We reply to Erlewine and Kotek’s (2018) claim that the phenomenon of covariation under focus (Tanglewood sentences; Kratzer 1991) is subject to syntactic islands and that it should therefore be handled by a focus movement theory (contra Kratzer’s view). We present novel data that are at odds with Erlewine and Kotek’s conclusions and demonstrate the necessity of an island-insensitive mechanism to capture focus covariation. We revisit Erlewine and Kotek’s main arguments against such a system and show that they are systematically confounded. Moreover, removing the confounds cancels the force of the arguments, corroborating the central point of this article.
Kratzer (1991) observes that a focus-marked XPF can covary with a (silent) copy of that XP in an elided VP, as exemplified in (1). On the relevant reading, the focused phrase Tanglewood covaries with another silent phrase in the ellipsis site. We call examples like (1) Tanglewood sentences, and the prominent reading they give rise to the covariation reading.
I only went to TANGLEWOOD because you did.
≈ Tanglewood is the only place x such that I went to x because you went to x
What mechanism derives the covariation reading? On the face of it, (1) might be taken to involve variable binding of the familiar sort, where Tanglewood undergoes covert movement and binds into the ellipsis site.1
As Kratzer argues, however, a movement-based analysis is not general enough; covariation readings also arise in cases like (3), where the relevant movement step would have to take place out of a relative clause island.
I only talked to the person who chairs THE ZONING BOARD before you did.
≈ The Zoning Board is the only x such that I talked to the person who chairs x before you talked to the person who chairs x
Kratzer observes, then, that a movement analysis of Tanglewood sentences suffers from the same shortcoming that originally motivated in-situ theories of association with focus more generally: it entails the existence of island-insensitive movement.
To account for these data without committing to the existence of island-insensitive movement, Kratzer proposes a special binding mechanism that is limited to focus constructions. Her theory augments the standard Roothian theory of association with focus (1985, 1992) by adding focus indices to all focus-marked phrases. Covariation arises in this system when two focus-marked phrases bear the same focus index (see Kratzer 1991 for the formal details). Tanglewood sentences thus involve focus coindexation between the (overt) focused phrase and a (silent) copy in the ellipsis site. For example, (3) is assigned the LF structure in (4).
I only [VP talked to the person who chairs the [Zoning Board]F1] before you did
[VP talk to the person who chairs [the Zoning Board]F1]
(based on Kratzer 1991:833)
Since coindexation is not subject to structural constraints, this system can capture the island insensitivity of covariation under focus.
Erlewine and Kotek (2018) (henceforth E&K) challenge Kratzer’s conclusions. They claim that the relevant covariation readings should be captured with a focus movement theory along the lines of (2). The apparent island insensitivity of the covariation readings, they contend, arises because of the previously overlooked possibility of covertly pied-piping a phrase bigger than the focused phrase itself. Specifically, E&K argue that whenever covariation readings appear to disobey island constraints, it is because the whole island pied-pipes along with the focused phrase and binds in the usual way into the ellipsis site.
E&K’s main line of argumentation is that once the option of covert pied-piping is considered, Tanglewood sentences do show island sensitivity. Kratzer’s focus coindexation system is, by design, not island-sensitive and thus overgenerates unattested readings. If E&K’s argument goes through, it suggests that the focus coindexation mechanism should be eliminated from the grammar.
In this article, we argue that E&K’s claims are ultimately unwarranted and that an island-insensitive mechanism, such as Kratzer’s focus coindexation or something like it, is necessary to capture the full array of covariation readings. Our response has two main components. First, we show that E&K’s main arguments are systematically confounded and that once these confounds are removed, the arguments no longer go through. Second, we present novel and old data that demonstrate the necessity of an island-insensitive mechanism.2
Throughout this article, we will use Kratzer’s (1991) system to illustrate an island-insensitive approach to focus covariation, but it’s worth noting that it’s not the only such account. Another theory consistent with our conclusions is that of Sauerland (2007), who captures covariation not with coindexation but by way of (island-insensitive) multidominant structures. We leave open the question of whether there are empirical arguments that can distinguish Kratzer’s mechanism from Sauerland’s.
The structure of the article is as follows. In section 2, we outline E&K’s island-sensitive focus movement theory of covariation under ellipsis and present their main arguments for it. In section 3, we discuss the problems in E&K’s argumentation, demonstrating with new data that focus covariation is island-insensitive (once the confounding factors are removed from E&K’s examples). In section 4, we address the case of focus covariation between two overt elements. Contra E&K, and following Sauerland (2007), we show that this type of data also can—and should—be captured by an island-insensitive system like Kratzer’s.
2 E&K’s Theory and Argumentation
2.1 E&K’s Alternative Theory
To begin, we briefly review E&K’s theory of association with focus and Tanglewood constructions. Following proposals by Drubig (1994), Krifka (2006), Wagner (2006), and others, E&K embrace a hybrid movement/in-situ theory of association with focus: the focus-marked XPF or a phrase containing it moves to the restrictor position of only, and focus alternatives are computed by applying the in-situ theory of focus projection to this moved phrase. E&K assume the following denotation for only, where α is the meaning of the covertly moved phrase and β is a λ-term that results from abstraction over the position of the trace (2018:447, (12)):
The free variable C in (5) is resolved to the set of focus alternatives of the restrictor of only. In the simplex case where just the focus-marked phrase moves to only, the resolution is straightforward. In the more complex case where the restrictor of only is a constituent containing the focus-marked phrase, the standard Roothian mechanics for compositionally computing alternatives is employed. Importantly, there are no indexed foci in this system.
Below, we illustrate how the theory works for a case that involves association with focus within an island, as in (6a). Since the hybrid approach allows more than just the focus-marked XPF to undergo movement to only, if XPF is contained within an island, the entire island is pied-piped to only (6b). The variable C in this case is then resolved to the focus alternatives of the island, (6c). The resulting interpretation is shown in (6d).
E&K can then capture covariation from within islands by pied-piping the island to only and having the resulting λ-abstraction bind into the ellipsis site.
The upshot is that E&K’s theory overcomes Kratzer’s primary criticism against focus-movement-based accounts, and derives the facts without indexed foci. The core of their article is devoted to showing that the focus coindexation theory overgenerates in ways that the focus movement theory does not. We now turn our attention to these arguments.
2.2 E&K’s Arguments from Covariation in Island Contexts
E&K’s main argument against Kratzer’s (1991) approach is that once the possibility of covert pied-piping is recognized, covariation under focus is sensitive to islands, as expected under a focus movement but not a focus coindexation account. To make this point, E&K present data from VP-ellipsis in three distinct island configurations, which we now review. These are the arguments that we will show to be confounded (see section 3), demonstrating that with more controlled examples, the overall facts bear out Kratzer’s predictions and not E&K’s.
2.2.1 Relative Clause Islands
(8) We only hired a nanny who speaks SPANISH because our son does.
≈ #A nanny who speaks Spanish is the only (person) x such that we hired x because our son hires x3
E&K claim that this pattern of judgments is predicted by their system but not by the focus coindexation system. On E&K’s account, the unattested reading in (8a) requires movement of Spanish to only, followed by binding into the ellipsis site. However, this movement is blocked because relative clauses are islands (see (9a)). Pied-piping the whole island to only is permitted (see (9b)), but it can only result in the implausible reading in (8b); as shown in (9c), attempting to pied-pipe the island and bind a variable in the object of speaks in the ellipsis site results in a nonsensical meaning.4
Under Kratzer’s approach, on the other hand, nothing blocks the LF structure in (10), where the focus-marked XP is coindexed with a copy in the ellipsis site. E&K thus claim that Kratzer’s theory incorrectly predicts the missing reading to be possible here.
The argument, then, is that covariation under focus is sensitive to islands (once we admit covert pied-piping into our theory), as predicted by a movement analysis.
2.2.2 Conditional Islands
E&K’s second example, (11a) (their (27)), is based on adjunct islands: an only taking sentential scope associates with a focus-marked XPF in an if-clause, which is an island. The context in (11) makes the covariation reading in (11b) natural, but such a reading is nonetheless unavailable.
Here too the unavailability of the covariation readings is predicted by E&K’s approach. Since the focus-marked StevensF is contained in an if-clause adjunct, it is not possible to covertly move just that phrase followed by binding into the ellipsis site. Pied-piping the whole if-clause is an option, but that would again not be helpful: for the relevant reading to arise, the elided VP needs to contain a variable bound by StevensF, not by if they invite StevensF.
In contrast, Kratzer’s theory purportedly overgenerates the missing reading. Since coindexation is not sensitive to islands, the focus index on StevensF can be duplicated in the ellipsis site.
2.2.3 Coordinate Structure Islands
E&K’s third example is based on the Coordinate Structure Constraint. The paradigm here features a matrix only that associates with a focus-marked XPF in the first conjunct of a coordinate structure. E&K observe that in (12) (their (28)), which exemplifies this construction, covariation of XPF with a phrase in the second conjunct is impossible.
The claim is once again that the impossibility of covariation supports E&K’s theory over Kratzer’s. Covert movement of the focused phrase [the Elm St. house]F to only is blocked by the Coordinate Structure Constraint, so it cannot bind into the ellipsis site. Importantly, E&K also reject an across-the-board movement of the focused phrase as in (13), which would potentially derive the covariation reading (see E&K 2018:453n13). They justify this on the grounds that covert movement is purportedly never across-the-board, as argued by Bošković and Franks (2000).
Under the above assumptions, then, E&K predict covariation to be impossible in coordinate structures.5 Kratzer’s theory wrongly predicts covariation to be possible using coindexation.
3 Our Criticism: Focus Movement Is Not Sufficient
In this section, we show that E&K’s conclusions are ultimately unwarranted. Our response divides E&K’s three island examples above into two classes: the relative clause and conditional islands (see sections 2.2.1–2.2.2) and the coordinate structure islands (see section 2.2.3). With respect to the first class, we argue that the key examples in (8) and (11) are confounded in a systematic way, and that the absence of the covariation reading is expected on independent grounds. Importantly, as we show with novel and more controlled test cases, once the confound is avoided the empirical picture turns against the focus movement theory. With respect to the argument from coordinate structure islands, we develop a somewhat different criticism: while we accept E&K’s basic data point in (12), we will argue—on the basis of a hitherto unnoticed observation about the interaction between VP-ellipsis and extraction from coordinate structures—that it does not in fact distinguish between the predictions of their theory and Kratzer’s.
3.1 Reevaluating the Arguments from Relative Clause and Conditional Islands
We begin with the relative clause and conditional islands. As mentioned above, our proposal is that the key examples in (8) and (11) are confounded. The confound is this: the intended ellipsis resolution is very difficult in such examples even when association with focus is removed from them. This is shown in (14) and (15), the baselines for (8) and (11); in these examples (and structurally parallel ones) that lack association with only, the intended ellipsis is almost inaccessible.
??We hired a nanny [who speaks Spanish] because our son does
?*I met the author [who wrote that book] before she did
write that book.6
*I told them they would regret it if they invite Stevens before they did.
*I suspected they would regret it if they invite Sally before they did.
Our basic criticism, then, is that the unacceptability of (8) and (11) cannot be taken as evidence for claims about the nature of association with focus, because the baseline of these examples is already marginal to impossible. This criticism does not depend on what makes (14) and (15) unacceptable, which is an interesting question in its own right. But if we want to create better baseline examples to test E&K’s predictions, it is useful to investigate the structural properties that cause the unacceptability of such examples. To this end, Hardt and Romero (2004) have studied similar constructions and proposed a theory of ellipsis that actually predicts (14) and (15) to be unacceptable. In a nutshell, Hardt and Romero argue that VP-ellipsis resolution is sensitive to the way discourse is structured, and that only those VPs that stand in an appropriate discourse-structural configuration can antecede elided VPs. In slightly more detail, their proposal is given in (16), of which (16b) is the crucial part.7
We now show how (16) rules out (14) and (15). Beginning with (14), Hardt and Romero (2004:380) argue that discourse particles like because and before are parsed into a discourse tree as in (17), and that quantificational determiners are parsed with restriction and nuclear scope as sisters in the discourse tree, as in (18).
In order for [VPEspeaks Spanish] to be elided, the Matching Condition in (16a) requires that this VP be dominated by a clause19) is to have F speaks Spanish and 16b) will not be satisfied, because 14).
Turning now to (15), repeated in (20a), the impossibility of ellipsis here is again predicted by Hardt and Romero’s Discourse Condition, assuming as they do that both before and if are parsed as distinct nodes in the discourse tree. As is clear from (20b), the VP invite Stevens in the antecedent of the conditional does not c-command the VP in the before-clause, so that ellipsis is correctly ruled out. We abstract away here from the discourse contribution of tell, which if anything would result in a further node dominating IF and hence a more deeply embedded antecedent VP.
At least on this theory of ellipsis resolution, then, the baselines to E&K’s crucial examples from relative clauses and if-clauses—as well as the examples themselves—are predicted to be independently unacceptable, in conformity with our judgments.8
If these islands are to serve as an argument in favor of focus movement and against a Kratzerian in-situ mechanism for covariation, the examples must therefore be modified to facilitate ellipsis in the basic case. We now proceed to show that once this is properly controlled for, the data support Kratzer’s theory.
Beginning with relative clauses, one way to improve the baseline ellipsis is to place the elided VP in the scope of the DP that contains the relative clause. The ellipsis is acceptable in such cases, (21a–b), as predicted on Hardt and Romero’s theory given the discourse tree in (21c).
Examples like (21a) and (21b) thus allow us to properly test the divergent predictions of E&K’s and Kratzer’s theories. Once focus association is added to the mix, E&K’s theory predicts that covariation should be impossible between a focus-marked XPF in a relative clause island and an elided copy in the consequent: the XPF cannot move out of the island (see (22a)), and the entire island is unable to coherently bind into the ellipsis site (see (22b)). Kratzer’s theory, on the other hand, predicts that covariation should be acceptable due to coindexation (see (22c)).
As the following examples demonstrate, covariation is readily available in such configurations:
I only said that the man who bought ASPECTS couldn’t afford to.
≈ Aspects is the only x such that I said that the man who bought x couldn’t afford to buy x
I only said that the woman who moved to FRANCE shouldn’t have.
≈ France is the only x such that I said that the woman who moved to x shouldn’t have moved to x
We conclude that once the confound is removed from E&K’s argument involving relative clause islands, the data in fact furnish an argument against their theory and in favor of Kratzer’s.
The same reasoning extends to the examples involving conditional islands: once we improve the baseline ellipsis, the covariation that E&K predict to be impossible is in fact readily available. To produce an acceptable conditional island baseline, we observe that a VP contained in the antecedent of a conditional can readily serve as the antecedent for ellipsis of a VP in the consequent of that conditional, as in (25).
If John doesn’t buy Celtics tickets, I’ll ask Bill to
buy Celtics tickets.
Bill said that if Mary submits her paper to a journal, he’ll ask Sue not to
submit her paper to a journal.
E&K’s theory therefore predicts that covariation should be impossible between a focus-marked XP in the antecedent of a conditional and an elided copy in the consequent: the XP cannot move out of the island (see (26a)), and pied-piping the whole island will not facilitate deriving the relevant reading (see (26b)). Kratzer’s theory, on the other hand, predicts covariation should be acceptable (see (26c)).
As (27) demonstrates, the covariation reading is indeed possible in such cases. The conclusion is once again that conditional islands furnish an argument against E&K’s theory and in favor of Kratzer’s.9
I only said that if Mary submits her paper to NATURE I will ask Sue not to.
≈ Nature is the only journal x such that I said that if Mary submits to x I’ll ask Sue not to submit to x
3.2 Reevaluating Coordinate Structure Islands
As E&K acknowledge, their conclusion hinges on the assumption that across-the-board (ATB) movement is impossible for covert movement (Bošković and Franks 2000). If covert ATB movement were possible, covariation would be predicted to be available in cases like (28), as the focus-marked phrase could covertly move out of both conjuncts to associate with only, as in (29).
The claim that covert ATB movement is impossible is in turn based on examples like (30), where Bošković and Franks (2000) observe that each book is barred from taking wide scope above someone, contrary to what would be expected if covert ATB were possible.
While we agree that the relevant reading is impossible here, (30) differs from examples like (28) in what turns out to be a crucial way: (30) features an overt occurrence of the quantifier in the second conjunct, whereas in (28), there is no overt quantifier in the second conjunct, which contains a VP-ellipsis site. Strikingly, when the second conjunct contains an ellipsis site, the quantifier can scope out of the coordinate structure and bind into both conjuncts.
Why does ellipsis facilitate the quantifier scoping out of the coordinate structure? We are aware of two analyses with independent support in the literature that can make sense of this apparently novel observation. The first is to assume that the ban on covert ATB movement is ameliorated by ellipsis. This is predicted under Citko’s (2005:sec. 4.2) multidominance approach to the Coordinate Structure Constraint, which holds that covert ATB movement is blocked because of a linearization conflict that arises due to the overtness of the quantifier in the second conjunct, and that ellipsis simply dissolves the linearization conflict. The second option, which is not necessarily committed to the view that covert ATB exists, is to assimilate (31) with Ruys’s (1992) observation that covert movement is possible out of the first conjunct of a coordinate structure if (and only if) the moved XP binds a pronoun in the second conjunct. (31) would then involve movement out of just the first conjunct, and a bound pronoun in the second conjunct that is deleted by ellipsis.
Whether the right explanation is Citko’s, Ruys’s, or something else, the upshot is that the coordinate structure E&K give in (28) should not be an island for covert movement. This is confirmed if we minimally change (28) to exhibit quantifier-scope interaction instead of association with focus, as in (32). The universal can easily take scope over the existential.
Both E&K’s and Kratzer’s theories, then, predict covariation to be possible in (28), as both can simply make use of a derivation like (29) (and Kratzer’s theory can also use coindexation to derive it). The conclusion is that coordinate structure examples do not support the focus movement account over the focus coindexation account.
So why is (28) marginal? We are unsure, but suspect that the judgment is due to some quirk of that particular example which is unrelated to coordinate structures in general. We find other structurally similar sentences, like those in (33), to be perfectly fine. Both E&K’s and Kratzer’s approaches predict this, as sketched in (34a) using the covert ATB analysis (and Kratzer’s approach can also use coindexation, as in (34b)).
I only said that Sue wanted to call BOB and couldn’t.
I only knew that Bill tried to fix the EDSEL and couldn’t.10
3.3 More Islands
To further corroborate our proposal that focus covariation is not sensitive to islands, we document three additional islands that allow the relevant reading. For reasons of space, we do not go over the details of how E&K’s theory predicts the relevant readings to be missing. The argument in each case is essentially identical to the ones in (22) and (26).
Complex NP Island
Only [the claim that all students liked ASPECTS] inspired Sue to find one who didn’t.
≈ Aspects is the only x such that the claim that all students like x inspired Sue to find one who didn’t like x
John only asked [who else likes ASPECTS] after learning that Mary does.
≈ Aspects is the only x such that John asked who else likes x after learning that Mary likes x
I only said that [for John to read ASPECTS] is more important than for Mary to.
≈ Aspects is the only x such that I said that John reading x is more important than Mary reading x
In summary, covariation with focus can take place out of just about every kind of island known in English, and this covariation does not depend on pied-piping the island. Moreover, E&K’s arguments to the contrary are confounded. With relative clause and conditional islands, their baseline examples do not readily license the necessary ellipsis resolution, independent of association with focus. With coordinate structure islands, their argument overlooks the independent fact that coordinate structures with VP-ellipsis are not islands to the sort of covert movement needed to derive the covariation reading. The net result is that an island-insensitive mechanism is needed to derive the full array of covariation readings.
4 The Case of Covariation with Overt Pronouns
As we mentioned in section 2, E&K present one additional argument in favor of their theory and against Kratzer’s, based on the possibility of covariation between a focused phrase and an overt pronoun, rather than with a phrase in an elided VP. In this section, we examine this argument in detail, showing that it does not go through as intended, and that the overall empirical picture concerning covariation with overt pronouns points to the opposite conclusion from what E&K argue for.
Kratzer (1991:831) explicitly proposes that focus coindexation is possible only in ellipsis constructions—in other words, that no two overt constituents can share the same focus index. This was needed in order not to overgenerate covariation readings of Tanglewood sentences in which the focused phrase is repeated overtly. In I only went to TANGLEWOOD because you went to Tanglewood, for instance, the covariation reading is absent (compare with (1)). In E&K’s system, where covariation is captured by ordinary binding, there is no similar reliance on ellipsis; all that is required is that the covarying element be a bindable thing. And indeed, Beaver and Clark (2008:112) have observed that covariation does not require ellipsis, as long as the covarying element is pronominal (i.e., a bindable phrase); see (38a). E&K’s claim, then, is that Kratzer’s reliance on ellipsis undergenerates the reading in (38b), whereas their own system captures it straightforwardly with the LF structure in (38c).
Moreover, along the lines of their logic outlined in section 2, E&K’s system predicts that in island contexts such overt covariation should be impossible: if neither XPF nor the island containing it can felicitously bind the pronoun, covariation cannot be generated. E&K argue on the basis of (39) (their (37)), which is an overt counterpart of (3), that this is borne out.
E&K’s point thus comprises two claims: (a) that Kratzer’s system undergenerates in (38), and (b) that such constructions are sensitive to islands, as shown by (39). We argue, however, that claim (a) does not follow, at least not under a particular view of pronouns (Elbourne 2002), and that claim (b) is empirically incorrect. We elaborate on this now, starting with claim (b).
While we agree that the covariation reading is marginal in (39), we suspect that its infelicity has nothing to do with islandhood; rather, it has to do with a strong preference to elide a large VP that repeats almost verbatim an earlier VP. If we sufficiently change the VPs to avoid this repetition, the example is completely acceptable.
I only promised that [if SUE had trouble at school] would I help her.
≈ Sue is the only x such that I promised that [if x had trouble at school] I would help x
I only said that the police officer [who arrested BILL] treated him fairly.
≈ Bill is the only x such that I said that the police officer who arrested x treated x fairly
(41) involves apparent binding out of the antecedent of a conditional, and (42) involves apparent binding out of a relative clause. In both cases, then, the covariation reading is possible in the exact configuration where E&K claim it should be blocked.
The data in (40), (41), and (42) thus appear to argue against, rather than in favor of, the focus movement account (while the infelicity of (39) plausibly has an independent source). This brings us back to claim (a): the question still remains how such examples are to be analyzed on a focus coindexation account, as both of the covarying elements appear to be overt, whereas focus coindexation should be conditioned by ellipsis of the second one.
We argue that these cases are not an obstacle for the focus coindexation + deletion view: there is a coherent way to generate them if we adopt Elbourne’s (2002) now-popular proposal that pronominal elements can be analyzed as DPs with an elided NP component. Our suggestion finds an explicit precursor in Sauerland 2007:33.13 To illustrate, on this analysis the covarying pronoun it in (40) has the underlying structure in (43), the NP component having been elided under identity with the overt NP in the focused phrase. Following Elbourne’s proposal, the resulting structure is spelled out as it at PF.
(43) it ≡ [DP the
[NP Zoning BoardF1]]
This ensures that the relevant focus index appears on only one overt phrase, maintaining Kratzer’s proposal in full generality.
In examples (41) and (42), the covarying expression is a proper name rather than a determiner-NP sequence. Following Elbourne (2002), this poses no special problems. In particular, there is significant evidence that proper names can at least optionally be analyzed as predicates that must combine with a definite determiner to appear in argument position. For example, Sue in (41) can be analyzed as having the underlying structure in (44).
(44) Sue ≡ [DP the [NP Sue]]
In English, this definite determiner is obligatorily silent, but in many languages it can be realized overtly, as in the German der Hans ‘the Hans’. Once we allow structures like (44) for proper names, we can capture examples like (41) on Kratzer’s theory as in (45), where again the overt pronoun her contains an instance of a focus-coindexed, elided NP.
(45) I only promised if [[the SueF1] had trouble at school] would I help [the
In sum, covariation between a focus-marked phrase and an overt pronoun raises no special challenge to the Kratzerian framework of focus coindexation followed by ellipsis. Moreover, the observed island insensitivity in this domain once again furnishes an argument in favor of this framework and against a focus-movement-based one.
2 It should be noted, though, that our results do not merit the stronger conclusion that covert focus movement (with pied-piping) never takes place. Covert focus movement with pied-piping in association with only has been argued to exist on independent grounds; see Krifka 2006, Wagner 2006, Kotek and Erlewine 2016.
3 While the paraphrase in (8b) suggests that the indefinite (a nanny) takes scope above because, there is arguably a truth-conditionally distinct paraphrase (i) with the reverse scope order (which is in fact also available for (8)).
Spanish is the only (language) x such that we hired a nanny who speaks x because our son hires a nanny who speaks x
We wish to abstract away from the meaning difference between (8b) and (i) (and whether it is really a matter of scope or something else). What currently matters is the fact that the reading in (8a) is not available for (8).
4 E&K reason that (9c) is nonsense because of conflicting semantic requirements on the bound variable x: in its first occurrence it ranges over people (given the lexical requirements of hire), but in its second occurrence it only ranges over languages (given the lexical requirements of speak). We note that this explanation cannot be the whole story, as it does not extend to parallel examples where the semantic conflict is removed. Consider, for example, the minimally different (i), where both occurrences of the variable now range over people (since the verb like can take human themes); still, the indicated reading is not available. The question for E&K’s system is then why the pied-piping representation in (ii) (which should give rise to that reading) is unavailable for (i).
We only hired the nanny who likes JOHN because our son does.
Unavailable reading: We didn’t hire the nanny who likes Bill because our son likes her (= that nanny)
The answer, we think, is that (ii) cannot feed VP-ellipsis: the VP likes [John]F is not identical enough to the variable-containing VP “likes x” for the purpose of ellipsis (see Heim 1997 for a set of definitions that derive this). We therefore believe that the more general problem with a representation like (9c) pertains to the conditions on VP-ellipsis rather than to what the variable can semantically range over.
5 Pied-piping the whole island (i.e., the entire coordinate structure) to only will of course not facilitate binding into the elided VP, because the elided VP is itself contained within the island.
8 Satoshi Tomioka (pers. comm.) has pointed out example (ia) to us, which appears structurally similar to (14) but does allow the relevant ellipsis resolution. However, our main point is not threatened, because (as Tomioka also observes) the relevant covariation reading is also available once focus association is added; see (ib).
Josh hired someone who can use LaTeX because his assistant canNOT.
Josh also hired someone who can use PREZI because his assistant cannot.
We currently have little insight into what might be the underlying difference between (14) and (ia). It is possible that the addition of contrastive negation might somehow facilitate the otherwise difficult ellipsis resolution (see also footnote 10).
9 A final note is in order concerning E&K’s conditional island data in (11). As Pullum (1987:264) points out, if-clauses associated with direct object it-pronouns are not islands for extraction. Thus, (ia), which involves a true conditional adjunct, differs strikingly from (ib), where the if-clause is associated with direct object it.
?*Which commitment will Joe quit his job [if we cannot keep e]?
Which commitment would Joe regret it [if we cannot keep e]?
It follows that E&K’s example (11) cannot be used to make their point even if the baseline ellipsis in (15) were possible, since there is no reason the focus-marked XP could not simply raise to associate with only. However, we decided to keep to E&K’s assumption that the if-clauses in (11) and (15) should be considered islands, for the sake of argument. Importantly, our example in (27) relies on a true conditional island.
10 Michael Yoshitaka Erlewine (pers. comm.) points out that both in (31) and in (33) (but not in (32)), the examples have some kind of contrast in polarity between the conjuncts (tried to . . . and failed to; wanted to . . . and couldn’t). Perhaps this type of contrast facilitates ellipsis compared with structures that lack it (see also footnote 8).
11 E&K themselves observe (p. 454) that a focus-marked XPF in a wh-island can license a covariation reading. They claim that this follows because wh-islands are weak islands that allow extraction in some cases. Note that none of the other cases discussed here are “weak” islands in this sense, as all robustly block overt movement.
12Tomioka (1999:219n4, 238n18) mentions that there is some variation in the acceptability of the covariation reading here. He cites a statistical study by Hirschberg and Ward (1991) which suggests that covariation out of a relative clause is possible for most native speakers of English, and we will follow his assumption that covariation across islands should be generated by the grammar. We do not have an explanation for the cross-speaker variation.
14Tomioka (1999) and Büring (2004) analyze (41) and (42), respectively, not in terms of NP-ellipsis but in terms of what Elbourne (2002) calls a D-type analysis, in which the covarying pronouns are a spell-out of contextually supplied descriptions that contain a bound variable (see Cooper 1979, Heim and Kratzer 1998). That analysis can capture overt covariation out of islands without necessitating focus coindexation, but it runs into well-known Formal Link problems (see Heim 1990, Elbourne 2002, Sauerland 2007). We therefore assume that a D-type analysis should not form the basis of a theory of covariation with overt pronouns.
We would like to thank Michael Yoshitaka Erlewine, Danny Fox, Irene Heim, Roni Katzir, David Pesetsky, Roger Schwarzschild, Satoshi Tomioka, and one anonymous reviewer for helpful comments. All errors are our own.