2.1 Syntactic analyses of P-omission

The possibility of using NPs or PPs as fragments and sluices has attracted considerable attention due to its potential to reveal something about the structure of these apparently nonsentential utterances. Merchant (2001: 92) proposes the Preposition-Stranding Generalization (PSG), given in (13), for sluicing (Merchant 2004 extends the PSG to fragments).

We should clarify at this point that ‘P-stranding’ should be understood as P-omission, or using NPs instead of PPs. The PSG predicts that this is a severely restricted option, as restricted as preposition stranding is, appearing only in Germanic. If the PSG is accurate we have support for a theoretical possibility that sluices are constituents in sentential structures that have survived deletion, and should exhibit all the relevant properties such constituents would have. For instance, the properties of the sluice in (11) should be equivalent to those of its counterpart in the sentential structure in (14), in which preposition stranding has applied.

Assuming that NP sluices are derived from sentential sources via preposition-stranding movement, then they should appear only in those languages that permit such movement. This possibility is defended in work following Merchant (2001; 2004), with current debate centered around the question of whether sluices and fragments survive deletion after leftward movement (e.g. Griffiths & Lipták 2014; Weir 2014; Shen 2018) or without it (e.g. Abe 2015; Griffiths 2019; Stigliano 2022). The latter kind of analysis, assuming deletion but not movement, has difficulty predicting why or when NPs can replace PPs under clausal ellipsis (but see Griffiths 2019 for a proposal), while the former kind doesn’t so long as NPs closely track the patterns we see in preposition-stranding movement, as per the PSG.

But the PSG’s predictions are far from perfect. The list of languages in which NPs alternate with PPs as fragments although no alternation should be permitted is long and growing: Bahasa Indonesia (Fortin 2007), Spanish and Brazilian Portuguese (Almeida & Yoshida 2007; Rodrigues et al. 2009; Stigliano 2022), Serbo-Croatian (Stjepanović 2008; 2012), Czech (Caha 2011), Polish (Szczegielniak 2008; Sag & Nykiel 2011; Nykiel 2013; 2018; Nykiel & Kim 2022b), Russian (Philippova 2014), Emirati Arabic (Leung 2014), Bulgarian (Abels 2017), Greek (Molimpakis 2019), Saudi Arabic (Alshaalan & Abels 2020), Mauritian (Abeillé & Hassamal 2019), French (Hassen & Abeillé 2025). On the one hand, these data have helped garner support for a competing nontransformational approach to ellipsis that sees fragments as genuinely nonsentential utterances, with no parallels to be expected between them and corresponding sentential constituents (Ginzburg & Sag 2000; Culicover & Jackendoff 2005; Sag & Nykiel 2011; Kim 2015; Abeillé & Hassamal 2019; Abeillé & Kim 2022; Nykiel & Kim 2022a).³ The nontransformational approach, unlike the deletion-based one, leaves no reason to think that NPs are ungrammatical while PPs are grammatical in any language. On the other hand, these data have also led researchers to look for explanations for why NPs should alternate with PPs under clausal ellipsis. These explanations range from syntactic ones whose aim is to defend the PSG along with the transformational approaches to ellipsis (e.g. Rodrigues et al. 2009; Abels 2017; Molimpakis 2019; Alshaalan & Abels 2020) to processing-based ones (Nykiel 2013; 2017; Nykiel & Hawkins 2020; Nykiel & Kim 2022b; Hassen & Abeillé 2025) whose aim is rather to advance our understanding of why NPs appear at all.

The focus has recently been on experimental and corpus data, which allows us to probe both the acceptability and frequency of P-omission across languages. The current findings are that PPs are generally better than NPs as fragments in the great majority of languages (Nykiel 2013; Merchant et al. 2013; Molimpakis 2019; Alshaalan & Abels 2020; Lemke 2021; Nykiel & Kim 2022b; Hassen & Abeillé 2025)⁴ and have higher frequencies in corpus data (Nykiel 2015). An exception to these patterns is present-day English, favoring P-omission in both corpus data (Nykiel 2017; Nykiel & Hawkins 2020) and experimental data (Lemke 2021) and French, which, however, favors P-omission only in corpus data (Hassen & Abeillé 2025). There thus appears to be crosslinguistic pressure to use PPs rather than NPs. It is observed even in a preposition-stranding language like English as a pattern weakening over time in response to a growing number of multi-word verbs (see Nykiel 2014; Nykiel & Hawkins 2020). But why should such pressure exist? Two scenarios are plausible. Given the nontransformational approach to fragments, both PPs and NPs are grammatical, but there is some reason to prefer PPs over NPs in some environments. That reason could be that PPs are generally easier to process, possibly by virtue of the amount of linguistic information they encode (this is proposed in Nykiel & Hawkins 2020 and discussed in detail in section 2.2). Given the transformational approach to fragments, we would assume that NPs are grammatical in some languages (or perhaps in some environments within a single language) and ungrammatical in others, with their degraded acceptability reflecting the pressure to avoid them in exactly those languages that fail to license them. It’s clear, however, that it’s not the PSG that can capture the crosslinguistic data, because the acceptability of NPs consistently falls above that of ungrammatical controls in experimental studies that have made such comparisons (Nykiel 2013; Molimpakis 2019; Alshaalan & Abels 2020; Nykiel & Kim 2022b; Hassen & Abeillé 2025). One could explore whether P-omission is a candidate for constituting yet another instance of acceptable ungrammaticality among those discussed in the literature (e.g. Wagers et al. 2009; Tanner et al. 2014; Parker & Phillips 2016; 2017), which includes an ongoing debate surrounding the acceptability and grammaticality of structural mismatches under a different elliptical construction, Verb Phrase ellipsis (VPE) (Arregui et al. 2006; Kim et al. 2011; Kim & Runner 2018; Parker 2018). In sum, all syntactic accounts face the task of explaining a disconnect between the acceptability and frequency of NPs vs. PPs and our theoretical assumptions about their grammaticality. Before we can determine how P-omission can be accommodated on existing formal theories of ellipsis, we need a better understanding of why and when exactly it is available at all. Whatever account of P-omission we articulate must also provide a motivation for the ban on P-omission under sprouting we saw in (12), as this ban appears to be the key to answering both the why and the when questions.

Only the transformational literature has offered syntactic explanations for the ban on P-omission under sprouting. The first is Chung’s (2006) proposal that deletion is subject to a constraint that prevents it from targeting lexical items that are absent from the antecedent. This has the effect that a preposition may never be deleted without an antecedent hosting an overt PP headed by the same preposition. Another explanation is a follow-up on this constraint but additionally incorporates some syntactic identity between the structure of a sluice and its antecedent. Chung (2013) proposes a requirement that the lexical head licensing the features of an NP sluice be the same as the lexical head licensing the features of its correlate (an argument structure identity constraint is also placed on the relevant predicates in the antecedent and the elided clause). This again has the desired effect of ruling out NP sluices that would be licensed by prepositional heads that are missing from the structure of the antecedents. Building on Rudin (2019), Anand et al. (2023) also defend the idea that argument-structure identity must obtain between the structures of a sluice and its antecedent, but they define identity over constituents smaller than entire antecedent clauses. For them, the relevant unit is what they call an ‘argument domain’, the largest phrase with an event-based semantics located within an extended projection and containing a lexical head and all of its arguments. Identity is calculated over the members of an argument domain located within the deleted material (out of which a sluice moves) and the members of a corresponding argument domain within the sluice’s antecedent. To see how this analysis handles sprouting and P-omission, consider the underlying structure for the ill-formed NP sluice in (12), repeated here as (15).

In deriving the sluice, we leave the preposition of behind at the ellipsis site, with the result that the argument domain it belongs to fails to fully match any argument domain in the antecedent because implicit phrases are not represented at all in syntactic structure on Anand et al.’s analysis. If, however, we pied-pipe the preposition with its complement what, forming a PP sluice, we eliminate the problematic mismatch at the ellipsis site. We would also not run into any mismatch-related difficulties if the antecedent contained an overt correlate for the NP sluice, since then a matching preposition would be present in the antecedent.

However, Anand et al.’s proposal faces two problems. First, their analysis, like Merchant’s (2001), relies on preposition-stranding movement to derive NP sluices, a move that compromises its accuracy as soon as NP sluices appear in languages without preposition stranding. Second, transformational accounts of P-omission have a tradition of postulating several kinds of underlying structure for NP sluices to avoid preposition-stranding violations. Besides the structure identical to its antecedent, a sluice may be embedded in a copular clause or a short cleft, neither of which involves illicit preposition stranding (e.g. Rodrigues et al. 2009; Van Craenenbroeck 2010; Vicente 2018). These structures are depicted in (16)–(17) for the NP sluice with an overt correlate in (11).

There is an obvious tension between Anand et al.’s argument-structure identity requirements and permitting copular clauses and short clefts, which can never satisfy such requirements, as sources for NP sluices. Thus, the transformational work on sprouting hasn’t brought us closer to formulating a satisfactory syntactic explanation for P-omission, even though it has highlighted a general need for argument structure-based parallelism between fragments and their antecedents whenever fragments are linked to correlates that are arguments to some lexical heads.⁵

2.2 Processing-based analysis of P-omission

The limited success of purely syntactic analyses of P-omission has led to a processing-based strand of research due to Nykiel & Hawkins (2020). The basic idea behind their proposal is that P-omission is grammatical (this is currently fully in line with the nontransformational approach to ellipsis and partially so with the transformational one) and that it serves as a way of shortening fragments in easy-to-process environments by removing contextually given information in the form of prepositions, much the same way that NP fragments can be shortened by avoiding repetition of their head nouns, as in (18).

This idea is tied to our current understanding of how fragments are processed online. They encode semantic and morphosyntactic cues for the parser to use in the process of locating their correlates, after which they can be integrated into the propositions expressed by their antecedents (Culicover & Jackendoff 2005; 2012; Culicover 2016; Goldberg & Perek 2019; Harris & Carlson 2019; Nykiel et al. 2023). The initial search for the correlate is executed by a direct-access mechanism, that is, by comparing a fragment’s properties against all extant memory representations simultaneously until the target is successfully identified, as Martin & McElree (2011) have proposed. The general process of resolving dependencies between current and previously encountered memory representations by engaging a direct-access mechanism is central to cue-based theories of sentence processing (Caplan & Waters 2013; Lewis & Vasishth 2005; Lewis et al. 2006; McElree 2000; McElree et al. 2003; Van Dyke 2011; Van Dyke & Johns 2012). These theories are cue-based in the sense that they don’t assume any serial search through existing memory representations and that successful retrieval relies on the information supplied by the constituent that initiates the retrieval (a retrieval probe). The retrieval probe carries information whose diagnosticity is defined in relative terms. If the retrieval probe is completely or partially similar only to the target, and there are no distractors with similar features, then the target is distinct and more easily accessible during retrieval. But if the retrieval probe, the target, and distractors all share some features, the cue diagnosticity of the retrieval probe is reduced, leading to potential cue overload, that is, failure to retrieve the target (Watkins & Watkins 1975; Nairne 2002; Öztekin & McElree 2007; Martin 2018). While we still don’t know with complete certainty what linguistic information is used as retrieval cues, there is evidence that it’s semantic, syntactic, and morphological (Parker et al. 2017). With this background in place, we can conclude that a fragment’s properties affect the success of the search for its correlate, but Nykiel & Hawkins (2020) argue that it is the properties of both the fragment and correlate that matter. Before proceeding, we should briefly look at the processing of expressions with different linguistic content.

Research on memory retrieval has documented a difference between the processing of lexical NPs as semantically and syntactically contentful expressions and the processing of less contenful expressions like pronouns. This difference is due to lexical NPs receiving mental representations that are more readily available for future retrieval (Craik & Lockhart 1972; Gallo et al. 2008; Hofmeister 2007; 2008; 2011; Hofmeister et al. 2007; Hofmeister et al. 2013; Karimi et al. 2014). In other words, the more contentful an expression is the more of an accessibility boost it receives on being processed. Differences in the accessibility of expressions are also connected to the manner in which we tend to refer to them. Highly accessible expressions are picked up by anaphors with less content (e.g. pronouns) and low-accessibility expressions are preferably picked up by anaphors with more content (e.g. lexical NPs) (Ariel 1990; Gundel et al. 1993; Karimi et al. 2014; Troyer et al. 2016; Nykiel et al. 2023). This is so regardless of whether we link accessibility to future retrieval from memory, as we do here, or to current availability for retrieval, as do Ariel (1990) and Gundel et al. (1993). These insights will be helpful in our discussion of how correlates impact the amount of linguistic information that fragments encode below and later in section 4.1.

Nykiel & Hawkins’ (2020) purpose is to account for both why and when P-omission is available. They take as their starting point the sprouting pattern we have already seen. A fragment linked to an implicit argument must satisfy the requirements imposed on that argument by the lexical head that selects it. This is possible because implicit arguments, regardless of whether they are taken to be actually present in syntactic structure,⁶ are activated in online processing as part of the full lexical entry for a licensing head and are available as correlates (see Culicover & Jackendoff 2005: 260–262 for extensive justification of this position). P-omission isn’t permitted if an implicit argument is a PP. If it was permitted, it would lead to the fragment being a smaller constituent than the required PP. But we can expand the fragment beyond a PP. These options are depicted in (19) (see also Weir 2017 for discussion of such expanded fragments).

Nykiel & Hawkins (2020) argue that, given the fragment’s task to efficiently point to a PP argument here, it is a processing advantage for it to encode the same syntactic category, as is including even more information by repeating antecedent material beyond the PP (the advantage of including redundant material in sluices has previously been documented by Harris 2015). Meanwhile, P-omission introduces potentially misleading cues: a different syntactic category than the PP we’re looking for and no repeated antecedent material whatsoever. The question this raises is why an overt PP argument in the antecedent should come with the options of both shortening a fragment to an NP and expanding it to a larger phrase than a PP, as shown in (20), without compromising the reliability of the fragment’s retrieval cues.

The answer lies in the accessibility of the PP argument being pointed to. Whenever it’s an implicit phrase, it lacks form and content, which renders its mental representation minimally accessible for future retrieval, according to the definition of accessibility we adopt here. That being so, the implicit PP is difficult to access and the burden is on the fragment to supply cues that are sufficient for linking it to that PP. We would expect expanded fragments to appear more frequently under sprouting than elsewhere, since the redundant information they provide can ease the processing difficulty in this context. If the fragment’s task is instead to point to an overt PP, which can be thought of as more accessible and, therefore, easier to access than any implicit PP on account of its content, fewer retrieval cues will be required of the fragment.

That possible changes to the form of fragments are a response to the context in which we find them follows from the principle of Minimize Forms (MiF) (Hawkins 2004: 31), given in (21).

Hawkins (2004; 2014) invokes MiF to explain various kinds of principled reduction, including some kinds of ellipsis. Nykiel & Hawkins (2020) invoke MiF as a principle dictating how the form of fragments rests on the ease with which their correlates can be accessed. For P-omission, MiF predicts that fragments may optionally carry less content by dropping prepositions if they are easy to pair with their correlates. But if correlates already have low levels of accessibility, P-omission removes redundant, and potentially helpful, material, impoverishing the set of retrieval cues that fragments encode, and hence is dispreferred. The fact that the processing of sluices with implicit correlates is indeed more demanding than the processing of sluices with overt correlates has been independently supported through experimental studies (Frazier & Clifton 1998; Dickey & Bunger 2011). That being so, the actual ban on P-omission under sprouting can be seen as the limiting case of MiF, with the grammar having stepped in to rule out a strongly dispreferred structure and thereby to ensure that fragments remain interpretable. In other words, we are assuming, following Hawkins (2004; 2014), that preferences governing syntactic variation interact with the grammar by being able to give rise to grammatical constraints that enforce structures that are strongly preferred or ban structures that are strongly dispreferred (see Hawkins 2014: 78–85 for more examples and discussion of how preferred structures can conventionalize into grammatical principles as dispreferred ones turn into grammatical violations). Recognizing the possibility that the grammar has responded to processing pressures, Nykiel & Kim (2022a: 92) follow up by formulating a grammatical constraint they term No Form Minimization, as in (22), to block P-omission under sprouting.

The processing-based motivation behind (22) is made clear by the explicit reference to MiF.

Beyond sprouting, Nykiel & Hawkins (2020) find support for the predictions of MiF in three kinds of context for fragments, of which I only discuss the one that is also relevant for pseudogapping (but see Nykiel & Hawkins 2020: 422–423, 432 for the entire set of contexts). There is a split between indefinite pronouns and lexical NPs when they serve as prepositional objects within PP correlates. Examples (23)–(24) from Nykiel & Hawkins (2020: 422) illustrate the two kinds of prepositional objects for sluices.

In Nykiel & Hawkins’ corpus data, P-omission is more frequent in sluices like (24) than in sluices like (23), because, they argue, a lexical NP (an old friend in (24)) is more contentful and accessible than an indefinite pronoun (someone in (23)). Crosslinguistic work reports the same pattern of preference in introspective judgments (see Caha 2011 for Czech, Rodrigues et al. 2009 for Brazilian Portuguese, Spanish, and French, and Stjepanović 2008 for Serbo-Croatian) and in experimental studies of fragments (see Sag & Nykiel 2011, Nykiel 2013 and Nykiel & Kim 2022b for Polish, and Hassen & Abeillé 2025 for French).

With MiF in place, Nykiel & Hawkins (2020) can account for the availability of P-omission in the languages in which it’s been reported, along with its more limited distribution compared to the distribution of corresponding PP fragments. But there are two factors that boost the frequency of P-omission if a language satisfies certain conditions, as does present-day English. These factors are numerous word-external dependencies between prepositions and other lexical categories and structural parallelism. I discuss them in turn.

Word-external dependencies between prepositions and lexical categories like verbs, nouns, and adjectives are what Nykiel & Hawkins (2020) single out as the cause for the crosslinguistically unusual behavior of present-day English, that is, its failure to favor PPs over NPs as fragments. Word-external dependencies can vary in number (the total number of dependent combinations a language possesses) and strength (whether only the preposition in a combination depends on the other category or whether both depend on each other), and, as Nykiel & Hawkins (2020) argue, they impact the human parser’s ability to successfully assign all the relevant properties to a PP fragment when it’s first encountered and before the process of searching for its correlate begins. Example (25) from Nykiel & Hawkins (2020: 424) (their original NP fragment has been turned into a PP here) illustrates the processing difficulty associated with a preposition that can’t be interpreted within the bounds of the PP in which it’s embedded and which constitutes the fragment, requiring simultaneous access to a verb located in the antecedent VP.

This is because the preposition on and the verb depend are involved in a two-way dependency, i.e., they depend on each other for their interpretation. Including the preposition on in the fragment here forces the parser to reach into the antecedent VP in order to retrieve the preposition’s semantics. This differs from the scenario where a preposition is semantically independent and can be assigned all of its semantic and syntactic properties within a fragment before the resolution proceeds any further, as shown by the A₂-fragment in (26) from Nykiel & Hawkins (2020: 420).

In Nykiel & Hawkins’ (2020: 419) terminology, the distance over which the constituents of a fragment are assigned their relevant properties is a lexical domain. That domain is coterminous with a fragment for all semantically independent prepositions, but stretches beyond the fragment in (25). With numerous verb-preposition dependencies in English (both Nykiel & Hawkins 2020 and Hawkins 2000 put the rate of semantically dependent combinations in corpus data at approximately 50%), Nykiel & Hawkins (2020) find that English fragments don’t contain those prepositions that would disrupt independent processability of PPs. This finding indicates that it’s the larger VP domain that pushes in favor of P-omission in English fragments, in effect working against the PP domain, which normally ensures that all constituents of a PP are adjacent to each other, making their locally determined properties easily processable within that PP. The VP and PP domains compete with each other, as Nykiel & Hawkins (2020) argue, and the reason that the VP domain wins in English fragments is the high number of semantically dependent verb-preposition combinations. If it wasn’t for that, a bias against P-omission would be observed, as it has been in several of the world’s languages, because then the PP domain would exert stronger influence over fragments than the VP domain (in addition to the effects of MiF).

Distance effects, whether associated with the VP or the PP domain, are in line with the principle of Minimize Domains (MiD) first defined in Hawkins (2004: 31) and given in (27).

MiD predicts that linguistic forms are preferably assigned all their semantic and syntactic properties over as minimal distances as possible, but we are specifically interested in semantic properties here. The primary area of application for MiD have been word-order preferences in contexts where alternative orderings of constituents are available. If combinations of semantically dependent constituents are involved, those orderings are favored that allow fastest recognition of all relevant semantically dependent constituents (Hawkins 2000; Lohse et al. 2004). As noted above, the distance over which it’s possible to compute such word-external dependencies between constituents is a lexical domain, originally defined in Hawkins (2004: 117). To illustrate, compare the VPs hosting the particle verb look up in (28) and (29) (see Lohse et al. 2004: 246).

Notice first that the particle’s interpretation is not literal but word-externally dependent on the verb, and, therefore, the parser requires simultaneous access to the verb to interpret the particle. The lexical domain required for interpreting the verb and the particle is seven words in (28), as the parser needs to process the intervening NP before reaching the sentence-final particle. In contrast, the parser requires only two words to compute the meaning of the verb and particle in (29). Thus the ordering in (29) shortens the lexical domain needed for computing the semantically dependent verb-particle combination by five words, being favored over (28). Nykiel & Hawkins’ (2020) results demonstrate that the application of MiD extends to fragments, and that the VP domain dominates over the PP domain there.

The second factor that boosts P-omission is structural parallelism (or structural priming), a general mechanism promoting reuse of previously encountered structure. Speakers tend to repeat structures they have themselves produced earlier (Bock 1986) or heard others producing (Bock et al. 2007), including as a form of linguistic co-operation between speakers in dialogue (Branigan et al. 2000). Structural parallelism is known to operate over a range of elliptical constructions, such as gapping (Carlson 2002; Kim et al. 2020; Overfelt 2024; Orth & Yoshida 2025), pseudogapping (Poppels & Miller 2023), sluicing and fragments (Frazier & Clifton 1998; Dickey & Bunger 2011; Nykiel 2017; Nykiel et al. 2023), VPE (Kim & Runner 2018; Parker 2018), and much less ellipsis (Carlson & Harris 2018; Harris & Carlson 2019). All of these constructions are consistent in favoring matching (morpho)syntax in antecedents and ellipsis sites, even though they vary in the kind of mismatch they tolerate and in the ease with which they tolerate mismatch. Gapping, for instance, appears to be the least able to tolerate any mismatch (e.g. Overfelt 2024; Orth & Yoshida 2025), whereas VPE is quite flexible (e.g. Kim & Runner 2018). Structural parallelism effects have also been observed outside of ellipsis, often in conjunction with coordination (Frazier et al. 1984; Frazier et al. 2000; Sturt et al. 2010; Poirier et al. 2012), but not necessarily so (Dubey et al. 2005; Callahan et al. 2010).

The structural parallelism effects we are concerned with here are induced by overt correlates’ ability to be continuous or discontinuous PPs. A continuous PP correlate (of what), with both constituents adjacent, is shown in (30) from Nykiel & Hawkins (2020: 431).

A discontinuous one is depicted in (31), also from Nykiel & Hawkins (2020: 431). Here the preposition is stranded.

Nykiel & Hawkins (2020) report a much stronger preference for P-omission for discontinuous correlates like (31) than for continuous correlates like (30), connecting it to speakers reusing the same structure in fragments as they have encountered in the correlates (see also Nykiel 2017 for more discussion of structural parallelism in English fragments). It’s important to note that structural parallelism promotes P-omission here because English permits preposition stranding in interrogative clauses, making available not just continuous but also discontinuous PPs as correlates. We could only expect to see structural parallelism promote P-omission in a language without preposition stranding if P-omission has reduced fragments from PPs to NPs, causing subsequent fragments to surface as NPs, as well.

I turn now to predictions that MiF, MiD and structural parallelism make for P-omission under pseudogapping.

3.1 Transformational analyses

Transformational analyses posit movement that evacuates a remnant out of a VP followed by deletion of all remaining material inside that VP, with deletion being subject to syntactic identity with an antecedent (Jayaseelan 1990; Lasnik 1999; Gengel 2013; Thoms 2016). It has been debated what kind of movement this should be, the theoretical possibilities for English ranging from heavy NP shift (Jayaseelan 1990) to A-movement (Lasnik 1999) to A’-movement, specifically, Focus Movement (Gengel 2013; Thoms 2016). Among other data, examples like those in (37)–(40) have served as support for the various kinds of movement proposed over the years. We will look more closely at Focus Movement here since it’s the most recent and least controversial of the three. Consider example (43) and a simplified tree representation for the right conjunct given in (44) from Thoms (2016: 291).

The remnant beer moves out of the VP and into Spec,ΣP, a focus-related projection, leaving behind the rest of the VP, which subsequently deletes. Should it be the case that we have an example like (1), repeated here as (45), we can move either the entire PP remnant of the private sector or just its complement (via preposition stranding) out of the VP. That being so, Gengel (2013: 73) appears to think that NP remnants are ungrammatical, although her Focus-Movement analysis can derive them alongside PP remnants.

This derivation, of course, depends on whether preposition stranding is permitted in a language, and it additionally predicts that there should be no misalignments between remnants and preposition placement under Focus Movement. That is, we wouldn’t expect to see NP remnants where preposition stranding is degraded in nonelliptical clauses and conversely, we wouldn’t expect to see PP remnants where preposition pied-piping is degraded. Exploring whether such misalignments exist requires careful, ideally, experimental testing, not least because problematic data have been reported for fragments (see Levin 1982, Fortin 2007, and Nykiel & Hawkins 2020 for more discussion). Furthermore, the hypothesis that preposition-stranding movement underlies P-omission under pseudogapping should be tested on crosslinguistic data, and in order to do so, we would have to work with languages with pseudogapping and P-omission but without preposition stranding. No such research exists yet.

To the best of my knowledge, it hasn’t been addressed within the transformational, or any other, approach to pseudogapping that correlates may be either implicit or overt, although the first examples of implicit correlates are reported in Miller (2014: 81).⁸ Nor has it been pointed out that the overt/implicit distinction has consequences for P-omission. Given the theoretical apparatus we have seen so far, it’s unclear why implicit PP correlates should never be accompanied by NP remnants (I return to and elaborate on this point in section 4). Consider the impossibility of dropping the preposition from (46) (in example (2), which we saw before, P-omission is equally impossible).

Given that pseudogapping has a sentential source and movement operations may apply to PPs and NPs alike in English, there is no mechanism that blocks the derivation of NP remnants just in sprouting cases, i.e. when their correlates are implicit PPs. However, we could imagine that some version of the constraints proposed by Chung (2006; 2013) and Anand et al. (2023) for sluices (recall the discussion in section 2.1) might be appropriate for pseudogapping, as well.

3.2 Nontransformational analyses

A key idea shared among all nontransformational analyses of pseudogapping is that no material inside a VP undergoes deletion. There is instead an empty element which is defined on some approaches as a pro expression serving as complement to an auxiliary verb, closely tracking analyses of VPE (Ginzburg & Miller 2018; Nykiel & Kim 2021; Kim & Runner 2022), or as an operator with a function analogous to a transitive verb (Kubota & Levine 2017). On the earliest analysis of pseudogapping due to Miller (1990) an auxiliary is itself a verbal proform which combines with a remnant and picks up its interpretation from a salient predicate present in the surrounding context. The signature pattern of this type of analysis is that remnants fail to behave like full-clause constituents that undergo movement and they violate syntactic identity conditions (for recent overviews of the evidence against transformational analyses of pseudogapping, see Miller 2014, Kim & Runner 2022, and Poppels & Miller 2023). We will see examples of syntactic-identity violations in (47)–(49) below and in fn. 9.

The sprouting example in (46) shows that pseudogapping remnants, like fragments, are impacted by the morphosyntactic specifications of their correlates. There is in fact more evidence than just sprouting for this being the case (e.g. Miller 2014; Kubota & Levine 2017; Poppels & Miller 2023). For instance, Poppels & Miller (2023) demonstrate experimentally that pseudogapping is degraded if remnants realized as PPs mismatch their correlates. The mismatch may be due to using different prepositions in the remnants and correlates (both PPs) when antecedent verbs select two prepositions for their PP complements. Example (47) illustrates matching prepositions and example (48) mismatching ones, both prepositions selected by the verb speak.

The mismatch may also be due to the dative alternation, that is, a PP remnant being paired with an NP correlate, as in (49). Notice that now we have a syntactic category difference between the two.

At the same time, mismatching prepositions or syntactic categories naturally occur in COCA, as demonstrated by Miller (2014), so they can’t be ungrammatical.⁹ All such cases of mismatch raise difficulties for transformational analyses of pseudogapping, as syntactic identity is violated between the trace left behind by a fronted remnant and its correlate (for more discussion, see Poppels & Miller 2023).

A recent analysis due to Kim & Runner (2022) aims to capture these facts by building on the analysis of fragments offered in Ginzburg & Sag (2000) within the Head-Driven Phrase Structure Grammar framework. For fragments, a correlate is supplied as part of the antecedent along with information that is relevant for establishing necessary morphosyntactic identity between it and the fragment. Ginzburg & Sag (2000) impose syntactic category identity on pairs of fragments and correlates (this constraint is abandoned in later work on fragments, e.g. Nykiel & Kim 2022a). Recall that the correlate and remnant are in a contrast relation under pseudogapping unless the correlate is implicit. Therefore, Kim & Runner (2022) introduce a requirement that there be a pair of contrasting elements, the first of which is the correlate and the second the remnant (they don’t include implicit correlates at all). Consider example (50) and a structural representation for the right conjunct in (51) from Kim & Runner (2022: 485).

In this simple case the correlate (magazines) and remnant (books) share the same syntactic category (NP) indicated by the identical values of their CAT features in (51). Syntactic category identity is in fact also imposed on all pairs of correlates and remnants by Kim & Runner (2022), which serves to capture remnants’ sensitivity to their correlates’ syntax exactly as desired. At the same time, Kim & Runner (2022) are forced to relax this requirement enough to allow constrained mismatch between prepositions used by the remnant and correlate in light of Miller’s (2014) corpus data. This move opens the door to abandoning the requirement of morphosyntactic identity between remnants and correlates and incorporating processing principles into theories of pseudogapping so that P-omission can be captured correctly. So long as we require the syntactic categories of fragments and correlates to match we can account for PP remnants, which would match PP arguments subcategorized by lexical heads and serving as correlates, but it’s less clear how we could deal with NP remnants, which would only match the complements of such PP arguments. In fact, this problem extends beyond P-omission and to examples like (49), where a PP remnant is linked to an NP correlate, leaving us with no possibility to match them on morphosyntactic grounds. The only way to match them is by referring to their semantics and to the subcategorization frame of the licensing head (the verb give).

In the next section I explore the hypothesis that P-omission has a processing motivation in pseudogapping (and by extension, across all constructions in which it’s observed). Toward this purpose, I test the predictions of the principles MiF and MiD on English corpus data.

4.1 Data and method

Data collected for this study constitute a subset of the pseudogapping data Philip Miller harvested from COCA and described in Miller (2014). The portion of Miller’s (2014) data that is of relevance to us consists of 1213 items with NP remnants and 293 items with PP remnants. Both sets are potentially nonexhaustive, the former less so as it contains most of the NP remnants available in the corpus at the time the systematic search for them was completed, i.e. in 2014 (Miller 2014; Philip Miller, p.c.). PP remnants were not as systematically searched for, so it remains a possibility that they have even higher frequencies in COCA than those reported here. I searched these items manually to select all antecedents that hosted overt PPs serving as correlates for remnants (both argument and adjunct PPs serving as correlates are included here and in the final dataset). As the final step, I filtered out items where PP remnants contained different prepositions than those present in the correlates. Mismatched prepositions are a possibility if correlates are adjuncts (52) (compare the correlate under Aquarius and the remnant at a natural reef overhang) and if they are arguments (53) (compare the correlate about his schoolboy career ending and the remnant of the season and recall our discussion of Poppels & Miller’s 2023 experimental data in section 3).

The size of the sample that remained is 350 items, and among these, 207 (59.1%) remnants are realized as PPs. This is our first finding: English pseudogapping fails to replicate the preference for P-omission that present-day English fragments have, while it aligns itself with fragments in the languages discussed in section 2.1. The second finding is that none of my 143 NP remnants appear in a sprouting environment with an omitted preposition. This is an unsurprising result, if MiF can be expected to apply consistently across all constructions where P-omission is available by discouraging it in difficult-to-process environments. We can conclude from it that a sprouting environment blocks P-omission altogether here, just as it does in fragments (as discussed in section 2.2), and this can be construed as a grammatical constraint with a processing-based motivation. For the sake of completeness, let me add that 62 (21.2%) of the 293 PP remnants collected by Miller (2014) are sprouted (and not part of the current analysis). Furthermore, 97.4% of all items (341 out of 350) are comparative clauses. While this distribution precludes any investigation of whether comparative vs noncomparative clauses impact P-omission, the predominance of comparative clauses, which reveals as a general characteristic of pseudogapping in Miller’s (2014) corpus study and Hoeksema’s (2006) experimental study, does have the effect of strengthening structural parallelism between the antecedent and the clause hosting a remnant, which I return to in the General discussion.

These data were processed for statistical analysis in order to test for effects of MiD and MiF in the ways described in the two sections below.

4.2 Results and discussion

The questions I ask in this section are to what extent P-omission under pseudogapping obeys the principles MiF and MiD and to what extent it shows the same patterns of preference here as it does in fragments. Toward this purpose, I fitted a mixed-effects logistic regression model using the lme4 package (Bates et al. 2015) and a Bayesian logistic regression model using the brms package (Bürkner 2018) to the sample of 350 items described above. The initial models predicting P-omission included two random intercepts (Items and Register) and four fixed effects: word-external dependencies (Dependency), antecedent attachment site (Antecedent), distance between remnant and antecedent lexical head (Distance), and correlate content. The random-effects structure was suggested by multiple occurrences of identical combinations of P and V (or N) defined as Items (for instance, the P-V look at appears 11 times, talk to 9 times, and speak of 7 times) and by the data being distributed across five registers (academic, fiction, news, spoken and magazines). However, model comparison procedures yielded the final models that converged and contained one random intercept (Items) and the first three fixed effects with no interactions between them. The fixed effects, with the exception of Distance, were reliable and statistically significant at p < 0.05 in the frequentist model.¹⁰ The frequentist model has very good predictive capacity (C-value = 0.96, Somers’ D_XY = 0.92) and good explanatory power (conditional R² = 0.52). Below I describe the main effects extracted from the frequentist model (Table 4), but they correspond with those of the Bayesian model (Table 6). All model results are summarized in Tables 4, 5, 6, 7.

The first factor in Table 4, i.e., Dependency, has the effects on P-omission that are expected based on the data we saw in Table 1. I used forward difference coding for this factor (as suggested by an anonymous referee), so the first contrast shows that, as we move from a level 0 dependency up to level 1, there is no reliable increase in the predicted probability of P-omission (p = 0.69). Moving from level 1 to level 2 does, however, lead to a significant increase, and the largest effect of all three predictors (the odds of P-omission are over 6 times higher here compared to the baseline, see Table 4), in the probability of P-omission (p < 0.001).¹¹ The second factor in Table 4 is Distance. This is a numerical factor, which I log-transformed for the analysis. It’s a marginally significant predictor showing a decrease in P-omission as the distance increases (p = 0.07). The final factor, Antecedent, reveals a significantly higher probability of P-omission for remnants whose correlates are attached at the highest verb level compared to the baseline, i.e., anywhere below that level (p < 0.001).

The results reported thus far offer support for MiD operating within the VP domain when semantically dependent prepositions tend to be removed from remnants in the event that their presence compromises the remnants’ independent processability. This effect is admittedly weaker than what Nykiel & Hawkins (2020) observed in fragments in affecting only the strongest dependencies at level 2, while both level 1 and 2 differ reliably from level 0 for fragments in Nykiel & Hawkins’ (2020) data (Nykiel and Hawkins apply treatment coding). It was, therefore, a well justified move to explore the operation of MiD in the PP domain by considering attachment sites for remnants’ correlates. The main effect of Antecedent confirms our suspicion that the PP and VP domains pull in separate directions much of the time; we can only be confident that their effects are additive and in favor of P-omission when we have a level 2 dependency and a correlate attached at the highest verb site. There is no interaction between Antecedent and Dependency, likely because items with lower attachment sites are concentrated at level 1 (57 out of 63 are at level 1). Figure 1 represents this graphically by cross-classifying the factors Dependency and Antecedent with the dependent variable. This distribution suggests that stronger Dependency effects might arise within items with attachment sites at the highest verb when all items with lower attachment sites have been removed. However, fitting a regression model with only the fixed factors Dependency and Distance and the same coding scheme as before to the subset of items with highest-verb attachment sites yields very similar results as far as Dependency is concerned: it is still only level-2 dependencies that reliably increase the probability of P-omission (β = 1.75, z-value = 3.09, p <0.001). In other words, lower attachment sites account for some, but not all, of the pressure to avoid P-omission. I return to the strength of word-external dependencies in the General discussion. Finally, the main effect of Distance speaks in favor of accessibility effects as per MiF, with larger stretches of material interpolated between remnants and antecedent lexical heads causing a degradation in the accessibility of antecedents and requiring better retrieval cues from remnants in the form of prepositions. In order to explore whether there are any VP domain-based distance effects that MiD predicts for semantically dependent and independent prepositions if they are retained in remnants, I conducted post-hoc analyses.

Figure 1

Frequency counts for NP and PP remnants by Dependency and Antecedent.

I fitted a mixed-effects linear regression model to the subset of data containing only PP remnants. The model predicted the distance between remnants and antecedent lexical heads from word-external dependency level. The latter factor was forward difference-coded, as before, but its original three levels were collapsed into two to compare the semantically dependent items with the independent ones. The Distance factor was log-transformed for the analysis. The results confirm a reliable decrease in distance as we move from independent to semantically dependent (β = –0.39, t-value = –5.18, p < 0.001). It’s unsurprising that an analogous model fitted to the NP remnants shows no reliable difference in distance between semantically dependent and independent items (β = –0.06, t-value = –0.68, p = 0.5); MiD predicts no differences here. Taken together, these results reveal that MiD cares not only about the processor having to reach outside a remnant to interpret a semantically dependent preposition, but also about how far into the antecedent the resulting dependency stretches, providing tangible evidence that MiD applies to ellipsis no less than constituent ordering. It will be interesting to examine whether these distance effects are correlated with processing speed in online processing, a question I leave for future inquiry.

With these findings, we have confirmed all of the predictions of MiD and MiF, which speak to the processing underpinnings of P-omission, but one, namely, correlate content effects. The data reported in Table 3 leave us wondering why personal and demonstrative pronouns behave the same as lexical NPs although their content differs. It’s helpful to evoke the definition of accessibility that is familiar from Ariel’s (1990) work on referring expressions (see also section 2.2). On Ariel’s Accessibility Marking Scale, personal pronouns are located toward the left end, because they are typically employed in contexts where their referents are salient and hence easy to retrieve. Demonstrative pronouns are located in between personal pronouns and lexical NPs, the latter picking out the least accessible referents of the three. If personal or demonstrative pronouns appear within correlates, we would assume that they index these correlates as (highly) accessible. A high degree of accessibility sets these pronouns apart from indefinite pronouns, which we see with fragments (recall example (23)) and which, by introducing new referents into the discourse, never index salient referents. What about lexical NPs appearing within correlates? Ariel’s (1990) Accessibility Marking Scale predicts a low degree of accessibility for such correlates, but the definition of accessibility adopted in section 2 leads us to the opposite conclusion, that is, that such correlates should receive strong mental representations by virtue of their content. These definitions are not irreconcilable: a lexical NP may be indexing a low-accessibility referent at the point it is first used but its own accessibility for future retrieval will be high from that point on. There is thus no contradiction in saying that all correlates are more or less equally accessible under pseudogapping. If this is so, then all correlates should promote P-omission, as MiF would lead us to expect. Yet they don’t. This brings us to differences between pseudogapping and fragments despite their similar behavior with respect to MiD and MiF, as discussed in the next section.