Attention and Locality: On Clause-Boundedness and Its Exceptions in Multiple Sluicing

Grano and Lasnik (2018) propose an analysis of clause-boundedness effects, cases where some dependency holds or process can take place within a minimal clausal domain, but not across a finite-clause boundary. Clause-boundedness effects have been observed in a variety of construction types, including inverse scope phenomena, gapping, pseudogapping, comparative deletion, antecedent-contained deletion, multiple questions, family-of-questions interpretations, reciprocal and anaphor binding, tough-movement, and multiple sluicing.

To give an example, consider the phenomenon of inverse scope illustrated in (1). A wide scope construal of the universally quantified argument is unavailable when a clause boundary separates it from a c-commanding existentially quantified argument. If we follow May (1977) in assuming that the universally quantified argument achieves wide scope via Quantifier Raising (QR) in (1a), the judgment in (1b) suggests that this operation cannot take place across a clause boundary. (See also Farkas 1981 and Farkas and Giannakidou 1996 for earlier treatments of clause-boundedness in QR.)

Grano and Lasnik (2018) discuss an exception to clause-boundedness that they dub the bound pronoun effect. When the subject of the embedded clause is a pronoun bound by an argument in the higher clause, the embedded quantifier is apparently able to scope out of the embedded clause. (The judgment reported here is from Grano and Lasnik 2018.)

The bound pronoun exception has been noted sporadically in the literature, usually in the context of discussing the existence of a clause-boundedness effect for some particular construction (see, e.g., Sloan 1991 for family of questions, Nishigauchi 1998 for multiple sluicing in Japanese, and Merchant 2001 for multiple sluicing in English and gapping). To our knowledge, Grano and Lasnik (2018) are the first to collect these various observations under the umbrella of a single generalization and attack them head on. Under Grano and Lasnik’s account, clause-boundedness is phase-boundedness. In examples like (1b), the embedded clause is a phase, blocking QR of the universal to a position above the embedded CP phase. However, under their account, in sentences like (2), the bound pronominal subject in embedded clauses renders its minimal CP a nonphase (for reasons concerning the feature content of bound pronouns that we discuss below), allowing the universally quantified argument to cross a (nonphasal) clause boundary.

In this article, we focus on one of the domains of clause-boundedness explored by Grano and Lasnik, namely, multiple sluicing. Sluicing (Ross 1969) is a construction that involves ellipsis in an interrogative clausal complement that leaves only the wh-phrase overt. Example (3) introduces some terminology we will use throughout. Following the convention in the literature on ellipsis, strikethrough represents unpronounced/elided material.

Sluicing usually requires a declarative clause, called the antecedent, that is parallel in some way to the interrogative clause in which ellipsis takes place, called the sluice. The antecedent typically contains a DP, called the correlate, that is the correspondent of the nonelided wh-phrase, called the remnant.¹

Multiple sluicing is sluicing with more than one remnant. In (4), for example, which student and with which professor count as remnants. For concreteness, we follow Park (2014) and Abels and Dayal (2017, 2021), among others, in assuming that multiple sluicing involves multiple wh-fronting, exceptionally licensed under ellipsis in English.²

Multiple sluicing is subject to clause-boundedness effects, where multiple remnants are only possible when their correlates are clausemates, as in (4); compare (5), where they are separated by a finite-clause boundary.

Since some professor is contained within CP₂, but some student is not, multiple sluicing is impossible.

Like the other constructions discussed by Grano and Lasnik (2018) in which clause-boundedness is observed, multiple sluicing is subject to the bound pronoun effect.

Below, we raise empirical challenges to Grano and Lasnik’s (2018) account, introducing new data that demonstrate that bound pronouns are neither necessary nor sufficient for avoiding clause-boundedness in multiple sluicing. That is, we show clear exceptions to clause-boundedness where no bound pronouns are involved, and cases where clause-boundedness is active when there is a bound pronominal subject in the embedded clause. To make sense of the pattern, we defend the generalization in (7).

This generalization encompasses the bound pronoun effect. In examples like (1b), the embedded subject is not construed with any argument introduced in the matrix clause. Hence, (1b) runs afoul of (7) and is correctly ruled out. Embedded bound pronominal subjects, on the other hand, do not usually shift attention away from the most salient argument in the higher clause, which serves as its binder. Our generalization in (7), then, captures the bound pronoun effect. Crucially, this generalization is couched in terms of discourse/information structure, and we propose to derive it from principles governing the flow of attention in discourse. This is perhaps surprising, since clause-boundedness appears, on its face, to be a syntactic phenomenon. If the generalization in (7) is valid, it provides an argument that syntax must be sensitive to such discourse notions.

In section 2, we briefly review Grano and Lasnik’s (2018) proposal, illustrating how it derives clause-boundedness in multiple sluicing and accounts for the bound pronoun effect. We also discuss Abels and Dayal’s (2017, 2021) proposal that aims to account for the same multiple sluicing facts. In section 3, we introduce a new dataset that demonstrates that bound pronouns are not necessary for suspending clause-boundedness, challenging the approaches of Lasnik (2014), Park (2014), Abels and Dayal (2017, 2021), and Grano and Lasnik (2018). In section 4, we sharpen our discourse-based characterization of the empirical pattern, adopting notions from Centering Theory (Grosz, Joshi, and Weinstein 1995), which has been proposed to model pronoun reference resolution and patterns of coherence between utterances in discourse. An important part of our proposal is that it requires syntactic constraints on movement and deletion to be sensitive to discourse coherence; accordingly, in section 5, we entertain plausible theoretical accounts of our generalization in (7), each making syntax sensitive to the relevant discourse relations between clauses. In section 6, we conclude.⁴

Grano and Lasnik (2018) advocate a view in which clause-boundedness is nothing other than phase-boundedness. In their account, bound pronouns optionally come with unvalued ϕ-features, which are valued when their binders are merged via a feature transmission mechanism (Kratzer 1998, 2009). They further assume that the presence of unvalued features in a candidate Spell-Out domain “voids” or “suspends” phasehood (following proposals in Felser 2004).⁵ Consequently, in sentences with a pronoun in an embedded finite clause that is bound by an argument in a higher clause, the embedded clause’s phasehood may be suspended because of as-yet-unvalued ϕ-features.

Applying this idea to the domain of multiple sluicing, Grano and Lasnik (2018:sec. 4.3) assume, following Lasnik (2014), that the movement of the second wh-phrase undergoes one-fell-swoop movement from its base position to its final landing site. In (8), an annotated version of (6), CP₂’s phasehood is suspended because of the unvalued features on the pronoun, thereby allowing the second remnant (henceforth Wh2) to escape CP₂.

Were CP₂ to constitute a phase—say, if the embedded subject were not a bound pronoun, as in (5)—a phase boundary would emerge, blocking movement of Wh2 out of CP₂.⁶

Grano and Lasnik note that the bound pronoun effect appears to obtain only with bound subject pronouns, and that even a bound object pronoun does not trigger the effect. This is illustrated in (9).

To account for this asymmetry, Grano and Lasnik follow a suggestion of Hisatsugu Kitahara that phase suspension only takes place when there are unvalued features on the head of the complement to a phase head. This proposal is outlined in (10).

Bound pronominal subjects with unvalued ϕ-features fail to value T⁰’s ϕ-features. Since T⁰ is C⁰’s complement, C⁰ fails to count as a phase head. When the subject is nonpronominal, as in (9b), its ϕ-features will always value the corresponding features on T⁰, giving rise to a phase boundary. A result of this is that only in a configuration like that in (11) (where [ϕ-:____] represents unvalued ϕ-features) does C⁰ fail to count as a phase.

Note that the discourse-based characterization of the bound pronoun effect in (7) explicitly (and apparently correctly) stipulates its subject orientation. We return in section 4 to how this subject specificity might be derived from discourse factors.

Abels and Dayal (2017, 2021) propose an alternative account of the bound pronoun effect in multiple sluicing. Specifically, they advocate a “short source” solution, where apparent cases of nonlocal multiple sluicing are, despite appearances, monoclausal. The basic idea is that an intuitively synonymous monoclausal presluice, as in (12), is available when the embedded pronoun is bound by the matrix subject correlate in the antecedent.

In the absence of an embedded bound pronominal subject, no synonymous short source will be available, as the external arguments of the upper and lower predicates are nonidentical.

In a similar way, this line of analysis also explains why the bound pronoun must be the subject. It is only when the embedded subject is coreferential with the main clause’s subject that a short source is available. This analysis requires a limited degree of nonidentity between the sluice and the antecedent, sufficient, at least, to introduce modality in the sluice. Whatever the degree of flexibility allowed, we expect that it will not be sufficient to permit the creation of a short source in the absence of a bound subject pronoun.

We introduce a new pattern of data that demonstrates that bound pronouns are unnecessary for suspending clause-boundedness in multiple sluicing. We further show that a short-source account is unavailable in such cases. Insofar as exceptions to clause-boundedness should have a unified analysis, such data are challenging for Grano and Lasnik’s account, as well as for short-source accounts (e.g., Lasnik 2014, Park 2014, Abels and Dayal 2017, 2021). In contrast, we will show that our generalization in (7) extends to these cases.

To begin, consider the following cases of biclausal multiple sluicing, this time with a different type of embedded subject, an expletive or a quantifier.⁷

In spite of the absence of bound pronominal subjects, both examples are substantially better than cases with referential DP subjects like (5). Grano and Lasnik’s (2018) approach has nothing to say about such cases since it is based on the presence of ϕ-features on the embedded T⁰ that are unvalued within the embedded CP. In both of these cases, however, T⁰ will be valued in the embedded CP, through agreement with either the postcopular DP or the quantifier.⁹

Abels and Dayal’s (2017, 2021) proposal likewise fails to account for such data, as no short source is available for the sluices in (9) and (15). To construct a monoclausal short source, both wh-phrases must be arguments or modifiers of a single predicate, and the result should be synonymous with the sluice. This works in cases where the embedded subject is bound by the matrix subject, as in (12), but there is no way to achieve this in examples like (9) and (15) (we leave it to the reader to try). The best we can do yields nonsynonymous (and pragmatically infelicitous/incongruent) sluices. For instance, a plausible argument position (possibly the only one) that which student may occupy in (9) is the one occupied by the expletive there: #which student was a problem with which professor. In (15), it is the position occupied by the quantified embedded subject: #which student talked about which topic.¹⁰

We now turn to another set of examples, where clause-boundedness is again suspended in the absence of a bound subject pronoun. As in the original bound pronoun cases, but unlike in the examples just discussed, the embedded subject in these cases is coreferential with a DP outside of the embedded clause. Crucially, though, the coreferential element does not c-command the embedded subject. Since binding is no longer at issue, we can use a broader range of possible anaphoric DPs, including pronouns, epithets, and R-expressions. Nonetheless, clause-boundedness is suspended exactly when coreference obtains.

Because of the absence of c-command between the coreferential elements, the ϕ-features of the embedded subject cannot be valued by the coreferential DP, under the standard assumption that such valuation takes place under an Agree relation that requires c-command. Consequently, Grano and Lasnik’s analysis does not extend to these cases. Similarly, no synonymous monoclausal short sources are available, rendering Abels and Dayal’s analysis unable to account for these cases as well.

We take the above data to show that it is not necessary for the embedded subject to be a pronoun bound by the matrix subject in order for clause-boundedness effects to be neutralized. Instead, as already anticipated above, we claim that normally porous embedded clauses become locality domains for multiple sluicing when the embedded subject is “shifty,” that is, when it shifts attention away from the most salient d-ref evoked in the matrix clause. All of the cases we have discussed in this section involve embedded subjects that do not induce such a shift in attention, either because they do not have reference (as in the expletive or quantifier examples) or because they are anaphorically dependent (even if not pronominal). In short, we emphasize here that bound pronouns are not necessary for suspension of clause-boundedness, which casts serious doubt on any accounts of clause-boundedness (or its exceptions) that rest on specific featural properties of bound pronouns. Below, we sharpen our definition of Shifty Subjects in (7) and demonstrate that bound pronouns are also insufficient to suspend clause-boundedness.

Thus far, we have been somewhat loose in the formulation of our generalization in (7), repeated here.

Specifically, we have not provided an explanation of what we take to be the most salient d-ref in a clause or of what it means to shift attention away from such a d-ref. In this section, we make our generalization more precise, adapting several notions from Centering Theory (Grosz, Joshi, and Weinstein 1995), a theory of the flow of attention in discourse across utterances and how the attention status of a discourse entity affects its syntactic realization.

Centering Theory, henceforth CT, is a theory of discourse coherence and salience, and the relationship between these notions. To illustrate the sort of phenomena CT is concerned with, consider the discourse in (17).

Clearly, the intended referent of he in (17e) is Tony, but this renders the discourse intuitively incoherent and somewhat misleading. Sentence (17e) has a garden path feel; it is only when we get to the information in the predicate that we can conclude that the subject refers to Tony, and not Terry. In other words, there is a strong preference for interpreting the subject in (17e) as coreferential with the subject in (17d). A more natural alternative to (17e), for instance, would be (18), where the subject pronoun persists in referring to Terry, as it does in (17b–d).

CT explains patterns like this by making reference to the realization of d-refs as linguistic expressions, both with respect to the form of these expressions (whether they are pronouns or definite descriptions, proper nouns, etc.) and with respect to their grammatical role in syntax, given that such roles are often associated with information packaging (borrowing a term from Chafe 1976). CT builds on the observation that there are clear patterns for how salient/topical entities in discourses tend to be realized in utterances. The discourse in (17) runs afoul of pressures having to do with a preference governing the realization of salient entities across utterances: salient entities tend to be realized in consistent ways. Discourses that maintain such consistency—that is, preserve the property that a narrative is about some specific topical entity—are easier to process and are preferable to discourses that do not (the discourse in (17) being a case in point).

We assume that discourses are broken up into discourse segments. In early work in CT, discourse segments were identified with utterances. We will instead assume a generalization suggested by Kameyama (1998), Miltsakaki (2003), and Poesio, Stevenson, and Di Eugenio (2004) whereby individual finite clauses constitute discourse segments. Given that discourses unfold over time, we take the ordering among these segments to be determined by their linear order, so that those associated with matrix clauses precede those associated with embedded clauses.¹¹

A fundamental tenet of CT is that each discourse segment is associated with a single d-ref that constitutes its focus of attention. This focus of attention corresponds to what the segment is intuitively “about”—that is, its topic. The notion of aboutness used in CT is grounded in terms of what came before the current segment, U; correspondingly, the focus of attention in segment U is called its backward-looking center, notated C_b(U). Additionally, each discourse segment U is associated with its forward-looking centers C_f(U), a (partially) ordered list of d-refs that are realized by some DP in U. This list contains the d-refs that are realized in U, and the ordering among the elements in the C_f list corresponds to the d-refs’ relative salience in U. Salience is determined by a number of factors—most prominently, grammatical role. For the most part, subjects are ranked above objects, which are in turn ranked above obliques—though as we will show, there are principled exceptions to this. The C_b of a discourse segment U_n is defined to be the d-ref realized in U_n that is ranked highest among the C_fs from the previous utterance, that is, C_f(U_n−1). The highest-ranked d-ref in C_f(U) is therefore the d-ref that is most likely to be the C_b in a subsequent discourse segment, and indeed it must be so if it is realized in the next segment. This d-ref is therefore given a distinctive name; it is the discourse segment’s preferred center, C_p(U).

We can now use these notions to distinguish two types of transitions between discourse segments.

In cases of attentional maintenance, the same d-ref remains at the top of adjacent discourse segments’ C_f lists. For attentional shift, the d-refs encoded by the two are distinct.¹² The intuition here is that coherent discourse keeps a single element, the focus of attention, in a grammatically prominent position across discourse segments. When this is not the case, the discourse is taken to be marked.

This attentional shift vs. maintenance distinction allows us to account for intuitions concerning short discourses like (20).

What goes wrong here is a violation of a preference for interpreting the subject pronoun in U₂ as coreferring with Jack, not Bill. The transition between these utterances involves a case of attentional shift: the C_p of U₁ is Jack, as it is the highest-ranked element of C_f(U₁) (as the d-ref realized by the subject), while the C_p of U₂ is Bill (as the d-ref realized by that clause’s subject). Looking back at the discourse in (17), repeated here, one can also see why the final transition between discourse segments gives rise to anomaly, while the other transitions are coherent.

For all but the final sentence, the highest-ranked d-ref in the C_f list is Terry. As a result, all the transitions constitute instances of attentional maintenance. However, in the last sentence, (17e), the only d-ref that is realized is Tony. As a result, the C_p for this discourse segment is Tony, which is distinct from the C_p of the preceding sentence, (17d), giving rise to attentional shift.

With this much in place, we can make our proposal precise: deviation from optimal discourse coherence in the form of attentional shift induces clause-boundedness.

We can now apply this condition to derive the core cases of clause-boundedness suspension, cases with and without bound pronouns. First, consider again example (6), repeated here as (22) with discourse segments annotated.

By the definitions in (19), the discourse in the segments hosting sluicing in (22)—namely, U₄ and U₅—is a case of attentional maintenance: the d-ref associated with the bound pronoun he is the C_p of the embedded discourse segment, C_p(U₅), and is identical to the d-ref realized by the subject of U₄, which student, the C_p of the interrogative discourse segment, U₄. Since clause-boundedness is only triggered in cases of attentional shift, we correctly predict clause-boundedness effects to be absent in examples like (6)/(22).

On the other hand, if the embedded subject in U₅ is not a bound pronoun, but instead a name or other R-expression, the d-ref it realizes, which is the C_p of that discourse segment, is now distinct from C_p(U₄). As a result, this transition constitutes an instance of attentional shift, triggering clause-boundedness effects, as in (5), repeated here as (23) with discourse segments indicated.

We will now illustrate how our proposal extends to the other cases of clause-boundedness exemptions introduced in section 3.

At first glance, it is not clear why existential clauses should be porous if attentional shift induces clause-boundedness. The relevant example, (14), is repeated here.

Assuming the expletive subject, there, does not realize a d-ref, two d-refs are realized in the existential clause, neither of which is a C_b: the d-ref realized by the postcopular nominal (hence-forth the pivot), the DP headed by problem; and the d-ref realized by the DP complement to the preposition, which professor. It is not entirely clear what metric one should use to determine their relative rank, since it is not clear what the relevant grammatical roles are in this case. We might instead appeal to structural notions other than grammatical role, such as hierarchical prominence: though plausibly neither c-commands the other, we might appeal to the containment relation and posit that it is the problem-headed DP that outranks which professor. Either way, the d-ref realized by the higher-ranked one will be the C_p of the existential clause, and will not be identical with the C_p of the matrix clause, which is the d-ref associated with which student. Consequently, it would seem that we have an instance of attentional shift.

It has been claimed in the literature on existential sentences that they are topicless, purely thetic/comment structures, or are entirely rhematic (see Mikkelsen 2011 for a survey and discussion). Their function in discourse is typically described as one of simply introducing a new d-ref. If attentional shift requires the introduction of a new topic (i.e., C_p ) distinct from the one immediately before, such sentences cannot trigger attentional shift.

To capture this property of existential sentences, we capitalize on the observation that although the existential pivot is not in the canonical subject position, the expletive locative pronoun there is. We assume then that this expletive has a place in the C_f list, and as a result the existential pivot is not the highest-ranked element of a C_f list (and consequently cannot serve as a C_p ). Since expletives are, by definition, not associated with referential content, there would seem to be no way to include them in the C_f list, which consists of d-refs. We modify the definition of C_f(U), then, and assume that its members are not d-refs but σ, δ pairs, where each σ is a syntactic constituent in U (typically of category DP) that realizes the d-ref δ. When a DP σ does not realize a d-ref, we take this to give rise to a pair σ, in the C_f list.

As before, we contend that the grammatical subject of U_n is typically ranked highest in C_f(U_n), so that even an expletive subject may occupy the highest position in the C_f list. The C_f(U_n+1) for the embedded existential clause in (8), then, has the pair there, as its top-ranked element. Recall that the C_p is defined to be the highest-ranked d-ref in a C_f list. Since such a C_p, as a candidate C_b, is a d-ref, it will be undefined in existential sentences: there is no d-ref present in the highest pair in the C_f list. In such a case, then, the transition between a discourse segment U_n and the one following it, U_n+1, cannot be categorized as attentional shift: there is no d-ref corresponding to C_p(U_n+1) that is distinct from C_p(U_n).

This refined characterization of attentional shift is supported by data concerning discourse coherence. In (24), which seems like a coherent discourse, the presence of the discourse segment U₂ does not induce attentional shift because of the existential structure.

There is no problem with interpreting the pronoun in U₃ as the d-ref Jack in spite of the intervening clause.

One might object, however, that this discourse is felicitous because the intervening existential clause does not introduce any new animate d-refs consistent with masculine-gender pronouns, which are necessary to induce ambiguity in the interpretation of the pronoun. Indeed, in an example like (25), where such a d-ref is introduced in the intervening clause by a DP in subject position, the ambiguity does give rise to a mild incoherence, if the pronoun in U₃ is to be interpreted as identical to Jack.

This pattern is as we would expect. The subject DP another man realizes a novel d-ref in U₂. Therefore, it is C_p(U₂), which is distinct from C_p(U₁), namely, Jack. As a result, this constitutes an instance of attentional shift. Attentional shift then takes place again in the transition from U₂ to U₃. Note, however, that introducing such a discourse referent in U₂ with an existential does not have the same effect.

Though the DP another man introduces a male d-ref in U₂, it is our judgment that the most natural interpretation of the pronoun in U₃ is Jack, and that the other interpretation is associated with mild incoherence. This is consistent with what we claimed above: if there is no attentional shift between the first and second (existential) discourse segments, then there is no attentional shift between the second and third. The existential sentence just adds nontopical information to the context in which Jack remains the topic.

The same reasoning we have outlined for existential sentences applies to cases where the embedded subject is a quantificational expression that fails to realize a discourse referent. In example (15) (repeated here), no professor fails to evoke a d-ref, and clause-boundedness is relaxed.

The highest-ranked element in the C_f list of the embedded clause in (15) is the pair no professor, . Consequently, there is no C_p for the embedded clause. As we noted for embedded existential sentences, this fact entails that the discourse transition between the matrix and embedded clauses is not defined. Since attentional shift does not obtain, clause-boundedness is correctly predicted to not be active.

As predicted, just as existential sentences do not induce attentional shift (see (26)), sentences with quantificational subjects/expressions that fail to evoke a discourse referent also fail to do so. The discourse in (27) seems coherent and felicitous with the indicated coreference relations.

Earlier, we presented cases in which an R-expression may be the subject of an embedded clause that does not induce clause-boundedness, noting that such cases are distinguished by the fact that the R-expression must realize a d-ref that is mentioned earlier (e.g., must be an epithet). As we showed, this kind of case also extends to referentially dependent pronouns that are not syntactically bound. The relevant examples, (16a–b), are repeated here.

Following Miltsakaki (2003), we assume that adjoined phrases, such as relative clauses, form a single attentional domain/discourse unit with the clause to which they are adjoined.¹³ If the d-ref realized by the embedded subject (whether pronoun, proper noun, or epithet) is identical to the C_p of the matrix domain, the example does not involve attentional shift, and clause-boundedness is predicted to be relaxed. However, this holds only if we relax the assumption that grammatical role is the determining factor for ranking in C_f(U_n). Take, say, (16a); here, Jack is contained inside the matrix subject one of the headhunters who interviewed Jack and is a direct object, whereas the matrix subject d-ref itself is also realized as the relative clause subject. If grammatical role alone determined ranking, we would expect the matrix subject to outrank Jack, and attentional shift would be predicted (incorrectly predicting clause-boundedness to be active). However, it is not a central claim of CT that grammatical role is the only factor determining ranking in C_f lists. Instead, it is possible that relevance broadly construed, lexical choices, world knowledge, and speaker intentions play a form-independent role. The topic at the time of utterance of (16a) may well most plausibly be Jack and his job prospects, favoring him as C_p.

We thank an anonymous reviewer for providing us with the following challenging example of a connected discourse where our predictions do not seem to be borne out at first glance, though we believe examples like this actually demonstrate how factors other than grammatical role can hinder or help discourse coherence.

There is a preference for construing the subject of (28c) as referring to the headhunter who interviewed Terry, realized as the subject in (28b) (bracketed), and not Terry himself—in other words, a strong tendency to analyze the transition between (28b) and (28c) as a case of attentional shift (accounting for the infelicity). Note that the subject in (28b) is the same as the subjects in the acceptable cases in (16), so that the incoherence in (28) is somewhat surprising if we are on the right track.

However, note that the judgment shifts with some manipulation of the lexical choices in (28c), where so instead of and then introduces the discourse segment. To our ears, this favors construing the subject pronoun with Terry while maintaining coherence.¹⁴

The switch from and then to so is pretty minimal, and we do not have much more to say about why this switch results in the effect demonstrated above.¹⁵ However, further manipulation of the lexical content of (28c) can more clearly favor one or the other construal. Given world knowledge about how headhunter/job-seeker interactions typically proceed, the most coherent reading of the following version of (28c) even more strongly favors construing the subject pronoun with Terry, since headhunters typically do not call a job-seeker to learn more about positions to which they themselves alerted the job-seeker.

Turning to our cases in (16), similar results can be achieved by manipulating lexical items as informed by world knowledge. Let us call the reviewer’s (28b–c) U₁ and U₂, respectively. The analogous segments in our sluicing antecedents in (16) are the main clause (corresponding to U₁) and the embedded clause (U₂) (following Miltsakaki 2003, Poesio, Stevenson, and Di Eugenio 2004). In (16a–b), the embedded clause in the antecedent favors construing the subject pronoun with the job-seeker, since headhunters do not report on their own job prospects in typical headhunter/job-seeker interactions.

We can also change the lexical content of U₂ in our sluicing antecedents to favor, instead, construing the embedded subject pronoun with the headhunter. Here, once again, world knowledge about typical headhunter/job-seeker interactions favors construing the embedded subject pronoun as coreferring with one of the headhunters.

Note that we correctly predict multiple sluicing (along with clause-boundedness suspension) to be possible whether the embedded subject pronoun is construed with one of the headhunters or with Jack (as in (16)). In (31), where the embedded subject pronoun is construed with the main clause subject, we have a run-of-the-mill bound pronoun case.

Note also that this appears to introduce some optionality as to whether the d-ref realized by the main clause subject or the d-ref realized by Jack gets to be the C_p in U₁. One fact supporting the idea that Jack (the job-seeker) is the C_p in the matrix discourse segment in (16a) is that the embedded subject, which is coreferential with Jack, is pronominal. As the only pronoun in the embedded segment, it must serve as the embedded discourse segment’s C_b.¹⁶ In CT, when the C_b of some utterance, U_n+1, is the C_p of U_n+1, there is a pressure toward attentional maintenance. The fact that (16a–b) are coherent supports the hypothesis that the transition between the matrix and embedded clauses is a case of attentional maintenance, which follows if Jack is the C_p of the matrix segment. The fact that (31) is coherent as well suggests, via the same reasoning, that the headhunter (the d-ref realized by the main clause subject) may instead, optionally, be the C_p of the matrix segment.

The upshot is that, beyond grammatical role, lexical choice and world knowledge can be equally important factors in determining discourse coherence and the nature of transitions between discourse segments. We do not endeavor to develop a full account of these facts here, as that would take us too far afield.

Thus far, we have explored the analysis of examples that demonstrate that bound subject pronouns are not necessary for the suspension of clause-boundedness. We now examine cases that demonstrate that bound subject pronouns are not even sufficient to defuse clause-boundedness. We will show that the behavior of such cases is as expected under the system developed above, once we incorporate an additional factor that influences the rankings in C_f lists.

Examples (33a–b) show that clause-boundedness is suspended only when the embedded subject pronoun is bound by the matrix subject. Examples (34a–b) show the opposite pattern, with binding by the object being necessary to defuse clause-boundedness.

Importantly, in both (33b) and (34a), the embedded subject is bound by an argument in the higher clause, yet clause-boundedness remains in effect.¹⁷

We suggest that the relevant generalization across all of these cases relates to the nature of the binder. When the pronoun is bound by the indefinite, the effects of clause-boundedness are obviated. But why should indefiniteness matter?

We take alternative-evoking expressions like indefinites to be sufficiently salient to factor into the ranking of arguments in the C_f list in the matrix discourse segment. That is, in the matrix clauses in (33)–(34), the indefinite is ranked highest regardless of its grammatical role or hierarchical prominence in the clause. Thus, when the embedded pronoun is bound by a DP other than the indefinite, attentional shift obtains, since the C_ps of the matrix and embedded clauses are not identical, leading to clause-boundedness.¹⁸

The examples in (33)–(34) show that bound pronominal subjects on their own are insufficient to defuse clause-boundedness. Such data, it turns out, are challenging for Grano and Lasnik’s (2018) analysis of the bound pronoun effect, which in fact leads to predictions opposite from what is found.

Recall from section 2 that Grano and Lasnik take clause-boundedness to be phase-boundedness. Thus, by the Phase Impenetrability Condition (PIC; see Chomsky 2001), movement across a finite CP boundary that does not pass through the CP’s edge will be ill-formed; under Grano and Lasnik’s assumption, this includes cases of movement subject to clause-boundedness. Grano and Lasnik further assume that bound pronouns are introduced into the derivation with unvalued ϕ-features and that because of their presence, a phase containing them remains “porous” to derivational operations.

Let us return now to the data in (33)–(34). As already noted, cases like (33a) are handled straightforwardly under these assumptions: the ϕ-features of the subject pronoun are not valued until the matrix subject is merged. Assuming with Grano and Lasnik that v does not count as a phase head, there will be no phase boundary between the base position of the second wh-phrase and its ultimate landing site in the higher CP, resulting in the well-formedness of the example. What is puzzling on this view is why an example like (34a) should behave differently. Here, the embedded pronominal subject is bound by the matrix subject, as before, which should lead to the voiding of the embedded phase, yet the example is unacceptable.

Next, consider the contrast in acceptability between (33b) and (34b). In both, the embedded subject is bound, which should suspend phasehood for the embedded CP because of the presence of unvalued ϕ-features. In each case, the binder is not the matrix subject, but a matrix (indirect) object. Assuming that v is not a relevant phase head and that the pronoun’s features are valued within the VP when the object is merged, the matrix CP will be the first active phase head encountered. Therefore, Wh2 should be able to reach the specifier of the matrix CP in one fell swoop without running afoul of the PIC in either case.¹⁹

To summarize, in this section we have built on results from section 3, where we established that bound pronouns are not necessary and indeed are insufficient for suspending clause-boundedness. We sharpened our generalization in (7) by appealing to notions from CT, where clause-boundedness is active whenever the transition between an embedded clause and the main clause counts as attentional shift. This characterization of the facts makes the right empirical cut and unifies bound pronoun cases with the other cases introduced here.

As we have discussed at length, our account ties a syntactic effect—clause-boundedness and its exceptions—to notions of discourse coherence that are connected with the flow of attention across utterances. Yet Grano and Lasnik (2018) advance some arguments against this very notion, namely, that the bound pronoun effect might have something to do with salience. Their general idea is that when the reference or construal of the embedded subject pronoun is easier to establish, perhaps because its referent is more salient at the time of utterance, the embedded clause boundary is somehow more permeable and establishing long-distance dependencies across it is less offensive. They report on results from a judgment study that involved embedded clauses with unbound first and second person pronoun subjects. Arguably, first and second person pronoun referents are salient in any discourse, and such test items were rejected just the same as clause-boundedness violations with unbound pronoun subjects more generally (see (35)). One might conclude from this result that salience and the flow of attention in discourse are irrelevant in accounting for the bound pronoun effect and that a purely syntactic account is on the right track.

However, our proposal evades this criticism through its implementation in CT. Our account makes crucial reference not just to salience and attention, but also to constraints on coherent discourse sensitive to information packaging across discourse segments. In short, examples like (35) count as attentional shift transitions under our assumptions, just like examples not subject to the bound pronoun effect, and so our account correctly predicts clause-boundedness to be active in such cases.²⁰ In other words, clause-boundedness (and its exceptions) are best understood as an interface phenomenon, where pragmatic constraints on the flow of attention across utterances in discourse have a syntactic consequence. In what follows, we suggest theoretical implementations of this idea.

The view we have endorsed thus far envisions a restriction on the syntactic locality of certain grammatical processes that is sensitive to the discourse transition between one clausal domain and another. If this view is correct, the architecture of grammar must incorporate some way in which the syntax can show such sensitivity to discourse notions.

Though we are not committed to any particular syntactic implementation of such an interaction, in this section we sketch two proposals that achieve this. The first is a modification of Grano and Lasnik’s (2018) approach, which preserves many of their core architectural assumptions but makes phasehood contingent on discourse properties. The second, couched in Murphy’s (2017) Harmonic Minimalism framework, jettisons syntactic modularity, instead allowing discourse properties to interact directly with syntactic derivations. The clause-boundedness of multiple sluicing across a shifty subject turns out to be an instance of cumulativity, where multiple wh-movement in sluicing and movement across a shifty subject are allowed individually but not in combination.

A simple and conservative way to capture the clause-boundedness paradigm as regulated by attentional maintenance is to incorporate our discourse notions directly into the determination of the phasehood of candidate phase heads in Grano and Lasnik’s (2018) system, abandoning reference to bound pronouns or ϕ-features. We provide such an implementation here as a sketch of how further work in this domain might proceed.

First, let us assume with Grano and Lasnik that clause-boundedness is phase-boundedness, along with a dynamic view of phases where candidate phase heads sometimes count as phase heads and other times do not. Following Lasnik (2014), we also assume that Wh2-movement in multiple sluicing proceeds in one fell swoop. Finally, we import Grano and Lasnik’s assumption that unvalued features on the head of the complement of a phase head, X, render X a nonphase (allowing for exceptions to clause-boundedness). To capture subject orientation in the suspension of clause-boundedness, we adopt Grano and Lasnik’s assumption that it is the valued/unvalued status of a candidate phase head that determines its phasal/nonphasal status. The innovation/ modification we suggest lies in the nature of the features that phasehood is sensitive to. Instead of being sensitive to just any unvalued features, phasehood is specifically sensitive to attentional features.

One question that immediately comes to mind concerns the locus of these attentional features. We have just suggested that it is phase heads that bear attentional features. An anonymous reviewer rightly asks what independent evidence there might be for the existence of such attentional features. We are not aware of any work that explicitly makes reference to attentional features per se, though we note that there is much work in the syntax-pragmatics interface literature that makes reference to the notion of topichood, and we suggest here an intimate connection between topichood and shiftiness.

Frascarelli and Hinterhölzl (2007) identify three kinds of topic constituents in Italian and German—shifting topics, contrastive topics, and familiar topics—with distinct syntactic and prosodic properties. In Italian, shifting topics are constituents that introduce a novel topic in a discourse (which, information-structurally, evokes our notion of shifty subjects) and always precede other kinds of topics when occurring with them in the same clause. Frascarelli and Hinterhölzl posit a left-peripheral ShiftP projection associated with shifting topics.²¹

Let us assume that this left-peripheral ShiftP projection contains a phase head and that Shift⁰ is the element with attentional features. When the lower clause includes a DP subject whose associated d-ref constitutes a new topic (i.e., when we have a case of attentional shift between the higher and lower clauses), this DP will value the attentional features on the embedded Shift⁰. As a result, ShiftP’s phasehood will remain intact, giving rise to clause-boundedness. In the absence of a shifty subject in the downstairs clause, the attentional features on the embedded Shift⁰ head will remain unvalued until the occurrence of the DP that constitutes the aboutness topic. As a result, the embedded clause will be porous to one-fell-swoop movement, yielding the possibility of long-distance multiple sluicing.

An anonymous reviewer argues that Grano and Lasnik’s approach predicts our result, assuming attentional features exist. That is, if phasehood is suspended whenever any features remain unvalued, then phasehood would be suspended were attentional features to remain unvalued. We would emphasize that things cannot be that general, and that phasehood (and its suspension) must be sensitive to attentional features in particular. If any unvalued feature may suspend phasehood, then we do not expect the puzzling cases discussed in section 4.3 to have the empirical pattern they do. In Grano and Lasnik’s proposal, unvalued ϕ-features associated with bound pronominal subjects in embedded clauses are sufficient to suspend phasehood of the embedded CP. This would make it mysterious why (33b), repeated here, is ungrammatical despite having an embedded subject bound by the matrix object.

Under Grano and Lasnik’s approach, she_j has unvalued ϕ-features, which keep the embedded CP phase open. Under our approach, indefiniteness plays a role in determining the C_p of the main clause, and (33b) constitutes a case of attentional shift, so that the embedded clause counts as a phase, correctly predicting ungrammaticality.

Additionally, example (34a), repeated here, arguably should be grammatical in Grano and Lasnik’s approach, since the embedded subject is bound by the matrix subject (therefore lacking valued ϕ-features until the matrix subject is merged). Under our approach, the indefinite direct object of the matrix clause is the center of attention in the matrix clause, and this example counts as a case of attentional shift, correctly predicting its ungrammaticality.

In other words, it cannot be the case that any unvalued feature manages to suspend phasehood. Instead, phasehood and its suspension must be restricted to attentional features.

To further support this point, we highlight an additional empirical pattern that is unexpected under Grano and Lasnik’s approach, focused on ϕ-features, but expected under our proposal, focused on attentional features. Grano and Lasnik discuss the relevance of the bound pronoun effect for island phenomena, showing that bound pronouns in islands yield more acceptable island violations than island violations without bound pronouns.

Our proposal also captures this pattern in the same way. Example (37c) constitutes a case of attentional shift. Attentional shift is avoided in (37b).

On the other hand, Grano and Lasnik’s proposal predicts that we should not see the same cline in acceptability in the following examples:

In both cases, the most deeply embedded clausal island, the after-clause, has a bound pronoun subject, which will not be valued until its binder is encountered in the main clause. This ensures that any higher potential islands, such as the because-clauses in these examples, will remain “open” as well, regardless of their own subjects, since they also contain within them the unvalued ϕ-features of the most embedded island. Therefore, there should be no difference in acceptability between these examples, contrary to fact. In contrast, under our approach the hierarchically intermediate because-clause’s subject (Bill) in (38b) induces attentional shift and phasehood, predicting unacceptability relative to (38a), where the subject of the because-clause is a bound pronoun, which does not shift attention.

Even accepting the empirical arguments for the distinctive role played by attentional features in establishing phase boundaries, one might ask why these features play this special role. We believe that one potential response relates to the original computational motivations underlying CT. Recall that the existence of phases has also been motivated by the need to reduce complexity in grammatical derivation. Typically, such complexity is taken to derive from the desire to reduce search during grammatical derivation. In the centering context, Joshi and Kuhn (1979) and Joshi and Weinstein (1981) suggest that the avoidance of shifts between centered entities in discourse is similarly motivated by the need to reduce the complexity—in this case, complexity of inference. Structuring utterances around a center effectively reduces the underlying logical predicates to a monadic structure. This restriction yields a logical system that allows more efficient inference as compared to unrestricted predicate logic. Though monadic logic is not sufficient to represent linguistic meaning in general, the idea that sections of discourse retain an attentional center will lead to a reduction in inferential complexity. If this is on the right track, under our view phases are indeed motivated by computational considerations, not of narrow grammatical computation but of inference.

Our second suggested way of syntacticizing our discourse-based understanding of clause-boundedness builds on Murphy’s (2017) framework of Harmonic Minimalism. Here, the well-formedness of syntactic representations is determined by the outcome of a Harmonic Grammar optimization. As Murphy argues at length, this framework provides a natural analysis of cumulativity phenomena: cases in which two syntactic processes A and B are permitted individually but cannot occur in combination. He documents a wide range of phenomena that fall under the cumulativity rubric. One example is found in Serbo-Croatian, Russian, Polish, Romanian, Bulgarian, and Czech, among other languages, which permit the fronting of multiple wh-phrases in a wh-question. The Serbo-Croatian example in (39) illustrates.

Some of these languages, including Serbo-Croatian, Russian, Polish and Czech, are exceptional in another way: they allow left-branch extractions, that is, the movement of a prehead modifier or possessor out of a DP. We illustrate with Serbo-Croatian once again.

Interestingly, however, even in languages that allow multiple wh-fronting and left-branch extraction to occur separately, they are not permitted to occur together. That is, still exemplifying with Serbo-Croatian, multiple left-branch extraction is not possible.

As is well-known, multiple wh-fronting and left-branch extraction are crosslinguistically marked; many languages do not permit either of them to occur, even by themselves. As a result, from the perspective of Optimality Theory (OT), there should be a constraint that is violated by each of these processes. Murphy adopts (42a–b).

Murphy follows other work in OT syntax in assuming that wh-movement is motivated by a constraint that is violated in the absence of movement.

In a traditional OT setting, if MULTSPEC is ranked above WHCRIT, the result will be a language that prohibits the fronting of multiple wh-phrases, while the reverse order will yield multiple wh-fronting. Similarly, if *LBE is ranked above WHCRIT, wh-movement out of a left branch will be blocked, and the reverse ordering will allow it. However, no strict ranking will suffice to derive the cumulativity pattern, where each pattern is allowed separately but they are not permitted together. Instead, to derive this, Murphy argues that each constraint should be assigned a numerical weight. To evaluate the relative optimality of each candidate, the violation weight of each constraint is multiplied by −1 times the number of violations, and the weighted violations are then summed to compute the candidate’s harmony, indicated by . The optimal candidate is the one with the highest (least negative) harmony. This is illustrated in tableaux (44) and (45).

In each case, because the weight associated with WHCRIT is greater than the weights associated with MULTSPEC and *LBE, violations of either of the latter constraints are preferred to violations of the former. As a result, either multiple wh-movement or left-branch extraction is preferred to lack of movement. The outcome for an input that includes multiple wh-phrases within left branches will be different: because of the cumulative cost, lack of movement is preferred to multiple left-branch movement.²³

Returning now to the analysis of sluicing, we will adopt from past work the following constraints that restrict the deletion of material:

The first of these constraints, proposed by Merchant (2008), favors the deletion of material that is marked as informationally backgrounded. The second constraint, first posited by Baković and Keer (2001), seeks to prevent the deletion of wh-elements. To these constraints, we add the following one, which penalizes movement out of a domain that includes a new focus of attention—in other words, that includes a shifty subject:

We recognize that SINGLEATT simply stipulates the ungrammaticality of extracting across a shifty subject. While we are not averse to purely formal constraints on syntactic derivations, in this case we believe SINGLEATT is the grammatical reflection of the aforementioned considerations of inferential efficiency that have motivated much of the work in CT.

With these constraints in place, we can evaluate an instance of sluicing involving a single wh-phrase that needs to move across a shifty subject to satisfy the requirements of the relevant C head.

Note that the correct output, involving non-clause-bounded movement, will be chosen so long as the weight of SINGLEATT is less than the sum of the weights of FAITH(WH) and WHCRIT.

Note also that if the subject of the lower clause were not shifty, the movement candidate would not violate any of the constraints we are dealing with and would therefore also have higher harmony than a structure in which the wh-phrase does not move. To allow such shifty-subject-crossing movement to take place in the absence of sluicing, so that a question like Who did A say B talked to? can be generated, a stronger condition is required on the weight of SINGLEATT.

The fact that this condition is stronger than what is necessary to allow phase-boundary-crossing movement to take place under ellipsis makes the prediction that there should exist languages that permit movement of a single wh-phrase across a shifty subject in sluicing contexts, but not in nonsluicing contexts. At present, we are unaware of any languages that show this pattern.

Let us now consider inputs that are both e-marked and contain multiple wh-phrases—that is, candidates for multiple sluicing. We will start with a monoclausal structure, so as to avoid issues of locality.

Here, we consider only candidates in which the e-marked TP is deleted, and we contrast candidates in which either zero, one, or two of the wh-phrases are fronted. The first candidate, in which no movement takes place, is harmonically bounded by the second (it suffers all of its violations and more) and hence will never be selected as optimal regardless of the constraint weights. For the third multiple sluicing candidate to be correctly chosen, the following condition on MULTSPEC must be respected:

In other words, under the grammar of English, the movement of multiple wh-phrases, though violating MULTSPEC, is less grammatically marked than failing to move and deleting one of the wh-phrases. Of course, if nothing is inducing the deletion of a wh-phrase that remains in situ, because of the absence of e-marking, the movement of a single wh-phrase will be grammatically preferred under the assumption that (54) obtains.

Let us turn finally to the evaluation of candidates involving multiple long-distance sluicing. Assuming that the embedded subject constitutes a new center of attention, movement out of the embedded clause will yield a violation of SINGLEATT.

As before, multiple wh-fronting will induce a violation of MULTSPEC. This time, because the locality constraint SINGLEATT is also violated, the example is an instance of the cumulativity pattern: either multiple wh-fronting or movement across a new attentional center may take place, in the right conditions, but they may not occur together. Such cumulativity requires that constraint weights abide by the following inequality:

A large number of solutions (indeed infinitely many) satisfy the inequalities over constraint weights, among which is the following:

Our discussion has focused entirely on multiple sluicing, and we have developed our analyses to account for the patterns found in that empirical domain. However, clause-boundedness and well-defined exceptions to it exist beyond multiple sluicing, as Grano and Lasnik (2018) argue extensively for the bound pronoun effect. Specifically, they argue that the same pattern holds across tough-movement, comparative deletion, QR, antecedent-contained deletion, gapping, pseudogapping, multiple questions, reciprocal binding, and family-of-questions readings. This immediately raises an important empirical question: does the discourse-related generalization that we have argued governs exceptions to clause-boundedness in the multiple sluicing case also apply in these other empirical domains? A positive answer would have an important theoretical consequence: the analysis of the multiple sluicing case should extend to the others as well. On the other hand, a negative answer would favor an analysis that treats multiple sluicing in a distinctive manner.

Interestingly, the analyses outlined in sections 5.1 and 5.2 are distinguished by the extent to which they generalize in this way. The dynamic phasehood approach extends to the full range of phenomena exactly as Grano and Lasnik propose, as it constitutes a minimal modification of their proposal. A CP containing an unvalued Shift⁰ head will fail to constitute a locality domain for all phase-sensitive operations or dependencies, and thus will be porous to any such operation or dependency. The Harmonic Minimalism OT analysis, in contrast, does not generalize in this straightforward fashion. The success of this analysis depends crucially on the cumulative weighting of two constraints, SINGLEATT (which penalizes movement across a clausal domain in the presence of attentional shift) and MULTSPEC (which penalizes multiple specifiers), as compared to the weights of constraints that induce movement of wh-phrases. While SINGLEATT would be relevant for any multiclausal structure and therefore to all potential cases of clause-boundedness, the other constraints will not necessarily apply. Most conspicuously, MULTSPEC plays a role only in structures containing multiple elements that are moved to the specifier of a single phrase, and it seems unlikely that this structural configuration obtains in all of the cases that Grano and Lasnik explore. Consequently, unless something specific is assumed about the linking of constraint weights, it should be possible to have a grammar that allows attentional-shift exceptions to clause-boundedness for some cases but not all.

To decide which of these approaches is correct, we must first resolve an empirical question: to what degree is attentional shift relevant to the broader landscape of clause-boundedness? In other work, we have begun to explore the relevance of attentional shift to the entire set of phenomena discussed by Grano and Lasnik, and we believe that the same pattern holds by and large for the other construction types they examine (except for reciprocal binding, which we believe should receive an independent account). If this is correct, it would favor the dynamic phasehood approach. For now, however, we leave this issue open.

Before concluding, we want to address Huang’s (to appear) argument that attentional shift is not the relevant factor governing clause-boundedness in comparative deletion. Huang conducted an acceptability study involving examples like the following:

Huang reports that examples with a bound pronoun subject like (58b) are judged as highly as the infinitival baseline in (58a). In contrast, the participants in his experiment judged the examples with expletives, quantifiers, and unbound pronouns (58c–e) to be considerably worse. On the surface, this constitutes substantial evidence against the generality of the pattern we have reported here. However, we believe there is reason to be skeptical of this conclusion. First of all, the canonical shifty-subject example in (58f) was judged to be better than all of our nonshifty-subject examples, which we find surprising. Indeed, we had constructed examples of comparative deletion that we judge to be as good as the bound pronoun case.

What explains the difference between these cases and Huang’s results, then? We believe the issue lies in confounds present in each of Huang’s examples, which we suspect have a negative impact on their acceptability, coupled with the way Huang assesses the modulating effect of lack of attentional shift. In the expletive subject case (58c), the comparative is attached to a DP that is nonfinal in the clause (unlike all of Huang’s other examples). For reasons that are not clear to us, finality has a significant impact on the acceptability of this construction.

When the comparative is attached to the final DP in an embedded clause with an expletive subject, the sentence improves considerably.

For the cases of quantifier subjects (58d), the confound relates to the impact of the presence of a negative quantifier on acceptability judgments. Because of the semantic complexity such quantifiers add, we expect that the following pair of sentences would yield an acceptability contrast in a judgment survey:

As a result, identifying the degree to which the presence of the quantifier subject modulates the clause-boundedness effect would require us to first control for the impact of the presence of the quantifier, and to determine whether the reduced acceptability in cases like (58d) is lower than this. The case of unbound pronouns in (58e) is similar: here, the confounding factor is the presence of the subject-modifying relative clause. Such complexity in a subject yields a center-embedded structure that induces a higher processing load, something we would expect to yield lower judgments. Again, in order to understand whether the unbound pronoun has any modulating effect, we need to compare such cases with examples that include a relative clause, but where the pronoun is bound by the subject.

Our judgment is that these examples are comparable in acceptability. In any case, it is only by comparing structurally comparable examples that we can assess the impact of the pronoun binding itself on clause-boundedness.

In this article, we have examined clause-boundedness and its exceptions in multiple sluicing. We have uncovered data that point to the inadequacy of past work on this topic that ties exceptions to clause-boundedness to the presence of a bound pronoun in the embedded clause. Our work instead introduces a new generalization, stated in (7), according to which clause-boundedness arises from constraints on forming dependencies across discourse transitions involving what we call “shifty subjects.” We have provided a characterization of this pattern in (21), building on ideas from Centering Theory, and we have proposed two possible syntactic implementations.

We take an implication of the pattern we have uncovered to be that syntactic locality must be sensitive to discourse factors. We have implemented such sensitivity by explicitly representing discourse features in syntax. However, as an anonymous reviewer observes, there is a reductionist alternative. Specifically, we might say that, in sluicing contexts, attentional shift across discourse segments results in degraded acceptability, obviating the need to invoke syntactic well-formedness. Such a discourse-based perspective on syntactic locality is in fact not novel, having been advocated by Erteschik-Shir (1973), Kuno (1987), Van Valin (1995), Goldberg (2013), and Abeillé et al. (2020), among others. Though these authors’ specific proposals vary, they share a central idea: islandhood derives from a conflict between the information status of a displaced element (i.e., it is focused) and that of the constituent out of which it is extracted (i.e., it is topical). For example, subjects, which are typically topics, constitute islands to extraction for question formation, which involves a focused wh-phrase, under the assumption that it is infelicitous to have a focus as part of a topic. While this approach is similar in some respects to the perspective we take here, it differs in important details. We do not focus on the information status of a complement clause; rather, we focus on that of the complement clause’s subject. Moreover, it is not obvious how the presence of a shifty as opposed to nonshifty subject would affect the information status of the entire clause containing it. All of the cases we have discussed here involve extraction from a complement clause, which these authors assume to be a topical constituent. A shifty subject, then, would not alter the information status of the clause containing it. In contrast, a nonshifty subject, though focal, should also not affect the information status of its clause (see Abeillé et al. 2020: 3n5). Even if there were some way of extending these previous discourse-based conceptions to the current context, it is not clear why the kind of clause-bounded locality described here should arise only in very specific contexts such as multiple sluicing.²⁴

Consequently, though we believe the ultimate motivation for clause-boundedness does indeed derive from properties of information structure and their impact on the efficiency of inference, we think that such an impact must arise through its grammaticalized structural reflection. We have provided two ways to model this grammaticalization. One involves the incorporation into syntactic representations of features that encode discourse properties. This has been a theme of much recent work (Frascarelli 2007, Frascarelli and Hinterhölzl 2007, Haegeman 2012), and we view our findings as providing potential support for this development. An alternative approach to the syntax-discourse interaction jettisons syntactic modularity and instead allows discourse properties to interact directly with syntactic derivations.

For insightful feedback and comments, we would like to thank Klaus Abels, Byron Ahn, Lisa Cheng, Veneeta Dayal, Michael Yoshitaka Erlewine, James Griffiths, Guliz Gunes, Larry Horn, Hadas Kotek, Dave Kush, Anikó Lipták, Jason Merchant, Gereon Müller, Andy Murphy, Dennis Ott, David Pesetsky, Catherine Rudin, Anna Szabolcsi, and Adam Szczegielniak, the members of the Yale Syntax reading group and Grammatical Diversity Project research team, and our anonymous reviewers. We are also grateful to conference audiences at the 2017 LSA Annual Meeting, the 40th GLOW Colloquium, and the workshop Multiple Questions about Sluicing at Yale University, and to colloquium audiences at the Department of Linguistics at Southern Illinois University, the NYU Syntax Brownbag, and the Institut für Linguistik at Universität Leipzig.

RF would like to dedicate this article to the memory of Aravind Joshi, with whom he began work on a planned dissertation on Centering Theory in 1987. RF got sidetracked by other topics during grad school, but is extremely pleased to return at long last to Centering Theory and to explore its significance for syntactic locality, another of his long-standing interests.

The authors contributed equally to this work and are listed alphabetically. All mistakes are, of course, our own.