2.1 Why is this a puzzle?

Although the Stranding Generalization seems at first glance unsurprising, it constitutes a genuine puzzle for many theories of ellipsis. The generalization seems to be that non-contrastive functional items require a prosodic host. But in other contexts, they are perfectly able to appear without such a host. Consider object pronouns again. The pronoun in (11a) has a verb to its left, and therefore cliticizes.⁹ But the pronouns in (11b–11c) have no material to their left. When the syntax delivers a structure with no possible prosodic host, an unreduced stress-bearing form is used. (11d–11e) shows the same pattern with possessor pronouns: when there is nothing to its right, the possessor pronoun appears in unreduced (in this case suppletive) form.

So non-contrastive functional items have two possible realizations: a weak/reduced form when there is an available host; and a strong/unreduced form when there is no available host. The puzzle of the Stranding Generalization, then, is that the strong/unreduced form cannot be used, even when ellipsis has deprived functional material of a potential host.

This is a problem for LF-copying theories of ellipsis. According to these theories, ellipsis involves insertion of a null pronoun, which retrieves its meaning anaphorically from the discourse. Consider (8b), repeated here as (12). According to LF-copying theories, the ellipsis site consists of a null pronoun, anaphorically related to the noun in the antecedent. Since there is no appropriate prosodic host for the pronoun, the prediction is that the unreduced form is used.

The same problem arises, though, for any PF-copying theory in which ellipsis precedes prosody. According to PF-copying theories, ellipsis sites contain syntactic structure, including potential prosodic hosts for functional material. But ellipis is widely assumed to take place before the prosodic computation takes place. If this is the case, then the input to the prosodic component will not contain the potential prosodic host in cases like (13). The prediction again is that the unreduced form will be used.

Thus, under many current theories, ellipsis is expected to derive a structure in which any potential prosodic host of functional material is absent, whether through deletion/non-pronunciation or insertion of a null element. Functional items have an unreduced form that is used when there is no appropriate host to cliticize to, whether as a result of movement or base-generation. These theories therefore fail to predict the Stranding Generalization: an unreduced form is predicted to be available, in the absence of material to cliticize to.

2.2 Contrast and Newness

In this section we argue that information structural constraints on ellipsis cannot, by themselves, explain the Stranding Generalization. As we will see, a requirement that all ellipsis remnants contrast is too strong, ruling out stranded pronouns but also a number of possible non-contrastive elements. On the other hand, a requirement that ellipsis remnants be merely new/non-given is too weak; ruling in non-contrastive elements in certain contexts will also allow non-contrastive pronouns, violating the Stranding Generalization.

We have seen that one way for a functional item to avoid the Stranding Generalization is to be contrastively focused. It is therefore tempting to attribute the effects of the Stranding Generalization to a generalization like that in (14):

For example, a number of linguists have argued that gapping remnants must contrast with parallel antecedents (Kuno (1976); Féry & Hartmann (2005); Gengel (2007); Winkler (2011), among others). It might then be that the pronoun cannot appear in (15a) because it doesn’t satisfy the Contrast Generalization. This would be supported by the impossibility of a non-contrastive lexical noun phrase appearing in the same position (15b). The effects of the Stranding Generalization would then be attributed not to the prosodic needs of function words, but an independent requirement on their information structure. We argue that this move will not work.

Firstly, as indicated by the diacritic, not all speakers find a non-contrastive lexical noun phrase unacceptable out of the blue in (15b), in contrast to the robust unacceptability of (15a) across speakers. More importantly, the speakers we consulted uniformly accept it in contexts like the following:

Crucially, although non-contrastive lexical noun phrases can appear in these contexts, pronouns cannot (17).

Thus, (17) shows that a pronoun cannot be stranded by ellipsis of the verb, even in a context where a non-contrastive noun phrase can appear. The Stranding Generalization therefore does not fall out of the Contrast Generalization: pronouns are, in some sense, special. Note that it will not help to treat Sheryl in (16a–16b) as contrastive. This would incorrectly predict that pronouns can appear in this position, since focused pronouns can be stranded by ellipsis.

The same divergence between pronouns and non-pronouns is found in NP-ellipsis cases.

The way the Stranding Generalization is formulated predicts that when pronouns do have an appropriate host, they can appear as gapping remnants, even when non-contrastive. This is correct: when a preposition is available for the pronoun to cliticize to, it can appear as a gapping remnant (19a–19b). Note the minimal pair of (19a) and (19c): although the pronouns are both non-contrastive, the pronoun in the double object construction violates the Stranding Generalization, while the pronoun in the to-dative construction does not.

Similarly, possessor pronouns can serve as ellipsis remnants when provided with an appropriate host, even when that host—like the numeral in (20)—does not contrast.

The depictive in (21a) and the adverb in (21b) are further examples of non-contrastive remnants. They do not violate the Stranding Generalization because they do not require a prosodic host.

One potential response is to suggest that the above non-contrastive remnants are actually part of a larger constituent that contains contrastive focus. For example, it might be argued that in (21a), the object and depictive form a constituent, [the fish raw]. Because this constituent contains the contrastively focused object the fish, the Contrast Generalization is satisfied. But even if such a constituent structure could be motivated in all the above examples, it could not explain the Stranding Generalization. If the Contrast Generalization could be satisfied by these putative constituents, then putting the contrast on the other element in these constituents should allow a pronoun to survive ellipsis. This is not the case: in (22a), contrasting the depictive does not allow an object pronoun to appear. The same logic carries over to the ditransitive example in (22b).

The same pattern holds in other varieties of ellipsis. It is possible to construct counterexamples to the Contrast Generalization in pseudogapping (23a) and fragment answers (23b). Mirroring the gapping pattern precisely, a given pronoun cannot appear where a given lexical noun phrase can. (23a–23b) are therefore further instances of the Stranding Generalization.

Again, in accordance with the Stranding Generalization, non-contrastive material can appear when provided with a prosodic host.

We conclude that the Contrast Generalization does not hold. It is too strong: it correctly rules out the unacceptable remnants covered by the Stranding Generalization, but also rules out a number of acceptable remnants.

Two reviewers suggest that, although the Contrast Generalization cannot be sustained, a weaker requirement might do the job. They argue that the non-contrastive remnants in the above examples are not given, but constitute new information. The Contrast Generalization, then, could be replaced with (25):

We think this is descriptively correct, and will adopt it in what follows. Note that the appearance of non-contrastive elements is improved by the richer discourses we provided above. For example, while only some speakers accepted the non-contrastive name as an ellipsis remnant in (15b), repeated as (26a), all speakers accepted it in a context like (16a), repeated as (26b).

Building on reviewers’ comments, we suggest that the name/pronoun in richer contexts like these (e.g. 19, 21, 24, 26b) can be treated as new information, not Given, even though its reference is salient in light of the first conjunct. The reason, we hypothesize, is that the relevant DP is part of a pair, which serves as an answer to the salient question under discussion, one of whose elements does contrast across the two conjuncts. For example, the answer in (26b) replies to a question under discussion (QUD) which makes salient pairs of names and days, such as <Li, on Monday> and <Sheryl, on Tuesday>. We suggest that Sheryl in (26b) can be considered new information in the context of the pair <Sheryl, on Tuesday>, because the pair itself contrasts with <Sheryl, on Monday>. We take the difficulty some speakers found in accepting (26a) in out of the blue contexts to result from a difficulty in accommodating a QUD which contrasts these pairs.

With this perspective in mind, the Newness Generalization indeed allows non-contrastive ellipsis remnants that the Contrast Generalization ruled out. However, allowing merely new ellipsis remnants can only be part of the solution. The problem is that the Newness Generalization by itself would also rule in all the remnants that the Stranding Generalization shows to be impossible. As we saw in (16)-(17), the richer contexts allow a non-contrastive name to appear as an ellipsis remnant but not a non-contrastive pronoun. (27) gives another illustration of this point. If the richer context allows Jian to be treated as new, it should allow the same possibility for him, contrary to fact.

The moral of this section is that the Stranding Generalization shows a split in behaviour, between functional and lexical elements, that manifests independently of information structure. Even when information structure is held constant—as in the richer contexts that allow non-contrastive ellipsis remnants—functional elements are still subject to the Stranding Generalization. A condition on information structure, though necessary in any account of ellipsis, is therefore not sufficient to explain the Stranding Generalization. Any explanation will have to invoke additional factors.

In section 3 we will argue that the Stranding Generalization follows from independently motivated prosodic constraints, if ellipsis is computed together with prosody. We will first show that the patterns we have just described pose a significant problem for Move and Delete theories of ellipsis.

2.3 Move-and-Delete theories

Many current theories analyze ellipsis as involving deletion of a syntactic constituent at PF. The ability of constituents to survive ellipsis is then tied to prior movement out of the deleted constituent. In many cases, this requires so-called ‘exceptional’ movements, which do not occur in corresponding non-elided utterances. For example, gapping is commonly analyzed as involving something like vP ellipsis, after movement of remnants from the ellipsis site (28a) (Jayaseelan 1990; Toosarvandani 2013; Sailor & Thoms 2014). Movements like in (28a) are not possible in non-elliptical utterances (28b).

Such theories almost always assume that it is focus that drives these movements (Jayaseelan 1990; Gengel 2007; Toosarvandani 2013). This is motivated by the Contrast Generalization (section 2.2), in gapping and other related ellipsis constructions, such as pseudogapping and fragment answers. This in turn has been argued to explain the exceptional nature of these movements. Boone (2014) and Weir (2015), for example, argue that they are ‘last-resort’ movements, made possible only to ensure that contrastively focused constituents do not remain in the ellipsis site.

However, we have seen that the Contrast Generalization does not hold: non-contrastive material can, in some circumstances, be remnants of gapping constructions. This is a significant problem for Move-and-Delete (M&D) theories. One possible move would be to allow optional movement of non-contrastive constituents like to-im in (29a). However, this would mean that movement out of ellipsis sites is not a last-resort phenomenon, nullifying the arguments of Boone (2014) and Weir (2015). It also makes it extremely difficult to account for the Stranding Generalization. It is not clear why optional movement would be allowed for to-im in (29a), and Jian in (29b), but not the pronoun in (29c). The relevant differences between them are prosodic, not syntactic.

We should also consider a theory in which either contrastive or new material moves out of the ellipsis site. This would derive the Newness Generalization in (25), and could potentially be motivated as a last-resort movement, as in Boone (2014) and Weir (2015). Such a theory would be too weak to account for the Stranding Generalization, for the reasons outlined in the previous section. It would allow the non-contrastive PP to-im to move in (29a), but also incorrectly allow the pronoun him to move in (29c). Something would need to be said to account for why movement is impossible precisely when it would violate the Stranding Generalization.

A reviewer suggests the possibility that weak pronouns cliticize in the syntax, bleeding the possibility of movement out of the ellipsis site. We are not aware of evidence that weak pronouns cliticize, or if they did, why they could not cliticize to a position outside of the ellipsis site.¹⁵ For possessors, this cliticization would have to involve an illegitimate syntactic lowering operation. More problematic still are cases like (20)—repeated as (30)—in which the possessor pronoun survives when provided with a prosodic host. The account here would have to be that the possessor first lowers onto the numeral, then raises together with it out of the ellipsis site. We do not think these movements can be motivated on independent syntactic grounds.¹⁶

The reviewer suggests alternatively that movement out of ellipsis sites is subject to Holmberg’s Generalization (Holmberg 1999), which prevents weak pronouns from moving past the landing site of the lexical verb. We can see a number of problems with this idea, but we will mention one that we take to be decisive. Holmberg’s Generalization applies to any constituent that can undergo object shift. In Icelandic, it prevents both weak pronouns and lexical noun phrases from moving past the verb. Thus, if Holmberg’s Generalization constrained movement out of ellipsis sites, it would prevent any object from appearing as an ellipsis remnant.¹⁷

We conclude that there is no promising syntactic account of why non-contrastive lexical noun phrases can move out of ellipsis sites, but non-contrastive pronouns cannot. Without such an account, a M&D theory cannot explain the Stranding Generalization. We will therefore pursue an account in which ellipsis remnants are not required to move.¹⁸ An additional benefit will be the possibility of eliminating ‘exceptional’ movements from the grammar.

3.1 Match Theory

We begin by briefly introducing Match Theory (Selkirk 2009; 2011), which is an ‘indirect reference’ theory. Indirect reference theories posit that the syntax-prosody mapping is governed by two competing forces: the pressure for the prosody to transparently reflect syntactic structure; and independent prosodic well-formedness constraints.

The pressure to reflect syntactic structure stems from Match Constraints, which force syntactic elements to map to categories in the prosodic hierarchy in (31).

These Match Constraints are as follows, with wording adapted from Weir (2012):

However, these constraints can be outranked by prosodic well-formedness constraints. For example, there have been argued to be constraints forcing ɸs to be both minimally binary (BINMIN) and maximally binary (BINMAX). If such constraints outrank MATCH PHRASE, the output will not transparently match the syntactic structure. For example, in (33), the need to satisfy BINMIN prevents the DP dogs from mapping to ɸ, so that the winning candidate violates MATCH PHRASE.

According to Match Theory, syntax-prosody mismatches arise out of interactions like these between Match constraints and prosodic well-formedness constraints. In the next sections, we will briefly introduce relevant constraints governing the prosodic behavior of function words and focused material.

3.2 Function words

Function words generally fail to map to ω. As a result, it is common to assume that Match Word only applies to lexical words (Selkirk 2011; Weir 2012). However, we will follow Tyler (2019), who argues that failure of function words to map to ω is instead a syntax-prosody mismatch of the sort just outlined. Specifically, he argues that many function words have violable prosodic subcategorization frames (Zec 2005; Bennett, Harizanov & Henderson 2018). For instance, he argues that object pronouns have the subcategorization (henceforth SUBCAT) frame in (34).

This SUBCAT frame requires that the pronoun have as its mother a node of category ω, and material within ω to its left. This requirement is satisfied by the structure in (35).¹⁹

A SUBCAT constraint will assign a violation mark to every pronoun that fails to map to the structure in (35). Note that for ease of reading, we have suppressed representation of any categories below the level of ω. Thus, any material not enclosed in (_ω) maps to a bare syllable in what follows, with (_ω (_ω saw) him) indicating cliticization of him to saw.

If SUBCAT constraint outranks MATCH WORD and MATCH PHRASE, the pronoun will always cliticize when possible (36). Candidates (a–c) are ruled out by SUBCAT: in (a) and (c), the pronoun has not cliticized at all; in (b) it has cliticized, but to ɸ instead of ω. Among the candidates that satisfy SUBCAT, the more Match-compliant candidate (e) wins, because candidate (d) unnecessarily fails to map the verb to ω. The result, therefore, is cliticization of the pronoun, as in (35).

As noted in section 2.1, the syntax will sometimes deliver a structure in which there is no appropriate host for a functional item (e.g. Him leaving surprised me, 11c). This receives a simple explanation in Tyler’s system: since there is no candidate in which SUBCAT is not violated, the most Match-compliant candidate wins (37).²⁰ The pronoun will map to ω, and as a result it will be stressed. Thus, the use of the unreduced non-contrastive pronoun is an ‘emergence of the unmarked’ effect.

An important fact about the prosodic behavior of object pronouns is that they cannot cliticize when embedded inside a conjunction, even when there is a linearly adjacent potential host.

We therefore need a way to rule out a parse like (39).

This shows that Tyler’s (2019) SUBCAT constrain for object pronouns is not quite sufficient. Note that what seems to be relevant here is the presence or absence of a syntactic relationship between the functional item and a potential host. Verbs and prepositions select and assign case to their complements; there is no such relationship between either V/P and a DP conjunct, or between and and a DP conjunct. In light of this, we propose to generalize (34) to (40). In this revised form, SUBCAT constraints do not encode prosodic requirements of individual lexical items such as object pronouns, as in Tyler (2019). Rather, they preserve syntactic relationships, like case assignment or selection, in the prosody:

The constraint in (40) correctly predicts that pronouns will not cliticize to elements like and, with which they are not in a featural relationship. It also captures in a parallel fashion the fact that prepositions do not cliticize onto elements like and ([tu and from], c.f. [*tə-and from]). And crucially, it applies to feature-chains, not specifically to function words as in Tyler (2019). This means that in addition to punishing function words that appear without their potential hosts (violations of the Stranding Generalization), SubCat will also punish hosts that do not appear with their potential function word clitics. It will also mean that SubCat will be satisfied vacuously if both elements delete, since the feature-chain will not be present in the output.

Conceived in this way, SubCat bears a strong resemblance to other constraints that have been argued to preserve syntactic relationships in the prosodic structure. For example, Richards (2016) proposes a family of ‘Contiguity’ constraints, which force the prosody to reflect syntactic relations such as selectional and Probe/Goal relations. Similarly, Clemens (2014) argues for an ARGUMENT-ɸ constraint, which specifically requires verbs and their arguments to phrase together.²² These constraints, as well as our revised SUBCAT, can be seen as minimizing the prosodic distance between elements in a feature-chain.

3.3 Focus

When a functional item is contrastively focused, it does not cliticize onto adjacent elements. This can be explained by a constraint like STRESS-FOCUS, adapted from Truckenbrodt (2006).

This constraint will require focused pronouns (marked by the feature F) to map to ɸ in order to bear a pitch accent.²⁴ (42a) thus establishes the ranking STRESS-FOCUS ≫ SUBCAT (42b).

3.4 Givenness

We call an element ‘given’ if it denotes an individual, a property, a relation or a proposition that has been made salient enough in the preceding discourse.²⁵

In cases like (43a), where both verb and object are deleted, we assume that both the verb and object are marked as Given (G-marked, see e.g. Fery & Samek-Lodovici (2006); Selkirk (2008); Büring (2016); Kratzer & Selkirk (2020)).²⁶

As discussed in section 2.2, rich contexts like (44a) allow even previously mentioned material to be treated as new information (if they are part of a pair that stands in a contrast with another pair). Following Fery & Samek-Lodovici (2006); Selkirk (2008) and others, newness is not marked in syntax, but is treated as the absence of G-marking. Thus in such contexts, only the verb is G-marked (44b).

3.5 Ellipsis

We propose that the pressure to elide comes from a constraint, ELIDE, which penalizes the realization of G-marked nodes.

If ELIDE is dominated by MATCH WORD, there will be no ellipsis (46b).

However, if the ranking is instead ELIDE ≫ MATCH WORD, G-marked nodes will have to delete. Thus, ellipsis arises from an optional reranking, in which ELIDE dominates Match Word.²⁸ This optionality in ranking corresponds to the optionality of inserting an E-feature in syntactic theories of ellipsis like Merchant (2001).

ELIDE refers specifically to G-marked elements, and therefore does not apply either to F-marked or merely new elements. Contrastive or new elements will always escape ellipsis, in line with the Newness Generalization (section 2.2). As noted, however, this is not sufficient to explain the Stranding Generalization, to which we now turn.

3.6 Back to the Stranding Generalization

The Stranding Generalization says that functional items cannot be stranded by ellipsis without their (potential) prosodic host. This means that an accusative pronoun cannot be stranded by deletion of the verb, but can appear if supported by e.g. a preposition. We can now see how this theory of ellipsis explains the Stranding Generalization. When ellipsis is evaluated at the syntax-prosody interface, candidates will contain potential prosodic hosts like verbs because, by hypothesis, deletion has not yet taken place. As a result, SUBCAT constraints play an active role, if ranked above ELIDE. This is crucially unlike cases in which the syntax simply does not provide a potential host, whether by movement or base-generation; we saw in (37) that in such cases, SUBCAT is irrelevant, because there are no candidates in which it can be satisfied. In ellipsis, the syntax does provide a potential host, and SUBCAT becomes relevant.

Consider the ungrammatical example in (48a) (in what follows, we will suppress the representation of ɸ and ɩ, except when required by e.g. STRESS-FOCUS). We are assuming, following section 3.4, that the pronoun can be treated as new in this context: if Given it will simply delete (with the ranking Elide ≫ MATCH WORD). The verb is G-marked, and would have to delete in order to satisfy ELIDE. Any structure without the verb, however, will violate SUBCAT (candidates c–d); recall independent evidence from section 3.2 that cliticization to and does not satisfy SUBCAT.²⁹ With the ranking SUBCAT ≫ ELIDE, then, ellipsis of the verb will be suppressed—violating ELIDE—in order to satisfy SUBCAT.³⁰

Now, consider the behaviour of lexical noun phrases. Again, we are considering contexts in which the object is treated as new information. Because lexical noun phrases do not have SUBCAT requirements, deletion of the verb will not result in a violation of SUBCAT, and ELIDE can be satisfied (candidates c–d). The more Match-compliant candidate (d) wins, since it does not fail to map Yusuf to ω.

This analysis carries over to cases like (50a). We are treating the PP to him as new, rather than given, made possible by the rich context. Deletion of the verb is forced by ELIDE, as in the above cases. Here, both the preposition to and pronoun him are required by SUBCAT to cliticize to each other. Following Tyler (2019), these requirements can be satisfied by the structure in (50b), forming a single phonological word in which both of their SUBCAT requirements are satisfied.³¹ Because there is no SUBCAT violation, the PP to him is an eligible ellipsis remnant.

This contrasts with the minimally different double-object construction in (51a). Here there is no way to satisfy the pronoun’s SUBCAT requirement, unlike the prepositional dative in (50a). Ellipsis is therefore suppressed by SUBCAT.

Finally, the ranking STRESS-FOCUS ≫ SUBCAT guarantees that contrastively focused pronouns—and contrastively focused material in general—will be realized.³² STRESS-FOCUS rules out the SUBCAT-compliant candidate in (d), as well as the Match-compliant candidate in (e). ELIDE ensures that the verb is elided, ruling out (c). Satisfying ELIDE and SUBCAT by deleting both the verb and pronoun is ruled out by STRESS-FOCUS. Thus, the verb will delete and the pronoun will map to ɸ (candidate a); satisfying STRESS-FOCUS will necessarily lead to a violation of SUBCAT.

We argued in section 3.2 that SUBCAT seeks to preserve the syntactic relationship between a functional head and another element. This was captured by making SUBCAT refer specifically to the feature-chain established by, for example, case assignment. Deleting a preposition to the exclusion of its complement, then, will violate SUBCAT, since the feature-chain will not be appropriately represented. Thus, when Wednesday is F-marked, and therefore prevented from deleting by STRESS-FOCUS, its case-assigning preposition escapes deletion in order to satisfy SUBCAT.

We thus have the following constraint ranking for ellipsis in English:

This analysis straightforwardly extends to NP ellipsis. We again consider contexts where non-contrastive remnants are possible, so as to give the pronoun the best chance of surviving. As discussed in section 2, possessor pronouns have SUBCAT constraints that force them to cliticize to their right. Since ellipsis would violate SUBCAT, the G-marked noun must be pronounced.

If the possessor is a lexical noun phrase, there are no SUBCAT requirements to satisfy, and therefore no need to violate ELIDE. The most Match-compliant of the remaining candidates wins.

If other material is present that allows the possessor pronoun to satisfy its SUBCAT requirements, ellipsis is once again possible.

As we showed in section 2.2, pseudogapping is subject to the Stranding Generalization. The explanation is exactly as before. Since the verb is in the input, deleting it would violate SUBCAT.³³

A lexical noun phrase can again appear, since SUBCAT would not be violated.

And once more, a pronoun can appear if it can find a host to satisfy its SUBCAT requirements.

Finally, we consider fragment answers. The analysis is completely parallel to gapping, NP-ellipsis, and pseudogapping. We assume a sentential analysis, since there are connectivity effects Merchant 2004). Given a sentential analysis, the verb is present in the input, and the pronoun’s SUBCAT requirements must again be satisfied.

A lexical noun phrase can appear in the same position, because it has no SUBCAT requirements to satisfy.

And again, a pronoun can appear if it has a host, allowing it to satisfy SUBCAT.

Thus, the ranking ELIDE ≫ MATCH WORD forces deletion of as much G-marked material as possible. But because SUBCAT in turn dominates ELIDE, the pressure to preserve the relationship between function words and their hosts takes precedence over the pressure to delete. The result is that ellipsis can be suppressed in order to preserve this relationship. The ranking SUBCAT ≫ therefore explains the Stranding Generalization.

3.7 Constituency conditions

As Johnson (2017) notes, ellipsis cannot just elide any string. For example, he cites the following, from Hankamer (1979):

These and other similar cases led Johnson to formulate the following condition on gapping remnants:

Something like (65) would follow from a theory in which remnants have moved out of the ellipsis site. Since the present theory does not rely on movement, it is important to show that it does not overgenerate. There thus must be high-ranked prosodic well-formedness constraints that rule out unacceptable strings like those in (64).

In fact, many of them straightforwardly follow from independently motivated SUBCAT constraints. For example, the determiner a prosodically cliticizes to the right (Itô & Mester 2009), diagnosed again by vowel reduction. This requires a rightward cliticizing SubCat frame, motivated in our terms by the selectional relationship between D and NP.

From this it follows, following the same logic as previously, that pen cannot serve as a gapping remnant to the exclusion of the determiner. Because the determiner is in the input, candidates will be evaluated with respect to its SUBCAT constraint. SUBCAT will make determiners and their hosts behave as a unit, just as verbs and accusative pronouns do. They will therefore either be pronounced together, or deleted together (with deletion in this case ruled out by STRESS-FOCUS).

This explanation carries over without modification to cases like (68). We saw in section 2 that possessor pronouns cliticize to the right. Vowel reduction also shows rightward cliticization of some and the.³⁴ SUBCAT will then rule out deletion of their potential hosts. There is no clear evidence that every cliticizes, but we suggest that SUBCAT is a general constraint that applies to all determiners: this would mean that every cliticizes as a foot, as Itô & Mester (2009) argue for portmanteau function complexes.

Finally, it also extends to the following contrast, originally from Ross (1970):

Vowel reduction again shows that the infinitival marker cliticizes to the right ([tə] leave would be rude/Leave, he wants [tu]). SUBCAT correctly rules out deleting the host of to, as in (69b). Note that in this case, the unacceptability of (69b) doesn’t follow from the Constituency Condition, since the complement of to is a constituent.

It was pointed out in section 3.2 that SUBCAT constraints generally have the effect of preserving syntactic relationships in the prosody. For example, determiners and possessors cliticize within the noun phrase, and accusative pronouns to their case-assigning heads. We suggested that SUBCAT constraints could be assimilated to other constraints that preserve syntactic relationships, like those in Richards (2016) and Clemens (2014). If this speculation is on the right track, then the fact that ellipsis remnants behave like constituents is expected, even in a theory that operates over prosodic constituents. But we expect that phenomena that do not have an explanation in terms of constituency, like the Stranding Generalization, to be the result of independent prosodic well-formedness constraints.

3.8 Summary

We have argued for an implementation of ellipsis at the syntax-prosody interface. This required only minimal assumptions. The constraint SUBCAT, which requires functional items to phonologically cliticize when possible, is ranked above ELIDE, which requires G-marked nodes to delete. As a result, deletion of a prosodic host in order to satisfy ELIDE will always be prevented by SUBCAT. This explains the Stranding Generalization. In particular, it explains why functional items can be deprived of a prosodic host by movement or base-generation, but not by ellipsis. Movement and base-generation take place prior to the syntax-prosody mapping, and will yield candidates in which satisfaction of SUBCAT is not possible. But because ellipsis does not take place in the syntax, the input of the syntax-prosody mapping will contain candidates in which SUBCAT can be satisfied.

4.1 The P-Stranding Generalization

A widely observed crosslinguistic characteristic of ellipsis is the P-stranding generalization (Merchant 2001; 2004). Merchant argues that languages like English, which have P-stranding, allow a fragment DP to appear ‘bare,’ with the preposition elided (70a). In languages where P-stranding is unavailable, like German, P must appear with its complement (70b). The same facts hold for sluicing.

The P-stranding generalization follows from a Move-and-Delete theory.³⁵ The DP fragment can appear ‘bare’ in (70a) because the necessary movement out of the ellipsis site is able to strand the preposition. Since preposition-stranding movement is unavailable in German, bare DP fragments are impossible.

This explanation is consistent with our theory. Ellipsis fragments are not required to move, for us, but nothing prevents them from doing so. To derive gapping and pseudogapping, Move-and-Delete theories require ‘exceptional’ movements, which are unable to occur outside of ellipsis contexts. But this is not the case for examples like (70a), which can be derived from attested movements like (71a). Since SUBCAT cannot be satisfied, failure to elide the G-marked preposition would be gratuitous.³⁶

Our account may explain some counterexamples to the P-Stranding Generalization. For example, Russian lacks P-stranding (with a few exceptions, Podobryaev (2009)). Nevertheless, there are a number of prepositions that, despite being unable to strand (72a), can appear ‘bare’ in sluicing (72b, from Ionova (2020)). The same facts hold in gapping constructions (Ionova 2020).

By contrast, other prepositions, like o ‘about,’ must be pronounced with their complements.

Similar counterexamples in Spanish and Brazilian Portuguese (Vicente 2008; Rodrigues, Nevins & Vicente 2009) and Polish (Szczegielniak 2008) have been explained away as involving ‘pseudosluicing.’ Ionova (2020) shows, on the basis of case-matching effects, that pseudosluicing cannot cover all of these cases. She goes on to argue for a prosodic generalization about the relevant prepositions. Those that can delete under sluicing and gapping (‘heavy’ prepositions) map independently to ω, while those that cannot delete (‘light’ prepositions) phonologically cliticize to ω.

In Tyler’s (2019) theory, that means that the ‘heavy’ prepositions do not have SUBCAT frames, or are forced to map to ω for independent reasons, while cliticization of ‘light’ prepositions is required by SUBCAT. If this is the case, then the ranking SUBCAT ≫ ELIDE will prevent light prepositions from deleting, even when given. Heavy prepositions, by contrast, will be allowed to delete, since they are not required to cliticize.

Ionova (2020) argues that, because of this sensitivity to prosody, the deletion of heavy prepositions is a post-syntactic process of ‘P-deletion,’ separate from ellipsis, which is for her a syntactic process. The problem for this account is that, as she notes, P-deletion is impossible in non-elliptical utterances.

This would be mysterious if P-deletion was a separate process from ellipsis. In our account, P-deletion follows naturally from the interaction of ELIDE with independent prosodic constraints. We conclude that our account offers a promising route to explaining exceptions to the P-Stranding Generalization in Russian, as well as Serbo-Croatian, which behaves similarly (Stjepanović 2008), and potentially other languages.

4.2 Irish responsive ellipsis

Bennett, Elfner & McCloskey (2019) examine the curious behaviour of subject pronouns in Irish ellipsis. ‘Simple’ subject pronouns (75)—which are necessarily unfocused—cliticize leftward onto the verb. They argue that this prosodic relationship is fed by subject incorporation, a ‘morphosyntactic rebracketing operation.’

In responsive ellipsis, the subject pronoun is deleted, disrupting its relationship with the verb (77a). However, if the verb is ‘emphasized,’ the pronoun must be realized, and carries a strong focal accent (77b).

The account in Bennett, Elfner & McCloskey (2019) follows the same logic as ours. They propose a constraint, ELIDE (X_[∅]), which penalizes the realization of nodes marked for deletion. Another constraint, SC(SPI) penalizes structures in which subject pronoun incorporation has not taken place. SC(SPI) outranks a faithfulness constraint, FAITH(S⟹M,X⁰) which prevents the phonological output from containing dominance relations not found in the syntax.

Pronouncing the subject pronoun, whether it is incorporated (candidate c, triggering allomorphy of the verb), or not (candidate b)–will violate ELIDE (X_[∅]). The ranking ELIDE (X_[∅]) ≫ SC(SPI) therefore guarantees that ellipsis bleeds the pronunciation of the subject pronoun se /ʃe/.³⁷ With respect to subject pronouns, then, Irish is the mirror image of English, preferring to disrupt a close prosodic relationship between verb and pronoun in order to satisfy the pressure to elide.

In turn, Bennett, Elfner & McCloskey (2019) argue that the behaviour of subject pronouns when the verb is emphasized follows from another constraint, BINARY FOCUS:

Because BIN-FOC outranks ELIDE(X_[∅]), the subject pronoun–which they show is semantically Given—is required to escape ellipsis, and map to ω. Failure to do so would render the focused verb non-binary.

Irish thus displays exactly the same kind of interaction between the pressure to elide and independent prosodic well-formedness constraints that are expected if ellipsis is calculated in the syntax-prosody mapping. If such constraints, like BIN-FOC, are sufficiently high-ranked, ellipsis will have to cater to their their needs.

4.3 St’at’imcets VP ellipsis

St’at’imcets VP ellipsis offers further support to this theory.³⁸ Davis (2014) notes an interesting interplay between ellipsis and the prosodic requirements of auxiliaries. In St’at’imcets, ‘light’ auxiliaries cannot be stranded at the right edge of a string after ellipsis, (81b). They instead undergo what he calls ‘rhetorical lengthening’ (81c).

Davis suggests that this lengthening takes place in order to satisfy the following condition:

He moreover suggests that this derives from the need for the auxiliary to procliticize to following material, noting that there is no rhetorical lengthening if, for example, the second position clitic t’u7 is present (83b).

Davis further notes the following pattern, with three auxiliaries in the antecedent: plan ‘already,’ wa7 ‘imperfective,’ and the motion verb p’an’t ‘to return.’ Again, a ‘light’ auxiliary cannot be stranded by ellipsis (84b). Instead of lengthening the auxiliary, the additional auxiliary p’an’t (‘return’) must be pronounced (84c).

This follows straightforwardly from the present theory. The procliticizing behavior of the auxiliary wa7 strongly suggests a right-cliticizing SUBCAT frame. This SUBCAT constraint can be satisfied by pronouncing given material to its right, for example ‘return’ in (84c); this establishes the ranking SUBCAT ≫ ELIDE. The St’at’imcets-specific twist is that, when SUBCAT cannot be satisfied by the pronunciation of given material, the auxiliary can instead lengthen. This would follow from a sufficiently low-ranked Dep constraint. Again, these are exactly the kinds of interactions between ellipsis and prosodic well-formedness constraints the present theory predicts. We expect that a close examination of ellipsis and prosody in other languages will uncover other such interactions.