2.1 The Island-Sensitivity Generalisation

In the ellipsis literature, syntactic isomorphism obtains in clausal ellipsis environments when an elliptical clause (henceforth, a source) and its antecedent have parallel phrase markers. Although it has been claimed that syntactic isomorphism is required to license some forms of ellipsis (Rooth 1992b; Fiengo & May 1994), research has shown that isomorphism is not required to license clausal ellipsis (Merchant 2001; Barros 2014a; among others). Support for this conclusion comes from the observation that isomorphic sources generate unattested readings (9), are ill-formed due to island-violating wh-movement (10–11), or simply cannot be formulated (12) for many clausal ellipsis constructions that are nonetheless judged as acceptable. Such constructions must therefore have syntactically anisomorphic (i.e. not isomorphic) sources.

As the Hungarian example from Barros et al. (2015) in (13) shows, clausal ellipsis is prohibited when no well-formed source is available. In this example, the copular clausal source in (13a) is unacceptable because the plural agreement suffix –ak, which obligatorily suffixes to predicative adjectives, is absent from magas, whereas the sources in (13b–13c) are ill-formed because they display island-violating wh-movement.

This example from Hungarian exemplifies the Island-Sensitivity Generalisation presented in (7), as the remnant’s correlate in the antecedent clause (i.e. magas) is contained in a syntactic island. This generalisation holds across languages, as Abels (2011) and Barros et al. (2014) have shown. With respect to wh-fragments such as milyen magas in (13), the pattern described by the Island-Sensitivity Generalisation is hardly surprising: Hungarian fronts its wh-phrases, and therefore an island-violation is expected when no “island-evading” elliptical source is available. What is surprising is that the Island-Sensitivity Generalisation also holds for non-wh fragments in languages that do not Ā-move presentational or contrastive foci in standard, non-elliptical environments. This is demonstrated by the English dialogue in (14). In this example, no anisomorphic source for ellipsis is available, as the adjective former cannot be used predicatively (i.e. Buckeye Bill is former is ungrammatical).

The observation that the Island-Sensitivity Generalisation holds for non-wh fragments in English is straightforwardly explained under the M&D approach, which maintains that all remnants of clausal ellipsis, regardless of whether they bear wh-features or not, undergo Ā-movement to escape ellipsis. Consequently, the only source available for (14B) – i.e. the isomorphic source – violates the Left Branch Condition (Ross 1976), and is therefore ungrammatical:²

2.2 The Preposition-Stranding Generalisation

Rodrigues et al. (2009) demonstrate that, in the Spanish example in (16a), the source for clausal ellipsis is the anisomorphic cleft clause in (16b).

Because these cleft clauses cannot host the exceptive modifier más ‘else’ in Spanish (as (17a) shows), its presence in a clausal ellipsis construction such as (16b) makes the cleft source unavailable. When this happens, an unacceptable utterance is generated, as (17b) shows. Unacceptability arises here because no isomorphic source is available, either. Generating the isomorphic source requires preposition-stranding wh-movement, which is forbidden in Spanish (see (17c)).

Although no anisomorphic copular clausal or cleft source for clausal ellipsis is available (see (19)) (Merchant 2001: 122), the English counterpart of (17b) in (18) is acceptable. This is because the isomorphic source in (20) is available, due to the fact that English allows preposition-stranding wh-movement.

The contrast observed between English and Spanish exemplifies the Preposition-Stranding Generalisation from (8). When the availability of anisomorphic sources is controlled for, one observes that English (a preposition-stranding language) allows for “bare” DP fragments that are interpreted as complements to P, while Spanish (a language without preposition-stranding wh-movement) does not.

With respect to wh-fragments, the pattern described by the Preposition-Stranding Generalisation is again unsurprising: the constraints on wh-movement that (fail to) generate bare DP wh-fragments are also observed in non-elliptical environments. What is surprising is that the Preposition-Stranding Generalisation holds for non-wh fragments in languages that do not Ā-move presentational or contrastive foci in standard, non-elliptical environments. For instance, English permits DP fragments that are interpreted as complements of P (see (21)), but German, which does not permit preposition-stranding Ā-movement, does not (see (22)). (Note that the dative case-marking on the German fragment rules out any anisomorphic sources for (22B).)

The observation that the Preposition-Stranding Generalisation holds for non-wh fragments is straightforwardly explained under the M&D approach, which maintains that all remnants of clausal ellipsis, regardless of whether they bear wh-features or not, undergo Ā-movement to escape ellipsis. This movement is permitted in English, which permits preposition-stranding, but prohibited in German, which disallows preposition-stranding (in the cases under consideration here, at least):

3.1 Recovering the meaning of clausal ellipsis from the Maximal QUD

Ginzburg (1994) defines the Questions Under Discussion (the QUD) as a partially-ordered set that specifies the currently-discussable questions at any point during a conversation. The maximal QUD (MaxQUD) is the most conversationally-prominent discussable question (or unordered subset of questions) in this set, and occupies the topmost position on the pushdown stack. Questions are introduced into the QUD by a number of processes. For instance, asking an explicit question q makes q the MaxQUD, whereas asserting p moves the polar question ?p into the MaxQUD position (Farkas & Bruce 2010; see Ginzburg & Miller 2017 for more ways to add questions to the QUD). In the first case, the MaxQUD is an explicit question, whereas in the second case the MaxQUD contains an implicit question. Adopting a “direct interpretation” approach, Ginzburg & Sag (2000: Chapter 8) argue that the propositional meaning of a fragmentary response is obtained by unifying the fragment with a single question within the (possibly singleton) set of questions that constitutes the MaxQUD. Ginzburg & Sag’s theory has been independently recast in Minimalist terms by Krifka (2006), Reich (2007), Barros (2014a), Weir (2014; 2017), and others. In this line of research, which adheres to the notion that fragments are remnants of clausal ellipsis, it is proposed that the meaning of an elided clause is semantically recovered from the meaning of a single question in the MaxQUD. The numerous benefits of appealing to a question in the MaxQUD to recover the meaning of the elided clause, rather than appealing only to the utterance that explicitly antecedes the elided clause, are well-documented in this line of research, and are therefore not repeated here (see Weir 2014 for an engaging overview).

Weir (2017) has recently demonstrated that a simple QUD-based recoverability condition such as (25), which states that clausal ellipsis is recoverable in a clause α if the focus semantic value of α and the denotation of a question in the MaxQUD match, is insufficient for capturing the patterns of acceptability found in a number of clausal ellipsis configurations.

Weir’s evidence comes from comparing questions such as those in (26) and (27) below. On a standard “propositional-set” view of questions and focus, which uses Hamblin/Rooth alternatives, these questions denote the same sets of sets of possible worlds. In other words, they are cointensional. This is because the set of worlds in which the machine sends two signals (for instance) is the same set of worlds in which it sends the signal twice. From a propositional-set viewpoint, the focus semantic values for the answers in (26) and (27) are cointensional in precisely the same way.

Because the questions in (26) and (27) are cointensional, and because their answers are cointensional too, the simple recoverability condition in (25) predicts that clausal ellipsis is recoverable in each of the answers. This prediction is not borne out. Only the cardinal fragment may answer the question in (26), and only the multiplicative adverbial fragment may answer the question in (27), as the examples in (28) and (29) show. Having convincingly shown that one cannot appeal to morphosyntactic constraints to explain the unacceptability of (28b) and (29a), Weir concludes that the recoverability condition in (25) is inadequate.

Weir argues that the contrast observed in (28) and (29) can only be captured by a recoverability condition on clausal ellipsis that is formulated in the Structured Meaning approach to questions and focus. In this framework, questions are viewed as simple functions that yield type-t expressions when applied to their answers. Both questions and statements containing foci are structured propositions that are arguments of speech-act operators such as QUEST(ION) and ANS(WER) (Reich 2002; Krifka 2006), as shown in (30) and (31). In these representations, the last member of the ordered set is the background, while those members that precede the background are collectively known as the focus. These representations are known as focus-background structures.

Adopting the Structured Meaning framework, the recoverability condition that Weir offers is reproduced in a modified form in (32).³ Unlike (25), which makes reference to the meaning of the entirety of the question that constitutes the MaxQUD and the entire elliptical clause, this condition makes reference only to meaning of their backgrounds.

Now consider the focus-background structures for the utterances in (28) and (29), which are given in a simplified form in (33) and (34) below. In (33), only the background of the first answer matches with the background of the MaxQUD. Conversely, only the background of the second answer matches with the background of the MaxQUD in (34). Therefore, according to Weir’s condition, clausal ellipsis is only recoverable in these matching configurations. As the data in (28) and (29) have already shown, this is correct.

Ginzburg & Sag (2000), Barros (2014a), and Weir (2014) argue that indefinite expressions invoke implicit wh-questions that enter the MaxQUD. This means that the assertions in (35A) and (36A) invoke implicit MaxQUDs that are equivalent to the explicit questions in (28) and (29) respectively. Because the pattern of ellipsis licensing observed in (36–35) is the same as the one in (28–29), one must conclude that the implicit questions that license clausal ellipsis also have structured meanings.

At this point, the question arises about the how implicit questions that license clausal ellipsis come to have meanings that are structured. There are two options: either these questions are linguistically derived but unpronounced, or they are not linguistically derived and our domain-general conceptual structures (i.e. Mentalese) are complex enough to mimic focus-background structures. Although the latter option is pursued in other frameworks (e.g. Dynamic Syntax with the Type Theory of Records), I suggest that the former option is better suited to Minimalist approaches to fragments, following (Reich 2004: §5.2). This is because these approaches analyse fragments as remnants of ellipsis, which has been characterised as a surface anaphor (Hankamer & Sag 1976). This characterisation can be maintained if the implicit questions that license clausal ellipsis are viewed as linguistically – i.e. syntactically – derived. For ease of reference, let us refer to this idea as the QUD-syntax Correspondence conjecture:

As I show in §4, the main consequence of adopting (37) is that one can use it to derive the Island-Sensitivity and Preposition-Stranding Generalisations without appealing to M&D-style mechanisms.⁶ Before doing this, I will first outline how pied-piping can be represented in the Structured Meaning framework, so that Weir’s Background-matching condition from (32) can be straightfowardly employed in configurations where pied-piping is observed.

3.2 Representing pied-piping in the Structured Meaning approach

In order to apply Weir’s Background-matching condition in configurations that display pied-piping (which will be necessary for deriving the Preposition-Stranding Generalisation in §4), I must develop a Structured Meaning approach of pied-piping. To accomplish this, I adopt Cable’s (2010) approach to pied-piping and translate his “Q-based” theory into the Structured Meaning framework.

Cable’s Q-based theory of pied-piping is a reflex of his theory of wh-questions. Cable argues that wh-phrases must be c-commanded by a Q-particle, which is phonologically realised in many of the world’s languages but is null in English. I rename the Q-particle as the f-particle here, as I will shortly put the Q-particle – or rather, the semantic choice function introduced by the Q-particle at LF – to work in both questions and assertoric utterances that display foci. In wh-movement languages such as English, it is the maximal projection of the f-particle (i.e. fP) that undergoes wh-movement, as shown in (38) below. From a semantic perspective, Cable views wh-words as having a focus semantic value but as having no ordinary semantic value, following Beck (2006). In other words, he views wh-words as introducing a set of Hamblin-alternatives. Through Pointwise Function Application (Hamblin 1973), this set of alternatives must become large enough to provide a focus semantic value for the syntactic sister of the f-particle. As already mentioned, the f-particle itself is treated as a choice function, which is any function that chooses a member from a non-empty set. This choice function takes the focus semantic value of its syntactic sister as its argument. For Cable, the choice function introduced by the f-particle is a variable, which is existentially-bound from the periphery of an interrogative clause. This existential binding is effectuated by the syntactic head Force_Q, which also introduces a λ-binder for the proposition and hence derives a question meaning. Putting these ideas together and incorporating them into the standard propositional-set approach to interrogatives yields the meaning in (39) for the wh-question in (38).⁷

Notice that the wh-phrase is sister to the f-particle in (38). Because it is fP rather than the wh-phrase itself that undergoes Ā-movement to form questions in English, a pied-piping configuration is generated whenever the f-particle is sister to larger phrase containing the wh-phrase, as (40) shows. For our purposes, we can take the view that the semantic formulae in (39) and (41) therefore differ only with regards to the size of the choice function’s argument. Put differently, the set of individual concepts derived through Pointwise Function Application is larger in (41) than in (39). As Cable (2010: Chapter 2) makes clear, an important benefit of this analysis is that it renders LF-pied-piping mechanisms superfluous, as (39) and (41) pick out the same set of propositions (Cable 2010: 73).

From a syntactic perspective, Cable assumes that f and C_Q must enter into syntactic AGREE relation (Chomsky 2001) in wh-questions, which is represented by a dashed line in (38) and (40). Because f selects for its complement in wh-questions in English and wh-movement is overt, this triggers the overt Ā-movement of fP to SpecCP. We may therefore understand wh-movement in English as arising from the fact that C_Q bears an EPP feature (Chomsky 2000: 102). Because Ā-movement cannot occur over island boundaries, this encodes the island-sensitivity of English wh-movement. Note that C_Q itself is semantically vacuous (Cable 2010: 79).

I will now recast Cable’s analysis in the Structured Meaning framework. The reader should be aware at this point that, in order to maintain clarity of exposition and ease of comparison, this recasting mimics Cable’s analysis as closely as possible, and therefore utilises Roothian (1985; 1992a) Alternative Semantics in a framework that typically eschews it. For a one-dimensional alternative to Cable’s analysis that is couched in the Structured Meaning framework, see Reich (2004).

To translate the import of Force_Q into Structured Meaning terms, I suggest that Force_Q’s introduction triggers the creation of a structured proposition that has the choice function as its focus, and a λ-abstraction over the choice function as its background (following Reich 2002: 175). Thus, our Structured Meaning versions of (39) and (41) are provided in (42).

Cable (2010: 201–205) suggests that his analysis can be extended to declarative utterances containing foci. I actualise Cable’s suggestion here. Following Reich (2004), I suggest that F-marked items are also semantic arguments to choice functions, and that these choice functions are also introduced via syntax, just as they are in Cable’s theory of wh-questions. To maintain the analogy with wh-questions, I will also refer to the syntactic instantiation of these choice functions as “f-particles”. A plausible Structured Meaning representation for the declarative utterance in (43) is presented in (44).

There is a notable syntactic difference between (43) and (38): fPs in assertoric responses are adjuncts in English (reflecting the adjunct status of the so-called squiggle operator from Rooth 1992a; see Reich 2002; 2004),⁸ whereas fPs in wh-questions take a phrase containing the wh-phrase as its complement.⁹ Aside from the larger structured proposition being the argument to a different speech-act operator (i.e. ANS(WER), which is introduced by Force_A, rather than QUEST(ION), which is introduced by Force_Q), the main semantic difference between (44) and the formulae in (42) is that the choice function in (44) is constrained. f_Pete is the constant function that, when applied to the set represented by its argument, chooses Pete from the set. This allows for the correct ordinary semantic value to be retrieved for (44), as such values are retrieved from structured propositions by applying the background to the focus (i.e. ⟨α, β⟩ = β(α)). For a detailed discussion of how choice functions can be constrained in this manner without recourse to syntactic movement of the focused phrase, see Reich (2004).

Just as the f-particle in questions is not required to take a wh-phrase as its direct complement (see (40)), the f-particle in assertoric responses is not required to directly adjoin to a presentational or contrastive focus. The declarative counterparts of a pied-piping configuration, which are known as Focus Phrase configurations (Drubig 1994), are generated in such circumstances. According to our simple Structured Meaning treatment, the presentational Focus Phrase configuration in (45) yields the focus-background structure in (46).

The Structured Meaning-style formulae presented above are admittedly rather complicated. For clarity’s sake, I will therefore use simplified notations whenever possible in the rest of this paper. Henceforth, constraints on choice functions are omitted, as are the restrictors for foci and wh-phrases. λ’s are also removed. Thus, structures without pied-piping are represented as in (47a), whereas structures with pied-piping are represented as in (47b).

By recasting Cable’s (2010) theory in Structured Meaning terms in §3.2, I have now extended Weir’s (2017) analysis to pied-piping and Focus Phrase configurations. The most useful outcome of this exercise is that one can now identify a common property shared by wh-questions and their responses, which is that each displays an fP. Furthermore, the utilisation of fP allows for a more accurate PF instruction for clausal ellipsis to be formulated. Recall from §1 that the rule for selective pronunciation (see (4)) specified that, in an elided clause, a designated phrase Y containing a wh-phrase or a focused element can be pronounced. One can now identify this designated phrase more accurately:

The distribution of the E-feature for English clausal ellipsis can also be specified now: it is optionally borne by Force_Q and Force_A.¹¹ Resultantly, a more accurate in-situ analysis of the fragments in (1), which is repeated below in (49), is provided by the structures in (50).¹²

Having now introduced a way to represent pied-piping in the Structured Meaning framework, and having now refined the notion of how clausal ellipsis is syntactically licensed under an in-situ approach, the stage is almost set for deriving the Island-Sensitivity and Preposition-Stranding Generalisations without requiring remnants to undergo exceptional movement from the ellipsis site in §4. However, one final issue must be addressed before we proceed to §4. Because it is novel (to my knowledge), readers receptive to the foundational idea in my analysis – namely that syntactically derived implicit questions can be invoked from explicitly-uttered assertions (see (37)) – might nonetheless be puzzled about what this “invocation” procedure consists in and whether there are restrictions on it.

Although my response to these questions is fully compatible with the current ellipsis literature and therefore wholly unsurprising to ellipsis specialists, expounding it is too space-consuming and tangential to this paper’s main aim, which is merely to derive the two generalisations from §2 using an in-situ approach to clausal ellipsis. Because of this, I choose to address these questions in an appendix to this paper. Stated very briefly here, the response given in the appendix runs as follows: “invocation” consists in the listener accommodating an implicit question when confronted with a fragmentary response which has a non-erotetic explicit antecedent. This accommodation procedure is syntactically specified (Fox 1999; Thoms 2013; 2015), and can perhaps be identified as Arregui et al.’s (2006) recycling procedure. Ignoring some important details, the accommodation procedure is constrained such that:

(51)	i.	Aside from its fP projection, an implicit wh-question cannot contain any morphosyntactic that are not also present in the explicit antecedent (see Rudin 2018 and references therein).
	ii.	Any pronoun present in the implicit wh-question must co-refer with a phrase in the explicit antecedent (cf. Vehicle Change, Fiengo & May 1994: 218).

Because the implicit questions discussed in the upcoming sections are either syntactically isomorphic to explicit antecedents from which they are invoked or are anisomorphic copular clausal questions that obey the constraints listed above, this partial clarification of what it means to “invoke a syntactically derived implicit question” is sufficient for current purposes. Readers interested in this aspect of the current framework are directed to this paper’s appendix for further discussion.

4.1 The Preposition-Stranding Generalisation without Ā-movement

If the QUD-syntax Correspondence conjecture is correct, and if one adopts the idealistic position that only syntactically well-formed structures in a particular language L can be used to advance the conversation in L (i.e. if one ignores performance errors), then a question q can only be in the MaxQUD in L if q is syntactically well-formed in L. This idea is the driving-force behind deriving the Island-Sensitivity and Preposition-Stranding Generalisations without reference to syntactic movement from within the ellipsis site. In this subsection, I derive the Preposition-Stranding Generalisation.

Cable (2010) offers a number of restrictions on the syntactic position of fP. One relevant restriction is that fPs cannot be interveners to selection by functional heads. This is encapsulated in his fP-intervention condition, which is presented in (52).¹³

At first glance, the preposition-stranding wh-question in (53) appears to violate the fP-intervention condition, as fP intervenes between DP and P that selects it before Ā-movement occurs. Adopting ideas from Abels (2003), Cable (2010: 111) argues that the fP-intervention condition is actually satisfied in (53), as English Ps are lexical heads, not functional ones.

According to the QUD-syntax Correspondence conjecture, this means that (53) can serve as the MaxQUD, and can therefore license clausal ellipsis in a clause whose background matches with the background of (53). The question in (53) is the MaxQUD in the dialogue below. In this case, clausal ellipsis is licensed because the background of the MaxQUD and the background of the elliptical utterance match, as (55) shows.

According to Cable, prepositions are functional heads in a language such as German. This means that fP cannot intervene between DP and P that selects it, which consequently rules out preposition-stranding wh-movement:

According to the QUD-syntax Correspondence conjecture, this means that (56) cannot serve as the MaxQUD. It is syntactically ill-formed, and is therefore not used in conversation. Consequently, ellipsis cannot be licensed in any clause that depends on (56) for satisfying Weir’s Background-matching condition, as (56) is simply absent from the QUD. Because B’s response in (57) does indeed depend on (56) being the MaxQUD for its ellipsis site to be semantically recovered, ellipsis is judged to be unacceptable in (57B). Note that the meaning of the ellipsis site in (57B) cannot be recovered from the pied-piping wh-question that is the MaxQUD because the background of this question and the background of (57) do not match, as (58) shows.

This QUD-based explanation for the contrast observed between English and German is fully extensible, and therefore derives the Preposition-Stranding Generalisation without reference to movement of the remnant. Non-P-stranding languages cannot generate syntactically well-formed P-stranding wh-questions, and therefore such questions cannot serve as MaxQUDs. Clausal ellipsis cannot be recovered from a question that is not in the QUD, and consequently an elided clause whose meaning can only be recovered from an ill-formed (and therefore “QUD-absent”) P-stranding question is judged to be unacceptable.

Before moving on, there is slightly more to say about the English P-stranding data. Recall from §3.2 that, according to Cable’s (2010) theory, two utterances u and u^′ pick out the same (sets of) sets of possible worlds if they only differ in how much pied-piping they display. I take this to mean that, if a question q is added to the QUD, all syntactically well-formed pied-piping equivalents of q are simultaneously added to the QUD. This allows us to capture mismatch cases in which pied-piping of P occurs in the explicit question but not in the response (see (59)) or vice versa. This is because the explicit question in (59A) invokes the syntactically well-formed implicit P-stranded question who did Peter speak with?, whose background may serve as the antecedent for clausal ellipsis in (59B).

Mismatches are disallowed for some speakers of English when an elided wh-question has an explicit wh-question antecedent, however, as the examples in (60) show. I suspect that this restriction is pragmatic in nature, as such wh-fragments function as verbatim repetitions of the antecedent question. This observation is encapsulated in the rule in (61).

4.2 The Island-Sensitivity Generalisation without Ā-movement

When coupled with the QUD-syntax Correspondence conjecture, the in-situ approach captures the Island-Sensitivity Generalisation straightforwardly. To see this, consider the dialogue in (62) below. After B answers with yes in (62), which moves the propositional content of A’s assertion the company wants to hire a hard worker into the Common Ground (Stalnaker 2002),¹⁴ a number of equally-maximal implicit questions are added to the QUD.¹⁵

We observe that (62a–c) are available as fragmentary responses to (62A), each licensed by a different implicit question. Conversely, the fragment in (62d) is not available as a fragmentary response because there is no question in the QUD available to license it. The only question that could license (62d), which is given in (62ii), is syntactically ill-formed in English, and is therefore unsuitable for use in conversation.

This example illustrates that a fragment with an island-bound correlate and for which no anisomorphic source is available is unacceptable because no question is available in the QUD to semantically license it. Thus, English non-wh fragments are indirectly island-sensitive: this “island-sensitivity” arises as an epiphenomenon. Such fragments do not themselves cross island boundaries.

4.3 Summary of §4

I demonstrated in this section that the independently motivated QUD-syntax Correspondence conjecture, which is the idea that the meaning of a clausal ellipsis site is always recovered from a syntactically derived question, can be utilised to capture the Preposition-Stranding and Island-Sensitivity Generalisations without resorting to the idea that movement always occurs from within the clausal ellipsis site.¹⁸ Adopting this conjecture therefore provides the required restrictiveness to the in-situ approach to clausal ellipsis but renders the M&D approach to clausal ellipsis superfluous, at least for capturing the Preposition-Stranding and Island-Sensitivity Generalisations.

5.1 Fragments without Ā-movement properties

5.2 Wh-in-situ languages

The remit of the Island-Sensitivity Generalisation in (7) is purposefully restricted to languages with wh-movement. This is because having an island-bound correlate typically has no effect on the acceptability of fragments in wh-in-situ languages such as Japanese (Abe 2015 and references in there), Korean (Kim & Park 2016: 27–28), and Turkish (İnce 2012), even when no island-evading anisomorphic source is available. This is exemplified by the non-wh presentationally and contrastively focused fragments in (86B) and (87B), which come from Turkish (İnce 2012: 206).

When one considers that no predicational copular clausal short source is available for the fragments in (86) and (87) (as the ungrammaticality of (88) shows), one must conclude that Turkish fragments are genuinely island-insensitive.

To account for these data within the M&D framework, İnce (2012) adopts Merchant’s (2004) idea that island-sensitivity is a phonological phenomenon, which is obviated when the lower links in an Ā-chain that crosses an island boundary are fully contained in a phrase that undergoes ellipsis (see Merchant 2004: 705–709 for details). Aside from being an unfalsifiable theory, İnce’s analysis fails to explain why some fragments are indeed island-sensitive in Turkish:

Let us now consider these Turkish data from an in-situ perspective, abstracting momentarily away from the technicalities of the particular in-situ analysis adopted in this paper but upholding the QUD-syntax Correspondence conjecture. When viewed from this perspective, the data receive a natural explanation. Because Turkish is a wh-in-situ language, the QUD that licenses (86B) and (87B) is syntactically derivable (see (90)) and therefore available for licensing fragments when invoked from assertions. Although having island-bound wh-phrases is usually no obstacle to forming questions in wh-in-situ languages such as Turkish, grammatical questions cannot be formed if the wh-phrase is an adjunct contained in an island (for reasons that are irrelevant here; see Huang 1982; Görgülü 2006; and Çakır 2016 for discussion). Consequently, (89B) is unacceptable because there is no question in the MaxQUD to license it, as the desired question is syntactically ill-formed (see (91)).²⁴

The ease with which the in-situ approach (as broadly construed) captures behaviour of fragments in wh-in-situ languages therefore highlights its cross-linguistic flexibility. The M&D approach, by comparison, is overly restrictive, as it requires the displacement of remnants in languages which typically front neither foci nor wh-phrases.

Although the island-(in)sensitivity of fragments in wh-in-situ languages is captured naturally by the in-situ approach as broadly construed (see Abe 2015 for a sustained example), the technical machinery behind the particular in-situ analysis advocated in this paper actually conspires to generate incorrect predictions about the distribution of fragments in wh-in-situ languages. In the remainder of this subsection, I describe the problem at hand (using Japanese ‘no da’ fragments as my construction of illustration) and then offer a solution.²⁵

In contrast to wh-movement languages such as English, in which the f-particle selects for a wh-phrase (or a phrase that contains it), the f-particle adjoins to a wh-phrase (or a phrase that contains it) in wh-in-situ languages (Hagstrom 1998; Cable 2010). From this adjoined position, the f-particle then undergoes overt or covert long-distance head-movement to C_Q. Because the f-particle cannot cross an island, question formation in wh-in-situ languages is therefore island-sensitive. In cases where question formation appears to be island-insensitive, the f-particle is actually base-generated as an adjunct to the island itself, as the simplified representation for (92) in (93) illustrates.

Because the f-particle is base-generated as an adjunct to the island, (92) yields a focus-background structure in which the choice function takes the entire island as its argument:

To satisfy the Background-matching condition and therefore license clausal ellipsis, fP must also be base-generated as an adjunct to a corresponding island in any candidate elliptical response to (92). Because deletion avoids the syntactic sister of f, this gives rise to the prediction that a fragmentary response to a question such as (92), in which the wh-phrase is contained in an island, must itself be island-sized (and no smaller).

Problematically, this prediction is incorrect. As the Japanese example in (95) shows, and as the Turkish examples in (86) and (87) have already shown, fragmentary responses to questions in which the wh-phrase is contained in an island can correspond to the wh-phrase itself, rather than to the island that contains it.

Although Abe (2015) has convincingly shown that fragments such as (95B) cannot be derived from island-evading anisomorphic sources (contra Saito 2004), one could still contend that, because the Japanese fragment in (95B) is a cleft-like structure, the clausal ellipsis that derives (95B) is governed by a different licensing constraint to the Background-matching condition.

Rather than dismiss fragments in wh-in-situ languages such as Japanese as somehow ontologically different from fragments in wh-movement languages, my current solution to the problem posed by (95B) for the current analyis (which is perhaps not the ultimate solution) is to posit an additional yet silent f-particle within wh-questions in such languages; one which does not enter into an AGREE relation with a clause-peripheral head. This f-particle selects for the wh-phrase, as the updated representation for (92) in (96) illustrates.

If a corresponding extra fP is inserted in the response to a question that fits the syntactic schema in (96), then matching focus-background structures are generated for the question and its response. This is demonstrated for the dialogue in (95) in (97) and (98) below (where g is the choice function introduced by the extra f-particle in the simplified focus-background structures in (98)).

According to this solution, there are two fPs contained in the underlying clause for (95B): one that adjoins to the island and one that adjoins to the phrase that corresponds to the eliciting question’s wh-phrase. Consequently, deletion can apply in such way that it either leaves the entire island pronounced, or only leaves pronounced the focused phrase contained within it (as in (97b)).

It is important to stress two aspects of this solution. Firstly, the fact that the additional fP in (96) selects for a complement means that Cable’s fP-intervention is active, which captures the observation that Japanese obeys the Preposition-Stranding Generalisation (once the generalisation broadened to include all adpositions and case morphology on nominals) (Takita 2009).²⁶ Secondly, the analysis exemplified in (96) is only available for languages in which wh-phrases are focused in standard single-wh constituent questions (i.e. wh-in-situ languages such as Japanese; see Truckenbrodt 2013). It cannot be extended to wh-movement languages such as English or German, as Ā-moved wh-phrases in wh-movement languages are inherently unfocusable in standard, single-wh constituent questions (ibid.). See Reich (2002: Footnote 21) for a similar analysis.

5.3 Summary of §5

I have shown in this section that, in addition to successfully deriving the Island-Sensitivity and Preposition-Stranding Generalisations from §2, the in-situ approach to clausal ellipsis defended in this paper makes correct predictions about certain types of fragments that can only be explained in the M&D framework by recourse to ellipsis repair. These include fragments that do not exhibit Ā-movement properties in English, and fragments in wh-in-situ languages.