Contrasting Contrastive Left-Dislocation Explications

Contrastive left-dislocation (henceforth CLD¹) is a member of a family of left-dislocation phenomena that also includes “ordinary” topicalization (TOP) and hanging-topic left-dislocation (HTLD).

In German and Dutch, the three construction types can usually be most easily told apart by verb-second (V2) and the presence, form, and placement of the element associated with the topic in clause-initial position. The fronted topic in TOP constructions triggers V2 and is not linked to a resumptive element. Both CLD and HTLD constructions feature a resumptive element, but its form and placement usually differ. In CLD, the resumptive is typically a d-pronoun, preferably (but not necessarily) placed in the left periphery (i.e., (1b) is more natural than (1bʹ)). In HTLD, it is always a personal pronoun, often (but again not invariably) placed in the middle field (i.e., (1c) is more marked than (1cʹ); HTLD is fully ungrammatical with weak ’m because the position to the left of the finite verb in V2 constructions accepts no weak nonsubject material). In neither CLD nor HTLD constructions is the fronted topic itself followed by the finite verb. When the resumptive element is fronted into the left periphery, this element triggers V2. (The examples in (1) are from Dutch. English lacks CLD, but it does have the other two left-dislocation constructions.)

Apart from the presence and form of the resumptive element, TOP, CLD, and HTLD are also distinct in that only HTLD can fail to exhibit case connectivity. In ordinary topicalization-cum-V2 and CLD constructions, the topic must always have the case expected on the basis of its relationship with the predicate in the clause. As Grohmann (2000:161, (24)) points out, HTLD does not preclude case connectivity between the topicalized constituent and the proform to which it is linked; but importantly, HTLD is unique in allowing the topic to surface with default case. The German examples in (2)–(3) illustrate this.

The sentences in (2), exhibiting dative case connectivity (helfen ‘help’ is a dative case assigner), are all grammatical (in the case of CLD and HTLD irrespective of where the resumptive element is placed in the clause); but in (3), with default nominative on the topic, only the HTLD examples are well-formed.

Even when the topic shows case connectivity, however, HTLD is generally taken not to allow binding connectivity between a pronominal variable contained in it and a constituent of the clause. In this regard, too, HTLD behaves differently from TOP and CLD, both of which exhibit binding connectivity effects. The German examples in (4) illustrate this for bound variable anaphora. Case connectivity does not guarantee binding connectivity: the HTLD sentences in (4c–cʹ) are grammatical with seinem ‘his’ interpreted referentially but do not support a bound variable reading for seinem, despite case connectivity.

Surveying the empirical landscape of left-dislocation, we discover that CLD and HTLD cannot be told apart on the basis of the placement of the resumptive element in the linear string, and that even the form of the resumptive (d-pronoun vs. personal pronoun) is not necessarily a foolproof discriminator because CLD can in principle also avail itself of personal pronouns as resumptives (Ott 2014:273n7). But the facts that (a) HTLD but not CLD can fail to show case connectivity and (b) CLD but not HTLD exhibits binding connectivity are two generally reliable diagnostics. In any doubtful case, therefore, we should be able to resort to these tests to settle the matter.

For each of the left-dislocation phenomena surveyed in section 1.1, the logical hypothesis space admits three different approaches to account for their properties. These are listed in (5) (where “RES” marks a resumptive correlate, and “Δ” represents an ellipsis site).

Inside CP1, option (5c) could in principle avail itself of either (5a) or (5b); that is, the topic could be assumed to be base-generated in the left periphery of CP1 (as in (5a)) or to be moved there (as in (5b)), binding a trace inside the ellipsis site Δ. Therefore, the difference between (5c) and (5a–b) does not lie primarily in whether or not movement is exploited for the placement of the topic. Rather, the key difference is that (5c) takes (some) left-dislocation phenomena to involve a biclausal structure, with the topic occupying the left periphery of a clause (CP1) whose TP is marked for ellipsis, and with the overt TP surfacing in a separate clause (CP2).

A prominent representative of a movement analysis along the lines of (5b) is that of Grohmann (2003), whose approach to CLD is schematized in (6). (The notation “Top+Vfin” is meant to abstract away from the internal constitution of the Top-head in V2 constructions.)

In Grohmann’s approach, both TOP and CLD involve movement of the topic into the left periphery, as in (6a), but in CLD the topic moves from this topic position into a position even higher in the left periphery (identified by Grohmann as Spec,CP; see (6b)). This movement operation takes place entirely within the confines of what Grohmann calls a “prolific domain” (in particular, the information-structural domain, “Ω”). Because movement of the topic is confined to a single prolific domain, it must be realized in such a way that both the moved copy in Spec,CP and the copy left behind in the topic position are spelled out separately. This is what gives rise to the emergence of the resumptive pronoun (“RP”) in the topic position: the copy of XP in Spec,TopP in (6b) must be vocalized (via late insertion) in order for movement within a single domain to be legitimate.² By contrast, in TOP constructions the fronted constituent moves to the topic position and stays there (see (6a)); no need for a resumptive element arises.

One might ask (as did one of our reviewers) what motivates movement of the topic within the single Ω-domain. Contrastiveness seems to us to supply a possible answer: the position in the Ω-domain that the topic reaches first is an “aboutness”-topic position, while the higher one is a contrastive topic position. This would be in line with Frascarelli and Hinterhölzl’s (2007) analysis of the German and Italian split-CP field. In their analysis, the locus for contrastive topics (CTopP) is above the position for familiar “aboutness”-topics (FamP); in principle, a topic may move first to FamP to license its [aboutness] property, and then on to CTopP to license its [contrastive] property. (To our knowledge, it is true that a contrastive topic in a CLD construction is always discourse-familiar.) If the topic phrase is derivationally related to two Ā-positions associated with complementary (rather than conflicting) features, the fact that it is spelled out in both positions is understandable. This picture meshes well with the case of sentence-medial resumptives, which are also in a noncontrastive topic position (see the discussion of (18) in section 3.1). It also explains why we cannot have both a sentence-medial and a sentence-initial resumptive: a blend of Dutch (1b) and (1bʹ) (*Die man,dieken ikdieniet ‘that man d-PRON know I d-PRON not’) would involve syntactically licensing the same noncontrastive topic property (familiarity or aboutness) twice within the same clause.

The prosody of the Hungarian CLD construction may lend further support to this analysis of the motivation for movement from Spec,TopP to Spec,CP in (6b). While the contour associated with the left-dislocated XP may be a rise (L* H), the contour associated with the resumptive cannot be. If (as seems highly likely) the possibility of being realized with a rising contour is specifically linked to contrastive topic status in Hungarian, then prosody tells us that the dislocated phrase is a contrastive topic, while the resumptive is not: it is an ordinary topic.

Because the topic, in both TOP and CLD constructions, originates inside TP and receives case there, the fact that both TOP and CLD exhibit case connectivity follows straightforwardly from the analysis in (6). Similarly, the presence of a silent copy of the topic-XP inside TP accounts for the binding connectivity effects.

Grohmann’s (2003) analysis needs just a small amendment to get HTLD under its purview. While in CLD constructions the left-dislocated topic arrives in Spec,CP via movement, in HTLD constructions it can be taken to be externally merged there, as in (5a). Not binding a trace inside the clause, the left-dislocate in HTLD generally exhibits no connectivity effects. Because the topic itself does not perform a role in the argument structure of the predicate of the clause, a proform is called upon to serve this purpose. The proform can optionally undergo topicalization to Spec,TopP.

In an interesting contribution to this journal, Ott (2014) argues in detail for an approach to CLD constructions that treats them as cases of clausal juxtaposition, as in (5c), with the first clause being stripped down to its contrastive topic via ellipsis.³ For Ott’s (2014:269) German example in (7a), this analysis is illustrated in (7b), where, as usual, strikethrough marks ellipsis.

The ellipsis approach handles many of the distinctive properties of CLD very handsomely. For instance, the case connectivity between the left-dislocate and its correlate follows from the fact that the left-dislocate originates in CP1 as the complement of the very same verb that takes the correlate den as its complement in CP2. Another interesting expectation that Ott’s analysis raises is that the resumptive element should in principle be allowed to be a topic quite different in form from the left-dislocated constituent. Ott (2014:273n7) mentions in this connection that epithets can serve as resumptives in CLD constructions, citing examples from German (see (8)), Dutch, and Icelandic. We could add here that Hungarian also allows epithets to resume a dislocated constituent (see (9a)).

For Hungarian, the fact that the resumptive epithet cannot occupy clause-internal position (see (9b)) may confirm that we are dealing with CLD here. A reviewer claims that there is no binding connectivity with an epithetic resumptive in German (whereas established cases of CLD, as we have shown, do display binding connectivity). But we doubt that this is systematic; in both Dutch (10) and Hungarian (11), bound variable connectivity is easy to get.

Left-dislocation constructions with an epithet as the resumptive that admit a CLD derivation pose a potentially serious question for a Grohmann-style movement approach, which treats the resumptive as a surface anaphor. Can an epithet be treated as a late spell-out of the feature bundle left behind in the topic position after the dislocated constituent has vacated this position? We do not consider it impossible to treat the epithets in (8)–(11) as surface anaphors; but there is no received wisdom in the late-insertion literature that would allow us to answer this question with confidence. In Ott’s approach, the existence of CLD with epithetic resumptives is directly expected. This will only plead in favor of clausal juxtaposition, however, if the analysis can limit the extent of the divergence between the topic in CP1 and the resumptive element in CP2. It is reasonable to suspect that there must at least be an interpretive link between the two elements. A hyponymy relation would a priori seem to be able to create such a link—and indeed, wa-marked topics in Japanese are famously allowed to be hyperonyms of an associate to their right (which may itself be wa-marked as well).

But hyponymy does not license CLD with a full-nominal “resumptive” in Germanic: Dutch (13a) is sharply deviant; a hanging topic introduced in an as for–type phrase must be used instead, as in (13b).⁴

We conclude, therefore, that in CLD constructions, hyponymy cannot license the presence of two different though related topics in the left peripheries of Ott’s juxtaposed CPs. This subtracts substantially from the prima facie appeal, in light of the epithet facts in (8)–(11), of the clausal-juxtaposition analysis.

Ott does not explicitly raise the question of whether TOP and HTLD could be analyzed along the lines of (5c) as well. The answer, a priori, is that they could—with movement of the topic being involved in the case of TOP (ensuring connectivity) but not in HTLD constructions, and with an overt resumptive element surfacing in CP2 in HTLD but not in TOP constructions. It is not obvious what a biclausal approach along the lines of (5c) could buy us for TOP constructions: everything we need to derive should be derivable in a single CP. So we will set (5c) aside as a viable analysis for TOP. But whether Ott’s clausal-juxtaposition-cum-ellipsis analysis could handle HTLD is an interesting question, particularly because HTLD shares with CLD what is perhaps Ott’s primary motive for developing a biclausal syntax for the latter: “verb third” word order. We now turn to this matter.

Both CLD and HTLD constructions violate the V2 constraint. The finite verb systematically appears in third position: the left-dislocated topic is followed by some constituent, which in turn immediately precedes the finite verb. That constituent can be the resumptive element (as in (14a)) or the subject (as in (14b)); it can even be some other major constituent of the clause, as shown in (14c).

Ott capitalizes on the verb-third (V3) problem at several points in his article to motivate the clausal-juxtaposition approach to CLD. His solution for this problem is clausal juxtaposition. Consider (15).

In CP2 in (15), there is just one constituent to the left of the finite verb; the left-dislocated topic is in an elliptical clause of its own and does not count as the occupant of the initial position in the V2 construction in CP2.⁵

HTLD constructions also exhibit a V3 pattern. So if Ott’s biclausal approach to CLD is to be the solution for V3 word order, it is imperative that it carry over to HTLD. If it turned out that the V3 problem for HTLD constructions could or even must be dealt with without an appeal to clausal juxtaposition, this would defeat a major portion of Ott’s rationale for the clausal-juxtaposition-cum-ellipsis analysis for CLD constructions. Let us examine, then, whether an approach along the lines of (5c) could be applied to HTLD.

Recall from section 1.1 that CLD and HTLD differ in that the latter but not the former can exhibit a lack of case connectivity, and that the former but not the latter is claimed to give rise to binding connectivity. The absence of binding and case connectivity would be difficult to reconcile with an analysis in which the left-dislocated topic in an HTLD construction originates local to its selector; it would follow from base-generation of the topic in its position in the high left periphery. So within Ott’s approach, one could try to capture the difference between CLD and HTLD by assuming that the elliptical TP inside CP1 in (15) contains a trace of the topic in CLD constructions but not in HTLD constructions. That would provide a common solution for the V3 problem shared by CLD and HTLD while still making the necessary distinctions for case and binding connectivity.

However, if we allowed the topic of CP1 to bind no trace inside TP, we would expect that when CP1 stands alone and has a nonelliptical TP, it should be able to contain a topic in the left periphery that does not bind a trace in TP and hence shows no connectivity. This expectation is not fulfilled: in ordinary topicalization constructions of the type in (1a), the topic always has the case that it would also have had if it were in clause-internal position. There is no grammatical alternative to (2a) (repeated in (16a)) in which the topic has default nominative case: (16b) is ungrammatical.

It seems impossible, therefore, to treat the left-dislocated constituent of an HTLD construction without case connectivity as the ordinary topic of an elliptical TP: doing so would massively overgenerate the lack of case connectivity beyond HTLD.⁶

The alternative is to analyze HTLD without an appeal to ellipsis. In (17), the left-dislocated topic is in a very high position in the left periphery of CP, one from which no V2 is triggered: “XP” is the specifier of a lower functional projection within CP (labeled “TopP” in (17), for concreteness), and the finite verb is raised to the head of that projection.⁷ The key difference between CLD and HTLD, on an approach along the lines of (17), lies in whether the topic arrives in Spec,CP via movement (as in Grohmann’s analysis for CLD constructions) or is base-generated there (which is what one would want to say for HTLD). The fact that CLD exhibits connectivity effects but HTLD does not can be directly related to the movement vs. base-generation dichotomy.

For the analysis in (17), solving the V3 problem posed by CLD and HTLD constructions comes down to ensuring that the C-position between the topic and the resumptive must remain silent. Ott rejects a Grohmann-style analysis of CLD precisely because it does not explain why this C must be silent. But it is not tremendously difficult to provide a descriptive perspective on the silence of C in (17) within a wider context. If highest-subject wh-questions in English are the result of movement of the subject from Spec,IP to Spec,CP (as is standardly assumed in the literature, even in Chomsky’s (1986) discussion of the Vacuous Movement Hypothesis), a generalization covering both Who left? vs. *Who did leave? and the silence of C in (17) can be couched in terms of the contiguity of the extraction and landing sites of movement: whenever what moves to Spec,CP is separated from its lower copy by just the C-head, this head must be silent. This of course is just a descriptive generalization—and one that would have to be tweaked to carry over to HTLD (which involves base-generation of the topic in Spec,CP, in light of the absence of connectivity). But once we establish a link between the silence of C in (17) and the silence of C in highest-subject root wh-questions, the fact that CLD constructions exhibit V3 word order does not seem fatal for a Grohmann-style analysis.

If there is indeed at least one position for topics that is not immediately followed by the finite verb in root contexts in the Germanic V2 languages, the V3 problem no longer compels us to embrace an Ott-style clausal-juxtaposition approach to CLD.

In Grohmann’s analysis of CLD, the sole purpose of the resumptive element is to salvage the otherwise illegal domain-internal movement of the topic. In this analysis, the surface anaphor is a copy of the left-dislocated constituent, its spell-out being tied directly to the very local movement of the dislocate, within a single prolific domain. Any occurrences of such resumptives in positions that are not local to the dislocate thus stand out as potential problems for the analysis. There are two situations in which such nonlocal resumption could arise, in principle. In the first, we are dealing with a single clause and the resumptive occurs somewhere inside the clause, not in its left periphery. In the second, we are dealing with biclausal constructions in which the left-dislocate is upstairs and the resumptive is in the left periphery of the embedded clause.

We have already shown cases of the first type: in the Dutch example (1bʹ) (repeated here as (18)), the d-pronominal resumptive does not obviously appear in the Spec,TopP position in (17).

But although the resumptive seems to be quite a distance away from the topic, it clearly has scrambled across niet ‘not’. And since the subject of (18) is itself a topic as well, we could devise an analysis of (18) that is compatible with Grohmann’s idea that the resumptive is spelled out for “antilocality” reasons, by placing ik and die each in the specifier position of its own TopP in the left periphery, and pronouncing the finite verb in the higher of the two Top-heads, as in (19).

Since the copy of XP in Spec,TopP2 is within the same prolific domain as its antecedent, spelling this XP out as a resumptive element is, by the logic of Grohmann’s analysis, just as obligatory in (19) as it is in (6b). Linear orders of the type in (18), in which the phrasal material intervening between the topic and the resumptive is itself topical, can thus be accommodated by Grohmann’s system.

The trouble with low resumptives becomes more serious, however, when we include sentences of the type in (20) in the discussion.

In these examples, the subject is a focus (marked by the focus particle zelfs ‘even’ and by focus prosody)—and in (20b), this focal subject linearly intervenes between the left-dislocated topic zo’n auto ‘such a car’ and the resumptive d-pronoun die. Now we cannot avail ourselves of TopP recursion. We could certainly place the subject of (20b) in a high focus position, and the resumptive in a relatively low topic position. But the facts from a variety of languages whose information-structural left peripheries have been studied in detail (especially Hungarian and Italian) have made it clear that within a single functional domain, FocP is always below TopP (see Puskás 2000, Belletti 2004, Benincà and Poletto 2004, Rizzi 2004, paceRizzi 1997). Though foci can occur to the left of topics, it is standard in the cartographic literature to take the FocP and TopP in such cases to belong to different structural domains: the FocP occurs outside TP, and the low TopP belongs to the VP’s left periphery.

The structure in (21) could accommodate the linear order of (20b) by placing the left-dislocated topic zo’n auto ‘such a car’ in the specifier position to the left of (silent) C, the focused subject in Spec,FocP, the finite verb in Foc, and the resumptive d-pronoun in the lower Spec,TopP. The CP whose specifier is occupied by the left-dislocated topic and the FocP whose specifier is occupied by the focused subject belong to the same functional sequence, and to the same prolific domain. The question for Grohmann’s analysis would be whether the TopP where the resumptive is spelled out belongs to this prolific domain as well.

The TopP below T is in a different functional sequence: it belongs to the left periphery of VP, not to that of TP. But this TopP is of the same nature as the functional projection harboring the left-dislocated topic (CP): both are discourse-related projections. As long as membership in a particular prolific domain is based on a projection’s nature and not on its position in the clause, we can maintain that CP and the lower TopP in (21) belong to the same prolific domain: the discourse domain (labeled “Ω” by Grohmann) of the root clause.⁸ If movements linking two positions included in the same prolific domain for a particular clause require both members of the chain to be spelled out, even if the positions in question are linearly far apart and do not belong to the same functional sequence,⁹ the obligatory pronunciation of the resumptive in (20b) can be accounted for.

Now, how does Grohmann’s approach to CLD fare in cases of biclausal sentences in which the left-dislocate appears upstairs and the resumptive is spelled out in the left periphery of the embedded clause? We will explore this question by juxtaposing Ott’s and Grohmann’s analyses.

We begin by constructing an example in which a topicalized element occurs to the left of the subject in an embedded clause. This is much harder in Dutch and German than it is in English—but it can be done, via what has been called “focus scrambling” (Neeleman 1994), as in Dutch constructions of the type in (22). Here the scrambled element is a contrastive topic, just as in CLD constructions.¹⁰

The examples in (20) already showed that zo’n auto ‘such a car’ can undergo CLD in the presence of a focus. Putting (20) and (22) together, we can now ask (a) whether CLD of zo’n auto within the embedded clause to a position to the left of the d-pronoun is possible, and (b) whether the topic can be placed in the matrix clause, leaving behind a resumptive d-pronoun inside the embedded clause.

The answer to the first question is that with the ordinary verb-final word order of subordinate clauses, the result is ungrammatical, as shown in (23).

The second question is answered by (24): long-distance left-dislocation of zo’n auto is grammatical as long as V3 order obtains in the root clause, as in (24b).¹¹

The key contrast is the one between (23) and (24b).

For Ott’s (2014) clausal-juxtaposition-cum-ellipsis analysis, accounting for this contrast is quite simple. (23) is underivable, as it would have to involve embedding the juxtaposed CPs below the complementizer dat ‘that’; such “CP recursion” structures are generally impossible with verb-final order.¹² In (24), CP-juxtaposition proceeds in the root clause, with the resumptive d-pronoun contained in the second CP, as depicted in (25).

The well-formedness of CP1 is entirely on a par with that of the long-distance topicalization case in (26). The fact that the left-dislocate is followed by the subject of the matrix clause (as in (24b)) rather than by the finite verb (as in (24a)) follows from the fact that CP2 is a subject-initial V2 construction. The placement of the d-pronoun to the left of the embedded subject is made possible by focus scrambling.

For Grohmann’s movement approach to CLD, a central question posed by (24b) is where the topic is launched from. Note that (24b) features embedded-clause word order, with the verbs clustered at the end of the string. CLD within a subordinate clause is not possible with verb-final order: (23) is ungrammatical. So the embedded clause of (23) cannot underlie that of (24b): we cannot take the topic to be left-dislocated within the embedded clause prior to its left-dislocation in the matrix clause. But a topic position is clearly available to the left of the focused subject of the embedded clause: the resumptive d-pronoun is spelled out right there in (24b). So we could in principle launch the left-dislocated constituent from the topic position of the subordinate clause.

Movement of the left-dislocate from the topic position of the lower clause should deliver the pronunciation of a d-pronoun in that lower clause.¹³ Under Grohmann’s approach, it can do so only if in the course of the derivation of (24b) the left-dislocate undergoes a movement entirely confined to one prolific domain within the embedded clause, prior to being extracted from that clause. On standard assumptions, movement from a finite clause proceeds via the Spec,CP position of that clause, leaving an intermediate copy behind in that specifier position. Movement from Spec,TopP to Spec,CP would certainly be very local. But does it take place within a prolific domain? The answer depends on whether the Spec,CP position in which an intermediate stopover is standardly taken to be made on the way out of an embedded clause counts as belonging to the discourse-related Ω-domain of that clause. If it does, the derivation of (24b) can proceed as in (27), which, by the logic of Grohmann’s work, will deliver the obligatory resumptive d-pronoun to salvage the movement within a single Ω-domain.¹⁴

While this accounts for the obligatory emergence of the resumptive in (24b), it does not yet explain the contrast between this example and its close counterpart in (24a). This contrast can be understood in light of the fact that the information-structural status of a particular constituent is specified once only in the course of the syntactic derivation. In the derivation in (27), representing (24b), the DP zo’n auto ‘such a car’ is identified as a topic in virtue of its move to the Spec,TopP position of the subordinate clause. Onward movement into the matrix clause must not add information-structural baggage to this DP; so precisely the only position in the matrix clause where zo’n auto is welcome to be spelled out is a Spec,CP position that is not given information-structural content, as in (24b). The redundancy of movement of zo’n auto from a Spec,TopP position in the embedded clause to a Spec,TopP position in the root clause (which would result in subject–finite verb inversion in the root) could deliver the marginality of (24a), compared to the woefulness of (23). This may reflect well on Grohmann’s analysis of CLD. It is unclear how Ott’s clausal-juxtaposition approach could capture the relative contrast between (23) and (24a): the latter seems radically impossible without resorting to scattered deletion in (25) ( preserving denk ik in CP1 and eliding ik denk in CP2).

To support the idea that prolific-domain-internal movement takes place within the subordinate clause in CLD constructions in which a clause boundary separates the dislocate and the correlate, we turn to Hungarian. First, we need to show that Hungarian has CLD. Example (28) shows this: it features a demonstrative correlate, and connectivity effects for both case and binding, very much as in Germanic. And just like Germanic, Hungarian allows the dislocate and the correlate to be separated by a clause boundary. We see this in (29); here, along with case connectivity, there is also connectivity for definiteness between the left-dislocate (indefinite rossz diákot ‘bad students.ACC’) and the predicate of the embedded clause (vesznek ‘admit’); see section 5.4 for more discussion.

The fact that the dislocate and the correlate in (29) are not within the same local domain on the surface can be accommodated within Grohmann’s analysis if in the embedded clause the dislocate moves from TopP to CP (within the same prolific domain), and if it is the (silent) copy of the dislocate in the embedded CP that triggers the PF rule that requires the copy in the embedded TopP to be spelled out. That the dislocate originates within the embedded clause and undergoes syntactic movement from the lower clause into the higher clause is shown by the emergence of island effects of the type in (30a), parallel to what we find under long-distance topicalization (as in (30b)).¹⁵

That the dislocate is indeed present at some derivational stage in the lower CP (right above the demonstrative in the left periphery of the subordinate clause) in sentences like (29) is also shown by the following pattern in infinitival embedding, where (31b) is the key example:

Infinitival complement clauses allow in principle for a topic within their bounds—even an anaphoric demonstrative one, as shown in (32). But they do not permit the simultaneous presence of both a dislocate and a correlate in their left periphery. This is demonstrated by (33B). Against this background, the ungrammaticality of (31b) suggests that when the dislocate and its correlate are severed by a clause boundary, the dislocate must topicalize within the lower clause and then move on to the local Spec,CP—which is possible in the finite clause in (29) but not in the infinitival one in (31b) (in light of (32B)), nor in (30a), with an adjunct island.

If all apparent cases of a nonlocal relation between the dislocate and its surface-anaphoric correlate can be accounted for along the lines pursued here, such that the dislocate and the correlate are in fact in the same prolific domain at the relevant stage in the derivation (sanctioning the spell-out of the correlate), the “antilocality” approach to surface-anaphoric correlates in CLD constructions, as in Grohmann 2003, will be upheld. For Ott’s (2014) clausal-juxtaposition approach, by contrast, it is not clear how the facts canvassed in this section can all be made to fall into place. We have already noted the relative contrast between (23) and (24a); the robust contrast between Hungarian (31b) and (31a)/(32B) seems particularly difficult to fit in with Ott’s analysis.

Ott likens the ellipsis process whereby the clausal chunk following the fronted topic is rendered silent to the ellipsis standardly postulated in the analysis of sluicing and fragment answers. For sluicing and fragment answers, the constituent marked for ellipsis is customarily identified as IP (or TP). Referring to work in progress, Ott (2014:270n2) suggests that “‘IP-ellipsis’ is a misnomer and that what undergoes deletion in all relevant cases is in fact the derived sister of the fronted operator”; he chooses to “use the noncommittal term clausal ellipsis . . . to designate the relevant ellipsis pattern.”

It is by no means easy to perform clausal ellipsis after a fronted topic, however.¹⁶ We will illustrate this on the basis of data from both Dutch and Hungarian, the latter being famously transparent when it comes to the syntax of the left periphery. Consider first the examples in (34) and (35), from Dutch and Hungarian, respectively. Here, clausal ellipsis following the contrastive topic in the second clause is grammatical—thanks to the fact that the topic here contrasts explicitly with the topic of the previous clause. The role of explicit contrast is brought out by the contrasts between the (a)- and (b)-examples in (36) and (37). Here, ellipsis after jou ‘you’ or neked ‘you.DAT’ is impossible. At least one alternative to the contrastive topic must be explicitly provided in the discourse in order for clausal ellipsis to be licensed; in the absence of explicit contrast, the sentence must be pronounced in full, as in the (b)-cases.

This requirement of explicit contrast imposed on the licensing of clausal ellipsis in topicalization constructions is conspicuously not in effect in CLD constructions. The fronted topic in CLD does not need to be explicitly contrastive: (1b), for instance (repeated as (38)), does not have to be preceded or followed in the discourse by a statement in which die man ‘that man’ is contrasted with some other topic. If CLD is to represent the left-dislocate as the topic of an elliptical clause, as in Ott’s analysis, the licensing of ellipsis must be more lenient here than in established cases of clausal ellipsis in contrastive topicalization constructions.

The ellipsis problem is compounded by the fact that in (38) there are two topics, each of which, in Ott’s analysis, is placed in the left periphery of its own clause: die man ‘that man’ in the elliptical CP1, and die, its correlate, in CP2. But although the two topics are presumably in the same structural position in their respective clauses, ellipsis is licensed only after the first. In (39), from Dutch, and (40), from Hungarian, the content of the TP is fully recoverable from speaker A’s statement. But despite its recoverability, and even though eliding TP after the topic is possible in a simple topicalization construction (recall (34) and (35)), ellipsis of TP is not allowed after the correlate in (39Bʹ,Bʺ) and (40Bʹ,Bʺ), regardless of whether there is ellipsis after the left-dislocate. The ungrammaticality of the utterances by speakers B' and B" presumably has the same source. Under Ott’s assumptions, that source remains unclear.

Contrasts between the distribution of sluicing (the standard case of TP-ellipsis) and contrastive left-dislocation abound. The ungrammaticality of ellipsis in (41B) and (42B) shows that in these embedded contexts (clauses embedded under factive verbs and ‘if’-clauses), sluicing is disallowed—both after contrastive topics (as shown in (41B)) and after foci (the more typical case of sluicing; see (42B)). But such embedded clauses do allow CLD, as the examples in (43) (corresponding to (41B) and (42B)) show.¹⁷

The following examples make the same point in the opposite direction. Infinitival complements are a structural domain in which a left-dislocated topic cannot occur (as already shown in (33), repeated as (44)); but sluicing (as in (45)) is fine here.¹⁸

It is interesting to note in this connection that left-dislocation of foci is grammatical in embedded infinitivals (and is then on a par with sluicing, as in (45)), as illustrated in (46).

Problematic for Ott’s approach to CLD, therefore, is that it postulates ellipsis in a context in which ellipsis is ordinarily impossible or restricted in ways that do not match the restrictions imposed on CLD constructions. Explicit contrast is normally a necessary condition for clausal ellipsis in contrastive topicalization constructions—but no such condition generally holds for CLD. And even when there is explicit contrast, CLD patterns distinctively differently from contrastive topicalization constructions with sluicing. What is needed to make Ott’s analysis fully operational is some explanation for the fact that contrastive topics that have a correlate in a clause that is juxtaposed to it sanction ellipsis of the TP that follows them under circumstances that are systematically different from the circumstances under which clausal ellipsis after contrastive topics is otherwise licensed. In the absence of such an explanation, the ellipsis approach to CLD becomes sui generis.

Also problematic for an ellipsis approach to left-dislocation is that sluicing and fragment answers (the clausal ellipsis phenomena that, by Ott’s own admission, are most directly akin to what is postulated in (2)) behave very differently from CLD when it comes to the possibility of multiple remnants. Once again, we will illustrate with Dutch and Hungarian data. This time, each language presents a different pattern—but importantly, in both languages CLD is consistently different from fragment answers and sluicing constructions, which are the standard cases of clausal ellipsis in the literature.

Consider first the contrast in Dutch between the sluicing and fragment answer examples in (47) (also grammatical in English, as the prose translations show) and the CLD cases in (48).

Dutch being a V2 language, (48a) is trivially ungrammatical because, in the clause following the comma, the finite verb comes in third place, following the two fronted correlates of the dislocates. But even when we leave the correlate of the PP-dislocate in situ, as in (48b), the result is ungrammatical. The deviance of (48b) contrasts markedly with the immaculate status of the bona fide ellipsis constructions in (47), where both the subject and the met-PP are successfully preserved under ellipsis. The contrast between (47b) and (48) is particularly striking. An ellipsis approach to CLD as in (5c) could account for this contrast only if it found a way to constrain ellipsis in the first CP of CLD constructions more stringently than in the case of question-answer pairs like (47b). It is entirely unclear how this could be achieved.

The two versions differ with respect to the placement of the demonstrative correlate of the first topic vis-à-vis the second topic. The difference in placement of azt goes hand in hand with a difference in interpretation of the sentence, which we have tried to bring out in the English paraphrases.¹⁹ (49a) features two independent left-dislocated topics and concomitantly gives rise to a recursive information structure.

So the sentence cannot mean that among pairs of individuals—say, twins—Mari and Juli are a pair of individuals who I have never confused with each other (although perhaps I have already confused another pair of individuals). In (49b), on the other hand, such a nonrecursive information structure, with a single pair of topics, is precisely what we do get.

Under Ott’s approach, (49b) is not difficult to analyze: Marit Julival ‘Mari.ACC Juli.COMIT’ can be treated as a pair of topics in the left periphery of the elliptical CP1, and azt azzal ‘DEM.ACCDEM.COMIT’ is the corresponding pair of correlates in the nonelliptical CP2.

The question is how the recursive information structure of (49a) can be accounted for, with the left-dislocated topics in different CPs (as they must be, on Ott’s assumptions). Ott’s analysis could treat (49a) in either of two ways. In the first, illustrated in (52a), each of the four constituents preceding még ‘yet’ is in a CP of its own, with CP1 and CP2 forming one pair of juxtaposed clauses and CP3 and CP4 another; ellipsis targets the IPs of all but the last of the four CPs. In the alternative, schematized in (52b), Marit is in an elliptical clause juxtaposed to a second clause featuring azt and Julival in the topic field and ellipsis following Julival, and that second clause in turn is juxtaposed to a third that has azzal in its left periphery, followed by the comment még nem kevertem össze ‘never yet confused’. Neither of these analyses delivers the information structure that (49a) has (as given in (50)).

There are additional problems as well. For (52a), one concern is that the CP1+CP2 clausal-juxtaposition pair features ellipsis of the TP in both clauses, something that is otherwise impossible, as shown earlier: recall (40B'). And as (40B') and (40B") taken together show, TP-ellipsis following a correlate demonstrative in a CLD construction fails generally. So the fact that azt ‘DEM.ACC’ in (52a) is followed by an ellipsis site is doubly problematic for this analysis. What hobbles both (52a) and (52b) is that the elliptical clauses are not (all) identical to one another or to the nonelliptical clause. And the forced postulation of an object-pro (coreferential with the topic of CP1 and the correlate pronoun in CP2) in the nonelliptical clause is problematic as well: Hungarian third person object-pro is limited to singulars (see Farkas 1987), but CLD constructions of the type in (49a) are grammatical with a plural left-dislocated object, as in (53).

The contrast between Dutch (48) and Hungarian (49) shows that languages clearly differ with respect to the availability of multiple CLD. We have not investigated this beyond the Dutch/ Hungarian contrast, but it seems to us likely that this variation is correlated with the availability of multiple (ordinary) topics: Hungarian allows these; Dutch does not. If indeed multiple CLD turns out crosslinguistically to imply multiple topicalization, Grohmann’s approach allows us to understand this: for his analysis to exclude (48) and rule in (49), one only needs to assume whatever one’s favorite syntactic mechanism is to model single/multiple fronting parameters in general. But Ott’s approach to CLD harbors little hope of capturing such an implicational relation: multiple topicalization is multiple fronting to the left periphery within one clause, but multiple CLD would, in this approach, involve multiple clausal juxtaposition.

The problems with ellipsis we have identified in this section undermine one of the pillars of Ott’s clausal-juxtaposition-cum-ellipsis analysis of CLD. Though Ott’s approach may very well be conceptually attractive, its appeal to clausal ellipsis would have to be fine-tuned very carefully in order for it to successfully model the syntax of CLD—and we doubt that it can be fine-tuned entirely satisfactorily.

The other pillar that Ott’s analysis rests on is the postulation of clausal juxtaposition. Ott explicitly seeks a parallel between CLD and clausal coordination: the two CPs that together constitute the CLD construction are in a relationship of asyndetic coordination ( parataxis). This parallel breaks down in at least two contexts. The lack of parallelism between CLD and coordination reveals in mutually reinforcing ways that the d-pronominal resumptive in CLD constructions is a surface anaphor (in the sense of Hankamer and Sag 1976) while its counterpart in coordination constructions is a deep anaphor.

Ott (2014:275) states that CLD of their complements disambiguates modals toward a deontic reading in German. We doubt that this is a general fact about CLD of the complement of a modal: in Dutch (54), the epistemic reading (made salient by haast wel ‘pretty much’) is easily available. But in the coordination examples in (55), involving a d-pronoun in the second conjunct but no CLD, the epistemic reading is very hard to get. Since the use of haast wel gives the modal a salient epistemic reading in the first conjunct, the unavailability of that reading in the second conjunct renders (55a–b) very awkward (regardless of whether the d-pronoun is topicalized in the second conjunct, as in (55a) (on a par with (54)), or not).

That (55a–b) are awkward is not difficult to understand: dat in the second conjunct can only be a deep anaphor (in the sense of Hankamer and Sag 1976), and when a deep anaphor represents the complement of a modal in Dutch, the modal’s interpretation is always deontic. Speaker B’s responses to A’s statement in (56) are absurd since we cannot plausibly give the weather an order or attribute a deontic ability to it.

The interesting thing is that (54) is not impossible. This is the first of a number of indications that the d-pronoun in CLD constructions is a surface anaphor. On Ott’s analysis of CLD, in which the d-pronoun and its associate are in different conjuncts, the fact that the d-pronoun can be anything other than a deep anaphor is highly surprising.

Ott (2014:292) claims that floating of argumental material under contrastive left-dislocation is impossible, supporting his case with a German example involving a clausal complement. But Dutch clearly allows argumental material of the fronted verb to be floated in the matrix clause, as in (57).²⁰ Hungarian CLD does, too: the Hungarian equivalent of (57a), given in (57a'), is grammatical as well.²¹

As Dutch (58a) and its Hungarian counterpart in (58b) clearly demonstrate, it is entirely impossible to have internal arguments floating around demonstrative pro-predicate dat/azt in coordination constructions, which Ott argues throughout his article to be the closest relatives to contrastive left-dislocation constructions.

In allowing argument float, CLD is very different not just from coordination but also from both HTLD and right-dislocation (RD): as (59a) and (60a) show, HTLD and RD both resist floating. (The (b)-examples are included to show that without argument float these HTLD and RD constructions are fine.)²²

Here again, the difference between deep and surface anaphors rears its head. When dat and azt ‘that’ are used as deep anaphors, they do not allow internal arguments to float around them.

There is something morphologically peculiar about the demonstrative correlates of indefinite dislocates in Hungarian CLD constructions that suggests once again that these are surface anaphors. In (62a), the inflection on the verb in the clause containing the correlate demonstrative is indefinite: ismerek rather than ismerem. In the absence of an indefinite left-dislocated topic, the demonstrative azt always brings about definite inflection on the verb. In the CLD construction in (62b), whose left-dislocated object is definite, we find definite inflection on the verb in the correlate clause. And in the simple example in (62c), the definite inflection is again a straightforward reflection of the demonstrative’s inherent definiteness. So it is precisely in CLD constructions whose contrastive topic is indefinite that we see azt combine with indefinite inflection. The choice of inflectional form of the verb in the correlate clause is apparently a function of the definiteness not of the correlate demonstrative itself but of its associate (i.e., the topic).²³

For Ott’s analysis of the CLD construction, this presents a conundrum. In this analysis, the demonstrative correlate is a deep anaphor, just as it is in clausal coordination constructions. But in coordination constructions, the demonstrative behaves differently: in (63), azt can assert its definiteness in the familiar way in the form of the verb.²⁴

Ott’s clausal-juxtaposition-cum-ellipsis approach could assign (62a) the structure in (64).

But to get from (64) to the desired surface result in (62a), one would need nothing short of a miracle: the verb of the sluiced TP in CP1 (ismerek ‘know’) would need to be preserved and copied over to the right, replacing the verb of CP2. It seems clear that the clausal-juxtaposition-cum-ellipsis approach cannot handle (62a).

The facts reviewed above are understandable if (a) the left-dislocated topic really is a constituent of the correlate clause at some level of analysis, and (b) the demonstrative proform in CLD constructions is a surface anaphor—an element that is not part of the numeration of the clause in which the correlate is spelled out but instead a piece of morphophonological matter that makes its first appearance in the PF component, after vocabulary insertion. If (a) and (b) are both met, what we get in the underlying representation of CLD constructions is effectively the same structure as that underlying topicalization constructions whose topic is moved from a clause-internal position into the left periphery.

Grohmann’s analysis of CLD directly embodies the treatment of the demonstrative correlate as a surface anaphor. On his approach, the difference between CLD and ordinary topicalization (TOP) lies in the fact that in CLD but not in TOP constructions, the topic moves from the topic position into a higher position in the left periphery. Floating of arguments, as in (65a), is then on a par with what has long been known to be possible in VP-topicalization constructions (see Den Besten and Webelhuth’s (1987, 1990) discussion of remnant topicalization); see (65b).

And the facts in section 5.3 fall out because the morphosyntax derivation and its continuation into LF do not have a demonstrative proform in them at all.

One property of CLD that stands out as potentially problematic for a movement analysis is P-stranding. In the present section, we will show that P-stranding actually poses problems for both of the major approaches to CLD represented by (5b) (instantiated by Grohmann’s proposal) and (5c) (Ott’s analysis), but that while (5b) stands a decent chance of overcoming this hurdle, the ellipsis approach has no way out for CLD—although it actually makes exactly the right predictions for right-dislocation.

Ott (2014:sec. 4.2) states that the ellipsis-based approach to CLD correctly predicts that Merchant’s (2001) P-stranding generalization, reproduced in (66), holds also for CLD.

But while apparently obeyed by CLD in several other Germanic languages (Ott gives examples from German, Icelandic, Norwegian, and Swedish), the P-stranding generalization is not obeyed by Dutch CLD: (67a) is perfectly grammatical even though, in Ott’s ellipsis approach, P-stranding movement would be involved in the derivation of the first CP (see (67b)).

Ott (2014:286n29) is aware of this, and downplays the problem with a two-pronged approach: (a) he points out that P-stranding in Dutch, while generally poor with anything other than Van Riemsdijk’s (1978) R-pronouns (like daar ‘there’ in (67a)), is relatively good with non-R material in topicalization contexts; and (b) he notes that because proper names do not inflect for case in Dutch, Peter could, for all we know, be a default nominative, which would make (67a) a case of HTLD, not CLD.

With respect to the first point, we certainly agree that the ordinary topicalization variant of (67a), given in (68a), is relatively acceptable. But this misses the point: (67a) is not just relatively acceptable, it is perfect. So CP1 in (67b), the Ott-style derivation of (67a), still behaves markedly differently from a case of ordinary topicalization of a proper name with P-stranding. The plot thickens when we substitute a pronoun for the proper name in (68a), as in (68b): while topicalization of a proper name with P-stranding is relatively good, pronouns do not allow this at all. This makes a derivation of (68c) qua CLD along the lines of (67b) problematic: after all, the nonelliptical version of CP1 for (68c) is ungrammatical (cf. (68b)). Moreover, in (68c), the only option is the explicitly case-marked form hem; the (default) nominative hij is impossible. This casts doubt on an alternative analysis of (68c) in terms of HTLD: (68d), an unmistakable case of HTLD (with the associate in situ), is fine with a nominative topic pronoun.²⁵

For Grohmann’s movement analysis of CLD, deriving (67a) is not strictly speaking impossible, but doing so will require a PF-based approach to the distribution of R-pronouns: Peter can extract from PP in syntax as long as, at PF (the earliest point at which the surface anaphor daar emerges), the immediate antecedent of the PP-internal trace is an R-pronoun. One could herald as an achievement of a surface-anaphor analysis of CLD the fact that it forces such a PF-centric account of R-extraction—an account that would fit in well with a late-lexical-insertion approach.²⁶

In CLD constructions, R-words can serve as contrastive topics in P-stranding contexts, as shown in (69).

For a Grohmann-style movement-cum-surface-anaphora approach to CLD, the grammaticality of (69) is unremarkable. On an Ott-style analysis, by contrast, (69) presents a major conundrum.

Under Ott’s approach, the derivation of (69) would be represented as in (70), with the R-word daar as an ellipsis remnant in a P-stranding environment.

But in well-established TP-ellipsis contexts, it is surprisingly difficult to preserve R-words as remnants when the elliptical constituent contains a stranded P; see (71) and (72).

Locative waar is a legitimate remnant of sluicing in sentences of the type in (71a); but when the R-word strands a preposition in the elliptical clause, as in (71b) and B’s utterances in (72), the result is sharply deviant. To make these ellipsis cases grammatical, the preposition must be piedpiped and thereby spared under clausal ellipsis. The contrast between (71b) and (71c) (for which the ellipsis literature to date has not found an explanation; see Van Craenenbroeck 2013 for discussion) is all the more striking because when no ellipsis takes place, P-stranding is preferred to pied-piping: the nonelliptical version of (71b) is better than that of (71c).

On Ott’s approach to CLD, the left-dislocate in (69) is the R-word remnant of an elliptical clause that contains a stranded preposition, as in (70). This structure is the same as that of the ungrammatical sluicing construction in (71b) and of the P-less versions of B’s utterances in (72). The fact that (69) is good while (71b) and P-less (72) are not is entirely unexpected on an Ottstyle approach to CLD.

As Ott (2014:283n25) suggests in passing and as Ott and De Vries (2016) argue explicitly, the clausal-juxtaposition-cum-ellipsis analysis can be applied to right-dislocation (RD): for a sentence like (73a), the right-dislocated element can be analyzed as the remnant of clausal ellipsis in a clause juxtaposed (this time on the right-hand side) to the clause containing the correlate pronoun, as shown in (73b).

Ott and De Vries point out that RD observes the P-stranding generalization in many languages. Indeed, in Dutch, RD robustly reveals the kinds of restrictions that well-established ellipsis cases impose on P-stranding in the language: (74a) (cf. (67a)) and (74b) (cf. (69)) are both ungrammatical without the preposition repeated in front of the right-dislocated constituent.

In this respect, RD in Dutch is robustly different from both CLD and HTLD: as shown earlier in this section, neither CLD nor HTLD resists a P-less constituent in sentence-initial position in the presence of a P-dependent correlate.²⁷

The sharp unacceptability of the P-less versions of (74a–b) and the consistency with which other languages obey the P-stranding generalization in their RD constructions is a strong selling point for a clausal-juxtaposition-cum-ellipsis approach to RD, such as that of Ott and De Vries. This analysis ties RD together with several other constructions involving forward ellipsis (with the ellipsis site following the antecedent), including sluicing and Den Dikken, Meinunger, and Wilder’s (2000) Type A specificational pseudoclefts.²⁸

Some forward ellipsis cases can be reversed (i.e., turned into backward ellipsis constructions): (76a) is grammatical alongside (75a). But not all of them can. While (76b) may be acceptable without ellipsis (as a copular amalgam sentence; see O’Neill 2015 for in-depth discussion of these kinds of sentences), eliding Nobody was eating makes it entirely impossible.

The contrast between (75b) and (76b) is by no means specific to English: as far as we are aware, in all languages that allow negative-polarity-item (NPI) connectivity in specificational pseudoclefts, this is subject to a strict word-order restriction. Backward ellipsis of clausal material²⁹ thus seems more restricted than forward ellipsis of such material—especially so if the constituent spared by ellipsis is not [+wh]. Since in CLD constructions the left-dislocated constituent is [–wh], a backward clausal-ellipsis approach to CLD would seem difficult to maintain in light of the problems encountered with backward clausal ellipsis with [–wh] remnants elsewhere.

For RD, this discussion has no adverse consequences for an Ott-style analysis because it involves forward ellipsis. But for CLD it casts significant doubt on the feasibility of an ellipsis approach. Since RD and CLD (as well as HTLD) also behave sharply differently with respect to the effects of the P-stranding generalization, this raises the suspicion that CLD presumably is not optimally analyzed in terms of clausal ellipsis.

At this point, let us take stock. We started in section 1 by laying out the hypothesis space (repeated below) for the syntax of CLD and closely related phenomena. Throughout this article, we have looked at two detailed analyses of CLD—Grohmann’s (2003) movement-cum-surface-anaphora analysis along the lines of (5b), and the clausal-juxtaposition-cum-ellipsis approach put forth by Ott (2014), representing (5c). We have shown that clausal juxtaposition, while quite plausible as an approach to right-dislocation, faces a variety of challenges, which the movement-cum-surface-anaphora analysis has little or no trouble with. Are there any empirical or conceptual considerations that would lead us to conclude that (5c) is necessary, as a member of the spectrum of analyses for CLD?

Empirically, the answer to the previous question seems to us to be negative. For clausal juxtaposition, we have been unable to find explicit and incontrovertible evidence in CLD. The V3 problem, which Ott advances prominently as an argument for it, turns out upon closer inspection not to be a major obstacle for (5b): it is possible, at least on a descriptive level, to account for the silence of the C-head between the left-dislocate and the demonstrative correlate by capitalizing on the contiguity of the two referentially linked constituents. The account of the silence of C seems principled enough, which suggests to us that the fact that CLD constructions exhibit V3 word order is not a decisive empirical argument for clausal juxtaposition as an ingredient of the syntax of CLD. As for the ellipsis operation that (5c) would need to resort to, we have shown that it would have properties that make it (and the analysis employing it) sui generis.

Conceptually, (5c) is eminently sensible as an approach to dislocation phenomena. There is no doubt that clausal juxtaposition (analyzable as asyndetic coordination) and clausal ellipsis exist. The general building blocks of Ott’s analysis of CLD are so uncontroversial that it would seem impossible to radically block a derivation of at least some dislocation constructions along Ott’s lines. And indeed, clausal juxtaposition cum ellipsis may work for right-dislocation. But although clausal ellipsis indubitably exists, it has become clear that it is not available precisely in the context in which Ott’s analysis of CLD needs it to be: after contrastive but not necessarily explicitly contrastive topics. The circumstances under which clausal ellipsis would have to be sanctioned in (5c) seem to be systematically different from the circumstances under which clausal ellipsis after contrastive topics is otherwise licensed.

Nothing in the data seems to unequivocally recommend the approach in (5c). What (5c) does well can also be handled by (5b); the movement-cum-surface-anaphora approach manages to handle important facts that (5c) cannot deal with, and it does not make the kinds of incorrect predictions that (5c) gives rise to.

What is left to discuss is whether the syntax of CLD constructions needs to avail itself in any way of (5a), an analysis in which the topic is base-generated in the left periphery, linked to a deep anaphor lower in the structure. An analysis along these lines is the gold standard for HTLD. It accounts naturally for the absence in this construction of connectivity for morphological features and binding dependencies. These connectivity effects are commonly taken to define the distinction between CLD and HTLD: whenever we find obligatory connectivity for morphology and binding, HTLD cannot be involved. In our view, it remains an open question whether any and all left-dislocation constructions that fail to exhibit morphological connectivity are necessarily not cases of CLD, and must instead involve HTLD. We will present one brief case study in closing, to illustrate why we think it will be good to keep an open mind: the (lack of) morphological connectivity in CLD constructions involving a left-dislocated verb phrase.

Referring to Danish facts brought to light by Mikkelsen (2011) (which we will not repeat here for reasons of space), Ott (2014:sec. 4.3) claims that the morphology of a contrastively left-dislocated VP must show connectivity with the selector of the correlate proform: when the correlate is selected by an auxiliary of the perfect, the verb heading the dislocated VP must have past-participial morphology. But it would be false to claim, and therefore wrong to try to derive, a general morphological connectivity constraint on left-dislocated VPs. In Dutch cases of fronting of the complement of verbs like proberen ‘try’, which normally select only te-infinitives, the infinitival marker te can be dropped, as in (77b) (Zwart 1993:263; for some speakers, the first author of this article included, dropping te is the preferred option). This is true both in ordinary VP-topicalization and in cases of left-dislocation with a d-word as the correlate.

The grammaticality of (77b) without te would be difficult to account for if we took an approach to it along the lines of (5b) or (5c). The legitimacy of the absence of te in (77b) strongly suggests that it is possible to base-generate a VP in the left periphery, with the bare infinitive being the citation form of the verb. The selectional mismatch that would arise if the te-less infinitive originated clause-internally is averted by base-generation of the bare-infinitival VP in the left periphery, associated with a d-word proform, which proberen can independently select (Ik hebdatnog nooit geprobeerd ‘I have never tried that yet’). That no overt correlate pronoun is necessary when the topicalized verbal constituent lacks the infinitival marker te (i.e., dat in (77b) is optional) suggests that the proform to which the base-generated topic is linked can itself be silent.

In the te-less infinitival topicalization/dislocation example in (77b), the proform (silent or overt) is not a surface anaphor: since the constituent in the high left periphery was base-generated there (and thanks to this, manages to escape the selectional restrictions imposed by the matrix verb), it cannot have left a surface anaphor behind further downstream. The proform that the clause-initial VP is associated with must hence be a deep anaphor—in other words, we must be dealing here with (5a).³⁰ Does this mean that the te-less version of (77b) must be treated as a case of HTLD?³¹

An HTLD approach would surely be problematic for the te-less version of (77b) that lacks the proform dat: it would open the door to HTLD derivations for “ordinary” topicalization, which would have all sorts of undesired consequences. A more cautious approach would be to brand (77b) as a case of CLD and to treat the proform as a deep anaphor, as in (5a). But this raises the question of how CLD with deep anaphora can be distinguished from HTLD. Also, if indeed CLD can avail itself of deep anaphora under certain conditions, the division of labor between (5a) and (5b) becomes a prominent item on the research agenda.

Since (5a) and (5b) have different numerations (the former features a deep proform, which is part of its numeration, while the latter does not), the two derivations will only be competitors in terms of economy of derivation if the competition is based on interpretive equivalence (as in Fox 1999, Reinhart 2006). If they are in competition, (5a) might be thought preferable to (5b) because the latter involves a domain-internal movement step whereas the former eschews this movement. But there are strong indications that (5b), exploiting only movement, represents the default option, and that the coindexation/coreference mechanism that (5a) relies on is a last resort.

The second approach would be a less radical departure from the standard view, according to which what appear to be core cases of CLD are indeed bona fide cases of CLD (involving movement), and it is only in special cases that a special construal may be unavoidable. The conceptual appeal of this second possibility is that it would resemble Reinhart’s (1983) Rule I (“NP A cannot corefer with NP B if replacing A with C, C a variable A-bound by B, yields an indistinguishable interpretation”). It would be more closely analogous to Reuland’s (2001) economy hierarchy, according to which syntactic encoding of coreference takes precedence over variable binding (which in turn takes precedence over discourse coreference). Taking (5b) to be the default would also be in line with the classic picture of resumptive pronouns in relative clause islands in English, according to which the resumptive (base-generation) strategy is only available when the movement strategy is not (see last-resort approaches like those proposed in Shlonsky 1992, Pesetsky 1998 (in the Optimality Theory framework), or Aoun, Choueiri, and Hornstein 2001). Sichel (2014) adds an interpretive dimension to the last-resort type of approach to resumption: a resumptive pronoun can be interpreted like a gap only when a gap (instead of a pronoun) would be ungrammatical. This goes along with Grohmann’s account of CLD, in which the resumptive (as a surface pronoun) apparently allows for reconstruction (the resumptive behaves as if the dislocated phrase occupied its position). The requirement (noted by Sichel) that this should be possible only when the gap is ungrammatical is met in Grohmann’s derivation, since the gap would violate his antilocality restriction on domain-internal movements.

To those whom this line of argument strikes as convincing or even compelling, it will be a reason to believe that surface anaphora, as in (5b), is the default, and that (5a) is to be resorted to only when the facts force it upon us: when there are reasons to treat the proform as a deep anaphor. We present this line of thinking here as a topic for further investigation in the syntax of left-dislocation phenomena.

This article grew out of a lively meeting of the syntax reading group at the Research Institute for Linguistics of the Hungarian Academy of Sciences. An earlier version was presented by the first author at the first Budapest Linguistics Conference (19 June 2015), under the title “Contrastive Left Dislocation: Why One Size Does Not Fit All.” We would like to thank the anonymous LI reviewers for their outstanding comments, which led to a complete reorganization of the text. We also gratefully acknowledge the essential financial support of the Hungarian Academy of Sciences (which made it possible for the first author to spend the 2014–2015 academic year in Budapest as a Distinguished Guest Scientist) and the Hungarian Scientific Research Fund (OTKA grant 84217, awarded to the second author). We are grateful to George Soros.