Pseudogapping is most typical with transitive verbs (with NP or PP complements).

Though both the comparative and the noncomparative variants are clearly acceptable in such simple examples, pseudogapping is a somewhat marginal phenomenon at best, and judgments are often unstable. For this reason, it is important to first clarify the factors that affect the felicity of pseudogapping and to control for them as much as possible.¹

The most fundamental property of pseudogapping, which is particularly important to bear in mind, is that, as noted by Hoeksema (2006), this construction must satisfy the Contrast relation in Kehler’s (2002) classification of discourse relations.² Thus, note that the highly marginal (6a) improves with the use of contrastive but in (6b) and becomes virtually unexceptionable with the use of the comparative structure in (6c).

Note moreover that in all these cases, contrastive emphasis on essays and novels increases acceptability of the sentence as uttered (other sources of increased acceptability include the use of the demonstratives this/that (see section 4.6), which corroborates the same point).

Indeed, Hoeksema (2006) notes a strong statistical association between pseudogapping and comparative constructions, where 87% of his attested examples involve comparatives or constructions for comparison (with expressions such as like and the way/manner). This makes sense given the tight correlation between pseudogapping and the Contrast relation.

Also, as noted by Levin (1979), Hoeksema (2006), and Miller (2014), keeping the subject of the antecedent and the pseudogapping clause identical greatly increases the acceptability of pseudogapping (in fact, Miller notes that 85% of the pseudogapping examples in his corpus sample contain a pronoun as the subject of the ellipsis clause). Thus, compared with (6a–b), (7a–b) are somewhat degraded.

The effect of the Contrast requirement and the “same-subject” preference is that the least acceptable example in this paradigm is (7a) with no contrastive stress on the remnants (and no discourse context suggesting that essays and novels are contrasted), and the best is (6c) with strong contrastive stress on the remnants. Thus, we do not regard (7a) as ungrammatical; it just fails to satisfy all the relevant discourse conditions affecting the felicity of pseudogapping.³ When presenting our examples below, we will control for these factors so that the examples will not violate these interfering discourse conditions. This is especially important for examples with more complex structures, in which such effects (unsurprisingly) tend to be aggravated. For example, even with single remnants, when the syntactic and semantic types are not the simple NP individualdenoting type as in (4) and (5), the acceptability noticeably drops, as in the following (but note that the comparative structure is consistently better than the noncomparative structure):

Beyond the “base cases” involving direct objects of transitive verbs as remnants, there exists a variety of more complex pseudogapping examples that are well within the range of acceptable patterns. We take all these examples to be generated in the syntax since doing so will make the overall analysis simpler. Wherever relevant, we offer some observations on the extragrammatical factors possibly affecting their perceived acceptability.

As noted in section 1, two major approaches have been taken to pseudogapping: (a) transformational analyses with movement + VP-ellipsis and (b) nontransformational analyses that rely on purely anaphoric mechanisms to retrieve the meaning of the missing verb. We now turn to data that prove to be especially difficult (or even intractable) for one or the other of these approaches.

There are two aspects to movement-based approaches to pseudogapping that need to be kept separate. One is the characterization of pseudogapping (and ellipsis more generally) as an operation that makes reference to purely syntactic information. The second is the specific implementation of this syntactic dependency via structure-changing operations.

The essential insight of movement-based approaches seems to lie largely in the first of these aspects. Movement-based approaches immediately explain the category-matching connectivity effect in pseudogapping, which can be accommodated only by an ad hoc stipulation in the interpretive approaches. At the same time, as we discuss in detail below, previous transformational analyses are unsatisfactory on both empirical and conceptual grounds: the various movement operations utilized for the analysis of pseudogapping either lack independent motivation, or (when an independently motivated movement is retooled) do not match the actual distributional properties of pseudogapping. Moreover, movement-based approaches do not by themselves illuminate the question of why we might expect something like pseudogapping to be a possible type of ellipsis in English.

The transformational literature has essentially followed Kuno (1981), who took pseudogapping to be a case of VP-ellipsis in which various constituents are moved out of the VP via adjunction operations, thus “surviving” VP-ellipsis. Adopting this general idea, Jayaseelan (1990) analyzes (20) (=(1)) as in (21), via heavy NP shift (HNPS).

However, this approach faces major empirical challenges. First, since HNPS cannot move the NP complement of a preposition, this analysis incorrectly rules out examples like the following (Lasnik 1999, Miller 2014):

Second, Jayaseelan attributes the ill-formedness of (23a) (the judgment * is Jayaseelan’s) to the impossibility of multiple rightward movements in HNPS. But this supposed prohibition is directly contradicted by data such as (23b) (see more examples in section 2).

Given that pseudogapping is much more acceptable in comparatives than in ordinary coordination, the contrast in (23) isn’t particularly surprising.

An extreme example of this kind is provided by (10), repeated here as (24).

On Jayaseelan’s analysis, the input to the movement prior to VP-ellipsis is the following:

In order to evacuate VP₀ of all its nonhead daughters, leaving only bet in place to be deleted, movement must apply successively to each of the complements of the verb.

But the rightward movements in (26) have serious empirical shortcomings. As we discuss below, when the verb is not elided, such rightward movements are ill-formed.

To see this, note first that neither of the two objects of bet can be right-shifted via HNPS.

The unacceptability of (27b) or (28b) cannot be attributed to the NPs themselves since they can be right-shifted (cf. In the past, I’d transferred to Terry’s account a ton of money).

The pattern just observed severely jeopardizes an account of (24) via rightward movement. Such an account would first take the leftmost complement an enemy to undergo HNPS to the right, followed by two further successive rightward movements targeting the remaining complements (below, x marks an operation shown to be inadmissible in (27b) or (28b)).

In short, the necessary rightward movements are precisely the prohibited ones.⁹

Finally, Jayaseelan argues that (30) supports the HNPS analysis since HNPS would not be able to apply to a weak definite pronoun such as it.

But this only shows that pseudogapping requires its remnant to carry stress. The stress requirement itself presumably follows from the required Contrast relation in pseudogapping along the lines discussed in section 2 (in fact, Jayaseelan himself notes this condition). Note that replacing it with that (with a marked stress on it) improves (30). (Incidentally, as a reviewer reminds us, the fact that pronouns other than it can readily occur as remnants in pseudogapping provides a strong argument against Jayaseelan’s proposal—one of the hallmarks of pronouns is that they are exempt from HNPS.)

Subsequent transformational analyses have added little to Jayaseelan’s main ideas. The only differences consist in whether the movement is taken to be A- or Ā-movement, and rightward or leftward movement. For example, (31) illustrates Lasnik’s (1999) alternative.

Here, Jayaseelan’s rightward HNPS is replaced by a leftward A-movement of the remnant NP Harry to Spec,Agr_O. As noted by Takahashi (2004), while this treatment avoids the difficulties of an exclusively HNPS analysis, it creates a new problem. Examples such as (32) require a structure in which everything in the VP except CDs is deleted.

The complex interactions of Lasnik’s assumptions about feature checking, derivational economy, and binary branching yield the following structure:

To delete both give and me, it would be necessary to delete VP₁, which would also delete CDs. Suppose, then, that we instead assumed a simpler initial structure, where the verb directly precedes me and CDs, and then deleted the partially evacuated VP, as in (34).

However, as noted by Takahashi, this derivation would also fail. The problem is that such a derivation requires a leftward A-movement of the indirect object CDs across the direct object—an operation that is blocked (except in British English) in nonellipsis contexts.

Thus, as noted by Takahashi, there is no available derivation for (32) on the assumption that pseudogapping involves exclusively leftward A-movement prior to VP-deletion.¹⁰

Finally, Culicover and Jackendoff (2005:294) note that Lasnik’s analysis, if applied to data such as (36), would require the clausal remnant to undergo A-movement to the left.

But, as they note, crosslinguistic evidence from Dutch and German, where overt leftward object shift is standard, shows that clauses do not undergo such movement.

In place of Jayaseelan’s (1990) exclusively rightward and Lasnik’s (1999) exclusively leftward movement analyses, Takahashi (2004) proposes a mixed analysis where both (leftward) object shift and rightward adjunction are available to partially evacuate VPs prior to deletion. It might seem at first that this “eclectic” approach would overcome the problems just noted for Lasnik’s analysis, as well as those noted earlier for Jayaseelan’s. For example, (32) and (36) can be generated just as they would be under Jayaseelan’s analysis, via a single application of HNPS. In the case of (9), Takahashi’s analysis would move the leftmost complement to the left via object shift, followed by rightward A-movement of the rightmost complement. In a sense, Takahashi’s approach can be seen as the limiting case of the movement strategy: given that neither the leftward nor the rightward analysis covers all cases, the next (and the last) analytic alternative is to combine all approaches that have worked in particular cases. Unfortunately, however, a wider set of data reveals problems similar to those that undermine the previous accounts.

In (24), for example, there are three remnants. Takahashi’s analysis would take the leftmost complement an enemy to undergo object shift to the left, followed by either two rightward movements targeting each of the remaining complements, or a second movement to the left, applying to euros, and a movement of the clausal complement to the right. But both of these possibilities are ruled out by Takahashi’s own respective arguments against Jayaseelan’s analysis on the one hand and Lasnik’s on the other. In the former case, the same problem arises as in (29): leaving aside the legality of multiple HNPS, the first of the rightward movements must move the indirect object euros over the clausal complement. But, as discussed above, this is prohibited; see (27). In the latter case, the first movement must move the indirect object over the direct object—again, an option precluded for Takahashi’s approach, since admitting such movement would incorrectly license the passivization of an indirect object in (35).¹¹

Given the discussion to this point, we have two kinds of evidence bearing on the movement hypothesis for pseudogapping: the general argument for movement in pseudogapping based on putative compliance with island constraints is undercut by the evidence from Miller 2014 given in section 2.2, while the arguments just reviewed against each specific movement-based analysis make it difficult to see how such analyses can be maintained. But the problems do not end here. Both a general conceptual problem and one specific empirical problem pose serious challenges to the general class of movement+VP-ellipsis analyses, regardless of the specific implementation of the movement and deletion operations involved.

We start with the conceptual issue. The main motivation for a movement+VP-ellipsis approach comes from the fact that pseudogapping can be subsumed under VP-ellipsis once some movement operation can be established to evacuate the remnant (but note that the latter component is actually a major weak point of this approach). A big advantage of such an approach in particular is that syntactic connectivity effects come for free (for a similar argument involving other types of ellipsis, see Merchant 2004). Despite these motivations, however, there are examples that pose serious challenges to a syntactic approach, such as the antecedentless and split-antecedent pseudogapping examples noted in section 2 (see (16) and (17)). These examples suggest that, despite the initial appeal of the movement+deletion strategy, descriptively speaking, the type of ellipsis involved in pseudogapping is anaphoric rather than being licensed syntactically. But then, the fact that pseudogapping leaves a remnant (displaying connectivity effects) is particularly troublesome, since, as noted by a reviewer, extraction “out of ” unequivocally anaphoric expressions is generally prohibited. Note for example the following contrast between antecedentcontained ellipsis and its counterpart involving do so–anaphora:

Previous syntactic accounts of pseudogapping remain silent about this tension between (apparent) evidence for a structural account and evidence against it.

Moreover, there is at least one empirical argument against the specific assumption (common to all movement+VP-ellipsis analyses) that a syntactic operation of VP-ellipsis underlies pseudogapping. This assumption leads to a striking incompatibility between the principal derivational analyses of pseudogapping and of gapping in view of data such as (38) involving an interaction of the two.

Consider the consequences of (38) for Johnson’s (2000, 2009, 2014) low-VP-coordination/acrossthe-board (ATB) verb movement analysis of gapping. On the one hand, under a VP-evacuation/deletion analysis of pseudogapping, the first conjunct of the than-clause you can ice cream is an output of VP-ellipsis, deleting a VP containing the verb and the trace of the remnant direct object. On the other hand, in order to get gapping in the right-hand conjunct, Johnson’s analysis requires ATB movement of the verb eat. Suppose, following Johnson (2000, 2009), we assume a structure for (38) along the lines of (39).

If eat undergoes ATB movement from this structure, where have the two tokens of this verb in each conjunct in (39) gone in (38)? Suppose the ATB movement for gapping applies first. Then, eat is removed from the first conjunct, can no longer be deleted by VP-ellipsis, and hence is necessarily visible in the comparative clause at the end of the derivation, contrary to fact. The only other option would be to start with pseudogapping in the left-hand conjunct. In this case, we would obtain the intermediate structure in (40).

Even allowing non-ATB movement from the VP, we still have nowhere to move eat to such that (38) is derived. Given these considerations, it seems fair to say that there is no straightforward analysis of the pseudogapping-gapping interaction in (38) consistent with the standard assumptions about the two phenomena in movement-based approaches.

Thus, previous movement-based approaches not only are problematic as analyses of pseudogapping itself, but also suffer from the implications of the fundamental premise: the assumption that the verb is elided by the syntactic operation of VP-ellipsis not only is undermotivated given an overall descriptive classification of ellipsis and anaphora, but also leads to mispredictions in interaction with analyses of other syntactic phenomena.

During the past three decades, an alternative approach to ellipsis has emerged, whose central claim is that ellipsis never involves covert structure (e.g., Schachter 1978, Sag et al. 1985, Miller 1990, Dalrymple, Shieber, and Pereira 1991, Hardt 1993, Culicover and Jackendoff 2005). Versions of this approach typically invoke some kind of anaphoric process based on the semantics of the antecedent clause. We illustrate this strategy by reference to Miller 1990, which offers the most explicit proposal of this sort to date for pseudogapping (see Culicover and Jackendoff 2005 for a similar idea, worked out in less detail).

The key idea of Miller’s (1990) nonderivational analysis of pseudogapping, couched in Generalized Phrase Structure Grammar, is that auxiliaries can appear as the head verb in the same set of phrase structure rules that license projections of lexical verbs. For example, in (1), reproduced here as (41), the auxiliary has is effectively treated as a transitive verb and directly combines with the remnant Harry.

Miller implements this strategy by assuming that auxiliaries can appear not only in subcategorization frames taking nonfinite VP complements, but also in frames instantiating any subcategorization frame of a lexical verb in English. This means that the auxiliary has is specified in the lexicon to be compatible with the [SUBCAT 2] specification, which is associated with the following phrase structure rule licensing lexically transitive verbs such as drink:

This rule licenses (41), and the meaning of the “missing” verb is then supplied by anaphoric reference to some “corresponding” verb in the preceding clause.

Elegant though it is, this analysis has one serious source of overgeneration. The problem, in a nutshell, is that Miller’s anaphora resolution procedure makes no reference to any syntactic information of the antecedent clause—in particular, to the syntactic selectional properties of the head verb, which must be matched by the auxiliary in the pseudogapped clause, as discussed above. This indeterminacy entails that if some complement in the pseudogapping clause has a denotation that corresponds to the denotation of a syntactically different complement in the antecedent clause, then it is in principle possible to obtain a coherent interpretation in Miller’s analysis even though the verb in the antecedent clause cannot actually combine with the pseudogapping clause complement. Thus, this account as it stands does not predict the anomaly of (19a), repeated here as (43).

Here, the individual denotation anne is a possible interpretation for for Anne (cf. John waited for Anne, where the preposition for is standardly taken to be meaningless). But then, the meaning of the auxiliary did can be anaphorically resolved as the meaning of the verb spoke in the antecedent clause (note that to in spoke to Mary is similarly meaningless), leading to the misprediction that (43) should be well-formed with the same interpretation as John spoke to Mary more often than he did to Anne.

Miller takes (43) to be ruled out by a semantic selectional restriction analogous to the gender restriction on pronouns. This selectional restriction applies to the anaphoric auxiliary and imposes the constraint that it is felicitous just in case the verb meaning that is anaphorically retrieved is compatible with the overt preposition that heads the PP that the auxiliary syntactically combines with. Thus, for example, (43) is predicted to be semantically anomalous since NP1 speak to NP2 and NP1 speak for NP2 mean different things (NP2 is a participant in the act of speaking in the former but not in the latter). Thus, when speak appears with for (as in the pseudogapping clause), the meaning of speak in the antecedent clause would not be the “appropriate” one, and anaphora resolution therefore fails. Though this approach seems in principle implementable in an interpretive approach, it is unclear to us what motivates the anaphoric auxiliaries (which are all identical in form in the relevant respect) to carry semantic restrictions based on the intended antecedent target, which according to Miller is no different from the gender restriction on pronouns (the latter of which has a clear morphological reflex on the overt form of the pronouns).¹² In the next section, we offer an alternative formulation of the syntactic connectivity restrictions that keeps the core insight of Miller’s proposal but implements the relevant constraint in a way we believe is much more straightforward.

This section presents a quick overview of Hybrid TLCG, pitched specifically to readers familiar with standard derivational approaches (for a more complete presentation discussing the logical underpinning of the theory in detail, see Kubota 2010, 2015 and Kubota and Levine 2014a). We start with a simple CG equivalent to phrase structure grammar, extend it first with a mechanism that models (and in fact generalizes) the notion of movement, and then extend it further by introducing flexible constituency.

Since we take pseudogapping to be a special case of VP-ellipsis, we start with an analysis of VP-ellipsis. In CG, auxiliary verbs are standardly analyzed as having the syntactic category VP∕VP (where VP is an abbreviation for NP\S), as in the following lexical entry for can:

We take VP-ellipsis to be licensed by an alternative sign for the auxiliary verb that does not subcategorize for a VP but instead anaphorically retrieves the relevant VP meaning in reference to the preceding discourse. For this purpose, we posit an empty operator that applies to the lexical sign of auxiliaries and saturates the VP argument slot of the latter. This “VP-ellipsis” operator is defined as in (56).

By applying (56) to (55), we obtain a derived auxiliary entry of category VP as in (57).

Then, a simple VP-ellipsis example such as (58) can be derived as in (59) (here and below, the syntactic category of the expression that serves as an antecedent of VP-ellipsis is shaded).

Note that, since the operator directly applies to the auxiliary to modify its subcategorization property, no phonologically empty verb is involved.

At this point, some comments are in order about our choice of an analysis involving an empty syntactic operator. There are at least three alternatives to this approach: (a) a bindingbased analysis in which a hypothetical VP is bound by an antecedent VP via a syntactic mechanism of variable binding (Morrill, Valentín, and Fadda 2011, Barker 2013); (b) an analysis that posits an empty VP (this would correspond most closely to a deletion-based analysis in derivational approaches); and (c) an analysis that posits an alternative auxiliary entry (identical to the output of our syntactic empty operator) in the lexicon (Jäger 2005).

We find these three alternatives less than optimal. The binding approach does not easily extend to intersentential anaphora; especially problematic are cases where VP-ellipsis takes place across speakers. The present approach is superior to an empty-VP approach in that it can straightforwardly capture the generalization that auxiliaries (including the “infinitive marker” to) are the triggers of VP-ellipsis.¹⁷ We believe that our approach is superior to a lexical approach along the lines of the third alternative in straightforwardly generalizing to the pseudogapping case (see below). It is not clear whether a purely lexical approach like Jäger’s (2005) can offer a general characterization of the set of alternative entries for the auxiliary necessary to license pseudogapping.

Interactions between VP-ellipsis and other phenomena such as quantifier scope and the strict/sloppy ambiguity of pronouns can be handled in essentially the same way as in previous analyses of VP-ellipsis in TLCG (Morrill and Merenciano 1996, Jäger 2005). (61) shows the sloppy reading of (60a) and (62) shows the every > before reading of (60b).

We assume the so-called binding-at-VP analysis of pronouns in (61) (see Bach and Partee 1980, 1984). In this analysis, after the binding of the pronoun to the subject NP, the right meaning (selfascription of the property of being a genius) is assigned to the VP, which the VP-ellipsis operator can then take as the antecedent.

In the quantifier-scope interaction case in (62), the VP-ellipsis operator takes the VP in the antecedent clause containing a free variable x (to be later bound by the universal quantifier) as the antecedent. The quantifier takes scope over the whole sentence after this anaphora resolution takes place, and binds the variable x in both the antecedent clause and the ellipsis clause.

We analyze pseudogapping in (63) via transitive verb (TV = (NP\S) ∕ NP) ellipsis (Jacobson (2014) independently arrives at the same conclusion).

In the present setup, this involves making only a minimal extra assumption. In fact, the only thing we need to do is to make the VP-ellipsis operator in (56) polymorphic. Polymorphism is a standard technique for generalizing the lexical definitions of semantic operators independently needed in the grammar, in the analysis of coordination and certain adverbial operators (see the “crosscategorial” analysis of focus particles in Rooth 1985).

Moreover, there is independent evidence that English allows for TV-ellipsis. Jacobson (1992, 2008) argues that ACD is to be analyzed in terms of TV-ellipsis rather than VP-ellipsis. The idea is that in (64), what is missing after had is just the transitive verb showed instead of a full VP.¹⁸

We refer the reader to Jacobson’s work for a detailed empirical justification and technical execution of this analysis of ACD (see also Jäger 2005 for a TLCG implementation of Jacobson’s analysis), but one big advantage should be immediately obvious: in this analysis, the notorious problem of “infinite regress” simply does not arise, since a VP containing a trace is not reconstructed in the ellipsis site to begin with.

Since pseudogapping is not restricted to transitive verbs but can involve ditransitive verbs, and so on, we make the VP-ellipsis operator polymorphic, employing Steedman’s (2000b) $-notation for polymorphic lexical entries.

VP∕$ is a metavariable notation for a set of categories where any number of arguments (of any category) are sought via ∕ (VP, VP∕NP, VP∕ NP∕ PP, etc.). The three occurrences of VP∕$ are to be instantiated in the same way. The key idea behind this extension is that the ellipsis operator is generalized to apply to any syntactic category that the auxiliary itself can be derived in (as will become clear momentarily). Thus, though the schema in (65) itself needs to be stipulated, it embodies a natural extension of the simpler version in (56).

The TV∕TV (=(VP/ NP) ∕ (VP∕ NP)) entry of the auxiliary that this operator applies to in the analysis of (63) can be derived from the lexically assigned VP/ VP entry and does not need to be posited separately. This is an instance of the Geach rule, which is a theorem in the Lambek calculus and TLCG (as long as the calculus is associative).

The analysis of a basic pseudogapping example like (63) is then straightforward.

Here, the auxiliary is in the derived TV∕TV category. The VP-ellipsis/pseudogapping operator in (65) takes this auxiliary category as an argument and saturates its TV argument by anaphorically referring to the transitive verb eat in the antecedent clause.

As discussed in section 2, the “deleted” material in pseudogapping does not necessarily correspond to a syntactic constituent in the traditional sense. The present approach straightforwardly handles such cases of “nonconstituent” pseudogapping (like those in (11)), by treating the “nonconstituent” strings in the preceding clause as syntactic constituents that can serve as antecedents in pseudogapping. We illustrate in (69) the derivation for (the antecedent clause of ) (68) (= (11a)).

Via hypothetical reasoning involving directional slashes, the string take the lining out of is derived as a syntactic constituent of category VP∕NP. This can then be identified as the antecedent of the relevant anaphoric process in the target clause. Examples like those in (11) are especially important in that they show the significance of the flexible notion of constituency available in CG in an empirical domain other than coordination.¹⁹ Note that these nonconstituent pseudogapping examples pose significant problems for many previous transformational accounts since deriving these examples via movement + ellipsis entails positing various otherwise unmotivated movement operations.

As discussed in section 3, pseudogapping with multiple remnants like the following is also highly problematic for movement-based approaches:

Multiple-remnant pseudogapping is straightforward in our approach. The key point is that the following PDTV∕PDTV (where PDTV = VP∕PP∕NP) version of the auxiliary can be derived from the lexically specified VP∕VP entry via the Geach rule:

Since the derivation is parallel to the one for the TV∕TV entry above (hypothesizing a PDTV, NP, and PP to the right of the auxiliary and withdrawing these hypotheses one by one after combining them with the auxiliary), we omit it here.

Since the VP-ellipsis/pseudogapping operator is polymorphic, it can take this derived auxiliary verb as an argument and anaphorically saturate the missing PDTV argument position, in the same way as in the simpler examples above. Here, we show only the derivation for the target clause of pseudogapping. The VP-ellipsis/pseudogapping operator makes reference to the ditransitive verb in the antecedent clause with category PDTV and semantics introduce (here we ignore the tense meaning of the auxiliary would).

The present analysis also correctly predicts the interactions between pseudogapping and strict/sloppy readings and quantifier scope in examples like the following:

We omit the derivations, which are parallel to the VP-ellipsis case in (61) and (62).

Since in CG the combinatorial properties of linguistic expressions (including those corresponding to nontraditional constituents) are represented explicitly in their syntactic categories, our approach overcomes the major problem for previous nontransformational approaches as well. Recall from section 3.2 that Miller’s (1990) interpretive approach has difficulties in explaining the ungrammaticality of preposition mismatch examples like (74) (=(19a)).

Our approach rules out this type of example straightforwardly. In the antecedent clause, we have an instance of the verb speak that subcategorizes for a to-PP (of syntactic category VP∕ PP_to). But in the target clause, we need to recover the meaning of speak associated with a different subcategorization frame, VP∕PP_for. Because of the syntactic category mismatch, the relevant anaphoric mechanism fails and hence (74) is correctly blocked.

Interestingly, the present proposal can also correctly capture cases of tolerated category mismatch, exemplified by data such as (75) (=(15)).

Miller (2014) makes the important observation that (75) is licensed despite the preposition mismatch because of the closeness of the lexical meaning of the verb in the different subcategorization frames. This condition is not satisfied in the minimally different (74), resulting in the degraded status of the latter.

To see how the contrast between (74) and (75) can be accounted for in the present approach, note first that exactly the same contrast is found in unlike-category coordination (UCC).

This contrast motivates assigning the category VP∕ PP_of VP∕PP_about involving the meet connective to the verb speak, following the general analysis of UCC by Morrill (1994) and Bayer (1996) (see also Kubota and Levine 2013). We assume that ∕ and \ associate more strongly than ; thus, VP∕ PP_of VP∕ PP_about is an abbreviation for (VP∕ PP_of ) (VP∕ PP_about). In (77), the two (related yet distinct) meanings of speak associated with different subcategorization frames are represented separately in the form of a tuple.²⁰

With this lexical assignment and the Meet Elimination rules in (78) (where π₁ and π₂ are the standard projection functions such that π₁ (‹α, β›) = α and π₂ (‹α, β›) = β), the analysis for (76a) is straightforward, as in (79).

The contrast in (76) then follows from the assumption that speak with a for-PP complement is simply listed as a separate entry in the lexicon. We take it that the “closeness” of meaning that Miller (2014) alludes to governs which subcategorization frames can be “packaged” into a single lexical entry involving the meet connective for any given verb.

The parallel contrast between (74) and (75) in the pseudogapping case follows from the same assumption. The preposition-mismatch pseudogapping apparently violating connectivity is licensed in the present analysis without any extra machinery, except that the anaphoric retrieval mechanism is a bit more involved in this case. We assume that the VP-ellipsis/pseudogapping operator can access either of the two category-meaning pairs stored in a linguistic sign involving the meet connective such as (77). With this assumption, examples such as (75) can be accounted for straightforwardly. We provide an analysis of (80), a slight variant of (75), in (81).

The ungrammaticality of (74) still follows, since the meaning of speak associated with the different lexical entry with syntactic category VP∕ PP_for cannot be anaphorically retrieved from an occurrence of the VP∕ PP_of VP∕ PP_about entry in the antecedent.

Furthermore, the following related example noted by Miller (1990:300), in which a ditransitive verb instantiates different subcategorization frames (V NP NP vs. V NP PP) in the antecedent and the pseudogapping clauses, can be analyzed in essentially the same way (see Kubota and Levine 2014b for a complete derivation):

The key assumption is the following entry for the ditransitive verb give involving the meet connective (which again is motivated by the pattern in UCC (Kubota and Levine 2013)):

We take voice-mismatch examples such as (84) (=(14)) to be licensed in a similar way.

Though the active/passive alternation is different from the argument structure alternation involving ditransitive verbs in that a morphological marking is involved (thus, the meet connective would be of no use here), there is an obvious similarity between examples like (82) and voice mismatch examples involving the active form in the antecedent clause licensing a passive pseudogapped verb or vice versa. The key in both cases is lexical relatedness and the mutual entailment of the two related meanings.²¹ Following the standard assumption in the nontransformational literature (Bresnan 1982, Pollard and Sag 1994), we take passivization to be a lexical operation. Since the argument structure and the morphological form are different, the passive form of a verb is listed in the lexicon as a distinct entry separate from the active form. However, they are related to each other via some explicit lexical operation (one standard way of formalizing this is in terms of lexical rules), and the active and the passive forms have identical meanings in terms of their truth-conditional entailments. It is then not unreasonable to assume that the pseudogapping operator can have access to the lexical entry of the passive form from the occurrence of the active form in the preceding clause and vice versa, because of this close relation between the lexical entries for the active and passive forms in the lexicon. Thus, voice-mismatch examples like (84) do not pose problems for the present approach.

Before moving on to more complex examples involving discontinuity, we would like to discuss how the interaction between pseudogapping and gapping discussed at the end of section 3.1 (and identified there as being problematic for transformational analyses) can be dealt with in the present approach.

Unlike movement-based approaches, the present approach properly handles the gappingpseudogapping interaction in (38), reproduced in a slightly modified form in (85) (involving coordination rather than a comparative structure).

The only difference between this example and simpler examples of gapping like (86) is that in (85), the verb eat is missing from the first conjunct of the gapping conjunction because of pseudogapping.

Thus, the proof for (86) works for (85) as well except for this extra complication. In our analysis in Kubota and Levine 2016a:app. 2.2, the following sign is derivable:

The rest of the derivation merely involves replacing the actual transitive verb eat with the pseudogapping operator supplying the meaning of this transitive verb for the auxiliary via anaphoric reference to the preceding clause.

As discussed in the previous section, our approach extends smoothly to quite complex types of data such as multiple-remnant pseudogapping (70), “nonconstituent” pseudogapping (68), “unlikecategory” pseudogapping (75) and (82), and the just-mentioned interaction between pseudogapping and gapping (85), all of which are highly problematic for many previous approaches. However, as it stands, the present analysis does not yet cover cases of discontinuous pseudogapping exemplified by the following data:

There are at least two possible approaches to this problem, and deciding between them is a delicate matter, given the somewhat marginal status of the pseudogapping construction itself and especially its discontinuous variant. Here, we simply lay out the two options and leave it for future work to determine which is the better alternative.

One approach would be to deal with discontinuity via the prosodic λ-binding mechanism already available in the grammar. Since this alternative does not involve adding any new machinery to the grammar, we will be relatively brief in sketching the main idea. (See Kubota and Levine 2014b for a more detailed demonstration of this approach with example derivations.) Essentially, the idea of this vertical-slash-based approach is that in order to license (89a), for example, we can derive the following expression of type VP↿NP and identify it as the antecedent for the VP-ellipsis/pseudogapping operator (whose syntactic category also needs to be changed slightly) in the target clause:

This would suffice to license (89a) ((89b) could be derived analogously).

One concern about this type of approach is its overgeneration. The present CG analysis is already very flexible, and it needs to rely heavily on pragmatic principles to rule out unacceptable examples (but note that all other alternative proposals share the same problem: they either undergenerate or overgenerate). Allowing for the possibility of unrestricted discontinuous pseudogapping would increase the set of strings that the syntax would overgenerate even more. One concrete case that the vertical-slash-based approach would overgenerate is (91).

Here, the same VP↿NP category as in (90) could be assigned to the string laughed when __ arrived. It is not immediately clear whether this example could be ruled out by any of the known pragmatic properties associated with pseudogapping.²²

While pushing the “syntax overgenerates, pragmatics constrains” approach to its limits is certainly an attractive option (especially if doing so avoids the need to introduce new theoretical machinery), we would like to offer another possibility, if only as a point of comparison for future investigations. The guiding intuition behind this alternative is the observation (which itself needs to be scrutinized, but which seems to match the overall empirical patterns we have been able to identify) that there is an intriguing overlap between the cases of discontinuous pseudogapping and patterns of (apparent) discontinuity traditionally analyzed by the wrapping operation in the CG literature.²³ Wrapping is a mechanism originally proposed by Bach (1979) and Dowty (1982) in the early Montague Grammar literature for treating discontinuous strings (such as make __ up, pull __ out in verb-particle constructions) as combinatoric units. For example, in a wrappingbased analysis, the verb-adjective pair found __ attractive in (89b) is analyzed as an “underlying” constituent, and it “wraps” around the object NP her coworker in the surface form of the sentence.

We now give a somewhat detailed sketch of this alternative, which treats discontinuous pseudogapping like that in (89) via an interaction between pseudogapping and wrapping. In contemporary TLCG, wrapping is modeled by enriching the prosodic component of the theory (roughly corresponding to PF) via the notion of “multimodality” (Moortgat and Oehrle 1994, Dowty 1996, 1997, Muskens 2007, Kubota 2010, 2014a, Mihaliček 2012). The notion of “modality” here pertains to different “modes” of composition in the prosodic component governing various reordering and restructuring operations related to surface morphophonological constituency; it has nothing to do with the notion of modality in the semantics literature. Following Kubota (2010, 2014a), we call this surface morphophonological component of grammar the prosodic algebra.

For our purposes, it suffices to distinguish between two modes of composition in the prosodic algebra: the ordinary concatenation mode (○) and the infixation mode (which we notate as ○.). Prosodic terms are ordered in the prosodic algebra by the deducibility relations between terms (where φ₁ ≤ φ₂ is to be read ‘φ₂ is deducible from φ₁’). Specifically, to model wrapping, we posit the following rule:

The intuition behind this is that when A and B are combined in the infixation mode, an expression C that combines with that unit at a later point in the derivation can be infixed in the middle by a surface morphophonological reordering operation. To refer to the deducibility relation in the prosodic algebra from the combinatoric component during the course of a derivation, we posit the following P(rosodic)-Interface Rule:

The syntactic rules are also revised to take into account the sensitivity to modes of composition (for space reasons, we only reproduce the rules for ∕ , but the rules for \ are similarly revised; the rules for ↿ remain the same as above).

In these revised rules, the modes encoded in the slashes match those that are used to combine the phonologies of the functor expressions with those of their arguments.

With this small extension, a simple wrapping example such as (95) can be analyzed as in (96).

The point here is that the (surface) discontinuous string found __ attractive behaves as a unit in the combinatoric component (motivation for this assumption comes from patterns of argument structure–sensitive phenomena such as passivization and binding; see, for example, Dowty 1982, 1997). The pseudogapping operator can then directly refer to the syntactic category and the semantics of this “underlying constituent” to supply the relevant subcategorization frame and meaning of the missing TV to the auxiliary, in exactly the same way as in the simpler examples above. Thus, (89b) is licensed as follows:

The wrapping-based alternative just sketched does not admit discontinuous constituents involving the vertical slash ↿ (such as VP↿NP), since VP∕ $ ranges over categories involving directional slashes only. With this restriction, it is predicted that discontinuous pseudogapping is possible only when the deleted discontinuous string corresponds to an “underlying” constituent in the combinatoric component involving wrapping. The data reported in the literature seem to conform to this prediction, but it remains an open question whether this analysis is adequate or whether true counterexamples can be found that would suggest that the other approach involving the vertical slash would be better.

The analysis of VP-ellipsis and pseudogapping given above is actually a bit too simplistic in assuming that there is always a syntactic antecedent that the ellipsis operator anaphorically refers to (see the side condition in (65)). This requirement is clearly too strong for VP-ellipsis and arguably also for pseudogapping. As noted by Miller and Pullum (2013), if appropriate discourse conditions are satisfied, purely exophoric VP-ellipsis is possible.

While it seems considerably more difficult to construct analogous purely exophoric cases of pseudogapping (presumably because of the requirement specific to pseudogapping that the remnant needs to be contrasted with some “corresponding” item),²⁴ as we have already noted, there are cases of pseudogapping in which there are no appropriate syntactic antecedents in the preceding clauses (Miller 2014), and also instances of split-antecedent pseudogapping, which essentially establish the same point.

While these examples clearly show that the condition encoded in (65) (which requires the existence of a syntactic antecedent) is too strong, purely interpretive approaches such as Miller’s (1990) would overgenerate radically, as Miller (2014) himself acknowledges.

We think the right empirical pattern can be captured by relaxing the condition on the VPellipsis/pseudogapping operator (reproduced in (100)) slightly, along the lines of (101).

With these conditions, the preposition mismatch case in (43), repeated here as (102), is still correctly ruled out.

The remnant PP_for forces the syntactic category of the derived auxiliary to be VP∕ PP_for, but then there is no matching syntactic antecedent in the preceding clause. Crucially, recovering the ‘speak to’ meaning of speak from the preceding clause via a purely anaphoric process (clause (101b)) is not an option either, since that meaning is associated with a distinct subcategorization frame VP∕ PP_to and thus is inconsistent with the VP∕ PP_for frame.²⁵

The revised condition in (101) is clearly in the same spirit as Miller’s (1990) selectional restriction–based treatment (see section 3.2) in embodying the intuition that essentially, (102) is ill-formed because the verb has distinct meanings depending on which of the two subcategorization frames it appears in. But it achieves the same effect by simply making the anaphora resolution process be sensitive to both the syntactic and the semantic information of the antecedent simultaneously, rather than by making the semantic restriction on the denotation of the anaphoric verb directly access the subcategorization frame of the antecedent.²⁶

The antecedentless and split-antecedent examples in (99) are no longer problematic for the revised formulation of the anaphora resolution condition in (101). In these cases, there are no syntactic antecedents matching in category with the “missing verbs.” However, unlike in the case of (102), the relevant relations appropriate as antecedents (such as ‘use’ for (99a) and ‘saw or heard’ for (99b)) are salient in the preceding discourse; moreover, there is no interference from a lexically associated conflicting subcategorization frame. Thus, anaphora is resolved by a purely semantic/pragmatic mechanism in these cases.

We believe that the above discussion has made it clear that our analysis of pseudogapping achieves better empirical coverage than any of the transformational analyses. At the same time, the flexible CG-based syntax-semantics interface enables us to formulate the restrictions pertaining to syntactic connectivity much more simply than in purely anaphoric approaches. Nonetheless, the present proposal leaves open one major issue, which we should note explicitly: overgeneration owing to the flexible architecture of CG. For example, on our account, nothing in the syntax predicts (103a) to be unacceptable. We take this to be the correct result, since the structurally parallel (103b) (=(11a)), an attested example from Levin 1979:77, is an acceptable example of pseudogapping.

But then, how can we account for the unacceptability of (103a)? Here too, we feel sympathetic to the general perspective advocated by Miller (2014), in which the syntax overgenerates somewhat wildly and additional processing-oriented and pragmatic factors constrain the acceptability of specific examples further. It is beyond the scope of this article to fully articulate these extragrammatical conditions, but we would like to note some potentially relevant factors, in the hope that our discussion will at least provide a starting point for further investigating this quite complex issue in more detail.²⁷

The acceptability of complex instances of pseudogapping (such as those in (103)) seems to be particularly sensitive to pragmatic factors such as prototypicality and plausibility of the event described by the sentence in view of general world knowledge.²⁸ For example, the intended interpretation of (103b) is presumably supported by the fact that linings are components of coats that are detachable for some types of coats, but not all. In (103a), by contrast, there is no such inherent part-whole relation between books and boxes.

Note further that the contrast in (103) becomes less clear if we manipulate certain lexical choices. (104b) is less natural than (103b) since skirts and dresses don’t normally have linings. By contrast, (104a) is more natural than (103a) since the use of the demonstratives this and that naturally invokes a contrast between the two remnants.

Given the diversity of the possible relevant factors, predicting the acceptability of specific examples in some precisely measurable manner is a huge open question, and we do not attempt to answer it here. But the overall conclusion from the above discussion should be clear: in general, one should be extra careful in assessing the acceptability of pseudogapping examples; in particular, when some example seems to sound bad, one should not immediately attribute its unacceptability to grammatical factors. Such a conclusion is justified only if the unacceptability cannot be ameliorated by carefully controlling for all possible confounding factors.