1 Introduction

Using data on Sluicing in Saudi Arabic, this paper is a contribution to the debate on the syntactic analysis of ellipsis. On the basis of three acceptability judgment experiments on Saudi Arabic elliptical questions with prepositional phrases we argue that there is syntactic structure at the ellipsis site and that the wh-remnant in these elliptical questions can – and sometimes must – relate to a resumptive pronoun in the ellipsis site. The argument rests on the observation that the acceptability of elliptical questions depends on the availability of a corresponding full question.

The formation of elliptical questions in which only the wh-phrase is pronounced is referred to as sluicing (Ross 1969). Typical examples are given in (1).

(1)

In example (1a) the embedded question is understood to mean which car John bought, although what is pronounced is which one. We will refer to the pronounced material in the question as the remnant. The sluice itself is arguably a clausal constituent and thus different from the remnant (Levin 1982; Ginzburg and Sag 2000; Merchant 2001; Culicover and Jackendoff 2005); unlike the remnant, the sluice is a clause. Under some theories, the sluice is made up of a remnant and an ellipsis site. The content of the sluice is recovered via a contextually given antecedent. In a canonical example like (1a) the remnant corresponds to an indefinite in the antecedent: the correlate¹ (Chung, Ladusaw & McCloskey’s (1995) ‘inner antecedent’). Chung et al. (1995) coined the term ‘merger type sluicing’ for examples like (1a), in which there is an overt indefinite correlate in the antecedent and the sluice questions its identity. When the sluice questions the identity of a different entity, we speak (following Merchant (2001)) of contrast sluicing. Example (1b) with what else as the remnant is an instance of contrast sluicing.

Following Dayal & Schwarzschild (2010), we refer to the full clausal structure hypothesized to fill the ellipsis site under some theories as the pre-sluice. In (1a) there are two plausible pre-sluices: which one he bought and which one it is.

Despite intense research, no consensus has been reached about whether the understood material is syntactically represented at the ellipsis site (Ross 1969; Lakoff 1970; Baker & Brame 1972; Chomsky 1972b; Chung et al. 1995; Merchant 2001; Lasnik 2005; Fukaya 2007; Van Craenenbroeck 2010a; 2010b; Müller 2011; Fukaya 2012; Barros 2014; Barros et al. 2014; Griffiths & Lipták 2014; Abe 2015; Abels 2017a) or not (Erteschik-Shir 1973; Levin 1982; Dalrymple et al. 1991; Ginzburg & Sag 2000; Culicover & Jackendoff 2005; Sag & Nykiel 2011; Barker 2013). That is, it is unclear whether the ellipsis site is occupied by a syntactically represented pre-sluice whose structure varies in some way with the antecedent clause.

The debate persists, because the evidence that bears on the structure of the ellipsis site is necessarily indirect and appears to point to different conclusions: Ross’s (1969) famous observation that sluicing repairs island violations has been taken to indicate (Culicover & Jackendoff 2005) that there is no structure at the ellipsis site (although see Barros et al. 2014; Abels 2017c for a different assessment). On the other hand, the observation – also due to Ross (1969) – that the correlate and the remnant match in morphological case is often interpreted as evidence not just for the presence of syntactic structure at the ellipsis site but for the presence of structure syntactically isomorphic to the antecedent (Ross 1969; Lasnik 2005 but see Ginzburg & Sag 2000; Culicover & Jackendoff 2005; and for discussion of more complex cases see Kim 2015; Vicente 2015; Wood, Barros, and Sigurðsson 2016; Abels 2017c; Kidwai 2018). Finally, Ross (1969) suggested that sluicing is derived by wh-movement in the ellipsis site. The basis of this claim is that sluicing obeys constraints on pied-piping in corresponding full questions. However, different authors have idealized the relevant facts in a number of ways and come to diverging conclusions.

The current paper picks up the theme of pied-piping. We set island amelioration and case matching between correlate and remnant aside. The issue of case matching is moot since Saudi Arabic has no morphological case distinctions. Islands are ameliorated under sluicing in Saudi Arabic, but the examples we are aware of do not allow us to choose between different theories of island amelioration.

As we alluded to above, Ross (1969) proposes a theory of sluicing according to which the ellipsis site contains syntactic structure identical to the antecedent and the remnant is extracted by regular wh-movement. This approach predicts Merchant’s (2001) second form identity generalization: ‘language L will allow preposition stranding, under sluicing iff L allows preposition stranding under regular wh-movement’ (Merchant 2001: 92). This is so because, under Ross’s account, constraints on pied-piping, and in particular pied-piping of prepositions, are enforced under sluicing (Abels 2019). A theory of sluicing without structure at the ellipsis site makes no such prediction (Sag & Nykiel 2011; Nykiel 2013; Kim 2015): whether the preposition appears or does not appear in the sluice is independent of syntactic constraints on pied-piping and determined by other factors. Merchant (2001) claimed that the prediction of the Ross-style theory is correct. Recognizing the importance of the issue, a lot of literature since has probed the truth of this generalization, and produced an impressive number of counterexamples to the generalization (Szczegielniak 2006; Vicente 2006; Almeida & Yoshida 2007; Fortin 2007; Stjepanović 2008; Szczegielniak 2008; Rodrigues et al. 2009; Sato 2010; Van Craenenbroeck 2010a; Sato 2011; Wei 2011; Algryani 2012; Stjepanović 2012; Vlachos 2012; Nykiel 2013; Adliene 2014; Philippova 2014; Leung 2014a; 2014b; Albukhari 2016; Alshaalan 2017; Abels 2017a; Stigliano 2018; Molimpakis 2019).

For ease of reference, we will refer to examples of sluicing where the remnant is a DP and the correlate DP is the complement of a preposition as OPUS, which the reader is invited to interpret as an abbreviation, depending on their theoretical predilections, as Ostensible P-stranding Under Sluicing or as Omission of Preposition Under Sluicing. OPUS is the name of a phenomenon in search of an analysis. There have been two predominant reactions to the existence of OPUS. Some researchers have analysed the phenomenon in terms of wh-movement in the ellipsis site but with pre-sluices that are semantically but not necessarily syntactically identical to the antecedent (Rodrigues et al. 2009; Van Craenenbroeck 2010a; 2010b; Algryani 2012; Barros 2014; Barros et al. 2014; Leung 2014b; Albukhari 2016; Abels 2017a). Others have taken OPUS to demonstrate the absence of structure at the ellipsis site (for fragment answers see Culicover and Jackendoff 2005; and for analyses on sluicing see Sag & Nykiel 2011; Barker 2013; Nykiel 2013). The two types of theory differ in their predictions. A theory that allows paraphrases of the antecedent as pre-sluices predicts that OPUS will not be acceptable unless there is a suitable paraphrase; a theory without structure at the ellipsis site doesn’t predict that the availability of a paraphrase should affect the judgments. The discussion in this paper addresses the predictions of both types of theory.

The logic of the experiments and the main thrust of the argumentation employed here is different from many previous experimental and psycholinguistics studies on ellipsis. These studies have dealt with structures for which there is a well-formed candidate structure that can potentially fill the ellipsis site but which differs from the antecedent in various ways. The studies have then measured how the structural distance between the overt antecedent and the hypothesized pre-elliptical structure affects a variety of variables. The literature reviewed in Frazier (2018) under the heading ‘Empirical evidence for syntactic structure at the ellipsis site’ is typical in this regard.² Our own experiments do not try to measure the effect of structural distance between antecedent and hypothesized elided structure. Instead we attempt to demonstrate that there must be some structure in the ellipsis site (isomorphic or not) by showing that if no candidate structure exists, there is an acceptability penalty.

One language that challenges the idea of isomorphic pre-sluices and antecedents is Polish. Nykiel (2013) showed that examples of OPUS are significantly more acceptable than P-stranding in non-elliptical structures. However, examples of OPUS are also less acceptable than pied piping under sluicing. She further shows (contra to Szczegielniak 2008) that wh-clefts cannot be the source for these P-less sluices. She claims that complexity of the wh-remnant and that of the correlate affect the acceptability of P-less sluices.³

Molimpakis’s (2019) work on Greek presents broadly similar results (see Table 1). In Molimpakis’s (2019) experiments there are no well-formed candidate pre-sluices for examples of OPUS. The puzzling result of her study is that OPUS is less acceptable than pied-piping (both in elliptical and non-elliptical conditions) but at the same time much more acceptable than P-stranding in non-elliptical conditions (or any other plausible pre-sluice).⁴

Such results lead to an argumentative impasse. On the one hand, proponents of a structural approach can claim that these results support their theory, since there is a clear decrease in acceptability when the preposition is missing. On the other hand, proponents of a non-structural account can claim that the results support their theory, since there is a clear and very large increase in acceptability of P-less sluices over overt P-stranding.

The current set of experiments on Saudi Arabic aim to overcome this impasse by explicitly contrasting structures with and without well-formed pre-sluices. Experiment 3 in particular probes structures that differ in the availability of a fully acceptable pre-sluice. By contrasting examples of OPUS with where versus when as remnants, we demonstrate that the judgments are modulated by the existence or lack of a grammatical pre-sluice. The results thus suggest that there is structure at the ellipsis site.

The experiments also bear on the question of what factors, apart from the existence of a well-formed pre-sluice, might influence the acceptability of OPUS. We focus on Nykiel (2014; 2015; 2017) as the most explicit model of such effects.

Working within a non-structural approach to ellipsis resolution, Nykiel in a series of papers develops a model of the factors influencing the acceptability and prevalence of what she calls ellipsis alternation, that is, the alternation between PP and DP remnants of clausal ellipsis (sluicing’s OPUS being but one case), where the correlate is a PP or the DP complement of a preposition, respectively. Given her non-structural approach to the ellipsis site, there is no a priori reason to expect the ability of a preposition to be absent under ellipsis to be correlated with its ability to be stranded under syntactic movement in any language (Nykiel 2012). In our view, considerations of learnability do indeed suggest that ellipsis resolution cannot be parameterized due to the lack of triggering evidence. Therefore, cross-linguistic differences in the phenomenology of ellipsis resolution should either be absent or reduceable to independently known properties of the languages in question.

Nykiel (2014; 2015) develops a model of the propensity of the preposition to be absent under clausal ellipsis based on four factors, which she claims to have cross-linguistic validity.

The first factor is construction type. Nykiel’s (2013; 2015) corpus work covers sluicing, bare argument ellipsis, reprise questions, and split questions. We will not consider this factor here, because we only use sluicing in our experiments and so the factor of construction type is held constant.

The second factor has to do with “the semantic and syntactic content of the wh-phrase serving as a remnant” (Nykiel 2017: 29). The idea is that less informative and structurally less complex remnants are dispreferred in examples of OPUS compared to more informative and structurally more complex remnants. Nykiel supports this claim based on qualitative comments in the syntactic literature, on her own (2013a) work on Polish (though see footnote 4 above for reasons to be skeptical), as well as corpus work on contemporary (Nykiel 2017) and previous stages of English (Nykiel 2015). It is difficult to judge the relevance of this factor for sluicing based on Nykiel’s corpus data, since the effect of what we will call informativity is only reported for the entire corpus which is heavily skewed towards fragment answers. Examples of sluicing make up only about one in seven examples in Nykiel’s (2015) historical corpora and, judging by the (2013) version of Nykiel (2017), there are only 9 examples of sluicing in total in the corpus work on contemporary English.

This second factor of informativity is manipulated in experiment 1, which used both bare indefinites and contentful DPs as correlates. In the experiment we did find a small effect of informativity but this effect failed to replicate in a second experiment, which we do not report here in full (but see footnote 18).

The third factor in Nykiel’s (2013; 2015; 2017) model is structural persistence. This factor distinguishes B’s utterance in (2), where the antecedent contains a PP but the elliptical utterance just contains a noun phrase (lack of persistence), from A’s final utterance, where the immediate antecedent contains no preposition and the noun phrase is structurally persistent:

The idea behind structural persistence is that the form chosen in the antecedent utterance conditions the form chosen for the elliptical utterance. A PP in the antecedent increases the likelihood of a PP as the remnant, while a DP in the antecedent increases the likelihood of a DP as the remnant. Moreover, on Nykiel’s assumption that corpus frequency correlates positively with acceptability, persistent sluicing remnants should be more acceptable than non-persistent ones. Structural persistence is, in essence, a syntactic priming effect. It has been found in psycholinguistic work on ellipsis (Levelt & Kelter 1982) and plays a role in predicting corpus frequencies in historical and contemporary English (Nykiel 2015; 2017). We should point out that since Nykiel’s corpora are heavily skewed towards fragment answers and since Levelt and Kelter’s (1982) paper also deals with fragment answers, the relevance of structural persistence to sluicing has never been demonstrated. The factor of structural persistence is held constant in experiments 1 and 3 of the present paper, but it is manipulated in experiment 2, where we compare structures that obey persistence with structures that disobey persistence. Structural persistence predicts that structures that disobey persistence should be less acceptable than structures that obey persistence. We did not find the predicted effect.

The fourth and final factor has to do with the semantic dependency between the preposition and the verb. Nykiel employs two tests taken from Hawkins (1999) for semantic dependency between PP and verb. The verb is dependent on the preposition if a sentence with the PP does not entail the sentence without the PP. The verb is independent if the sentences with the PP does entail the one without the PP: I counted on Peter does not entail I counted, thus, count is dependent on the PP in the first sentence. On the other hand, I read on Tuesday entails I read. Thus, read is independent of the PP. Dependency of the PP on the verb (i.e., dependency in the other direction) is tested by checking whether the sentence with the verb and PP entails a sentence with the same PP and a generic predicate. For example, I read on Tuesday entails Something happened on Tuesday and thus the PP is independent of the verb. On the other hand, I counted on Peter does not entail Something happened on Peter or I did something on Peter, etc. The PP is therefore dependent on the verb. In Nykiel’s (2015; 2017) corpus data, a dependency between verb and prepositions in either direction makes the preposition less likely to appear in the remnant. Employing again the idea that corpus frequency is positively correlated with acceptability, OPUS is predicted to be more acceptable when there is a dependency than when there is none. Experiments 1 and 2 largely used verbs and PPs that have a dependency; experiment 3 used verbs and PPs that – by Nykiel’s own criteria – are independent of each other. Since in experiment 3 all factors that enter into Nykiel’s model are held constant, the model would seem to predict a null result. The fact that we did find an effect thus demonstrates the insufficiency of the model.

The remainder of the paper is structured as follows. In section 2, we review what has been proposed about OPUS in other Arabic dialects and provide background on question formation in Saudi Arabic. In sections 3–5 we report the results of three experiments. The methods and design of these experiments are modelled closely on Molimpakis’s (2019) study of OPUS in Greek. In section 3 we report the results of an experiment showing that OPUS is permissible in Saudi Arabic. To block the availability of a wh-cleft in the ellipsis site, which the previous literature had argued to be the source of OPUS in Arabic, we employ contrast sluices in the experiment. We conclude that under a structural approach to the ellipsis site, resumption must be available. The experiment reported in section 4 compares judgments on OPUS with judgments for canonical sluices with DP-remnants. All examples in the experiment have acceptable pre-sluices, although they are not necessarily isomorphic to the antecedent. We find that in this type of setup there is no degradation coming from the lack of a preposition in the remnant. Hence, what is interpreted under a structural approach as the construction of an alternative pre-sluice does not seem to come with a measurable degradation. Finally, experiment 3 in section 5 shows that sluices without a plausible paraphrase as the pre-sluice are less acceptable⁵ than those with a suitable paraphrase that could serve as the pre-sluice. We interpret these results to show that there is syntactic structure at the ellipsis site. Various further conclusions and issues are discussed in the concluding section.

2 On sluicing and question formation in Saudi Arabic

OPUS has previously been noted for other Arabic dialects: Jordanian (Albukhari 2016), Emirati (Leung 2014a; 2014b), Libyan (Algryani 2012). Working within structural approaches to sluicing, these authors suggest that sluicing in Arabic can be derived from two (roughly) synonymous types of pre-sluices: wh-movement structures that are structurally identical to the antecedent and wh-clefts. As illustrated below, wh-clefts do not require preposition pied-piping. The above authors use this observation to suggest that OPUS in Arabic does not argue against the existence of structure at the ellipsis site but only against the idea that the structure at the ellipsis site must be syntactically identical to the antecedent.⁶ We should point out that while the data presented in the papers cited is compatible with a structural approach to sluicing, this approach is not forced, because the papers do not probe what happens when no well-formed pre-sluice is available. Furthermore, the authors consider only wh-clefts as an alternative pre-sluice, but we will demonstrate the relevance of resumptive pronouns in experiments 1 and 2 under a structural analysis.

In order to understand this criticism of the existing literature on Arabic and the experiments below, it is important to realize that Arabic, including Saudi Arabic, has three⁷ clearly distinct, relevant strategies of wh-questions formation: wh-movement⁸ (3a), wh-clefts (3b), and wh-resumption (3c). The wh-movement strategy is characterized by the fronting of a wh-phrase, the appearance of a gap in the question nucleus, and the absence of a relative complementizer or (pronominal) copula between the wh-phrase and the question nucleus. The wh-clefting strategy is characterized by the fronting of a wh-phrase, the optional presence of a pronominal copula, the obligatory presence of the relative complementizer (aly), and the presence of an obligatory overt resumptive pronoun in most syntactic positions except for the local subject position (see McCloskey 1990; Shlonsky 1992).⁹ Finally, the resumptive strategy is characterized by the absence of the relative complementizer, the absence of a pronominal copula, and the presence of a resumptive pronoun.

These three strategies interact in different ways with contrastive wh-phrases, with preposition pied-piping, and with wh-type. Preposition pied-piping is always enforced with wh-movement (4a) but it is optional with wh-clefts (4b), and wh-resumption¹⁰ (4c).

Contrastive wh-phrases on the other hand, are possible with wh-movement (5a) and wh-resumption (5c) but not with wh-clefts¹¹ (5b). We agree with a reviewer, who suggests that the deviance of (5b) is likely owed to the exhaustivity of clefts.

Contrast interacts in the expected way with pied piping of prepositions: wh-movement requires pied-piping, wh-clefting is incompatible with contrast whether pied-piping occurs or not, and resumption is compatible with contrast and pied-piping. Relevant examples are given below in section 3, Table 4.

We should note that other kinds of contrast are compatible with the cleft strategy. The following context contrasting boys with girls does not distinguish the three strategies:

Since the aim of the experiments is to distinguish the three strategies from each other, we use examples like (5) and set examples like (6) aside as uninformative.

Finally, wh-movement is compatible with all kinds of wh-phrases (nominal, d-linked, non-d-linked, and adverbial). Wh-clefts (7a–9a) and resumption (7b–9b) are somewhat restricted: they do not occur with true adjuncts how and why nor with when. Both do allow where, as we show immediately below.

The literature on other Arabic dialects reports that wh-clefts are restricted to nominal wh-phrases only (Shlonsky 2002), but we do not find this restriction in Saudi Arabic. The literature on resumption (Aoun et al. 2010) also leads one to further expect non-d-linked phrases such as what or where to be excluded from the resumption strategy. However, this expectation is thwarted in Saudi Arabic. Examples (10) and (11) show that resumption and clefting are possible with what and with where. These judgments are corroborated in experiments 1 and 3, respectively. As expected, resumption of which-phrases and of who are unproblematic in Saudi Arabic. The situation is summarized in Table 2.

Two themes in the literature on resumption are island (in-)sensitivity and reconstruction, and their interactions (see Aoun et al. 2010). In Saudi Arabic, the cleft and the resumptive strategy are both island insensitive while the wh-movement strategy is, expectedly, island sensitive. We have not systematically investigated the reconstructive properties of the three question types. Island (in-)sensitivity is not considered in this paper, because the target construction of our investigation, sluicing, is independently known to lack island effects (Ross 1969). Reconstruction effects of the type studied in Aoun et al. (2010) are very difficult to investigate experimentally. Our experimental aims here are much more modest.

We should note that a descriptive three-way distinction between wh-question types characterizes other Arabic dialects as well. Some researchers (e.g. Algryani 2012) suggest that the structures underlying what we call the cleft-strategy and what we call the resumptive-strategy are identical. Under this view the resumptive strategy – like the cleft strategy – is a bi-clausal copular structure with a null copula and a null relative complementizer. We do not follow this line of analysis, as it fails to explain the differences between the two strategies noted above and offers no explanation for why the relative complementizer should be null just in this particular kind of structure; Saudi Arabic does not usually allow null relative complementizers.

We leave open the question of whether resumptive pronouns (outside of island contexts) can or must be the residue of movement.

3 Experiment 1: Are clefts the only non-isomorphic pre-sluice in Saudi Arabic?

In experiment 1 we investigated whether wh-clefts are the only possible non-isomorphic paraphrase in the ellipsis site as has been suggested by Algryani (2012); Leung (2014b); Albukhari (2016). As mentioned in section 2, some of the previous literature on Arabic dialects had observed the acceptability of OPUS under merger type sluicing and analysed it in terms of an elided wh-cleft.¹² Recall that in Saudi Arabic contrastive wh-phrases are possible with wh-movement and wh-resumption but not with wh-clefts (Table 2). We focused here on contrast sluices to block the availability of wh-clefts in the ellipsis site. If wh-movement and wh-clefts are the only possible pre-sluices in the ellipsis site, OPUS under contrast sluicing should be impossible: wh-movement is impossible because Saudi Arabic disallows P-stranding, and wh-clefts are impossible because clefts are incompatible with contrastive semantics (see Rodrigues et al. 2009 for just this claim regarding OPUS in Spanish). The results did not conform with the expectation of an account where OPUS necessarily derives from a wh-cleft. OPUS under contrast sluicing is acceptable. The results are compatible with a structural approach to ellipsis resolution only if a different structure is available in the ellipsis site, which we will argue to be resumption later on. The result is also compatible with a non-structural approach.

To investigate these questions, we conducted a web-based acceptability judgment experiment. Items were constructed crossing three factors: left-peripheral P (absent vs. present), ellipsis (sluicing vs. non-elliptical structure), and type of wh-phrase (simplex vs. complex). If wh-clefts are the only possible pre-sluice, we expect a significant main effect of the factor left-peripheral P, with the conditions without a left-peripheral preposition being uniformly unacceptable no matter whether ellipsis takes place or not.

The third factor concerns the type of wh-phrase and contrasts the behaviour of simplex wh-phrases (who and what) with that of which-phrases.¹³ This factor was included to test whether the informativity of the remnant would affect the judgments on OPUS, as suggested by Nykiel (2013; 2015; 2017).

For ease of reference, Table 3 summarizes the theories mentioned and their predictions. ‘Wh-cleft’ in the table refers to the account according to which OPUS in Saudi Arabic rests exclusively on wh-clefts as pre-sluices. ‘Wh-resumption’ in the table refers to the account according to which OPUS might in addition involve a resumptive structure in the pre-sluice. No structure refers to the fine-grained predictions made by Nykiel’s model. The latter predicts an interaction between type of wh-phrase and ellipsis. As explained in the introduction, we focus on Nykiel’s work because it stands out among non-structural accounts in developing fine-grained predictions.

3.2 Results

Our research question was whether wh-clefts are the only possible non-isomorphic pre-sluice for OPUS in Saudi Arabic. We tackled this question by asking whether contrast sluicing would allow OPUS. Table 5 shows the average rating provided by participants for each of the eight Conditions. In general, OPUS under contrast sluicing is as acceptable as corresponding non-elliptical sentences both with simplex and complex wh-phrases. P-stranding in non-elliptical structures is significantly less acceptable than any of the other conditions.

Table 5

Experiment 1: Mean acceptability rating by Condition (n = 29).

	Condition	Mean	SD
Condition 1	Left-peripheral P absent, sluicing, simplex	4.7913	1.96918
Condition 2	Left-peripheral P absent, sluicing, complex	5.3247	1.84467
Condition 3	Left-peripheral P present, sluicing, simplex	5.4217	1.81946
Condition 4	Left-peripheral P present, sluicing, complex	5.3391	1.93062
Condition 5	Left-peripheral P absent, non-elliptical, simplex	2.3074	1.74851
Condition 6	Left-peripheral P absent, non-elliptical, complex	2.2121	1.68446
Condition 7	Left-peripheral P present, non-elliptical, simplex	5.2445	2.03727
Condition 8	Left-peripheral P present, non-elliptical, complex	5.1810	2.03476

This is obvious in Figure 1. The conditions without left-peripheral preposition with sluicing are significantly more acceptable both with simplex wh-phrases (M = 4.7) and complex wh-phrases (M = 5.3) than the non-elliptical P-stranding conditions with (M = 2.3) and (M = 2.2), respectively. As for the pied-piping conditions, both simplex wh-phrases and complex wh-phrases were rated as highly acceptable under sluicing (M = 5.4) and (M = 5.3) and in non-elliptical conditions (M = 5.2) and (M = 5.1), respectively.

Figure 1

Experiment 1: Mean acceptability rating by Condition (n = 29). Error bars represent the standard error of the mean (SEM).

The data of 29 participants (232 datapoints per condition; a total of 1856 datapoints) were analysed using a linear mixed-effects model using ‘lme4’ package in R (Bates et al. 2014). Prior to the analysis, raw ratings were z-score transformed to eliminate some of the forms of scale bias that potentially arise with rating tasks (see Schütze and Sprouse 2014 for a review).¹⁷ Fixed factors were left-peripheral P, ellipsis, and wh-type with two-levels each. Random intercepts and slopes were assumed for participants and items as random effects in order to control for individual variation as well as for any variation in the dependent variable across items. All reported P-values were estimated using the Satterthwaite approximation; R package ‘lmerTest’ (Kuznetsova et al. 2017).

As indicated in Table 6, results show a highly significant main effect of ellipsis (t = –5.740, p = 2.56e-08), with sluiced conditions rated on average as much more acceptable than non-elliptical conditions. We also found a significant main effect of wh-type (t = 3.444, p < .001).¹⁸ Although there was no significant main effect of pied-piping (t = –1.050, p = 0.29), results show a highly significant interaction between left-peripheral P and ellipsis (t = 4.152, p = 4.91e-05) reflecting the fact that P-stranding in non-elliptical conditions is substantially degraded, while OPUS is acceptable to a degree comparable to pied-piping both under sluicing and in the non-elliptical conditions. The results also show a significant three-way interaction and two two-way interactions between wh-phrase type, pied-piping, and ellipsis. This is driven by the different behaviours of simplex and complex wh-phrases under sluicing and P-stranding. (See footnote 18 for discussion).

Table 6

Experiment 1: Summary of linear mixed effects models P-values estimated using the Satterthwaite approximation (***p < .001).

	t-value	Pr(>|t|)
(Intercept)	7.386	3.26e-12***
Left-peripheral P	–1.050	0.294580
Ellipsis	–5.740	2.56e-08***
Wh-type	3.444	0.000766***
Left-peripheral P * ellipsis	4.152	4.91e-05***
Left-peripheral P * wh-type	–2.856	0.004619**
Ellipsis * wh-type	–2.937	0.004302**
Left-peripheral P * ellipsis * wh-type	2.356	0.020677*

• Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

Since the data was not normally distributed, a non-parametric paired Wilcoxon tests was used to determine whether there is a difference between left-peripheral P present and left-peripheral P absent under sluicing with different types of wh-phrases. We found no significant difference between condition 4 ‘left-peripheral P present under sluicing’ and condition 2 ‘left-peripheral P absent under sluicing’ with complex wh-phrases (p = .91). However, we did find a significant difference between condition 3 ‘left-peripheral P present under sluicing’ and condition 1 ‘left-peripheral P absent under sluicing’ with simplex wh-phrases (p < 0.001), both conditions were highly acceptable (M = 4.7) and (M = 5.4), respectively. We also found a significant difference between OPUS with simplex wh-phrases (condition 1) and complex wh-phrases (condition 2).

4 Experiment 2: Do resumptive pronouns lead to measurable degradation in Saudi Arabic?

In experiment 2 we probe the idea that OPUS relies on a resumptive structure in the ellipsis site. This can only be done if there is a discernible signal from resumption. For example, if wh-resumption were generally less acceptable than wh-movement, then all structures with resumption should be less acceptable than minimally different structures without resumption (i.e. conditions 1, 5, 6, 7 and 8 in the experiment should be less acceptable than conditions 2, 3, and 4). Alternatively, if wh-resumption was less acceptable than wh-movement unless a resumptive pronoun is obligatory in a particular position, then the unforced use of optional resumptive pronouns should lead to a measurable degradation (i.e. conditions 5, 6, 7 and 8 in the experiment should be less acceptable than conditions 1, 2, 3, and 4). Finally, if resumptives had a mild semantic incompatibility with contrast,¹⁹ then conditions with both resumption and contrast (i.e. conditions 1, 5, and 7) should be less acceptable than resumptive conditions without contrast (i.e. conditions 2, 4, 6, and 8) and contrastive conditions without resumptive pronouns (i.e. condition 3).

Some of the literature on resumption treats resumption as a last resort phenomenon (Rizzi 1990; Shlonsky 1992; Aoun 2000; Aoun et al. 2001). This may suggest – but by no means forces – the hypothesis that structures with resumptive pronouns are less acceptable than structures without them. The idea that there might be an acceptability cost of resumption (or of unforced resumption or of resumption in conjunction with contrast) thus underpins the logic of the experiment. The experiment also tries to shed light on the question of what drove the low acceptance of all eight conditions by some speakers in experiment 1. In particular, these speakers might have rejected the sentences based on the way contrast is expressed or based on some other unknown property of the examples. We probe this by explicitly introducing contrast as a factor. Finally, the non-contrastive conditions additionally serve to verify that Saudi Arabic allows OPUS under merger type sluicing in line with the other Arabic dialects discussed in the literature cited above.

To investigate the question of whether there is a resumptive penalty and whether it shows up with examples of OPUS, we conducted a web-based acceptability judgment experiment. Items were constructed crossing three factors: Preposition (present vs. absent in the antecedent), ellipsis (sluicing vs. non-elliptical structure), and contrast (contrastive vs. non-contrastive). See Table 8. The left peripheral wh-phrase in all conditions was a complex DP. We chose complex (d-linked) wh-phrases since they are uncontroversially able to antecede resumptive pronouns in all Arabic dialects. All conditions involved a left peripheral DP in the target clause; there were no conditions with pied-piping of prepositions in this experiment. The factor of preposition distinguished instead between cases where that left-peripheral DP corresponds to the complement of a preposition in the antecedent (preposition present) and cases where it does not (preposition absent), i.e., where the DP acts as the complement of a verb in the antecedent instead of as the complement of a preposition. The factor of ellipsis distinguished between sluices and non-elliptical structures, but in contrast to experiment 1, the non-elliptical structures do not contain gaps but resumptive pronouns. Finally, the third factor distinguished contrastive examples like in experiment 1 from non-contrastive ones.

We hypothesized that the ellipsis site can, and at least under contrastive OPUS must, contain a resumptive pronoun. We consider a variety of hypotheses that assign no cost to resumption, a cost to resumption in all cases, a cost to resumption unless forced, or a cost to resumption in contrastive environments. Furthermore, from the perspective of Nykiel’s model, the experiment manipulates the factor of structural persistence. In the conditions without a PP in the antecedent, a DP remnant is forced (and obeys persistence), while in the conditions with a PP in the antecedent a DP remnant violates persistence and may alternate with a PP remnant. To the best of our knowledge, this is the first systematic test of structural persistence in a judgment task.

Briefly, persistence predicts that in experiment 1 and the experiment described in footnote 18 DP remnants should be less acceptable than PP remnants and that in experiment 2, DP remnants in the condition with a preposition in the antecedent should be less acceptable than in the condition without a preposition in the antecedent. The expected effect of persistence did not materialize in the first experiment.²⁰ No prediction about the non-elliptical conditions are made.

A theory which assumes structure at the e-site predicts that if resumption is not costly, no main effect or interactions are predicted. That is, all conditions should be rated equally. If resumption is costly in all cases, we would expect a main effect of preposition and an interaction with preposition and ellipsis only (elliptical conditions with no preposition in the antecedent don’t require a resumptive in the ellipsis site). That is, sluices with no preposition in the antecedent should be more acceptable than the other conditions because they do not force or overtly involve resumption. If resumption is costly unless it is forced, the unforced resumptive in conditions 7 and 8 below should give a strong signal. Finally, a theory that assumes that resumption is costly only in the contrastive environment, would predict only a main effect of contrast but no main effect of preposition or ellipsis. Under this view, all non-contrastive conditions are expected to be rated higher than all contrastive conditions.

4.2 Results

Our research questions asked whether there is a resumptive penalty and whether it shows up under sluicing with OPUS. The experiment also tried to shed light on the question of what drove the low acceptance of all eight conditions by some speakers in experiment 1.

The overall results (n = 70) are indicated in Figure 2. We again found that there is a class of speakers who reject all contrastive conditions.²³ The results for these speakers (n = 13) are given separately in Figure 3. Given the small dataset, we were unable to investigate these data very deeply. Figure 3 indicates that these participants generally reject the contrastive conditions (with or without preposition and in non-elliptical and elliptical structures alike). These speakers’ judgments were excluded from the rest of the analysis. We have no insight to offer into what exactly drove these speakers to reject all contrastive conditions.²⁴

Figure 2

Experiment 2: Mean acceptability rating by Condition (n = 70). Error bars represent SEM.

Figure 3

Experiment 2: Mean acceptability rating by Condition (n = 13). Error bars represent SEM.

The data of the remaining 57 participants (see Figure 4, 342 datapoint per condition; a total of 2736 datapoints) were analyzed further. Table 8 shows the average rating provided by participants for each of the eight conditions. In general, the presence/absence of the preposition in the antecedent did not affect the acceptability of the conditions.

Figure 4

Experiment 2: Mean acceptability rating by Condition (n = 57). Error bars represent the standard error of the mean (SEM).

Table 8

Experiment 2: Mean acceptability rating by Condition (n = 57).

	Condition	Mean	SD
Condition 1	Preposition present, sluicing, contrast	6.0351	1.47071
Condition 2	Preposition present, sluicing, non-contrast	6.4647	1.32658
Condition 3	Preposition absent, sluicing, contrast	6.0265	1.49607
Condition 4	Preposition absent, sluicing, non-contrast	6.3655	1.49615
Condition 5	Preposition present, non-elliptical, contrast	5.9735	1.50590
Condition 6	Preposition present, non-elliptical, non-contrast	6.2375	1.51246
Condition 7	Preposition absent, non-elliptical, contrast	5.9471	1.53335
Condition 8	Preposition absent, non-elliptical, non-contrast	5.9971	1.70552

As illustrated in Figure 4, OPUS is highly acceptable both with contrastive (M = 6.03) and with merger type sluicing (M = 6.4). Conditions without a preposition were also rated as highly acceptable both under contrast sluicing (M = 6.02) and merger sluicing (M = 6.3). This result is mirrored in the non-elliptical conditions; with all four conditions being highly acceptable.

As in the previous experiment, raw ratings were z-score transformed prior to analysis. A linear mixed-effects model was fitted with preposition, ellipsis, and contrast as fixed factors and subject and item as random effects with random slopes and intercept.²⁵ As indicated in Table 9, no significant main effects or interactions were found. This is due to all eight conditions being highly acceptable. The fact that there was no significant main effect of presence/absence of a preposition in the antecedent (t = 0.103, p = 0.918), no interaction between ellipsis and the presence/absence of a preposition in the antecedent (t = 0.148, p = 0.882), nor a three way interaction (t = –0.417, p = 0.677), indicates that the presence or absence of the preposition in the antecedent did not affect the acceptability of the results in either elliptical or in non-elliptical structures.

Table 9

Experiment 2: Summary of linear mixed effects models P-values estimated using the Satterthwaite approximation (***p < .001).

	t-value	Pr(>|t|)
(Intercept)	–0.530	0.597
Preposition	0.103	0.918
Ellipsis	0.203	0.839
Contrast	1.369	0.173
Preposition * ellipsis	0.148	0.882
Preposition * contrast	–0.053	0.957
Ellipsis * contrast	–0.360	0.719
Preposition * ellipsis * contrast	–0.417	0.677

• Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1.

5 Experiment 3: OPUS when no well-formed pre-sluice is available in Saudi Arabic

Based on the results of experiments 1–2, we hypothesized that OPUS specifically with merger type sluicing should be acceptable whenever an acceptable synonymous non-elliptical structure (a cleft or a sentence with resumption) is available as a pre-sluice (see Barros et al. 2014; Abels 2017a).²⁶ To test this hypothesis we set out to compare examples where clefts and/or wh-resumption are possible with examples where neither is possible. Recall from section 2 that Arabic when can only be used in the wh-movement strategy and is incompatible with both the cleft and resumption strategies. (The same is true of how and why, but since these do not occur as the complements of prepositions, we do not consider them further.) Where, on the other hand, is compatible with all three question formation strategies. (These claims from section 2 are verified by the non-elliptical conditions in this experiment.) On a structural approach relying on a well-formed structure at the ellipsis site, only where should give rise to fully acceptable cases of OPUS. Examples with when should be less acceptable due to the lack of a well-formed source. That is, if the acceptability of OPUS in the previous experiments is attributed to the availably of a possible source in the e-site, we expect a significant effect of the type of wh-remnant such that where as OPUS remnant should be fully acceptable and when should be unacceptable. Under Nykiel’s approach, OPUS should be equally (un)acceptable with where and with when as remnants; this is predicted because all factors that enter her model are held constant: all examples involve sluicing (construction type), have a PP in the antecedent (structural persistence is violated throughout), involve contentful correlates (informativity of the correlate), and involve PPs that have no semantic dependency with the verb. Recall that under Nykiel’s (2015) criteria, a verb depends on the preposition if the sentence without the PP does not entail the sentence with the PP. In the experimental stimuli, all verbs are independent of the PP by this criterion. Moreover, prepositions are said to depend on the verb, if the predicate cannot be replaced by a bland do or happen type predicate without changing the meaning of the preposition. Again, the stimuli have independent PPs by this criterion. Thus, Nykiel’s model predicts the absence of an effect of wh-type. The results confirm the predictions of the structural approach. Surprisingly though, we found the examples with when to be more acceptable than their hypothesized pre-sluice, a point to which we return in the discussion.

We conducted a web-based acceptability judgment experiment. Items were constructed crossing three factors: Left-peripheral P (absent vs. present), ellipsis (sluicing vs. non-elliptical structure),²⁷ and wh-type (where vs. when). For ease of reference, Table 10 summarizes mentioned theories and their predictions.

5.2 Results

Our research questions asked whether the acceptability of OPUS is driven by the availability/lack of an alternative source at the e-site. Table 12 shows the average rating provided by participants for each of the eight conditions. In general, lack of a left-peripheral preposition (both in elliptical and non-elliptical conditions) was significantly more acceptable with where than with when. Recall that in condition 5 and condition 6 ‘left-peripheral P absent, non-elliptical’, the non-elliptical structures involved wh-clefts and resumptive structures in equal numbers. As we expected, both wh-clefts and wh-resumption are equally compatible with where (means of M = 5.09 and M = 5.1, respectively). Similarly, both wh-clefts and wh-resumption are equally incompatible with when (means of M = 2.8 and M = 2.2, respectively). Thus, we report the results for clefts and resumption together in what follows.

Table 12

Experiment 3: Mean acceptability ratings by Condition (n = 74).

	Condition	Mean	SD
Condition 1	Left-peripheral P absent, sluicing, where	6.1471	1.55142
Condition 2	Left-peripheral P absent, sluicing, when	4.6230	2.23204
Condition 3	Left-peripheral P present, sluicing, where	6.2235	1.49726
Condition 4	Left-peripheral P present, sluicing, when	5.8081	1.79409
Condition 5	Left-peripheral P absent, non-elliptical, where	5.1318	2.08058
Condition 6	Left-peripheral P absent, non-elliptical, when	2.5339	1.97030
Condition 7	Left-peripheral P present, non-elliptical, where	5.5271	1.94142
Condition 8	Left-peripheral P present, non-elliptical, when	4.7201	2.21946

As illustrated in Figure 5, the conditions lacking a left-peripheral preposition are significantly more acceptable with where both under sluicing (M = 6.1) and in non-elliptical structures (M = 5.1) compared to their counterparts with when (M = 4.6 for sluicing and M = 2.5 for the non-elliptical condition). As for the conditions with left-peripheral prepositions, both where and when were rated with comparable acceptability scores in sluicing conditions (M = 6.2) and (M = 5.8) and in non-elliptical conditions (M = 5.5) and (M = 4.7).

Figure 5

Experiment 3: Mean acceptability ratings by Condition (n = 74). Error bars represent the standard error of the mean (SEM).

As in the previous experiments, raw ratings were z-score transformed prior to the analysis. A linear mixed-effects model was fitted with left-peripheral P, ellipsis, and wh-type as fixed factors and subject and item as random effects with random slopes and intercept. As indicated in Table 13, no significant main effect of any of the three factors was found. However, the results revealed a highly significant interaction between left-peripheral P and the wh-type (t = –4.483, p = 1.07e-05). This reflects the fact that elliptical and non-elliptical conditions without left-peripheral prepositions are much more acceptable with where than with when. The results also revealed a significant three-way interaction (t = 2.758, p = 0.00646). This is due to the improvement of the ratings specifically for when in the OPUS condition. No comparable effect is present for where, for which both elliptical conditions are highly acceptable (M = 6.1 and M = 6.2, respectively). Indeed, these two values for where are not significantly different from each other.

Table 13

Summary of linear mixed effects models; P-values estimated using the Satterthwaite approximation (***p < .001).

	t-value	Pr(>|t|)
(Intercept)	3.929	9.86e-05***
Left-peripheral P	–1.426	0.15471
Ellipsis	0.643	0.52052
wh-type	–1.647	0.10084
Left-peripheral P * ellipsis	–0.922	0.35769
Left-peripheral P * wh-type	1.419	0.15700
Ellipsis * wh-type	–4.483	1.07e-05***
Left-peripheral P * ellipsis * wh-type	2.758	0.00646**

• Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

6 Conclusion

To summarize, we conducted three acceptability judgment studies on OPUS in Saudi Arabic. They were driven by two main ideas: First, the ellipsis site contains silent syntactic structure. Second, the syntactic structure at the ellipsis site, the pre-sluice, may but need not contain a resumptive pronoun rather than a gap. The results overall support these ideas.

We showed that OPUS is acceptable in Saudi Arabic (all three experiments). Thus, Saudi Arabic behaves in agreement with what has been reported for other Arabic varieties but in conflict with Merchant’s (2001) P-stranding generalization. We have also shown that OPUS is acceptable not only in merger type sluicing (experiments 2 and 3) but also in contrast sluicing (experiments 1 and 2), which is problematic for approaches like Algryani (2012); Leung (2014b); Albukhari (2016), where clefts are assumed to be the sole pre-sluice giving rise to OPUS in Arabic. Experiment 2 probed whether the use of a resumptive pronoun in non-elliptical conditions leads to a degradation in the judgments, which should, under our hypothesis, correlate with a similar degradation in contrastive OPUS structures, where, it will be recalled, resumptive pronouns are forced. However, we did not find a degradation coming from the use of resumptive pronouns – either in elliptical or non-elliptical structures. Experiment 3 showed that the judgments on OPUS depend on the acceptability of a grammatical pre-sluice in the ellipsis site. This was in line with structural expectations.

Together, we take these three experiments as strongly supporting an approach to sluicing that assumes a silent pre-sluice at the ellipsis site and which allows questions with resumptives as pre-sluices. We do not see how to square this finding with a non-structural approach for sluicing.

The results are moreover problematic for Nykiel’s model, because the effects expected under her account did not reliably influence the results in our experiments. While the first experiment did show an effect of the informativity of the correlate, this size of the effect was too small to exclude a false positive. Indeed, the effect failed to be reproduced in the follow up study mentioned in footnote 18. Furthermore, the effect of structural persistence predicted by Nykiel’s model in experiment 2 failed to materialize. Finally, in experiment 3, where Nykiel’s model predicts a null result, we found a robust effect of wh-type.

While experiment 3 showed a degradation in OPUS conditions with when compared to OPUS conditions with where, there was an additional interaction with ellipsis here: the elliptical condition, while degraded, is substantially improved compared to its non-elliptical counterpart. How does Saudi Arabic fit into the cross-linguistics picture? The results of experiment 3 are reminiscent of the situation reported in Molimpakis (2019) for Greek. Recall that in Molimpakis’s experiment there were no plausible acceptable pre-sluices for the OPUS conditions. While Molimpakis found a degradation, she also found the strong ameliorating influence of ellipsis. Similarly, Nykiel (2013) argues that OPUS in Polish (contra Szczegielniak 2008) lacks a grammatical pre-sluice. And again examples of OPUS are judged worse than sluices with PP remnants but at the same time strikingly more acceptable than overt P-stranding. This is the behaviour of OPUS with when in our experiment 3. On the other hand, we found that when there is a grammatical pre-sluice (experiments 1, 2, and where in experiment 3), there is no penalty from the lack of a left-peripheral preposition at all. Studies involving OPUS where a grammatical pre-sluice is available have not been conducted for Greek and Polish, so we do not know how such examples would behave. The findings from Saudi Arabic suggest that the OPUS penalty in Molimpakis’s and Nykiel’s experiments should be interpreted as the signal of the lack of a grammatical pre-sluice, that is, as a signal that the sentences are ungrammatical. The improvement compared to the non-elliptical versions, under this view, must be attributed to other, possibly extra-grammatical factors. We must leave the question of what those factors might be unanswered.

The experiments have shown that resumption is compatible in Saudi Arabic non-elliptical questions with all nominal wh-phrases and with where but not with when. This finding contrasts with what has been reported for other Arabic dialects (Aoun et al. 2010), where resumptive pronouns have a substantially more limited distribution. Experiment 2 further showed that resumption in cases where it is optional does not have a negative effect on acceptability. It seems to us that this finding should be of some interest for debates on resumption (Rizzi 1990; Shlonsky 1992; Aoun 2000; Aoun et al. 2001).

Given the results on OPUS in Saudi Arabic, it may be worth revisiting the analysis of OPUS in other Arabic dialects. As mentioned above, the literature has never brought contrastive sluices to bear on the question of the possible pre-sluice and has, in some cases, conflated the cleft and the resumptive strategy for question formation. We hope that our focus on contrast sluices may provide an impulse for comparable work on other Arabic dialects. We expect that existing variation in the distribution of resumptive pronouns across Arabic dialects should be reflected in the acceptability patterns of OPUS, which would further strengthen the core thesis of this paper.

Overall, we hope that the current results can help overcome the current theoretical impasse in the literature on OPUS. Our results suggest that there is structure at the ellipsis site but that it need not be isomorphic to the antecedent. Generalizing the results from Saudi Arabic, we would expect that for other languages and structures we should find the following type of pattern: OPUS should be fully acceptable just in case a well-formed and semantically appropriate pre-sluice is available. OPUS should be degraded when no well-formed and semantically appropriate pre-sluice is available. In the case where no pre-sluice is available, the extent and cause of any amelioration under ellipsis needs to be further studied.

Abbreviation

1 = first person; 2 = second person; 3 = third person; F = Feminine; M= Masculine; SG= singular; PL= plural; NEG = negation; OPUS= Ostensible P-stranding Under Sluicing or Omission of Preposition Under Sluicing.

Competing Interests

The authors have no competing interests to declare.