CHAPTER 23
ANTECEDENT-CONTAINED DELETION
23.1 INTRODUCTION TO THE PHENOMENON
IN a canonical instance of Verb Phrase Ellipsis (VPE) in English, as in (1), the site of ellipsis (indicated by did in bold) marks the absence of a Verb Phrase (VP). The ellipsis is interpreted by ‘looking back’ to a salient antecedent in the discourse (Hankamer & Sag, 1976; Sag, 1976)—in this example, the VP fell in love in the first clause, as shown in (2).
(1) Lady Sybil fell in love, and Lady Mary did, too.
(2) Lady Sybil [VP fell in love], and Lady Mary did 〈fall in love〉, too.
Antecedent-Contained Deletion (ACD) is a special case of VPE in which the site of ellipsis is contained in its antecedent, as illustrated in (3) (Bouton, 1970).
(3) Daisy eventually [VP knew how to prepare every dish that Mrs Patmore did].
This situation gives rise to two well-known issues. First, as long as the site of VPE is contained in its antecedent, the elided VP cannot be parallel, or identical, in form with the antecedent (Merchant, 2000b; Fox, 2002). Second, if we assume that the ellipsis is to be resolved by copying in the antecedent, then we face a second problem. Because the antecedent contains the VPE site, whenever the antecedent is copied in, the VPE is as well, resulting in an infinite regress, as shown in (4).
(4) Daisy eventually [VP knew how to prepare every dish that Mr. Patmore did [〈know how to prepare every dish that Mrs Patmore did〉 [〈know how to prepare every dish that Mrs Patmore did〉…]]]
If we were stuck with this situation, the sentence would be uninterpretable, and our sentence processors would be occupied indefinitely.
However, there is a solution, first proposed by May (1985), and fleshed out in subsequent years by Fiengo & May (1994); Kennedy (1997); Merchant (2000a); and Fox (2000, 2002); among others. This solution relies upon Quantifier Raising (QR), a covert form of Ā movement in which a quantificational phrase is displaced in the structure at Logical Form (LF), as illustrated in (5).1,2 (For the time being, I will use notation suggesting that QR targets a landing site above the subject position, but I will revise this assumption shortly.)
(5)a. Base structure:
Daisy eventually [VP knew how to prepare [QNP every dish that Mrs Patmore did]].
b. QR of the QNP:
[QNP every dish that Mrs Patmore did]
Daisy eventually [VP knew how to prepare [QNP t]]
c. Interpretation of the VPE via copying:
[QNP every dish that Mrs Patmore did 〈know how to prepare〉]
Daisy eventually [VP knew how to prepare t]
An alternative to (5) is to follow Fox’s (2002) proposal that there is covert rightward movement of the Quantified Noun Phrase (QNP) followed by adjunct merger of the relative clause (Lebeaux, 1988), thereby skirting a violation of parallelism. This step is followed by trace conversion and lambda extraction, as shown in (6). Either way, we can assume that the QNP is covertly displaced at LF in order to resolve the issues that ACD imposes on interpretation.
(6)a. Base structure:
Daisy eventually [VP knew how to prepare [QNP every dish]].
b. Rightward movement of the QNP:
Daisy eventually [VP knew how to prepare [QNP every dish]] [QNP every dish]
c. Adjunct merger of relative clause:
Daisy eventually [VP knew how to prepare [QNP every dish]]
[QNP every dish [RC that Mrs Patmore did 〈know how to prepare dish〉]]
d. Trace conversion and lambda abstraction):
every dish λx. Mrs Patmore 〈knew how to prepare the dish x〉
λy. Daisy eventually [VP knew how to prepare dish y]
e. Interpretation of ellipsis:
‘For every dish that Mrs Patmore knew how to prepare,
Daisy eventually knew how to prepare that dish’
The movement operation that covertly displaces the QNP in ACD is semantically, and not syntactically, motivated by needs related to compositionality and sentence interpretation. This movement targets a landing site that is higher in the structure. The landing site of QR indicates the site at which the QNP takes scope, and also has consequences for the binding principles, which are evaluated at LF.
As the reader might surmise, there are also non-movement-based approaches to ACD resolution that do not rely upon variables and traces. Going forward, I will henceforth adopt the covert movement approach presented in (5), not only through personal preference, but also because this approach to ACD resolution is, at this point, widespread in the field. It should be possible to translate the implications of the collective findings I review here into an alternative approach; however, where the distinction becomes most relevant is in the second part of section 23.2, where we evaluate whether evidence from psycholinguistic studies with adults can be interpreted as directly indicating whether QR has taken place in ACD resolution.
Let us take binding relations first. While at first blush, it might look as though (7) would incur a Principle C violation (since the R-expression in bold is not free), the binding relation is evaluated at LF, after the QNP has undergone QR, and as a result, ACD bleeds Principle C.3 However, the subsequent examples incur a violation for reasons related to the interpretation of ellipsis and the landing site. In (8), the pronoun from the surface is copied into the ellipsis site, resulting in a Principle B violation (Fiengo & May, 1994). Perhaps the most informative violations for QR, however, come from those in (9) and (10), since they reveal that the landing site that is targeted must be below the subject position. In both examples, a pronoun in subject position cannot be co-construed with an R-expression that it c-commands in the structure below it (Fox, 1995a,b, 2000; Johnson & Tomioka, 1998; Merchant, 2000a,b).
(7)a. Lady Rosamundk told herj everything that Lady Edithi thought shek would.
b. Lady Rosamundk [QNP everything that Lady Edithi thought [CP shek would 〈tell heri〉]] told her
(8)a. *Lord Grantham mentioned himi to everyone that Tom Bransoni did.
b. *Lord Grantham [QNP everyone that [CP Tom Bransoni did 〈mention himi to〉]] mentioned himi to
(9)a. *Hei told the Inspector everything that Mr Batesi thought hei should.
b. *Hei [QNP everything that Mr Batesi thought [CP hei should 〈tell the Inspector〉]] told the Inspector
(10)a. *Shei recounted every story Corai’s housemaid did.
b. *Shei [QNP every story Corai’s housemaid did 〈recount〉] recounted
The scope-taking ability of the QNP based on QR is also illustrated in (11) and (12). In (11) there is a Negative Polarity Item (NPI): a damn thing, modified with a relative clause with ACD. Because the NPI is only licensed as long as it falls under the scope of negation, a downward entailing environment (Fauconnier, 1975; Ladusaw, 1979, 1980), the QNP must QR to a position below negation (Merchant, 2000a).4 Thus, there is further evidence that the landing site of QR is below the subject, and is adjoined to, for example, vP. In (12), the sentence is ambiguous, because ACD is embedded within another VP. As a result, the QNP could QR out of the embedded VP, resulting in the embedded reading indicated below, or QR out of the matrix VP, resulting in either the matrix reading or the embedded reading indicated below. (Once the QNP is no longer contained in either VP, either one is a candidate for antecedent status.)
(11) Thomas won’t do a damn thing Mr Carson asks him to without an attitude.
Thomas won’t [a damn thing Mr Carson asks him to 〈do〉 without an attitude] do
(12) Isobel wants to know about every decision about the hospital that the Dowager does.
Embedded:…every decision about the hospital that the Dowager does 〈know about〉
Matrix:…every decision about the hospital that the Dowager does 〈want to know about〉
Now, in (12), ACD is embedded in a non-finite clause (wants [to know…]), and there is no question that the QNP should be able to move via successive cyclic movement to a position higher than both VPs. Until fairly recently, there has been an assumption that QR is subject to a clause-boundedness constraint, whereby a QNP cannot raise outside of a tensed clause. In (13), for example, the QNP every place setting… is claimed not to be able to raise out of the embedded clause, because the tensed say takes a finite complement clause (that Mr Carson inspected…). Thus, while the surface scope reading is possible, the inverse scope reading with the QNP taking scope over the indefinite subject is questionable, as indicated below the example.
(13) A footman [VP said [CP that Mr Carson inspected every place setting]]
a > every, ?every > a
If this is the case with QR in ACD, we would not be able to access the intended interpretation of (14), and would only be left with the embedded clause reading of (15). However, observations by Wilder (1995, 1997) and Fox (2002), and further theoretical proposals about successive cyclic movement of QR through the syntactic structure to allow such readings by Cecchetto (2004), have strongly suggested that this is not, in fact, the case. This empirically-based disagreement with earlier theoretical assumptions thus leaves room for experimental work to provide relevant informative evidence about constraints on QR, and therefore possible limits on available interpretations of ACD sentences.
(14) Lady Cora expects that everyone will fancy Lady Rose that Lord Grantham does. 〈expect that those people will fancy Lady Rose〉
(15) Spratt said he noticed every imperfection that Danker did.
Embedded: 〈noticed〉
Matrix: 〈said she noticed〉
This review has uncovered a host of characteristics and puzzles connected to the phenomenon of ACD. First, ACD is a form of VPE in which the site of ellipsis is contained in its antecedent. To resolve this situation, the QNP must undergo semantically-motivated QR to a position outside of the antecedent VP (most likely adjoined to vP). Second, this landing site is not only above the VP, but below the subject position, a fact that we can diagnose with binding relations and licensing of NPIs. Moreover, the landing site is not restricted to one location: ambiguous ACD sentences are such because there are multiple landing sites, correlated with different interpretations. Finally, there are cases where the accessibility of a landing site has been questioned, based on a priori assumptions about QR, but ACD sentences have the potential to provide data bearing on such claims.
In the sections 23.2 and 23.3, I will summarize a growing number of experimental studies designed to tap into precisely these aspects of ACD. In section 23.2, we will evaluate what we know about children’s interpretation of ACD sentences, and what the experimental results say about the mechanisms that are active in the child grammar. In section 23.3, we will evaluate evidence from studies investigating adults’ interpretation of ACD sentences, which have direct bearing on linguistic theory. In the first case, evidence from ACD sentences provides us with a basis for questioning the apparent clause-boundedness of QR. In the second case, evidence from reaction times and acceptability ratings is taken to bear on theoretical assumptions regarding the nature of ACD resolution and our linguistic representations. The combined results across all tasks illustrate how experimental research on ACD sentences provides us with a unique opportunity to investigate the composition of our grammar, and push the implications and limits of linguistic theory.
23.2 INVESTIGATIONS OF ACD IN CHILD LANGUAGE ACQUISITION
One of the main goals that researchers of language child acquisition have is to determine the level of continuity in language acquisition and development. To that end, we seek to ascertain the interpretations that children assign to specific linguistic constructions as a means of crystallizing statements about the nature of their grammar, identifying the mechanisms they deploy for sentence interpretation, and determining the division of labour between what is learned and what is not. For reasons outlined in the Introduction, ACD provides us with ideal testing grounds.
23.2.1 Evidence for QR in child grammar
Given that QR is a key component of the adult grammar (under the theoretical framework being assumed here), one might ask whether children have QR as a part of their grammar. While a host of studies have examined children’s ability to interpret quantificational sentences (e.g. Lidz & Musolino, 2002; Miller & Schmitt, 2004; Musolino & Lidz, 2006), few (if any) of these studies have provided unambiguous support for the presence of QR in child grammar. For example, the ability to interpret a scopally ambiguous sentence such as (16), which could mean either that every horse does not have the property of jumping over the fence or that it is not the case that every horse has this property, could arise from reconstruction of the QNP into the VP-internal base position. The ability to access the inverse scope reading of (17) might also arise via a choice function, rather than scope shifting (e.g. Reinhart, 1997; Winter, 1997).
ACD, however, requires QR for resolution. So if children are able to successfully interpret sentences with ACD, this provides us with more reliable evidence that QR is part of the child grammar.
This was the starting point for Syrett & Lidz’s (2009) investigation of ACD in language acquisition. Four-year-olds and adult controls participated in a Truth-Value Judgement Task (TVJT; Crain & McKee, 1985; Crain & Thornton, 1998).5 Because all of the acquisition experiments reported in this section—and some of the adult experiments reported in section 23.2.2—make use of these age groups and this methodology, I will take some space here to lay out the details of such a paradigm.
In a TVJT, there are typically two experimenters: one tells the child a series of short stories using toys and props or computer images, as a puppet (played by the second experimenter) watches alongside. The child is told that the puppet is learning, and that the child’s job is to watch and listen to the story carefully along with the puppet, so that he can help the puppet learn. At the end of each story, the puppet makes a statement about what she thinks happened in the story. This statement includes the target construction—in this case, a sentence with ACD. The child assesses whether the puppet’s statement was accurate or not, given the context presented in the story. That is, the child must judge the truth-value of the puppet’s sentence, but instead of responding ‘true’/‘false’, the participant responds by saying ‘right’/‘wrong’ or ‘yes’/‘no’. If the puppet is right, she gets a special treat (e.g. a cupcake or cookie), and if she is wrong, she gets something lesser (e.g. milk or a banana); either way, she gets something. Participants are encouraged to provide a justification for the response, but with children, this is not mandatory.
In Syrett & Lidz (2009), participants were randomly assigned to one of four experimental conditions, depending on the context and the linguistic construction they heard. The construction was either an ACD sentence or a sentence with coordinated conjunction. This contrast was chosen based on previous evidence that when children appear to misinterpret a relative clause construction (which ACD contains), they seem to default to conjunction (Tavakolian, 1981).6 There were multiple trials for each condition, as there were in all TVJT experiments reviewed here.
Each of these sentence types was presented in one of two contexts. In one, there was one set of objects (e.g. frogs that two characters had to jump over), and in another, there were two sets (e.g. two sets of frogs, one for each character to jump over). In the one-set condition, each character attempts to jump over all of the frogs in succession. However, in the end, neither one does so, and consequently, neither wins a prize. In the second condition, each character jumps over all of his frogs (but none of the other character’s frogs), and each therefore receives a prize.
In the one-set condition, the ACD sentence is true, and the conjunction sentence false, while in the two-set condition, the ACD sentence is false, and the conjunction sentence true. Accordingly, both child and adult participants consistently accepted the ACD sentence in the one-set condition, and rejected the conjunction sentences, and demonstrated the reverse pattern in the two-set condition. The only mild surprise was that some adult controls were not as willing as the children were to restrict the domain of quantification in the two-set condition, and rejected the sentence, because the characters had not jumped over every frog.
The results of this experiment thus demonstrate that 4-year-old children, like adults, correctly interpret ACD sentences, and therefore that they must have the QR mechanism in their grammar. Presented with evidence that children can perform QR and interpret ACD sentences, one might further probe just how adult-like their QR and ACD resolution are. That is, do they target the same landing sites as adults? The next two experiments reported in this section were designed to address precisely this question.
23.2.2 Evidence for the landing site of QR in child grammar
Earlier, we observed that the landing site of QR in ACD sentences must be below the subject position. There are a few consequences to targeting this position. First, the QNP lands below negation, thereby allowing for NPIs in the ACD construction to be licensed. Second, ACD bleeds Principle C violations when there is a pronoun in the object position and an R-expression in the relative clause, since the R-expression raises to a position outside of the VP and therefore higher than the object, allowing for co-construal. Finally, ACD feeds Principle C violations when there is a pronoun in subject position and an R-expression in the relative clause, since even after movement, the QNP remains below the subject, and the pronoun therefore remains in a position to c-command the R-expression, preventing co-construal.
Kiguchi & Thornton (2004) took these observations from Fox and Merchant, and sought to determine how children interpreted ACD sentences in which binding principles were invoked. Their test sentences across a set of experiments had structures as indicated in (20)-(22).
(20)a. Dora gave himj the same color paint the Smurfj’s father did.
b. Dora [the same color paint the Smurfj’s father did <give himj>]
[VP gave himjt]
(21)a. *The Mermaid baked himi the same food that Cookie Monsteri did.
b. *The Mermaid [the same food that Cookie Monsteri did <bake himi>]
[VP baked himi t]
(22)a. *Hei jumped over every fence that Kermiti tried to.
b. *Hei [every fence that Kermiti tried to <jump over>]
[VP jumped over t]
In each of the scenarios, the experimenters manipulated the story so that there were two possible antecedents for the pronoun: the character denoted by the target R-expression in the sentence and another salient character of the same gender (e.g. Mickey Mouse in (20), Jabba the Hutt in (21), or Cookie Monster in (22)). Moreover, when the puppet delivered the target sentence, she mentioned the key players in the scenario, with the possible co-referent from the sentence listed last. In the ungrammatical sentences, this order of mention was intended to highlight the possibility of this character being co-construed with the pronoun.
Children responded correctly to control ACD sentences without pronominal reference (e.g. Kermit found Emily the same color egg that Gonzo did) at a rate comparable to that of adults: close to 100 per cent of the time. Children accepted sentences like (20), but rejected those like (21) (which has a Principle B violation) and (22) (which incurs a Principle C violation). (22) is the key sentence for demonstrating that the landing site for QR is below the subject, while all three sentence types demonstrate something about how children perform QR and interpret elided material.7
23.2.3 Evidence for multiple landing sites for QR in child grammar
Having established that children can correctly interpret ACD sentences, and moreover that they appear to target the same landing site as adults, we can probe their grammar further and determine whether children have the same range of accessible landing sites as adults do, or whether they are restricted to, for example, the closest landing site (arrived at via the shortest move). In Syrett & Lidz (2005, 2011), the goal was to present child and adult participants with ambiguous ACD sentences, where the ambiguity stemmed from the ACD being embedded within multiple VPs. In one set of experiments, ACD was embedded in a non-finite clause, with matrix verbs that selected for an infinitival complement (want, need, ask, invite), including possible restructuring and non-restructuring verbs.8 An example is shown in (23).
(23) The cowgirl needed to jump over every frog that the old cowboy did.
Embedded:… every frog that the old cowboy 〈jump over〉
Matrix:… every frog that the old cowboy 〈need to jump over〉
In a second series of experiments, the matrix verb took a tensed clause as a complement, resulting in ACD being embedded in a finite clause, as in (24). There is thus a crucial difference between the two cases in terms of a possible tensed clause boundary present in the latter case. This point will become relevant shortly, and will be taken up again in section 23.3 in a series of studies conducted exclusively with adults.
(24) Clifford said that Goofy read every book that Scooby did.
Embedded:… every book that Scooby 〈read〉
Matrix:… every book that Scooby 〈said that Goofy read〉
In each case, the target sentence was presented in one of two between-subject conditions: one that made the embedded reading true and the matrix reading false, and one that made the embedded reading false and the matrix reading true. The crucial question in both cases was whether children could access the matrix reading, since this would then indicate that they can perform so-called ‘long QR’ and are not restricted to the first possible landing site and can QR to a position outside of the matrix VP (Larson & May, 1990; Kennedy, 1997).
When children were presented with sentences like (23), where ACD is embedded in a non-finite clause, in either the ‘embedded’ or ‘matrix’ context, they—like adults—were able to access either of the two readings. Furthermore, participants either accepted the target sentence, justifying their response with the reading made true in the context, or rejected it, appealing to the reading that was made false. Thus, children are able to perform both ‘short’ and ‘long’ QR. This conclusion was further bolstered by a follow-up study conducted by Sugawara et al. (2013), in which a verb and tense mismatch between the matrix and embedded clause disambiguated the readings.
When all participants were presented with sentences like the one in (24), the adults and the children showed a very different pattern. Adults, for the most part, patterned as one would expect if QR were subject to a clause-boundedness constraint: they accepted the sentence in the context that made the embedded reading true (and the matrix reading false), and rejected it in the context that made the matrix reading true (and the embedded reading false). By contrast, children were prone to rejecting the target sentence, flying in the face of expectations and the principle of charity. This pattern is not only curious, because it seemed as though children are being contrary, but because if children really are overgenerating interpretations of these ACD sentences, and are too lenient with the QR operation, it seems nearly impossible to arrive at a learnability story whereby they might encounter evidence that would allow them to prune away the grammatically barred readings of these ACD sentences, which are rarely encountered in adult speech, let alone child-directed speech.9 Thus, we are faced with a classic Poverty of the Stimulus argument.
However, a closer look at the adult responses revealed that a small subset of adults were patterning in the same way, consistently accessing the matrix reading of sentences in which ACD was embedded in a finite clause. Thus, for these adults, it would appear that QR was not subject to a clause-boundedness constraint. Consequently, there would be no learn-ability story to tell: children and these adults have in common that QR is not subject to clause-boundedness. If this is the case, then two questions arise: (a) what licenses the matrix reading of these ACD sentences (i.e. what allows QR to target a matrix-VP external site), and (b) what was preventing the vast majority of adults from accessing the matrix reading in such sentences? These questions are pursued in section 23.3.
23.3 INVESTIGATIONS OF ACD IN ADULT GRAMMAR
23.3.1 ACD as evidence against clause-boundedness of QR
If QR is a covert form of Ā movement, on a par with wh-movement, it seems unexpected that QR should be subject to a clause-boundedness constraint, whereas wh-movement is not. A simple comparison between the two cases (as in (25)-(26)) highlights this difference. In (25), the wh-phrase can undergo successive cyclic movement out of the object argument position of a verb in a finite embedded clause to land in a position above the subject, whereas in (26), the universally quantified phrase in object position seems to be unable to take wide scope over the indefinite in subject position.
(25) What did a footman [VP say [CP that Mr Carson inspected t]]?
However, enough counterexamples have been presented over the years, among them those in (27) and (28), to lead to the conclusion that this is not a hard and fast constraint. (See also Farkas & Giannakidou, 1996.)
Perhaps more to the point here, the matrix reading of sentences in which QR must occur out of a finite clause appears not only to be accessible, but sometimes forced, as in (29) (a comparative construction in which the degree phrase undergoes QR) and (30)-(31), ACD sentences.
Faced with these suggestive examples paired with the data from children and adults in Syrett & Lidz (2011), Syrett (2015a) undertook a series of studies with adult participants in order to probe the availability of the matrix reading of sentences in which ACD is embedded in a finite clause, as in (32). In this case, the matrix reading would be Woody said he jumped over every frog that Jessie said that he jumped over.
(32) Woodyi said hei jumped over every frog that Jessie did.
To facilitate the availability of this reading, the experimental contexts were manipulated to support the discourse conditions that would make the matrix reading felicitous. First, the event denoted by the embedded verb (e.g. jumping) was never shown. Second, the act of reporting was made especially salient, since both characters mentioned in the sentence reported their actions to a third party (Buzz Lightyear), and these speech events were accompanied by a speech bubble above the animated images on a computer screen. Next, instead of saying, ‘I know what happened!’ at the end of each story, the puppet exclaimed, ‘I know what X said!’
In addition to manipulation of the context, the target linguistic sentence was also manipulated to facilitate processing. First, the embedded CP did not have a that complementizer. (However, a follow-up experiment demonstrated that even with the complementizer present, the matrix reading was robustly available.) Second, the embedded subject was a pronoun in lieu of a full name. Next, contrastive focus was placed on the two main characters (see Wilder, 1995, for a related observation). Finally, the ACD constructions were extended beyond just the universal quantifier to those with the same, as in (33).
(33) Jakei said (that) hei found the same treasure as Captain Hook did.
The response pattern obtained across experiments, accompanied by full-fledged justifications, leaves no doubt that the matrix reading is indeed available to adults, and therefore that QR can cross a tensed boundary. The percentage of matrix readings ranged from 37 per cent to 48 per cent at the lowest, and 94 per cent to 95 per cent at the highest. Thus, speakers allow not only ‘long QR’ out of non-finite clauses, but also finite clauses as well, and do so robustly when the discourse conditions support the matrix reading.
This ability to access the matrix reading of ACD embedded in a finite clause is actually predicted under Cecchetto’s (2004) proposal, in which QR proceeds in a successive cyclic fashion, motivated by Scope Economy and constrained by the Phase Impenetrability Condition (Chomsky, 1999/2001), whereby each sub-link of the movement is semantically motivated and crosses just one instance of v or C (each a phase). In fact, under Cecchetto’s account, once the QNP lands outside the matrix VP, it could raise even higher, again as long as the QR is semantically motivated and respects movement constraints. The prediction is that if there is an indefinite in subject position that could interact scopally with the QNP, and generate a distinct semantic interpretation, then this would license movement. This prediction is borne out.
Syrett (2015b) followed a methodology similar to the previous TVJTs, and similar to Syrett (2015a) in particular. Participants were presented with sentences such as (34) in a context that supported an ‘extra wide scope’ reading of the QNP over the subject indefinite and a matrix reading of the VPE.
(34) Someonei said hei could jump over every frog that Jessie did.
In the story corresponding to this target sentence, Jessie claims to Woody she can jump over a set of frogs. Following her departure from the scene, a group of young boys, who have overheard her boasts, enters. Each one approaches one of the frogs and makes a claim that he can jump over that frog. When Jessie returns, Woody reports back to Jessie what the boys have said. Irate, she storms off. The puppet then recaps, ‘I know why Jessie was so mad! Someone said he could jump over every frog that Jessie did.’ Adult participants accepted the sentence over 40 per cent of the time, with close to 39 per cent of the written justifications provided unambiguously pointing towards the extra wide scope reading.
23.3.2 ACD and possible evidence for theoretical frameworks
There is robust evidence that both young children and adults are able to provide offline judgements indicating their interpretation of ACD sentences, and the role of the grammar, the discourse, and the processor in licensing various interpretations. We might, then, ask whether on-line data—that is, data that are collected from real-time sentence processing studies—can provide us with further insight into the nature of the representations of ACD sentences. This was the goal of Hackl et al. (2012).
Hackl et al. (2012) presented participants with sentences as in (35)-(36) in a self-paced reading study. The factors that were manipulated included the head of the object phrase containing the relative clause (the/every), the presence of a lexical verb or a site of ellipsis (e.g. funded vs. was), and (in Experiment 2) whether the intended antecedent for ellipsis was a small or large VP (e.g. did = treat… or was reluctant to treat…) The authors predicted that, since the universal quantifier would have to undergo QR for reasons of type mismatch anyway, the semantically-motivated movement would facilitate processing in the case of ACD relative to the definite determiner, but that there would only be an advantage in the case of the embedded VP (which would implicate short QR) and not with the matrix VP (which would require long QR, and involves a more complex VP to process).
(35) The understaffed general hospital was negotiating with {the/every} doctor that the nonprofit medical organization {funded/was} in order to arrange for free vaccination clinics.
(36) The doctor was reluctant to treat {the/every} patient that the recently hired nurse {admitted/did/was} after looking over the test results.
The results for Experiment 1 demonstrated an initial slowdown for the sentences with a lexical verb in the ‘spillover region’ immediately following the main verb or auxiliary in the matrix sentence, followed by a crossover with a slowdown for the ellipsis cases. In both experiments, the ‘determiner’ the sentences incurred longer processing time than the quantificational every sentences, and the results yielded a determiner-ellipsis interaction in Experiment 1 and a determiner-ellipsis size interaction in Experiment 2 (with the-was sentences taking longer). Hackl et al. (2012) thus concluded that the presence of the universal quantifier facilitates resolution of ACD downstream, and that this finding supports a QR approach to ACD resolution.
However, the interpretation of these findings was subsequently challenged by Szabolcsi (2014), who has pointed out that under a variable-free approach, the slowdown with the can be accounted for, because a default ‘individual-forming’ interpretation of the definite determiner will not work for the interpretation of the sentence. The reinterpretation will trigger a reprocessing of the preceding material, and result in a slowdown of reading time. Moreover, if the default processing interpretation for the antecedent of ellipsis is a small VP antecedent, then this preference may account for the RTs observed in Hackl et al.’s Experiment 2.
Gibson et al. (2015) also took issue with the interpretation of Hackl et al.’s conclusion that the results support a QR theory of ACD resolution, arguing that the target sentences tipped the scales in favour of a slow-down with the, since in such cases a speaker would normally provide an additional expression highlighting the sameness of the action performed on the matrix and relative clause objects, such as same or also (their ‘sameness’ hypothesis), and that the most natural production in the target sentences would be one where there were two different lexical verbs. In Gibson et al.’s experiments, they manipulated the type of determiner (every vs. the), the presence or absence of the words same and also, and the presence or absence of a full VP and whether or not it was different (e.g. negotiate, funded), as illustrated in an abbreviated form in (37). They also excluded the spillover region, since they were obtaining acceptability ratings.
(37)a. The understaffed general hospital was negotiating with {every/the (same)} doctor that the non-profit medical organization {(also) was/funded/was negotiating with}.
b. The choreographer evaluated {every/the (same)} ballerina that the lead dancer {did/evaluated}.
Across tasks, they replicated the determiner-ellipsis interaction, but further demonstrated a significant effect of verb type (with the different verb increasing acceptability at a comparable rate for both every and the), and an effect of the presence of same and also, with both terms increasing acceptability in the sentences with the, and the presence of also lowering acceptability with sentences with every. (Same cannot be used alongside every.)
Thus, the authors concluded that the original processing data do not unambiguously signal a QR analysis of ACD sentences, and that the effects could arise via other means. Clearly, this lively exchange regarding the processing and acceptability judgements of ACD sentences highlights the reasons why ACD is a hotbed of research, and how the results arising from experimental work in this area may directly bear upon theoretical proposals about the nature of our linguistic representations.
23.4 CONCLUSIONS AND FUTURE DIRECTIONS FOR RESEARCH ON ACD
The experimental evidence collected over a range of experiments demonstrates the following. First, children as young as 4 years of age can arrive at the correct interpretation of ACD sentences, thereby indicating that the QR mechanism is part of the child grammar. Second, when children perform QR to resolve ACD sentences, they target a landing site that is outside of the antecedent VP but also below the subject, and are not restricted to the nearest landing site. Given an ACD sentence made ambiguous by ACD being embedded in multiple VPs, children and adults are able to access both candidate interpretations. This finding holds not only for ACD embedded in non-finite clauses, but also for ACD embedded in finite clauses—but in the latter case is dependent upon discourse and processing factors being controlled for successfully. Finally, on-line processing and off-line acceptability tasks with ACD sentences allow us to obtain evidence that bears potentially not only on how we represent and resolve such configurations, but also on the theoretical framework we adopt. Future research on ACD should further probe the limits of ACD configurations and interpretations, investigate the interaction of wh-words and ellipsis in ACD sentences, examine the role of binding principles in ACD configurations, and determine at a more fine-grained level what our real-time processing of ACD sentences says about the nature of our grammar and the role of extra-grammatical factors.
ACKNOWLEDGEMENT
This writing of this chapter was made possible by startup funding from Rutgers, The State University of New Jersey—New Brunswick.
1 This assumption is in contrast to an earlier proposal by Hornstein (1994) that A-movement was implicated. By now, it is standard to assume that this is not the case.
2 See section 14.4.2.1. ‘Logical Form revision’ of Brasoveanu & Dotlačil (Chapter 14 in this volume) for further discussion of QR and scope of quantification.
3 The fact that these binding relations are evaluated at LF, and not on the surface, has been observed by Fiengo & May (1994), among others.
4 For extensive discussion of the linguistic properties of NPIs and the environments that license them, see Xiang (Chapter 26 in this volume).
5 See also Gordon (1996) for further details about this methodology.
6 Based on these misinterpretations by children, as exhibited in act-out tasks, early researchers concluded that children could not actually represent relative clauses in their grammar (Tavakolian, 1981). However, this conclusion based on performance data was unwarranted, since children have no difficulty comprehending relative clauses when the felicity conditions governing the use of a relative clause are satisfied, and as control items from the TVJT experiments reported here indicate (see, in particular, Hamburger & Crain, 1982).
7 But see discussion in Syrett & Lidz (2009) concerning how a conjunction interpretation and pragmatic reasoning could yield a similar pattern. Thus, the combination of Kiguchi & Thornton (2004) and Syrett & Lidz (2009) provides evidence that children correctly interpret ACD sentences, and therefore that QR is part of child grammar.
8 See discussion in Kennedy (1997) related to claims by Hornstein (1994).
9 One need not rely on this precise kind of ACD construction, of course, to learn the limits of application, but then it is extremely challenging to identify an analogue with another linguistic construction that would allow the learner to arrive at the adult state. (See related discussion across Pullum & Schulz (2002) and Legate & Yang (2002).)