Selective vulnerability in very advanced L2 grammars: Evidence from VPE constraints

Abstract

This study examines whether the second language acquisition (L2A) of syntactic properties at the interfaces is problematic for L2 learners. English verb phrase ellipsis (VPE) was tested as an interface property which involves feature interpretability. Two subtle contrasts of VPE at different grammar-internal interfaces were examined: (a) copula be and lexical verb ellipsis at the syntax–lexicon interface (*John is here, and Mary will too vs. John slept, and Mary will too), and (b) progressive be and perfective have stranding at the syntax–semantics interface (*John slept, and Mary was too vs. Peter saw your parents last week, but he hasn’t since). Unlike (a), the contrast in (b) requires identifying (un)interpretable features on -ing and -en to recover the semantics of the elided material. Since Saudi Arabic does not license VPE, (very-)advanced English L2 speakers whose first language (L1) is Saudi Arabic were tested. The results from a bimodal timed acceptability judgment task showed they were more target-like on the contrast at the syntax–lexicon interface than they were on the syntax–semantics interface. They particularly deviated from target-like judgments on have stranding which requires recognizing the perfective uninterpretable feature on -en. These results suggest selective vulnerability at grammar-internal interfaces whenever uninterpretable features are involved.

Keywords

feature interpretability selective vulnerability syntax–lexicon syntax–semantics verb phrase ellipsis

I Selective vulnerability

Understanding the phenomenon of selective vulnerability in second language acquisition (L2A) is still an unresolved puzzle. Selective vulnerability here is understood as failure to develop target-like L2 grammars in certain domains but not others. In the current version of the Interface Hypothesis¹ (Sorace, 2011; Sorace and Filiaci, 2006; Tsimpli and Sorace, 2006), the locus of L2 vulnerability is at the interface between the syntax and grammar-external cognitive domains, such as discourse and pragmatics. This version developed from comparing the acquisition of formal properties at grammar-internal interfaces and the task of integrating these formal properties with other cognitive domains at the grammar-external interface. According to Tsimpli and Sorace (2006), properties relating to pronominal subjects at the syntax–discourse interface are more subject to vulnerability than properties such as object-fronting at the syntax–semantics interface. They argued that ‘the distinction between the two interfaces is based on the assumption that the syntax–discourse interface is a “higher” level of language use, integrating properties of language and pragmatic processing, whereas syntax–semantics involve formal properties of the language system alone’ (Tsimpli and Sorace, 2006: 653). One problem with this account is that the status of the L2A of grammar-internal interfaces is not addressed in any detail. It is not made explicit what role formal properties play in rendering internal interfaces less vulnerable. Also, it is presumably expected that linguistic properties at the same internal interface are equally vulnerable, which is not necessarily the case.

The aim of this article is to supplement the study of interface phenomena with predictions of the Interpretability Hypothesis (Tsimpli and Dimitrakopoulou, 2007) about feature interpretability in the syntax. According to the Interpretability Hypothesis, properties bearing uninterpretable features not activated in L1 are vulnerable, whereas those bearing interpretable features are not, even when they are not activated in L1. Two subtle cases for English verb phrase ellipsis (VPE) are examined in partial morphological identity with the antecedent verb: (a) ellipsis of copula be and lexical verbs following a stranded modal (e.g. *John was here, and Mary will ~~be here~~ too vs. John slept, and Mary will ~~sleep~~ too), and (b) progressive be and perfective have stranding (e.g. *John slept, and Mary was ~~ing sleep~~ too vs. Peter saw your parents last week, but he hasn’t ~~en see~~ since) (examples from Lasnik, 1999; italics and strikethrough by author). To acquire the contrast in (a), L2 speakers need to distinguish between lexical verbs and copula be in terms of their representation in the lexicon (non-inflected vs. inflected, respectively) (Lasnik, 1999).² To acquire the contrast in (b), L2 speakers need to distinguish the formal features encoded on the stranded affixes -ing and -en in the elliptical site (interpretable progressive features vs. uninterpretable perfective features, respectively) (Rouveret, 2012). Whilst the contrast in (a) is captured at the syntax–lexicon interface, the contrast in (b) is implicated at the syntax–semantics interface and requires identifying (un)interpretable features to recover the semantics of the elliptical material. The aim of comparing the contrasts (a) and (b) is to uncover whether structures at different grammar-internal interfaces (i.e. syntax–lexicon vs. syntax–semantics) are equally problematic. This study proposes that feature interpretability could help us understand diverging performance at grammar-internal interfaces. The contrast between -ing and -en at the syntax–semantics interface is a test of whether structures at the same interface could also be equally problematic. Overall, these subtle contrasts illustrate a poverty of the stimuli phenomenon because they are under-determined by the input and not taught to L2 speakers. Since Saudi Arabic does not license VPE, L1 Saudi Arabic speakers were targeted in this study in order to examine the extent to which they can acquire such subtle distinctions.

The remainder of this article is organized as follows: first, a parametric description of the licensing of VPE is given, followed by a syntactic explanation of the core VPE contrasts. Then, existing studies on these contrasts are briefly reviewed. After that, the research questions and the methodology chosen for this empirical study are presented. Finally, the results are reported and discussed with suggestions for future research.

II The licensing of VPE

VPE is a PF-deletion process of a complete syntactic structure (Lasnik, 1999; Rouveret, 2012). In English, the licensors must be hosted by Tense through movement or base-generation (Aelbrecht, 2010; Rouveret, 2012).³ Thus, the licensing condition involves realizing the contrast in verb-movement constraints between English auxiliaries and lexical verbs, as in (1–3), taken from Aelbrecht (2010: 166–167). The licensors are marked in bold, and the elided material with a ~~strikethrough~~ throughout the examples.

(1) Uriel was drinking coffee and Aviad was [~~drinking coffee~~] too.

(2) *Ed doesn’t like cats and dogs, but Chris likes [~~cats and dogs~~].

(3) Ed doesn’t like cats and dogs, but Chris does [~~like cats and dogs~~].

The ungrammaticality of (2) has been explained in terms of verb-movement constraints on English main verbs. Rouveret (2012: 948) proposed that main verbs move to little v so they can c-command the VP and in principle could license ellipsis of the material in the verb’s complement. However, English main verbs (unlike auxiliaries) cannot move further to Tense, which must be lexically filled to license VPE.⁴ If Tense is lexically empty, do support is obligatory for lexicalizing the tense feature, and thus, licensing ellipsis of the material in VP (3). Among the licensors for English VPE are modals (4) and infinitival to (5), which are lexically hosted by Tense. The examples are taken from Aelbrecht (2010: 166).

(4) Jasmin can draw an elephant, but Ryan can’t [~~draw an elephant~~].

(5) Bettina couldn’t make it, but she really wanted to [~~make it~~].

Different from English, Arabic does not license VPE regardless of it being a verb-movement language. Unlike English auxiliary verbs and modals, Saudi Arabic verbs are not VPE licensors and thus cannot be stranded, as in (6–7). Instead, the whole TP, including the moved verb, needs to be elided (i.e. TP-ellipsis), as in (8). The examples in (6–8) are from a Hejazi variety. The lack of VPE in Saudi Arabic has also been reported with a Najdi variety (Abdulkarim and Roeper, 1997).

(6) * Nizaar kaan yi-gra bass Omar maa kaan [~~yi-gra~~].

Nizar was reading but Omar not was reading

‘Nizar was reading, but Omar wasn’t.’

(7) * Nizaar gaaʕid yi-gra bass Omar mu gaaʕid [~~yi-gra~~].

Nizar Prog reading but Omar not Prog reading

‘Nizar is reading, but Omar isn’t.’

(8) Nizaar yi-Hubb Ɂs-safar wu Omar [~~yi-Hubb Ɂs-safar~~] kamaan.

Nizar loves traveling and Omar loves traveling too

‘Nizar loves traveling, and Omar too.’

The correlation between VPE and verb-movement is not attested cross-linguistically (Cyrino and Matos, 2005; Rouveret, 2012). Instead, Cyrino and Matos (2005) proposed that the licensing of VPE correlates with the degree of aspect grammaticalization in a given language, and therefore the projection of an independent aspectual functional category (AspP) below Tense. For English, they proposed that AspP is an extension of the vP (labeled as Asp/vP), and thus Tense holds a local c-command relation over vP, where Asp is part of the elliptical site, as in (9b). On the other hand, in verb-movement languages without VPE (e.g. French, Spanish, Italian and German), they proposed that AspP is an independent projection intervening between Tense and vP because Asp is highly grammaticalized. To illustrate with French, in (10b), the verb which has moved to Tense does not locally c-command the elliptical site because of the intervening AspP, and thus VPE is not possible. The examples (9a–10a) are from Cyrino and Matos (2005: 106).

(9) a. John has seen his friends lately and Peter has ___too.

b. … and

(10) a. * Tu vas au cinéma ce soir et Paul va ___ aussi.

you go to the cinema this evening and Paul goes too

b. * … et

Like other verb-movement languages without VPE, the verbal system of Saudi Arabic shows aspect grammaticalization,⁵ which suggests the projection of AspP intervenes between Tense and vP. It is assumed that because AspP violates the locality of the c-command relation between Tense and the elliptical site, VPE is not allowed in Saudi Arabic.

III The identity and recoverability conditions in VPE

Two subtle contrasts of VPE from Lasnik (1999) are discussed as properties at grammar-internal interfaces. The first contrast (11–12) is between lexical verb and copula be ellipsis when the antecedent and the elided verbs are morphologically non-identical: the antecedent verb is marked in italics throughout the examples.

(11) John slept, and Mary will ~~sleep~~ too. (Lasnik, 1999: 108)

(12) *John was here, and Mary will ~~be here~~ too. (Lasnik, 1999: 109)

The contrast in acceptability here is assumed to be captured at the syntax–lexicon interface, where copula be (unlike lexical verbs) is inflected in the lexicon. Under Lasnik’s (1999) account, lexical verbs are inflected in the syntax through morphological merge, whereas auxiliaries are inflected in the lexicon. The identity condition, which assures that the antecedent and the elided verbs are identical, is thus computed at the root level before morphological merging with affixation (Lasnik, 1999). In (11), verb ellipsis applies at the root level [sleep] before affix hopping [_T -ed] + [_v sleep]. VPE here is possible because the antecedent and the elided verb, sleep, shared identical roots when they entered the derivation. Conversely, ellipsis of copula be in (12) is disallowed because the antecedent was and the target form be in the ellipsis site were non-identical forms when they entered the derivation. Unlike copula be, lexical verbs can ignore morphological parallelism because they are inflected in the syntax not in the lexicon. Although Lasnik (1999) captured this contrast in terms of violating morphological identity, Rouveret (2012) proposed a lexical identity account. Whilst both forms of sleep in (11) have the same lexical entry, the forms was and be in (12) belong to distinct lexical entries, which violate the condition in (13):

(13) Lexical Identity Condition: ‘Every lexical item in the numeration of the elided VP must be identical to an item in the numeration of the antecedent VP’ (Rouveret, 2012: 930).

Turning now to the second contrast (14–15), which is between progressive be and perfective have stranding when preceded by a finite antecedent:

(14) *John slept, and Mary was ~~sleeping~~ too. (Lasnik, 1999: 111)

(15) Peter saw your parents last week, but he hasn’t ~~seen them~~ since. (Lasnik, 1999: 113)

Although VPE is possible with a non-identical finite antecedent, it is disallowed with a stranded auxiliary be [-ing]+[~~sleep~~] (14), but allowed with a stranded have [-en]+[~~see~~] (15). The contrast in acceptability between be and have stranding is captured at the syntax–semantics interface. Rouveret (2012) explained this contrast by considering feature interpretability on the stranded affixes, proposing the recoverability principle in (16).

(16) Principle of Recoverability: ‘A VP-elided constituent cannot contain any non-recoverable interpretable feature’ (Rouveret, 2012: 911).

Rouveret (2012) argued that the two aspectual affixes differ in feature interpretability: whereas -ing encodes a progressive interpretable feature, -en encodes a perfective uninterpretable feature. His argument was based on the meaning contribution of these aspectual affixes. If the grammatical morpheme can distinguish meanings, then it is assumed to encode an interpretable feature specified for that meaning. In English, the -en affix occurs in perfective verb forms (e.g. have eaten) and passive forms (e.g. be eaten). The contrast in meanings between (have eaten vs. be eaten) is resolved by the auxiliary rather than -en. Therefore, it is assumed that -en does not convey the perfect meaning and thus bears an uninterpretable feature. In contrast, the -ing morpheme only occurs in progressive verb forms (e.g. be eating). Therefore, it is assumed that -ing rather than be must carry the progressive interpretation. Whilst the perfect meaning is encoded in auxiliary have, the progressive meaning is not encoded in auxiliary be. Instead, be encodes a kind of ‘stative meaning’ with passive (e.g. be eaten), progressive (e.g. be eating) and copular (e.g. be happy) constructions.

Given the above reasoning, the contrast in (14–15) is explained as follows: according to the principle of recoverability (16), any affixes with interpretable features, such as -ing, cannot be deleted unless they are recoverable from the antecedent clause. Interpretable features function at the syntax–semantics interface because they are required at LF for semantic interpretation. When the -ing is deleted from the ellipsis site, the progressive meaning is non-recoverable from either the stranded auxiliary was or the antecedent (14). Conversely, when the -en is deleted from the ellipsis site, the perfective meaning is still recoverable from the stranded auxiliary, have (15).

A subtle contrasting behavior of -ing and -en also arises in the antecedent clause, as in (17–18), taken from Lasnik (1999: 109). The (?) mark is understood to show that the status of (17) is marginally unacceptable.

(17) ^?John was sleeping, and Mary will ~~sleep~~ too.⁶

(18) John has slept, and Mary will ~~sleep~~ too.

According to the lexical identity condition, the contrast in the degree of acceptability of the lexical verb ellipsis between (17) and (18) should be surprising. If identity is assessed before morphological merger and the verb forms sleep are lexically identical, why should verb ellipsis be less acceptable when the antecedent contains a progressive participle [was]+[ing]+[sleep] compared with a perfective participle [has]+[en]+[sleep]? Hagstrom (1994: 4), wrote that the reason why the status of (17) is questionable is probably because it can also be interpreted as ‘John was sleeping, and Mary will ~~be sleeping~~ too’ (footnote 5). However, in this interpretation, was and be are non-identical forms, and thus (17) should be completely ruled out if both forms entered the derivation fully inflected (compare with 12). It is plausible to assume that the relative unacceptability of (17) is due to a semantics mismatch in the computation of identity. Because of its meaningful content, -ing (as opposed to -en) requires strict morphological identity to generate parallel semantic/aspectual interpretations between the antecedent and the elliptical material. For -ing, violation of the identity requirement, as in (17) seems less severe than the recoverability requirement, as in (14). Hence, (17) is marginally unacceptable, whereas (14) is completely ungrammatical.

To summarize, the identity and recoverability conditions in English VPE are quite complex. The identity condition is assessed at the syntax–lexicon interface, whereas the recoverability condition is assessed at the syntax–semantics interface where feature interpretability needs to be identified on the stranded auxiliaries. Unlike interpretable features, uninterpretable features are syntax-internal and do not contribute to meaning, and are thus deleted before the derivation is transferred to the semantics (Rouveret, 2012).

IV Existing L2A studies

There has been considerable psycholinguistic research investigating parallelism effects on VPE compared with VP anaphora, from a semantics-processing approach (e.g. Duffield and Matsuo, 2009; Duffield et al., 2009), or a syntax-processing approach (e.g. Arregui et al., 2006). None of the cited studies have investigated morpho-syntactic parallelism of the type tested in this current study. Nonetheless, Hawkins (2012) did examine the same contrasts (among other cases) in L2 English by L1 Arabic and Chinese speakers. The L2 speakers were of advanced, intermediate and low-intermediate English proficiency. Adopting a syntactic-processing approach, Hawkins assumed that although these VPE contrasts are syntactically-determined, the gradience effect in the acceptability of non-parallel structures is a performance rather than a grammar phenomenon. In a sentence completion judgment task, the VPE stimuli appeared following a two to three sentence-long context to challenge the participants to identify the elided material from the given context. The participants had to judge the VPE stimuli by choosing one of three options (perfect, possible or impossible). The task was untimed and had 60 items presented in written form. In target-like judgments, the L2 speakers rejected copula be ellipsis (e.g. Jill is very successful … #Sue will too), but accepted lexical verb ellipsis (e.g. Jack sent Jill a letter … Mary will too), where both structures had finite antecedents. However, the L2 speakers gave non-target-like judgments by equally rejecting perfective have stranding (e.g. Jack wrote Jill a letter … Mary has too), and progressive be stranding (e.g. Jack sent Jill a letter … #Mary was too), where both structures had finite antecedents. Hawkins’ (2012) explanation was that the L2 speakers treated the participle affixes -en and -ing as bearing interpretable features which are non-recoverable. Since neither Arabic nor Chinese have an uninterpretable perfective feature, they potentially did not recognize it in -en. Hawkins (2012: 434) called for L2 data from near-native L2 English speakers to verify whether ‘the same insensitivity to the uninterpretable status of -en’ would eventually disappear and thus challenge the Interpretability Hypothesis.

An earlier study by Al-Thubaiti (2009, 2010) examined these specific contrasts with L1 Saudi Arabic instructed English L2 learners. They were college students of various proficiency levels: beginners, low-intermediate, high-intermediate and advanced. Their accuracy results from a timed acceptability judgment task showed that perfective have and progressive be stranding with a finite antecedent were equally rejected (e.g. John plays football, but Tom hasn’t recently vs. *Heather visits her mother, but Mary isn’t yet). This finding was taken to indicate a potential lack of uninterpretable features on -en which is not available in L1 Saudi Arabic. Nonetheless, L2 data from highly-proficient speakers is required to verify the current findings, which the present study is designed to test.

V Research questions and predictions

Since Saudi Arabic does not license VPE, the following research questions were examined with (very-)advanced English L2 speakers whose L1 was Saudi Arabic:

Can L2 speakers recognize aspectual auxiliary verbs and modals as opposed to simple lexical verbs as possible syntactic licensors for VPE?

(a) John is sleeping, and Mary is ~~sleeping~~ too.

(b) John will sleep, and Mary will ~~sleep~~ too.

Can L2 speakers distinguish (un)grammatical conditions of VPE, where the antecedent and the elided verbs are non-identical?

(d) *John was here, and Mary will be too.

(e) John slept, and Mary will ~~sleep~~ too.

(f) *John slept, and Mary was ~~sleeping~~ too.

(g) Peter saw your parents last week, but he hasn’t ~~seen them~~ since.

The first question tests whether L2 speakers could distinguish modal-stranding and auxiliary-stranding as possible VPE constructions as opposed to *verb-stranding. Unlike English, Tense in Saudi Arabic is instantiated with a strong verbal uninterpretable feature which requires verb-movement. If resetting feature strength is possible, the L2 speakers should be able to restrict verb-movement to auxiliaries, and thus distinguish between non-raising lexical verbs and raising aspectual auxiliaries. Since VPE licensors must be in Tense, they should be able to realize aspectual auxiliaries as well as modals as legitimate syntactic licensors, unlike lexical verbs. The Interpretability Hypothesis (Tsimpli and Dimitrakopoulou, 2007) does not make explicit predictions about resetting the feature strength of uninterpretable features. Instead, the claim is that acquiring uninterpretable features not activated in L1 is predicted to be problematic. Hence, this question examines whether resetting the strength of uninterpretable features is as problematic as acquiring uninterpretable features from scratch.

The second question examines L2 performance on two grammar-internal interfaces, which is the focus of the study. The first contrast between copula be and lexical verbs in (d–e) is constrained by the lexical identity condition at the syntax–lexicon interface. The second contrast between be and have stranding in (f–g) is constrained by the recoverability principle at the syntax–semantics interface and involves distinguishing (un)interpretable features on -ing and -en. As discussed earlier, the Interface Hypothesis (Sorace, 2011) does not offer testable predictions about either interface but instead makes broad claims about interface properties. The present study was designed to understand whether grammar-internal interfaces are equally problematic. According to the Interpretability Hypothesis, the tested conditions should not be equally problematic. Non-target-like performance is predicted on the syntax–semantics interface because it involves uninterpretable features not activated in L1 (as with -en), but target-like performance is predicted on the syntax–lexicon interface contrast because it does not involve distinguishing feature interpretability on copula be or lexical verbs.

VI Method

1 Participants

Forty-nine participants took part in this study, comprising 34 highly-proficient Saudi Arabic L2 speakers of English and 15 English native speakers (NSs) as controls. The average age at testing was 26.80 years (19–56) for the controls, and 27.18 years (20–37) for the L2 speakers. The English controls were graduate students at UK universities at the time of testing. The L2 speakers were English professionals recruited from the population of university instructors in Saudi Arabia. They each had a minimum of eight years of continuous experience of using English in their profession (M = 20.06, SD = 6.88).

The English proficiency of the L2 speakers was evaluated by the Quick Oxford Placement Test (QOPT; Max = 60) and Nation’s (2001) Vocabulary Levels Test (VLT; Max = 90). Each participant was assigned a proficiency score out of 150, which was the total of their scores on both tests. All 34 participants had scored in the (very-)advanced levels of the Association of Language Testers in Europe (ALTE). To form two proficiency groups based on their combined score, a hierarchical cluster analysis was run on the 34 eligible cases. Applying Ward’s method, two groups were produced and labelled as advanced and very advanced. In an independent samples t-test, the two groups differed significantly with a large effect size, t (32) = 11.81, p < .001, d = 4.08. Their details are summarized in Table 1.

Table 1.

L1 Saudi Arabic-L2 English speakers’ details (n = 34).

Groups	n	Proficiency score (out of 150)	Age at first exposure	Years of exposure
		Mean (SD)	Mean (SD)	Mean (SD)
Very advanced	15	143.13 (4.49)	5.73 (3.79)	23.93 (7.32)
Advanced	19	128.53 (2.67)	8.21(3.74)	17.00 (4.76)

2 Material

An Acceptability Judgment Task (AJT) of 164 stimuli (half of them grammatical and half ungrammatical) was constructed to test a variety of morpho-syntactic structures (resumption, verb-raising effects and VPE) which differ between Saudi Arabic and English. For VPE, there were 48 stimuli and eight ungrammatical distracters to balance the distribution of (un)grammatical conditions (28 grammatical and 28 ungrammatical) (see Table 2). The full list of VPE material is given in Appendix 1.

Table 2.

Summary of VPE constructions by elided and antecedent verb forms (identical vs. non-identical) with examples and numbers of tokens.

Elided form	Antecedent form		Tokens
Elided form	Identical	Non-identical	Tokens
1. Modal-stranding:
Non-finite copula be	Non-finite, e.g. John will be here, and Mary will ~~be here~~ too.	Simple finite, e.g. *John is happy, and Mary will ~~be happy~~ soon.	4g+4ug
Non-finite lexical verb	Non-finite, e.g. Bruce can design a webpage, and George can ~~design a webpage~~ too.	Simple finite, e.g. Mike revises his work, and Jennifer should ~~revise her work~~ as well.	4g+4g
		Progressive participle, e.g. ^? Bill is writing an essay, and Tom should ~~write an essay~~ as well.	4ua
		Perfective participle, e.g. John has baked a cake, and Mary will ~~bake a cake~~ too.	4g
2. Auxiliary-stranding:
Progressive	Progressive, e.g. John is watching TV, but Mary isn’t ~~watching TV~~ yet.	Simple finite, e.g. Heather visits* her mother, but Mary isn’t ~~visiting her mother~~ yet.	4g+4ug
Perfective	Perfective, e.g. Claire has bought a new house, but Mary hasn’t ~~bought one~~ yet.	Simple finite, e.g. John plays football, but Tom hasn’t ~~played football~~ recently.	4g+4g
3. Lexical verb-stranding:
Simple finite	Simple finite, e.g. Neal grows* ~~vegetables~~, and his wife grows vegetables too.		8ug
4. Ungrammatical distracters:
Progressive		Perfective, e.g. Tom has slept* on the sofa, but Bill isn’t ~~sleeping~~ yet.	4ug
Perfective		Progressive, e.g. Bob is reading* a novel, but Mary hasn’t ~~read one~~ yet.	4ug

Note. For illustration, the antecedent appears in italics, licensors in bold and elided verbs with a ~~strikethrough~~. ug = ungrammatical, ua = marginally unacceptable, g = grammatical.

As shown in Table 2, three types of VPE were tested in conjunct sentences with morphologically (non)identical verb antecedents. The elided verb was assessed in the context of the antecedent conjunct. To examine the condition of lexical identity, there were two sets of modal-stranding constructions involving ellipsis of non-finite copula be and lexical verbs, each under conditions of identical and non-identical verb antecedents to the elided verb (non-finite vs. simple finite). To further test the assumption that lexical verb ellipsis takes place before morphological merger, two other sets were constructed with two types of participle antecedents non-identical to the elided lexical verb: progressive and perfective participles. For the recoverability principle, there were two sets involving progressive be and perfective have stranding tested under (non)identical conditions (progressive/perfective participle vs. simple finite). To test the licensors, eight items on lexical verb-stranding were constructed to be compared with modals and aspectual auxiliary stranding with identical antecedents. All stimuli were counter-balanced in terms of coordinator type (and, but) so that half of the stranded modals and auxiliaries were followed by negation.

3 Procedure and analysis

Two pseudo-randomized versions of the AJT were administered individually. Using a power-point slide show, the stimuli were presented aurally and in writing. The transition time between the slides was fixed and the participants were not allowed to move forward from one slide to another. This dual modality of presentation was chosen to control for response time. It was crucial that the task was timed in order to reduce the chances of participants relying on explicit knowledge in a judgment task (Ionin, 2012; Sorace, 1996). Research has also shown that L2 speakers perform better on ungrammatical structures when they are presented using a dual modality (Murphy, 1997). Each stimulus was read twice by a female English NS and displayed for eight seconds on the computer screen. During the display time, the participants had to circle their judgments on a prepared paper. They were asked to rate the stimuli on a five-point scale ranging from 1 (definitely impossible) to 5 (definitely possible). The mid-point 3 was interpreted as an intermediate judgment which lay between both extremes on the scale (Cowart, 1997). To ensure eliciting ‘sentence intuitions’, the participants had a thorough instruction and practice session. Using the set of instructions compiled by Bley-Vroman et al. (1988), they practiced how to rate sentences with examples from Saudi Arabic followed by English. Two issues were stressed in the practice session: (a) what they felt about the sentence; and (b) their ‘true’ judgment regardless of what they were taught. The practice items were unrelated to the tested conditions. To control for fatigue and order effects, the 164 stimuli were presented in four blocks (with 41 items per block), and each block had one item of each condition (see Ionin, 2012). For all participants, the actual judgment task was completed in almost 21 minutes, excluding the practice session and interval breaks between blocks. In a different session, they were given a proficiency test and a background questionnaire.

All statistical analyses (with alpha set at p < .05) were conducted on the mean acceptance and rejection ratings per condition. Within-group analyses were chiefly performed to test whether the L2 speakers could distinguish the different contrasting conditions. Therefore, the L2 speakers’ judgments were not directly compared with those of the NSs, but with their pattern of performance on the paired conditions.

VII Results

1 The licensing of VPE

Before examining any effects for morphological parallelism, the L2 speakers were examined on whether they could distinguish (un)grammatical stranding structures. In target-like grammar, lexical verb-stranding is ungrammatical (e.g. … *and his wife grows ___ too), whereas modal-stranding (e.g. …, and Mary will __ too) and auxiliary-stranding (e.g. …, but Mary isn’t __ yet) are possible manifestations of English VPE. In the tested cases, the antecedent conjunct and the elliptical material were identical. Table 3 presents the mean ratings for the NSs and the (very-)advanced L2 groups.

Table 3.

Mean ratings (SD) of three types of stranding structures with morphologically identical verb antecedents.

Groups/conditions	Modal-stranding (k = 8)	Auxiliary-stranding (k = 8)	*Verb-stranding (k = 8)
Native speakers (n = 15)	4.43 (.84)	4.55 (.76)	2.55 (1.11)
Very advanced (n = 15)	4.36 (.45)	4.53 (.56)	2.55 (.87)
Advanced (n = 19)	4.36 (.42)	4.36 (.45)	2.53 (.96)

Note. * indicates ungrammatical VPE.

For all the groups, repeated-measures ANOVAs were conducted with stranding type (modal-stranding, auxiliary-stranding and *verb-stranding) as the within-participant factor. The results showed a significant effect of stranding type with a very strong effect size on the mean ratings achieved by the NSs group, F (1.08,15.11) = 54.71, p < .001, $η_{ρ}^{2}$ = .80, the very advanced group, F (2,28) = 55.72, p < .001, $η_{ρ}^{2}$ = .80, and the advanced group, F (1.31,23.62) = 67.16, p < .001, $η_{ρ}^{2}$ = .79. In Bonferroni comparisons, the results for all groups consistently showed that the mean ratings on *verb-stranding were significantly lower than those on modal-stranding (p < .001) and auxiliary-stranding (p < .001). In contrast, the results showed no significant differences between the acceptance ratings of modal-stranding and auxiliary-stranding by the NSs group (p = .17), the very advanced group (p = .75) and the advanced group (p = 1.0).

Although all groups gave significantly lower ratings to *verb-stranding, the ratings seem higher than expected for an ungrammatical structure (see Table 3). Hawkins (2012: 206) raised a situation in which these structures could be felicitous if the deleted material were interpreted as arbitrary null objects rather than VPE. In English, some transitive verbs (e.g. read, write) optionally allow arbitrary null objects. For example, the deleted material in the second clause of Jack wrote Jill a letter. Mary wrote ___too, can be interpreted as (Mary wrote something else) and not as (Mary wrote #Jill a letter too). The possibility of such an alternative interpretation could explain why the ratings were higher than expected. Nonetheless, what matters most here is that all groups distinguished between *verb-stranding on the one hand, and modal-stranding and auxiliary-stranding on the other. They also gave significantly comparable high acceptance ratings to both modal-stranding and auxiliary-stranding as grammatical manifestations of VPE.

Having shown through within-group analyses that the L2 speakers could distinguish (un)grammatical VPE constructions, the following analyses were conducted to examine the lexical identity condition and recoverability principle.

2 The lexical identity condition: Modal-stranding

The first analysis examined the contrast between non-finite copula be and lexical verb ellipsis following a modal in (non)identical conditions (see Table 4). Whilst ellipsis of copula be is only licensed with an identical verb antecedent, lexical verb ellipsis is possible with a non-identical antecedent.

Table 4.

Mean ratings (SD) of copula be and lexical verb ellipsis with morphologically (non)identical antecedents.

Groups/conditions	Copula be		Lexical verb
Groups/conditions	Identical (k = 4)	*Non-identical (k = 4)	Identical (k = 4)	Non-identical (k = 4)
Native speakers (n = 15)	4.27 (.85)	1.33 (.37)	4.58 (.89)	4.20 (.89)
Very advanced (n = 15)	4.02 (.70)	1.95 (.86)	4.70 (.33)	4.42 (.51)
Advanced (n = 19)	4.09 (.66)	2.36 (.73)	4.63 (.32)	4.16 (.64)

Note. * indicates ungrammatical VPE.

For each group, repeated-measures ANOVAs were conducted with verb type (copula be and lexical verb) and morphological parallelism with the verb antecedent (identical and non-identical) as within-participant factors. The results showed a significant main effect for verb type with higher acceptance ratings of lexical verb ellipsis by the NSs group, F (1,14) = 106.76, p < .001, $η_{ρ}^{2}$ = .88, the very advanced group, F (1,14) = 95.37, p < .001, $η_{ρ}^{2}$ = .87, and the advanced group, F (1,18) = 92.52, p < .001, $η_{ρ}^{2}$ = .84. The results also showed a significant main effect for parallelism with higher acceptance ratings of identical antecedents by the NSs group, F (1,14) = 114.45, p < .001, $η_{ρ}^{2}$ = .89, the very advanced group, F (1,14) = 79.43, p < .001, $η_{ρ}^{2}$ = .85, and the advanced group, F (1,18) = 50.78, p < .001, $η_{ρ}^{2}$ = .74. Most important, the results showed a significant interaction effect between verb type and parallelism on the ratings by the NSs group, F (1,14) = 112.39, p < .001, $η_{ρ}^{2}$ = .89, the very advanced group, F (1,14) = 65.26, p < .001, $η_{ρ}^{2}$ = .82, and the advanced group, F (1,18) = 25.46, p < .001, $η_{ρ}^{2}$ = .59. To uncover the source of the interaction effect, Bonferroni pairwise comparisons were performed for each group on three paired conditions as described next.

Looking first at copula be, all groups gave acceptance ratings to copula be ellipsis with identical antecedents, but rejection ratings to non-identical antecedents (see Table 4). In the Bonferroni comparisons, there were significant differences between the acceptance and rejection ratings of the NSs group (p < .001, d = 3.10), the very advanced group (p < .001, d = 2.71) and the advanced group (p < .001, d = 1.68). The magnitude of the mean difference in the ratings of the L2 groups showed improvement with increased proficiency.

Turning to lexical verbs, all groups gave acceptance ratings to lexical verb ellipsis with both (non)identical antecedents, but they gave slightly lower acceptance ratings to non-identical ones (see Table 4). In the Bonferroni comparisons, there were significant differences between the acceptance ratings of the NSs in favor of identical antecedents (p = .003, d = .93). The L2 groups’ ratings varied statistically by proficiency. Whilst the very advanced showed marginally non-significant differences between (non)identical antecedents (p = .059, d = .55), the advanced gave significantly higher acceptance to identical than to non-identical antecedents (p = .006, d = .76). Nonetheless, the effect size across all three groups was comparably small, suggesting no substantial differences.

In the third paired condition, the acceptance ratings of lexical verb ellipsis and rejection ratings of copula be ellipsis with non-identical antecedent were compared (see Table 4). As expected, there was a significant difference reflecting higher acceptance of lexical verb ellipsis by the NSs group (p < .001, d = 3.24), the very advanced group (p < .001, d = 2.99) and the advanced group (p < .001, d = 2.82).

These findings offer some evidence that the parallelism requirement is applied before morphological merger in the case of lexical verbs (unlike auxiliary verbs) and therefore morphological mismatch between the antecedent and the elided verbs is acceptable. Following up on this assumption, a series of repeated-measures ANOVAs was conducted for each group on their acceptability ratings of ellipsis of non-finite lexical verbs with three types of verb antecedents (simple finite, perfective participle and progressive participle) as the within-participant factors (see Table 5). The results showed a significant effect for antecedent type with a moderate effect size on the acceptability ratings of the NSs group, F (1.43,19.96) = 7.11, p = .009, $η_{ρ}^{2}$ = .34, the very advanced group, F (2,28) = 11.74, p < .001, $η_{ρ}^{2}$ = .46, and the advanced group, F (2,36) = 11.86, p < .001, $η_{ρ}^{2}$ = .40.

Table 5.

Mean ratings (SD) of non-finite lexical verb ellipsis after modal with three types of morphologically non-identical antecedents.

Groups/conditions	Simple finite (k = 4)	Perfective participle (k = 4)	^? Progressive participle (k = 4)
Native speakers (n = 15)	4.20 (.89)	4.17 (.93)	3.70 (.73)
Very advanced (n = 15)	4.42 (.51)	3.83 (.99)	3.48 (.99)
Advanced (n = 19)	4.16 (.64)	3.82 (.75)	3.45 (.87)

Note. ? indicates marginally unacceptable VPE.

In pairwise comparisons (LSD), the NSs gave comparable ratings on simple finite and perfective participle antecedents (p = .758, d = .08), but gave significantly lower ratings on progressive participle antecedents compared with the antecedents of simple finite (p = .003, d = .96) and perfective participle (p = .026, d = .67) respectively (see Table 5). Although the NSs’ ratings on the simple finite and perfective participle suggest that lexical identity is computed before morphological merger, their ratings on the participle forms (progressive and perfective) are problematic for this assumption. However, their ratings do conform with the contrasting behavior noted in the literature between the progressive aspect (-ing) and the perfective aspect (-en) in the antecedent conjunct (Lasnik, 1999).

On the other hand, the L2 groups’ ratings showed varied degrees of acceptability across verb antecedent types, with the simple finite antecedent receiving the highest ratings followed by the perfective and progressive participles, respectively (see Table 5). The simple finite was rated significantly higher than the perfective participle by the very advanced (p = .008, d = 1.00) and advanced groups (p = .041, d = .51). Also, the simple finite was rated higher than the progressive participle by the very advanced (p = .001, d = 1.34) and advanced groups (p < .001, d = 1.11). Comparing the progressive and perfective participles, the very advanced group showed comparable ratings on both participle antecedents (p = .073, d = .50) but the advanced group gave significantly lower ratings on the progressive participle compared with the perfective (p = .009, d = .69). Although the pattern of the L2 groups’ ratings was different from that of the NSs’, they showed that lexical verb ellipsis does not require strict morphological parallelism between the antecedent and the elided verb. Unlike the NSs, at advanced levels, they do not distinguish between participle forms. Regardless of the type of aspectual auxiliary, they seem to have a grammar which distinguishes simple forms from participle forms.

To summarize, the (very-)advanced L2 groups recognized modal-stranding as one manifestation of VPE. They also realized the contrast in morphological parallelism constraints between copula be and lexical verb ellipsis. Whilst they accepted ellipsis of copula be with only identical antecedents, they accepted ellipsis of lexical verbs with both identical and non-identical antecedents.

3 The recoverability principle: Auxiliary-stranding

The second analysis tested two types of stranded auxiliary (progressive be and perfective have) with (non)identical antecedents to the elliptical verb (see Table 6). Whilst it is required that the antecedent verb must match the elliptical progressive verb following a stranded be, it is not a requirement with perfective have stranding. Therefore, with non-identical antecedents, progressive be stranding is ungrammatical whereas perfective have stranding is grammatical.

Table 6.

Mean ratings (SD) of progressive be and perfective have stranding with morphological (non)identical verb antecedents.

Groups/conditions	Progressive be		Perfective have
Groups/conditions	Identical (k = 4)	*Non-identical (k = 4)	Identical (k = 4)	Non-identical (k = 4)
Native speakers (n = 15)	4.45 (.81)	1.38 (.45)	4.65 (.75)	3.83 (.67)
Very advanced (n = 15)	4.27 (.75)	1.37 (.79)	4.80 (.47)	3.22 (1.15)
Advanced (n = 19)	4.22 (.58)	1.46 (.63)	4.50 (.53)	2.45 (.83)

Note. * indicates ungrammatical VPE.

For each group, repeated-measures ANOVAs were conducted with stranded-auxiliary types (be and have) and morphological parallelism (non-identical and identical) as within-participant factors. The results revealed a significant main effect for type of stranded-auxiliary reflecting higher acceptance of auxiliary have by the NSs group, F (1,14) = 166.64, p < .001, $η_{ρ}^{2}$ = .92, the very advanced group, F (1,14) = 74.63, p < .001, $η_{ρ}^{2}$ = .84, and the advanced group, F (1,18) = 46.70, p < .001, $η_{ρ}^{2}$ = .72. The results also showed a significant main effect for parallelism with higher acceptance ratings of identical antecedents by the NSs group, F (1,14) = 195.08, p < .001, $η_{ρ}^{2}$ = .93, the very advanced group, F (1,14) = 117.47, p < .001, $η_{ρ}^{2}$ = .89, and the advanced group, F (1,18) = 139.51, p < .001, $η_{ρ}^{2}$ = .89. Most important, the results revealed a significant interaction effect between the stranded-auxiliary and parallelism on the ratings of the NSs group, F (1,14) = 98.90, p < .001, $η_{ρ}^{2}$ = .88, the very advanced group, F (1,14) = 16.23, p = .001, $η_{ρ}^{2}$ = .54, and the advanced group, F (1,18) = 6.80, p = .018, $η_{ρ}^{2}$ = .27. Also, the interaction effect size improved with increased proficiency. To uncover the interaction effect, Bonferroni pairwise comparisons were carried out with each group. Three paired conditions were examined: progressive be and perfective have stranding each with identical and non-identical antecedents, and progressive be and perfective have stranding with non-identical antecedents.

The results for the progressive be stranding showed a significant difference between the acceptance ratings on identical antecedents and the rejection ratings on non-identical antecedents in the judgments of the NSs group (p < .001, d = 3.69), the very advanced group (p < .001, d = 2.89) and the advanced group (p < .001, d = 2.77). The magnitude of the mean difference was comparably large for all groups. Conversely, the results for the perfective have stranding were not categorical as expected. Although both conditions should be grammatical, the NSs gave significantly lower acceptability ratings on non-identical antecedents than on identical ones (p < .001, d = 1.85). Comparable results were also shown by the very advanced (p < .001, d = 1.82) and advanced groups (p < .001, d = 1.83). Again, the magnitude of the mean difference was comparably large for all groups.

Comparing the judgments on non-identical antecedents by auxiliary type, the results showed that the NSs gave significantly lower ratings on progressive be stranding than on perfective have stranding (p < .001, d = 3.32). Likewise, the very advanced gave significantly lower ratings on progressive be stranding (p < .001, d = 1.92) and the advanced group did too (p < .001, d = 1.31). However, the magnitude of the mean difference for the NSs was larger than those of the (very-)advanced L2 groups. In fact, by displaying the ratings of all groups on boxplots, we noted different patterns of judgements on have vs. be stranding (see Figure 1). Although all the groups gave rejection ratings on be stranding, they produced different patterns of judgments on have stranding. To uncover the nature of their judgments given to have stranding, their mean ratings were compared with mid-point 3 (see Table 6). In one-sample t-tests, the NSs’ ratings were significantly higher than the mid-point rating, t (14) = 4.80, p < .001, suggesting acceptance ratings. Conversely, the ratings of the very advanced group were equal to the mid-point rating, t (14) = .73, p = .48, and the ratings of the advanced group were significantly lower than the mid-point rating, t (18) = 2.91, p = .009. Unlike target-like judgments, the very advanced group seemed uncertain in their judgments, whereas the advanced group gave rejection ratings. In fact, the very advanced group compared with the other two groups showed much individual variation as indicated by their large SD which stretched over both regions of acceptance and rejection (see Figure 1).

Figure 1.

Boxplots for mean ratings on VPE with non-identical antecedents: (a) perfective have stranding, and (b) progressive be stranding.

To further verify that the L2 groups’ judgments truly represent the contrast inherent between be and have stranding, their mean ratings with non-identical antecedents were divided by construction type (negative and positive) to uncover any potential effect of the presence of negation following the stranded auxiliary (see Table 7).

Table 7.

Mean ratings (SD) of progressive be and perfective have stranding with non-identical antecedents by construction type: negative vs. positive.

Groups/conditions	*Progressive be		Perfective have
Groups/conditions	Negative (k = 2)	Positive (k = 2)	Negative (k = 2)	Positive (k = 2)
Native speakers (n = 15)	1.27 (.46)	1.50 (.65)	3.93 (.70)	3.73 (.96)
Very advanced (n = 15)	1.30 (.68)	1.43 (.96)	3.67 (.90)	2.77 (1.59)
Advanced (n = 19)	1.47 (.72)	1.45 (.85)	2.92 (1.00)	1.97 (.98)

Note. * indicates ungrammatical VPE.

Repeated-measures ANOVAs were run for each group separately with construction type (negative and positive) and auxiliary type (be and have) as within-participant factors. The results confirmed our findings regarding the effect of auxiliary type by the NSs group, F (1,14) = 157.08, p < .001, $η_{ρ}^{2}$ = .92, the very advanced group, F (1,14) = 50.09, p < .001, $η_{ρ}^{2}$ = .78, and the advanced group, F (1,18) = 30.81, p < .001, $η_{ρ}^{2}$ = .63. The results also showed a significant interaction effect between auxiliary type and construction type on the ratings by the very advanced, F (1,14) = 10.82, p = .005, $η_{ρ}^{2}$ = .44, and advanced groups, F (1,18) = 9.63, p = .006, $η_{ρ}^{2}$ = .35, but not the NSs group, F (1,14) = 2.94, p = .109, $η_{ρ}^{2}$ = .17. Unlike target-like judgments, the (very-)advanced L2 groups’ ratings seemed to be affected by negation. The negative marker improved the acceptance ratings of perfective have stranding unlike progressive be stranding. As shown by the SD, individual variation in perfective stranding decreased with negation among the very advanced group, but increased among the advanced group (see Table 7). Bonferroni pairwise comparisons confirmed that the ratings were significantly higher for perfective have stranding with negation (e.g. John plays football, but Tom hasn’t __ recently) than in positive construction (e.g. Neal cooked dinner, and Sarah has __too) by the very advanced (p = .011, d = .88) and the advanced groups (p = .001, d = .87). Conversely, the same effect did not emerge with progressive be stranding, where results showed comparable ratings across negative and positive constructions (e.g. *Heather visits her mother, but Mary isn’t__ yet vs. *Julie sleeps late, and Debbie is__ too) by the very advanced (p = .334, d = .30) and the advanced groups (p = .905, d = .03).

To summarize, the results from the different analyses suggested that although the (very-)advanced L2 groups significantly distinguished between identical and non-identical conditions irrespective of auxiliary type, their ratings were non-target-like on perfective have stranding compared with progressive be stranding. The L2 groups of (very-)advanced proficiency seemed to be affected by negation in their ratings of perfective stranding, which made them appear target-like.

VIII Discussion and future research direction

Two subtle contrasts of English VPE were examined at grammar-internal interfaces. The first contrast was between copula be and lexical verb ellipsis (e.g. *John is here, and Mary will too vs. John slept, and Mary will too), which is constrained by the lexical identity condition at the syntax–lexicon interface. The second contrast was between be and have stranding (e.g. *John slept, and Mary was too vs. Peter saw your parents last week, but he hasn’t since), which is constrained by the recoverability principle at the syntax–semantics interface and involves distinguishing (un)interpretable features on -ing and -en. These contrasts were tested with English NSs and L2 speakers of L1 Saudi Arabic whose L1 does not allow VPE. A bimodal timed AJT was used to elicit judgment data under controlled time response to reduce the chances of them accessing explicit knowledge as much as possible. As predicted, performance on the syntax–semantics interface revealed non-target-like judgments compared with the syntax–lexicon interface. The diverging performance at the syntax–semantics interface was associated with uninterpretable features. Of all the tested conditions, the results uncovered non-target-like judgments on perfective have stranding which required recognizing the perfective uninterpretable feature on the stranded affix -en. The results will be first discussed in terms of the licensors to answer RQ1. Then, the L2 judgments will be discussed answering RQ2 in terms of adherence to the lexical identity condition at the syntax–lexicon interface, followed by the recoverability principle as implicated at the syntax–semantics interface.

To start with the licensors, although Saudi Arabic does not license VPE, the results showed that (very-)advanced L2 speakers successfully identified auxiliaries and modals as VPE syntactic licensors (see Table 3). In target-like judgments, they gave lower ratings to lexical *verb-stranding (e.g. *Neal grows vegetables, and his wife grows __ too)⁷ compared with modal-stranding (e.g. John will be here, and Mary will __ too) and auxiliary-stranding (e.g. John is watching TV, but Mary isn’t __ yet). They also showed comparable high acceptance ratings of both modals and auxiliaries as VPE licensors. Since their ratings were in favor of stranded auxiliaries and modals, it could be suggested that they have recognized that the VPE licensors must be hosted by Tense. Their ratings from auxiliary-stranding and *verb-stranding also suggest that they have recognized the verb-movement constraints implicated in the licensing of VPE. As an effect of resetting the strength of verbal uninterpretable features, verb-movement to Tense is restricted to auxiliaries. To confirm that Tense in their L2 grammars is the host of English VPE licensors, further research is needed with a wider range of licensors, including do support and infinitival to. However, two remarks are in order: first, the ratings of the L2 groups and the NSs on lexical *verb-stranding seemed higher than expected for an ungrammatical construction. As discussed earlier, this is due to the possibility that deleted material following the stranded verb can be interpreted as null arbitrary objects (Hawkins, 2012). In this case, verb-stranding is structurally ambiguous in English. Nonetheless, what matters is how *verb-stranding was rated compared with auxiliary-stranding and modal-stranding. Second, it could be argued that their acceptance ratings of modal-stranding and auxiliary-stranding reflect a simple deletion strategy on surface structure because ellipsis was tested under strict parallelism with the antecedent. Therefore, robust evidence is required from much more complex structures to verify that their judgments reflect an underlying representation which syntactically licenses VPE. This evidence is discussed next and it bears on their sensitivity to the lexical identity condition and recoverability principle (Rouveret, 2012), where morphological parallelism between the antecedent and the elliptical verb is manipulated in terms of identical and non-identical conditions.

According to the lexical identity condition, ellipsis of copula be, unlike lexical verbs, requires strict parallelism with the verb antecedent. The NSs’ judgments were consistent with this condition (see Table 4). Although they rejected ellipsis of non-finite copula be preceded by a finite non-identical antecedent (e.g. *John is happy, and Mary will ___ soon), they accepted ellipsis of non-finite lexical verbs in the same condition (e.g. Mike revises his work, and Jennifer should ___as well). Unlike copula be, morphological parallelism is irrelevant to lexical verbs because the identity of the antecedent and the elliptical verb is assessed at the root level before morphological merge (i.e. revise+-s). As discussed earlier, this contrast bears on differences in the lexical representation of verb type in the lexicon. Whilst copula be enters the derivation already inflected from the lexicon, lexical verbs are inflected in the syntax (Lasnik, 1999). Inflected and non-inflected forms of copula be are thus assumed to be stored in separate entries in the lexicon. Acquiring this contrast requires integrating verb type information at the syntax–lexicon interface. In target-like judgments, the (very-)advanced L2 groups successfully distinguished between ungrammatical copula be ellipsis and grammatical lexical verb ellipsis preceded by finite verb antecedents (see Table 4). Further supporting evidence arises when considering their judgments on each verb type in paired identical and non-identical conditions (see Table 4). In terms of copula be, the (very-)advanced L2 groups were target-like in their distinction. They rejected non-finite copula be ellipsis with finite non-identical antecedents (e.g. *John is happy, and Mary will ___ soon), but accepted it with non-finite identical antecedents (e.g. John will be here, and Mary will __ too). Their judgments show that they recognized that copula be requires the verb antecedent and the elliptical verb to be lexically/morphologically identical. In terms of lexical verbs, the very advanced L2 group gave comparable acceptance ratings to the ellipsis of non-finite lexical verbs with identical and non-identical verb antecedents (e.g. Bruce can design a webpage, and George can ____too vs. Mike revises his work, and Jennifer should ___as well). Their performance shows that they realized that lexical identity is assessed at the root level and thus morphological parallelism is irrelevant with lexical verbs. In contrast, the NSs showed statistically lower acceptance ratings of non-identical conditions compared with identical ones. There is evidence from psycholinguistic experiments that non-parallel structures are usually less acceptable than parallel structures (Duffield et al., 2009). This native speaker – non-native speaker (NS–NNS) different behavior is interesting and requires further investigation.

The results from copula be and lexical verb ellipsis in identical and non-identical conditions lend converging evidence that very advanced L2 speakers are sensitive to the lexical identity condition assumed to be captured at the syntax–lexicon interface. The results also indicate that they recognized modals as legitimate syntactic licensors for VPE. Had they been applying a simple deletion strategy at the surface, they would not be able to accept lexical verb ellipsis in non-identical conditions. These results are in line with earlier findings reached by Al-Thubaiti (2009, 2010), where L2 English college students of various proficiency levels were tested. Although the college students showed vague knowledge of constraints on VPE at the early stages of development, they gradually started to realize, with some individual variation, that ellipsis of copula be (unlike lexical verbs) requires strict identity with the antecedent verb. In fact, college students of advanced English proficiency still performed less well than the very advanced L2 speakers in the present study. Eventually, the contrast between copula be and lexical verbs seems acquirable, as was also found by Hawkins (2012) when testing L1 Arabic and Chinese English L2 speakers.

The identity condition as it applies to lexical verbs offers some evidence for deletion before morphological merger. This assumption was further tested by comparing judgments on non-finite lexical verb ellipsis with three types of non-identical antecedents (simple finite, progressive participle and perfective participle) (see Table 5). Although the NSs gave comparable acceptance ratings to ellipsis with the antecedents of simple finite and perfective participle (e.g. Mike revises his work, and Jennifer should __ as well vs. John has baked a cake, and Mary will ___ too), they showed lower ratings with the progressive participle (e.g. ^?Bill is writing an essay, and Tom should ___ as well). Whilst the NSs’ ratings are indicative of deletion before morphological merger, the lower ratings on the progressive participle are problematic for this assumption. In fact, these lower ratings were expected given the relative contrast in acceptability between -ing and -en when they occur in the antecedent clause (Lasnik, 1999). On the other hand, the L2 groups gave a slightly different pattern of judgments from those of the NSs. They rated both perfective and progressive participles as each being less acceptable than simple finite antecedents. In terms of the participle forms, with increased proficiency, the difference in the degree of acceptance ratings between both forms disappeared. Whilst the very advanced group showed comparable ratings on the progressive and perfective participles, the advanced group showed lower ratings on progressive participles than on perfective participles. At very advanced levels, the L2 group seemed to have developed a grammar which distinguishes the simple form, on the one hand, from the participle forms with auxiliary, on the other. Unlike the NSs, their ratings do not reflect the relative contrast in acceptability between the progressive and perfective participles (Lasnik, 1999). However, they recognize that strict morphological parallelism with the antecedent is not a requirement for lexical verb ellipsis. In a processing approach to VPE, reaction time data⁸ could be quite informative in explaining the underlying contrast among different verbal constructions. The ‘gradience effect’ in acceptability is an open question which requires interdisciplinary work that combines processing and theoretical approaches.

Moving now to the recoverability principle, perfective have and progressive be stranding illustrate a subtle contrast in ellipsis between the stranded affixes -en and -ing. This contrast bears on feature interpretability and is captured at the syntax–semantics interface. The -ing cannot be elided unless its meaning is recoverable because it carries an interpretable progressive feature, whereas the -en can be elided regardless of recoverability because it carries an uninterpretable perfective feature (Rouveret, 2012). On this contrast, the NSs performed less categorically than expected (see Table 6). Although they gave acceptance ratings to perfective verb ellipsis following a stranded have in both identical and non-identical conditions, their ratings were significantly lower on finite non-identical antecedents (e.g. John plays football, but Tom hasn’t ___ recently) compared with identical perfective antecedents (e.g. Claire has bought a new house, but Mary hasn’t ____ yet). Their acceptance ratings were found to be statistically higher than mid-point ratings. Conversely, they gave categorical judgments on progressive verb ellipsis following a stranded be. Their acceptance ratings were significantly higher on identical progressive antecedents (e.g. John is watching TV, but Mary isn’t __yet) compared with their rejection ratings on finite non-identical antecedents (e.g. *Heather visits her mother, but Mary isn’t __yet). Most important, the NSs were categorical in their distinction between both auxiliary types when preceded by finite non-identical antecedents. Their judgments are taken to indicate that the two structures have a different underlying representation which bears on the presumed feature interpretability. Having seen how the NSs performed on this subtle contrast, we need to understand the L2 speakers’ performance (see Table 6). The (very-)advanced L2 speakers gave comparable judgment patterns to those of the NSs on identical and non-identical conditions for each auxiliary type. They were also categorical in rejecting progressive verb ellipsis and accepting perfective verb ellipsis when both were preceded by finite non-identical antecedents. Thus far, it could be argued that the L2 speakers of very advanced proficiency managed to acquire the underlying contrast.

Nonetheless, the L2 speakers’ judgments on perfective verb ellipsis with finite non-identical antecedents showed considerable individual variation (see Figure 1). In deeper analysis of their judgments, an effect of negation was uncovered (see Table 7). By breaking down their judgments by construction type (negative vs. positive), the presence of negation clearly improved their preparedness to accept perfective verb ellipsis with non-identical antecedents but had no effect on progressive verb ellipsis in the same condition. This is illustrated by the fact that the L2 speakers gave significantly higher ratings to (e.g. John plays football, but Tom hasn’t __ recently) than to (e.g. Neal cooked dinner, and Sarah has __too).⁹ This same pattern of judgments did not emerge with the NSs. There is no theoretical account which would associate negation with perfective have stranding, and it was not even predicted to have an effect. The question now is how to explain the effect of negation. Following Tsimpli and Dimitrakopoulou (2007), it could be argued that L2 speakers of very advanced proficiency can approximate target-like grammars by resorting to interpretable features.¹⁰ In this case, they resorted to negation. There is some evidence that negation can act as a licensor for VPE (Aelbrecht, 2010; Johnson, 2001; Lobeck, 1995). To illustrate with an example from Aelbrecht (2010: 168), the presence of ‘not’ rescued the ungrammatical derivation containing a contracted auxiliary and made it grammatical (compare *He says he’s not working, but he’s vs. He says he’s working, but he’s not; bold by the author). However, it could be argued that the ungrammaticality of the stranded ‘he’s’ is independent of VPE.¹¹ This article does not argue for the role of negation as a licensor but presents it as an interesting observation which deserves further investigation. It is not clear how the L2 speakers came to make this association and whether the presence of negation as a full form vs. clitic made a difference. Since this result emerged from breakdown analyses, and was therefore based on only a few items, further research is indeed required to verify the observed effect of negation as an independent factor. Overall, the present findings indicate that the contrast between stranded -ing and -en in the elliptical site can be overcome at very advanced levels but by different means. Therefore, these results are different from those reached by Al-Thubaiti (2009, 2010) and by Hawkins (2012), who tested less-proficient L2 speakers. Further research is required to follow up on the potential effect of negation.

Taking stock, what do these findings tell us about the domain of vulnerability in L2A? To what extent do predictions of the Interpretability Hypothesis (Tsimpli and Dimitrakopoulou, 2007) help us understand ‘selective vulnerability’ at grammar-internal interfaces? We have seen how L2 performance on the syntax–lexicon and syntax–semantics contrasts were not equally problematic. Whilst the contrast between copula be and lexical verb ellipsis at the syntax–lexicon interface was unproblematic, the contrast between perfective have and progressive be stranding at the syntax–semantics interface turned out to be problematic. A consideration of the role of formal features as in the Interpretability Hypothesis can explain why different grammar-internal interfaces are not equally problematic. As predicted, highly-proficient L2 speakers had problems with the contrast which involved distinguishing (un)interpretable features at the syntax–semantics interface unlike the contrast at the syntax–lexicon interface. They showed problems recognizing uninterpretable features on the perfective form, -en, but no problems in recognizing interpretable features on the progressive form, -ing. Such contrasting performance is taken to indicate a case of ‘selective vulnerability’ at grammar-internal interfaces, where vulnerability arises with uninterpretable features. Nonetheless, L2 speakers of very advanced proficiency can overcome this difficulty in acquiring uninterpretable features with the help of interpretable features. Unlike target-like judgments, they seem to have developed a grammar in which the presence of negation correlates with the acceptance ratings of the perfective form. This effect was subtle when computing the group mean, and thus made the L2 groups appear target-like in the first set of paired comparisons. Only through detailed analyses was this subtle effect revealed in their L2 judgments. If this analysis holds true, it should be replicated in future research. Since this article is the first to test such subtle contrasts, a strict experimental research design is required to follow up on the current findings.

In conclusion, this current study has shown that vulnerability at the interface domain cannot be understood without considering the role of formal features. Much more research is required to validate the findings on VPE especially because it is under researched.

Footnotes

Appendix 1

VPE stimuli and distracters.

Acknowledgements

I would like to thank first and foremost three anonymous Second Language Research reviewers for their valuable and meticulous feedback that helped in improving the article in many ways. I would like also to thank Roger Hawkins for his advice and comments on an earlier draft of the article. I am so grateful to Rex Sprouse for discussing the topic of VPE with me, and to Mahasen Abu-Mansour for her advice on the transliteration of the Saudi Arabic examples. I am so thankful as well to Heather Marsden for encouraging me to submit to Second Language Research.

Declaration of Conflicting Interest

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Notes

ORCID iD

Kholoud A Al-Thubaiti

References

Abdulkarim

Roeper

(1997) Economy of representation: Ellipsis and NP reconstructions. In: Shillcock

(ed) Language acquisition: Knowledge representation and processing. Edinburgh: Human Communication Research Center.

Aelbrecht

(2010) The syntactic licensing of ellipsis. Amsterdam: Benjamins.

Al-Thubaiti

(2009) The acquisition of uninterpretable features: The case of vP ellipsis. In: Bowles

Ionin

Montrul

et al . (eds) Proceedings of the 10th Generative Approaches to Second Language Acquisition Conference (GASLA 2009) (pp. 192–200). University of Illinois at Urbana-Champaign. Somerville, MA: Cascadilla Press.

Al-Thubaiti

(2010) Age effects in a minimal input setting on the acquisition of English morpho-syntactic and semantic properties by L1 speakers of Arabic. Unpublished PhD dissertation, University of Essex, UK.

Al-Thubaiti

(2015) L2 acquisition of English aspect by L1 Arabic speakers. In: Ayoun

(ed) The Acquisition of the Present (pp. 185–213). Amsterdam: Benjamins.

Arregui

Clifton

Frazier

et al . (2006) Processing elided verb phrases with flawed antecedents: The recycling hypothesis. Journal of Memory and Language 55: 232–46.

Bley-Vroman

Felix

Ioup

(1988) The accessibility of Universal Grammar in adult language learning. Second Language Research 4: 1–32.

Cowart

(1997) Experimental syntax. London: SAGE.

Cyrino

Matos

(2005) Local licensers and recovering in VP ellipsis. Journal of Portuguese Linguistics 4: 79–112.

10.

Duffield

Matsuo

(2009) Native speakers’ versus L2 learners’ sensitivity to parallelism in VP-Ellipsis. Studies in Second Language Acquisition 31: 93–123.

11.

Duffield

Matsuo

Roberts

(2009) Factoring out the parallelism effect in VP-ellipsis: English vs. Dutch contrasts. Second Language Research 25: 427–67.

12.

Hagstrom

(1994) Issues of affix hopping in an attract-F framework. Unpublished manuscript. Cambridge, MA: MIT.

13.

Halle

Marantz

(1993) Distributed morphology and the pieces of inflection. In: Hale

Keyser

(eds) The view from Building 20: Essays in Linguistics in honour of Sylvain Bromberger (pp. 111–76). Cambridge, MA: MIT Press.

14.

Halle

Marantz

(1994) Some key features of distributed morphology. In: Carnie

Harley

(eds) MIT working papers in linguistics 21: Papers on Phonology and Morphology (pp. 275–88). Cambridge, MA: MIT Press.

15.

Hawkins

(2012) Knowledge of English verb phrase ellipsis by speakers of Arabic and Chinese. Linguistic Approaches to Bilingualism 2: 404–38.

16.

Ionin

(2012) Formal theory-based methodologies. In: Mackey

Gass

(eds) Research methods in second language acquisition: A practical guide (pp. 30–52). Malden, MA: Wiley-Blackwell.

17.

Johnson

(2001) What VP-ellipsis can do, and what it can’t, but not why. In: Baltin

Collins

(eds) The handbook of contemporary syntactic theory (pp. 439–79). Oxford: Blackwell.

18.

Lasnik

(1999) Minimalist analysis. Malden, MA: Blackwell.

19.

Lobeck

(1995) Ellipsis: Functional heads, licensing, and identification. Oxford: Oxford University Press.

20.

Murphy

(1997) The effect of modality on a grammaticality judgement task. Second Language Research 13: 34–65.

21.

Nation

(2001) Learning vocabulary in another language. Cambridge: Cambridge University Press.

22.

Rouveret

(2012) VP ellipsis, phases and the syntax of morphology. Natural Language and Linguistic Theory 30: 897–963.

23.

Sorace

(1996) The use of acceptability judgments in second language acquisition research. In: Ritchie

Bhatia

(eds) Handbook of second language acquisition (pp. 375–409). New York: Academic Press.

24.

Sorace

(2011) Pinning down the concept of ‘interface’ in bilingualism. Linguistic Approaches to Bilingualism 1: 1–33.

25.

Sorace

Filiaci

(2006) Anaphora resolution in near-native speakers of Italian. Second Language Research 22: 339–68.

26.

Tsimpli

I-M

Dimitrakopoulou

(2007) The interpretability hypothesis: Evidence from wh-interrogatives in second language acquisition. Second Language Research 23: 215–42.

27.

Tsimpli

I-M

Sorace

(2006) Differentiating interfaces: L2 performance in syntax–semantics and syntax–discourse phenomena. In: Bamman

Magnitskaia

Zaller

(eds) Proceedings of the 30th Annual Boston University Conference on Language Development (BU) (pp. 653–64). Somerville, MA: Cascadilla Press.