Quantifier float (Q-float) occurs when a quantifier is separated from its associate nominal, as has been first observed by at least Postal (1974) and discussed in much later work (e.g., Kayne 1975; Fiengo & Lasnik 1976; Sportiche 1988; Shlonsky 1991; Bobaljik 1995; Doetjes 1997; Benmamoun 1999; McCloskey 2000; Bobaljik 2003; Bošković 2004; Fitzpatrick 2006; Spector 2009; Jenks 2013; Lacerda 2016; Zyman 2018). The example in (1b) is representative:

Different accounts were offered to explain the Q-float phenomenon. One account, the movement/stranding account, argues that Q-float results from leftward movement of the associate nominal out of a complex constituent that contains both it and the quantifier (e.g., Giusti 1990; Shlonsky 1991; Merchant 1996; Cinque 1999; McCloskey 2000; Zyman 2018). Another account is the adverbial/adjunct analysis which simply treats floating quantifiers (FQs) as adverbial elements that semantically modify the predicates they combine with, or that modify their associate nominals (Kayne 1975; Dowty & Brodie 1984; Miyagawa 1989; Baltin 1995; Bobaljik 1995; Torrego 1996; Brisson 1998; Benmamoun 1999; Reed 2010). A more recent analysis, a hybrid analysis, argues that in some languages, both stranding and adverbial modification are available (e.g., Fitzpatrick 2006).

This paper investigates Q-float in Modern Standard Arabic (and related languages like Hebrew), and presents empirical facts that argue in favor of a movement analysis such as the one proposed by Ott (2011; 2012; 2015) for split topicalization and Q-float in German. This analysis does not involve stranding of the FQ in the strict sense. Rather, the FQ and its associate are merged together in a particular syntactic position as a set of autonomous phrases, where a member of the set (the associate) moves to allow the set to be labeled and integrated into the structure. This means that although the associate and the quantifier are in a movement dependency, they have not formed a continuous constituent at any stage of the derivation (see section 4 for a definition of a continuous constituent). I will show that the analysis captures many of the peculiarities of Q-float in natural languages in general and in Arabic specifically. Also, if the analysis is in the right direction, it provides support for the labeling framework that emerged from Chomsky (2013) and related work.

I begin by presenting the facts of Q-float in Arabic in section 2, many of which have not been documented before, as far as I know. In section 3, a brief critique of previous accounts will be presented, showing that they do not capture all the facts of Q-float in Arabic. In section 4, I present an analysis of the facts, adopting the assumptions and analysis of split topicalization and Q-float in German proposed by Ott (2011; 2012; 2015) and show that this analysis solves many of the puzzles of Q-float in Arabic. Section 5 is a conclusion.

This section presents a description of Q-float in Arabic. The description includes the nature of elements that may float, the distribution of these elements, and case matching and phi agreement between floating elements and their associates. Unless otherwise indicated, the variety of Arabic to be investigated throughout the paper is Modern Standard Arabic. All the Arabic examples were confirmed with native speakers.

Before presenting my analysis of Q-float in Arabic, a few remarks on previous accounts are in order. Two accounts have been proposed in the literature: the movement/stranding analysis and the adverbial/adjunct analysis. Proponents of the stranding analysis claim that Q-float results from leftward movement of the associate nominal away from the quantifier (Giusti 1990; Shlonsky 1991; Merchant 1996; Cinque 1999; McCloskey 2000; Zyman 2018, among others).8 One major version of this analysis in the literature on Semitic languages like Arabic and Hebrew is Shlonsky (1991) in which it is proposed that Hebrew kol ‘all’ (and its Arabic equivalent) is a functional head that selects a DP complement. According to Shlonsky, building on Sportiche’s (1988) movement analysis, Q-float is derived by moving a DP out of a QP, resulting in the quantifier being stranded, as illustrated in (43) for (42):

Here the quantifier kol ‘all’ floats as a result of leftward movement of ha-yeladim ‘the children’ through the specifier of QP. Ha-yeladim and kol enter into spec-head agreement, which is manifested by the agreement clitic -am. (In Shlonsky’s analysis, VP-internal subjects are projected to the right. Presumably, in his analysis, all the other specifiers should project to the left.)

The stranding analysis has received much criticism. First, it does not capture the cases in which FQs may not appear in NP trace positions, as with passive and unaccusative verbs (e.g., (17)).9 Second, the stranding analysis does not explain why in some languages FQs cannot associate with Ā-positions (e.g., (20)). More particularly, as pointed out by Bobaljik (1995; 2003), Fitzpatrick (2006), and others, one problem for the stranding analysis is accounting for the anaphor-like locality restrictions on FQs in languages like English: a nominal which has undergone A-movement may be the associate to a FQ, while a nominal that has undergone Ā-movement may not. As pointed out by Bobaljik (2003), this is hard to explain under the stranding analysis given that well-attested stranding phenomena are not restricted to A-movement. As has been shown in section 2.2, the two criticisms outlined above are inapplicable to Arabic: Arabic allows FQs to appear in the object position of passives and unaccusatives, and allows FQs to associate with nominals occupying Ā-positions.

There are other facts that argue strongly against the stranding analysis of Q-float. More particularly, in many cases, it is impossible for the associate and the FQ to have formed a continuous constituent at any stage of the derivation (see section 4 for a definition of a continuous constituent), as pointed out by many (e.g., Sportiche 1988; Bobaljik 2003). This is true for Arabic and for languages like English and French:

In each of the pairs above, the non-floating sentence requires a preposition (‘of’) between the quantifier and the nominal that follows it, a fact that is hard to explain under the stranding analysis.

Similarly, Arabic FQs may associate with a coordinate phrase. Non-floating versions are grammatical only with a preposition:

This is the case in English as shown by Bobaljik (2003) (note, however, that even with an intervening preposition, a quantifier cannot precede a coordinate phrase in English):

Moreover, a FQ may associate with a quantified phrase, but a non-floating version is unavailable for this case:

Furthermore, and as shown in section 2.3, the best indication of the fact that a FQ cannot have formed a continuous constituent with the associate is the fact that in floating constructions, the floating element and its associate must have matching case, whereas in non-floating constructions the associate invariably gets GENITIVE case. This indicates that floating and non-floating quantification are not derivational variants.11

All of the above facts lead to the conclusion that the stranding analysis does not explain all the cases of Q-float in Arabic. However, a number of facts suggest that there is a movement dependency between a FQ and its associate: a FQ may associate with A- or Ā-positions (21)–(22) and is sensitive to islands (31)–(33). In addition, a FQ shows reconstruction effects for the binding conditions:

It therefore seems that Q-float cannot result from movement of the associate out of a continuous constituent that includes it and the quantifier. It rather appears that the associate moves out of a projection that dominates both it and the quantifier. In section 4, I will present an analysis I essentially adopt from Ott (2015) which reconciles these apparently conflicting facts.

The adverbial analysis, on the other hand, was proposed as an alternative to the movement analysis. Proponents of the adverbial analysis argue that FQs are adverbial elements that semantically modify the predicates they combine with, or in some versions of the analysis, that modify their associate NPs (Kayne 1975; Dowty & Brodie 1984; Miyagawa 1989; Baltin 1995; Bobaljik 1995; Torrego 1996; Brisson 1998; Benmamoun 1999, among others). The type of argument used to support the adverbial analysis is the same one that rendered the movement analysis problematic: FQs cannot appear in thematic positions of some verbs like passives. As shown in section 2.2, this criticism is mainly based on facts from English and French, and is inapplicable to Arabic.12

The conclusion that I reach, then, is that existing accounts of Q-float do not explain the Arabic facts perfectly. The distribution and the nature of the elements that float indicate that FQs cannot have formed continuous constituents with their associates at any stage of the derivation, nor are they projected as adverbial adjuncts.

In this section, adopting Ott’s (2011; 2012; 2015) analysis of split topicalization and Q-float in German, I propose that in Arabic, a FQ and its associate are merged together in a particular syntactic position as a symmetric set of autonomous phrases from which the associate must move leftward to allow the set to be labeled and integrated into the structure (Chomsky 2013).

Ott’s analysis rests on a number of assumptions. First, he assumes that a FQ and its associate hold a predicate-argument relation at the semantic level. Second, as argued by Chomsky (2004; 2008; 2007; 2013), the operation (symmetric) merge combines two syntactic objects into an unordered set (assuming that linear order is computed in the post-syntax). This set must be labeled in order for the constructed unit to be syntactically and semantically integrated into the surrounding structure. According to Chomsky, the label of the set is identified via a simple algorithm which identifies the head through a specific feature of that head. To put it simply, the label of {A, B} is A if A is a lexical item and B is an XP. A set that results from merging XP with YP is a symmetric set or a locally unstable combination for which no lexical item can be identified as a head (Moro 2000; Chomsky 2013). For the derivation to converge, it is crucial for the combination to be labeled. One solution is for one of the members of the set to move out of it via internal merge (Moro 2000; Chomsky 2013). The result is that only one phrase remains properly contained within the set, and, as a result, it determines the label of the set.

To illustrate the analysis for German, consider the following example from Ott (2015):

In this sentence, the DP ‘the children’ merges with the QP ‘both’ together in a symmetric set, {DP, QP}. Since the labeling algorithm cannot detect a unique lexical item that can label the set, movement of DP is forced to allow the set to be labeled. As a result of this movement, QP becomes the only phrase that is properly contained within the set, and consequently the set should be labeled as QP:

In principle, moving either of QP and DP would asymmetrize the set; however, according to Ott, it is DP rather than QP that should move for pragmatic considerations. In particular, fronting DP results in a structure that is pragmatically felicitous in which the DP acts as a frame setting (in the sense of Jacobs 2001) with respect to which the proposition that follows it can be interpreted. This is compatible with the view that syntax is not crash-proof, and that merge applies freely, but is restricted indirectly by constraints at the interfaces (e.g., Chomsky 2004).

Note also that it is crucial in Ott’s analysis to treat FQs (and the remnants of split topicalization) as elliptical nominals, based on the fact that German allows free NP-ellipsis. This is also the case in Arabic:

Given this, I propose that all the examples of Q-float in Arabic involve merger of QP with a null nominal in them. In non-floating constructions, on the other hand, the quantifier is merged with a DP as a lexical item, giving rise to an asymmetric set that can be labeled as QP. The two merging options are illustrated in (58) for the examples in (57) = (9), repeated below:

The first merging option in (58a) gives rise to a continuous constituent, while the second in (58b) gives rise to two autonomous phrases. Given the labeling framework outlined above, I propose the following definition of a continuous constituent:

This definition should be accompanied by the assumption that the labeling algorithm computes the labels before internal merge of objects from the current constituent applies. It follows, then, that a movement that is forced by labeling breaks the continuity of constituents because, at the point it applies, the set has not been labeled yet. In contrast, a movement that is not forced by labeling, say wh-movement of a DP from within a PP = {P, DP}, for instance, does not break the continuity of the PP constituent, because at the point it applies through a phase edge, all the constituents inside that phase have been labeled.

Now turning to the cases of Q-float in Arabic, consider the example in (5b), reproduced below:

The labels of ‘the students’ and ‘all’ are DP and QP, respectively. The quantifier ‘all’ semantically takes ‘students’ as an argument. The quantifier and its associate combine into a set, {DP, QP}. Because the set is symmetric, the labeling algorithm cannot identify a label for it. Without a label, the set cannot be integrated syntactically and semantically into the structure, and consequently the derivation will crash. One way to break the symmetry is for one of the members of the set to move, which is the DP (‘students’) for the pragmatic reasons explained briefly above. The label of the set will then be the label of the phrase that is properly contained in the set, which is the label of the quantifier ‘all’ (QP). Assuming that derivations proceed in phases, the moved phrase moves via the edge of the VP phase, where it may undergo further movements:

Case matching between the quantifier and the associate nominal follows under the analysis presented. Any head proping for case features values its features via Agree with all members of its goal, which guarantees that each member is marked with the same case (Ott 2015). I suggest that the clitic appearing on the FQ is a result of matching between the nominals prior to movement. It should be noted that the shared phi and case features between the members of the set {DP, QP} are insufficient to label the set via the mechanism of labeling by feature-sharing (e.g., Chomsky 2013). Following Chomsky (2013), mere feature matching is not sufficient to render a set labelable; only agreement is, which is not what see here.

Also, the analysis accommodates the non-floating cases. These are the cases in which QP is not elliptical. Consider the example below again:

Here the quantifier and the nominal following it are not merged together in their base position as a symmetric set. Rather, they form a complex phrase (i.e., construct state) in which the quantifier selects the nominal phrase as a complement. ‘All students’ is merged as an asymmetric set of {Q, DP}. I adopt the structure proposed by Shlonsky (2004) for construct states headed by a Q, as illustrated below (details left out):

The case mismatch between the quantifier and the DP follows from the fact that the quantifier values its case features via Agree with an outside case probe, being the head of the projection, while the DP values its case features within the construct state nominal. The mechanism through which case is valued inside the construct is not crucial, but I assume, following the assumptions in Benmamoun (1999) and Bošković (2004), that the DP is assigned GENITIVE case via Agree with the quantifier kull.

The analysis also explains the impossibility for examples like (64) = (47a) to have a non-floating version:

This sentence is derived by merging the coordinate structure and the quantifier together in a set of {&P, QP} (I follow Collins 1988, Johannessen 1998, and others in assuming that coordinate structures have the label &P. But see for instance Zhang 2010, Al Khalaf 2015; 2017, and Bruening & Al Khalaf 2019 for more recent perspectives on coordinate structure). Note that in this example, the QP is an elliptical nominal. The coordinate structure moves out of the set and allows the set to be labeled as QP. Consider the non-floating version in (65) = (47b) again, on the other hand:

The ungrammaticality of the sentence (without the preposition) can be explained as follows. The coordinate structure and the quantifier are merged together in subject position as a symmetric set. Again, the QP is an elliptical nominal. Neither of the members of the set moves, and the combination fails to be labeled as a result, causing the derivation to crash. Note, however, that the sentence becomes grammatical with the preposition ‘of’. The grammaticality follows if the quantifier merges with the PP as a lexical item (i.e. a Q), forming an asymmetric set of {Q, PP}, as shown above for non-floating constructions. This clearly does not pose any issues to the labeling algorithm. Another way to explain the unacceptability of Q &P sequence and the acceptability of Q PP sequence is to posit selectional restrictions on ‘all’, banning it from co-occurring with &P and allowing it to occur with PPs headed by of. This would capture the facts, but it remains an ad hoc stipulation that needs to be explained.

Moreover, the analysis captures the island sensitivity of FQs. The associate moves from a projection that dominates both it and the floating element. It also follows that this symmetry-breaking movement should in principle create A- and Ā-chains, accounting for the fact that associates can occupy A and Ā- positions in Arabic (and languages like West Ulster English).

It should be noted that the analysis presented predicts that Arabic FQs may appear in thematic positions or where nominals are externally merged (this does not mean the FQs cannot appear in non-thematic positions. That depends on whether an unlabeled set can move to a non-thematic position prior to asymmetrization). As was shown in section 2.2, Arabic FQs appear in the positions that are known to be NP trace positions, including the complement position of passives and unaccusatives where FQs are banned in other languages. One class of exceptions to this generalization, however, are the cases in which FQs appear sentence-finally (66) = (18):

As indicated earlier, English (and French) FQs, in contrast, are banned in these positions, and are allowed only if followed by a PP or a secondary predicate (19). I suggest that the contrast between English-type languages and Arabic is due to the fact that in Arabic, word order is freer than the word order in those languages. For instance, in Arabic, the subject may precede the verb or follow it, and may even be separated from the verb by VP adjuncts when it follows the verb:

One could simply say that in these examples, the subject is merged in spec-VP as a symmetric set from which the associate moves; the verb moves resulting in the FQ appearing sentence-finally or near-sentence-finally. This would capture the passive and unaccusative examples perfectly. However, it predicts that with transitive verbs FQs should not appear sentence-finally after the object, contrary to fact:

Thus, it seems to be more plausible to say that the peculiarity of the distribution of FQs in Arabic arises from the fact that subjects may appear at the left or right edge of VP (cf. Shlonsky 1991 who argues that in Hebrew subjects can be right-peripheral).

Before concluding this section, I should point out a potential problem illustrated by the examples below.

In these examples, the FQ associates with a preceding pronominal possessor within a construct state nominal, which may be a problem for the analysis given that this pronominal clitic is not obviously a DP and consequently has not necessarily moved from a symmetric set that contains both it and the quantifier (note that the same problem arises in the cases in which a FQ associates with an object pronominal clitic as in (35b)). The current analysis would explain the floating pattern in (70) as follows. Following Shlonsky (2004), nominal construct states have the structure of NP in which N selects DP/QP as a complement. In (70), the head beit- merges with a symmetric set of {QP, DP}. In order for the set to be labeled, the associate DP (-hum) must move. Here it moves and cliticizes on the N beit-. The clitic appearing on Q is a result of agreement.

This is how the analysis would capture these cases, but there remains the question whether analyzing pronominal clitics as DPs is a desirable option. I will leave this issue open for future research, however.

In this paper, I presented a detailed description of the syntax of Q-float in Arabic. The facts suggest a movement dependency between a FQ and its associate, but also show that it is impossible for them to have formed a continuous constituent at any stage of the derivation. To account for these two conflicting facts, following Ott (2012; 2015), I proposed that Arabic Q-float constructions involve merger of a symmetric set of two autonomous phrases. In order for the set to be labeled, it should be asymmetrized via movement of the associate out of the set. A major result of this study is that the distribution of FQs in Arabic serves as a powerful diagnostic of the distribution of lower copies of displaced NPs (NP trace positions in traditional terms). It also provides support for the labeling framework that emerged from Chomsky (2013) and related work.