2.1. Movement and PF-deletion accounts

Merchant (Reference Merchant2004) proposes that fragments derive from sentential sources, such as (7B, B′), with fronting of the fragment (here either as PP or NP) and PF-deletion of the rest of the clause, but otherwise, there is only a semantic entailment condition on ellipsis.Footnote ⁵ It follows that there are no syntactic constraints on fragments themselves.

In later work, Merchant (Reference Merchant2005, Reference Merchant2013) recognizes that certain argument structure mismatches (e.g. voice mismatch, as in (8)) are unavailable for clausal ellipsis, but his analysis of such mismatches relates to the underlying structures that embed fragments, and not to fragments themselves.

Case-matching effects are also the consequence of fragments being embedded in underlying sentential structures, with case being assigned by appropriate lexical heads, which must then be the same as the lexical heads assigning case to fragments’ correlates. Whether fragments surface as PPs or NPs is similarly linked to the underlying structures that embed them. Two kinds of underlying structures are found in the literature: declarative clauses (or interrogative clauses in the case of sluicing) that match the structure of antecedents (Merchant Reference Merchant2004) and copular clauses that do not (Rodrigues, Nevins & Vicente Reference Rodrigues, Nevins, Vicente, Torck and Wetzels2009; Van Craenenbroeck Reference Van Craenenbroeck2010; Barros Reference Barros2014; Griffiths & Lipták Reference Griffiths and Lipták2014). One of these structures is expected to apply to derive NPs or PPs as fragments, as needed. For instance, in languages that do not tolerate fronting of prepositional objects for lack of a preposition-stranding rule, and hence cannot derive NPs as fragments from declarative clauses like (7B′), copular clauses (9) could in principle be used as underlying sources.Footnote ⁶

Implicit in these accounts is an assumption that fragments may freely take either NPs or PPs as their correlates (but see Kobele Reference Kobele2015 for an explicit statement to this effect). This assumption, however, leaves it unclear why such freedom exists for merger fragments, but not for sprouting fragments. Indeed, accounting for the merger/sprouting difference has proven difficult, since it is well known that sprouting requires that a fragment share the semantic and syntactic features of the parallel position in the structure of its antecedent or, in other words, that it be no less complex a phrase than that position allows (see Chung Reference Chung, Cover and Kim2006, Reference Chung2013; Culicover & Jackendoff Reference Culicover and Jackendoff2012). To capture this requirement, Chung (Reference Chung2013) proposes a syntactic condition on DP fragments such that they be case-licensed by the same lexical heads that case-license their correlates (see further Section 4.3). Chung (Reference Chung2013) imposes further argument structure identity on antecedents and underlying structures that embed fragments, but her proposal conflicts with the possibility of using copular clauses as underlying structures, which clearly violate any kind of identity requirement based on argument structure (see e.g. Thoms Reference Thoms2015 for a critique of Chung Reference Chung2013).

An attempt to reconcile merger and sprouting fragments with a more relaxed identity condition than Chung’s (Reference Chung2013) is found in Abels (Reference Abels, Mueller-Reichau and Guhl2017). His fit condition, as stated in (10), is formulated for sluicing.

(10) requires by implication that all remnants fit into the structure of the antecedents as well as into the structure of the underlying clauses that host them, because Abels (Reference Abels, Mueller-Reichau and Guhl2017) admits the possibility that underlying structures are either interrogative or copular structures. The data that constitute the basis for formulating (10) come from Bulgarian case markers available for pronominal objects. To illustrate, (11) shows two case-marking options for the verbal object in the antecedent clause: general case (gc) or non-subject case (non.s). Abels reports that the same options are inherited by the sluicing remnant, with a preference for case match such that both the remnant and the verbal object are marked for either gc or non.s.

The Bulgarian data are interesting for two reasons. First, they show that the morphosyntactic form of a remnant and the morphosyntactic form of its correlate are both constrained by the argument structure of the lexical head that licenses the correlate. Abels (Reference Abels, Mueller-Reichau and Guhl2017) makes this point only indirectly by pointing to the necessity of a remnant fitting into the structure of its antecedent. Second, the remnant tends to follow the morphosyntactic form of its overt correlate once a choice has been made, that is, once one of the available cases has been selected by the correlate.

As far as P-drop, Abels reports that a remnant may be realized as an NP only if it is marked for gc, as in (12), and otherwise, it will surface as a PP.

Abels’s analysis of this pattern is that remnants marked for gc, unlike remnants marked for non.s, could derive from copular clauses as underlying sources (see example (9) above), since copular clauses require gc-marked subjects. Abels argues that an NP remnant marked for gc fits into the structure of its antecedent as well. Taking the syntactic category of the remnant as our guide, we locate the phrase that matches that category in the antecedent (njakoi), replace it with the remnant, and confirm that the resultant structure is well-formed (we leave it to the reader to check that (10) is also satisfied if the remnant in (12) is realized as a PP).

The fit condition also predicts correctly that for sprouting cases, nothing smaller than a PP fits the structure of an antecedent if there is an implicit PP slot in it, and hence a remnant must be a PP, as in (5) repeated here as (13).

However, we will see examples of sprouting fragments in Section 4 that do not fit into the structure of their antecedents and are therefore clearly problematic for the fit condition.

To summarize, the movement and PF-deletion approach relies on two kinds of underlying structures to be available for clausal ellipsis, while at the same time attempting to reconcile the behavior of merger fragments with the behavior of sprouting fragments. The result is a tension between allowing a variety of underlying structures (but recall the problem that legitimate underlying structures are not available in every language, see Footnote fn. 6.) and constraining how fragments may be realized, which an account like Abels (Reference Abels, Mueller-Reichau and Guhl2017) is an attempt to solve.

2.2. Direct interpretation accounts

The central idea behind DI accounts of fragments is that they have no more structure than is visible and hence are not constituents of full clauses (Levin Reference Levin and Bresnan1982; Ginzburg & Sag Reference Ginzburg and Sag2000; Culicover & Jackendoff Reference Culicover and Jackendoff2005, Reference Culicover and Jackendoff2012; Sag & Nykiel Reference Sag, Nykiel and Müller2011; Ginzburg Reference Ginzburg2012; J.-B. Kim Reference Kim2015a,Reference Kimb; Kim & Abeillé Reference Kim and Abeillé2019). We view this characteristic of DI accounts as an advantage over movement-and-deletion accounts. We can implement the idea that the argument structure of lexical heads present in the antecedent determines in all relevant cases the range of XPs that are available to serve as fragments regardless of whether we find appropriate underlying sources for the fragments. Movement-and-deletion accounts cannot do the same without losing the ability to derive fragments that require nonisomorphic sources, as discussed in Section 2.1.

The semantic and morphosyntactic features of a fragment are licensed non-locally: they are inherited from the fragment’s correlate in the antecedent clause (which Ginzburg & Sag Reference Ginzburg and Sag2000 term salient utterance and which Culicover & Jackendoff Reference Culicover and Jackendoff2005 term target) via a mechanism that integrates the fragment into a proposition constructed on the basis of the antecedent. Fragments are aligned with overt (merger) or implicit (sprouting) phrases functioning as their correlates, and the transfer of relevant semantic and syntactic features (e.g. case and preposition choice) is an integral part of this process (see, in particular, Culicover & Jackendoff Reference Culicover and Jackendoff2005). An early proposal that makes use of the idea that fragments inherit the grammatical functions of their correlates along with the relevant morphosyntactic features non-locally is found in Levin’s (Reference Levin and Bresnan1982) Lexical Functional Grammar analysis of sluicing. In the proposals of Levin (Reference Levin and Bresnan1982) and Culicover & Jackendoff (Reference Culicover and Jackendoff2005) there is no requirement that the syntactic category and/or case of a fragment match those of its correlate; rather, featural match follows from the fact that a constituent with a certain grammatical function must carry the features that are licensed for that function. However, Culicover & Jackendoff (Reference Culicover and Jackendoff2005) distinguish between direct NP arguments and oblique NP arguments (NPs functioning as prepositional objects), which they use to ensure that fragments whose correlates are implicit PPs are realized as oblique NP arguments together with the correct prepositions. This makes syntactic category a syntactic feature that is relevant for fragments.

An important aspect of Culicover & Jackendoff’s analysis is their treatment of optionality in the argument structure of lexical heads in the antecedents. There are fragments that cannot be matched with any overt correlates in the current argument structure of a head. For instance, we must interpret the fragment in (14B) as an oblique NP argument of the verb flirt despite the fact that no such oblique argument is licensed by the intransitive use of flirt in (14A).

Culicover & Jackendoff propose that the lexical entry of flirt includes the optional oblique NP argument and is activated in memory along with the intransitive variant whenever this verb is encountered in discourse. This lexical entry of flirt can therefore license the fragment in (14B) as an oblique NP argument. This idea will be relevant for our discussion of sprouting fragments in Section 4.3.

Ginzburg & Sag (Reference Ginzburg and Sag2000), working within the framework of Head-driven Phrase Structure Grammar (HPSG), constrain fragments to be the category nominal (i.e. NPs or PPs), and correlates, whether implicit or overt, must match this category. This is a requirement that consistently reappears in much later work (Sag & Nykiel Reference Sag, Nykiel and Müller2011; J.-B. Kim Reference Kim2015a,Reference Kimb; Jacobson Reference Jacobson2016; Abeillé & Hassamal Reference Abeillé and Hassamal2019). However, in none of these works is it clear how one would go about separating out merger fragments from sprouting fragments with respect to the (im)possibility of dropping prepositions (or case markers in Korean). The usual assumption that NP fragments, unlike PP fragments, take NPs as correlates (see e.g. Sag & Nykiel Reference Sag, Nykiel and Müller2011) leaves it unclear why NP correlates are possible only for merger fragments. J.-B. Kim (Reference Kim2015a) attempts to block case drop under Korean sprouting by formulating an additional constraint, which has a similar function to the lexico-semantic constraint of Chung (Reference Chung, Cover and Kim2006) blocking elements not present overtly in the antecedent from being dropped under clausal ellipsis. We will formulate a constraint that is sensitive to the merger/sprouting distinction and provide a processing-based motivation for it.

The general consensus in the DI literature is that syntactic constraints are not enforced on fragments across the board. Discussions of this point center on fragments that have adjuncts as their correlates and whose syntactic features therefore do not depend on any lexical heads in the antecedent (see Culicover & Jackendoff Reference Culicover and Jackendoff2005; Ginzburg Reference Ginzburg2012). To illustrate, consider (15) from COCA.

The fragment in (15A $ {}_2 $ ) represents sprouting of a VP adjunct, excluding the possibility that any syntactic features of the fragment are licensed by the verbal head. Similarly, Ginzburg (Reference Ginzburg2012:236) notes that locative and temporal expressions impose no syntactic category restrictions on merger fragments. A variety of syntactic categories are permitted as fragments corresponding to the adverb when, as in (16).

We will likewise not enforce any syntactic constraints on all fragments, but rather ensure that relevant fragments have access to all the morphosyntactic information available from the argument structure of lexical heads in the antecedents.

In the remainder of this paper, we defend a syntactic condition on fragments involving both syntactic category and case features, and we let processing principles interact with it to capture the merger and sprouting facts. Our analysis, articulated in Section 4, follows the DI approach, which has been convincingly argued both for English (Ginzburg & Sag Reference Ginzburg and Sag2000; Culicover & Jackendoff Reference Culicover and Jackendoff2005, Reference Culicover and Jackendoff2012; Jacobson Reference Jacobson2016) and for Korean (J.-B. Kim Reference Kim2015a,Reference Kimb; J. Kim Reference Kim2017). It remains unclear how structural identity of this kind could be implemented on movement-and-deletion approaches without sacrificing the possibility of deriving fragments from a variety of underlying sources. Like Ginzburg & Sag (Reference Ginzburg and Sag2000), we use the framework of HPSG. This choice is dictated by the explicitness of representing various aspects of the argument structure of lexical heads, on the one hand, and the admissibility of (extra-grammatical) construction-specific constraints on fragments, on the other (see Section 4).

3.1. Sprouting fragments

It is well known that sprouting fragments are subject to strong syntactic identity constraints if they pick out arguments as their correlates (Ross Reference Ross, Binnick, Davison, Green and Morgan1969; Chung, Ladusaw & McCloskey Reference Chung, Ladusaw and McCloskey1995, Reference Chung, Ladusaw, McCloskey, Gutiérrez-Bravo, Mikkelsen and Potsdam2011; Chung Reference Chung, Cover and Kim2006, Reference Chung2013; Culicover & Jackendoff Reference Culicover and Jackendoff2012). We mentioned in the introduction that this holds true of both English and Korean. Consider again (2) and (5) repeated below:

The English fragment in (17B) corresponds exactly to a PP argument and the Korean fragment in (18B) to an NP marked for source, and if these correspondences do not hold, ungrammaticality arises, as in (17B′) and (18B′). In the oft-cited examples of sprouting fragments the correlate is usually an argument within VP, but sprouting can reach into constituents that are more deeply embedded in the antecedent structure. Consider (19)–(20) from COCA.

In both examples, modifiers within NP (case and trembling) license the material being sprouted, which again must preserve the syntactic category restrictions imposed by the licensers.

These patterns demonstrate clearly that sprouting fragments must be structurally no smaller than the implicit positions in the current argument structure of given lexical heads in the antecedents, with some flexibility in terms of which heads fragments actually access. To the best of our knowledge, there are no challenges to this generalization that is termed Chung’s Generalization, after Chung (Reference Chung, Cover and Kim2006), as far as languages with prepositions are concerned (though see Section 4.3).

As for Korean, Nykiel et al. (Reference Nykiel, Kim, Sim, Kim and Farrell2018) reported, as expected, severely degraded acceptability ratings for caseless sprouting fragments like (18B′), and this also applies to omission of structural case markers, as seen in the degraded caseless fragment corresponding to an implicit accusative-marked NP in (21B′).

Korean sprouting fragments thus constitute a further robust argument for requiring morphosyntactic identity between them and their correlates in the event that the correlates are members of the argument structure of some lexical head in the antecedent.

However, sprouting reveals yet another aspect of the correspondence between fragments and their antecedents. Two case markers -lopwuthe and -eykey can express the source case in Korean. Either of them may be marked on the sprouting fragments in (22B, B′). This example indicates that the status of these fragments is radically different than the status of caseless sprouting fragments like (18) and (21), which are judged as unacceptable. We take the behavior of the sprouting fragments in (22) as evidence that sprouting permits variation in case marking only if this variation is licensed by the argument structure of a given lexical head.

Similar variation, this time in terms of syntactic category, is found in Polish sprouting fragments. Consider (23), where either a dative-marked NP (23B) or a PP (23B′) is available as a fragment.

These possibilities follow from the fact that either a dative NP or a PP may realize the indirect object left implicit in (23A). The indirect object, whether appearing as an NP or PP, can either precede or follow the direct object drinki (‘drinks’) in (23A), and this may be contributing to the acceptability of both of the Polish fragments in (23B, B′) compared to their English counterparts. The English counterparts of these fragments depicted in (24B, B′) show a different acceptability pattern: the fragment in (24B′) seems unacceptable although an indirect object NP alternates with a PP (e.g. Harvey is buying everybody drinks vs. Harvey is buying drinks for everybody) in full clauses (see also Chung et al. Reference Chung, Ladusaw and McCloskey1995 and Merchant Reference Merchant2001 for related examples from English).

We assume that this difference between English and Polish is due to the fact that, while in Polish a single argument slot permits two syntactic categories without any accompanying syntactic rearrangement of the other arguments (see Section 3.2 for similar data from Dutch), the choice between an NP and PP in English is linked to different word orders and two different kinds of argument structure. Hence, the structure in (24A) commits the indirect object to being realized as a PP.

In sum, the sprouting facts discussed above indicate that a fragment has access to all the slots in the current argument structure of lexical heads in its antecedent, including to slots filled by elements optionally realized overtly and by elements that alternate with other elements in terms of case or syntactic category. Recall from Section 2.2 that Culicover & Jackendoff’s (Reference Culicover and Jackendoff2005) analysis of fragments already permits optionality in the realization of particular slots in the argument structure of verbal heads to influence the range of fragments that are licensed.

3.2. Merger fragments and variation in argument structure

This section is concerned with fragments of the merger type similar to examples like (22) and (23) above. Judgments about fragments corresponding to alternating phrases change when one of the options is selected explicitly in the antecedent. We illustrate this point first with Korean merger fragments in (25B, B′) and then with Polish merger fragments in (26B, B′). The source argument is expressed by means of the -lopwuthe case marker (not -eykey) in the antecedent in (25A), and the mismatching fragment (25B) is degraded (though not ruled out completely) compared to the matching one (25B′).

The Polish fragment in (26B), which mismatches the recipient argument’s syntactic category (the PP dla nich ‘for them’ in (26A)) is also degraded.

The Bulgarian sluicing data we discussed in Section 2.1 fall into the same category. Recall Abels’s (Reference Abels, Mueller-Reichau and Guhl2017) example in (11), repeated here as (27).

The remnant may surface as marked for either of the cases licensed for the verbal object, which functions as the correlate. Abels (Reference Abels, Mueller-Reichau and Guhl2017) notes that there is a preference for matching the case features of the remnant and correlate, but mismatch is also possible.Footnote ⁷

There is experimental support for the matching effects noted above. Levelt & Kelter (Reference Levelt and Kelter1982) reported on Dutch data where they manipulated the syntactic category of arguments selected for by the Dutch verb laten ‘let’ in questions like (28) and (29). This verb selects for an NP (28A) or a PP (29A) and permits either NP fragments or PP fragments to be used as answers, as in (28B) and (29B). Levelt & Kelter’s main finding is what they term a ‘correspondence effect’ such that matching syntactic categories (the NP fragment in (28B) and the PP fragment in (29B)) were consistently favored over mismatching ones.

These data show that a fragment has access to the same argument structure options as those available to its correlate at the same time as it preferably replicates the morphosyntactic features chosen for its correlate. The data thus strengthen our argument that some fragments are subject to structural constraints imposed on their correlates on the basis of the argument structure of the head that licenses the correlates.

We suggest that the degradation associated with mismatching fragments and correlates (whether it is case mismatch, as in Bulgarian and Korean, or syntactic category mismatch, as in Dutch or Polish) follows from a violation of the preference for matching features, even if variation is permitted. It is important to note that the preference for matching features is a preference that can be violated only in cases where the correlate may be realized in more than one way in the antecedent; otherwise, the features of the fragment and the correlate will always match (we are not aware of any acceptable examples of fragments mismatching the features of their correlates when the argument structure of a given lexical head does not license any variation). Matching overt features can be interpreted as facilitating the process of locating a fragment’s correlate. This process is known to involve a direct-access mechanism that relies on the features of a fragment and finds the best match for it among content-addressable linguistic representations previously stored in memory (Martin & McElree Reference Martin and McElree2011). More generally, this type of direct search for the target phrase among stored memory representations is part of the cue-based models of sentence processing (see McElree Reference McElree2000; McElree, Foraker & Dyer Reference McElree, Foraker and Dyer2003; Lewis & Vasishth Reference Lewis and Vasishth2005; Lewis, Vasishth & van Dyke Reference Lewis, Vasishth and Van Dyke2006; Van Dyke Reference Dickey and Bunger2011; Van Dyke & Johns Reference Van Dyke and Johns2012; Caplan & Waters Reference Caplan and Waters2013) and is affected by the ‘cue specificity’ of the anaphor that initiates the search (here, the fragment) such that the more features that are shared between the target and the anaphor, the easier the search. To illustrate, Harris (Reference Harris2015) offered evidence for processing benefits associated with matching correlate/remnant features under sluicing. Sluicing remnants with fully specified nominal heads like which tourists in (30) were better cues to their correlates (i.e. were read faster) than remnants with partially specified nominal heads like which ones in (30).

It follows from this research that the search for the correlate for a fragment becomes harder in the event that the morphosyntactic features of the fragment are underspecified or mismatched with the features of the correlate.Footnote ⁸

We see the bias toward matching features under merger even in code switching. Gonzalez-Vilbazo & Ramos (Reference Gonzalez-Vilbazo and Ramos2014) found a preference for case matching in sluicing remnants where an antecedent was in Spanish and a clause that embeds a remnant, and the remnant itself, in German. In (31), the preference is for the accusative-marked remnant (wen), which matches the accusative-marked Spanish correlate. The dative-marked remnant (wem), which would be selected for by the German equivalent of amenazar (‘threaten’), is degraded in comparison (as is the nominative-marked remnant (wer), which neither the Spanish nor the German verb selects for).

We conclude that while variation in case/syntactic category features is generally permitted for merger fragments whenever it is also permitted by the argument structure of an appropriate lexical head, features mismatching the features of the correlates are degraded for independent reasons to do with the ease of processing.

3.3. Merger fragments and form minimization

In this section, we return to the patterns of case drop and P-drop that we saw in the Korean and English examples in (1)–(5) in Section 1. We argue that both of these phenomena have a purely processing-based motivation and are restricted to easy-to-process environments.

Given the data we have reviewed thus far, we propose that in the event that a fragment picks out as its correlate a member of the argument structure of a lexical head in its antecedent, it may pick out any member of the current argument structure, including optional arguments and alternations within a single argument slot involving case or syntactic category. The fragment must then inherit all the semantic and morphosyntactic features that are appropriate for the correlate.Footnote ⁹ However, we also need to take into account that fragments can vary in size, moving along the structure of the antecedent. For instance, all three XPs in (32B, B′, B″) and (33B, B′, B″) are possible fragments, but only the B″-fragments avoid repeating any parts of the antecedents.Footnote ¹⁰ The B-fragments are authentic (from City of Lies, R. J. Ellory Reference Ellory2010 and COCA respectively), and the rest are constructed.

The fragments in (32B″) and (33B″) correspond to the arguments of the prepositional heads about and in. They can be seen as minimal XPs that elaborate on the propositions raised by the antecedents by picking out specifically the focused elements as their correlates (the wh-phrase what and the NP something). In contrast, the B- and B′-fragments pick out larger phrases containing the focused elements. We follow Barton (Reference Barton, Leffel and Bouchard1991, Reference Barton, Progovac, Paesani, Casielles and Barton2006) in assuming that the sizes generally permitted for fragments are maximal categories that the grammar can generate, i.e. NP, PP, VP, AdjP, or AdvP, which also explains why certain reductions in fragment size (object drop in (34B) and Nominal ellipsis in (35B)) are not allowed in English.

We could also invoke here Lambrecht’s (Reference Lambrecht1994) assumption that a fragment must be phrasal, not lexical, in form, corresponding to the constituent of the antecedent that he terms the ‘focus domain’, a role that may be filled by a predicate, an argument, or an adjunct.

To return to examples (32) and (33), the minimal NPs in (32B″) and (33B″) are grammatically well-formed fragments, but the question is why they are permitted only as merger-type fragments. We are particularly concerned with P-drop, that is, the difference between the B′- and B″-fragments, as this difference is criterial for the merger/sprouting distinction. Before we proceed, note that (36) illustrates a similarly wide range of sizes for Korean fragments, from the largest (36B) to the smallest, a caseless NP (36B′″).

Now, recall that the direct-access mechanism that retrieves a correlate for a fragment, which we discussed in Section 3.2, relies on the specificity of the cues provided by the fragment. This means in turn that the fragment that contains more overt material also carries more retrieval cues that match the antecedent, reducing the difficulty of first finding the antecedent and then integrating the fragment into the proposition expressed by the antecedent. The smallest fragments, which carry minimally few retrieval cues, are expected to be the hardest to process from this perspective. This expectation finds clear crosslinguistic support in the sense that languages employ the smallest fragments (NPs) less than they do larger-sized PPs (e.g. Stjepanović Reference Stjepanović2008 for Serbo-Croatian, Szczegielniak Reference Szczegielniak, Abner and Bishop2008 for Polish, and Rodrigues et al. Reference Rodrigues, Nevins, Vicente, Torck and Wetzels2009 for Spanish and French), and acceptability judgment experiments show a degradation for NP fragments compared to PPs (Nykiel Reference Nykiel2013 for Polish, Merchant et al. Reference Merchant, Frazier, Clifton, Weskott and Goldstein2013 and Lemke Reference Lemke2016 for German, and Molimpakis Reference Molimpakis2018 for Greek). Footnote ¹¹ Footnote ¹² If we consider that the fragment is also a whole processed independently of its antecedent, we could argue that there are additional advantages to using PPs rather than NPs. Certain syntactic and semantic dependencies that are resolved within PP (i.e. case assignment, subcategorization relations and various semantic ambiguities) might prefer its constituents to be adjacent within the fragment (for more detail, see Nykiel & Hawkins Reference Nykiel and Hawkins2020). But then, why is P-drop permitted at all (and more widespread than what movement-and-deletion accounts would predict based on the pattern of preposition stranding, recall Footnote fn. 6)?

To explain the possibility of P-drop we appeal to Hawkins’s (Reference Hawkins2004) principle of Minimize Forms, one of his principles of efficient language processing, given in (37).

Nykiel & Hawkins (Reference Nykiel and Hawkins2020) argue that MiF applies to fragments by reducing their complexity from a PP (or a larger-sized XP) to an NP if the required syntactic and semantic properties can be easily assigned to them. Crosslinguistic data relating to both fragments and anaphora more generally show that the ease with which these properties are assigned is mediated by the accessibility of the antecedent. One way to think of the accessibility of the antecedent is in terms of the semantic and syntactic richness of the phrase serving as the correlate for a fragment. For instance, the minimal NP fragment has the indefinite pronoun someone as its correlate in (38) but the lexical NP a lawyer in (39). This difference results in the lexical NP correlate in (39) being a semantically and syntactically richer phrase, which makes it a more accessible correlate for the minimal fragment. There is a greater preference for an NP fragment in (39) than in (38) due to the greater ease of matching it with its accessible correlate.

This preference has been reported in several studies of fragments (Stjepanović Reference Stjepanović2008 for Serbo-Croatian; Szczegielniak Reference Szczegielniak, Abner and Bishop2008, Sag & Nykiel Reference Sag, Nykiel and Müller2011, and Nykiel Reference Nykiel2013 for Polish; Rodrigues et al. Reference Rodrigues, Nevins, Vicente, Torck and Wetzels2009 for Brazilian Portuguese, Spanish, and French; Caha Reference Caha2011 for Czech; and Nykiel Reference Nykiel2015, Reference Nykiel2017 for English) and aligns well with independent research results on memory retrieval and anaphora.

The accessibility difference between pronouns and lexical NPs can be thought of in terms of the strength of the mental representations that they receive once they have been processed. During processing semantically and syntactically richer expressions like lexical NPs receive stronger mental representations than expressions that are less contentful (e.g. pronouns), which boosts their accessibility for future retrieval from memory (Craik & Tulving Reference Craik and Tulving1975; Fisher & Craik Reference Fisher and Craik1980; Marks Reference Marks1987; Gallo et al. Reference Gallo, Meadow, Johnson and Foster2008; Hofmeister et al. Reference Hofmeister, Jaeger, Sag, Arnon, Snider, Featherston and Sternefeld2007, Reference Hofmeister, Jaeger, Arnon, Sag and Snider2013; Hofmeister Reference Hofmeister2011). We would expect therefore that phrases with stronger mental representations are more accessible antecedents for anaphors, with the result that fewer cues are required for their successful retrieval and hence less contentful anaphors are preferred. Consistent with this expectation, Karimi et al. (Reference Karimi, Fukumura, Ferreira and Pickering2014) found that the length (in words) of a nominal antecedent affects the choice of anaphor (lexical NP vs. pronoun) that refers back to it: the longer the antecedent the less contentful the anaphor.

When we think of accessibility this way, the burden of providing sufficient retrieval cues shifts from the fragment to the correlate in cases where the fragment is minimally contentful (an NP). The correlate can support the retrieval process here if it is contentful, and we expect to see the kind of gradience that is reported in the crosslinguistic literature on P-drop, with lexical NPs being preferred as correlates for NP fragments. But if the correlate has no content at all, as is the case with sprouting, then its degree of accessibility is at its lowest and the retrieval process becomes difficult, if not impossible, on account of neither the correlate nor the NP fragment providing unambiguous retrieval cues. More generally, we would predict an overall processing advantage for overt material over non-overt material that precedes ellipsis regardless of P-drop. Such an advantage has indeed been reported in reading comprehension studies of ellipsis, such that sprouting fragments were read slower than the corresponding merger fragments (Frazier & Clifton Reference Frazier and Clifton1998; Dickey & Bunger Reference Dickey and Bunger2011). The crosslinguistic ban on reducing sprouting fragments to NPs could thus be seen as a construction-specific constraint on fragments that has arisen as a conventionalization of what MiF heavily disfavors (i.e. pairing minimal fragments with low-accessibility correlates) as a grammatical principle,Footnote ¹³ along the lines of (40).Footnote ¹⁴

We will keep NFM as a constraint in our proposal, but, alternatively, we could leave it to MiF to rule out minimal fragments under sprouting.

By applying MiF to fragments, we can also account for English merger examples like those in (41)–(42).

In (41B) the PP fragment is a larger phrase than what its correlate appears to be (the NP what), and conversely, the NP fragment in (42B) is a smaller phrase than its apparent correlate (the pied-piped PP on what) (note that the status of (42B) is not the same as that of the NP fragment in the Dutch example in (29B), because an NP cannot replace a PP in the argument structure of the verbal head depend in (42A)). This is possible because form minimization is optional here: the fragment may be realized either as a larger maximal category than the minimal phrase (41) or it may simply be the minimal phrase (42). The question that remains, and which we address in Section 4, is how to determine what the fragment’s correlate is.

Before concluding this section, we offer additional evidence that P-drop is an instance of form minimization and that this kind of form minimization is specific to fragments. The best indicators of form minimization are found in languages with overt case marking, such as Greek. Molimpakis (Reference Molimpakis2018) demonstrated experimentally that in the Greek example in (43), the NP fragment must still be marked for the caselicensed by the dropped preposition (and hence, nominative is illicit here).

This pattern is expected on the assumption that the fragment undergoes purely processing-based form minimization. But it is problematic for Abels’s fit condition. If there is a well-formed underlying structure for the fragment in (43), it must include a PP within which the case on the fragment is licensed and out of which the fragment can move, but such a structure is unavailable in Greek.

While it is clear that the kind of P-drop discussed in this section is specific to anaphora, there is evidence that although Korean case drop is not, it behaves differently in anaphoric contexts than in non-anaphoric ones. Nykiel et al. (Reference Nykiel, Kim, Sim, Kim and Farrell2018) demonstrated experimentally that semantic case markers may be dropped from Korean merger-type fragments, as in (44).

Ratings for caseless fragments trended in the direction of lower acceptability than ratings for case-marked fragments, although this difference failed to reach statistical significance. Nykiel et al. (Reference Nykiel, Kim, Sim, Kim and Farrell2018) obtained similar results for Korean structural case markers. The fragment in (45) may be realized as either an accusative-marked NP or a caseless NP, with no reliable acceptability difference between the two options.

There is one configuration in which minimal (caseless) fragments are better than larger ones. In examples like (45A), the fragments’ correlate (mwues-ul ‘what-acc’) can undergo case ellipsis such that it is realized as the caseless NP mwues. The most acceptable way to combine the fragment and the correlate is by matching their structural case features by making the fragment caseless as well (Nykiel et al. Reference Nykiel, Kim, Sim, Kim and Farrell2018) (this result resembles the matching effects we discussed in Section 3.2).

The fact that Korean structural case, as opposed to semantic case, tolerates case ellipsis is relevant for our argument. Structural case is limited to accusative, genitive and nominative and is arguably specified at the constructional level in the sense of Goldberg (Reference Goldberg1995, Reference Goldberg2006) rather than at the lexical level (Kim Reference Kim2016). In other words, an NP that fills a given slot in a syntactic structure receives the case required by that slot. It is therefore possible, given a certain syntactic structure, to predict what structural cases will be assigned to particular slots in that structure. The predictability of structural case is a likely reason why NPs can drop their structural case markers in full clauses (see Lee Reference Lee2016). However, Korean distinguishes between the structural cases accusative and genitive, on the one hand, and nominative, on the other. While the former may be dropped from full clauses, the latter may not. Regardless of this difference, merger fragments can drop nominative case markers in addition to accusative and genitive, as shown for the fragment in (46B′) (see Morgan Reference Morgan, Music, Graczyk and Wiltshire1989 for this observation regarding the nominative case).

This pattern suggests that merger fragments can undergo form reduction that is not always available to NPs in full clauses. This is an important property of fragments because we could assume that the possibility of dropping structural case that exists for fragments simply follows from the fact that argument NPs have the same possibility in full clauses. That is, case drop in fragments could be reduced to the way argument NPs can be realized in full clauses (see Merchant Reference Merchant2004 for this suggestion). However, (46) demonstrates that there is no such correspondence between fragments and argument NPs, and hence no possibility of deriving caseless fragments from legitimate underlying sources here. On this basis, Morgan (Reference Morgan, Music, Graczyk and Wiltshire1989) argues for a hybrid analysis of fragments, where case-marked fragments derive via deletion, while caseless fragments are base-generated. However, we do not pursue such an analysis here.

4.1. Theoretical apparatus

We follow Ginzburg & Sag (Reference Ginzburg and Sag2000) in assuming that a fragment contains no unpronounced material. The fragment in (47B) therefore is a type of finite clause with the simple structure consisting of the NP Harvey shown in (48). This means in turn that fragments can appear in all the contexts that permit finite clauses.Footnote ¹⁵

Despite its simple structure, the fragment has propositional semantics supplied by the surrounding context in the form of a Maximal-Question-under-Discussion (see below). The resolution of the fragment is supported by the following additional discourse-based machinery. We adopt the idea common in Construction Grammar frameworks, including in HPSG, that all levels of linguistic description are understood as pairings of form with semantic or discourse functions, and the grammar is a system of constraints that license these pairings (e.g. Goldberg Reference Goldberg2006; Sag Reference Sag, Boas and Sag2012). Pairings of form and meaning are defined as constructions in some frameworks (Goldberg Reference Goldberg2006), and in other frameworks the term constructions is extended to refer to constraints on classes of these pairings (Sag Reference Sag, Boas and Sag2012). Regardless of how constructions are understood, it is expected that there may be particular discourse- or processing-based constraints that are limited to only some constructions. The availability of form minimization, as supported by the processing principle of MiF we evoked in Section 3.3, and its limiting case represented by sprouting fragments are examples of just such construction-specific constraints on fragments.

The interpretation of a fragment depends on the notion of ‘Question-under-Discussion’ (qud) in the dialogue. Dialogues are described via a Dialogue Game Board (dgb) which provides a record of who said what to whom, and what/who they were referring to (Ginzburg Reference Ginzburg and Lappin1996, Reference Ginzburg2012; Ginzburg & Fernandez Reference Ginzburg, Fernandez, Clark, Fox and Lappin2010). In other words, the dgb monitors which questions are under discussion, constantly updating the value of qud as the dialogue progresses. This context offers the basis for the interpretation of fragments. The dgb has two attributes, Maximal-Question-under-Discussion (max-qud) and Salient Utterance (sal-utt), that are relevant for fragments.

The attribute max-qud, representing the question currently under discussion, takes as its value questions defined as propositional abstracts (see Ginzburg & Sag Reference Ginzburg and Sag2000 for a defense of this proposal and a critique of Groenendijk & Stokhof’s Reference Groenendijk and Stokhof1984 early account of the semantics of questions in terms of ‘exhaustive answerhood conditions’). We assume, with Ginzburg & Sag (Reference Ginzburg and Sag2000), that questions are distinguished from other messages in terms of the feature params (parameters or variables), whose set value is empty for polar questions and nonempty for wh-questions. Each wh-phrase is associated with one parameter, which means that the params set of a wh-question will minimally be a singleton set. This can be reinterpreted using the $ \lambda $ -abstraction format, as in Sag (Reference Sag2010). The notations in (49) are sample semantic representations for interrogatives in this format: a unary wh- question with a singleton params set in (49a) and a polar question with an empty params set in (49b).

As can be seen here, wh-questions are open propositions with a nonempty params set, while polar questions are closed propositions with an empty params set (see Sag Reference Sag2010). A MAX typically arises from an interrogative clause, in which case its content is the content of that interrogative clause, as in (47A), or from a declarative clause hosting an indexed XP, usually an indefinite NP or a quantified NP, as in sluicing (see (38A) and (39A)), so long as the result is a question that remains unresolved.

Although part of max-qud, which is a semantic object, sal-utt is also the locus of certain elements of prior syntactic structure. Ginzburg & Sag (Reference Ginzburg and Sag2000) define it as an object of type local, which encodes both semantic information and morphosyntactic information like syntactic category and case via its cat (‘syntactic category’) feature. This information is used by Ginzburg and Sag for licensing the morphosyntax of fragments, as sal-utt serves as a fragment’s correlate. For instance, uttering the question Who left the party? in (47A) will activate the feature structure with the appropriate dgb information in (50).

The lexical entry for the verb leave furthermore includes the arg-st list of two ordered NPs, as in (51), in addition to the spr and comps lists. The arg-st list is an ordered list of all the arguments of a lexical head, whether realized canonically or non-canonically (e.g. as gaps). The spr list contains the subject argument, and the comps list contains remaining arguments that are realized canonically (the comps list thus may be empty if no non-subject arguments are realized canonically). The spr and comps lists are part of the val (‘valence’) feature of a lexical head, a head feature whose values are shared between the mother node and the head daughter node.

The max-qud in (50) asks for the identity of the person that left the party, and this information functions as the sal-utt whose syntactic category is NP (see the sal-utt’s cat feature value in (50)) and which is co-indexed with the first NP on the arg-st list and the only member of the spr list in (51). The fragment Harvey in (47B) supplies a value for the variable (x) and in this case agrees with the sal-utt in syntactic category. What we now need is a better understanding of what can serve as a sal-utt, especially given the patterns relating to argument structure that we discussed in Section 3.2.

Note that the value of the sal-utt may be determined in a forward- or backward-looking strategy (Ginzburg & Sag Reference Ginzburg and Sag2000; Ginzburg Reference Ginzburg2012). The forward-looking strategy preselects candidates for the sal-utt, while the backward-looking strategy uses a fragment as a guide to what the sal-utt is. We will use the backward-looking strategy since it appears to be more consistent with the direct-access mechanism via which fragments are resolved and for reasons addressed in Section 4.2 (but see Ginzburg Reference Ginzburg2012 for discussion of the viability of both strategies). Settling the value of the sal-utt via the backward-looking strategy is also entirely consistent with Culicover & Jackendoff’s (Reference Culicover and Jackendoff2005) indirect licensing mechanism, whereby the fragment is matched to an explicit or implicit phrase in the antecedent clause and inherits the syntactic features assigned to that phrase, without the two being explicitly required to share morphosyntactic features.

It is well known that fragments can pick out overt or implicit adjuncts as their correlates. Why remnants are the most frequent among direct sluices (= non-clarificational remnants), and when remnants are more frequent than both who remnants and which-NP remnants, whose correlates are arguments (Ginzburg Reference Ginzburg2012:244). We must then assume that overt or implicit adjuncts can give rise to max-quds and are available to serve as the sal-utts within them.Footnote ¹⁶ The morphosyntactic features of adjunct correlates are not licensed by any lexical heads in the antecedent, and hence it is not necessary to impose any identity constraints, beyond matching index values, on pairs of fragments and their adjunct correlates (for discussion, see Culicover & Jackendoff Reference Culicover and Jackendoff2005 and Ginzburg Reference Ginzburg2012).

However, we need to rule in the set of elements of the arg-st list of lexical heads in the antecedent that are permitted to function as the sal-utt in case the fragment picks out an argument correlate. We do so in two steps, first by co-indexing the val values of the fragment with those of the sal-utt, and second by requiring that the spr and comps lists are empty. These steps ensure that the sal-utt and the fragment are both maximal categories present in the argument structure of a lexical head, including optional arguments and case and syntactic category alternations, as discussed in Section 3.2. We flesh out this idea in the next section.

4.2. The head-fragment construction

Given the current HPSG analyses of sluicing (Ginzburg & Sag Reference Ginzburg and Sag2000; Sag & Nykiel Reference Sag, Boas and Sag2012; J.-B. Kim Reference Kim2015a) and of Korean fragments (J.-B. Kim Reference Kim2015b), we take the view that there is no syntactic structure at the ellipsis site and fragments are the sole daughters of an S-node. This view has been formalized in the HPSG feature system as follows (see Ginzburg & Sag Reference Ginzburg and Sag2000):

This constraint allows the mother to be of the same type as finite clauses and the head daughter to serve as a stand-alone XP. The head daughter may be any syntactic category but it must match the syntactic category and case features specified by the sal-utt, as indicated by their identical cat feature values. However, this constraint is too strong, incorrectly ruling out any case or syntactic category alternations, even though they may be licensed by the argument structure of a given lexical head in the antecedent. We propose a way to relax it by means of the constraint given in (53).

(53) requires that the val feature of both the sal-utt and the fragment consist of empty spr and comps lists. This ensures that the sal-utt and the fragment are saturated phrases, that is, maximal categories present on the arg-st lists of lexical heads (usually verbal heads, but see below) in the antecedent. There is no requirement of cat-feature identity between the fragment and the sal-utt, but only of val-feature identity, so that we can derive pairs of sal-utts and fragments that may be non-identical in terms of syntactic category or case features but are co-indexed with the same slot in the arg-st list of an appropriate lexical head in the antecedent clause. However, the syntactic category and case information that sal-utts and fragments carry as part of their cat feature provides a means of checking that they have the morphosyntactic features that are appropriate for the slot in the arg-st list that they are co-indexed with.

sal-utts and fragments have access to all members of the subcategorization frame of a lexical head, including those realized non-canonically, in a given syntactic arrangement. But they do not have access to alternative syntactic arrangements of the elements on the arg-st list that arise from the application of lexical rules and result in different arg-st lists. For instance, the alternants that are part of the benefactive alternation (e.g. Harvey is buying drinks for everyone vs. Harvey is buying everyone drinks) are linked to different arg-st lists, as in (54a, b).

This aspect of the theory correctly predicts that fragments attempting to retrieve arguments that are part not of the current arg-st list, but of an alternating arg-st list (e.g. (54b) if the antecedent features (54a), recall the discussion around Example (24) in Section 3.1) are ill-formed.

By (53), the sal-utt may minimally be the content of the params set of the max-qud. Thus, if the params set contains an expression co-indexed with an argument derived from the arg-st list of a prepositional head in the antecedent, then the sal-utt may just correspond to that argument, as in (32B′), repeated here as (55B″). But the sal-utt can be a larger maximal category that contains the content of the params set (e.g. the PP licensed by the verbal head talk in (55B′)), and it may include a lexical head and the argument it licenses, as in (55B). All of these possibilities are indicated by the size of the fragments, pointing to the backward-looking strategy, and are available here because all three sal-utts are overt phrases and there is nothing to block the application of MiF to these fragments (they will differ in acceptability and/or usage frequency, though, as a function of the accessibility of the antecedent, see Section 3.3).

Furthermore, if there is only one way of realizing a particular argument, then the sal-utt will simply carry the morphosyntactic features that are licensed for that argument, with the fragment matching these features. Otherwise, we may see case and/or syntactic category variation between the sal-utt and the fragment, as discussed in Section 3.2.

Further reasons to believe that the backward-looking strategy of settling the value of the sal-utt is correct come from data like (56)–(59) harvested from COCA.

It would be difficult to single out the correct phrases for being the sal-utts using the forward-looking strategy. It seems instead that these sprouting fragments are picking out arguments licensed by nonverbal heads embedded deeper in the structure of the antecedents in a manner consistent with the backward-looking strategy. Note also that these examples speak in favor of the DI approach to fragments in that they cannot be captured by Abels’s fit condition (see Section 2.1), which requires that fragments fit into the antecedent structure. These fragments clearly do not do so (e.g. * Are you a frustrated by something performer?, * But the Fed is an independent of something agency).Footnote ¹⁷

Finally, recall the fragments in (41)–(42) repeated here as (60)–(61).

In both examples the sal-utts can only be identified on the basis of what the fragments are. In (60A) the sal-utt is a PP, and hence a larger phrase than the content of the params set of the max-qud. In (61A) the sal-utt is just the value contributed by the wh-phrase despite the fact that the WH value of the wh-phrase is passed on to the fronted PP in cases of pied-piping (see Ginzburg & Sag Reference Ginzburg and Sag2000:195–200) and the entire PP is thus co-indexed with the params set of the max-qud, being a plausible candidate for the sal-utt.

We now turn to ways of licensing sprouting fragments in our account.

4.3. Licensing sprouting fragments

We argued in Section 3.3 that fragments that pick out implicit phrases as their correlates must be semantically and syntactically more complex than fragments whose correlates are explicit phrases. This is to ensure that implicit correlates, which are harder to retrieve than explicit phrases due to their low accessibility, are efficiently identified based on the information provided by fragments. Our specific task is to block the generation of fragments smaller than PPs in English and smaller than case-marked NPs in Korean. The PF-deletion accounts designed to specifically deal with sprouting attempt to ensure that all elements that undergo deletion in the IP-clause have counterparts in the antecedent clause. This idea is expressed for sluicing in Chung (Reference Chung, Cover and Kim2006:11, Reference Chung2013:30) as stated in (62) and (63), respectively.

These conditions guarantee that remnants do not surface as NPs if the corresponding NPs are not part of the overt structure of the antecedent clauses. Hence data like (64)–(65) can be captured correctly if we extend the conditions in (62) or (63) to fragments.

The DI accounts, on the other hand, cannot establish comparisons between the antecedent structure and the underlying structure of a fragment. A proposal articulated in Kim (Reference Kim2015a: 284) on the basis of Korean sluicing data is that material not present overtly in the antecedent structure has no ability to serve as a correlate for a sluicing remnant. Kim’s constraint is given in (66).

We abandon (66) in favor of NFM, which we motivated in Section 3.3 and which differentiates between fragments with adjunct correlates and fragments with argument correlates. It appears that adjunct sprouting, unlike argument sprouting, permits form minimization, as in (67) from the Santa Cruz Ellipsis Project. The fragment is the minimal NP which stop rather than the larger PP to which stop (the intended interpretation is provided in square brackets).

To make NFM work correctly we must first ask whether implicit arguments should be represented in the syntax of antecedents. There are two options here. First, implicit arguments could appear on the arg-st lists of lexical heads. This idea can be implemented following the proposal of Ruppenhofer & Michaelis (Reference Ruppenhofer, Michaelis, Bourns and Myers2014) that lexical heads can subcategorize for non-instantiated indefinite null (ini) arguments (see also Kim Reference Kim2015a). For instance, the lexical entry for the adjective angry in (68) would include an implicit PP argument annotated as ini on the arg-st list which can optionally be realized as a complement and can be activated in context whether or not realized overtly.Footnote ¹⁸ This PP is thus defined as a type of non-canonical object on the arg-st list.Footnote ¹⁹

Thus in an exchange like (65), uttering the statement that Harvey is angry can activate the implicit PP argument and introduce a max-qud questioning the value of the variable introduced by the PP argument ( $ {\lambda}_x $ [ $ angry. with\left(h,x\right) $ ]). The fragment with Mimi offers a value for this variable. The following structure is thus projected for it in (69):

Although featural match is not required between the fragment and the sal-utt, they both are PPs since they are co-indexed with the same slot in the arg-st list of angry (but if this slot in the arg-st list of angry permitted another syntactic category alternating with a PP, the fragment could optionally be realized as that category, recall Section 3.1).

Korean sprouting fragments are accounted for in a similar manner. We could assume that Korean, like English, allows ini arguments (see Ahn & Cho Reference Ahn, Cho and Müller2012 for detailed discussion) and that such arguments differ from dni arguments. The Korean fragment in (22) can function as a response to the statement A letter came, licensing a structure like the following:

Uttering the statement that a letter came evokes the max-qud asking who (x) the letter came from. This variable is activated from the arg-st of the verb o-ss-e ‘come-pst-decl’ that licenses an indefinite NP argument bearing the source semantic case (scase) and being annotated as ini in the syntactic representation of the statement A letter came. Footnote ²⁰ The fragment in (70) supplies a value for this variable, which serves as the sal-utt. Since two case markers -eykey and -lopwuthe can indicate source, either can be selected for the fragment (J.-B. Kim Reference Kim2015a). However, the ini annotation on the sal-utt provided by the antecedent triggers the application of NFM, blocking the fragment from being caseless.

The second option is to allow implicit arguments to be absent from the syntax of antecedents, while recognizing that they are activated in online processing. One implementation of this idea is found in Culicover & Jackendoff’s (Reference Culicover and Jackendoff2005) proposal that fragments may be matched to implicit phrases that are licensed as part of subcategorization frames of lexical heads and activated together with those heads in online processing. The processing of a lexical head will activate its entire set of arguments, including optional ones, which will then be available as correlates for fragments. This means that a fragment’s semantic index is not linked to any arguments encoded in its antecedent’s syntax but instead points to a grammatical function that is fulfilled by a certain implicit phrase with appropriate morphosyntactic properties. An anonymous referee points out that for a max-qud to arise from such an implicit correlate might require a certain degree of accommodation, which is costly. To see this, consider (65) again. The salient max-qud is the polar question of whether Harvey is angry, but the fragment with Mimi requires that the dgb be updated with the new max-qud Who is Harvey angry with? This additional step arguably introduces processing difficulty that is absent from examples where no update to the dgb is needed. However, it is equally plausible that processing difficulty arises here from having to retrieve a low-accessibility correlate (one with no linguistic content), as predicted by MiF, and also from retrieving the new max-qud that this correlate gives rise to. All fragments used in this context will be required to be realized as phrases exactly matching their implicit correlates, with no form minimization permitted, as per NFM.

Whichever option we choose, we are dealing with phrases with no linguistic content that make less accessible correlates, and hence phrases that trigger the application of NFM. An advantage of the second option, as noted by an anonymous referee, is that it is more consistent with the backward-looking strategy than the forward-looking one, given that implicit arguments are not represented in the antecedent syntax.