Structural intervention effects in the acquisition of sluicing

ABSTRACT

Experimental studies show that children have greater difficulty with wh-extraction from object position than subject position, arguably an intervention effect (e.g., Relativized Minimality). In this study we provide additional evidence of a S/O asymmetry in A’-dependencies from a novel source—sluicing. The results of our first comprehension study show that English-speaking 3–6-year-olds obey the “identity condition” on sluicing—that is, they disallow interpretations in which the elided verb or arguments are distinct from their antecedent. Importantly, our results also show a subject > object asymmetry and thereby support syntactic theories of sluicing that posit a fully articulated (but unpronounced) TP at the ellipsis site from which the wh-phrase has been extracted, e.g., Someone wrote this paper, but I don’t know who <_ wrote this paper>, as opposed to certain semantic/pragmatic theories that posit no such structure. Our second comprehension study investigates the role of animacy. We find that children’s comprehension of object sluices, but not subject sluices, improves significantly when there is a mismatch in animacy features. Our results are incompatible with models that are solely frequency based but rather provide evidence for structure-based intervention effects. We conclude that subject > object asymmetries can be found even in instances in which the intervener is not overt, such as sluicing, and that [animacy] may be a feature involved in the computation of intervention.

Acknowledgement

We would like to thank Angeliek van Hout, Charlotte Lindenbergh, and Kyle Johnson for their suggestions and discussion, to Mackenzie Lighterink for help testing; Alex Rodriguez for assistance in the corpus study; and all the parents and children for participating in this study.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 As per convention, the elided material is given between angled brackets.

2 Ervin-Tripp (1970:88-89) defines “mastery” as “the first of successive appropriate replies […], allowing one ‘performance’ error if it followed more than two successes.”

3 Note that a number of studies have not found an asymmetry in comprehension of unrestricted wh-questions, i.e., those with no NP following the wh element (e.g., Cairns & Hsu 1978; Friedmann, Belletti & Rizzi 2009; Hanna & Wilhelm 1992; Varlokosta, Nerantzini & Papadopoulou 2015). We will address this question in more depth in the discussion.

4 Of course, all else may not be equal. We return in the following to other possible causes of a S > O asymmetry.

5 Children often chose to resolve the sluice by pointing to the relevant hidden or visible person, rather than responding with “yes” or “no.” We counted their answers as correct if they pointed to the appropriate character on the screen.

6 We used a “circle” setting, as in Figure 8, to elicit a “yes” answer to (11a), as opposed to a “lineup” setting like the one in Figure 11, which would also elicit a “yes” answer under an adultlike interpretation, because in the latter case it would be pragmatically infelicitous for children to be asked that question if they interpret (11a) as (11b), as no one is brushing the girl in yellow in Figure 11. Moreover, these scenarios also detect if children allow a reversal of arguments, e.g., “I can see someone is brushing Ben, can you see who <Ben is brushing _>?”

7 The fact that children do marginally (though not significantly) better in the sluices than the wh-question controls seems to be an effect of Condition 3 (see Table 1). The somewhat more depressed performance may be related to the fact that these control sentences were longer and involved modified NPs (e.g., I can see that Ben is brushing someone; can you see who the girl in green is brushing _?). Indeed, in Experiment 2, where we do not have this additional NP modification, we see children do very well overall in the control wh-questions.

8 It is interesting to observe that several adult processing studies of sluicing (Frazier & Clifton 1998; Harris 2015; Lawn & Harris 2017) have found an object advantage, arguably due to the shorter linear distance between the (overt) remnant wh-element and the more local, object correlate. We believe that for young children, any potential linear locality effect between the indefinite and the remnant is being masked by the difficulties they have with structural intervention in A’-dependencies.

9 This proposal raises the following theoretical question: If sluicing can repair island violations (see section 3), why does it not also repair the intervention violation in child grammar? The answer may relate to cyclic linearization and the view that island violations are conflicts between linearization statements and spell-out (Fox & Pesetsky 2005). Assuming RM does not interact with spell-out (or linearization), and that what ellipsis removes linearization statements, and consequently, ordering contradictions, that would explain why sluicing voids islands but RM does not. In other words, impenetrability/cyclicity and intervention are separate phenomena, and while ellipsis salvages phase theory/linearization violations, it does not repair RM violations, at least in children. (Our thanks to Jeff Lidz for first bringing this question to our attention and to Kyle Johnson for discussion of this issue.)

¹⁰ More precisely, children may be unable to implement the grammatical operation (e.g., smuggling) necessary to circumvent intervention when other processes are competing for computational resources. For example, Mateu (2016) tested children’s comprehension of subject-to-subject raising seem and subject control promise and found two groups of children. One group lacked the grammatical mechanism to circumvent the intervening argument (experiencer or benefactive) and thus performed consistently poorly. The other group, which showed more variable performance, had an adultlike rule system (by hypothesis) but lacked the computational resources to parse those structures. The latter, but not the former, group showed a positive correlation with verbal processing memory score.

¹¹ Other factors may also contribute to computational complexity. Aravind, Hackl & Wexler (2018) find that children do significantly better on object clefts such as It’s a cat that the dog is chasing in felicitous (presupposition satisfaction) versus infelicitous (presupposition failure) contexts. Clefts involve an A’-dependency and previous studies found the now familiar S > O advantage (Bever 1970; Dick et al. 2004; Lempert & Kinsbourne 1980). Aravind, Hackl & Wexler attribute children’s poor performance on object clefts in the earlier studies to the fact that sentences were presented to children without a felicitous supporting context. Interestingly, however, they also show that felicity is not a factor in subject clefts where children perform well in both felicitous and infelicitous contexts. They therefore conclude that object extraction and presupposition failure incur independent processing costs and the combination overburdens the child processor. We would argue the processing cost of objects clefts is the cost associated with crossing an intervening subject.

¹² Subject sluices, on the other hand, though far fewer in number, were evenly divided between matching (2) and mismatching (2) animacy features on subject and object.

¹³ See footnote 20.

¹⁴ Varlokosta, Nerantzini & Papadopoulou (2015) examine Greek-speaking children’s performance on subject wh-questions that have an inanimate subject and an animate object, but not object wh-questions that have an inanimate subject and an animate object.

¹⁵ It is worth noting that in Experiment 1, the subject > object asymmetry in sluices essentially disappears by age 6, while in Experiment 2 even the 6-year-olds experienced greater more difficulty with object sluices than subject sluices (compare Figures 12 and 21). Experiment 2 involved more characters and included inanimate subjects, which are very infrequent in children’s input. We believe that these factors may have made Experiment 2 more cognitively demanding, resulting in asubject > object asymmetry even in the oldest children tested. In this regard we note that even with respect to the subject sluices, which are overall much easier for children than object sluices, the younger children did better in Experiment 1 than in Experiment 2.

¹⁶ In an analogous way, normal adults show a subject > object asymmetry in relative clauses in online tasks (by their nature more cognitive taxing) in, for example, reading time measures, often accompanied by slightly higher error rates in comprehension questions after reading object relative clauses (Gordon, Hendrick & Levine 2002; Just & Carpenter 1992; King & Just 1991; Mak, Vonk & Schriefers 2002; Traxler, Morris & Seely 2002; Warren & Gibson 2002, a.o.), and also show more performance errors in lexical decision and word recall tasks taking place during or after reading object relative clauses (David & Waters 1999, a.o.; Waters, Caplan & Hildebrandt 1987).

¹⁷ See Appendix C for a breakdown of the results by subcondition, i.e. (i) [+animate] subject and [+animate] object, (ii) [−animate] subject and [−animate] object, (iii) [+animate] subject and [−animate] object, and (iv) [−animate] subject and [+animate] object, for both the subject and object sluices.

¹⁸ Though not included in this article, we also conducted a Relative Clause experiment manipulating animacy in the same way as Experiment 2, with the same children in a separate session. We found that children performed significantly worse with object RCs than subject RCs and that animacy mismatches improved their performance in the object RCs but not subject RCs, corroborating the hypothesis that [animacy] mismatches improve children’s comprehension of object A’-constructions. We also found a significant correlation in performance between the two, r_S(58) =.501, p <.001, further supporting the hypothesis that sluices are also derived via A’-movement (Merchant 2001).

¹⁹ Processing (e.g., memory)-based approaches such as Similarity-based Interference (Gordon, Hendrick & Levine 2002; Lewis, Vasishth & Van Dyke 2006; Van Dyke & McElree 2006; 2011, a.o.) also appeal to a notion of intervention but differ from RM in that they are nonstructural and do not rely specifically on morphosyntactic features. Such proposals argue instead that it is simpler to keep dissimilar expressions in memory than similar ones and importantly, the dissimilarities may be along any dimension, i.e., morphosyntactic, or purely semantic or phonological. Our study was not designed to disentangle a grammar-based approach from such processing-based approaches. However, to our knowledge, processing approaches do not predict interference if the distractor is not overtly realized, as is the case in sluiced wh-questions.

²⁰ Some structural-based accounts of intervention, such as RM, propose only features associated with T and overtly realized on the verb can figure into the computation of intervention in a specific language (Belletti et al. 2012). Insofar as [animacy] is not a movement-triggering feature in English, these claims are incompatible with our results. They are similarly incompatible with findings in other languages in which [animacy] is not active in the T system but which nevertheless find that animacy mismatches improve comprehension of intervening structures (e,g., Arosio, Guasti & Stucchi 2011; Durrleman, Bentea & Guasti 2016 for Italian; Costa, Grillo & Lobo 2012; Corrêa 1995 for Portuguese; Adani 2012 for German; Varlokosta, Nerantzini & Papadopoulou 2015 for Greek).

Additional information

Funding

This work was supported by the National Science Foundation [BCS-1451589]; UCLA Faculty Senate COR Grant.