Various views exist in the literature about whether parentheticals should be accommodated by syntactic or discursive means (or both). Some (e.g., Safir 1986, Fabb 1990, Haegeman 1991) have suggested, in various ways, that parentheticals are not syntactically integrated into their hosts, but associate with them at some extragrammatical level of (discourse) representation; this has come to be known as the orphan approach to parenthesis. The opposing view (represented by, e.g., Jackendoff 1977, Potts 2005, De Vries 2007, 2012) holds that parataxis is a bona fide syntactic phenomenon and hence that parentheticals are represented as integral constituents of their hosts; call this the integration approach.Griffiths and De Vries (2013) (henceforth, G&V) have argued that fragment answers furnish evidence for this latter approach, specifically for the syntactic integration of appositive relative clauses (ARCs). G&V claim that their data could not be captured by an orphan approach. In this squib, I show that their argument is inconclusive and fails to support the integration approach.
1 G&V’s Arguments
G&V’s baseline data are the following:
John stole [Mary’s computer], which crashes all the time.
[John stole Mary’s computer], which got him arrested.
In (1a), the ARC associates interpretively with the DP anchor Mary’s computer; call this (following G&V) the constituent reading. By contrast, the ARC in (1b) has a propositional reading; that is, it associates with the entire bracketed clause. G&V go on to note a contrast between B’s responses in (2) and those in (3).
A: What did John steal?
B: Mary’s computer, which crashes all the time.
B': Mary’s computer, which got him arrested.
B": Mary’s computer, which is awful.
A: Who stole Mary’s computer?
B: #John, which crashes all the time.
B': John, which got him arrested.
B": John, which is awful.
While (3B') is analogous to (2B') in permitting the propositional reading, (3B) is not analogous to (2B): it only permits a nonsensical constituent reading (with John coerced into an inanimate interpretation), not the meaningful constituent reading of (1a).1 Similarly, the ARC in (3B") only permits a propositional reading, whereas the one in (2B") is ambiguous. On the basis of these facts, G&V ( p. 335) establish the following descriptive generalization:
(4) An ARC can only surface next to an anchor [the ARC’s host constituent] that is at least partially overt.
G&V claim that (4) follows directly from a Par-Merge approach, according to which an ARC and its anchor are related by a functional head Par, creating a configuration that is said to be abstractly analogous to coordination (more on this in section 3).
(5) [ParP [XPanchor] [Par' Par ARC]]
By contrast, G&V argue, (4) could not be accounted for on an orphan analysis of ARCs, according to which ARCs are not structurally connected to their anchors. On the Par-Merge approach, the propositional reading in (2B') and (3B') is derived as follows (here and throughout, angled brackets indicate deletion; trace notation is used for expository convenience):
[ParP [CP [Mary’s computer]i [〈John stole ti〉]] [Par' Par
[ARC which got him arrested]]]
[ParP [CP Johni [〈ti stole Mary’s computer〉]] [Par' Par
[ARC which got him arrested]]]`
Following Merchant’s (2004) analysis of short answers, G&V assume that the DP anchor undergoes evacuation movement out of TP, which subsequently deletes; the ARC is coordinated with the entire clause. To derive the constituent reading in (2B), G&V assume that the entire ParP evacuates from TP, and the ARC is coordinated with the DP Mary’s computer.
(7) [CP [ParP [Mary’s computer] [Par' Par [ARC which crashes all the time]]]i [〈John stole ti〉]]
Crucially now, G&V argue that deriving the meaningful constituent reading in (3B) would require nonconstituent deletion. As shown in (8), this reading would require the ARC to be ‘‘coordinated’’ with the DP Mary’s computer (as in (5)); as a result, the ARC could only be exempt from deletion by eliding a nonconstituent.
*[CP Johni [〈ti stole [ParP [Mary’s computer]〉 [Par' Par
[ARC which crashes all the time]]]]]
G&V’s claim is that an orphan approach cannot exclude the constituent reading of (3B) in this way, since the ARC would be a separate expression not forming a constituent with its anchor.
2 Problems with G&V’s Analysis
I contend that G&V’s argument does not support the conclusions they draw. In what follows, I show that G&V’s analysis fails even when its disputable implicit premises are adopted. I then sketch a simpler but more principled approach to explaining the data, adumbrated but dismissed by G&V.
The first thing to note is that G&V’s argument crucially rests on the presumed nonexistence of nonconstituent deletion (following the style of analysis in Merchant 2004). While G&V are squarely within the mainstream in assuming this restriction, there are in fact strong reasons to doubt its general validity (see Bruening 2015, Ott and Struckmeier 2016). Note that anaphoric destressing applies to all given, noncontrastive elements in a sentence, without requiring them to form a single constituent (or even a continuous string). For instance, a nonelliptical version of (2B) in the same context would not show—in fact, would prohibit—fronting of focused material (small italics indicate anaphoric destressing; capitals, pitch accent; and subscript F, focus marking).
(9) John/He stole [Mary’s COMPUTER]F, [which CRASHES all the time]F.
There is no a priori reason why deletion should differ in this respect, especially if Chomsky and Lasnik’s (1993) characterization of deletion as ‘‘extreme destressing’’ is on the right track (see also Tancredi 1992, Sauerland 1996).2 Without assuming special movements that are licensed only under ellipsis, G&V’s argument no longer goes through: since the ARC is focused, it is exempt from deletion by virtue of this property alone; but if so, G&V’s explanation for the absence of a (meaningful) constituent reading in cases like (3B) cannot be stated in the first place.
Even if we accept G&V’s presupposition about the impossibility of nonconstituent deletion, however, it is not clear that the issue arises at all in cases like (3B). The reason is that, as De Vries (2006) and others show, ARCs can undergo extraposition (pace, e.g., Emonds 1979), as shown in (10). It is unclear, then, why (3B) could not receive the parse in (11), where the deleted material is a (remnant) constituent.
(10) John stole [Mary’s computer tk] yesterday, [ARC which crashes all the time]k.
(11) Johni [〈ti stole [Mary’s computer tk]〉] . . . [ARC which crashes all the time]k
Assuming that extraposition can target adjunction sites at least as high as TP (as in Culicover and Rochemont 1990), this would permit the syntactic separation of the ARC and the material deleted in (11). Indeed, the idea that extraposition can extract constituents from clausalellipsis sites in just this manner is defended at length in Lasnik 2014 for multiple sluicing in English.
But even if we assume that deletion targets C' rather than TP (see footnote 2) and it could be shown that extraposition cannot target a position this high, this conclusion would not help G&V’s argument. As they note ( p. 335), the facts in (3) are exactly analogous when the main clause is reduced by VP-ellipsis rather than clausal ellipsis.
A: Who stole Mary’s computer?
B: #John did, which crashes all the time.
B': John did, which got him arrested.
B": John did, which is awful.
Even extraposition to a position as low as VP (or perhaps vP) would thus be capable of evacuating the ARC from the ellipsis site, yielding the unattested constituent reading of (12B).
John did [VP 〈[VP steal [Mary’s computer tk]]〉 ( yesterday)]
[ARC which crashes all the time]k
In short, G&V’s account in fact rules in the reading they seek to exclude. Not only do they have to assume that deletion is unlike destressing in requiring the prior formation of a remnant target constituent, they also need to prohibit extraposition of the ARC.
Importantly, G&V’s argument for an integration analysis of ARCs fails even when these additional assumptions are adopted. The reason is that G&V’s generalization (4) is a subcase of the more general behavior of relative pronouns (RPs) in ARCs, which anaphorically resume the most prominent anchor in the preceding discourse that matches the RP in person/number/gender features.3 A focused anchor is maximally prominent, whereas a topical/pronominal anchor is maximally nonprominent even when linearly adjacent to the ARC (14)–(15). Where multiple prominent anchors are available (16), the RP resumes the temporally most recent one ((16Ba) vs. (16Bc)). Ambiguity arises where equally prominent candidate anchors are equally recent; this is the case in (16Ba–b), where the sentence-final DP John renders both itself and the containing proposition prominent, by virtue of being the exponent of both focal and sentential stress.
What was stolen by Johnj?
[Mary’s COMPUTER]i (not her purse) was stolen by himj,
whichi crashes all the time.
#whoj denies it vigorously.
Who stole [Mary’s computer]j?
JOHNi (not Peter) stole itj,
whoi denies it vigorously.
#whichj crashes all the time.
What happened?/Who stole what?
[[Mary’s COMPUTER]i was stolen by JOHNj]k,
whoj denies it vigorously.
whichk is awful.
#whichi crashes all the time.
Note that the sequence of (15B), (15Bb) is in fact the natural nonelliptical variant of G&V’s central example (3B) (in the context of (3A) _ (15A)). Given that RPs in ARCs favor prominent anchors over nonprominent/topical ones, this alone makes the continuation in (15Bb) infelicitous, for reasons that are entirely independent of deletion.4
Since RPs are sensitive to discourse prominence in the way indicated by the above paradigm, it does not come as a surprise that a nonovert constituent can never serve as an anchor for a RP (i.e., the RP requires a linguistic antecedent; see Del Gobbo 2007).
(17) [[Mary’s COMPUTER]i was stolen]j, (AGENTkimplicit)
whichi crashes all the time.
whichj is awful.
#whok remains at large.
This, of course, just illustrates G&V’s generalization in (4), but we see now that it is part of a larger pattern generated by the RP’s conditions of use. These conditions can be observed in other contexts as well. German d-pronouns function as RPs in ARCs but also as regular free pronouns (Wiltschko 1998), in which case they show the same preference for prominent anchors, unlike their personal-pronoun counterparts (see Bosch and Umbach 2007).5
Hans1 kümmert sich ungern um [seinen
Hans cares REFL reluctantly for his
Bruder]2. . . . (all-focus context)
‘Hans is reluctant to care for his brother. . . .’
Er1/2 ist Alkoholiker.
‘He is an alcoholic.’
Der#1/2 ist Alkoholiker.
‘. . . who is an alcoholic.’
Bosch and Umbach’s (2007) generalization is that d-pronouns associate with an ‘‘antitopical’’ antecedent, automatically excluding (among other things) pronominal and elided anchors. While English lacks d-pronouns, it is easy to show that personal pronouns and RPs in ARCs differ in the same way.
A: Who knows [an Italian]i?
B: Peterk (does), hei/k’s a rich one.
B': Peterk (does), [a rich one]i/k.
B": Peterk, whok/#i is a rich one.
The predicative appositive in (19B') ambiguously associates with Peter or the specific an Italian, showing that its underlying structure corresponds to (19B) (see Ott to appear). By contrast, the ARC’s RP in (19B") unambiguously resumes Peter, showing that it associates with the most recent nontopical anchor. The RP thus behaves exactly analogously to the German d-pronouns in (18) and is thus equally subject to Bosch and Umbach’s generalization, subsuming (4).
To be sure, these remarks are merely a first approximation of the full picture, and we would of course ultimately like to know why RPs behave in this way. But it is already clear at this point that G&V’s generalization (4) is merely a subcase of the sensitivity of RPs to their anchor’s discourse prominence, undermining G&V’s argument for syntactic integration of ARCs.
As a final piece of evidence in favor of this conclusion, note that the absence of the constituent reading in cases like (3B) obtains even when the ARC occurs as an independent utterance, as in (20) (where A, B, and C are separate speakers).
A: Who stole Mary’s computer?
C: (Yeah,) Which got him arrested.
C': #(Yeah,) Which crashes all the time.
As expected, (20C) resumes B’s elliptical statement John 〈stole it〉. But (20C') has the same nonsensical (but transparent) reading as (3B)/ (12B), even though the ARC here occurs as an independent utterance. There is no plausibility to the suggestion that (20B) and (20C') ever form a constituent (even when we admit the possibility of structural integration), which would require C’s responses to include an elliptical root clause. Apart from the fact that deletion of entire root clauses is neither independently attested nor compatible with standard assumptions about recoverability, this solution would contradict G&V’s generalization (4) (since it would mean that the syntactic anchor in (20C), John stole Mary’s computer, is nonovert) and hence could not be called on to support their argument in favor of integration. By contrast, an orphan approach, in conjunction with the descriptive generalizations about reference identification of RPs sketched above, directly accounts for both G&V’s facts and those in (20).
G&V reject this line of reasoning, stating that ‘‘constraints [of this kind] can be easily overridden’’ ( p. 338n7) and noting that focus can render anchors accessible to linearly distant RPs, as shown above in (14)–(15). While this is true, there is no reason to conclude from this that RPs should be capable of resuming nonovert anchors, this being precisely the case where there is no way of making the anchor more prominent (less topical) than any overtly expressed constituent. G&V further contend that the asymmetry between (21) and (22) militates against the general claim that RPs cannot resume nonovert anchors.
A book about WHICH English queen won the award?
*Elizabeth I, which was published in hardback.
Whose film won the award?
Fellini’s, which is truly a masterpiece.
G&V assert that ‘‘[(21B)]’s unacceptability cannot be due to the incompleteness of the DP anchor, as incomplete DPs . . . make for acceptable anchors in other environments, as [(22B)] illustrates’’ ( p. 340). The comparison does not hold up, however, and it misses the crucial point. The focal anchor in (21B), Elizabeth I, introduces a prominent discourse referent, but not the one intended to be resumed by the RP (i.e., a book about Elizabeth I). As a result, it can only be interpreted as quoting the title of the book in question, yielding an incongruent answer. By contrast, the anchor in (22B), Fellini’s 〈film〉 (with ordinary NP-ellipsis), does introduce the referent resumed by the following RP, providing a suitably prominent anchor. These cases thus do not argue against the conclusion that RPs resume prominent/ nontopical antecedents, which automatically excludes elided ones, as witnessed in (3B) and (12B).
3 Par-Merge: Conceptual Considerations
Building on De Vries 2007, 2012, G&V assume that ARCs, and parentheticals more generally, are integrated into their host sentences by a special type of Merge operation, called Par(enthetical)-Merge. Par-Merge mimics the effects of an orphan analysis: Par-merged constituents are linearly but not hierarchically integrated ( parenthetical inclusion), making them immune to c-command relations. As a result, parentheticals behave like structurally independent expressions. Hence, despite the similarity suggested by the label Par-Merge, this operation is entirely distinct from—in fact, virtually the inverse of—Merge as defined by Chomsky (1995 et seq.). As a theoretical primitive, Par-Merge thus constitutes a significant enrichment of Universal Grammar.
Given this state of affairs, we should ask if the introduction of such a device is justified. The conceptual motivation offered is the following:
[P]arentheses, like any other linguistic material, have both sound and meaning. That is, they are interpreted as well as pronounced; therefore, they must be present at the LF interface and the PF interface. According to standard assumptions about the organization of the grammar, there is only one way to get at these interfaces, namely via the overt syntax. . . . If a parenthesis were to be added at or after the LF interface . . . , there is no way it can be pronounced. (De Vries 2007:220)
This argument rests on a misunderstanding of the orphan approach, implicit in the quotation. The orphan approach claims that parentheticals are generated separately from their hosts; as generated objects, they are mapped onto both interfaces, hence ‘‘have both sound and meaning.’’ Hypothesizing their linear interpolation into and interpretive relation to their hosts to lie outside the scope of syntax in no way amounts to denying this fact; rather, it merely acknowledges the capacity of speakers to interpolate expressions in discourse (hardly a controversial assumption) and that they do so on the basis of prosodic and pragmatic factors governing the organization of discourse (see Onea 2016), just as independent sentences in discourse are related rhetorically rather than compositionally. Saying that parentheticals must be structurally integrated in order to ‘‘have both sound and meaning’’ is simply a non sequitur.
The integration approach to parenthesis based on Par-Merge introduces a construction-specific, primitive syntactic operation designed to preclude the effects of syntactic integration (see De Vries 2007, 2012 for relevant facts). Whatever range of facts can be described in this way, the logic underlying the approach—‘‘Parentheticals have the properties they do because they are Par-merged, and Par- Merge has the properties it does because those are the properties of parentheticals’’—exposes it as ad hoc. An orphan approach captures the structural autonomy of parentheticals while avoiding any such special mechanisms.
The facts presented by G&V do not establish the need for syntactic integration of ARCs. Rather, their generalization (4) follows from independently observable conditions of use imposed by relative pronouns. I have shown that G&V’s arguments fail even when their underlying assumptions are accepted, and that their analysis in fact falsely rules in the readings they seek to exclude.
I have argued on general conceptual grounds against the syntacticization of parenthesis endorsed by G&V, which rests on a construc- tion-specific mode of syntactic integration. This approach, based on a misinterpretation of the orphan alternative, falls short of attaining explanatory adequacy, irrespective of its descriptive merits.
Thanks to James Griffiths, Volker Struckmeier, Mark de Vries, two anonymous LI reviewers, and the LI Squibs and Discussion editors for comments and suggestions.
1 I use ‘‘#’’ here instead of G&V’s original ‘‘*’’ to indicate that (3B) has a nonsensical interpretation while being grammatical (in the technical sense, meaning ‘generated by the grammar’).
2 The assumption that deletion targets a dedicated constituent is intimately tied to Merchant’s (2001, 2004) theoretical implementation, in which deletion of XP is triggered by an E-feature borne by a head taking XP as its complement. See Ott and Struckmeier 2016 and references therein for some critical discussion of this approach. Strikingly, G&V in fact implicitly reject Merchant’s approach, since they assume that C' (not TP) deletes. This assumption is at variance with Merchant’s claim that deletion is licensed by a head bearing E (unless, as a reviewer points out, further functional structure is stipulated).
3 I follow Del Gobbo (2007), Onea (2016), and others in assuming that ARCs are semantically propositional, just like independent sentences (but unlike restrictive relatives, which are predicates), and that the RP is a referential (E-type) pronoun. It follows that ARCs are syntactically, semantically, and prosodically autonomous: they permit no subextraction, they do not compose but express side information, and they form independent intonation phrases with sentential stress and ‘‘comma intonation.’’
4 While pronominalization is, of course, not grammatically obligatory, it is strongly favored over repetition of the full anchor DP in the above contexts (and sanctioned under ellipsis by a semantic identity condition; see Merchant 2001, 2004). The pressure to pronominalize is strong here, owing to the convergence of prosodic and pragmatic preferences for sustaining the low-flat intonation of discourse-given material (everything following John in (15B)) no longer than necessary and for the overt marking of discourse-givenness, respectively. As noted by a reviewer, examples such as (14Bb) improve when their anchor (in this case him=John in (14B)) is realized as a full DP rather than pronominalized. This is unsurprising: the speaker’s being (mildly) uncooperative by foregoing pronominalization permits the hearer to accommodate the RP’s antecedent even if discourse-old, especially in this particular scenario where the animate anchor John is the only compatible antecedent for the RP who. Once we control for the latter confound, we see that a focal anchor trumps a topical full-DP anchor even when the latter is most recent.
Didn’t John know that Mary was going to get him [the book]j ?
No, [she SURPRISED him with the bookj /itj ]i, whichi/#j really moved him.
The RP in the continuation in (iB) is naturally interpreted as associating with the entire proposition, showing that the generalization in the text holds independently of pronominalization per se.
5 A similar locality condition appears to be imposed by the as heading as-parentheticals. Witness:
What did Mary say?
That Peter is stupid, as everybody knows.
Who said that Peter is stupid?
Mary, as everybody knows.
(iB) is ambiguous: the as-parenthetical is construed either as attaching to the remnant that Peter is stupid or to the underlying larger sentence Mary said that Peter is stupid. But (iiB) is unambiguous: the as-clause cannot take scope under the elided predicate say, showing that as-parentheticals behave like ARCs. Everything I say below appears to apply to this case as well, so that these facts should not be taken to support an integration analysis of as-parentheticals.