences of the same features get interpreted in different ways in different locations (e.g., [Latt] on Num° (filtering the denotation) and Pl° (creating a plural or non-plural denotation)).¹⁹ From a standard-theoretical view, the distribution of interpretable and uninterpretable features is indeed of utmost importance; the need to eliminate uninterpretable features from the structure to prevent a crash at LF is what supposedly powers the derivational motor. The occurrence of multiple interpretable occurrences of the same feature then begs the question as to how the requirements of interpretability are met in such a setup.

As for 1), the solution is quite simple: to give up on the implication between valuation and interpretability. As mentioned previously, exactly this has been suggested before by Pesetsky and Torrego (2007) and I will henceforth adopt this part of their theory. For clarity’s sake, I will also adopt their claim that the trigger for probing is the unvalued feature on the probing head, although as will become evident in the next section, nothing hinges on it.

With regards to 2.), I will refer to Brody’s (1998) claim that the system is more dynamic than the simple one-to-one correspondence of uninterpretable and interpretable features in standard minimalism would have you believe. More concretely, I modify Pesetsky and Torrego (2007) in the following way: While uninterpretable features can in principle be rendered invisible to LF by establishing a feature chain between the uninterpretable feature and an interpretable counterpart, this state of things is not set in stone. The occurrence of yet another interpretable feature that c-commands the aforementioned uninterpretable instance will render the latter visible LF again, i.e., uninterpretable relative to the ‘new’ interpretable feature (cf. Illustration 2 below):

Illustration 2

Interpretability.

I claim that in such a configuration, the presence of a c-commanding, interpretable matching feature will render the uninterpretable feature visible to LF again, although it is already part of a feature chain with another interpretable instance. In order to remedy this situation, the feature chain must be extended to the c-commanding, interpretable feature on X.

The problem of multiple interpretable instances is akin to a problem Brody notes about Chomsky’s analysis of wh-questions in English (Brody 1998: 162):

A +wh feature will be checked only if it is “strong” and then overtly (some of the problems with the notion of strength used here were discussed in section 2 above). Thus in English the wh-feature on C is strong and hence it can be checked either by (T +)did as in (36) or by a wh-phrase as in (34). Since a strong feature can be satisfied by a single element, the analysis raises the question of why (38a) is unacceptable. Here the strong wh-feature of C is satisfied by the hosted verbal element.”

1. (34)
1. a.

1. I wonder who +WH Bill saw (who)

1.  
1. b.

1. Bill saw who

1. (36)

1. Did +WH John see Mary

1. (38)
1. a.

1. did John give which book to Mary

1.  
1. b.

1. +WH John gave which book to Mary

This leads Brody to propose to model the data via his Bare Checking Theory, which forces all wh-features to merge (Brody 1998: 162):²⁰

which book in (38a) must form a chain linking it to the wh-feature of the auxiliary (and perhaps also of C). Further, the chain must be a full category copy chain, that is one that corresponds to overt movement of the minimalist framework, since English C is strong, i.e. it licenses a specifier in addition to an element in the word (Xo)-internal checking domain.

Wh-items are optional but if they are present, they need to be part of the agree-relations.²¹ This is akin to the situation in D[N1&N2] constructions, where multiple interpretable occurrences of [Latt] (for Heycock and Zamparelli) or [Plur] (for our purposes) are present in the structure. However, out of the box, Brody’s framework, in as far as its predictions are specific enough, is too strong.

Implemented as such, we would predict no difference between Romance and Germanic, as the number and gender bearing determiner would have to be connected to all other heads bearing gender and number, i.e., to both nouns of a given D[N1&N2] construction.

Applied to other syntactic configurations, such as Subject – Verb – Object structures, we would expect feature chains to minimally span across the verb (bearing uninterpretable phi-features), subject and object (both have interpretable phi features).²² Admittedly the latter would not necessarily be problematic in a predictive sense but instead be unnecessary.²³

The same reasoning obtains with respect to coordinated subjects and objects (both D°s have uninterpretable phi features, interpretable counterparts are to be found in both conjuncts) with the additional problematic that, depending on the locality constraints one assumes, chain formation might be altogether impossible yielding undergeneration. In the present framework this would be true, e.g., for any DP coordination, since neither D° c-commands the other, thus preventing Pl° in one conjunct to agree with the noun in the other.²⁴

I will therefore assume a modified version of Brody’s radical interpretability and thus that feature sharing à la Pesetsky and Torrego serves the purpose to make uninterpretable feature occurrences invisible to LF. In this sense, then it is a priori permissive to have multiple feature chains of a given feature in one derivation, so long as every chain is at least comprised of one interpretable instance. However, invisibility of a given feature is not a state achieved once and for all. Rather, membership in a given feature chain makes an uninterpretable instance of a given feature only invisible relative to the interpretable instances that are also members of the chain.

I will assume that an interpretable occurrence of a given feature [iF₁], which is not a member of a feature chain [iF₂]…[uF], makes the uninterpretable feature inside the feature chain visible to LF (relatively uninterpretable to that feature) and thus causes a crash iff the former c-commands any member of the chain.²⁵ This will force agreement in the relevant configurations, but crucially not require it in structures where it is impossible. This will also make sure that – as far as the specific semantics allow it – interpretable instances are optional in the sense discussed here. The intuition behind this mechanism is that the presence of a further interpretable occurrence of a given feature reveals uninterpretable occurrences of a given feature as uninterpretable relative to that one feature. Consider the following schema (… signals c-command):²⁶

1. (40)

1.   [iF_chain:1] … [uF_chain:1]

1. (41)

1. *[iF] … [iF_chain:1] … [uF_chain:1]

1. (42)

1.   [iF_chain:1] … [uF_chain:1] … [iF]

With these base assumptions laid out, I will proceed to the analysis of the two agreement patterns in Germanic and Romance. The difference between these two patterns will be derived from the assumption that Romance Num° lacks gender, which Germanic Num° however possesses. The latter will be tied to a morphological property, which allows to model plural as the fourth gender and thus to collapse number and gender in German and Dutch.

5 Deriving left conjunct agreement in Romance

Romance data derive straightforwardly. Let’s start with the gender-mismatched cases in Illustration 3, below:

Illustration 3

Romance Gender Mismatch Derivation.

After CoordP has been derived, Pl° merges. Since Pl° matches with two goals – the two conjunct nouns scuola and centro– agree applies. At this juncture, the question is whether a converging derivation will include an instance of multiple agree or of monogamous agree. Since Pl° bears the only interpretable [Latt] feature, and the two Ns bear an uninterpretable [Latt] feature each, a monogamous agree derivation would leave the uninterpretable feature of the furthest noun (i.e., the noun in the right conjunct centro) outside of the feature chain and would thus lead to a crash upon transfer. Multiple agree will apply and form two feature chains spanning over [Latt] and [Plur] across Pl° and the two N°s.

Next, Num° merges and matches with three goals – Pl°, N° scuola and N° centro. Again, the question is whether a converging derivation will include multiple agree or monogamous agree. Since there is already a chain, connecting all instances of [Plur] c-commanded by Num°, the result of the two distinct variants of agree will be the same: The existing chain is extended to [Plur] on Num and thus the uninterpretable instances on N° remain invisible to LF.

Finally, D° merges bearing unvalued [Plur] and unvalued [Gender] and finds two matching goals, i.e., the two nouns. Again, the question arises regarding which variant of agree will lead to a converging derivation. Since the Italian lexicon does not possess syncretic forms of the demonstrative, and multiple agree requires spellout by a syncretic form, only monogamous agree will lead to a converging derivation. Since agree is always agree with the closest c-commanded goal, D° will agree with the noun in the left conjunct scuola. Since scuola is already part of a feature chain for [Plur], no issue pertaining to interpretability will arise.²⁷

Next, we review the derivation of number mismatched cases, as shown in Illustration 4 below. First bear in mind that number mismatched cases are instances of PlP coordination, an intuition derived from their semantics. How is the formation of NP coordination in these cases prevented? As argued above, in NP coordination Pl° is the only source of interpretable [Latt] and thus must multiply agree with all nouns. As a side effect of this process, also Pl°’s [Plur] feature is valued multiply, which would lead to conflicting instructions to LF and thus to a crash, e.g., deriving a mass and a plural count noun derivation at the same time.

Illustration 4

Romance Number Mismatch Derivation.

Thus, a converging derivation will consist of PlP – conjuncts. After those have been derived, and each Pl° has agreed with its respective noun, forming two (independent) LF-invisible feature chains in terms of [Latt] and [Plur], Num° merges. Since Num° bears an interpretable occurrence of [Plur], it will render all c-commanded feature chains visible to LF. Those include uninterpretable instances on N° and thus must be agreed with by Num°. Hence, Num° multiply agrees with all c-commanded goals. This time [Plur:__] on Num° will receive two distinct values, which, as argued above, leads to a representation such as the following: Num° [Plur:–] [Plur:+]. As laid out, these instructions to LF are not problematic in English-type languages (since in English [Plur:–] on Num° does not trigger any filtering) but lead to a crash in Italian-type languages (because [Plur:–] and [Plur:+] filtering will result in an empty denotation), as desired.²⁸

Finally, D° merges and matches with both nouns. As before, both multiple agree and monogamous agree are viable options. However, only monogamous agree will yield a converging derivation, since the result requires a number syncretic D-element, which is unavailable in Romance. Thus, monogamous agree applies, yielding left conjunct agreement.

6 Deriving morphologically resolved agreement in English

Although arguably the language with the least number of relevant cases pertaining to the phenomenon (due to the absence of overt gender marking in the DP), we will start with English, shown in Illustration 5 below. Remember that English, due to the fact that it has no morphological reflex for gender within DP, only exhibits resolved morphological agreement with number mismatches, i.e., PlP coordination.

Illustration 5

English Number Mismatch Derivation.

Up to the formation of the CoordP, the derivation runs parallel to the one shown for the Romance languages: Each Pl° agrees with its respective noun. Next, Num° merges and minimally agrees multiply with both Pl°s, satisfying interpretability in terms of [Plur]. Since [Plur] on Num° is unvalued, it receives two values from both Pl°s.

When D° finally merges, its closest goal will be Num° (as opposed to the Romance case, where D° does not match with Num° due to the presence of gender on D°), and thus D° will receive multiple [Plur] features. The derivation will converge if a syncretic form can be found that will lexicalize D°.

7 Deriving morphologically resolved agreement in Dutch and German

Applying the system to the other two Germanic languages requires a further stipulation. If we apply the previous derivation to Dutch and German, inserting a gender feature – that English lacks – left conjunct agreement would be predicted. The reason for this is that D°’s closest goal in Dutch and German would not be Num° (or Pl°); Num° isn’t relatively phi-complete for Dutch and German D°, which are specified for Gender. The logic of the present approach suggests, then, two solutions: (i) Either D° possesses an interpretable or valued feature, whose counterpart is situated on N°, thus forcing multiple agree of D° with both N°s or (ii) Num° is a matching goal for D°, i.e., Num° has gender.

That Dutch and German Num° is specified for gender while Romance Num° is not is supported by the availability of strings of the form D > a\one > N (“>” is indicating linear precedence), which are permitted in Dutch and German but not in Romance:

One may wonder whether gender is really the feature at stake here – after all even in German the paradigm is rather impoverished, exhibiting no differences between genders in nominative. However, the accusative paradigm clearly shows that we are in fact dealing with gender:²⁹

While any of the morphologically resolving languages here has no issue in producing such a string, the Romance varieties under scrutiny harshly reject it. Assuming that the overt D° forces the numeral to occupy Num°, eine/eene can lexicalize Num° if German and Dutch Num° are specified for gender. The Romance versions however cannot, because gender information is lacking in Num° but is necessary for spellout. Note that this construction becomes well-formed in Romance once a cardinal element different from one is chosen:

This makes sense, as one is the only cardinal element which morphologically marks – and therefore requires – gender.³⁰

However, the question at this point is, why should this be so? Or rather how should the child acquire this? After all, constructions such as these are not exactly frequent in discourse, which is already implied by the fact that they are restricted to the numeral one (other cardinal numerals do not inflect for gender in the languages under scrutiny).

I want to suggest that the reason this is so, is related to a very special property of the German and Dutch D°-inventory: German and Dutch D-elements “lose” their gender in the plural, i.e., plural forms of D-elements exhibit a perfect syncretism in terms of gender. This long standing observation has led some authors (Krifka 2009; Sternefeld 2006) to argue that the German DP does not have a number feature to begin with, but that the plural is the 4^th gender.³¹ If this idea is on the right track, then Num° is relatively phi-complete for D° because both of them do not have a dedicated number feature ([Plur] in our terms) but only gender, which implies the information conveyed by number. This would also explain how children acquire the feature specification of a mostly null head and how grammatical judgements for D[N1&N2] constructions, although relatively rare, are typically expressed with much confidence.

At this juncture, it also becomes evident why English should pattern with Dutch and German and not with Italian and French. English, like Dutch and German, is featureally ‘impoverished’ in contrast to Romance languages. English lacks gender and Dutch and German lack number.³² This total absence of gender is what prevents English D° from probing beyond the multiply agreeing Num°.

For the present purposes, I keep a separate number feature [Plur] in the tree in order to keep the structure clearer and simply add the feature [Gender:___] to Num°. With Num° bearing [Plur:___] and [Gender:___], morphologically resolved agreement follows in a straightforward fashion. I start with number-mismatched cases (cf. Illustration 6 below):

Illustration 6

German Number Mismatch Derivation.

After the conjunction of PlPs has been derived, the two Pl°s agree with their respective N°’s in terms of [Plur] and [Latt], establishing feature chains respectively. Next, Num° merges, bearing [Plur:___] and [Gender:___]. Since Num° bears an interpretable instance of [Plur] and c-commands both feature chains, it has the potential to render these chains visible at LF. Since Num°, by virtue of bearing [Gender], does not match with Pl° but with the individual nouns, it will have to probe every N° to maintain LF-invisibility, which will allow the derivation to converge.

Finally, D° merges, bearing both [Plur:___] and [Gender:___]. Thus, its closest matching goal will be Num°, which it will minimally have to agree with. Both multiple and monogamous agree will lead to a converging derivation. In any event, morphologically resolved agreement will obtain, because D° will receive multiple valued features from Num°, which will require spellout by a syncretic form.

Last, we take a look at what happens in Dutch and German gender-mismatched cases, as visualized in Illustration 7 below.

Illustration 7

German Gender Mismatch Derivation.

Remember that in these cases we are dealing with NP coordination.³³ After the NP conjunction has been derived, Pl° merges and, should the derivation converge, necessarily multiply agrees with all nouns, since Pl° is the sole locus of interpretable [Latt]. Also, since [Latt] on Pl is interpretable and uninterpretable on N, a feature chain needs to be established between both Ns and [Latt].³⁴ Next, Num° merges, bearing [Plur:___] and [Gender:___]. As before, Num° does not match with Pl° because Pl° is relatively phi incomplete to Num°. However, Num° matches with both N°s. Since [Plur] on Num is interpretable, a converging derivation will include multiple agree of Num° with all N°s, thus rendering the uninterpretable occurrences of [Plur] on N° invisible to LF.

Finally, D° merges. Irrespective of whether multiple or monogamous agree applies, D will receive multiple valued features, because its closest matching goal is Num°, which has just multiply agreed. The result is again spellout in a syncretic form. If such a form is unavailable, the derivation will crash.

8 Spell-out: Syncretic forms cannot be the product of intersection – Evidence from German

Up to this point, the answer to how multiple feature bundles lead to the spellout of syncretic forms has been rather vague. In this section, I argue that an approach in terms of intersection as proposed by Hein and Murphy (2019) is unsuitable for deriving the correct outcome due to the existence of a potential elsewhere case in the German article paradigm (der). I then suggest the implementation of the approach described in Bjorkman (2016)

One might ask why a multiply agreeing probe should even carry so much ‘garbage.’ In fact, Hein and Murphy (2019) have suggested handling cases such as these by intersecting feature bundles. Under the assumption that syncretic forms are underspecified, i.e., feature intersections of the fully specified contexts in which they can appear, and that spellout is governed by the subset principle, such an approach would lead to the correct outcome in the majority of the cases. However, German presents a problem to this approach in the form of the determiner form der. Der can surface in a variety of contexts, which do not reduce to natural classes:

If der were to be analyzed as one form it would be the radical elsewhere case, void of any phi-featural content. For instance, according to the feature system in (Sternefeld 2006) we would get the following:

The intersection of all these features would yield the empty set. Such a conclusion would be undesirable, as it would lead to the prediction that there are no ungrammatical cases of D[N1&N2] constructions in German, given that der could always be inserted in the absence of a more specified competitor, counter to the fact:

To prevent this, the most natural move would be to argue that der is at least partially an accidental syncretism and not a systematic one.

Albright and Fuß (2012) suggest that these two types of syncretisms can be teased apart by testing whether they license certain syntactic constructions: If so, we would be dealing with a systematic case, and if not the form is identical by accident. Lamoure (2020) shows that for der, neither of these two cases can be shown to be true, resulting in an epistemological draw.

I suggest circumventing the problem entirely and thereby reducing the number of assumptions necessary for syncretic spellout to work by adopting Bjorkman’s (2016) approach to spellout.

Bjorkman (2016) faces a similar issue albeit in a different context. Her main focus lies on the English go get construction, which is only possible in environments calling for an uninflected or bare verb (62) or for a form that is syncretic to the bare verb (63). Otherwise, the result is ungrammatical (64):

Bjorkman (2016) suggests that in go get constructions, the [INFL] feature of the two verbs is valued twice, once by [uINFL:DIR] on v° and once by e.g., [iINFL:pres] on T°. In order to account for the fact that the result is only well-formed if a bare form or a form syncretic with the bare form is inserted, the author makes use of multiple applications of vocabulary insertion (VI), which is triggered whenever a single head has multiple features of the same type.³⁵ In order to converge, multiple applications of VI must converge on the same result, as there is still only one (head-)position.³⁶

How does this work in the present context? Consider the following syntactic output, which would correspond to the ungrammatical (61) (*Der Mann und Frau):

D° {[Fem :–] [Masc :+] [object :–] [oblique :–]} {[Fem :+] [Masc :–] [object :–] [oblique :–]}

	d-{[Fem :–] [Masc :+] [object :–] [oblique :–]}	d-{[Fem :+] [Masc :–] [object :–] [oblique :–]}
-er	der	–
-ie	–	die

Just as in Bjorkman’s case, the relevant heads (the article in this case) are associated with two different vocabulary items, which represents an ‘[…] impossible morphological representation, precisely because it associates two different vocabulary items with a single position of exponence’ (Bjorkman 2016: 74).

In contexts where a syncretic form is available, the approach correctly predicts a converging VI:

In this case, the multiple application of VI will converge, as VI will insert only one form into the relevant position of exponence (D°).

This solution derives the correct outcome for der, since this spellout mechanism remains agnostic towards the true nature of the syncretism. At the same time, no features are lost during the derivation, thus satisfying inclusiveness, which may be at risk in Hein and Murphy’s suggestion. Finally, this approach does not require a DM-Style subset-principle approach to syncretisms, but could in principle be modelled by any approach that predicts the correct distribution of syncretic forms in non- D[N1&N2] contexts.

9 Conclusion

In this paper I have shown that D[N1&N2] constructions exhibit different agreement patterns in Romance and Germanic languages. I have presented an analysis based upon a modified version of Heycock and Zamparelli’s (2000; 2005) pioneering analysis that allows us to model both the observed morpho-syntactic and semantic properties of this construction. In terms of syntactic assumptions, this analysis requires multiple agree yet implements it in a way that is solely constrained by the interface conditions of the semantics and phonology – the output of the syntax and hence agree must be semantically and phonologically computable. The hypothesized locus of variation between Romance and Germanic was suggested to be the availability of strings of the form D > one > N, which turns out to be as cross-linguistically stable as the observed agreement pattern.

The system as it is set up right now predicts two types of languages with respect to morpho-syntactic patterns: left-conjunct agreement languages in which D > one > N strings are ungrammatical, and languages with morphologically resolved agreement in which D > one > N strings are well-formed. I speculate that the latter property derives from an impoverished feature inventory in these languages. Further data from other languages is required to further falsify this analysis – more specifically, it would be interesting to gather data from a different language family, such as Slavic, and investigate languages from the Germanic branch, whose feature systems do not lend themselves to an analysis parallel to the one in Dutch and German.³⁷

Abbreviations

SG     =  Singular

PL     =  Plural

F       =  Feminine

M      =  Masculine

N       =  Neuter

GEN  =  Genitive

DAT  =  Dative

ACC  =  Accusative

Competing interests

The author has no competing interests to declare.