This article puts forward two distinct arguments regarding the condition on identity between antecedent and ellipsis site that governs the grammaticality of sluices. The first argument is that the viability of a requirement of syntactic identity has been too hastily dismissed. Such a condition is viable if syntactic identity is not assessed over the entire deleted constituent, but instead is assessed head-by-head for each head stranded in the ellipsis site. This allows syntactic differences associated with material that has moved out of the ellipsis site to not affect the calculation of syntactic identity. The second argument is that the bestiary of possible mismatches under sluicing can be given a uniform syntactic characterization: all and only material originating outside of the verbal complex can be mismatched under sluicing. The restriction of identity conditions to the verbal complex is implementable in many (but not all) approaches to ellipsis identity; I provide a concrete application of it to the proposed head-based syntactic identity condition.
In (1a), a lone wh-element seems to occupy a position in which we would expect to find a full CP. (1b) illustrates the standard theory of what is going on in (1a): the sentence is interpreted as if the wh-element were followed by the crossed-out material; the ability to make the inference that the wh-element should be interpreted as scoping over that material (and therefore the well-formedness of the entire sluicing construction) is determined by the presence of a suitable antecedent clause (boldfaced).
I will assume here, following the seminal work by Merchant (2001), that the crossed-out material in (1b) represents unpronounced syntactic structure.1 This unpronounced constituent is present in the narrow syntax, but is eliminated from the phonological realization of that string; for this reason, I will refer to the crossed-out content as elided material or as the ellipsis site. I assume, still following Merchant (2001), that sluicing is always deletion of a full TP. Sluicing is restricted to clauses headed by null [+wh, +Q] Cs (see Merchant 2001:sec. 2.2.1). In this article, I do not analyze the procedure by which TP-deletion is triggered, nor why this triggering is restricted to such Cs. Rather, my exclusive focus is the identity condition on the well-formedness of sluicing; that is, my goal is to characterize the relation that must hold between the antecedent and the ellipsis site in order for a sluice to be grammatical.
To be more precise, my goal is to characterize specifically the syntactic aspect of the identity condition on sluicing. Sluicing, like all ellipsis, is a process the full understanding of which requires consideration not just of syntax, but also of phonology and of meaning (whether it enters the picture via semantic identity conditions, as proposed by Merchant (2001) and AnderBois (2014), via pragmatic restrictions on the availability of interpretations for the ellipsis site, as proposed by Kroll and Rudin (2017) and Kroll (to appear), or perhaps both). Because what I offer here is only a theory of syntactic identity restrictions on sluicing, it must be married with accounts of the role of sound and meaning in sluicing before we can have a full picture of the restrictions on its acceptability. That said, I will suggest that a syntactic identity condition can supplant extant semantic identity conditions, though I stop short of arguing that it must do so. The worldview most compatible with what I will argue for here is that the role meaning plays in sluicing is primarily pragmatic.
This article is split into two main parts, each of which puts forward one idea about sluicing. These two ideas are in principle independent of each other; however, they interact in a way that justifies their inclusion in a single article. Viewed together, I believe they comprise a fuller sketch of a theory of the identity condition on sluicing than they would if considered separately.
In section 1, I present empirical arguments that strongly suggest that some form of syntactic identity between antecedents and ellipsis sites must be enforced, and I examine contrasting arguments that have previously been taken to suggest that syntactic identity cannot be a condition on the well-formedness of sluices. I propose a novel way to think of syntactic identity that is flexible enough to avoid ruling out the grammatical cases taken to be problematic for syntactic identity conditions, without being so flexible as to avoid ruling out the ungrammatical cases that made syntactic identity conditions seem prudent in the first place.
In section 2, I present a broad range of data, much of it novel, about possible mismatches under sluicing that are problematic for all extant identity conditions. I propose that these mismatches can be given a unified syntactic characterization as being restricted to the portion of the clausal spine above the verbal domain, and I give a general solution: namely, that the domain of identity is strictly smaller than the ellipsis site—that sluicing privileges material originating within the verbal domain. The proposal that the domain of identity is strictly smaller than the ellipsis site can be ported into a wide variety of accounts of the identity conditions on sluicing, including semantic approaches to the identity condition. In this article, I show in detail how this idea can be implemented given the syntactic proposal in section 1.
The proposal of section 2 makes a concrete and novel prediction: that identity will be strict (all mismatches will be ungrammatical) within the verbal domain of the sentence, but that interpretations associated with mismatches above the verbal domain will come into and out of availability as mediated by a large variety of pragmatic factors (including the salience of the interpretation, its relevance to prior discourse, its relationship to the local context, and so on). This prediction has so far been borne out by corpus work carried out at the University of California, Santa Cruz under the aegis of NSF grant 1451819: The Implicit Content of Sluicing.
In section 3, I supplement these two main proposals with some brief thoughts about the role of sound and meaning in sluicing—though, as already stated, the purview of this article is predominantly syntactic.
Though this article focuses solely on sluicing, reflecting its origins in the corpus work mentioned above, in section 4, in addition to summarizing the arguments of the article, I discuss the potential for extending its proposals to other well-studied forms of ellipsis, such as VP-ellipsis.
1 Head-Based Syntactic Identity
1.1 Entailment-Based Approaches and Their Problems
The identity condition on sluicing that is most commonly assumed in the literature is a semantic condition originally proposed by Merchant (2001). On this account, a TP can only be deleted if it is e-GIVEN.2 The notion of e-GIVENness is an extension of Schwarzschild’s (1999) GIVENness; my presentation here is a mixture of Schwarzschild’s own formulation and Merchant’s deployment of his ideas.
An expression E counts as GIVEN iff E has a salient antecedent A and, modulo ∃-type shifting, A entails the F-closure of E.
∃-type shifting raises expressions to type
The F-closure of a constituent α (F-clo(α)) is the result of replacing F-marked parts of α with ∃-bound variables of the appropriate type (modulo ∃-type shifting).
For Merchant (2001), the role GIVENness plays in determining the grammaticality of TP ellipsis is this:
Focus condition on TP-ellipsis (e-GIVENness) (Merchant 2001:31, (61)–(62))
A TP E can be deleted only if it has a salient antecedent A and, modulo ∃-type shifting, E entails F-clo(A) and A entails F-clo(E).
In other words, modulo ∃-type shifting and F-closure, the elided TP must stand in a relation of mutual entailment to its antecedent. As expressed above, this mutual entailment condition is a necessary condition, but not a sufficient condition, on the well-formedness of a licitly triggered sluice. Recent work (e.g., Chung 2005, 2013, Merchant 2013b) has argued that this semantic identity condition must be supplemented by some sort of lexicosyntactic identity condition as well. Chung (2005) brings up data like the following:
Data like these appear problematic for a purely entailment-based account of the identity condition on sluicing: the only difference between (4a) and (4b) is that in the former a preposition has been pied-piped and in the latter it has been stranded—a purely syntactic difference—and yet the sluice in (4a) is licit and the sluice in (4b) is not.3
In both cases, the stranded wh-phrase has no overt correlate in the antecedent clause—what was christened “sprouting” by Chung, Ladusaw, and McCloskey (1995). If we add an overt correlate to the antecedent, (4b) becomes grammatical.
The generalization is this:
A preposition can be stranded in an ellipsis site only if it has an overt correlate in the ellipsis-licensing antecedent.
Chung (2005) proposes a lexicosyntactic condition to supplement the semantic identity condition; her condition states that sluices are well-formed only if the ellipsis site contains no lexical items that are not present in the antecedent. This condition directly stipulates the generalization in (6).
Though Chung acknowledges the provisional and unsatisfying nature of such a condition by characterizing it as “the beginnings of a proposal” (p. 82), it has been used elsewhere in the literature (e.g., in AnderBois 2014) as a ready-made lexicosyntactic patch on an otherwise purely semantic account of the identity condition.
This piece of data is an appropriate empirical starting point because it shows that there are nontrivial problems facing the idea that the identity condition on sluicing is based only on entailment relations between antecedents and ellipsis sites. That some sort of lexicosyntactic condition on sluicing is necessary has been clarified and expanded in its empirical scope in work by Chung (2013) and Merchant (2013b) (further relevant data will be encountered in section 1.2), but on both accounts the syntactic condition supplements a mutual entailment condition in the style of Merchant (2001). Such a split system, relying on both semantic and syntactic constraints on the grammaticality of sluicing, replicates the situation as it has stood for quite some time in the literature on VP-ellipsis, giving rise to the same uncomfortable feeling of redundancy—that the semantic and syntactic constraints must inevitably duplicate each other to a large extent, given standard ways of understanding how syntactic and semantic composition parallel one another. My goal in this section—given the apparent necessity of reference to syntax in the condition on ellipsis licensing—is to state the syntactic condition on sluice licensing in as powerful and general a way as possible, in hopes that doing so will provide us with a unitary, nonredundant conception of the identity condition on sluicing.
1.2 The Syntactic Condition on Sluicing
In this section, I will present a syntactic condition on the well-formedness of sluices. One of the goals of this article is to push this condition as far as it can go—as it seems a syntactic condition is necessary, perhaps a syntactic condition of the appropriate kind could supplant the entailment-based semantic condition on sluicing entirely, giving us a cleaner, simpler account.
Merchant (2001:19–20) cites sluices like the following as an argument against syntactic identity as a requirement on the grammaticality of sluicing:
Here and throughout, I present simplified syntax trees that abstract away from irrelevant complexities, and I represent the tails of movement dependency chains as traces, though the theory as presented is agnostic with respect to the theory of movement it assumes, and is perfectly compatible with a copy theory of movement. In (7), although the antecedent and the ellipsis site contain the same lexical items in the same order, in the antecedent the verb ate, an optional transitive, takes no internal argument; in the ellipsis site, ate does take an internal argument. The TPs are not syntactically identical, due to the presence of a structural object position in the elided TP, and therefore a condition that would require syntactic isomorphism between the antecedent TP and the elided TP fails.
In the arguments presented in Merchant 2001 against syntactic identity conditions in sluicing,4 as well as in the arguments presented in Merchant 2005, it is assumed that syntactic identity conditions for sluicing would be stated as conditions on identity between entire TPs. The syntactic identity condition that I propose here instead assesses syntactic identity head by head5 and is enforced only over items that are not tails of movement dependency chains; that is, lower copies/traces can be freely elided.6
Syntactic condition on sluicing (to be revised)
Given a prospective ellipsis site E and its antecedent A, ellipsis of any head h ∈ E is licit only if either h is a nonhead member of a movement dependency chain or h has a structure-matching correlate n ∈ A.
This definition is parasitic on two other definitions. We can define structure matching as follows:
A node n in domain d structure-matches a node n′ in domain d′ iff n and n′ are dominated by an identical sequence of immediately dominating nodes within d and d′.
This definition takes some unpacking. First, sequence of immediately dominating nodes for a given node n refers to the node n1 that immediately dominates n, followed by the node n2 that immediately dominates n1, and so on. Consider, for instance, the following tree:
In this tree, the sequence of immediately dominating nodes that dominates the head slap is as follows:
(11) TP - T′ - VP - V′ - V - slap
Talking about things in terms of the sequence of immediately dominating nodes that dominates a particular head, then, is a way of talking about its position in a tree regardless of what other lexical items may be in that tree.
Second, the domains d and d′. We can assume for the time being that the domain of identity d is the entire ellipsis site—the full elided TP—and that the domain d′ is the full antecedent TP.7 In section 2.1, this conception will be revised to a smaller constituent of those TPs. It is important to note here that the calculation of structure matching is based on the positions of the elided and antecedent heads in the surface syntax, regardless of their external Merge sites or the positions they end up in at LF. This is of a piece with an implementation of ellipsis as phonological deletion: if ellipsis happens at PF, the relevant representations must be surface representations.
For the moment, I will give the following characterization of what can serve as the correlate of a head:
Correlate (to be revised)
A head h can be a correlate for a head h′ iff h and h′ are tokens of the same lexical item.
This characterization will be revised in section 1.3.
To see how this criterion operates, let us revisit the case in (7), shown again here:
The antecedent and the ellipsis site are not syntactically isomorphic to each other; the difference lies in the presence of a structural object position in the ellipsis site. The content of that object position, however, is a trace (or unpronounced lower copy), which is to say the tail of a movement dependency chain; therefore, the rule in (8) does not require it to have a structure-matching correlate. I assume that the verb ate in the antecedent and the verb ate in the ellipsis site are lexically identical, with optional transitivity being a lexical feature of each. For it to be grammatical that the ate in the ellipsis site goes unpronounced as a result of sluicing requires only that it have a structure-matching correlate in the antecedent. To calculate structure matching in the sense defined above, it is only necessary to look at the head and at the chain of nodes dominating it.
In each case (abstracting away from irrelevant complexity), the verb is the head of a VP that itself is dominated by a TP. The elided verb has a structure-matching correlate in the antecedent (the same reasoning applies to the elided subject, which is not shown), and so the ellipsis is well-formed. The fact that one instance of the verb ate has an object and the other instance does not plays no role in the calculation; all that matters is that they are lexically identical and that they both occupy the same position within their respective TPs. Object position is irrelevant to the well-formedness of this particular sluice: it is too low to alter the domination chains of anything else in the clause, and its content has been “moved out,” obviating the need to find it a structure-matching correlate.8
It should be clear how an analysis like this accounts for Chung’s generalization in (6): if the preposition is stranded in the deleted TP, as in (4b), the ellipsis is ill-formed, because the preposition has no correlate in the antecedent; if the preposition is pied-piped out of the ellipsis site, as in (4a), then it needs no correlate, because it has “moved out,” leaving only a trace (or unpronounced lower copy). If a correlate for the preposition is present in the antecedent, as in (5), the preposition can be safely stranded in the ellipsis site.
This account has the added benefit of providing a clean explanation of the impossibility of various other mismatches in the verbal domain that are not possible under sluicing, despite not obviously following from an entailment-based account of identity. The clearest such cases are active/passive mismatches.
*Joe was murdered, but we don’t know who
*Someone murdered Joe, but we don’t know by who
Joe was murdered
Take, for example, (15b). In this case, the meaning of the antecedent clause—namely, that someone murdered Joe—is identical to the meaning of the elided clause—namely, that Joe was murdered (given the assumption that murder entails that the killer is a person). On its own, a mutual entailment account does not explain the unacceptability of the examples in (15), though perhaps it could be supplemented by a pragmatic explanation of their infelicity. The syntactic identity condition proposed here, however, explains why these voice mismatches are systematically impossible in sluicing. For one, there are significant differences in the structural positions of the arguments of verbs in active and passive sentences in English that result in those arguments’ not structure-matching each other. However, the most general way in which this account explains the impossibility of voice mismatches under sluicing is in terms of mismatching v’s. A passive v cannot serve as a correlate to an active v, as they are not lexically identical, and so we predict the impossibility of voice mismatches under sluicing crosslinguistically regardless of the morpho-syntactic reflexes of the passive in any particular language.
That causative/inchoative alternations under sluicing are impossible was also observed by Merchant (2005:17–18).
Data like this are less clearly problematic for entailment-based approaches to the identity condition, as there is potentially one subtle semantic difference between the antecedent and the ellipsis site here: the inchoative antecedent does not entail that the jug was broken by an agentive action, whereas the causative ellipsis site does. Regardless of whether cases like this are insurmountably problematic for entailment-based accounts, their ungrammaticality follows from the syntactic proposal at hand: the lexical items in the ellipsis site are not in the same structural positions as the correlating lexical items in the antecedent.
*They embroidered something with peace signs, but I don’t know what on
they embroidered peace signs
*They embroidered something on their jackets, but I don’t know with what
they embroidered their jackets
*They loaded something onto the truck, but I couldn’t quite make out with what
they loaded the truck
*They loaded one of the trucks with hay, but I couldn’t quite make out onto which truck
they loaded hay
(based on Merchant 2005:9, (80)–(81))
These examples also may not be insurmountably problematic for entailment-based accounts due to subtle differences in meaning between the relevant argument structure variants (Rappaport and Levin 1988, Dowty 1991), but again, the syntactic proposal here has no problem accounting for their ungrammaticality, given that there are differences in structural position between elided heads and their antecedent correlates in all cases.
Though these argument structure mismatches are all ruled out by the failure of elided heads to structure-match their antecedents, just as with the case of voice mismatches a more general formula for ruling out such mismatches is provided by the theory here: insofar as argument structure mismatches can be traced to different v’s, we expect argument structure mismatches to be impossible, as the elided v will not be able to find a correlate in the antecedent.9
The way the current account connects the impossibility of various argument structure mismatches, among other things, to the impossibility of mismatching lexically different v’s receives independent support from the fact that similar explanations have been proposed within accounts of ellipsis identity that are otherwise quite different from the one proposed here; for instance, Merchant (2013b) also invokes the role v plays in argument structure in his explanation of the impossibility of such argument structure mismatches.
1.3 Correlation and Coreference
Above, I proposed a syntactic condition to supplant entailment-based semantic conditions on sluicing. I caution against interpreting that as endorsing the notion that meaning plays no role in the identity condition that governs the well-formedness of sluices. I will propose, however, that the role of meaning in the identity condition is limited to the way in which reference is encoded syntactically: via indexation. Indexation plays an important role in the identification of the correlates that antecede elided heads. In this section, I will incorporate two indexation-related conditions on ellipsis from prior literature, arguing that they should be thought of as expansions on the notion of correlate.
The syntactic condition on sluicing presented in (8) relies crucially on the notion of a correlate of a head. The original definition of correlation proposed in (12) was that the only valid correlate of a head is a lexically identical head. The following data show that that view is unsustainable (cf. comparable data in Merchant 2001:chap. 5).
In (19a), we see a sluice with an interrogative antecedent; the ellipsis site, however, does not receive an interrogative interpretation. Rather, the interpretation of the ellipsis site is something like ‘why those people who said those things said the things that they said’. To put it less convolutedly, the interpretation of the ellipsis site is ‘why the antecedent questions have the answers that they have’. The antecedent is interrogative, but the ellipsis site receives a definite interpretation.
Likewise, in (19b) we see an indefinite in the antecedent, but the ellipsis site does not receive an indefinite interpretation. If there were a corresponding indefinite in the ellipsis site, an interpretation would be possible in which a guy I know could be referred to in the ellipsis site who is not the same guy I know as the one referred to in the antecedent. Such an interpretation is not available. The interpretation of the ellipsis site is something like ‘when that guy I mentioned won a gold medal’. Again, the ellipsis site receives a definite interpretation.
Finally, (19c) shows that it is possible for quantificational antecedents to receive definite interpretations in ellipsis sites as well. The reading of the ellipsis site indicated in (19c) is one in which the interpretation of the ellipsis site is something like ‘who ate at those five restaurants’. We see that even with quantificational antecedents, ellipsis sites can receive definite interpretations.
I have glossed all of these elided definites as pronouns coindexed with elements of the relevant ellipsis antecedent, anticipating my solution. These data suggest that, for the correlate relation to capture the data effectively, it should allow the elision of heads to be licensed relative to correlates that are traces or wh-elements (19a), and full DPs (19b), including quantified DPs (19c).12 All of these data can be captured by the current proposal if correlation is defined in terms of indexation, as follows:13
A node n can be a correlate for a head h iff at least one of the following conditions holds:
n is a head and n and h are tokens of the same lexical item.
n is coindexed with h.
In addition to the ability of elided heads to be correlated with lexically identical antecedent heads, an antecedent DP can serve as a correlate for an elided head if the two are coindexed. The relation of correlation defined above is a relation between a head and a (potentially nonterminal) node. So, for example, in (19b) the deletion of the pronoun he in the ellipsis site is licensed in virtue of the fact that the pronoun’s correlate, the full DP a guy I know, is a DP node at the same position within the antecedent TP as the pronoun is within the elided TP.
In contrast with previous fully or primarily semantic theories of sluice licensing, I hope to have shown that we can define an empirically adequate identity condition in which the role coindexation plays in identifying correlates for deleted heads is the only point at which meaning affects the grammaticality of a sluice. Coreference and binding, as expressed in the syntax by indexation, play an important role in identifying possible antecedent correlates for elided heads. That this is only the syntactic expression of coreference and binding that matters is of a piece with the syntactic nature of the identity condition on sluicing that I have proposed here.14
1.4 Interim Conclusion
In this section, I have proposed a syntactic condition that can account for several facts about sluicing that have been held to necessitate lexicosyntactic identity conditions on the grammaticality of sluicing, while avoiding pitfalls claimed to be problematic for such conditions. As noted at the start of the section, some sort of lexicosyntactic component has often been taken to be required in the identity condition that governs the grammaticality of sluicing, even by analysts arguing for a predominantly semantic approach. Though I do not take myself to have disproven the possibility of a purely semantic approach in principle, I see no reason why the syntactic condition described here could not (or should not) supplant prior semantic identity conditions. The syntactic identity condition presented here captures the data that have been taken to necessitate reference to syntax by prior analysts, and enforces stringent enough lexicosyntactic identity requirements that a high degree of semantic identity between antecedent and ellipsis site emerges as a consequence, rendering it unclear why an independent semantic identity condition would be necessary.
In the next section, I introduce novel data that are problematic for all current theories of sluicing, and I make a general proposal to account for those data. The novel data concern mismatches high in the clausal spine, above the verbal domain; the general proposal is that sluicing requires syntactic identity within the verbal domain but tolerates mismatches above it.
2 Mismatches above the Verbal Domain
2.1 The Empirical Terrain of Mismatches under Sluicing
In the previous section, I presented a syntactic condition on sluicing and showed that it can account for some of the data that have been argued to necessitate syntactic conditions on ellipsis without running aground on other data argued to be problematic for syntactic conditions on ellipsis. In this section, I will present data that are equally problematic for both syntactic and semantic identity conditions, and make a general proposal about how they can be accounted for in a (fairly) theory-independent way. I will then present a concrete implementation of that theory-independent solution couched in the terms of the syntactic account developed above.
In the data that follow, the ellipsis sites are glossed in a manner that anticipates my analysis of these cases. I do not mean to claim that the indicated ellipsis sites represent the only possible way of glossing the interpretation of the elided material, just that the indicated ellipsis sites represent a plausible way of glossing the interpretation of the elided material. One crucial assumption that I have made, guided by the notion that ellipsis is phonological deletion of (part of) a grammatically acceptable syntactic structure, is that the sentences must be grammatical if unelided.15
I argue that it is possible to give glosses of all of the ellipsis sites below in terms of lexicosyntactic mismatches that are restricted to material originating in the portion of the clausal spine above the verbal domain, with vPs that are syntactically identical to their antecedents (in the sense developed in section 1). I have glossed these problematic data in terms of mismatches having to do with tense, finiteness, and modality in order to show that they can be given a unified empirical characterization in terms of the location of the mismatches within the clausal spine.
Many of these examples are adapted from corpus examples aggregated and annotated using funds from NSF grant 1451819: The Implicit Content of Sluicing.
In all cases above, mismatches of material higher than the verbal domain result in elided TPs that are not lexicosyntactically identical to their antecedent TPs. These data are not just problematic for syntactic accounts of sluicing; the elided TPs do not stand in a relation of mutual entailment with their antecedent TPs either. In fact, there is often no entailment in either direction. To see this clearly, note the modal mismatches in (23), in which modality can appear, disappear, or change in both modal flavor and modal force, in every case unarguably leading to an elided TP that means something substantially different than its antecedent.18 The case in (23c), as well as the cases in (23a) and (21b), illustrate a common property of these left-peripheral mismatches under sluicing: their interpretation is somewhat nebulous. In (23c), it is quite clear that the ellipsis site is not interpreted relative to the antecedent modal should; the modals might and could both yield plausible interpretations, as does will; no one interpretation seems clearly correct.19
2.2 Restriction to Eventive Cores
Mismatches of the kinds encountered above are frequently attested in corpora, yet despite the broad range of possible mismatches demonstrated above, no examples have yet been encountered that threaten the claim that argument structure mismatches under sluicing are impossible. The empirical generalization is this: verbs and their arguments cannot be mismatched under sluicing, but what can be mismatched are the finiteness of the clause, the time at which the event described by the verb occurs, the modality that lets us know whether that event is actual, hypothetical, desirable, and so on. The fact that some aspects of the interpretation of the ellipsis site, such as the verb and its arguments, are fixed, but other aspects, like modality, are slippery and indeterminate suggests that the explanation of these left-peripheral mismatches is that identity requirements on sluices apply only to some subset of the elements in the elided TP. Specifically, I claim that identity conditions on ellipsis apply only to elements that originate inside what I call the eventive core of the elided clause—roughly speaking, the verb and its arguments.
The notion that there is a syntactically privileged core of a clause was explored at length by Langacker (1974) in his discussion of the “objective content” of a sentence. Here’s Langacker, following up a discussion of how difficult this notion is to define positively:
It is perhaps easier to say what is EXCLUDED from the objective content of a sentence. The illocutionary force of a sentence is excluded, as are specifications of tense, aspect, and modality. . . . In short, the objective content of a sentence specifies a situation that the sentence is to deal with, and the remainder of the sentence—what I will call the non-objective content—specifies the location of this situation along various dimensions, such as temporal/aspectual, real/unreal, speaker’s affective reaction, and speech act. (Langacker 1974:645–646)
Langacker goes on to express skepticism that “objective content” is a syntacticizable notion. This is understandable given the state of syntactic theory at the time; 40 years of syntactic theory later, however, it has become quite an easy concept to syntacticize. I associate the eventive core with the vP of a clause—the complete verbal complex, including the origin sites of verbs and their internal and external arguments.
Though it might seem odd at first glance, the claim that identity in sluicing is enforced only over the vP of the elided TP is actually quite natural on a view like Langacker’s: the vP is the constituent inside of which the “objective content” of a sentence enters the derivation. Sluicing is grammatical only if the elided clause encodes the same content as its antecedent in terms of the basic event described, the “who did what to who,” so to speak, though the two clauses may mismatch each other in terms of the status of that event relative to the current state of the world (polarity, tense, aspect, finiteness), or the status of that event relative to the speaker’s information or desires (various flavors of modality), both encoded by material that enters the derivation above the vP. It is suggestive that this domain of identity corresponds to a phase; this, in concert with Langacker’s notion of the split between objective and nonobjective content, could be taken as substantiating Chomsky’s (2001) claim that the vP phase corresponds to a semantically meaningful subclausal unit that could be thought of as somehow propositional.20
Eventive core (to be revised)
The eventive core of any TP is its highest vP.
Due to the possibility of recursive clause embedding in natural language, any clause can have arbitrarily many vPs within it. The eventive core of the clause is its highest vP—the vP associated with its matrix verb. There is one further complication with the notion of eventive core that will be discussed in section 2.3.1.
The claim that the eventive core is the domain of identity calculation in sluicing can be implemented in different ways in different frameworks. For example, given a semantic, entailment-based identity condition such as Merchant’s (2001), this could play out by requiring that there be mutual entailment between the eventive cores of the antecedent and the ellipsis site, instead of between the full TPs21 (this would work most smoothly assuming a copy or multidominance theory of movement where moved-out material is still in the eventive core, so the semantic value of the eventive core can be calculated in isolation—the alternative would be to force obligatory reconstruction). For an LF copying account along the lines proposed by Chung, Ladusaw, and McCloskey (1995, 2011), this could play out by copying only the antecedent’s eventive core and decorating it with whatever other material results in a plausible interpretation. I set these other implementations aside here, as fully developing a semantic implementation of these ideas is far outside the scope of this article. In the rest of this section, I will assume the syntactic identity condition developed in section 1 and develop an implementation of the claim that the eventive core is the domain of identity calculation for sluicing couched in the terms of that analysis.
In a nutshell, given the syntactic approach detailed in section 1, the claim that the eventive core is the domain of identity calculation can be implemented by taking the identification of a structure-matching correlate as a requirement on the grammaticality of a sluice only for the deletion of heads that entered the structure within the vP phase.
This account generalizes to all of the examples in section 2.1. Because mismatches in finiteness, tense, modality, polarity, and illocutionary force do not affect the eventive core of a sentence, such mismatches will be acceptable on my account insofar as the eventive cores of the antecedent and the ellipsis site are still identical.22
The syntactic identity condition developed in section 1 can be revised in the following way to restrict its domain of application to the eventive core:
Syntactic condition on sluicing (final)
Given a prospective ellipsis site E and its antecedent A, nonpronunciation of the phonological content associated with any head h ∈ E is licit if at least one of the following conditions holds:
h did not originate within E’s eventive core.
h has a structure-matching correlate i ∈ A.
In other words: when deletion of a constituent is triggered, for each head in that constituent the deletion mechanism checks whether that head is inside the eventive core of the elided TP or is a member of a movement dependency chain whose tail is inside the eventive core of the elided TP. If no, the head is deleted with no further questions asked; if yes, a search for a structure-matching correlate of the head in the ellipsis-licensing antecedent is initiated, and the derivation crashes if none can be found.
A demonstration of the revised condition in (28) is given below:
In (30), lexical heads that originated within the eventive core of the ellipsis site are shown in gray; those are the only heads to which the syntactic condition on sluicing applies. The content of T, which did not originate within the eventive core, can be freely mismatched.
I would like to close this section with brief comments on four complications, three tractable and one not so clearly so.
2.3.1 Modal Verbs
The first complication has to do with the definition of an eventive core in (27). This definition identifies the eventive core of a clause with the highest vP in that clause; but in some languages, elements whose content Langacker (1974) would call “non-objective” occur as verbs. For instance, all modals in German are V heads, not T heads. Even though in German, modality is expressed within the verbal complex, modal mismatches under sluicing still occur. The opposite of this is predicted on the eventive core account of modal mismatches. The crucial cases here are German analogues of the modal mismatches in (23), because these are cases in which there are elided modals with no structure-matching correlate in the antecedent. In German, even though modality is expressed below T, the same sorts of phenomena are possible (Andreas Walker, pers. comm.).
Das war ein Problem, dass die Physik lösen musste, aber für lange Zeit war es
that was a problem that the physics solve must but for long time was it
wie die Physik es lö sen könnte
not clear how the physics it solve could
‘This was a problem that physics had to solve, but for a long time it wasn’t clear how
physics could solve it
Walker reports that (31) is perfectly grammatical with the indicated modal mismatch interpretation. In other words: a German modal verb can be elided even when it has no correlate in the antecedent. As it stands, the present account predicts this to be impossible; it only allows modal mismatches because identity is not enforced for elements whose external Merge site is above the verbal domain.
To fix this, we must characterize the syntactic notion of eventive core in a somewhat more complex way than simply saying that the eventive core of a clause is its matrix vP. The solution to this problem is also the reason why eventive core is a more enlightening term than verbal domain or just vP: instead of taking the term eventive core simply to pick out the matrix vP of a clause, we can define the eventive core of a clause as the highest vP in that clause that is associated with an event-introducing predicate.
Eventive core (final definition)
The eventive core of a clause is its highest vP that is associated with an event-introducing predicate.
Modal verbs only communicate something about the modal status of their prejacent proposition; thus, they are parasitic on the event introduced by the matrix verb of that prejacent. This revision of the definition of an eventive core is quite of a piece with the notion of objective content in Langacker 1974.23
Note also that this revision to the notion of eventive core clarifies an issue with English auxiliary verbs, pointed out to me by Christopher Tancredi (pers. comm.). Consider the following:
This case shows that auxiliary verbs can be mismatched as well. This fact follows from the idea that the eventive core of a clause is the vP associated with the highest event-introducing predicate.
As an auxiliary verb, have does not introduce an event; hence, it sits above the eventive core of the clause, allowing it to be grammatically mismatched.
2.3.2 Putative Case Mismatches
Chung (2013) discusses data like the following:
*In this monastery, it’s possible in principle to sing, but the abbot has to specify which monks
In this monastery, it’s possible in principle to sing, but the abbot has to specify which songs
(modified from Chung 2013:27, (56); 28, (57))
These data are prima facie problematic for the account developed herein. In (34a), all of the elided heads that originated within the eventive core are structure-matched by their correlates in the antecedent. The account developed here therefore incorrectly predicts it to be grammatical. The generalization that Chung derives from facts like these is that the sluice-stranded wh-phrase must be assigned Case by a head that is identical to the head that assigned Case to its correlate; she suggests that a principle that enforces such a generalization may supplement entailment-based theories of sluicing. However, the data in (34) do not give the full picture. Further scrutiny suggests that the actual generalization does not have to do with Case. Observe the following:
The fact that (35) is grammatical casts severe doubt on the notion that parallel Case assignment is required in order for sluicing to be grammatical. In this case, just as in (34a), the wh-phrase is being assigned Case by a head that is not identical to the head that assigns Case to its correlate; however, the sluice is well-formed.24 In reaction to this, one might conclude, following Thoms (2015), that the sluiced wh-remnant must be correlated with a certain kind of DP in the antecedent and that that requirement fails in the case of a null subject, as in (34a). However, a more careful consideration of the empirical facts indicates otherwise. Observe:
In this case, the antecedent (to sing) and ellipsis site (can sing) are the same as in the ungrammatical example (34a). However, bringing monks to salience prior to the sluice renders the sentence significantly less degraded. This suggests that the ill-formedness of (34a) stems neither from Case mismatch, as Chung (2013) claims, nor from the inability of null subjects to serve as correlates for sluiced wh-remnants, as Thoms (2015) claims; instead, it is a result of the pragmatics of the recoverability of the interpretation of a sluice: (34a) is ill-formed because the context does not adequately prepare the ground for the question Which monks?25
2.3.3 Split Antecedents
Sluicing, just like VP-ellipsis, allows split antecedence, as the following examples demonstrate:
Sally punched Jeff or Steven. I can’t remember which (one).
Sally went to the park, or she went to the movies. I can’t remember which (one).
Sue punched Jeff, or Jeff punched Sue. I can’t remember which (one).
Cases like these are discussed in depth by AnderBois (2014) and Barros (2014). They may or may not pose a problem for the theory presented in this article, depending on what one’s analysis of the content of the ellipsis site is in each case. I argue that (37a) and (37b) are easy to account for using the analysis presented here. (37c) may be trickier.
Given the hypotheses presented in (38) about the content of the ellipsis sites of (37a) and (37b), the theory presented here predicts both cases to be grammatical. In (38a), the content of the ellipsis site is structure-matched by the antecedent; nothing more needs to be said. (38b) is more complicated. In this case, we have what seem to be two distinct antecedents. However, crucially, there is only one head in the ellipsis site that originated within its eventive core: she. This head is structure-matched in both antecedents. We might, then, account for split antecedence in the following way: in the case that an ellipsis site has multiple antecedents, the relevant heads in the ellipsis sites must find structure-matching correlates in both antecedents.
I believe that the explanation for the trickier case, (37c), lies in the nature of anaphoric one. The case in (37a) involves coordinated DPs, and the one in the sluiced wh-remnant is anaphoric to those DPs; the case in (37b) involves coordinated VPs, and the one in the sluiced wh-remnant is anaphoric to those VPs; and the case in (37c) involves coordinated TPs, and the one in the sluiced wh-remnant is anaphoric to those TPs. It might be the case that when one is a TP anaphor, it lexicalizes the entire TP span, in which case the content of the sluiced CP prior to wh-movement is simply which one. If that were the case, the content of the ellipsis site in (37c) would simply be the trace of the moved wh-element, which would not need a structure-matching correlate by virtue of being the tail of a movement dependency chain; there would be no content needing a structure-matching correlate in the antecedent. The development of such a theory of anaphoric one is far outside the scope of this article. It may instead be that the true explanation for sluices like (37c) is that they are not true sluices; for instance, Barros (2014) presents what amounts to a pseudosluicing analysis of such sentences.
2.3.4 Exceptive Antecedents
Sluicing is possible in exceptive constructions like the following:
This sentence is quite problematic on every account of sluicing of which I am aware. It is problematic for mutual entailment accounts, because there is only a one-way entailment, from the antecedent to the ellipsis site; it is problematic for LF copying accounts, because the ellipsis site does not contain a copy of the antecedent (crucially, the interpretation of the sentence is not ‘I don’t know why nobody other than Jerry liked it’); it is problematic for my syntactic account, because the subject of the antecedent clause is nobody, not Jerry.
This observation goes back to Merchant (2001:22, (32i)), who presents the following sentence:
Merchant sources this sentence to the 1993 film Thunderheart; he presumably recreated it from memory. I revisited the scene in the film, transcribed the dialogue in question, and checked my transcription against the official subtitle track. The precise exchange in the film is actually somewhat more interesting than Merchant’s recollection.
A: They don’t want you here. Ain’t nobody gonna talk to you.
A: Except for the wica’sa wakan, and I don’t know why.
B: The who?
A: The man who sent me to find you. Says he has some information for the FBI.
There is a substantial pause between each conversational turn; B’s lines do not interrupt A’s. A’s middle turn is presented prosodically as a complete and self-contained sentence. A’s final turn, for our purposes, serves to disambiguate the interpretation of the ellipsis site in the sluice from A’s middle turn—it makes it clear that the intended interpretation of the ellipsis site is indeed something like
This dialogue clearly demonstrates that the solution to the problem of exceptive sluices is parasitic on the analysis of exceptive constructions. One trend in the literature on exceptives is to treat them as quantifier modifiers: except Jerry modifies nobody to produce quantification over the set of non-Jerry humans (see, e.g., Reinhart 1991, von Fintel 1993, Gajewski 2008). The fact that an exceptive can be used in a separate sentence from the quantifier that it modifies causes difficulties for a compositional semantic account of exceptives as quantifier modifiers. This phenomenon is robust even in cross-speaker cases:
A: Nobody liked that movie.
B: Yeah, except Jerry.
The exact nature of the problem that the possibility of exceptive sluices like (39) poses for the theory of sluicing will not be clear until we have a clear syntactic theory of exceptive phrases. It may be that the possibility of sluices like (39) can serve as a useful diagnostic in the development of such a theory.
2.4 Interim Conclusion
In this section, I have proposed a theory-independent way to account for the possibility of a wide variety of left-peripheral mismatches under sluicing: the domain of identity for TP deletion isn’t the full TP, it’s the eventive core of the elided TP. This conception is of a piece with Langacker’s notion that the subset of sentential content that refers to the concrete particulars of an event has a privileged status relative to content expressing the relation of that event to the current state of the world, and to the speaker’s beliefs and desires. Syntactic identity of that privileged “objective” content is required in order for sluicing to be grammatical.
3 Beyond Syntax: Sound and Meaning in Sluicing
In this section, I offer brief thoughts on ways that sound and meaning fit into the syntactic identity condition presented here. I argue first that the role meaning plays is predominantly pragmatic, serving to restrict the set of grammatically possible interpretations of sluices to those that are actually available in a particular context. Second, I argue, following work by Bennett, Elfner, and McCloskey (to appear), that the impossibility of mismatches involving verum focus could be profitably analyzed in terms of phonological grammar, rather than syntax or semantics.
3.1 The Role of Meaning: The Pragmatics of Interpretation Recoverability
One possible objection to the proposal developed in section 2—namely, that mismatches under sluicing that occur outside of the eventive core of the elided TP are grammatical—goes something like this: Why, says the skeptic, given this account, can’t I get left-peripheral mismatches willy-nilly? Why, says the skeptic, is the following interpretation not available?
The sluice in (43) is perfectly grammatical with the indicated interpretation of the ellipsis site if mismatches above the eventive core are freely possible. Doesn’t the account presented here predict that such interpretations will be available? I believe that this objection confuses conditions on the well-formedness of sluices, which is a grammatical notion, with conditions on the availability of interpretations of sluices, which is a pragmatic notion. To put it another way: undergeneration and overgeneration are very different sorts of problems.
When reasoning about what structural relationships between antecedents and ellipsis sites are grammatically possible, the most powerful evidence is positive evidence: if a sluicing construction can have a particular interpretation, it must be the case that it is grammatical to perform ellipsis on a structure with that interpretation. However, though the set of interpretations that interpreters will assign to sluicing constructions is presumably limited to the set of grammatically possible sluices, there is no reason to suppose that no ellipsis-independent pragmatic factors restrict the set of interpretations interpreters are willing to propose for specific sluices. To put it concisely: if an interpretation of a sluice is available, a sluice with that interpretation must be generated by the grammar of sluicing; if an interpretation of a sluice is unacceptable, it does not necessarily follow that it cannot be generated by the grammar of sluicing, as there is a huge variety of potential reasons why an interpretation might not be available in a particular context. The status of negative evidence in relation to grammatical theory is less clear than the status of positive evidence.
A prediction made by this basic line of reasoning, which goes back to the earliest work on generative grammar—namely, that the grammar produces a proper superset of what is actually acceptable in context—is that we should expect the set of interpretations an interpreter is willing to assign to a sluice in context to be a proper subset of the possible interpretations of grammatical ellipsis structures. The data bear out this prediction: a huge variety of mismatches above the eventive core of the elided TP are grammatical, but not every interpretation traceable to such a mismatch is available for every sluice in every context.
What I have done in this article is present an account of the role that syntactic identity between antecedent and ellipsis site plays in determining the well-formedness of sluices. The syntactic identity condition that I have developed predicts that some mismatches will be impossible and that others will be possible: mismatches inside the eventive core will be impossible, mismatches above it will be possible. This gives us, for any sluice, a set of grammatically available interpretations: all interpretations involving mismatches of material above the vP of the elided TP are made available by the grammar of sluicing. This gives listeners a meaning recovery problem: which interpretation did the speaker intend to communicate?
I argue that this problem is a problem for pragmatics, not a problem for grammatical theories. Given an ellipsis site that could potentially mean a large variety of different things, interpreters use commonsense pragmatics to weed out many interpretations of the sluice that are highly implausible, despite being technically grammatical in terms of grammatically possible structural relations between antecedents and ellipsis sites. Listeners know that when speakers choose to elide content, they are assuming that their interlocutors will be able to recover the meaning of the ellipsis site on the basis of what they have said; this follows quite simply from completely general pragmatic notions, like the cooperative principle (Grice 1975). If in a given elliptical construction an ellipsis site that perfectly matches the antecedent yields a plausible interpretation, it would seem quite uncooperative for a speaker to utter that construction with the intention of its being interpreted as an imperfect match, in the absence of very strong contextual factors militating against a perfectly matching interpretation. I express this principle as follows:26
Pragmatic principle of ellipsis interpretation27
If a perfectly antecedent-matching ellipsis site yields an interpretation that is plausible in context, that interpretation should be strongly preferred to interpretations generated via imperfectly antecedent-matching ellipsis sites.
This pragmatic principle is of course not a full theory of the pragmatics of ellipsis interpretation, though it goes a long way toward explaining the unavailability of mismatches in most normal contexts; I consider the development of a formal pragmatics of ellipsis interpretation to be far outside the scope of this article. I refer the interested reader to Kroll and Rudin 2017 and Kroll to appear. In the account developed in those papers, a sluice is only felicitous if the ellipsis site is a uniquely salient entailment of the local context. That account could be thought of as an account of the recoverability of the elided content of sluices: an interpretation will only be available for a sluice if it is a uniquely salient contextual entailment at the time of utterance. Christopher Tancredi (pers. comm.) points out that, as we saw in (23), often various interpretations are available for modal (and other) mismatches under sluicing, making “unique salience” seem perhaps to be too strong a condition on the recoverability of sluices. Nonetheless, that for an interpretation to be available for a sluice, the elided content must be a salient contextual entailment, whether unique or not, is a promising and prima facie sensible idea, with a clear connection to the long history of work on the role givenness plays in ellipsis.
The idea that semanticopragmatic factors that are highly sensitive to context have a role to play in the acceptability of elliptical interpretation is an old one, going back at least to the work of Rooth (1992), Tancredi (1992), and Heim (1997), all of whom explore the notion that ellipsis may be possible only under strictly stronger conditions than those under which deaccenting is possible. Such a system works along quite similar lines to the system outlined here: the acceptability of an ellipsis is determined by structural restrictions on relations between antecedent and ellipsis site, supplemented by the more general meaning-related considerations that govern the possibility of deaccenting. The acceptability of deaccenting is generally taken to be determined by whether or not the deaccented material is given—a notion that has been treated as a strict semantic one (see, e.g., Schwarzschild 1999, Büring 2016), but that is often treated as a somewhat looser pragmatic or information-structural notion as well (for a recent overview, see Vallduví 2016). It might well be that givenness, in some form or another, is indeed a relevant factor in determining the availability of a particular interpretation of a sluice, though it is hard to see how a strict semantic conception like Schwarzschild’s GIVENness could allow the kind of mismatches observed for sluicing.
What is crucial for my purposes here is this: the fact that left-peripheral mismatches are possible when strict matching between antecedent and ellipsis site results in an implausible interpretation demonstrates that such mismatches are grammatically possible; that such mismatches do not seem to be available interpretations for ellipsis constructions in which strict matching between antecedent and ellipsis site results in a perfectly plausible interpretation is a fact about the pragmatics of the availability of elliptical interpretations, not a fact about the grammar of ellipsis identity. The distinction between these two different approaches to studying ellipsis may also explain the very stark differences between deletion-based approaches and inference-based approaches to ellipsis. Purely grammar-internal deletion-based approaches may overreach by trying to explain the unavailability of interpretations that are properly ruled out by the pragmatics of ellipsis recovery in context, whereas purely inferential nongrammatical approaches to ellipsis may overreach by dismissing the syntactic grammar of ellipsis due to an over-narrow focus on the pragmatics of recoverability.
I have proposed here an account of the syntactic component of sluicing. My primary concern has been to develop a syntactic conception of the identity condition on sluicing that is capable of generating all grammatical cases, while ruling out cases that are strictly impossible, as mismatches in the verbal domain seem to be. That mismatches above the verbal domain are sometimes acceptable and sometimes not, I argue, is not a problem to be solved within the grammar of sluicing; rather, understanding exactly what pragmatic factors influence the availability of different interpretations in different contexts should be taken to be a research program of acute interest to elliptologists.
3.2 Phonological Grammar and Verum Focus
A class of unacceptable mismatches has been pointed out to me by Christopher Tancredi (pers. comm.). Consider the following cases:
These sluices are unacceptable, but are not clearly ruled out on the basis of the pragmatic reasoning above. The indicated interpretations of the ellipsis sites are allowed by the syntactic identity condition I have proposed, there are no more faithful matches that are plausible in context, and the indicated interpretations are perfectly plausible in these contexts—in fact, the context for the sluice in (45) entails that somebody ate, and the context for the sluice in (46) entails that somebody left.
(47) There were five people at the party, and the food disappeared rather quickly. However, John, Mary, and Sue ate nothing. That makes me wonder who DID (eat).
(48) There are fewer people here than there were just a minute ago, so it must be the case that someone left. John didn’t leave, so who DID?
Each of the (unacceptable) ellipsis sites in (45) and (46) contrasts with its antecedent in terms of polarity, making them excellent candidates for hosting verum focus, as (47) and (48) demonstrate. Both examples are degraded without verum focus. I argue that this explains the unacceptability of the sluices in (45) and (46). In an extremely interesting paper, Bennett, Elfner, and McCloskey (to appear) discuss interactions between verum focus and Irish responsive ellipsis. The conclusion that the data they discuss lead them to is that verum focus is unelidable for phonological, not syntactic or semanticopragmatic, reasons. If this is indeed the case, the explanation for Tancredi’s observations in (45) and (46) might well lie neither in the syntax of ellipsis identity nor in the pragmatics of the acceptability of elliptical interpretations, but rather in the domain of phonological grammar—an understudied aspect of the grammar of ellipsis.28
In this article, I have put forward two proposals.
The first proposal is a novel way of assessing syntactic identity that I believe can supplant the role semantic identity conditions have commonly been taken to play in the well-formedness of sluicing. The crucial innovation of this proposal is to take the relevant grain size of identity to be individual deleted heads, not entire deleted constituents. Enforcing syntactic identity deleted head by deleted head is rigid enough to rule out ungrammatical syntactic mismatches in voice and argument structure and to capture the generalization that ellipsis sites must contain no lexical material not present in their antecedents, but is flexible enough to evade criticisms that were fatal to conditions enforcing syntactic identity at the level of the entire deleted constituent.
The second proposal has to do with a novel empirical generalization: that an extremely broad variety of mismatches above the verbal domain are possible under sluicing, in stark contrast to the total impossibility of syntactic mismatches within the verbal domain. These data can be captured with the proposal that the domain of identity for sluicing is the eventive core, or highest event-introducing vP phase, of the elided TP: in order for sluicing to be grammatical, deleted heads originating within this vP phase must be anteceded by structure-matching correlates. In other words, the identity condition that governs the well-formedness of sluices is evaluated over a proper subpart of the deleted constituent. Such a restriction makes good on early attempts to explain within syntactic theory the privileged status of certain elements of the clause (Langacker 1974).
The worldview that results from adopting these two proposals can be restated in the following way: the identity condition that governs sluicing is fundamentally syntactic, but that does not mean that considerations of meaning have no role to play in the acceptability of sluices. Mismatches are strictly impossible within the eventive core, but available in principle above it; the availability of a particular interpretation in a particular context is conditioned by a wide variety of subtle pragmatic factors. This is both a prediction of the theory at hand and a robust fact about the empirical landscape. Better understanding the factors that make an interpretation available or unavailable in a given context, then, becomes an exciting avenue for future research; what is completely clear is that an account of the role of syntax in the grammaticality of sluicing should not rule such mismatches out.
This article has been concerned exclusively with sluicing, but I believe that the proposals presented here could be profitably applied to other well-studied forms of ellipsis, such as VP-ellipsis (VPE). I conclude with a brief remark on how such an extension might proceed.
Of particular interest is the idea that the domain of identity for ellipsis can be strictly smaller than the ellipsis site. One immediate connection this suggests is to the phenomenon of voice mismatches under VPE (Merchant 2013b). One way to account for the possibility of voice mismatches is to take VPE to delete only the VP, not the entire verbal domain, allowing the Voice head to avoid identity requirements by virtue of not being inside the ellipsis site. This is the analysis Merchant pursues. Another strategy, however, would be to take VPE to delete the entire verbal domain, but to take the domain of identity in VPE, just as in sluicing, to be strictly smaller than the ellipsis site. It is not clear at first glance which of these approaches would be preferable; the latter would require us to give up the justification of the privileging of the eventive core in terms of its association with objective content.
The idea that the eventive core is privileged above higher material in ellipsis identity could also be brought to bear on a variety of problems brought to light in recent literature on verb-stranding VPE. For instance, McCloskey (2017) discusses the possibility of mismatches in tense morphology under responsive ellipsis in Irish, which follows straightforwardly from an extension of my second proposal to VPE.
I leave a fuller investigation of the appropriateness of applying the ideas presented here to VPE to future work.
2 Recall that there are other restrictions on which TPs can be deleted—for instance, the restriction to TPs that are the complements of [+wh, +Q] Cs.
3 Christopher Tancredi (pers. comm.) points out that, despite pied-piping being a purely syntactic phenomenon, the problem for a mutual entailment approach rests on the question of whether the preposition stranded in the ellipsis site in (4b) is semantically contentful. If to is semantically contentful, then the ellipsis site will not necessarily stand in a relation of mutual entailment with its antecedent. I note only that for this argument to defang Chung’s (2005) argument, it would need to be extended to all strandable prepositions, as Chung observes this effect with a wide variety of prepositions (see her examples (18) and (19)).
4 The argument presented here is not Merchant’s only argument against syntactic identity conditions; in section 2, I discuss more of his arguments.
5 The decision to treat ellipsis identity head by head has its roots in Wasow’s (1972) theory of ellipsis, famously adopted by Williams (1977), in which ellipsis sites are fully articulated syntactic structures populated entirely by null heads. However, in Wasow’s theory, ellipsis sites are born empty and their antecedents are copied into them; Merchant (2001) argues forcefully against such theories on the basis of Case-matching facts.
6 That ellipsis identity ignores lower copies/traces is not necessarily as stipulative as it might seem. This follows straightforwardly, for example, from a suitable implementation of ellipsis as phonological deletion. If ellipsis is phonological deletion, then it is perfectly natural that it would ignore lower copies/traces, as they are associated with no pronounced material—there’s nothing there for the phonology to delete. See Bennett, Elfner, and McCloskey to appear and Rudin and Kalivoda in preparation for formal phonological implementations of ellipsis.
7 Note here that the TP that is identified as the antecedent that determines the domain of structure matching may well be an embedded TP within a larger structure. Thanks to Christopher Tancredi for discussion of this point.
8 Note that this explanation assumes that movement does not involve the creation of additional syntactic structure within the domain of identity.
9 For an argument for the breadth of the role v plays in regulating word order within VPs, see González-Vilbazo and López 2012.
10 Christopher Tancredi (pers. comm.) points out that a similar interpretation could be derived with an indefinite in the ellipsis site. If he is replaced with a guy I know in the ellipsis site, one available interpretation of the sluice would be ‘For no guy I know do I know when he won a gold medal’. This would entail the relevant interpretation of the sluice. However, the sluice in (19b) is felicitous in a context in which the speaker knows when everybody won a gold medal except the specific guy she’s talking about. Observe that the following is not a contradiction:
I think a guy I know won a gold medal, but I don’t know when. This other guy I know won a gold medal for sure, at the 2016 Olympics.
The acceptability of this example shows that the availability of the reading indicated in (19b) cannot be reduced to the interaction of an indefinite in the ellipsis site with negation outside of it.
I do not want to dismiss out of hand the idea that there might be a way of deriving these kinds of interpretations while maintaining lexical identity between the antecedent and the ellipsis site, allowing us to maintain a simpler notion of correlate, but it is difficult to see how to do so.
11 This is to be contrasted with another potential interpretation of this sluice.
Someone ate at five burger restaurants, but I don’t know who
ate at five burger restaurants
Data like this were originally discussed by Merchant (2001:chap. 5). That both readings are indeed available has been disputed in subsequent literature; Chung, Ladusaw, and McCloskey (2011) argue that the reading in (19c) is unavailable, contra Merchant; Elliott and Sudo (2016) argue that both readings are indeed possible.
John didn’t see anyone, but Mary did
On the account Merchant develops to explain this, NPIs like anyone and their non-NPI correspondents, like someone, are lexically identical: the NPI is the reflection of agreement with polarity. To the extent that NPI mismatches are possible under sluicing, that account could be adopted here unproblematically.
13 The notion that indexation is relevant to ellipsis identity goes back to the very earliest work on the subject and has been a standard assumption ever since; see in particular Sag 1976, Rooth 1992, Fiengo and May 1994, and Heim 1997 (cf. critical discussion in Merchant 2001). (20b) is a straightforward restatement of the notion of vehicle change in Fiengo and May 1994.
14 Further modifications may be necessary to make the definition in (20) empirically adequate. Consider the following grammatical sluice:
Many prominent representatives still have not endorsed the candidate. In a moment, two of them will explain why
they still have not endorsed the candidate
The relevant interpretation of (i) is the one in which the elided they is interpreted as referring back to the two representatives doing the explaining, not the DP many prominent representatives in the antecedent. In this case, we have a grammatical sluice in which an elided pronoun meets neither of the conditions on correlate identification in (20): material in the relevant portion of the antecedent is neither lexically identical nor referentially identical (coindexed).
Though I consider a fuller discussion of these data to be outside the scope of this article, I will note that Merchant (2001) proposes an analysis of data like this in terms of the rebinding of the elided pronoun by the “local, c-commanding quantifier” (Merchant 2001:214). Merchant notes further interesting empirical phenomena associated with rebinding readings, such as the requirement that the set the rebinding quantifier is associated with be a subset of the set quantified over by the antecedent quantifier (see Merchant 2001:215, (158)–(161)), which are desiderata for the explanatory success of a full theory of rebinding readings.
17 An anonymous reviewer expresses skepticism about the acceptability of this example. Such examples are very widely attested, however, and the relevant examples have been scrutinized under the aegis of the Santa Cruz Ellipsis Project by at least six undergraduate annotators each, as well as by various graduate students and faculty members working on the project, to say nothing of the fact that they reliably make it past newspaper copyeditors at the New York Times. I provide a corpus example with a more elaborated context here, to show how natural such cases sound in actual usage:
Context: On Dec. 10, [Senator] McCain sent a letter to the FCC urging the five-member board to end two years of deliberations and decide whether Paxson Communications should be given a license for a Pittsburgh station. Angela J. Campbell, an attorney for opponents to the deal, told the Globe that McCain’s letter likely “tipped” the scales in favor of the decision.
Sluice: “Senator McCain said, ‘Do it by December 15 or explain why,’ and the commission jumped to it and did it that very day,” Campbell told the Globe.
(Corpus example 22987, Santa Cruz Ellipsis Project)
See Kroll to appear for extensive discussion of polarity mismatching sluices.
19 Christopher Tancredi (pers. comm.) notes that the modal mismatch cases are not necessarily problematic for semantic accounts of the identity condition if one does not make the assumptions that I have made about the syntactic structure of the ellipsis site, and if one instead compares LF representations. For instance, he notes that (23a) could be analyzed as follows:
Sally knows that at all times t, it is possible that awful things happen att, but Sally doesn’t know for which t
awful things happen at t
This is a considerably more abstract level of representation than the one that must be assumed to be relevant given the assumption that lexicosyntactic identity plays an important role in the identity condition on sluicing, but it is not an unreasonable level at which to analyze semantic identity conditions. However, whether such an approach might extend to all of the modal mismatches discussed here remains to be seen, and it remains unclear at best how the approach might extend to the polarity mismatches.
22 Crucially, the account of illocutionary mismatches only goes through on the assumption that there is a syntactically present but unpronounced subject in the imperative that can serve as a correlate for the declarative subject in the ellipsis site. This is a standard assumption in the literature on both the syntax and the semantics of imperatives (Potsdam 1998, Zanuttini 2008, Kaufmann 2012, though cf. Portner 2004).
23 It is worth noting that the distinction between event-introducing and non-event-introducing verbs is prima facie a semantic distinction and thus not obviously of a piece with the predominantly syntactic approach to ellipsis identity pursued here. However, if one takes the semantic type of events to be syntactically encoded, this distinction could be expressed in purely syntactic terms. Thanks to Christopher Tancredi for discussion of this point.
24 An anonymous reviewer points out that (35) might instead be analyzed as having the following structure:
In this monastery, it’s possible in principle for monks to sing, but the abbot has to specify which monks
it’s possible for t to sing
In this structure, which monks has been extracted from subject position. This would normally be a Comp-trace violation, but if one assumes, following Merchant (2001), that the impossibility of at least some extractions is due to PF considerations and that therefore some extractions are possible under ellipsis that would not be possible unelided, this is a plausible structure that poses no problem for Chung’s (2013) account. No such analysis, however, is possible for (36), which is equally problematic for Chung’s account.
25Chung (2013) also notes that possessor extraction is impossible under Chamorro sluicing with no possessor in the antecedent.
*Malingu ga’lågu, lao ti hu tungu’ håyi.
AGR.disappear dog but not AGR know who
‘A dog disappeared, but I don’t know whose.’
This is the case even though possessor extraction is grammatical in Chamorro.
Håyi un fåhan karetå-ña?
who AGR buy car-AGR
‘Whose car did you buy?’ (lit., Who did you buy a car of ?)
Chung explains the ungrammaticality of possessor remnants in sluiced clauses in the absence of a correlate possessor in the antecedent in terms of her requirement that Case assignment at the extraction site of the sluice remnant be identical to Case assignment at the corresponding position in the antecedent. The account developed here can just as easily explain these facts if there is a structural distinction between possessive DPs and nonpossessive DPs. Evidence that such differences indeed exist in Chamorro comes from the optionality of order of arguments in Chamorro possessive DPs. In standard Chamorro possessive DPs, the possessor is to the right of the head noun; however, under certain circumstances the ordering can be reversed, and the possessor comes first.
yanggin mampus atdit chetnot-ña i patgun
if extremely AGR.severe illness-AGR the child
‘when a child has a very serious illness’
yanggin mampus atdit i patgun chetnot-ña
if extremely AGR.severe the child illness-AGR
‘when a child has a very serious illness’
Such ordering reversals between elements of nonpossessive DPs are impossible. These ordering reversals have the peculiar property of only being possible in possessive DPs headed by null determiners, which I have no explanation for. However, the very possibility of variable ordering within possessive DPs suggests that possessive Chamorro DPs have an additional layer of structure to which the DP can move, resulting in the anomalous ordering.
26 There may be finer-grained distinctions at work than the principle in (44) suggests. Take, for example, the following asymmetry (Pranav Anand, pers. comm.):
I need to get some gas. Can you tell me how
I can get some gas
?I need to get some gas. Can you tell me where
I can get some gas
The judgments here are subtle, but if this asymmetry is genuine, then more needs to be said than what I have said in this section. It is not entirely clear whether or not the account presented here is satisfactory; this empirical terrain is complex and largely unexplored. It may be that the best account of the distribution of modal mismatches is a syntactic account that assumes covert modality, such as Bhatt’s (2000); it may be a semantic account involving modal subordination, such as Brasoveanu’s (2010). These facts may also find a purely pragmatic explanation; Kroll’s (to appear) work on the felicity of sluices is a promising foray in that direction. I leave a fuller exploration of these facts to future work.
27 A comparable principle was proposed previously by Merchant (2010), in the context of a discussion of syntactic, semantic, and pragmatic varieties of ellipsis. To quote his own summary of his generalization:
[T]he basic intuition is that when there is a parallel syntactic antecedent available, it must be used (leading to the case and voice effects discussed). When a script is available, its modes must be used. When none is available, then and only then can other mechanisms (for case assignment, etc.) be used, and then and only then is the semantic ellipsis device triggered. (Merchant 2010:48)
In this context, script refers to a pragmatically available though unarticulated prompt of the kind that allows us to understand an utterance of “Uptown, please” upon entering a taxi to be implicitly answering the question, “Where would you like to go?”
28 Of course, there are potential nonphonological explanations available here as well. See, for instance, the discussion of circumstances under which VP-ellipsis is preferred to sluicing (and vice versa) in Takahashi and Fox 2005.
Thanks are due above all to Jim McCloskey, for never wavering in his belief that I had something to say that was worth saying, even in the most nascent stages of this project. Thanks are due as well to Sandy Chung and Pranav Anand for their enthusiastic intellectual stewardship of this project; to Christopher Tancredi for going far beyond the call of duty in his comments on various drafts of this article; to Erik Zyman, Jason Merchant, Steven Foley, an anonymous reviewer for LI, my compatriots in the Santa Cruz Sluice Crew (Margaret Kroll, Kelsey Kraus, and Karen Duek), all participants in the 2015–2016 UCSC graduate research seminar cycle, and audiences at LASC 2016, NELS 47, and LSA 91 for helpful discussion; to all annotators funded by National Science Foundation grant 1451819, The Implicit Content of Sluicing (which supported my own work on this project), in particular lead annotators Bern Samko, Chelsea Miller, and Rachelle Boyson, for deluging me with data and judgments; and finally to Samuel Jay Keyser, Sheelah Elizabeth Ward, and Anne Mark for efficiently shepherding this article along its path to publication. I take full responsibility for anything obviously wrong contained herein, and it should go without saying that the NSF having funded my work does not entail that it or any other federal agency endorses the ideas that I’ve used government money to develop.