1 Introduction

Sluicing, first described by Ross (1969), is a type of clausal ellipsis where a wh-phrase by itself is understood to correspond to a full interrogative clause. Sluicing is attested both in matrix clauses, as in B’s response in (1), and in embedded clauses, as in (2). The focus of this paper will be on constructions like that in (2), which constitute 72% of all natural instances of sluicing (Anand, Hardt & McCloskey 2021).

Conventionally, the first clause (i.e., “John was talking to someone on the phone”) is called the antecedent clause, and the indefinite (i.e., “someone”) in the antecedent clause is referred as the correlate. The wh-phrase in the second clause (i.e., “who”) is called the wh-remnant. The correlate and the remnant in (1) and (2) correspond to the same person in the discourse whose identity is unknown. The wh-remnant is interpreted as a full clause, as Who was John talking to on the phone? in (1), and as who John was talking to on the phone in (2).

There are different kinds of sluicing that are identified based on the properties of the correlate. We will refer to sluicing constructions where the correlate is an indefinite, as in (1) and (2), as regular sluicing. The indefinite correlate in regular sluicing can be a pronominal element (1–2), or it can be a full Noun Phrase (NP) as in (3).

The correlate can also be a definite expression (Chung et al. 1995; Merchant 2001; 2008). Sluicing constructions where a definite correlate bears prosodic prominence (e.g., it is contrastively focused) are referred to as contrast sluicing.

In contrast sluicing, the wh-remnant involves an extra linguistic element such as “other” or “else”, commonly referred to as an exceptive modifier (see Barros 2014b). The exceptive modifier ensures that the referent(s) of the remnant are the other candidate(s) in discourse that is/are distinct from the correlate.

Contrast sluicing is claimed not to be uniformly acceptable in every kind of linguistic environment (e.g., Merchant 2008; Barros et al. 2014; Griffiths & Lipták 2014). Contrast sluicing is observed to be unacceptable, for instance, when the correlate (“the cake” in 5a) resides in an if-clause (bracketed), as shown in (5a). This is unlike regular sluicing with a correlate in the same structural configuration in (5b) (e.g., Ross 1969; Merchant 2001).

Interestingly, these types of environments are the same environments as those which disallow wh dependency formation in constructions such as wh questions. Environments that block wh dependencies are known as islands. Islands have long been taken to be structural domains out of which displacement or movement of a phrase is not allowed (e.g., Chomsky 1973; Ross 1967). Displacement causes the sentence to be unacceptable, as in (6).

In comparison with (6), the acceptability pattern in (5a) versus (5b) appears to be puzzling. The unacceptability of (5a) makes sense, if the elided clause is isomorphic to the antecedent clause, since the wh-remnant must have undergone extraction out of an island (…what else Ben will be angry [if you don’t try]). But then we would expect (5b) to be unacceptable, as well. Griffiths & Lipták (2014) generalize the contrast between (5a) and (5b), and state that contrastive ellipsis like contrast sluicing is island sensitive, but non-contrastive ellipsis like regular sluicing is insensitive to islands. However, recent experimental work has challenged this, for instance, by finding contrastive stripping to be island insensitive (Yoshida et al. 2019), which we detail in Section 4. No previous work has experimentally investigated the acceptability pattern in (5a–b)—contrast sluicing versus regular sluicing—, a gap which we seek to fill in the present study.

Both regular sluicing and contrast sluicing have generated a lot of interest in the linguistic literature, because – unlike (6) – there are no pronounced strings of words after the wh-remnant in these constructions. In other words, meaning is somehow computed from silence. To account for how this interpretative process plays out, there are two strands of approaches: structural and non-structural approaches. Broadly speaking, non-structural approaches consider sluicing to involve sentential “fragments” with no covert syntax (e.g., Culicover & Jackendoff 1995; Ginzburg & Sag 2000; Barker 2013; Nykiel & Kim 2022). The interpretation of sluicing is achieved via featural properties of the syntactic node that dominates the wh-word (Culicover & Jackendoff 1995; Ginzburg & Sag 2000), or via an anaphoric relation between the antecedent clause and a null element in the ellipsis site (Chung et al. 1995; Barker 2013).

The other approach, the structural approach, maintains that there is covert syntax at the ellipsis site (e.g., Ross 1969; Lasnik 2001; Merchant 2001). For instance, there might be a structure matching the antecedent clause, as was illustrated in example (6) above. One popular way of modeling this, pioneered by Merchant (2001), is to give the head C⁰ an ellipsis feature. This feature requires non-pronunciation of C⁰’s complement (i.e., TP, as shown in gray ink in (7)).

In this approach, sluicing deletes a clausal constituent and leaves behind only the wh-phrase that escaped from the ellipsis site via movement (Ross 1969; Merchant 2001; and much other work).

As shown in (5a–b), regular and contrast sluicing seem to respond differently to islands: regular sluicing appears to be insensitive to islands, as in (5b), but contrast sluicing is sensitive to islands (as in 5a). To explain the insensitivity of regular sluicing to islands, some structural approaches have argued for an ellipsis repair mechanism where the full isomorphic structure is present but unpronounced, and this lack of pronunciation repairs the island violation (Ross 1969; Merchant 2001; 2008). Others have proposed an evasion strategy (Barros et al. 2014), where only a simpler structure, called a “short source,” is present, and movement in this simple elided “short source” structure does not involve movement out of an island. Importantly, both repair and evasion approaches predict contrast sluicing to be island sensitive but differ in their explanation for why (we lay out the full details of these two approaches in Section 2).

With this backdrop in mind, the primary goal of this paper is to evaluate the predictions of the non-structural, repair, and evasion approaches using novel experimental data. We do so by examining how regular sluicing and contrast sluicing behave when their antecedent is embedded either inside a subject NP or an object NP, like with “a monster” in (8).

Extraction out of object NPs as in (9b) is generally considered to be grammatical, but extraction out of subject NPs like “stories about” in (9a) is not. (9a) constitutes a type of island known as a subject island (Ross 1967).

Subject islands have been a topic of widespread debate in linguistics and cognitive psychology, independently of sluicing. This debate centers around the question of whether subject island violations arise due to formal constraints in the syntax proper (e.g., Chomsky 1973; Huang 1982; Stepanov 2007), or instead, whether the syntax itself allows the extraction, but the resulting output is judged to be ill-formed because it creates a meaning which is anomalous in most kinds of discourse contexts (e.g., Erteschik-Shir 1973; Goldberg 2006; Chaves & Dery 2019; Abeillé et al. 2020). In light of this ongoing debate, a secondary goal of our current study is to lay out what kinds of data from sluicing the two competing approaches must take into account in order to be successful.

The paper proceeds in the following way. First, Section 2 presents more background on how the repair and evasion approaches explain island (in)sensitivity in clausal ellipsis (with a focus on subject islands). Section 3 presents Experiment 1, which examines what the nature of the unpronounced structure in sluicing is. Our results suggest that the ellipsis source differs in regular and contrast sluicing: contrast sluicing does not allow a short source, while regular sluicing does. Considering this strict syntactic isomorphism in contrast sluicing but not in regular sluicing, we next test subject island effects in these constructions. Experiment 2, presented in Section 4, quantitatively evaluates how regular and contrast sluicing respond to subject islands in English. Contrary to predictions (e.g., Merchant 2008), we find no island effect in contrast sluicing, just like in regular sluicing. In view of this finding (i.e., insensitivity of contrast sluicing to islands), we further examine subject islands in Experiment 3 as presented in Section 5. Experiment 3 explores subject island effects in embedded wh-questions, which are taken to be the pronounced counterparts of sluicing. The findings verify that the subject island effect is active in embedded wh-questions. Since contrast sluicing does not have a short source available to it (Experiment 1), this means that ellipsis must be repairing the island effect. The implications of these findings will be discussed further in Section 6.

2 Background: Islandhood under sluicing and how it has been accounted for

If one assumes that sluicing is a form of regular wh-movement, differing only in that the rest of the clause after the wh-phrase is unpronounced, then sluicing is expected to be constrained by the same set of rules that regulate wh-movement. Hence, an island effect is expected to be found in sluicing when the antecedent clause hosts a correlate inside an island (i.e., the sentence should be unacceptable). Starting with Ross (1969), sluicing has been extensively explored in relation to various kinds of islands, like the complex NP constraint and the adjunct island constraint as exemplified in (10) (island domains are bracketed).

Interestingly, although the structural approach expects island effects in sluicing, they have in fact been observed to be absent (e.g., Ross 1969; Merchant 2001). An example involving a complex NP object island is shown with a regular sluicing sentence in (11). In the structural approach, the wh-remnant is taken to have moved out of the exact same domain as in (10a) above (shown in gray in (11)). However, while (10a) is unacceptable, (11) is perfectly acceptable.

Interestingly, this insensitivity to islands is claimed not to hold of contrast sluicing (e.g., Merchant 2008; Barros et al. 2014; Griffiths & Lipták 2014). Contrast sluicing is judged to be unacceptable when the correlate (e.g., “Greek” in (12)) is embedded in a complex NP object.

Two accounts have been proposed so far to explain the observed pattern of island (in)sensitivity under sluicing (11 vs 12). These accounts differ in their view of syntactic parallelism (i.e., structural matching) under sluicing. The repair approach (Chomsky 1972; Lasnik 2001; Merchant 2001; 2008) argues that the elided clause involves a structure isomorphic to the antecedent clause in both regular and contrast sluicing. This means the elided clause involves an island when the antecedent clause has one, as was shown in gray in (11) and (12). According to the repair approach, island violations occur due to PF (Phonological Form) constraints (e.g., Merchant 2001). The ellipsis operation is proposed to repair island violations via non-pronunciation (i.e., deletion at PF). Repair succeeds and sluicing becomes insensitive to the island effect in the case of regular sluicing (11). When the ellipsis is of the contrastive type, such as contrast sluicing (12), however, the sentence is unacceptable due to constraints on movement: the correlate is not allowed to move to the highest clausal node at LF and therefore fails to license the ellipsis (Merchant 2008); or, the correlate is banned from being extracted from an island at Logical Form (LF) (Griffiths & Lipták 2014).

The evasion approach adopts a more flexible view of the structural matching that is required to hold between the elided clause and the antecedent clause (Vicente 2008; Barros 2014a; 2014b; Barros et al. 2014; Griffiths 2019; Wu 2022). Syntactic mismatch is permitted because ellipsis can take place when a semantico-pragmatic identity condition is met according to this strand of approaches. Importantly for our purposes, these identity conditions are loose enough to allow structural non-parallelism between the antecedent clause and the elided clause. Accordingly, the elided clause can fail to match the syntax of the antecedent clause in regular sluicing; for instance, it can contain a copular clause, as shown in (13).

The evasion approach, hence, naturally explains the acceptability of regular sluicing in relation to islands. There is no island violation in (13): note that pronouncing the material in gray in (13) is acceptable, in contrast with pronouncing the gray material in (11).

As for contrast sluicing, the evasion approach argues that the elided clause has to be syntactically isomorphic to its antecedent clause in contrast sluicing, otherwise, infelicity results due to independent focus-background matching constraints (Barros 2014a; Weir 2014; Griffiths 2019). This required matching forces the elided clause to contain an island if the antecedent clause has one, which then leads to unacceptability when the wh-remnant is extracted from it, as in (14).

Non-structural approaches to ellipsis predict no island effects whatsoever (Culicover & Jackendoff 1995; Ginzburg & Sag 2000; Barker 2013; Nykiel & Kim 2022), because there is simply no elided structure present at all – instead, the intended meaning of the ellipsis is reconstructed through featural properties of the syntactic node that dominates the wh-word (Culicover & Jackendoff 1995; Ginzburg & Sag 2000), or via an anaphoric relation between the antecedent clause and a null element (i.e., e) in the ellipsis site (Chung et al. 1995; Barker 2013). These approaches face a challenge from informal judgements like those in (12) and (14), where it appears that island effects do emerge in contrast sluicing. Thus, it is important to establish precisely what the empirical landscape of islandhood under regular and contrast sluicing looks like through fully controlled acceptability judgements.

To set the scene for investigating subject islands, we first provide new insights into what the ellipsis source is in regular and contrast sluicing by taking one step back from islandhood. In our first experiment (Section 3), we use a sentence continuation study to examine what kind of material is elided from the position after the wh-remnant in regular and contrast sluicing. We then use the results from this experiment to specify what predictions each approach (non-structural, repair, and evasion) makes regarding the effect subject islands should have on regular sluicing and contrast sluicing.

3 Experiment 1: Investigating the source under sluicing

The purpose of this experiment is to investigate what the structure of the unpronounced material may be in regular sluicing and in contrast sluicing. Formal research takes the (felicity of) potential non-elliptical continuations into consideration when deciding what the elided material must be (Merchant 2001; Barros 2014a; 2014b; Barros et al. 2014; among others). No prior study has experimentally investigated such heuristic approaches that rely on native speaker intuitions about “what the sluice would sound like had it not been elided” (Barros 2014b: 4). According to Barros (2014b: 4):

… the heuristic method of checking speaker intuitions about plausible continuations has proven to be useful in many domains in the investigation of ellipsis. It is, in fact, this intuition that motivates the isomorphism hypothesis to begin with, as simple examples of sluicing … are intuitively synonymous with a non-elliptical paraphrase that is the Wh-question version of the antecedent. Speaker intuitions about the meaning of the sluice, in particular, are an unquestionably valid source of empirical evidence in constraining the hypothesis space for theories of the E-site’s content.

In the same spirit, we adopt the view that a pronounced felicitous continuation will mirror the structure that undergoes deletion under sluicing. Hence, we ask experimental participants to pick among possible continuations in order to tap into intuitions about what the elided material is likely to be.

We conducted a three-alternative forced-choice sentence completion experiment where participants had to read a sentence and choose the best continuation from three options. In a 2 × 2 (within-subjects, within items) design, we crossed two factors. The first was the sluice type: the stimuli involved either regular sluicing sentences (e.g., a(n) x in the first conjunct, which x in the second conjunct), or contrast sluicing sentences (e.g., the x in the first conjunct, which other x in the second conjunct). The second manipulation was the grammatical position of the NP correlate, as subject or direct object (NP_subj; NP_obj). The correlate was not embedded in any larger structure (like an island). This allowed us to assess the availability of each continuation independently of the presence of an island. The grammatical role manipulation was included in case there was an effect which might be relevant in testing subject islands in Experiment 2. This 2 × 2 paradigm is shown in Table 1 (top part).

After each sentence in the experiment, three options were presented as a potential continuation as shown in Table 1, bottom part (note that these options do not have island violations, because the correlate NP is not embedded in any larger NP structure). Of the three continuation options, one involved a copular clause in the form of a pronoun “that”, followed by a copular verb (i.e., “that is”, “that was”). In evasion approaches, copular clauses are considered to be one of the alternative non-isomorphic structures that can appear in the ellipsis site under regular sluicing. They are thought not to be available in contrast sluicing, because of the independent focus-background constraint on contrastive focus (see section 2). Indeed, the two native speaker authors of this study judged the copular continuation to be unacceptable in the contrast sluicing condition, in line with what has been reported in the literature (Merchant 2001; Vincente 2008; Barros 2014a; 2014b; Barros et al. 2014; Griffiths 2019; a.o.). Our goal was to verify this experimentally. In the present study, “that” was used as the subject in copular clauses rather than “it” to prevent ambiguity; “it” in this environment could also be the subject of a reduced cleft (Barros et al. 2014) rather than a simple copular clause. Using “it” rather than “that” could have introduced undesirable complexity to the stimuli; hence, the non-isomorphic copular continuation option consistently involved “that”. We highlight that “that” is interpreted as an e-type pronoun in this environment (Merchant 2001; Barros et al. 2014): its interpretation is anaphoric to the meaning of the antecedent clause, which basically corresponds to the meaning of a relative clause headed by a definite NP derived from the correlate (e.g., the superhero that amazed Mike).

The second continuation option involved a syntactically isomorphic clause that matched the structure in the antecedent clause. Because antecedent clauses were created as simple English sentences with a transitive verb and its core arguments, the isomorphic continuation options had the same transitive verb as in the antecedent clause in addition to that verb’s internal or external NP argument, both repeated as they appeared in the antecedent clause. The third option was a grammatically ill-formed continuation (which served as an attention-check). This experimental paradigm is shown in Table 1.

3.3 Results

The results are shown in Figure 1. Note that, since the dependent variable of the present study is binary, the remaining proportion of choices – the reverse of the pattern in Figure 1 – corresponds to the copular continuation responses.

Figure 1

Proportion of choice of isomorphic continuation across contrast sluicing and regular sluicing. The dashed line represents chance level (i.e., 50%).

Sluice type was a significant predictor of participants’ source choice: the isomorphic continuation was chosen significantly more in contrast sluicing conditions compared with regular sluicing (0.92 vs 0.2; β = –6.34, SE = 0.87, z = –7.29, p < .0001; BF₁₀ = 9 × 10⁷). There was also a significant effect of position in that the isomorphic continuation was chosen significantly more with object correlates compared with subject correlates (0.58 vs 0.53; β = –0.83, SE = 0.32, z = –2.56, p = .01; BF₁₀ = 27). The interaction of sluice type and position was marginal (β = –1.17, SE = 0.65, z = –1.81, p = .07), but the Bayes Factor analysis revealed that we have at best only weak evidence to believe that this marginal interaction corresponds to a meaningful effect (BF₁₀ = 1.82). Thus, we leave it aside for present purposes.

Lastly, to evaluate whether the proportion of choice of isomorphic source was significantly different from chance level (i.e., 50%, see the dashed line in Figure 1) for each of the two sluicing levels, we conducted a further analysis by recoding the predictors (following West et al. 1996) and fitted a separate model for each level, focusing on the model intercept: because 0 corresponds to 0.5 (or 50%) in logit space, a significant positive intercept means that the response is statistically above chance, and a significant negative intercept means that it is below chance. Here, we found that the isomorphic continuation option was picked above chance level in contrast sluicing conditions (0.92; β = 4.18, SE = 0.80, z = 5.17, p < .001), but below chance level in regular sluicing. The pattern was in the opposite direction as indicated by the negative values of the intercept of the model (0.2; β = –2.18, SE = 0.42, z = –5.2, p < .001). In sum, a short (copula) continuation was preferred for sentences with regular sluicing, and an isomorphic continuation was preferred for sentences with contrast sluicing. We also found a higher rate of isomorphic continuation choice when the antecedent was in object position compared with subject position.

4 Experiment 2: Subject island (in)sensitivity in sluicing

Recall that regular sluicing and contrast sluicing have been reported to interact with islands differently. Regular sluicing with an indefinite correlate is characterized as being insensitive to islands; however, the island effect is argued to be active in contrast sluicing where the correlate and the remnant are contrasted (Merchant 2008). This pattern has been claimed to be a property of clausal ellipsis in general: fragment answers and stripping have also been claimed to be sensitive to islands (Merchant 2005; 2008; Griffiths & Lipták 2014). The common factor for island sensitivity is argued to be “contrastiveness” (see Griffiths & Lipták 2014). However, some contrastive fragment answers involving islands have been informally reported to be acceptable, as shown in (16) (see Barros 2014b; Weir 2014; contra Griffiths & Lipták 2014; this example involves a relative clause island).

Moreover, some experimental studies have found mixed results regarding the island sensitivity of stripping, a type of clausal ellipsis which is commonly analyzed as akin to sluicing and fragment answers (Potter 2017; Yoshida et al. 2019; Wu 2022). Potter (2017) suggests that contrastive stripping, as illustrated in (17a), is “partially” sensitive to island effects (e.g., rated 4.8/7, in comparison to its baseline (17b), which is rated 5.2/7). But on the other hand, Yoshida et al. (2019) found that the acceptability of contrast sluicing sentences like (17a) was almost as high as the baseline in (17b) (i.e., 5.02 vs 5.21). They conclude that contrast stripping is, in fact, not sensitive to islands.

The empirical evidence for island effects under clausal ellipsis is controversial. The current paper aims to experimentally investigate how regular and contrast sluicing interact with subject islands. We chose to investigate subject islands for two reasons: First, no prior study has quantified the behavior of subject islands under sluicing; and second, with subject islands, it is possible to isolate the effect of the island by comparing extraction from subject with extraction from object, both of which involve very similar structural complexity but only one of which involves an island. We conducted an acceptability rating study with a 3 × 2 factorial design, crossing Sluice Type with NP position. In the NP position manipulation, the correlate is embedded either inside a subject (“NP_subj”), from which extraction should give rise to an island effect, or inside an object (“NP_obj”), which serves as a non-island baseline. Sluice Type had 3 levels. An example item set is shown in Table 2. The Contrast condition consists of sluicing sentences that contrast a definite correlate with a remnant that includes an exceptive modifier (“which other x”). The Regular condition includes regular sluicing sentences with an indefinite correlate and a wh-remnant (“which x”). We included a baseline level for Sluice type, which had an embedded clause with no wh movement in it (called “NoSluice”). The NoSluice condition involves sentences conjoined by the subordinator but and followed by an articulated embedded clause (no ellipsis).

4.2 Results

We report the results of seventy self-reported L1 English speakers (1680 critical trials); two participants were excluded because they did not report English as their L1. The mean of the ratings of the grammatical filler sentences was 6.10 (SD = 1.30); the mean of the ratings of the ungrammatical fillers was 1.55 (SD = 0.98). This indicates that the acceptability scale was used reasonably.

Sentences in the Contrast condition received lower ratings (4.16) compared with the Regular (5.54) and No Sluice (5.31) conditions, as shown in Figure 2. Overall, sentences in NP_subj conditions, where the correlate is embedded in an NP in subject position, were rated lower than the ones in NP_obj conditions (4.86 vs 5.15). Among the three levels of Sluice Type, the difference between NP_subj and NP_obj was largest for contrast sluicing (3.91 vs 4.40).

Figure 2

Mean ratings across No Sluicing, Regular Sluicing, and Contrast Sluicing conditions for Subject NPs (gray bars) and for Object NPs (black bars). Error bars represent 95% confidence intervals. The dashed line at the top represents the mean ratings to the grammatical fillers (6.10). The dashed line below is for the ungrammatical fillers (1.55).

The raw ratings were normalized via z-score transformation prior to statistical analysis (with means and standard deviations estimated by participant). The z-score transformed data was fitted to a linear-mixed effects regression model with crossed random effects for participants and items (Baayen et al. 2008), implemented in lme4 in R (Bates et al. 2015). The factor of Position was sum coded. The three levels of Sluice Type were coded with centered Helmert contrasts: The first contrast compared the No Sluice condition (coefficient: 2/3) with the two sluicing conditions (regular and contrast), pooled (coefficient: –1/3 for each). The second contrast compared Regular Sluicing (coefficient: –1/2) with Contrast Sluicing (coefficient: +1/2; the coefficient for No Sluice in this comparison was 0). We followed the same statistical protocols as for Experiment 1, likewise supplementing our frequentist results with Bayes Factors.

The results of the linear mixed-effects regression model are shown in Table 3. Our analysis revealed a significant effect of Position, which manifested as lower ratings to sentences when the correlate was embedded in an NP in subject position as compared to object position (4.86 vs 5.15; β = –0.13, SE = 0.04, t = –3.36, p = .003; BF₁₀ = 4.11). No Sluice conditions were also rated as significantly better than the two sluicing conditions, pooled (5.54 vs 4.74, β = 0.2, SE = 0.07, t = 2.9, p = .007; BF₁₀ = 3). Furthermore, Contrast Sluicing was overall rated significantly lower than Regular Sluicing (4.16 vs 5.31; β = –0.61, SE = 0.05, t = –12.01, p < .0001; BF₁₀ = 2.2 × 10¹¹).

Table 3

Results of the linear-mixed effects regression model, Experiment 2. Significant effects are bolded.

EFFECT	β	SE	t	p
Position	–0.13	0.04	–3.36	.003
NoSluice vs. (Regular + Contrast)	0.2	0.07	2.90	.007
Regular vs. Contrast	–0.61	0.05	–12.01	<.0001
NoSluice vs. (Regular + Contrast) × Position	0.076	0.07	1.1	.27
Regular vs. Contrast × Position	–0.12	0.08	–1.6	.12

Interestingly, however, neither interaction was significant (ps > .12), meaning we find no evidence that the effect of Position (subject vs object) changed according to Sluice Type. In fact, the Bayes Factor analyses indicated that we have moderate to strong evidence for accepting the null hypothesis; namely, that the effect of position is the same for non-sluicing as for regular and contrast sluicing pooled (BF₁₀ = .09) and the same for regular sluicing and for contrast sluicing (BF₁₀ = .25). Thus, we find that there is no subject island effect in either regular sluicing or, surprisingly, in contrast sluicing (although contrast sluicing is rated worse than regular sluicing generally).

5 Experiment 3: Subject island effects in embedded wh questions

The goal of this experiment is to determine how the elided material in sluicing patterns with respect to the subject island effect, when it is pronounced in the form of an embedded wh question as in the bolded part of the example in (18).

Recent work by Abeillé et al. (2020) has shown that subject island violations are not judged uniformly. In matrix wh-questions, subject island violations are judged to be unacceptable. However, in relative clauses, subject island violations are not judged to be unacceptable. Here, we ask about whether they are acceptable in embedded wh-questions: Abeillé et al. (2020) did not test embedded wh-questions. Our question is whether they will pattern with matrix wh-questions, where subject island violations are unacceptable, or with relative clauses, where they are not. To this end, we designed a 3 (Clause Type) × 2 (Position) acceptability rating study to test subject island effects in embedded wh-questions. The Position factor had two levels: NP_subj involved wh phrases extracted from an NP in subject position (hence, a potential subject island violation). NP_obj involved wh phrases extracted from an object NP, which served as the baseline. The Clause Type factor had three levels: matrix wh question, embedded wh question, and relative clauses. None of the stimuli contained any form of ellipsis. Sample stimuli in all six conditions are presented in Table 4.

5.2 Results and discussion

We report the results from all participants, which involve 1296 critical trials. The average rating for the grammatical filler sentences was 6.25 (SD = 1.10), while the ungrammatical fillers received an average rating of 1.71 (SD = 1.07). The findings show that subject conditions were overall rated lower than object conditions (3.78 vs 4.74). The difference in ratings between subject and object conditions was largest in matrix wh question conditions (3.74 vs 5.04), suggesting a subject island effect, whereas this difference was smallest in relative clause conditions (3.98 vs 4.44). Importantly, embedded wh questions showed a pattern similar to matrix wh questions (3.63 vs 4.75, see Figure 3).

Figure 3

Mean ratings across matrix wh question, embedded wh question, and relative clause conditions for extraction from subject NPs (gray bars) and for extraction from object NPs (black bars). Error bars represent 95% confidence intervals.

Like with Experiment 2, raw acceptability ratings were z-score transformed prior to statistical analysis. Position was sum coded. The three levels of the Clause Type predictor were coded with repeated contrasts: The first coefficient contrasted the acceptability of sentences in the matrix wh question condition (coefficient: –2/3) with the embedded wh question (coefficient: +1/3 for embedded wh question conditions; +1/3 for relative clause conditions). The second coefficient contrasted the embedded wh question condition (coefficient: –1/3) with the relative clause condition (coefficient: +2/3 for relative clauses; –1/3 for matrix wh questions). We then followed the same statistical procedure as for Experiment 2. Table 5 shows the results.

Table 5

Results of the linear mixed-effects regression model, Experiment 3. Significant effects are bolded.

EFFECT	β	SE	t	p
Position	0.46	0.03	14.38	<.0001
Matrixwh vs embeddedwh	–0.1	0.04	–2.54	.011
Embeddedwh vs. RC	0.01	0.04	0.37	.7
Position × Matrixwh vs embeddedwh	–0.1	0.08	–1.35	.18
Position × Embeddedwh vs. RC	–0.29	0.08	–3.75	.0002

Our analysis showed that Position was a significant predictor: Sentences where a wh-phrase is extracted from a subject NP were rated significantly lower compared with extraction from an object (3.78 vs 4.74; β = 0.46, SE = 0.03, t = 14.38, p < .0001; BF₁₀ = 4562).³ There was a significant overall difference between ratings for matrix wh questions and embedded wh questions (4.39 vs 4.19; β = –0.1, SE = 0.04, t = –2.54, p = .011); however, the Bayes Factor analysis provided at best weak evidence in support of this difference (BF₁₀ = 1.3). The overall difference between embedded wh questions and relative clauses was not significant (p = .7; BF₁₀ = .04).

However, our key question about whether embedded questions pattern with matrix wh questions or with relative clauses with respect to the subject island effect concerns the interactions. Importantly, the interaction of position with the comparison between matrix wh questions and embedded wh questions was not significant (p = .18), and the Bayes Factor analysis returned a BF₁₀ of 0.21, thus providing moderate evidence that matrix and embedded wh questions do respond in the same way to the manipulation of position. However, the interaction of position with the comparison between embedded wh questions and relative clauses was significant (β = –0.3, SE = 0.08, t = –3.75, p = .0002; BF₁₀ = 91.3), providing very strong evidence that the subject island effect is different in relative clauses vs embedded wh questions. In short, matrix wh questions and embedded wh questions show a subject island effect, and this is attenuated in relative clauses (cf. Abeillé et al. 2020).

Thus, this experiment shows that embedded wh questions pattern with matrix wh questions regarding the subject island effect. Recall that the types of sluices we looked at in our Experiment 2 are structurally akin to embedded wh questions (e.g., Merchant (2001) and the work following) with an unpronounced embedded clausal complement. The current findings weaken the argument that the nature of subject islands causes the lack of island effect under sluicing because embedding a wh question does not attenuate the subject island effect, the way relative clauses do. At the same time, it is still an open question whether non-pronunciation (i.e., ellipsis) repairs island violations under contrast sluicing as claimed in Yoshida et al. (2019). We will revisit this next in our General Discussion.

6 General discussion

This study is the first to quantitatively explore how regular and contrast sluicing respond to (subject) islands in English. First, we found distinct patterns in regular sluicing and contrast sluicing concerning the syntax of the elided clause. In sentences compatible with contrast sluicing, there was a notable preference for sentence continuations that were syntactically isomorphic to the antecedent clause. This supports the common structural view that the elided clause in contrast sluicing must be isomorphic with its antecedent. On the other hand, regular sluicing permits and even prefers non-parallel syntactic structures: Experiment 1 found a strong preference for a copular continuation over an isomorphic continuation for sentences compatible with regular sluicing. We interpret this finding as indirect evidence in favor of the evasion approach, which says that island violations can appear to be repaired in sluicing because they can in fact be evaded. Under this approach, the clause that is elided does not actually contain an island (e.g., Barros et al. 2014). If this is correct, a special ellipsis repair mechanism would not be needed to explain the acceptability of regular sluicing with islands (cf., Merchant 2008). However, we acknowledge that our Experiment 1 findings could reveal preferences for pronounced continuations only, and may not be indicative of the unpronounced structure in regular sluicing. Before making any strong conclusions, we next aimed to quantify the island sensitivity of sluicing, particularly contrast sluicing. Contrary to predictions (e.g., Merchant 2008; Barros et al. 2014; Griffiths & Lipták 2014), Experiment 2 found that contrast sluicing is not sensitive to the subject island constraint, while Experiment 3 showed that pronounced versions of these sluicing clauses are. Experiment 3 found that embedded wh questions pattern with matrix wh questions and against relative clauses (Abeillé et al. 2020) in being sensitive to subject islands. Experiments 2 and 3 together challenge the consensus view in the literature according to which contrastive ellipsis is island-sensitive (see also Barros 2014b; Weir 2014; Yoshida et al. 2019). We suggest that previously reported informal judgments of island sensitivity in contrast sluicing may reflect a general degradation in acceptability compared to regular sluicing, rather than any effect of islands in particular. This general degradation was evident in Experiment 2, where non-island violating contrast sluicing sentences were rated lower in acceptability than their regular sluicing counterparts.

The findings of Experiments 1 and 2/3 seem to point in opposite directions. Experiment 1 suggests that evasion approaches to the island insensitivity of regular sluicing are on the right track. Since contrast sluicing does not allow a non-isomorphic continuation, we would then expect contrast sluicing to be sensitive to islands, as has been reported in the literature. This is not what Experiment 2 showed, however; that experiment found that contrast sluicing is not unacceptable when the presumed extraction takes place from a subject island. Experiment 3 verified that subject islands do constrain embedded question formation. Experiments 2 and 3, then, indicate that it is something about ellipsis in particular that gets around an island violation, as is proposed by ellipsis repair approaches. Taken together, our results suggest that the evasion approach and the repair approach might both be correct. If what is elided in sluicing is a simpler clause with no island (a copular clause, for instance), then that will certainly evade an island violation in regular sluicing. At the same time, it appears that ellipsis can also repair an island violation, since we found that contrast sluicing is not sensitive to subject islands, but contrast sluicing requires a syntactically isomorphic continuation. Since evasion is not a possibility in contrast sluicing (e.g., Barros et al. 2014), only repair-by-ellipsis could be behind the lack of an island effect.

One might think that allowing both repair by ellipsis and evasion is not the most parsimonious approach. However, once we allow repair by ellipsis, allowing evasion in addition does not require any additional theoretical machinery. So long as a non-isomorphic structure is possible for the elided clause (and it seems to be, if continuations are indicative), then it follows that there will be no island violation, since the structure contains no island. It is the repair-by-ellipsis approach itself that needs additional theoretical machinery. We believe that our results can actually help to reduce this machinery: If our findings are correct and there is no difference between contrast sluicing and regular sluicing in sensitivity to islands, then the different structures and mechanisms that have been proposed in the repair-by-ellipsis approach to account for the presumed contrast can be done away with. There only needs to be one structure and one repair mechanism. For example, Merchant (2008) proposes that in contrast sluicing, the focused correlate (e.g., Captain America) can only move as high as VP. In this position, it is not parallel to the wh-phrase in Spec-CP of the elided clause (e.g., which other superhero), and so ellipsis of IP is not licensed. This is in contrast with regular sluicing, where an indefinite correlate (e.g., a superhero) can apparently move as high as Spec-CP. Given our findings, this unmotivated distinction can be done away with: In both cases, the correlate is capable of moving to (or taking scope at) Spec-CP, providing the needed parallelism for ellipsis licensing.

An alternative to allowing both repair by ellipsis and evasion is to say that there is no evasion, there is only a general repair-by-ellipsis mechanism. On this alternative, pronounced continuations do not give us evidence about the structure of the elided clause in (regular) sluicing. Pronounced continuations are only evidence of the structure of pronounced continuations. We will leave this as an open possibility, although we do think that pronounced continuations are a clue to elided structure (because speakers have to be able to spell out an elided structure when required to).

Finally, non-syntactic approaches to islands, like the discourse-based analysis of Abeillé et al. (2020), will need to explain why ellipsis remedies whatever non-syntactic problem is supposed to give rise to an island effect. For Abeillé et al., this is a discourse clash in the case of subject islands: wh-phrases are focused, but originating inside the subject makes the wh-phrase backgrounded, and this conflict is what causes unacceptability. In sluicing and contrast sluicing, we cannot see that eliding the clause out of which the wh-phrase has been extracted should change either focus or backgrounding. In fact, the ellipsis is consistent with what is assumed to be the focus structure: The focused phrase is pronounced (the wh-remnant), while the backgrounded material is elided (the rest of the clause). In addition, it is clear that the meaning of the elided clause can be computed (it must be, to be interpreted). We do not see any obvious way to account for the saving effect of ellipsis in this discourse-based approach. It is therefore incumbent upon proponents of this approach, or any other non-syntactic approach to subject islands, to show how their approach can capture the saving effect of ellipsis.

Overall, the recent uptick in research which explores clausal ellipsis phenomena and island phenomena experimentally has helped to clarify the empirical picture and to direct analytical theory. By looking at how regular and contrast sluicing respond to subject islands, our findings have contributed in this same way. We expect that this trajectory should lead to much new exploration of the issues involved in clausal ellipsis and extraction phenomena.

Data availability

The stimuli and results for Experiments 1, 2 and 3, along with Bayesian models results, are publicly accessible through an OSF repository: https://osf.io/e5agt/?view_only=e6ee86c26d1049f7908322c9e5561ad9.

Acknowledgements

We thank Satoshi Tomioka, Rui Chaves and the audience at the Syntax Semantics Lab at the University of Delaware for their insightful comments and discussions. All remaining errors are ours.

Competing interests

The authors have no competing interests to declare.