French clausal ellipsis: types and derivations

1 Introduction

Like many other languages, French has, in its lexicon, pro-forms that substitute for a proposition. The most common of these is le ‘it/so’, a weak (clitic) pronoun that refers back to a previously introduced proposition. This type of anaphora can only be effected by means of an overt pronoun; that is, le never alternates with pro or with an elided CP or TP category in such cases. This is illustrated in (1) where the material within brackets is taken to be the pronoun’s antecedent.

In addition to the clitic le, the strong demonstrative pronoun ça ‘that’ can substitute for a proposition given the proper context, for example when the speaker experiences negative affect towards the referent, as in (2). Like le, ça does not alternate with a silent category.

Given these properties, it is somewhat surprising that there are contexts in which le does alternate with a silent category and never alternates with the strong pronoun ça. One such context (others will be discussed below) is that involving a propositional complement to a modal verb. This is illustrated in (3).^[1]

Regarding the paradigm in (3), some preliminary remarks are in order. First, the examples in (3) have the same interpretation with le as they do without le. Second, as is often the case when the linguistic grammar of a language makes two options available, speakers vary as to which option is to be preferred over the other in a particular sentence/construction according to factors that are not always clear. There does seem, however, to be a tendency for speakers to prefer the use of le in formal styles of speech (so-called français soutenu) and to favor not using le in colloquial speech (so-called français familier).^[2] On the syntactic side, such sentences, when used without le, have been argued by Dagnac (2010) and Authier (2011) to involves ellipsis of an infinitival complement to the modal head, giving (3a) without le, the structure illustrated in (4), where the crossed-out material is structurally present but remains unpronounced in PF.

A possible interpretation of the paradigm in (3) would then be that the absence of le signals clausal ellipsis, while the presence of le signals that propositional anaphora is achieved pronominally, just like it is in (1). There are, however, two facts that militate against this view. First, it does not explain why the use of the strong pronoun ça is sharply ungrammatical in (3). Second, one of the main reasons to assume that sentences like (3) involve ellipsis is that extraction is possible out of the ellipsis site, as shown in (5).

On the pronominal analysis of the le that appears in (3), we thus predict that such extraction operations should be incompatible with the presence of le since pronouns have no internal structure. This, however, turns out to be the incorrect prediction: le can, in fact, be used in sentences involving wh-extraction like (5), as (6) shows.^[3]

Thus, the examples in (6) strongly argue against treating le in modal ellipsis contexts as a propositional weak pronoun. While this might explain why the strong pronoun ça remains unavailable in such examples, namely because they disallow any kind of pronominal anaphora, it leaves the syntactic status of what I will from now on call ‘special le’ in (3) and (6) unexplained. In this article, I first argue that special le is the overt realization of the E-feature hypothesized by Authier (2023) to occur in French predicate ellipsis.^[4] I then provide evidence for the existence of two types of clausal ellipsis in French and propose an analysis that accounts for their distribution. The rest of this article is organized as follows. Section 2 establishes that the presence of special le systematically correlates with contexts in which clausal ellipsis occurs. In Section 3, I take stock of the evidence uncovered in Sections 1 and 2 and I establish the existence of two distinct types of clausal ellipsis: Type 1, which may, but need not, involve the extraction of an ellipsis remnant, and type 2, which must involve extraction of a remnant to be licensed. Section 4 lays out the dynamic derivational theory of ellipsis assumed in this article. In Section 5, I propose a new analysis of French modal ellipsis (type 1 clausal ellipsis) that accords with the assumptions made in Section 4. Sections 6 and 7 propose an account of the properties that characterize type 2 clausal ellipsis found in comparative standards and sluices. Section 8 seeks to provide an explanation of why type 2 clausal ellipsis is licensed in ACD relatives, but not in standard relatives. Finally, in Section 9, I provide a summary of the findings and proposals made in this article.

2 The distribution of special le

In this section, I will provide evidence that the special le whose existence was uncovered in Section 1 is not restricted to modal ellipsis sentences. Specifically, I will show that the presence of special le generalizes to contexts in which clausal ellipsis occurs. These include the standard of comparatives, Antecedent Contained Deletion (ACD) sentences, and sluicing with long distance wh-extraction.

3 Empirical synopsis

In the preceding sections, two main empirical patterns involving special le have been uncovered. First, special le must co-occur with ellipsis. Contrasts like those in (34) and (35) show this very clearly.

Second, special le co-occurs with what appears to be two distinct types of clausal ellipsis. The first type (hereafter type 1) may, but need not, involve wh-extraction out of the ellipsis site. French modal ellipsis is of this type. The second type (hereafter type 2) is exemplified by clausal ellipsis in comparatives, ACD constructions, and long-distance sluicing, all of which must involve wh-extraction to be licensed. Before proposing an analysis that accounts for the presence of le in both type 1 and type 2 clausal ellipsis, however, I will first lay out the theoretical assumptions regarding ellipsis that I will be using throughout the remainder of this article.

4 A theoretical framework for ellipsis

The approach to ellipsis phenomena I will adopt here is that spelled out in Authier (2023). This approach draws heavily from Albrecht (2010) and Bošković (2014), both of which start with the premise that, in a derivation by phase framework, ellipsis occurs at the point of transfer, which leads us to expect that the target of ellipsis is always a phasal domain. Additional assumptions are listed in (36).^[11]

Authier (2023) combines these assumptions with the view that the le clitic that appears in French predicate ellipses like (37) is the phonological realization of an unvalued E-feature on faire ‘do’, which is argued to be a low semi-auxiliary that selects a vP phasal complement and serves as the licensing head for predicate ellipsis. The ellipsis in question, which Authier argues is vP ellipsis, is thus assumed to proceed as illustrated in (37).^[12]

(37)

(Manon	n’a	pas	lavé	son	linge,	mais)	Axel	l’a	fait.
Manon	Neg-has	not	washed	her	clothes	but	Axel	Cl-has	done
‘(Manon didn’t wash her clothes, but) Axel did.’

As in Authier’s analysis of French predicate ellipsis, I will assume that the special le that appears in the clausal ellipsis phenomena under discussion here is the phonological realization of an E-feature on an ellipsis licensing head, and I will further assume that the same E-feature can also exist without a phonetic matrix.^[13] In what follows, I will explore the consequences of the assumptions laid out in this section for the syntax of special le and the licensing of the two types of French clausal ellipsis.

5 Type 1: French modal ellipsis

Authier (2011: 202) argues that French modal ellipsis is licensed by topicalization. That is, elided infinitival clauses in modal ellipsis contexts are silent copies of infinitival clauses standing in a topic position, with the copy occupying the first-merge position being silent as well. This is illustrated in (38b).

As Authier (2011: 203) points out, this hypothesis correctly predicts that French modal ellipsis is possible under deontic but not epistemic modals, namely because infinitival clausal complements to deontic modals can be topicalized but infinitival clausal complement to epistemic modals cannot, as shown in (39).

While this analysis works well for clausal ellipsis under modals, it does not appear to be extendable to the other type of clausal ellipsis under investigation in the present work. For example, when clausal ellipsis occurs in a comparative sentence like (40a), the elided clause has no grammatical topicalized counterpart like (40b) that could serve to justify such an analysis.

I would therefore like to suggest a reinterpretation of the fact that deontic but not epistemic modals license clausal ellipsis in French; that is, as I will argue, clausal complements to deontic modals are phasal but clausal complements to epistemic modals are not and, given that ellipsis can only target phasal projections (cf. (36d)), it follows that only clausal complements to deontic modals can be elided.

There is a tradition going back to Jackendoff (1972) that takes epistemic modals to be raising predicates and deontic modals (a.k.a. root modals) to be control predicates. In the syntactic literature on modals, epistemic modals are uncontroversially assumed to be raising predicates. However, the claim that deontic modals are control predicates is not universally accepted. Nauze (2008) offers evidence that deontic modals are often control predicates, but he also points out that the syntactic properties of modals differ significantly cross-linguistically. In the syntactic literature on French, the traditional view argued for in Dubois (1969: 119), Kayne (1975: 259), and Huot (1974: 171–172) takes epistemic modals selecting an infinitival clause to be raising verbs, and deontic modals to be control verbs. This view is largely substantiated by the work of Reed (2016, 2019), but she shows that unlike deontic falloir ‘be necessary’, deontic devoir ‘must’ does not fit the traditional mold in that it is ambiguous between a control and a raising entry. The latter allows deontic devoir, but not falloir, to merge with infinitival complements headed by a verbal element that is not theta-associated, as (41) illustrates.^[14]

Reed analyzes the raising option for deontic devoir in (41) with a structure in which devoir takes a bare vP infinitival as its complement. Since vP is standardly assumed to be phasal, we therefore expect ellipsis to be possible. This expectation is fulfilled as shown in (42).

As Reed shows, deontic devoir also participates in a control structure and since control infinitives, unlike raising infinitives, are standardly assumed to be phases (see e.g., Chomsky 2000: 105) we again predict ellipsis to be possible. This prediction is borne out, as (43) illustrates.

While the assumed size of control infinitives like that elided in (43) varies from author to author, depending on the particular theory of control they assume, what is important for the view of clausal ellipsis defended here is that control infinitives be phasal.^[15] Raising infinitives of the TP variety, on the other hand are headed by defective T, a non-phasal category, from which it follows that modals that embed them; namely epistemic modals, do not license clausal ellipsis, as illustrated in (44).

Thus, French modal ellipsis requires two ingredients. First, the modal verb must be associated with an E-feature (either covert or phonologically realized as special le) and second, the clausal complement targeted for ellipsis by this modal verb must be phasal. The latter requirement in fact extends beyond modal ellipsis (i.e., type 1 ellipsis). Consider, for example, the case of a raising predicate like paraître ‘appear’, which belongs to the class of verbs that license type 2 clausal ellipsis in comparatives. As shown in (45), paraître can take as a clausal complement either a tensed CP, or a raising infinitival TP. Assuming that ellipsis only targets phases, we then expect their tensed CP complement, but not their raising infinitival complement to be elidable. That this is, in fact, what happens, was noted by Vergnaud (1975). The relevant contrast is illustrated in (46).

Thus, we now have a strong connection between type 1 and type 2 clausal ellipsis. Both display a closed set of verbs that act as ellipsis licensors, both involve the optional spell out of an E-feature on the licensor as special le and both involve ellipsis affecting a phasal projection. My assumptions regarding modal ellipsis are illustrated by the (partial) derivation in (47b), which corresponds to the bracketed portion of the sentence in (47a). I follow Chomsky (2000) in assuming that control infinitives project up to CP, but my analysis is compatible with other projections proposed for control infinitives as long as such projections are assumed to be phasal.

(47)

a.

Ils	n’utilisent	pas	ses	arguments,	mais	[ils	(le)	pourraient].
they	Neg-use	not	these	arguments	but	they	(LE)	could
‘They don’t use these arguments, but they could.’

b.

In (47b), the ellipsis licensor, which is the modal pourraient ‘could’, bears an E-feature (optionally realized as le) and tags for ellipsis the complement to the phasal head of the infinitival CP it selects.^[16] However, the actual transfer of this TP to the PF component as unpronounced is not triggered until the next phase head (the higher v in (47b)), is introduced in the derivation and values the E-feature on pourraient. Consequently, elements that move to the edge of CP prior to the valuation of the E-feature on the licensing head can escape the ellipsis site. This correctly predicts that wh-extraction out of the TP elided in French modal ellipsis is allowed, yielding sentences like (5b), repeated here for convenience.

Note finally, that the overt option of the E-feature, namely special le, is to be understood as its morphological realization on the licensing head in very much the same way overt case morphology on a noun realizes the abstract Case feature it is associated with. Special le is, however, slightly different in that it is additionally a clitic and as such, it must attach to the highest verbal element in its clause.

In the next section, I turn to a closer examination of clausal ellipsis in comparative constructions.

6 Type 2: clausal ellipsis in comparatives

Type 2 clausal ellipsis, found in comparatives standards, has three basic properties that any analysis should account for. First, it requires the movement of a phrase to the left edge of the elided clause. This is illustrated by contrasts like (48).

In (48a), the le that may appear in the comparative standard has a silent counterpart and is therefore special le. In (48b), on the other hand, we see that in the absence of wh-movement, the le that occurs in the second sentence has no silent counterpart and is therefore a pronoun. Thus, special le on a verb like croire ‘believe’ signals clausal ellipsis only when there is a remnant (ce que ‘what’ or its silent counterpart in (48a)) that has escaped the ellipsis site.

A second property of this type of ellipsis is that it cannot follow an overt complementizer, as shown in (49).

Finally, the class of verbs that can embed this type of ellipsis is restricted. This class seems to correspond to so-called ‘bridge verbs’, a class that includes verbs like penser ‘think’, croire ‘believe’, imaginer ‘imagine’, escompter ‘expect’, supposer ‘suppose’, prétendre ‘claim’, prévoir ‘anticipate’, parier ‘bet’, espérer ‘hope’, etc. Verbs that are incompatible with clausal ellipsis (cf. (50)) include ‘manner of speech’ verbs like murmurer ‘murmur’, factive verbs like savoir ‘know’, découvrir ‘discover’, or être surpris/content ‘be surprised/happy’, and so-called ‘response stance’ verbs like nier ‘deny’, accepter ‘accept’, or être d’accord ‘agree’.

Interestingly, the distinction between these two classes of verbs is not just relevant to the licensing of embedded clausal ellipsis. Indeed, it has been shown to affect the availability of verb-second word order in the subordinate clauses of some Germanic languages. For example, Vikner (1991) argues that embedded verb-second in Danish and the other Mainland Scandinavian languages is limited to the complements of bridge verbs, and de Hann and Weerman (1985) make similar observations regarding embedded verb-second in Frisian. These authors argue that bridge verbs are special in that they license CP-recursion, and this recursion provides an additional embedded C head and [Spec, CP] positions to which the finite verb and the topic phrase can move (see also Iatridou and Kroch 1992). The idea that clausal complements to bridge verbs are cross-linguistically structurally more complex than those of non-bridge verbs has been echoed by many since (see de Cuba and Ürögdi 2010; Haegeman 2006; McCloskey 2006; Weir 2014; Yasui 2021; among others). Why this should be so has received various explanations, but a leading idea is that bridge verbs embed a speech act and therefore a double complementizer (a.k.a. CP-recursion) structure, which I will represent, following Weir (2014), as in (51).

In (51), CP2 denotes a proposition without illocutionary force. CP1, on the other hand, is required to create a speech act and only it can license ellipsis (i.e., can be endowed with a E-feature). For de Cuba and Ürögdi (2010: 45), CP1 (which they call cP) corresponds to a non-referential semantic object denoting a speech act, which they define as an unresolved proposition or an open question. Weir (2014) simply states that non-bridge verbs denote relations between individuals and propositions, while bridge verb embed an assertion speech act syntactically encoded by C1. I will assume here that C1 contains ASSERT, a silent category that denotes a speech act in which a state of affairs is claimed by the speaker to hold. Thus, if p is a proposition, then ASSERT p puts p as true in the Discourse Commitment Set of the speaker and puts it up for ‘discussion’ in the QUD (Question under Discussion). Only C1 can be associated with an E-feature. Whenever this happens, C1 becomes a licensing head for ellipsis and, being a phasal head, it targets its complement CP2 for non-pronunciation at PF. The E-feature on C1 is then valued by the ellipsis trigger, which is the first v dominating CP1, at which point CP2 is sent to PF as unpronounced. This allows wh-elements to move to [Spec, CP1] to escape ellipsis and, assuming with Weir (2014) that complementizers head C2, we correctly predict that this type of ellipsis can never be preceded by an overt complementizer (cf. (49)).^[17] Finally, I assume that the overt realization of the E-feature as le originates on C1 but, being a clitic, special le must move to attach to a verbal element, in this case the bridge verb that selects CP1.^[18] Note that this is an instance of head movement that obeys the Phase Impenetrability Condition (PIC) because the first merge position of special le is C1, which stands at the edge of a phasal projection and is therefore available for further computation. A partial derivation that sums up my analysis of CP2 ellipsis with special le is provided in (52).

(52)

a.

Soline	gagne	plus	d’argent	que	(ce qu’)	elle	le	dit	[qu’elle en  gagne].
Soline	earns	more	of-money	than	(what)	she	LE	says	that-she of-it earns
‘Soline makes more money than she says.’

b.

The only remaining question is why ellipsis, when licensed by C₁, requires the movement of a phrase to the edge of its licensor. Requiring that the E-feature on C₁ be accompanied by the insertion of an EPP feature (as Merchant 2001, 2004 does for fragment answers and sluicing) may seem like a move that just restates the problem. However, there are other cases in which two features function in tandem (i.e., introducing one in the derivation cannot be achieved in the absence of the other). For example, whatever feature triggers V to T movement in French can only be associated with T if the latter also has a tense feature. I will therefore assume that features sometimes cluster in this manner on particular heads, the clustering of an E-feature and EPP on C₁ being just one example of a larger phenomenon.

7 Clausal ellipsis in sluiced sentences

Sluiced sentences license type 2 clausal ellipsis in two distinct contexts. The first context is that of long distance wh-movement and only applies to the embedded clause when the latter is selected by a bridge verb. The example in (53) illustrates this phenomenon.

The ellipsis that occurs in (53) straightforwardly falls out of the analysis developed in the previous section, the only difference being that the combien ‘how much’ number wh-operator that transits through the specifier of CP₁ in (53) cannot be silent or be spelled out as ce que ‘what’ in the superordinate clause as it does in comparatives.

The second context in which clausal ellipsis occurs in sluiced sentences is what is commonly known as (non-matrix) sluicing. The examples in (54) illustrate two potentially distinct facets of this phenomenon.

Indeed, in the literature on indirect questions, some authors have argued that the examples in (54a) and (54b) have distinct semantic and syntactic properties (e.g., Berman 1990; Matos and Brito 2013; Plann 1982; Suñer 1999), while others (e.g., Ginzburg and Sag 2000; Kartunnen 1977; Lahiri 2002) do not share this view. At the heart of the controversy is whether indirect questions like (54a) (called ‘proper indirect questions’ by Suñer 1999) differ semantically from indirect questions like (54b) (called ‘improper indirect questions’ by Suñer 1999) and whether this difference is reflected in their syntax. Berman (1990) espouses the view that wh-complements of predicates like se demander ‘wonder’ and savoir ‘know’ differ semantically as well as syntactically. The former are questions (i.e., denote Hamblin sets) and are syntactically realized with a Q-morpheme in their C position while the latter are propositions and lack such a morpheme. A similar characterization is offered in Matos and Brito (2013). The assumption is mainly based on the fact that paraphrases of proper indirect questions include the expression ask the question, but paraphrases of improper indirect questions include know/discover/guess/reveal the answer. In what follows, however, I will present evidence that this view is misguided.

As pointed out by Matos and Brito (2013), the strict view advocated by Plann (1982) and Suñer (1999) that only verbs with an inquiring content like ask and wonder can select proper indirect questions is untenable because predicates that normally select improper indirect questions like know (cf. (55a)) may license proper indirect questions when they are in the scope of modal verb like want or would like, as (55b), which is nearly synonymous with (55c), shows.

The problem exhibited by (55a) versus (55b) is that, given that syntactic selection is local, at the stage when the wh-complement merges with savoir, both derivations must be syntactically identical and this means that if the semantic distinction between proper and improper indirect questions is indeed warranted, it has no syntactic correlate. In fact, Berman’s (1990) claim that only proper indirect questions contain a Q-morpheme can be falsified by observing the distribution the overt French Q-morpheme est-ce que. As shown in (56), est-ce que can appear in both proper and improper indirect questions, suggesting that they undergo similar derivations.^[19]

Additional evidence that proper and improper indirect questions are syntactically the same is provided by Lahiri (2002) who points out that Right Node Raising does not distinguish between the two types of predicates. This is illustrated for French in (57).

Finally, Krifka (2001) argues that both proper and improper questions embed question acts, but that verbs like know (or savoir in French) type-shift this question act to the set of true answers. An advantage of this view is that it can account for phenomena relating to quantification into embedded interrogatives. To explain, Krifka (2001) proposes that the pair list reading of questions with quantifiers, as expressed by the paraphrase of (58), reduces to quantification into speech acts.

As he argues, the interpretation of quantifiers rests on conjunction, and we should therefore expect universal quantifiers, which are generalized conjunctions, to be able to scope over speech acts. With this background in mind, consider the contrast between (59a), in which know embeds a that-clause and (59b), in which know embeds an improper indirect question.

As (59) makes clear, the apparent wide-scope reading of the universal quantifier is possible when know embeds an improper question, but not when know embeds a that-clause. The contrast can be explained if we assume that when know embeds a that-clause, it embeds a proposition/sentence radical and therefore a single CP-layer, but when know embeds an improper indirect question, it additionally embeds a speech act, syntactically encoded via CP-recursion (i.e., an improper question involves C₁ containing QUEST and C₂ either covert or phonologically realized as est-ce que in French).^[20] If this is on the right track, then ellipsis in embedded questions falls under the same type of analysis developed in the previous section; that is, ellipsis of an embedded clausal complement is possible only in the context of CP-recursion. The higher C (i.e., C₁) is phasal and is associated with an E-feature, which allows ellipsis of the lower CP projection.^[21]

There is, however, one last remaining issue to be addressed, namely why clausal ellipsis of embedded questions does not, in French, allow the E-feature on C₁ to be overtly realized as the clitic le, as (60) shows.

This may appear surprising in view of the fact that the (argumental) pronominal le can substitute for an embedded interrogative, as (61) shows, from which it follows that the restriction observed in (60) cannot be due to a constraint on cliticization with verbs like savoir ‘know’ and se demander ‘wonder’.

Before we resort to invoking general grammatical principles to account for the constraint illustrated in (60), however, we’d best remember that, as Authier (2023) points out, the E-feature now obligatorily spelled out as le in French predicate ellipsis, used to be phonologically unrealized up to the first half of the XX^th century (see also Eriksson 2006 for a variety of examples involving predicate ellipsis without le). In other words, the availability of the overt versus covert realization of E-features in French appears to be subject to time and construction specific fluctuations. Thus, I will assume that the absence of le in elided interrogative contexts like (60) is the result of an idiosyncratic choice favoring the use of a silent E-feature in sluiced constructions rather than a grammatical constraint on the overt realization of an E-feature dictated by a particular syntactic derivation. When the grammar of a language allows two options, as French does with overt and covert E-features, arbitrary choices of this type are not uncommon. A case in point is the competing derivations found in French causatives with the matrix verbs laisser ‘let’, faire ‘make’, entendre ‘hear’ and voir ‘see’. In European French, up to the first half of the XXth century, all four verbs could embed both an ECM infinitival and a structure in which the infinitival verb raises over the embedded subject, as (62) illustrates.

The ECM option in (62a) has, however, been abandoned for faire in Modern European French, though it subsists in some Canadian French dialects (see Reed 1992). The most natural explanation for this is that modern speakers of European French reject sentences like (63) not because they violate some principle of the linguistic grammar of French, but rather because their dialect arbitrarily restricts the two options made available by the grammar to one for this particular verb.

What I would like to suggest is therefore that a similar arbitrary restriction applies to sluices. That is, the E-feature that licenses clausal ellipsis can, in principle, be silent or phonologically realized as the clitic le. However, when it appears on a higher C corresponding to QUEST, the overt option is arbitrarily discarded, hence the E-feature is always silent.

8 Clausal ellipsis in ACD relatives

In Section 2.2, it was established that type 2 clausal ellipsis (with and without special le) can be found in ACD free relatives that embed a bridge verb. This is illustrated in (64).

Clausal ellipsis in ACD contexts is not, however, limited to free relatives; it can also occur with headed relatives, as shown in (65).

The sentences in (64) and (65) all involve the ellipsis of the clausal complement to a bridge verb embedded in the ACD relative. That the presence of this bridge verb is crucial to the licensing of type 2 clausal ellipsis in such cases is evidenced by the fact that so-called relative sluicing is prohibited in French, as shown in (66).

The ungrammaticality of sluices like those in (66) can be attributed to the fact that relatives do not encode a speech act and do not, therefore, have a recursive CP structure. This correctly filters out sentences like (66) while predicting the grammaticality of sentences like (64) and (65). That is, in a sentence like (65a), clausal ellipsis is licensed by the presence of a higher CP-projection selected by prétend along with the presence of a silent copy of the relative pronoun in its specifier.^[22]

As it stands, this explanation is, however, unable to explain why relatives with embedded ellipsis like (64) and (65) are contingent upon the presence of an ACD configuration. That is, standard relatives do not license clausal ellipsis even when they embed a bridge verb, as (67) shows. This is so despite the fact that the bridge verbs in (67) provide the requisite double CP structure for type 2 ellipsis to take place.

It thus appears that ACD relatives differ from standard relatives in such a way that they create the proper environment for clausal ellipsis to be licensed under a bridge verb. Importantly, this cannot be assumed to follow from a general condition on ellipsis licensing because type 1 (modal) ellipsis turns out to be perfectly acceptable within standard relatives, as shown in (68).

As compared to standard relatives, ACD relatives have two unique properties which could potentially help us solve this puzzle. I will use the rest of this section to examine each of them in turn and assess whether they play a role in the restricted distribution of clausal ellipsis in relatives under a bridge verb.

The first special property attributed to ACD relatives is that they cannot be interpreted in situ. It is well known that ellipsis of a phrase XP is only possible if the elided XP is identical to its antecedent XP at LF.^[23] This parallelism requirement suggests that ACD relatives must vacate the VP that contains them at LF, as first proposed by Sag (1976). To see why, consider the sentence in (69), which illustrates the case of an English ACD relative with an elided VP.

If the relative headed by every book were interpreted in its base position, the antecedent VP would contain the elided VP and the two VPs would not be identical. Further, replacing the elided VP with the antecedent VP would lead to infinite regress; that is, a regress into an infinite sequence of propositions in an attempt to interpret the sentence. If, however, the ACD relative is interpreted in a VP-external position, the parallelism requirement is satisfied (cf. (70)) and the infinite regress problem disappears.

The classical analysis of ACD (cf. Sag 1976 and much subsequent work) implements this idea by assuming that the DP hosting the relative clause undergoes covert movement to create an LF configuration where the elided VP is no longer contained inside the antecedent VP and both VPs have traces in their object position. However, while this movement has frequently been assimilated to QR, it is important to note, as Hackl et al. (2012) do, that the movement of an object relative clause hosting an ACD site is independent of the quantificational status of its DP head, as evidenced by the fact that it can occur even when the head of the relative hosting the ACD site is a definite DP. This is illustrated in (71), where (71b) is a possible LF representation of (71a).^[24] Thus, the LF movement of ACD relatives cannot be QR in its usual sense.

This has led to analyses that share the assumption that ACD relatives undergo late merger to circumvent a violation of Parallelism and avoid the problem of infinite regress (and thus “antecedent contained deletion does not exist”, as first argued in Baltin 1987). The details of such analyses vary, however. Fox (2002) proposes that the first step in the derivation of an ACD relative is rightward movement of the relative head. This movement is followed by the late merger of the ACD relative as an adjunct, an operation which is needed to eliminate antecedent containment. In other words, ACD relatives involve a special, more complex derivation of relative clauses so as to ensure convergence at LF. Fox’s (2002) derivation of ACD relatives is schematically illustrated in (72).

A slightly different version of this type of analysis is Ding’s (2013) Split Relative Clause Account. As illustrated in (73), Ding’s proposal does not involve rightward movement of the external head of the ACD relative. It assumes instead that the relative head is merged within the antecedent vP while the relative clause containing the ACD site is late merged as a vP-adjunct.

Thus, late merger approaches to ACD structures eliminate antecedent containment and satisfy Parallelism, but do they predict clausal ellipsis in examples like (64) and (65)? It seems possible to argue that they do if we assume with Sakamoto (2016) and Takahashi (2024) that (a) different clauses can be constructed independently within the workspace, (b) the elements used in a phase that undergoes Transfer can be reused to construct another phasal category which is then merged countercyclically, and (c) reused elements do not have phonetic content because they have been transferred. Under such assumptions, a sentence like (74a) could plausibly be derived in the fashion schematically illustrated in (74b).

In this derivation, the step in (74c) is the crucial one. The occurrence of the embedded CP₁ is reused and therefore remains phonologically unrealized when the structure in (74c) is countercyclically merged with that in (74b). Thus, clausal ellipsis in sentences like (74a) can be made to follow from a derivational copying view of ellipsis. I will, however, discard this type of analysis on grounds that it makes use of an ellipsis mechanism that both undergenerates and overgenerates: It undergenerates because it says nothing about clausal ellipsis in sluiced sentences, which do not involve late merger, and it overgenerates because, assuming that clausal adjuncts in general must (Stepanov 2001), or at least may (Abe 2018), undergo late merge, it wrongly predicts that sentences like those in (75) should license clausal ellipsis. That is, as we have seen in previous examples, type 2 clausal ellipsis under a bridge verb is only licensed if a remnant is extracted from the ellipsis site. As this is not the case in (75), clausal ellipsis is prohibited, as evidenced by the fact that the le that appears in such examples must be overt and is therefore a pronoun.

I conclude that late merge alone cannot correctly predict clausal ellipsis under bridge verbs in ACD relatives. However, ACD relatives exhibit a second unique property, which I will argue plays a role in licensing type 2 ellipsis. I am referring here to the quantifier scope constraints unique to ACD contexts discussed in Koster-Moeller and Hackl (2008) (hereafter KM&H).

KM&H start with the observation that in standard relatives, both surface scope and inverse scope are possible for a matrix subject (a professor in (76)) and what they call the host DP object (every article) regardless of the scope properties of the subject DP inside the relative. This is illustrated in (76).

However, in ACD relatives containing an elided VP, while the inverse scope reading remains available if the subject inside the relative is an existential DP, it becomes difficult to access if the subject is a proper name or a universal DP. This is shown in (77), where the # symbol indicates that the inverse scope reading is much more difficult to access than it is in (76).

As it turns out, the facts uncovered by KM&H are not specific to VP-ellipsis. Similar facts obtain in French ACD relatives containing an elided clause complement to a bridge verb, as illustrated by the contrast between (78a) and (78b–c).

As KM&H point out, paradigms like (76) versus (77) support the generalization that the factor that determines the presence versus absence of the inverse scope restriction in ACD relatives is whether the host DP is scopally commutative or non-commutative with the subject DP inside the relative clause.^[25] They state this as the (slightly reworded) ACD-scope generalization in (79).

As they point out, this generalization hinges on the fact that the host DP interacts scopally with QP₂, something that can only happen if the quantificational determiner of the host DP is also found inside the relative. This, they argue, suggests that ACD relatives are derived via an amended version of the matching analysis of relatives in which there is a full copy of the host DP inside the relative that raises to [Spec, CP] and undergoes deletion under identity when the relative is countercyclically merged with the external copy of the host DP. Further, as they point out, the presence of an ACD seems to be the ‘driving force’ behind the special structure exhibited by this type of relative.^[26] Before discussing the details of KM&H’s derivation of ACD relatives, let us consider the implications of their proposal in view of reaching a better understanding of the difference between standard relatives, which do not license ellipsis under a bridge verb, and ACD relatives, which do. Let us begin by discussing the nature of the so-called relative pronoun in standard relatives like (80), abstracting away from whether (80) is the result of a raising or a matching derivation.^[27]

First, we must assume that que ‘that’ in (80) is a complementizer rather than a wh-element as the contrast between (81a) and its interrogative counterpart, given in (81b), suggests.

However, assuming the presence of a silent wh-operator (OP) in (80) is crucial because this operator is a necessary ingredient to explain the interpretation of the relative: only via its movement can a trace/variable be created that turns the CP relative clause into an unsaturated expression; that is, a predicate of type <e,t>. This predicate is then semantically conjoined with the head noun (cf. Heim and Kratzer 1998) and the conjunction of these two predicates results in an intersection such that the restrictor of the head noun is enriched with more semantic content. Finally, semantically, connecting the relative clause to its head amounts to making an identity claim; that is, a sentence like John knows the girl that I saw says that the girl John knows is the same girl as the one I saw. I will assume that this identity relation is encoded by a feature on C, as suggested by the fact that it can be explicitly realized as as in e.g., the English dialect of Somerset County in Southwest England (cf. 82).^[28]

Concerning the operator/relative pronoun in (80), several remarks are in order. First, it is standardly assumed to move from an argument position to the [Spec,CP] position in the relative. Second, and more importantly, relative pronouns are minimal pronouns in the sense of Landau (2015: 23); that is, they bear unvalued φ-features in addition to their wh-feature. These unvalued features are then valued in the course of the derivation via Agree between the head noun and the relative pronoun (cf. Heck and Cuartero 2013 among others). Morphologically, this translates into their being silent (as in (80)), or displaying an invariant form – what in the English dialect of Somerset (cf. (83)), or fully reflecting the result of featural valuation as illustrated for French in (84).

Thus, the picture that emerges is that while binders generally bear unique indices, in standard relatives, relative pronouns bear unvalued features that are valued by the features of the DP relative head via Agree in the syntactic component. In contrast, the evidence uncovered by KM&H suggests that in ACD relatives, the full host DP of the relative clause is scopally active and must therefore be part of the internal derivation of the relative, or to put it slightly differently, the fully valued determiner of the host DP needs to be present in both the ellipsis antecedent and the elided constituent. In view of this crucial difference between standard and ACD relatives, I would like to propose that type 2 clausal ellipsis requires that the wh-remnant that transits through [Spec, CP₁] be fully valued and that this is the case in ACD relatives, but not in standard relatives. Why this should be so can then be made to follow from the assumptions in (85).

With this in mind, I now turn to a critical evaluation of KM&H’s derivation of English ACD relatives with VP-ellipsis, the steps of which are sketched in (86). In (86a), the matrix object, DP₁, is built separately from the ACD relative, which also contains a copy of DP₁, labeled DP₂. HM&K argue that it is at this point in the derivation that ellipsis is licensed, giving both the antecedent constituent and the elided constituent access to identical copies the host DP/relative head. As shown in (86b) the relative clause is then late merged into the host DP, giving rise to a single tree in which the relative clause internal copy, DP₂, is deleted under identity with DP₁. Finally, to produce a coherent semantic interpretation, further operations must take place, namely trace conversion of DP₂ (cf. Fox 2002), going from type <et,t> to e, followed by type shifting of the converted DP₂ from type e to type <e,t>, a predicate.

(86)

Mary read every book John did.

(a)

(b)

HM&K’s proposed derivation of ACD relatives raises some unanswered questions, however.

Consider first the process by which PF-deletion of DP₂ occurs in (86b) under identity with DP₁. This process of deletion under identity is not implausible. It is, in fact, reminiscent of Bresnan’s (1973) so-called Comparative Deletion. This similarity is not entirely surprising since comparatives of equality equate the degrees to which an object possesses a property (I ate as many fries as you did) and relatives equate the elements for which properties hold. However, given the structure in (86b), DP₂ and DP₁ cannot plausibly be identical (under any definition of identity) because DP₁ does not form a constituent that excludes DP₂ or, to put it slightly differently, DP₁ and DP₂ cannot be identical because the former contains the latter.

A second issue concerns the movement (internal merge) of DP₂ to the specifier of CP in (86a). Given that DP₂ is, in this particular case, a universally quantified DP, such movement is surprising because it can create unbounded dependencies, as (87) suggests, yet DP₂ does not overtly display the morphological marking of a wh-element.

A more accurate representation of DP₂ in (86) might therefore be as in (88), that is, a wh-phrase headed by a silent wh-determiner (OP) that takes the universally quantified DP₂ as its complement.

The overall makeup of what I labeled WhP in (88) may seem unusual in that it contains both a wh-element (OP) and a quantificational determiner. There is evidence, however, that this syntactic combination is, in principle, allowed by UG. This evidence comes from wh-questions in those dialects of English that display the so-called what/who/where all phenomenon. These include dialects of Northern Ireland (McCloskey 2006) as well as American dialects found in North Carolina and the Ozarks (Lindemann 2008). Lindemann’s (2008) examples include the ones in (89), both of which overtly display the syntactic sequence ‘wh-Q-restriction’ posited for the WhP in (88).

I therefore would like to propose that ACD relatives are countercyclically adjoined to VP and that the relative clause internal copy (DP₂ in (86a)) is deleted by the operation Form Copy (FC) argued for in Chomsky (2021). FC is an operation that selects an element X and searches its c-command domain for a structurally identical element Y. As it identifies Y, it assigns the relation <X, Y> and Y deletes. Because FC has no access to the derivational history of a phrase marker, it applies to identical elements regardless of how they were introduced in the derivation, and it can therefore apply to identical elements that were independently introduced into the derivation by external merge. I will further assume with Saito (2024) (and contra Chomsky 2021) that if FC selects an element X in an A-position, its search for Y is not restricted to A-positions and that it can therefore be connected by FC to an element Y in an A′-position. Finally, I will assume that direct objects undergo internal merge to form [Spec, VP] so that VP can be properly labeled, as proposed in Chomsky (2015).^[30] Given these assumptions, the (partial) derivation of a French ACD relative like (74a), repeated here as (90a), is assumed to be as in (90b).

(90)

a.

Il	a	obtenu	tous	les	livres	qu’il	(l’)espérait	[qu’il   obtiendrait].
he	has	gotten	all	the	books	that-he	(LE)expected	that-he would-get
‘He got all the books he expected (he would).’

b.

In (90b), [OP tous les livres] in the relative clause externally merges as the object to obtiendrait then internally merges in [Spec, CP₁] and finally, internally merges in the specifier of the CP relative clause. After two applications of FC, the lower copies delete, leaving [OP tous les livres₃]. In the main clause, tous les livres externally merges as the object to obtenu ([tous les livres₂]), then internally merges in [Spec, VP] as [tous les livres₁], at which point FC applies and deletes [tous les livres₂]. Finally, once the relative is late merged, FC applies to [tous les livres₁] in the matrix and [tous les livres₃] in the relative and the latter deletes.^[31]

In summary, my goal, in this section, was to provide an answer to the question of why ACD relatives, but not standard relatives, allow clausal ellipsis under an embedded bridge verb. I argued that, in order to avoid infinite regress, ACD relatives force a derivation that differs from both the raising and matching derivations that have been shown to be available for standard relatives. Specifically, ACD relatives are late merged. This more complex derivation is necessary for convergence; that is, to avoid an LF crash due to infinite regress, and is not available to standard relatives for economy reasons; that is, it is not necessary for convergence and therefore prohibited. I further argued that late merged ACD relatives contain a full copy of the relative head, a conclusion reached independently by KM&H based on evidence pertaining to quantifier scope. Finally, I argued that this full copy constitutes, in the specifier of C₁, the type of remnant that licenses type 2 clausal ellipsis; that is a remnant merged with fully valued φ-features. Because they undergo late merge, ACD relatives are built separately in the workspace and consequently, the remnant that undergoes A-bar movement must be fully valued as it cannot rely on cyclic merge to have its features valued through Agree. In standard relatives, on the other hand, relative pronouns are minimal pronouns that bear unvalued φ-features (in addition to their wh-feature). These unvalued features are then valued in the course of the derivation via Agree between the head noun and the relative pronoun. As such, run-of-the-mill relative pronouns, having unvalued features when they transit through the specifier of C₁, are not appropriate remnants for type 2 clausal ellipsis.

9 Concluding remarks

In this final section, I would like to take stock of the main findings of this article. First, an examination of French clausal ellipsis confirms the existence in French of an overt morphological correlate to the abstract E-feature argued for by Merchant and much subsequent work on ellipsis, namely, special le. Authier (2023) was the first to argue that le currently functions as an E-feature in French predicate ellipsis but used to be phonologically unrealized at previous stages of the language. An examination of clausal ellipsis has revealed that, with the exception of sluices, both overt and covert versions are available, though in different registers. This suggests that the overtness versus covertness of the E-feature in French ellipsis varies along register and construction specific lines. Second, it was established that there are, in French, two distinct types of clausal ellipsis. Type 1, exemplified by so-called modal ellipsis, only requires the pairing of an E-feature with a modal verb that takes a phasal complement. Type 2, exemplified by sluices and the ellipsis of the complement to a bridge verb, is instantiated by the pairing of an E-feature with the higher C-projection (C₁) of a CP-recursion structure that is licensed by the presence of a speech act. It was further argued that when C₁ is endowed with an E-feature, it must also be associated with an EPP feature, which is but one example of feature clustering among many others. A direct consequence of this is that type 2 clausal ellipsis requires, rather than allows, wh-extraction of a remnant in order to be licensed. Finally, it was shown that relative clauses that embed a bridge verb license type 2 clausal ellipsis only when they are of the ACD type. Regardless of any analysis, this observation entails that standard and ACD relatives undergo distinct derivations, a conclusion previously reached Koster-Moeller and Hackl (2008) on the basis of the quantifier scope restrictions exhibited by English ACD relatives involving VP-ellipsis. It was argued that, in order to avoid infinite regress, ACD relatives force a more complex/costly derivation: they must be late merged. It was further argued that late merged ACD relatives contain a full copy of the relative head which constitutes, in the specifier of C₁, the type of remnant that licenses type 2 clausal ellipsis; that is, a remnant merged with fully valued φ-features. In standard relatives, on the other hand, relative pronouns are minimal pronouns that bear unvalued φ-features that are valued in the course of the derivation via Agree between the head noun and the relative pronoun. As such, run-of-the-mill relative pronouns, having unvalued features when they transit through the specifier of C₁, are not appropriate remnants for type 2 clausal ellipsis.