Ellipsis in appositives

This paper investigates the internal and external syntax of non-restrictive nominal appositives (NAPs), such as John Smith in I met an old friend, John Smith, at the pub . It is shown that the internal constitution of NAPs bears directly on the analysis of the irrelation to the surrounding host sentence, in that a rich internal syntax obviates the need for any direct syntactic connection between host and NAP. It is shown that NAPs are structurally independent sentence fragments that can be freely employed either sequentially (as ‘afterthoughts’) or as interpolated, supplemental speech acts, autonomous from the host in prosody, interpretation, and syntax. The analysis renders superfluous powerful extensions of core syntax/semantics proposed in previous work to capture the syntactic properties of NAPs and other parentheticals, concluding instead that NAPs warrant no enrichment of UG.


Introduction
This paper investigates the syntax of non-restrictive nominal appositives (NAPs), 1 such as John Smith in the following.
(1) As indicated, I refer to the clause linearly surrounding the NAP as its host clause and to the NAP's host-internal associate as its anchor. Throughout, NAPs will be enclosed by commas and printed in boldface for convenience.
From a formal point of view, NAPs raise two central questions: i. What is the structural make-up of NAPs (their internal syntax)?
ii. What is the relation between NAPs and their host clauses (their external syntax)?
Focusing on NAPs in German and English, Is how in this paper that an answer to the first question has direct bearing on the second, in that a rich internal syntax obviates the need for any direct structural connection between host and NAP. The issue is of wider 1 I have nothing to say here about restrictive (or 'close') appositions, such as my friend John (see Lekakou & Szendrői 2011 for an approach). To see the difference between the two types, consider the status of John in My brother John is an idiot and My brother, John, is an idiot. In the first case, John not not prosodically isolated and restricts the potentially multi-membered set of brothers to one, i.e. the speaker asserts that of what may be multiple brothers, John is an idiot. I also exclude unaccented appositive epithets as in Peter [that jerk], which likewise do not share central characteristics of genuine NAPs, as observed by Truckenbrodt (2014). theoretical relevance, since rather powerful extensions of core syntax/semantics have been proposed to deal with the peculiar properties of NAPs. A secondary goal of the present article is to show that such mechanisms can be sidestepped, and hence that NAPsand hopefully parentheticals more generally -warrant no enrichment of UG.
I proceed as follows. Section 2 establishes two basic types of NAPs, following previous work. Expanding on an analysis of peripheral after thought expressions (Ott & de Vries 2014;, I argue that both types are un-derlyingly sentential: 2 specifying NAPs reduplicate the host clause, while predicative NAPs are interpolated elliptical predicational copular clauses. The example in (1), a specifying NAP, is analyzed as shown in (2) (where strike through indicates PF-deletionand '⇑' marks the eventual linear position of the resulting fragment). (2) [ CP 1 I met an old friend ⇑ at the pub today ] [ CP 2 I met [John Smith] at the pub today ] Each CP 1 (= the host clause) and CP 2 (= the NAP clause) is an independent root clause. In discourse, the elliptical fragment functions as a supplemental speech act relative to its non-elliptical host. This configuration is a surface variant of a corresponding 'afterthought' construction (3a) and ultimately of an overt reformulation (3b).
(3) a. I met an old friend at the pub today: John Smith. b. I met an old friend at the pub today: I met John Smith at the pub today.
The central claim of the present paper is that (1), (3a), and (3b) are syntactically identical (modulo PF-deletion) and differ only in terms of how the syntactic ingredients are discursively arranged. 3 I show how this analysis, supplemented with a slightly different deletion analysis of certain predicative appositions, accounts for central properties of NAPs. Section 3 demonstrates the benefits of treating NAPs as concealed root clauses, capitalizing on their semantic, prosodic, and syntactic autonomy. In Section 4 I consider the question of syntactic integration of NAP fragments (their external syntax). I argue that, contrary to some recent proposals, such integration is neither necessary nor desirable, and conclude that NAPs interpolate discursively rather than syntactically. Section 5 concludes.

Two types of NAPs
I distinguish two basic types of NAPs: reformulating NAPs (r-NAPs) and predicative NAPs (p-NAPs). r-NAPs elaborate on their anchor by specifying it; by contrast, p-NAPs uniformly attribute a property to their anchor's referent. In this section, I introduce each type and propose an analysis, with an eye mainly to the internal sytax of NAPs. Their external syntax will be the focus of Section 3. 2 Döring (2014) cites Raabe (1979), Altmann (1981), and Schreiter (1988) as precursors of this idea; others will be mentioned below. 3 Note that this is very different from claiming that the sequences are transformationally related. Since transformations only apply to structures underlying individual root clauses (or even subclausal domains, such as phases), there is no sense in which (3a) is transformationally derived from (3b), or (1) from (3a). At most we could say that the after thought in (3a) and the NAP in (1) relate transformationally to the (structure underlying the) second sentence in (3b), if we construe deletion as a syntactic transformation.

Reformulating NAPs
Let us first consider some basic properties of r-NAPs; variants of (1) are given below.
(4) a. I met an old friend, (namely/#formerly) John Smith, at the pub today. b. German Ich habe einen alten Freund, (nämlich/#übrigens) den I have an.acc old friend namely/incidentally the.acc Peter (nämlich/*übrigens), in der Kneipe getroffen. Peter in the pub met 'I met an old friend, (namely/incidentally) Peter, at the pub.' The NAP John Smith in (4a) referentially identifies the indefinite anchor an old friend; as indicated, it is naturally accompanied by the specificational connective namely (on which see Blakemore 1993;Onea & Volodina 2011) but is incompatible with a temporal adverb (cf. Griffiths 2015a: 68; more on this below). Similarly, the NAP in (4b) permits inclusion of nämlich 'namely' but prohibits übrigens 'incidentally,' an indicator of supplemental information (Altmann 1981: 98). 4 Note that coreference (or, more generally, co-construal) of anchor and NAP rules out an analysis in terms of coordination, unless a specialized type of 'coordination' is stipulated (as in Griffiths 2015a; b). Furthermore, as indicated in (4b) above, German nämlich can precede or follow an r-NAP, which is problematic for the idea that reformulation markers of this kind are realizations of an 'abstract coordinator' in the schema [anchor & NAP] (Griffiths 2015b: 72).
(i) I met an old friend, (?incidentally) John Smith, at the pub today.
Following a suggestion by the same reviewer, I assume that this differential behavior of incidentally and its German counterpart übrigens is due to the interference of pseudo-reformulations in English. As discussed extensively in Van Craenenbroeck (2012) and Barros (2014), languages like English which do not show morphological case distinctions on nominals consequently allow a wider array of (semantically sufficiently parallel) underlying sources of fragments. For (i), this means that in addition to a faithful reformulation of the host, the NAP can be assigned the underlying structure of a copular clause corresponding to Incidentally it was John Smith. (As noted by the reviewer, this explanation immediately accounts for the fact that (i) is fully felicitous only if the addressee is familiar with the referent of John Smith, exactly as with the hypothetical alternative copular source.) As discussed in Griffiths (2015a: Chapter 2), reformulation markers such as namely are required to exclude such alternative non-reformulating sources in English. Note that the temporal adverb in (4a) leads to a pragmatically nonsensical source in either case: (ii) I met [an old friend]i at the pub today: a. … I met (#formerly) [John Smith]i at the pub today. b. … It was (#formerly) [John Smith]i.
By contrast, the NAP in the German (4b) could not be derived in this way, since its morphological case marking renders it incompatible with a pseudo-reformulation (more details on case below). As a result, the inclusion of übrigens clashes with the presupposition that the addressee can identify the referent of the specific indefinite anchor, exactly as in a corresponding reformulation: The fact that the overt reformulations in (iia) and (iii) equally exclude incidentally/übrigens supports the reviewer's suggestion that the relative naturalness of (i) is due to accommodation via pseudo-reformulation, excluded for the non-elliptical cases.
The specificational function of r-NAPs is not necessarily the result of referential identity. The NAPs in the following examples are non-referential expressions that specify their anchors by elaborating on their descriptive content: (5) a. German Ich habe den Peter, (#nämlich / also) einen alten I have the.acc Peter namely that is an.acc old Freund, in der Kneipe getroffen. friend in the pub met 'I met Peter, (#namely/that is) an old friend, at the pub.' b. I keep meeting scientists, especially linguists, at the pub.
As indicated by (5a), namely is restricted to referential NAPs (as in (4)). The specifying character of the NAP is brought out, however, by inclusion of the connective also 'that is,' which in German indicates elaboration. A similar function is performed by especially in (5b), indicating partial identification. For further detailed discussion of different kinds of apposition markers, see Heringa (2012a: Chapter 2).
As shown by the examples above, r-NAPs in German obligatorily match their anchors in morphological case. The following illustrates for dative case: The derivation I propose for r-NAPs is sketched below. The host clause (7a) and the NAP (7b-i) are generated as independent root clauses. The two clauses (call them CP 1 = the host clause and CP 2 = the NAP clause) are truth-functionally equivalent and morphosyntactically identical modulo the difference between anchor and NAP. Given this redundancy, CP 2 is reduced by deletion at PF (7b-ii). Clausal ellipsis as in (7b-ii) is a familiar phenomenon and requires no constructionspecific assumptions. The NAPs in (4) and (5a) above are derivationally equivalent to B's fragment responses in the following: See Merchant (2004) and references therein for extensive justification of a PF-deletion analysis of fragment answers (but see also footnote 16 for some differences between Merchant's and my implementation).
The analogy extends to embedded fragments as in (10), which German permits (cf. Temmerman 2013); expectedly, an analogous fragment can be employed as a NAP, as shown in (11) If the r-NAP here were a plain DP, it could not satisfy the selectional requirements of the matrix predicate glauben 'think' (at least not on the intended reading). By contrast, on the present approach the NAP is an underlying reformulation, so that the matrix predicate is complemented by CP in (10) and (11)  Note that the embedded fragment here receives case from the elided verb verhaftet 'arrested'; more on this in Section 3.4.1. Embedded NAPs as in (11) cannot be derived on a theory that takes r-NAPs to be syntactically coordinated with their anchors, as acknowledged by Griffiths (2015b). 6 Such 5 For reasons that are not clear to me, embedding of p-NAPs is highly unnatural (ia). The fact that the same is true of analogous free-standing predicative fragments (ib) shows that this peculiarity is not specific to NAPs, however. There is no plausibility to the suggestion that the NAP is coordinated with either one of its anchors (which, note, do not form a constituent), on any meaningful interpretation of the term 'coordination.' By contrast, analogous fragments are perfectly admissible as short answers to multiple questions: These facts lend substantial support to my thesis that r-NAPs involve clausal ellipsis while militating strongly against Griffiths's 'WYSIWYG' coordination approach.
If NAPs are indeed derived from full sentential structures, we expect them to perform similar rhetorical functions as corresponding non-elliptical or elliptical sentences discourseadjacent to the host. 7 To see that this expectation is borne out, compare (4) to (15)  namely/thus an old friend '…(#namely/that is) an old friend.' The a-continuations feature non-elliptical versions of the sentences underlying the NAPs in (4) and (5a), respectively; the overall arrangement expresses the same specification of the anchor as the corresponding NAPs (albeit at the cost of pragmatic redundancy due to ellipsis avoidance). In the b-continuations, CP 2 is PF-reduced by clausal ellipsis but not linearly interpolated -it surfaces as an 'afterthought' (AT). The present study can be considered a direct extension of the analysis of ATs developed in Ott & de Vries (2014;, advocating a full syntactic unification of NAPs and ATs. 8 Note that both the unreduced continuations in the a-examples and the AT variants in the b-examples track the felicity of namely/also witnessed in (4) and (5a), suggesting that these elements function as sentential connectives in all cases. In light of the above it seems wrong, and at the very least theoretically inelegant, to dissociate NAPs grammatically from ATs and non-elliptical reformulations. Griffiths (2015a;b) does just this by assimilating r-NAPs to ordinary coordinates with no underlying sentential structure. His primary motivation is the concern that ellipsis in r-NAPs on the approach adopted here is obligatory, while in the general case (inluding ATs) it is optional. That is, elided material can be felicitously realized with deaccenting (again represented by small italics below) when CP 2 is juxtaposed (17a), but not when it is interpolated (17b). 9 (17) a. I met an old friend at the pub, I met John Smith at the pub.
b. #I met an old friend, I met John Smith at the pub, at the pub. c. I met an old friend, John Smith at the pub, at the pub.
Consequently, Griffiths reasons, the analysis in (17c) requires obligatory deletion. This criticism is based on a confusion concerning the notion of 'obligatoriness,' however. I maintain that from a grammatical point of view, deletion is optional in all cases, 10 but extraneous felicity conditions dictate the use of elliptical fragments rather than nonelliptical sentences when discursive interpolation takes place. Since the speaker has both reduced and unreduced forms at her disposal, she resorts to the elliptical forms in this case -which however are but one option proffered by her grammar. What are the relevant extraneous conditions? Contrary to what Griffiths seems to presuppose, they are not conditions on deletion but conditions on deaccenting of the redundant material in CP 2 . Unlike deletion, which can apply 'backward' (18b), deaccenting is inherently anaphoric, i.e. backward deaccenting is prohibited (18c). 11 (18) observing Mary, known to be reluctant to go on dates, buying flowers: a. She must be dating someone, but I don't know who she's dating. b. I don't know who she's dating, but she must be dating someone. c. #I don't know who she's dating, but she must be dating someone. 8 This analytical unification entails that the distinction between ATs and NAPs is purely based on linear position and thus essentially arbitrary, as is particularly obvious when the anchor happens to be in a sentence-final position. Consequently, I will assume that such cases are bona fide instances of apposition. The unity of ATs and NAPs is recognized already in Altmann (1981). 9 A reviewer points out that a 'restart' version of (17b) is acceptable: (i) I met an old friend, (that is to say) I met John Smith, at the pub today. This is expected in light of what I will show immediately below, namely that backward deaccenting is prohibited. Unlike (17b), (i) involves no such cataphoric deaccenting due to the omission of the PP modifier. 10 In fact, it has to be, if the present approach is on the right track. Since CP1 and CP2 are outputs of independent derivations, it would be a contradiction in terms to assume that one can impose the application of grammatical operations on the other. 11 As a reviewer observes, deaccented material is prone to licensing by accommodated antecedents (cf. Fox 1999). This is why (18c) is presented in an out-of-the-blue context. In the elliptical case, the endophoric material is absent from the overt signal; suspension of its resolution permits cataphoric use of the fragment (or partially cataphoric use, as in (17c)). Deaccented material, by contrast, is not absent from the signal -it must be licensed (given) at the time of mention, precluding cataphoric use. The problem with (17b), distinguishing it from its linear-temporal variant in (17a), is thus the attempted cataphoric use of the deaccented PP at the pub, voided when deaccented material is deleted (17c), enabling its resolution 'later on' (see also note 9). To rephrase in terms suggested by a reviewer, inclusion of the deaccented PP would signal informational oldness of this PP, whereas its subsequent accented counterpart in the host implies novelty. In this way, backward deaccenting leads to a clash (the same holds, mutatis mutandis, in (19) and (20)), which however can be overcome by deletion where the contents of the ellipsis site are inferred subsequent to the processing of the host.
There is, then, no grammatically obligatory deletion in NAPs; rather, the grammar licenses elliptical and non-elliptical forms alike, but use of the latter is restricted by independent properties of prosodic reduction. As a result, there is no reason to introduce an analytical asymmetry between interpolated and non-interpolated fragments.

Predicative NAPs
Let us now turn to p-NAPs, which typically function as predicative supplements. In English, (non-referential) r-NAPs and p-NAPs are potentially ambiguous due to the absence of case morphology; however, German provides the relevant surface distinctions. Compare the following to (5a):  (Zifonun et al. 1997. Furthermore, as indicated, p-NAPs are compatible with the adverb übrigens 'incidentally, by the way.' Besides case, übrigens thus offers a further way of distinguishing p-NAPs from r-NAPs, as it is incompatible with the latter (recall (4)); compare (21) to the following: (24) German a. Sie hat ihrem Bruder, (also / #obwohl noch) einem she has her.dat brother that is although still a.dat Kind, Schnaps geschenkt! child schnapps given 'She gave schnapps to her brother, that is to a child.' b. Sie hat ihrem Bruder, (#also / obwohl noch) ein he has his.dat brother that is although still a.nom Kind, Schnaps geschenkt! child schnapps given 'She gave schnapps to her brother, (#that is/although still) a child.' A simple explanation for this asymmetry suggests itself. The connectives namely and German also indicate reformulating elaborations, as seen in (15) and (16) above. Unlike r-NAPs, however, p-NAPs are not reformulations of their hosts, but separate propositions expressing a predication. This explains their compatibility with temporal adverbs (23b) and certain prepositional clause markers (24b), a point to which we return briefly at the end of this subsection. The derivation I propose for p-NAPs is shown below. As with r-NAPs, the p-NAP and its host are independently generated root clauses. In this case, however, CP 2 is not a reformulation of the host clause but a predicational copular clause 12 whose subject is a 12 A further possibility, going back to Smith (1964), is that the pronoun is equivalent to the wh-phrase heading appositive relatives, assuming that such pronouns are indeed referential (as in Del Gobbo's 2003 analysis). Again, not much hinges on this question from the point of view of the present analysis. Note, however, that the subject of p-NAPs can be overt (Heringa 2012a;Koev 2013), in which case only the copula is deleted and an appositive-relative parse is excluded.
(i) The representatives, most of them were women, wore fancy attires.
This shows that at least not all p-NAPs could be derived from appositive relatives. In fact, it is unlikely that the appositive-relative source is ever preferred to the copular-clause source. The reason is that relative referential (or E-type) pronoun anaphorically resuming the anchor, 13 and whose predicate complement is the p-NAP. 14 (25) a. Note that in this case the NAP can be pronounced even when interpolated, given that this, unlike what we saw in connection with (17b) above, does not clash with the anticataphoric character of deaccenting: (27) I met John Smith, he is my best friend, at the pub today. Heringa (2012a), who develops a similar proposal with a somewhat different implementation, 15 traces the basic idea implemented here back to Motsch (1966) and Klein (1976) (see also Acuña-Fariña 1999). This approach to p-NAPs has likewise been adopted by Griffiths (2015a;b).
Unlike what we saw for r-NAPs, deletion in p-NAPs is thus not contingent on parallelism of the two clauses, i.e. p-NAPs are not reformulations but reduced copular clauses. Merchant (2004) shows that this kind of copular-clause reduction, which he dubs limited ellipsis, is licensed contextually, as shown by equivalent discourse-initial utterances (28a). Ott & de Vries (2014; argue that predicative fragments of this kind can be used as ATs, i.e. non-interpolated p-NAPs (28b). pronouns impose stronger discursive locality conditions than personal pronouns (Bosch & Umbach 2007;Ott 2016). Consider the following, where CP1 itself is elliptical: (ii) A: Who wants to marry an Italian? B: Sue, a rich one. B′: Sue, who is a rich one.
The relative pronounin (ii)B′ must be anteceded by Sue. By contrast,the predicative NAP/AT in (ii) B associates ambiguouslywith Sue or an Italian,which follows if its underlying structure corresponds to He is a rich one rather than to (ii)B′. 13 Tim Hirschberg (p.c.) points out that in some cases a demonstrative pronoun rather than a personal pronoun appears to make a more natural copular-clause subject (cf. Merchant 2010). This is certainly the case when the subject is kind-referring, as in the second NAP in (ia) (from Potts 2005), analyzed as shown in (ib): (i) a. Leo, a lion, a mighty species, swallowed the trainer whole. b.
Since none of what I say below hinges on the choice of the subject pronoun, I abstract away from this detail. 14 Note that predicational copular clauses permit not only nominal, but also AP and PP predicates.
Consequently, these, too, can be used appositively (Potts 2005;Koev 2013): (i) a. The guest, visibly angry at the host, left the studio. b. Ed, in trouble with the law once again, has altered his identity.
I see no reason to treat these cases as syntactically different from their nominal counterparts, which I focus on in this paper for reasons of space. 15 Heringa (2012a) recognizes that this approach does not account for 'identificational' appositions (referential r-NAPs), which he consequently excludes from the scope of his proposal. In his subsequent discussion he nevertheless repeatedly fails to distinguish p-NAPs and non-referential r-NAPs, and at some point argues that r-NAPs, too, are in fact underlying specificational copular clauses. That this cannot be correct is shown by the case-matching property of r-NAPs and the connectivity effects discussed in Section 3.4.
(28) a. pointing at a picture: An old friend. (= [ CP he/that is an old friend]) b. I met John Smith at the pub today, an old friend.
As with clausal ellipsis in r-NAPs, I thus take limited ellipsis in p-NAPs to be independently motivated. 16 The presence of a pronominal subject can be diagnosed in familiar ways (cf. Heringa 2012a). Thus, we find that p-NAPs are not licensed in environments where E-type anaphora fails: (29) # {Every/No} climber, an experienced adventurer, was found sipping hot cocoa at the lodge.
Examples like the above (due to Potts 2005) require no further explanation on the assumption that the NAP is a separate root clause that links back to the anchor discourseanaphorically. On this view, the oddness of (29) simply reduces to the oddness of the following: (30) # [{Every/No} climber] i was found sipping hot cocoa at the lodge. He i is an experienced adventurer.
Compare this to a context (again from Potts 2005) where E-type anaphora succeeds, regardless of whether CP 2 is realized as a subsequent full sentence (31a) or reduced and interpolated (31b).
(31) a. [Every climber] i made it to the summit. They i were all experienced adventurers. b. Every climber, all experienced adventurers, made it to the summit.
Such facts thus follow straightforwardly on the present analysis, where the NAP is an independent root clause, and the anchor-NAP dependency is an instance of discourse anaphora. By contrast, Potts (2005;, who treats NAPs as syntactically adjoined to their anchors (see Section 4 below), needs to resort to special stipulations to rule out cases like (29). 17 This analysis of p-NAPs directly explains their compatibility with temporal adverbs such as formerly and, conversely, their incompatibility with the specificational connective namely (23b): (32) I met [John Smith] i at the pub today. (#Namely) He i was (formerly) my best friend.
The copular clause in (32), just like the NAP in (23b), is not specificational but a supplemental predication; as a result, it is incompatible with namely. Conversely, the same 16 Note that unlike both Merchant (2004) and Ott & de Vries (2014) I am here not assuming that deletion is obligatorily fed by leftward movement of the remnant but take it to be maximal prosodic reduction, as proposed by Chomsky & Lasnik (1993) (see Bruening 2015 and Ott & Struckmeier 2016 for recent discussion). While it the present analysis could be straightforwardly restated in terms of movement of the NAP and subsequent remnant deletion, there appears to be little motivation for such a move, in general but also more narrowly for the case at hand. NAPs are generally not sensitive to locality constraints on movement (anchors can be located within any type of island), which are sometimes taken to constitute evidence for movement in elliptical contexts. I put further discussion of these matters aside here. 17 The fact that p-NAPs fail to attach to anchors that contain a bound pronoun (as noted by Potts 2005),illustrated by (ia), follows in the same way from the impossibility of pronominal resumption, as shown by (ib).
(i) German a. [Jeder Insasse]i spricht mit seineri Mutter, eine fürsorgliche Person, every inmate speaks with his mother a caring person mehrmals täglich. multiple times daily 'Every inmate talks to his mother, a caring person, several times a day.' b. [Jeder Insasse]i spricht mit [seineri Mutter]k mehrmals täglich. every inmate speaks with his mother multiple times daily #Siek ist eine fürsorgliche Person.
she is a caring person 'Every inmate talks to his mother several times a day. #She is a caring person.' temporal adverb yields a nonsensical reformulation when included in an r-NAP (23a), analogously to the corresponding reformulation: (33) I met [my best friend] i at the pub today. I (# formerly) met [John Smith] i at the pub today.
By the same token, früher 'formerly' is compatible with nominative p-NAPs only (34a-i), whereas its inclusion in a case-matching r-NAP in (34b-i) and (34c-i) gives rise to an illicit reformulation. 18 (34) German a. (i) Sie hat heute den Peter, (früher) ihr bester she has today the.acc Peter formerly her.nom best Freund, schwer enttäuscht. friend heavily disappointed 'Today she badly disappointed Peter, (formerly) her best friend.' (ii) [CP 2 er war früher ihr bester Freund] 'He was formerly her best friend.' b. (i) Sie hat heute den Peter, (#früher) ihren besten she has today the.acc Peter formerly her.acc best Freund, schwer enttäuscht. friend heavily disappointed 'Today she badly disappointed Peter, her best friend.' (ii) [CP 2 sie hat (früher) ihren besten Freund schwer enttäuscht] 'She formerly badly disappointed her best friend.' c. (i) *Sie hat heute ihren besten Freund, (#früher) she has today her.acc best friend formerly den Peter, schwer enttäuscht. the.acc Peter heavily disappointed 'Today she badly disappointed her best friend, Peter.' (ii) [CP 2 sie hat heute (#früher) den Peter schwer enttäuscht] 'She formerly badly disappointed Peter today.' The fact that inclusion of the temporal adverb leads to an interpretive clash in (34b-i) and (34c-i) butnotin (34a-i) highlightstheformer'sreformulative-sentential character. By the same token, the analysis furnishes a straight forward explanation for the contrast in (24) concerning elaborative also, which is preserved by 'disentangled' variants: (35) German a. Sie hat [ihrem Bruder] i Schnaps geschenkt. … she has her.dat brother schnapps given 'She gave schnapps to her brother.' b. continuation: (i) Sie hat (also) [einem Kind] i Schnaps geschenkt. she has thus a.dat child schnapps given '… That is, she gave schnapps to a child.' (ii) Er i ist (#also) ein Kind. he is thus a child '… (#That is) He is a child.' In (35b-i), just as in (24a), the reformulation of the host clause specifies the anchor's referent, as highlighted by elaborative also. By contrast, the follow-up sentence in (35b-ii) (and hence the derivative NAP in (24b)) provides a supplemental attribution rather than a reformulative elaboration, rendering the use of also infelicitous in this case. For the same reason, the prepositional connective obwohl 'although' 19 in conjunction with the adverb noch 'still' is felicitous only with the p-NAP variant (24b), deriving from (36a), while the case-matching r-NAP variant (24a) could only derive from the nonsensical reformulation in (36b): (36) continuation of (35a) (German): a. Obwohl er i noch ein Kind ist. although he still a.nom child is '… Although he is still a child.' b. #Obwohl sie noch [einem Kind] i Schnaps geschenkt hat.
although she still a.dat child schnapps given has '… Although she still gave schnapps to a minor.' The distribution of cross-sentential connectives thus follows directly from the NAP's underlying structure -a reformulation in the case of r-NAPs, a copular clause in the case of p-NAPs 20 -and the resultant rhetorical relation to the host.
The differential analysis of r-NAPs and p-NAPs proposed here makes certain predictions concerning the syntactic make-up of nominal fragments used in this way. For instance, i-within-i expressions such as [an enemy of himself i ] i cannot be used referentially and are consequently only permissible as nominal predicates (Williams 1982); this rules out (37b) as the structure underlying the NAP in (37a), showing it to be equivalent to (37c). Given that p-NAPs bear invariant predicative case while r-NAPs match their anchor in case, German reveals the unavailability of the structure in (37b) vs. the availability of (37c) on the surface: 21 (38) German a. Man konnte Peter i , (stets) [sein i eigener größter one could Peter always his.nom own worst Feind], nur bedauern. enemy only regret b. *Man konnte Peter i , (also) [seinen i eigenen größten one could Peter thus his.acc own worst Feind] i , nur bedauern. enemy only regret 'Peter, (always/*thus) his own worst enemy, could only be regretted.' 19 I refrain from using the label 'subordinator' here, since the clause headed by obwohl must be an independent root clause (as in Haegeman's 1991 analysis). This is unproblematic, given the fact that such clauses occur naturally as free-standing utterances. 20 Since NAPs are iterable, r-NAPs and p-NAPs can occur simultaneously (cf. Such examples simply consist of multiple amalgamated sentences (in the above case, the host, a reformulation, and a copular clause); nothing more needs to be said. 21 Expectedly, the case-matching r-NAP in (38b) can be used felicitously once eigenen is dropped and the possessive pronoun is interpreted as disjoint from Peter, in which case the NAP is no longer an i-within-i expression.
Similarly, article-less NPs in German are licit as predicates in copular clauses but not in argument positions. Consequently, the NAP in (39a) must have the underlying p-NAP parse in (39b) but cannot be assigned an illicit reformulation structure corresponding to (39c). We thus have substantial evidence for the analysis adopted here, according to which p-NAPs are elliptical copular clauses. For extensive further discussion of this approach, see Griffiths (2015a: chapter 3).

Parallelism in r-NAPs
What dictates that CP 2 be a faithful reformulation of CP 1 , i.e. how is missing material in the NAP resolved against its host? Since I propose that deletion in NAPs is ordinary clausal ellipsis, the question is more general than the proposal entertained here, and much ongoing research is devoted to it. 23 While I refrain from entering into this debate, I adopt a version of Merchant's (2001) semantic parallelism condition based on e-givenness for the sake of explicitness.
Note, first, that it would be too simple to require CP 1 and CP 2 to be truth-functionally equivalent (⟦CP 1 ⟧ ↔ ⟦CP 2 ⟧). This would work for r-NAPs that referentially identify their anchors but fail for other cases of reformulation, such as the following: (42) a. German Ich habe Peter, einen alten Freund, in der Stadt I have Peter.acc an.acc old friend in the city getroffen. met 'I met Peter, an old friend, in the city.' b. She wants to marry an Italian, (preferably) a rich one.
(Now she just needs to find one.) We want to ensure that CP 1 and CP 2 are identical in meaning, modulo anchor and NAP. Merchant's e-givenness condition allows us to do just this.
(43) e-givenness An expression E counts as e-given iff E has a salient antecedent A and, modulo ∃-type shifting, A entails the focus closure of E, and E entails the focus closure of A.
The focus closure of an expression α, F-clo(α), is the result of replacing F(ocus)-marked constituents of α with existentially-bound variables. We can now spell out the deletion rule as follows: 24 (44) Clausal ellipsis (optional rule) Delete all and only e-given material in CP.
Let us see how this applies to NAPs and their anchors. NAPs are never given or presupposed, i.e. they are F-marked in Schwarzschild (1999) sense (small caps indicate prosodic prominence; I ignore nuclear stress and additional F-markings beyond NAP and anchor throughout): 24 Note that I diverge from Merchant's formulation here. Being concerned with sluicing, he takes deletion to target IP (a conventional but highly questionable assumption, given the fact that C-heads never survive clausal ellipsis) and thus phrases his conditions in terms of IP-ellipsis. My statement of the parallelismcondition as applying to sentences resembles that of Rooth (1992a) and Reich (2007) in this regard (see also footnote 16). Ott & Struckmeier (2016) point out that a deletion rule of this kind fails to account for the fact that unfocused particles can survive clausal ellipses (as in (63) and (65)  It is now easy to see how deletion in CP 2 satisfies e-givenness. Since both anchor and NAP are F-marked constituents, they will be replaced by ∃-bound variables in each F-clo(CP 1 ) and F-clo(CP 2 ); felicitous deletion then requires everything else in the two sentences to be semantically identical, so that mutual entailment obtains. Consider (45) Trivially, CP 1 entails F-clo(CP 2 ) and CP 2 entails F-clo(CP 1 ), and deletion in CP 2 satisfies parallelism. Conversely, if CP 2 is semantically non-equivalent to CP 1 , it will fail to entail the latter's F-closure, and deletion will be infelicitous. By adopting Merchant's (2001) semantic parallelism condition, we permit truth functionally vacuous morphosyntactic non-isomorphism between the two sentences. 25 That this is a welcome consequence is shown by cases such as the following (from Potts 2007): (52) There was a former linguist, Ed Witten, at the party.
a. *There was Ed Witten at the party. b. Ed Witten was at the party.
If the elliptical CP 2 were required to be syntactically isomorphic to CP 1 (modulo NAP and anchor), it could only be assigned the deviant structure in (52a). By assuming a semantic parallelism condition, we permit (52b) as the underlying structure of the NAP. 26 We will encounter a further case of morphosyntactic non-isomorphism in Section 3.3. While NAPs are always F-marked, the requirement that their anchors also be appears to be lifted in two cases. The first is non-referential indefinite anchors, as in (42b). Here, no F-marking appears to be required: (53) A: Is it true that Sue wants to marry an Italian? B: Yes, she does want to marry an Italian, [a rich one] F . Now she just needs to find one.
I assume that the indefinite anchor is an existentially bound variable in the F-closure, as assumed by Merchant (2001) for standard cases of sluicing with indefinite anchors. We obtain F-clo(CP 1 ) = F-clo(CP 2 ) = ∃x : she wants to marry x, and e-givenness is satisfied.
The second case is that of p-NAPs. Here, too, we find that the anchor can be contextually given. Since German reliably distinguishes case-matching r-NAPs and nominative p-NAPs on the surface, we can illustrate with the following minimal pair: 25  In the context set up by speaker A, only the nominative p-NAP in B′ is felicitous, indicating that its addition, unlike that of the predicational r-NAP in B, does not require F-marking of the anchor. This is just what we expect, given that I have argued that p-NAPs are not derived by deletion under identity with CP 1 , but rather by a contextually licensed reduction (following Merchant 2004).
Merchant's parallelism condition based on e-givenness thus appears to yield the right results, allowing us to make precise the sense in which the sentence underlying an r-NAP reformulates the NAP's host. Note, however, that for those cases where anchor and r-NAP are referential expressions, nothing in the identity condition requires anchor and NAP to be coreferent. But disjoint reference of anchor and NAP is plainly ruled out: 27 (55) #I met [an old friend] i , Peter k , at the pub today.
I propose that obligatory coreference is not enforced by parallelism, but by the rhetorical relation between CP 1 and CP 2 . To see this, note that disjoint reference is as infelicitous in (56) as it is in (55).
(56) a. I met an old friend i at the pub today: I met Peter i/#k at the pub today. b. I met an old friend i at the pub today: Peter i/#k . The second, parallel sentence is naturally understood as rephrasing the first (as we saw before,this rhetorical relation can be made explicit by connectives such as namely or that is); as a result, we interpret Peter as identifying the referent of an old friend. Coreference of anchor and r-NAP in such cases is thus not enforced grammatically, but by text/discourse coherence. 28 I will leave a precise characterization of the relevant 27 Ignoring here a vocative interpretation of Peter, which is due to accidental surface ambiguity. 28 Since this is the case, coreference can be easily overriden by sentential connectives introducing CP2 that serve to indicate a non-specificational type of reformulation: (i) a. I met an old friend, and also John, at the pub today. b. I met an old friend, but unfortunately not John, at the pub today. c. I think I'll meet Peter, or perhaps John, at the pub today. d. I didn't see anybody, not even John, at the pub today.
As a result, the approach developed here extends naturally to such coordinative supplements, restatements, corrections, etc. While these constructions do not fall under the traditional definition of apposition, they bear obvious parallels to bona fide NAPs owing to their reformulative character. This is also recognized by Heringa (2012a: 20), based on related discussion in del Saz Rubio (2003) and Jasinskaja (2009), and appears to be assumed by Döring's (2014) general approach to parentheticals. Despite these obvious extensions, I continue to restrict my focus to conventional cases of apposition in what follows, in order to keep the discussion within manageable proportions. The full range of constructions encompassed by the current approach thus remains to be explored in future work. discourse relation to future work, since it is orthogonal to the syntactic focus of the present paper. 29

Interim summary
NAPs come in two basic varieties. r-NAPs are underlyingly parallel reformulations of their host clauses: felicitous deletion requires truth-functional equivalence of the two underlying sentences, modulo anchor and NAP. p-NAPs are underlyingly non-parallel copular clauses with a coreferent/E-type pronominal subject; deletion of this subject and the copula does not require a linguistic antecedent, as shown by other instances of limited ellipsis. A by-product of the reformulating character of r-NAPs is their case-matching property; p-NAPs, by contrast, bear invariant predicative case. R-NAPs and p-NAPs are thus derivationally equivalent to fragment answers: 30 (57)  The specificational/predicational function of the NAP reflects the rhetorical relation between CP 1 and CP 2 . If CP 2 is reduced by ellipsis and juxtaposed to the host clause in discourse, it is realized as an AT; if it is linearly interpolated into CP 1 , it surfaces as a NAP. All ingredients of the analysis are thus independently motivated (see Ott 2012; in press and Ott & de Vries 2014; 2016 for related discussion).

NAPs as autonomous root clauses
The preceding section presented a number of reasons to believe that NAPs contain more structure than meets the eye. This section investigates in detail further properties of NAPs that reveal their status as structurally independent root clauses. As root clauses, NAPs are shown to be compositionally, prosodically, and structurallly autonomous from their hosts -that is, NAPs are separate, 'interrupting' speech acts.

Interpretive autonomy
The claim that NAPs are autonomous root clauses makes direct predictions concerning their interpretation: NAPs are expected to be independent propositional units, rather than entering into the semantic composition of their hosts. As observed by Dever (2001) and Potts (2005;, NAPs and their hosts are truth-functionally separate entities. To see this, consider the following (from Dever 2001): (59) Plato, the greatest metaphysician of antiquity, wrote the Cratylus.
Dever and Potts note that the NAP here contributes a propositional meaning that is dissociated from the host's truth conditions: we can consistently assert that Plato wrote the Cratylus, while denying that he was the greatest metaphysician of antiquity. On the 29 On the approach developed in Ott & Onea (2015) and ongoing work (based on Onea 2013; 2016), coreference of anchor and NAP in cases like (55) is imposed by question-answer congruence: only if Peter is understood to be coreferential with an old friend does it answer a relevant potential question (viz., Which old friend did you meet at the pub?) On such an approach, an identity condition based on e-givenness can be replaced by Rooth's (1992b) notion of question-answer congruence, following observations of Reich (2007). I will not pursue these issues in the present paper. 30 Griffiths (2015b) points out that this analogy does not extend to corrective fragments, as in The wind blows abaft, or behind, the boat. I set this issue aside here, since corrections appear to have special properties; furthermore, right-node raising may be involved in such cases. present analysis, these two separate propositions of course correspond simply to the underlying bisentential source in (60) (61) Plato i wrote the Cratylus. He i was the greatest metaphysician of antiquity.
The fact that the p-NAP contributes a separate truth value thus follows automatically on the present analysis. Given that we analyze (59) as a surface variant of the sequence in (61), we avoid the need for the extra semantic machinery devised by Potts (2005) to account for the propositional nature of NAPs (see Section 4).
Consider now the r-NAP in (62), which on the present analysis is (62b) interpolated into (62a). Given that anchor and NAP are interpreted as coreferent for purposes of specification, differential truth values of host clause and NAP intuitively give rise to inconsistency. Nevertheless, we can reject this specification, e.g. by responding: No, that's not true. It's her brother John who has a girlfriend. 31 Note that we thereby do not deny the truth of the proposition asserted by the host, showing that even in the case of r-NAPs two separate (but truth-functionally equivalent) propositions are being asserted. Further support for the sentential propositional character of NAPs derives from their compatibility with sentential adverbs and modal particles (in German), illustrated in the following: (63) a. One of them, probably Peter, eventually got arrested.
b. German Sie haben einen Mann, vermutlich ein Obdachloser, they have a.acc man presumably a.nom homeless.man am Dom verhaftet. at.the cathedral arrested 'They arrested a man, presumably a homeless man, near the cathedral.' c. German Er ruft ja jemanden, angeblich wohl seinen Vater, he calls prt someone.acc allegedly prt his.acc father zweimal täglich an. twice daily up 'As we know he calls someone, allegedly his father, I suppose, twice a day.' Sentential adverbs and modal particles alike express speaker-oriented extra-propositional information. In the cases above, their function is to mitigate the speaker's commitment to the truth of the propositions expressed by the NAPs: that he is a homeless man and that he calls his father twice a day, respectively. Note that as with other adverbs, the scope of the 31 Note that the NAP in (59) cannot be directly denied in this way. However, Koev (2013) shows that this is a mere linearity effect: predicative ATs (= non-interpolated p-NAPs) or p-NAPs following a sentence-final anchor can be directly denied. sentential adverb/modal particle is restricted to the NAP, as expected if the latter is an autonomous root clause, shown below for (63c). The ellipsis analysis of NAPs thus requires no analytical devices beyond what is needed anyway to derive modified fragments in general (viz., clausal ellipsis). 32 Generally speaking, the distribution of modal particles is restricted to root clauses that function as speech acts (Jacobs 1991;Reis 1997;Coniglio 2012). The occurrence of modal particles (and sentential adverbs) in NAPs can thus be taken as an indication of their status as independent speech acts, as also argued by Truckenbrodt (2014 In (66a), the p-NAP is outside the scope of the host's interrogative force, asserting that (rather than asking if) Jane is the best doctor in town; in (66b), the host clause is declarative while the r-NAP is interrogative. This follows directly from the respective underlying structures of r-NAPs and p-NAPs: 32 Note, however, that ellipsis remnants of this kind are not easily accommodated by the theory of parallelism assumed in Section 2.3, since neither sentential adverbs nor modal particles are focused. An alternative that takes such ellipsis remnants into account is sketched in Ott & Struckmeier (2016). 33 Truckenbrodt's claim is restricted to p-NAPs, however. He assumes that r-NAPs are not compatible with sentence adverbs, but I believe this to be incorrect, as shown by examples like (63c). As is clear from these structures, the r-NAP in (66b) reformulates its host as a question (68), 34 while the p-NAP in (66a) introduces a predication as speaker-oriented side information (69). 35 In each case, the NAP is an autonomous speech act, deriving from an autonomous root clause. The Romanian example (67) is particularly clear in this regard, since the modal particle oare in this language is licensed in genuine question acts only (Coniglio 2012). Note that an analysis in terms of plain coordination (without ellipsis), as proposed by Griffiths (2015a;b), necessarily falls short of capturing facts of this kind, unless it is enriched with ancillary stipulations permitting it to emulate what follows automatically on the present approach.
Overall, then, there are strong indications that NAPs of both types are semantically autonomous from their hosts, used by the speaker as independent speech acts. See Dillon et al. (2014) for some recent evidence supporting this conclusion from a psycholinguistic perspective.

Prosodic autonomy
The structural separation of NAPs from their hosts is reflected not only in their interpretation as separate, supplemental propositions and autonomous speech acts, but also in their prosodic realization. NAPs manifest intonational isolation ("comma intonation") characteristic of parentheticals, i.e. they are flanked by prosodic breaks, often, but not necessarily, realized as pauses (see, e.g., Emonds 1976;Altmann 1981;Bolinger 1989;Taglicht 1998;de Vries 2007;Selkirk 2011).
The prosodic autonomy of NAPs is corroborated by the fact that they do not affect host-internal stress placement but constitute independent domains of sentential-stress assignment. To see this, consider the host clause in (70a) (uttered in a What happened? context), where friend bears nuclear stress (NS); if the optional adjunct is added, it attracts NS. A subsequent reformulation substituting a specificational NAP for the anchor will be prosodically realized as indicated in (70b), i.e. with everything but the specifying DP deaccented (indicated by small italics), so that the latter comes to bear NS. Ellipsis of the deaccented material and interpolation of the NAP yield the result in (70c), which conserves the stress patterns of the two original sentences, rather than shifting phrasal/ nuclear stress to the NAP (70d). 34 This entails that clausal force is irrelevant to parallelism computation, which only takes propositional content into account. That this must be the case is independently shown by run-of-the-mill instances of sluicing (A: Peter kissed someone. B: Who?). 35 An interesting issue that I will not investigate in detail here is to what extent the tense specification of the copular clause can differ from the host. Judging by examples such as the following, present tense appears to be default interpretation of the NAP, but can be overriden (unsurprisingly, since deletion in p-NAPs is not subject to a parallelism requirement).
(i) a. Chomsky, a linguist at Harvard, wrote LSLT. b. Chomsky, then a linguist at Harvard, wrote LSLT.
The most natural interpretation of (ia) is that Chomsky is currently a linguist at Harvard, i.e. he is a linguist at Harvard (note that this does not render (ia) false as a whole, cf. supra). By contrast, in (ib) the underlying clause must be past tense, owing to the presence of the temporal adverb. I leave further investigation of this aspect to future work; see Heringa (2012a) for related discussion.
(70) a. I met an old friend (at the pub). b. I met John Smith (at the pub). c. I met an old friend, John Smith (, at the pub). d. #I met an old friend, John Smith (, at the pub).
The same is true for r-NAPs: here, too, NS within the host is preserved (71c) rather than shifted to the NAP (71d) (cf. de Vries 2007).
(71) a. I met John Smith (at the pub today). b. He's an old friend. c. I met John Smith, an old friend (, at the pub today). d. #I met John Smith, an old friend (, at the pub today).
NAPs differ in this regard from integrated constituents such as conjuncts (72b) and adjuncts (72c), which do affect NS placement: (72) What happened? a. I met John today. b. I met John and Bill today. c. I met John at the pub today.
Unlike integrated constituents, NAPs realize an additional, autonomous NS (see also Truckenbrodt 2014). Since the domain of NS assignment is typically taken to be the in to nation phrase (ιP), we conclude that NAPs are ιPs at the level of prosodic representation. A 'WYSIWYG coordination' approach to r-NAPs, as pursued by Griffiths (2015a;b), must resort to ancillary stipulations to distinguish the prosodic realization of such NAPs from that of ordinary conjuncts. Selkirk (2011) deals with the ιP status of NAPs by adopting Potts's (2005) proposal that NAPs bear a special feature, termed "comma-feature." Selkirk assumes that this feature triggers the formation of a separate ιP in the PF-mapping -a property that commamarked constituents share with root clauses, by stipulation. On the present approach, the ad hoc comma-feature can be dispensed with: NAPs behave prosodically like root clauses simply because they are root clauses, the syntactic correlate of (as per, e.g., Gussenhoven's 2004 Align(S,ι) and Selkirk's 2011 Match(Clause,ι)). From this perspective, intonational breaks flanking NAPs are prosodic correlates of sentence boundaries, exactly as in a sequence of root clauses. Schematically: (73) a. (I met an old friend at the pub) ι P(I met John at the pub) ι P → interpolation b. (I met an old friend) ι P(I met John at the pub) ι P (at the pub) ι P Given that ιPs are domains of sentential-stress assignment, the autonomous prosodic status of NAPs is accounted for. In all cases, the prosodic contour is a faithful blend of the underlying sentences. 36 36 Two additional remarks are in order here. First, NAPs are not always set off by pauses; the left edge of r-NAPs in particular is prone to blending into the preceding tonal unit. This does not come as a surprise, however, since this prosodic behavior is familiar from other types of parentheticals such as comment clauses, as surveyed by Dehé & Kavalova (2007) citing Taglicht (1998: 196): "parentheticals may, in intonational phrasing, group to the left, but not to the right." Second, there appears to be a slight asymmetry between r-NAPs and p-NAPs, in that the latter tend to be set off more strongly from the surrounding host clause. As a result, pauses at both edges of p-NAPs appear to be virtually obligatory. In the light of the present approach it is tempting to speculate that these differential degrees of prosodic marking reflect differential degrees of redundancy between host clause and NAP clause: while r-NAPs are fully redundant with the host clause in their underlying structure, p-NAPs are non-parallel propositions whose content is recovered contextually.
Note that this explanation generates a prediction concerning NAP placement: each of the ad hoc ιPs engendered by NAP interpolation must realize NS. This explains why the NAP can be interpolated in the way shown above, but not when at the pub is given by context and replaced by a proform in the host: (74) What happened at the pub? a. (I met an old friend there) ι P(I met John there) ι P b. #(I met an old friend) ι P(I met John there) ι P(there)??
Being inherently destressed, the locative proform is incapable of supporting a separate ιP, precluding interpolation of the NAP to an anchor-adjacent position. Consequently, the fragment is juxtaposed rather than interpolated, i.e. realized as an AT. In much the same way, the prosody precludes NAP interpolation to a position adjacent to the anchor in (75a), isolating the prosodically weak verb particle: (75) German a. #Ich rufe morgen einen Freund, den Peter, an. I call tomorrow a.acc friend the.acc Peter up b.
Ich rufe morgen einen Freund an, den Peter. 'I'm going to call up a friend tomorrow: Peter.' This prosodically conditioned 'niching' behavior is entirely expected of NAPs if these are a species of parentheticals, as on the present approach (cf. Peterson 1999). By contrast, such behavior is entirely unexpected if NAPs are merely syntactically integrated conjuncts, as maintained for r-NAPs by Griffiths (2015a;b).

Syntactic autonomy
Parentheticals are linearly interpolated into their hosts but otherwise syntactically inert (see, e.g., Espinal 1991;Peterson 1999;de Vries 2007;. In this section, I show that the same is true for NAPs: no syntactic dependencies can straddle the host/NAP boundary. Where NAPs do seem to be structurally connected to their hosts, this is shown to be the result of the underlying parallelism of host and NAP. NAPs are invisible to host-internal agreement processes. This is illustrated for number agreement by the following (cf. Quirk et al. 1985; also Acuña-Fariña 1999 for Spanish gender agreement): 37 (76) a. The loot, fourteen pure diamonds, was/*were worth millions.
b. German Peter und Maria, (schon lange) ein Paar, haben / *hat Peter and Maria already long a couple have has beschlossen zu heiraten. decided to marry 'Peter and Maria, a (long-time) couple, have decided to get married.' Leaving a detailed empirical and theoretical investigation to future work, I suggest tentatively that it is this redundancy that enables a certain degree of prosodic coherence in the case of r-NAPs. Note that the formation of ιPs is known independently to be subject to non-structural factors (Nespor & Vogel 1986;Truckenbrodt 1999;Selkirk 2011). 37 Quirk et al.'s (1985) examples, such as their Lands, brains, wealth, technology -(in other words) everything we need -are/*is plentiful in our country are less than ideal, in that the pre-NAP phrase is amenable to a hanging-topic parse, in which case the NAP is in turn parsed as host-internal, giving rise to the opposite agreement pattern. I have tried to control for this confound in the examples given in the main text.
On the present analysis, the NAP is a separate root clause and as such outside the c-command domain of any host-internal functional head; the invisibility of NAPs to Agree or equivalent operations follows: Note that differential agreement within CP 1 and CP 2 does not impinge upon semantic parallelism of the two clauses, similar to the syntactic mis-match observed for cases like (52).
Further more, as observed by McCawley (1998), NAPs are in accessible to movement operations: they can neither be extracted (78b) ((78a) serves as a control showing that extraction is possible in principle) nor do they permit subextraction into the host clause (78c). This opacity is a further illustration of the fact that NAPs are parentheticals. Parentheticals are known to be robustly opaque in this way (Peterson 1999;de Vries 2007;; compare the attempted extraction out of the interpolated clausal parenthetical in (80a) to long extraction in (80b).   Griffiths (2015a;b) claims that r-NAPs, which are not parenthetical on his analysis, do in fact permit subextraction, based on examples like the following: (82) Which country do you hate the motorways of, or the 'highways' of, the most?
Griffiths claims that which country is ATB-extracted from anchor and NAP. There is reason to doubt, however, that ATB-extraction ever reorders material out of non-initial conjuncts, rather than applying asymmetrically (Munn 1993;Salzmann 2012). Thus, even if (82) did involve genuine ATB-movement, this would not establish the NAP's syntactic permeability. 38 Setting aside these concerns, the approach developed here is perfectly compatible with Griffiths's observation: whether we take the NAP to be a reformulating question or a declarative reformulation of the host question, in neither case would there be a syntactic dependency spanning host and NAP. To decide between the two options, consider the following example from German: Von welchem ihrer Freunde hat Maria einen Verwandten, of which of her friends has Maria a relative angeblich ja den Vater, auf den Mund geküsst? allegedly prt the father on the mouth kissed 'Which of her friends did Maria kissa relative of, allegedly the father of?' Assume that within the host clause, the wh-phrase is sub extracted from the NP headed by Verwandten 'relative.acc' (84a). But the elliptical sentence inserted as a NAP is declarative (84b), 39 as shown by the presence of the particle ja (roughly, 'as we know'), which is incompatible with interrogative force (84c). There is no reason, then, to assume that subextraction from the NAP into the host is involved in the derivation of cases like (82) and (83). 40 In fact, on an ATB analysis it would remain entirely mysterious why equivalent 'extractions' can occur across speakers, as in the following: (85) German A: Von welchem ihrer Freunde hat Maria einen Verwandten auf den Mund geküsst? (= (84a)) B. Angeblich ja den Vater.
'Allegedly the father (as we know).' B supplements A's question with a fragment equivalent to the NAP in (83); the interpretation of the supplement is identical to that of the NAP in (83), showing that both derive from the common source in (84b). Griffiths's (apparent) cases of ATB-movement thus provide no reason to weaken the conclusion drawn previously, according to which NAPs are invisible to syntactic operations launched from within the host.

Connectivity
If NAPs are structurally autonomous from their hosts, we expect them not to partake in quintessentially syntactic relations underlying case-marking, binding, and scope-taking. This section shows that this expectation is borne out. While p-NAPs show the structural disconnect on the surface, r-NAPs appear to be amenable to syntactic dependencies originating in the host, owing to their underlyingly reformulative nature.

Case and θ-roles
As mentioned before in Section 2, r-NAPs match their anchors in morphological case: (86) German a. Einer von den Typen, der Peter, hat sie schwer one.nom of the guys the.nom Peter has them heavily beleidigt. insulted 'One of the guys, Peter, insulted them badly.' b. Sie haben einen von den Typen, den Peter, schwer they have one.acc of the guys the.acc Peter heavily beleidigt. insulted 'They badly insulted one of the guys, Peter.' c. Sie haben einem von den Typen, dem Peter, geholfen they have one.dat of the guys the.dat Peter helped zu flüchten. to escape 'They helped one of the guys, Peter, to escape.' Such case matching appears to be a fairly robust crosslinguistic property of r-NAPs (see the review in Heringa 2012a). 41 Compare the following cases from Basque and Romanian: 40 And recall from Section 3.1 that Griffiths's 'WYSIWYG coordination' approach to r-NAPs fails to account for the autonomous illocutionary force of NAPs, as evidenced in (83) Less conspicuous but no less important is the fact that r-NAPs whose anchor is an argument match the latter in θ-role. Prima facie at least, these relational case and thematic properties suggest that r-NAPs are structurally connected to the infrastructure of their hosts after all. As also mentioned in Section 2, p-NAPs differ from r-NAPs in bearing invariant nominative case, irrespective of the case specification of their anchor. To the examples in (21), repeated in (89) (i) German Sie war im Besitz zweier Kleidungsstücke der Ermordeten, einem she was in possession two.gen pieces of clothing the.gen murdered.fem a.dat Mantel und einem roten Kimono. coat and a.dat red kimono 'She owned two pieces of clothing of the murdered woman, a coat and a red kimono.' Van Riemsdijk notes that this behavior is observed only with oblique anchors and dative NAPs; in nonoblique contexts, such 'deviant' dative NAPs are always excluded, and case matching is obligatory. As far as I can see, these exceptional dative NAPs are as problematic for my account as for any other. Surely one could come up with underlying structures that would accommodate these NAPs, but it is not clear at all how such source structures could be made compatible with identity conditions on deletion more generally, which usually enforce strict case matching (see Barros 2014 and works cited there). I am forced to leave an account of these peripheral cases to future research.

(90) Romanian
Astronauţii au dat Arabelei, un păianjen de grădină, astronauts.def aux gave Arabella.dat a garden spider.nom apă şi carne. water and meat 'The astronauts gave Arabella, a garden spider, water and meat.' This differential behavior of r-NAPs and p-NAPs presents a non-trivial challenge to all extant approaches to NAPs. Potts (2005; takes case matching as in (86) to support the adjunct status of NAPs; his claim is that the anchor's case feature is transmitted to the adjoined NAP. Apart from the fact that this is achieved by mere stipulation, the approach has no way of accounting for the absence of case matching in p-NAPs (indeed, Potts fails to properly distinguish the two types). While Heringa (2012a) excludes NAPs that specify their anchors from the scope of his analysis, he nevertheless attempts to account for casematching effects in non-referential r-NAPs by likening them to multiple case-assignment to coordinates (see also de Vries 2007;Griffiths 2015a;b). This idea, however, relies crucially on the highly idiosyncratic notion of coordination invoked in his analysis. In any event, as with Potts's approach, Heringa's provides no principled reason for why different types of NAPs should behave differently with regard to case. Finally, Griffiths's (2015b) 'WYSIWYG coordination' approach to r-NAPs fails to account for case matching with embedded NAPs (11) and complex NAPs specifying multiple anchors (13), both of which are falsely predicted to be impossible in general on his proposal, as pointed out above.
By contrast, the ellipsis approach handles the case facts without resorting to any special mechanisms. Recall that r-NAPs are underlying reformulations, i.e. deletion applies under identity with the host clause. As a result, anchor and NAP are case/ θ-marked in parallel, each in their respective (root) clause and thus in compliance with the Theta Criterion and standard assumptions about case assignment. (91) illustrates for (86c). In fact, as discussed extensively in Barros (2014: chapter 2), such case matching under ellipsis is "stubborn," i.e. it obtains even when semantic identity would tolerate mismatching case specifications of the ellipsis remnant and its correlate. By contrast, p-NAPs are predicate complements in underlying copular clauses and as such receive predicative nominative case. (93)  Case and θ-properties of NAPs are thus fully determined within CP 2 , in a run-of-the-mill, local fashion. The apparent structural connectedness of r-NAPs is illusory, reducing to clausal parallelism.

Binding
Analogously to what we saw for case in the previous section, r-NAPs appear to be capable of entering into interactions with host-internal binders. First, prounouns inside (object) r-NAPs can be bound by host-internal (subject) QPs, as shown by the following examples: (94) a. Every inmate i talks to one person, (probably) his i mother, once a week.
b. German Jedem Linguisten i ist ein Artikel, (nämlich) sein i erster, every linguist is one article namely his first besonders wichtig. especially important 'Every linguist considers one article, (namely) their first one, especially important.' Second, reflexives inside r-NAPs can be bound by host-internal R-expressions (and conversely, pronouns are not free in corresponding positions): (95) a. John i found something odd, a book about him*(-self) i , at the store. 42 b. John i 's mom found something odd, a book about him(*-self) i , at the store. c. German Peter i hat jemanden, (nämlich) {sich selbst i / *ihn i }, im Peter has someone namely himself him in Spiegel gesehen. mirror seen 'Peter saw someone, (namely) himself, in the mirror.' Third, Condition C is violated when a NAP-internal R-expression is (seemingly) c-commanded by a coreferent host-internal NP: (96) a. *He i found something odd, a book about John i , at the store. b. German *Sie i hat Hans, Marias i alten Vater, in der Stadt getroffen. she has Hans Maria's old father in the city met 'She met Hans, Maria's old father, in the city.' Like case matching, these binding patterns are straightforward indications of syntactic connectivity that could, at first glance, suggest the conclusion that r-NAPs are structurally embedded within their hosts, in apparent contradiction to the facts reviewed in Section 3. Fortunately, however, the ellipsis analysis of r-NAPs preempts this paradox: the relevant 42 A reviewer notes that the example is fine with an epithet like the bastard coreferent with John replacing the pronoun. A possibility raised by the reviewer is that this case might involve a predicational source ( This reduction of apparent connectivity to parallelism is familiar from analyses of fragment answers (Merchant 2004), ATs (Ott & de Vries 2014;, and other elliptical constructions (see, e.g., den Dikken et al. 2000;Arregi 2010;Kluck 2011). With these constructions we find analogous binding patterns mediated by ellipsis of parallel structure: (98) A: Who does [every inmate] i talk to once a week? B: (Probably) His i mother.
(99) [Every inmate] i talks to one person once a week, (probably) his i mother.
The fragment (probably) his mother is formally the same creature in each (94a), (98), and (99), employed in different ways in discourse. In German, the fragment shows casematching in all instances.
With regard to such binding interactions, r-NAPs behave strikingly different from surface-sentential parentheticals, which show no connectivity (Peterson 1999;de Vries 2007;. Thus, in (100a) and (100b) the host subject cannot bind the parentheticalinternal pronoun, and the configuration in (100c) fails to elicit a Condition C effect.
(100) a. *[Every guest] i -he i had just arrived -was complaining.
b. German *[Jeder Schüler] i hatte, wie er i schließlich zugeben musste, every student had as he eventually admit had.to beim Test geschummelt. at.the test cheated 'Every student had cheated in the exam, as he later had to admit.' c. She i told her uncle -Jane i had always hated him -to leave.
Here, unlike in the examples involving r-NAPs above, there is no parallel elided structure supporting the binding dependency. If the present approach is on the right track, the interpolated clauses in (100) and NAPs alike can be treated as syntactic disjuncts, i.e. genuinely extra-sentential expressions.
If r-NAPs show connectivity effects by virtue of being reformulations, we expect p-NAPs, being non-parallel predicational copular clauses, to test negative for binding connectivity, paralleling the observations in the previous section. This expectation is borne out. The examples in (101) differ sharply from both those in (96) above and those in (102), where an R-expression internal to the integrated constituent gives rise to a standard Condition C violation. As before, the impossibility of binding follows from the disjunct status of the NAP and the absence of a parallel binder from its underlying structure. Consider, e.g., the representation of the NAP in (101a)  'He is a narcissistic impostor.' Here as well as in cases like (94) above, the NAPs are no more transparent to external dependencies than the clausal parentheticals in (100); all binding is internal to the elliptical CP 2 . This accurately and directly predicts both the anti-connectivity effects in (101) and (103) and the (apparent) connectivity effects in (105).

Scope
In the light of the above findings concerning binding connectivity, we expect r-NAPs and p-NAPs to differ in analogous ways with regard to scopal interactions. This prediction is borne out.
It has been claimed that NAPs are generally scopeless, i.e. outside the scope of any host-internal elements (see, e.g., Potts 2005;Nouwen 2007 (109) Sheila thinks that Chuck i is fit to watch the kids. He i is a confirmed psychopath.
For Potts, who treats NAPs as syntactic adjuncts, the scopelessness of NAPs motivates a complex semantic system that maps the adjunct NP onto a supplemental proposition (corresponding to the denotation of the second sentencein (109)). On such an account, there is no direct connection between a NAP as in (107) and an overt reformulation as in (109); as a result, it is left to the syntax-semantics mapping to align the two cases (see Section 4 below). On the ellipsis approach, by contrast, the supplemental-propositional nature of NAPs reduces to their underlying syntax.
Even more damaging for Potts's approach, which does not distinguish between r-NAPs and p-NAPs, is the fact that r-NAPs are not scopeless in this way but do show scopal interactions with the host clause. 46 The following case is a variant of (107) with a specificational r-NAP: (110) Sheila thinks that a confirmed psychopath, her brother Chuck, is fit to watch the kids.
In this case, the NAP is interpreted as though it were embedded under the intensional matrix predicate: it is necessarily Sheila's belief that her brother Chuck is fit to watch the kids (as brought out by the infelicity of the continuation . . . but she doesn't think that Chuck is fit to watch the kids). The NAP is not actually within the scope of the host's matrix predicate, of course, but embedded under its counterpart within CP 2 : By virtue of reformulation, (112a) asserts that she didn't invite Peter, further specified by the reformulation that she didn't invite her best friend: the NAP takes scope below negation, within CP 2 (112b). By contrast, the minimally different (113a) conveys that Peter formerly was her best friend, not the negation thereof. Thus, here, as in (107) above, the NAP's 'scopelessness' is due to its non-parallel clausal structure (113b).
A related phenomenon is the potential of NAPs to project presuppositions. Wang et al. (2005) note that (114) has both a de dicto and a de re reading; on the latter, the anchor an Italian projects an existential presupposition (there exists a certain Italian, of whom it is then asserted that Mary wants to marry him). 47 (114) Mary wants to marry an Italian, a rich one.
On the present approach, the two readings are a direct result of the NAP's surface ambiguity. If the NAP is construed as a reformulation (115a), it permits a de dicto interpretation. By contrast, the copular-clause parse in (115b) gives rise to the de re reading (with an Italian construed as the referential antecedent of he). As expected, in a German translation of this example matching case on the NAP correlates with the interpretation of (115a) while nominative identifies the NAP as a nominal predicate, as in (115b). The English case can be disambiguated by varying the example slightly: (116) She i wants to marry an Italian, one of Mary i 's high school friends.
Here, only a de re interpretation is available. This is so because the NAP-internal R-expression (construed as coreferent with the host subject) precludes the reformulation reading, blocked by Condition C (117a). Hence, only the r-NAP parse is available (117b), which in turn requires the indefinite an Italian to act as a referential anchor for the elided subject pronoun. These intricate scope/projection behaviors of NAPs follow automatically on the assumption that host-NAP connectivity (where it obtains) is illusory, owing to the presence of elided parallel structure, and that NAPs of both types are structurally disjunct root clauses.

Interim summary
In this and the preceding sections, I have shown that the analysis proposed in Section 2 straightforwardly accounts for the seemingly inconsistent syntactic behavior of NAPs. All NAPs are parenthetical disjuncts and as such dissociated from their host's syntactic, prosodic, and compositional-semantic make-up, very much unlike arguments and adjuncts. Nevertheless, both r-NAPs and p-NAPs appear to show differential degrees of connectivity into their hosts. In the case of p-NAPs, this illusory connectivity is limited to the binding potential of the subject of the underlying copular clause. We saw that r-NAPs behave differently, as expected given their underlying parallel structure. Due to this redundancy, r-NAPs inherit all morphosyntactic properties of their anchors, resulting in case and θ-role matching and full binding/scope connectivity (as familiar from sluiced wh-phrases and fragment answers). In no case does the approach require a syntactic dependency between host and NAP, reconciling connectivity effects with the structural autonomy of the disjunct NAP clause.

NAPs in syntax?
In this section, I briefly compare the approach suggested here to the two most explicit syntactic analyses of NAPs I am aware of, i.e. those of Potts (2005; and De Vries (2007;, the latter adapted to NAPs by Heringa (2012a). 48 Both approaches share the assumption that NAPs are syntactically integrated into the host clause. First, I show that each approach is flawed independently of how it holds up against the ellipsis approach. I then show that a conception of NAP interpolation as an extra-grammatical process bears greater plausibility and should thus be preferred on grounds of parsimony, although no pragmatic account of NAP interpolation will be developed in this paper. Potts (2005; argues that NAPs are syntactic adjuncts that bear a special feature, which he dubs "comma," as shown in Figure 1 for the p-NAP in (118). 48 The 'WYSIWYG coordination' analysis of r-NAPs proposed in Griffiths (2015a;b) will not be discussed here, as I believe it to be untenable for reasons given throughout the discussion in previous sections.  Potts's main concern is accounting for the apparent semantic multidimensionality of sentences containing NAPs, i.e. the fact that such sentences can have multiple truth values (recall the discussion of (59) in Section 3.1). In his system, NAPs give rise to conventional implicatures (CIs) that correspond to the propositional meanings of NAPs. He sets up a system in which semantics employs both at-issue types (e a , t a , s a ) and CI types (e c , t c , s c ). His system involves two basic types of semantic composition: at-issue application (stan-dard functional application) and CI application, which creates dually-typed mother nodes. The comma-feature signals a shift from at-issue content to CI content, mapping 〈e a , t a 〉 expressions onto 〈e a ,t c 〉 expressions. In this way, Potts argues, the NAP in (118) gives rise to the implicature Edna is a fearless leader, brought out by the continuation in the following: (119) Edna, a fearless leader, started the descent. #Edna is not a fearless leader.
The approach developed in the present paper yields the same net result, by taking the NAP to be a surface-reduced sentence denoting the proposition corresponding to Potts's CI content. That is, on the present analysis the effect in (119) follows from the analogous effect in (120), where the bracketed part is the 'disentangled' version of (118).
(120) [Edna started the descent. She is a fearless leader.] #Edna is not a fearless leader.
I have argued that the NAP is simply the result of reduction and subsequent linear interpolation of the copular clause. Neither multidimensional composition nor comma-features are required on this approach. Furthermore, Potts's approach has a number of empirical and conceptual weaknesses. For one thing, Potts fails to distinguish r-NAPs from p-NAPs; he typically only considers instances of the latter but generalizes to all NAPs, which leads to empirically false claims, such as the general "scopelessness" of NAPs (recall the discussion in Section 3.4). Nothing in his system explains or even describes correctly the differential behavior of r-NAPs and p-NAPs discussed above.
Recall that the present approach unifies clause-medial NAPs and right-peripheral ATs, the only difference being their linear positioning (interpolation vs. juxtaposition). Since Potts takes NAPs to be adjuncts to their hosts, his analysis fails to capture the common core of NAPs and ATs. While Potts does not explicitly draw any connection whatsoever, his analysis permits only one option, namely to analyze ATs as rightward-moved NAPs. But such an analysis is implausible at best. First, if NAPs could move to the right periphery of their host, their incapacity to undergo any kind of leftward movement (illustrated in Section 3.3) would remain mysterious, and set them apart rather sharply from ordinary adjuncts. Second, such an analysis would beg the question why other DP adjuncts, such as free datives, resist being shifted to the right: This criticism applies more generally to Potts's reliance on the comma-feature, a descriptive device that at best restates the striking asymmetries between NAPs and conventional adjuncts in syntax, interpretation and prosody. While Potts's system is designed to capture the semantic autonomy of NAPs, it is only the comma-feature that accounts for their syntactic and prosodic behavior. No such ad hoc devices are required on the present account, which assumes that NAPs are independently generated expressions that may or may not be linearly interpolated to a position adjacent to their anchor.
De Vries (2007; 2012) develops a general framework for the syntactic analysis of parentheticals, which Heringa (2012a) adapts specifically to NAPs. 49 The basic idea is that parentheticals, including NAPs, are merged into the primary structure by means of a specialized operation. De Vries (2012) defines this operation, dubbed par-Merge, as follows: (122) par-Merge(A, B) yields C such that a. C directly par-includes A, b. C directly par-includes B, and c. A is the merge-mate of B.
Par-inclusion is defined as inclusion without dominance; as a result, par-Merge integrates a constituent into the structure linearly but not hierarchically. Par-Merge is thus a novel operation designed to incorporate parataxis into narrow syntax. NAPs and other parentheticals are then analyzed as complements of a functional head, called Par, which par-Merges with the NAP, yielding the boxed portion of the structure in Figure 2. As a result, only the NAP's linear position is specified; it does not enter into any dominance or command relations with other elements, as per (122). De Vries's original approach (see also O'Connor 2008) makes false predictions concerning NAP connectivity. Since NAPs are simply par-merged DPs, and par-Merge by stipulation precludes any hierarchical dependencies involving the NAP, the (apparent) connectivity effects reviewed in Section 3.4 directly refute the approach. More generally,the approach is ill-equipped to account for any of the root-clause properties of NAPs reviewed in Section 3: it either cannot account for them at all (e.g., adverbial modification) or else only by sheer stipulation (e.g., comma intonation as a correlate of par-Merge). Worse yet, there is no basis for distinguishing different types of NAPs, which are uniformly analyzed as complements of Par. 49 A slightly different implementation of this approach is developed in Heringa (2012b), which I will not discuss separately here. Heringa (2012a) adopts de Vries's basic assumptions but argues for an internal syntax of NAPs that is quite similar to what I have argued here for p-NAPs. That is, Heringa analyzes NAPs as copular clauses with a pro subject and an empty copula, roughly as shown below:

(123) [ NP Edna i [ ParP pro i BE [ NAP a fearless leader ]]]
Regarding p-NAPs, this proposal bears an obvious resemblance to the analysis proposed in Section 2.2. Note, however, that the deletion analysis requires no reference to otherwise unattested pro subjects (in languages like English and German). It is not entirely clear whether Heringa excludes r-NAPs from the scope of his analysis or intends them to be subsumed under (123) (or a variant thereof). In any event, it is clear that r-NAPs cannot be analyzed in this way, which would falsely predict the absence of connectivity effects. As presented, the analysis is not equipped to model the differential behavior of r-NAPs and p-NAPs.
More important to my mind is the central conceptual flaw of the de Vries-Heringa approach: like Potts's (2005) comma-feature, par-Merge is an ad hoc device introduced to encode the peculiar properties of parenthetical expressions (cf. Ott 2016). Even if the proposal were empirically tenable, it would hardly reduce the complexity of the original problem in an enlightening way. While Potts advocates a significant enrichment of the syntax-semantics mapping, de Vries and Heringa considerably enrich narrow syntax by devising a novel composition operation, over and above what is minimally needed for unbounded recursive composition. What both approaches show, in effect, is that a nonsyntactic analysis of NAPs and parenthetical expressions in general ought to be the null hypothesis, certainly if we follow Chomsky's (2007) methodological maxim that the complexity of UG ought to be kept to an irreducible minimum.
The analysis of the internal syntax of NAPs developed here has two immediate consequences for the analysis of their external syntax: it renders syntactic integration of NAPs both problematic and, fortunately, unnecessary. The latter consequence is due to the fact that the ellipsis approach undermines any putative argument for the syntactic integration of NAPs, in that all apparent indications of NAP-host connectedness are fully confined to the syntactic domain of CP 2 . This is a welcome result, since it entails that NAPs can be treated as extra-syntactic disjuncts just like other parentheticals (which, recall, show no signs of connectivity whatsoever); no differential treatment is required. The ellipsis analysis of NAPs finds a natural ally in the orphan approach to parenthesis (marshalled by Safir 1986;Haegeman 1991;Peterson 1999; Burton-Roberts 2005, among others), which holds that parataxis is an extra-syntactic, discursive phenomenon.
In fact, the essence of the analysis militates strongly against the syntactic integration of NAPs. The reason is that any such containment relation will necessarily render deletion in r-NAPs antecedent-contained. To see this stronger point, consider the two competing scenarios depicted graphically in Figure 3 (∆ is shorthand for deleted structure; linear position irrelevant). In scenario 1, the elliptical CP 2 is structurally embedded within the host clause; in scenario 2, it is not, and interpolation only occurs in production (as per the orphan approach). Given that deletion in CP 2 is resolved under identity with CP 1 (= the host clause), scenario 1 entails that deletion in r-NAPs is antecedent-contained, hence predicted to be irresolvable. 50 The fact that deletion in NAPs is resolvable thus strongly suggests that NAPs are not structurally embedded within their hosts, but rather interpo-lated extra-grammatically, at the level of discourse organization where rhetorical relations are established. 51

Conclusions
In this paper, I have argued that NAPs are interpolated elliptical sentences, identifying them as surface variants of ATs as analyzed in Ott & de Vries (2014;. Like ATs, NAPs come in two basic varieties: one that constitutes an underlying reformulation of the host (an insight going back to Burton-Roberts 1975), and another that provides side information in the form of an underlying predicational copular clause (as proposed in different ways in Potts 2005 and Heringa 2012a and earlier work).
In the case of r-NAPs, deletion applies under identity with the host, reducing the reformulation to a fragment that is optionally interpolated in discourse. Schematically: (124) a. reformulation I met an old friend at the pub. I met John Smith at the pub.
b. afterthought I met an old friend at the pub. I met John Smith at the pub.
c. NAP I met an old friend at the pub. I met John Smith at the pub.
In the case of P-NAPs, deletion of the subject pronoun and the semantically empty copula is recoverable without an explicit antecedent: These cannot mean I wonder why I wonder and I asked when I asked, respectively, which would require resolving the embedded sluice against the matrix clause that contains it, giving rise to an infinite regress. Compare: I wonder. But why?, I asked. When? 51 To resolve the problem arising in scenario 1, an integration approach would need to assume QR-like covert movement of the NAP to avoid antecedent-containment, on the assumption that the derived position could somehow be discounted for purposes of parallelism calculation. Recall from Section 3.3 that NAPs consistently resist any kind of overt displacement (or any other dependencies involving elements of the host), making it implausible that they can enter into analogous covert dependencies. Furthermore, as Fox (2002) and Chomsky (2004) have observed, the scoping-out solution to ACD is undermined by the Copy Theory of Movement. The alternative suggested by the present approach sidesteps the issue entirely, by denying that NAPs are ever represented as constituents of their hosts (cf. Chomsky 2004 on ACD in VP-ellipsis). (125) a. supplemental attribution I met John Smith at the pub today. He's an old friend.
b. afterthought I met John Smith at the pub. He's an old friend.
c. NAP I met John Smith at the pub. He's an old friend.
I have shown that implementing the distinction between the two types of NAPs in this way generates accurate predictions concerning differential connectivity of r-NAPs vs. p-NAPs. In both cases, the resulting sentence fragment is linearly interpolated into the externalized form of the host clause in discourse as an interrupting speech act, rather than integrated syntactically. 52 The conclusions arrived at in this paper converge not only with those of Ott & de Vries (2014; concerning ATs (and their interrogative variants, studied in Arregi 2010) but also the analysis of left-dislocation developed in Ott (2012;in press). In all cases of dislocation, including what we might refer to as medial dislocation in the case of NAPs, we find peripheral 'satellite' phrases that, despite their extra-sentential status, show syntactic connectivity into their host -a situation termed Cinque's Paradox by Iatridou (1995). Ellipsis resolves the paradox in all cases: by taking dislocated XPs to be elliptical fragments, we can attribute putative effects of structural connectedness to silent parallel structure. 53 To illustrate this synthesis, the DP den Peter 'Peter.acc' comes to bear the case (and θ-role) it does when dislocated in any of the ways shown in (126), because in all cases it is endophorically connected to its host qua elliptical reformulation there of (127), while simultaneously being cataphorically ((126a), (126b)) or anaphorically (126c) connected to a host-internal correlate.  (124) and (125) involves two independently generated sentences, one of which may be elliptical. The speaker can use a sentence fragment either as an AT or as a NAP (ceteris paribus), but these are two equivalent usage options involving identical syntactic ingredients, not forms related by syntactic transformations -and indeed they could not be, given that syntactic transformations only relate structures underlying (at most) individual sentences. 53 Modulo predicative fragments deriving from copular clauses, which, as we have seen, show no signs of connectedness and differ in their interpretation.
In all cases, the dislocation is the result of juxtaposition or interpolation of the fragment in discourse, relying on the resources furnished by the sentence grammar; all that distinguishes NAPs from left-dislocated XPs and ATs, or indeed fragment answers to questions, is their extra-syntactic linear interpolation into another utterance, whereas left-dislocation and AT are peripheral uses of fragment XPs. Taken together, these works thus resolve Cinque's Paradox by removing dislocation from the purview of syntax proper. In doing so, the approach furthermore obviates the need for construction-specific mechanisms devised in previous works to rationalize the paradoxical properties of dislocated elements, such as 'big-XP' base structures and 'binding chains' in the case of peripheral dislocations (see Ott 2014; and Par-Merge in the present context. If, as I have argued in Section 4, the linear intercalation of NAPs into their hosts is a matter of discourse rather than syntax, this result calls for a pragmatic characterization of the relation between host and NAP (working in tandem with prosodic conditions on niching, adumbrated in Section 3.2). Having homed in on the structural properties of NAPs in this paper, I do not develop such a theory here, but refer the reader to Ott & Onea (2015) and ongoing work building on the syntactic foundation developed here. In a nutshell, Ott & Onea characterize NAPs as answers to potential questions that are incrementally licensed in discourse by the host clause. To illustrate this in the most rudimentary fashion, consider the following host and r-NAP: (128) German a. Ich ✘ habe ✘ einen Freund ✔ gebeten ✔ die Akten ✘ zu I have a.acc friend asked the files to vernichten. ✔ destroy 'I asked a friend to destroy the files.' b. [CP 2 ich habe [den Peter] gebeten die Akten zu vernichten] 'I asked Peter to destroy the files.' The NAP can be felicitously inserted at any of the positions marked ✔, whereas the positions marked ✘ are unavailable. Building on Onea (2013;, Ott & Onea pursue the intuition that the permissible positions are those and only those at which a relevant question is licensed (by virtue of its presuppositions being satisfied) and salient -Which friend?, Which friend did you ask to do something?, and Which friend did you ask to destroy the files?, respectively. This approach captures basic facts about NAP interpolation: NAPs can never interpolate to a position linearly preceding their anchors; NAPs need not be string-adjacent to their anchors; and other potential anchors can 'intervene' by licensing potential questions on their own (which is why, in (128a), the NAP could not surface right-adjacent to the files). For a more detailed outline of the approach, which is developed in ongoing work, see Ott & Onea (2015). As I have shown in this paper, analyzing NAPs as locally adjoined to, par-merged to, or coordinated with their hosts falls short of capturing their autonomous status in prosody, interpretation, and syntax. NAPs are independent fragments that can be freely employed either sequentially (as ATs) or as interpolated, supplemental speech acts. A goal of the present paper was to show that the non-trivial enrichments of core/syntax semantics postulated by extant analyses of NAPs can be avoided, and that a more parsimonious view of NAPs as independent syntactic domains is not only tenable, but in fact provides a more accurate explanation of their structural properties.