2.1 General background

Dargwa languages constitute a separate branch of the Nakh-Daghestanian language family and are spoken primarily in the central part of the Republic of Dagestan (Russian Federation). It has been traditionally assumed that different varieties of Dargwa constitute dialects of a single language. Nowadays, however, there is a consensus that it is necessary to distinguish between several Dargwa languages, the number of which ranges from six (Mutalov 2021) to fifteen (Koryakov 2021). According to Koryakov (2021), Muira Dargwa, the language under discussion, belongs to the north-central group and is one of the largest languages of the branch, spoken by approximately 40.000 people.

As for linguistic characteristics, all Dargwa languages are left-branching with SOV as the unmarked word order; it should be noted, however, that the word order enjoys a lot of freedom (Forker 2021). The languages allow for extensive pro-drop, employing null pronominals for direct objects as often as for subjects; null realization of indirect objects is also attested, although it is a much less frequent phenomenon. Typically for Nakh-Daghestanian languages, Dargwa languages are morphologically ergative, both in terms of case and (gender) agreement, and syntactically accusative. They are characterized by rich agglutinating morphology, both in the nominal and verbal domains.

The data in this article come from author’s own field work with a dialect of Muira Dargwa spoken in the village of Kalkni (Dakhadaevski District, Republic of Dagestan, Russia). All acceptability judgments were elicited between 2022 and 2023 from at least three different native speakers of the language, aged between 30 and 65, and subsequently confirmed in different sessions.

2.2 Complex predicates

In Dargwa languages, the verbal lexicon is traditionally divided into three classes of predicates:

• i. Simplex verbs

• ii. Preverb verbs

• iii. Complex verbs (aka CPrs)

The class of simplex verbs is closed and numbers no more than 200 stems. These stems have a very simple phonological structure, in most cases consisting of two or three phonemes; some of them imply the presence of a gender agreement prefix. Absolute majority of underived stems in Dargwa languages come in pairs that express the aspectual opposition between perfective and imperfective. Due to the paucity of verbs in this class, most of Dargwa verbs are either CPrs or derived via the addition of spatial prefixes (preverbs) to simple stems. The formation of both preverb and complex verbs is highly productive, the two classes comprise an ever expanding segment of the verbal lexicon.

CPrs are comprised of two parts: a light verb (LV) and a non-verbal element (NV). LVs have rather bleached semantics (causation, initiation, etc.), function as verbalizers and play a crucial role in determining the argument structure of the resulting CPrs. Most of the LVs are homophonous with simplex verbs, which bear full lexical meaning – these are traditionally called “heavy verbs”. In comparison to other languages of the Caucasus and the world, the number of LVs in Muira Dargwa is quite high – so far, no less than 20 LVs have been encountered (Kalyakin 2023). However, most of them have rather limited distribution, combining with just a few (or even one) NVs. The list of the most commonly used LVs glossed with their heavy version interpretation is presented in (4).³ In what follows, I will primarily focus on CPrs with LVs -arq’ / -irq’ ‘make, do’ and -Ø / -irh ‘be, become’.

NVs are usually represented by nominal and adjectival roots, as well as ideophones, adverbs, numerals and borrowed verbs (primarily from Arabic, Persian and Russian), which are also used as NVs. The main function of an NV is to determine the lexical meaning of a CPr. Examples of CPrs with different types of NVs are provided in (5).

Now we can turn to the structure of CPrs. Many syntactic accounts of CPr formation have productively employed a decomposed verbal structure à la (Hale & Keyser 2002). Despite the wide variety of differences, these accounts share the notion that the LV originates at a point in structure higher than the NV, see (Folli et al. 2005; Megerdoomian 2011) for Persian, (Manetta 2018; 2019) for Hindi-Urdu, and (Tabak & Demirok 2024) for Turkish.

One of the most influential analyses was developed in (Folli et al. 2005) for Persian. Under this account, argument structure of a CPr is dependent on both the LV and NV: the former determines whether an external argument will be projected, while the latter selects for the internal argument(s). On the basis of this division of labor, Folli et al. (2005) suggest that the structure in (6), where both the external and internal arguments are projected, underlies transitive (causative) CPrs.

An intransitive (inchoative) CPr is formed by substituting a transitive LV with a corresponding intransitive one. A minimal pair is presented in (7): the LV in (7a) selects for an external argument, while the LV in (7b) does not. Similar examples from Muira Dargwa are shown in (8).

Due to a number of similarities between Persian and Muira Dargwa CPrs, I partially extend the analysis of Folli et al. (2005) to Muira Dargwa, treating the LV as an overt realization of the v head. I depart from their analysis in a number of important details, however. First, I take the NV to be the realization of a bare root (Marantz 1997) that corresponds to a phonological word. This explains the “deficient” morphosyntactic nature of NVs, which can be observed in a number of respects (Kalyakin 2023; 2025b). Most importantly, despite the fact that “nominal” NVs superficially resemble bare noun stems and that absolutive case is zero-marked, these NVs bear no case and are invisible to case calculus. This effectively explains the presence of another zero-marked nominal, which actually bears absolutive case, and the inability of these NVs to participate in the agreement relation – as (9) shows, the NV gap ‘praise’ is not in absolutive, since there is another zero-marked DP Rasul with which the verb agrees in gender and number. Besides that, the element cannot host modifiers (10) and it is not referential (11).

Furthermore, I adopt the now widely accepted view that roots are devoid of syntactic information, viz. they do not project and do not subcategorize (Alexiadou 2014; De Belder & van Craenenbroeck 2015; Merchant 2019). With respect to the position in the syntactic structure, roots can be attached to v (12), or merged with the direct object (13) – the exact position is determined by the root’s conceptual content. The intuition here is that the roots that modify the event introduced by v, for example imposing a manner on an activity event ( $\sqrt{\text{RUN}}$ , $\sqrt{\text{HAMMER}}$ ), are attached directly to v, while roots that describe the end state that is associated with a direct object of a change-of-state event ( $\sqrt{\text{BREAK}}$ , $\sqrt{\text{OPEN}}$ ) are attached to the direct object DP (Marantz 2009a; b; 2013; Irwin 2012; Levinson 2014). This dichotomy will prove important for the current study of vVPE, see §4.1.

Finally, concerning the structure of the verbal domain, I follow (Pylkkänen 2008; Harley 2013) and postulate two functional heads, Voice and v, which share functions that are usually associated with a single head, e.g. v in (Chomsky 1995). Voice is endowed with the ability to host external arguments in its specifier, such as agents or causes, while v functions as a categorizer and introduces eventive semantics.

Given all this, the structure of a causative change-of-state CPr, e.g. wana agr-arq’- ‘to warm smth. up’ in (3), will look as in (14). As can be seen, the structure here is an extended version of the tree (13); I will return to the semantic interpretation of this structure in §4.2. The inchoative counterpart of this CPr will have the same structure, except that the Voice head disallows merging of a specifier, viz. it does not introduce an external argument and does not bear any thematic information.

3.1 Pragmatic control

In their seminal work, Hankamer & Sag (1976) observe that ellipsis must have a linguistic (syntactic) antecedent rather than a merely contextual one. In other words, ellipsis does not allow what Hankamer & Sag call pragmatic control, see (15a). On the other hand, interpretation of deep anaphora depends on the particular discourse model, so pragmatic control is possible, as illustrated in (15b).

Similarly to English VPE, Muira Dargwa vVPE prohibits pragmatic control. Example (16a) shows that despite the presence of a non-linguistic antecedent, it is impossible to delete the internal argument χureg ‘food’ along with the NV wana ‘warm’ of the CPr wana agr-arq’- ‘to warm smth. up’. Note that the omission of only the internal argument is possible (16b), so it is reasonable to assume that in this case we are dealing with deep anaphora, though the pronoun has no phonological realization.

An additional example with the CPr gap agr-arq’- ‘to praise’ is provided in (17).

The relevance of the diagnostic to differentiate between surface and deep anaphora has been challenged in literature, see (Schachter 1977; Merchant 2005; Weir 2014) for critical discussion. In some cases, surface anaphora may be felicitous without a linguistic antecedent, and deep anaphora may fail to find a contextual antecedent. Nevertheless, so far all the judgments I have obtained show quite a strong contrast. Moreover, I have encountered no instances of vVPE without a linguistic antecedent.

3.2 Bound variables

Johnson (2001) points out that anaphoric elements in the antecedent and ellipsis clauses can be interpreted by two different mechanisms – coreference and binding. Under coreference, strict identity emerges – pronouns pick the same referent in the antecedent or refer to some other individual in the discourse. Binding, in turn, interprets each pronoun as a variable, which acquires its value from a c-commanding DP (usually, subject) in the local domain (usually, clause). Binding is said to produce sloppy identity readings.

English VPE is known for permitting both sloppy and strict identity readings, as exemplified in (18), see (Kehler 2002) for discussion and important qualifications. That is, the sentence does not commit one to thinking that John slapped Fred; it may well be the case that John slapped himself. On the other hand, deep anaphora, having no complex internal syntactic structure and thus no elements to be interpreted as variables, is unable to provide sloppy identity readings.

Muira Dargwa vVPE aligns with English VPE once again. Examples (19) and (20) demonstrate that both strict and sloppy readings are readily available.⁴

Like virtually all languages of the Nakh-Daghestanian family, Muira Dargwa has no dedicated third person pronouns. Instead, a number of demonstrative pronouns are employed. Importantly, as (21) shows for the demonstrative il ‘this, close to the addressee’, these pronouns prohibit sloppy identity readings. Given the contrast between (19)–(20) and (21), it is reasonable to conclude that the former examples are derived via a different mechanism, namely, ellipsis.

As was the case with pragmatic control, the availability of sloppy identity readings has been criticized for not being exclusive to ellipsis (Hoji 1998; Merchant 2013a). So, even though the diagnostic seems to be useful at the moment, from a cross-linguistic perspective, its reliability should be taken with a grain of salt.

3.3 Inverse scope

Quantification elements inside the ellipsis site can take wide scope over the operators outside the elided constituent. Such inverse scope readings are expected if we deal with quantifier raising out of a regular, though unpronounced, syntactic structure. On the other hand, pronominal anaphors, being an instance of surface anaphora, are not expected to allow inverse scope, cf. (22).

As can be seen from (23) and (24), which are parallel to (22a), Muira Dargwa vVPE mirrors English VPE by allowing for both surface and inverse scope readings.

3.4 Extraction

Extraction seems to be one of the most reliable diagnostics that help identify an active, though invisible, syntactic structure. The logic behind this diagnostic is very simple: if an ellipsis site contains internal syntactic structure, it should be possible to extract something from that structure before the ellipsis occurs, thus saving the moved constituent from being deleted. All types of extraction from ellipsis sites – A- (Aelbrecht 2010), A′- (Schuyler 2001), and X⁰-extraction (Goldberg 2005) – are attested and discussed in the literature.

Example (25) illustrates A′-movement out of the elided VP. As can be seen in the structure of the second clause in (26), the wh-phrase is base-generated as an internal argument of the verb and later displaced to the left periphery of the clause, vacating the ellipsis site.

In turn, Null Complement Anaphora, providing no position to move out from, fails to license the extraction of the object of an understood complement predicate, cf. (27). The same goes for pronominal anaphors like do it in English, cf. (28).

Example (29) from Muira Dargwa demonstrates that constituents can escape ellipsis by undergoing rightward movement to the postverbal position (this movement is driven by information structure purposes, see Forker 2021).⁵ In such cases, vVPE starts to resemble pseudogapping in English, which is often analyzed as a subtype of VPE (Jayaseelan 1990), but see (Johnson 2009) for a dissenting view.

A′-movement can also target the position somewhere above the verbal domain (30a). If the internal argument is not contrastive focus-marked, then there is nothing to sanction this movement and vVPE associated with the pseudogapping-like derivation fails, as (30b) exemplifies. This fact will be essential in proving that the process responsible for the non-pronunciation is vVPE, I will return to it in §3.6.

Muira Dargwa wh-phrases, similarly to their English counterparts, can be moved out of the silent constituent, cf. (31).

Finally, A-movement out of the ellipsis site is also permitted, as (32) shows. Some additional examples of this type of movement will be presented in §5.

3.5 Agreement

All the examples presented so far contain additional evidence for surface anaphora – agreement. Since ellipsis is an operation that removes only phonological feature matrices of terminals, while their formal feature matrices are left untouched, elements inside the ellipsis site can participate in formal operations that follow the process of deletion, though see (Aelbrecht 2010; Baltin 2012). Examples (29), (30a) and (32) above show that LVs agree with missing internal arguments in gender and number. Person agreement is also available (33), see also (17).

3.6 Against alternative analyses

In the previous subsections, we saw that all diagnostics converge on the fact that clauses missing both an internal argument and an NV are derived via ellipsis, or, in other words, they instantiate surface anaphora. However, as the discussed properties are not limited exclusively to vVPE, it is worth exploring whether the phenomenon we are dealing with is really vVPE, rather than, for example, Argument Ellipsis (AE), a more targeted ellipsis process which can elide syntactically fully present individual arguments independently. For instance, Hoji (1998) shows quite convincingly that what appear to be sloppy identity readings are not definitively characteristic of VPE in Japanese. Moreover, the very existence of VVPE, a phenomenon which seems to be quite similar to vVPE, has recently attracted a lot of skepticism (Landau 2020a; b).

Returning to the baseline example, the second clause of (3), repeated here as (34), could in principle have one of three possible derivations represented in (35).⁶ It has already been established that (35c) cannot be the correct analysis. Now, it remains to provide evidence that (35b), where the two elements undergo deletion separately, is also on the wrong track, thus leaving (35a) as the only right way.

Usually, it is quite challenging to isolate vVPE from AE, and, as mentioned before, because of this there is a lot of disagreement about how verb-stranding constructions are derived. The reasoning here should be as follows: if there is a constituent which cannot be deleted independently in the language but in some cases can nevertheless undergo ellipsis along with the internal argument(s), then it must have been silenced by virtue of VPE. The archetypal candidates for this role of non-deletable constituents are adverbs, so the interpretive possibilities of these elements within ellipsis sites should be diagnostic of VPE, see (Goldberg 2005; Simpson et al. 2013; Funakoshi 2016; Manetta 2018; Haldar 2021; among many others).

Fortunately enough, in Muira Dargwa, not only are adverbs non-deletable on their own, but also NVs. For reasons that will become clear later, I defer the discussion of the former until §4.1, focusing now on the deletion of NVs. The following paradigm demonstrates that it is permissible to silence either the internal argument and the NV (36a), or the internal argument alone (36b). Crucially, it is impossible to silence the NV independently (36c).

Given the impossibility of eliding an NV alone, the analysis (35b) must be incorrect, as it treats the non-pronunciation of an internal argument and an NV as the result of application of two independent operations, thus predicting the availability of NV-drop in (36c), contrary to fact. However, in §3.4, we have encountered examples of sentences where only the NV was left unpronounced, cf. (37). These data, I believe, not only do not pose a problem, but actually provide additional evidence for the existence of vVPE. An important difference between (36c) and (37) is that in the former example, the internal argument of the ellipsis clause is the same as that of the antecedent clause, while in the latter example they are contrasted. So, the possibility of silencing an NV alone strongly correlates with the presence of a contrastive focus stress on an internal argument, which should allow it to move out of the ellipsis site, surviving silencing. If, however, an internal argument is unable to escape the ellipsis site, as in (36c), where the direct objects of both clauses are the same, its fate is to be unpronounced. These facts, I believe, allow us to conclude that (37) must also be derived via vVPE.

Now, one might wonder why the deletion of an NV is not obligatory and how sentences like (36b), in which only the internal argument is missing, are derived. I suggest that, despite the superficial similarity, (36b) might actually be derived via AE, rather than vVPE. Evidence for yet another ellipsis mechanism is partly negative: I was unable to find any clues pointing to the possibility of an NV incorporating into v or any other evidence of an NV moving out of the domain of vVPE. Besides that, the fact that AE may be responsible for object-drop is confirmed by the presence of sloppy identity readings (38).⁷

Simpson et al. (2013) provide data from three South Asian languages – Bangla, Hindi-Urdu and Malayalam – and demonstrate that VVPE and AE can coexist peacefully within the same language. So, there is nothing exceptional about the fact that both processes are at work in Muira Dargwa, although in some contexts it may be extremely difficult to differentiate between the two. For instance, (39) with a simplex predicate may well be derived either by vVPE or AE, as the result of applying these processes would be indistinguishable, see also §4.1.

In this subsection, I have shown that CPrs, and more specifically the impossibility of eliding NVs individually, provide crucial evidence for the existence of vVPE in Muira Dargwa. An important lesson from this discussion is that NVs, once their morphosyntactic status has been established, can be productively employed to diagnose (different types of) VPE in other languages as well.⁸ To my knowledge, this has only been done so far for Bangla in (Haldar 2021). Crucially, as can be seen from the paradigm presented in (40)–(43), Bangla demonstrates exactly the same pattern as that observed in Muira Dargwa, thus providing additional support for the reliability of the NV-drop test.

4.1 Licensing requirement

The first part of (44) refers to the local, idiosyncratic features of the head that triggers the non-pronunciation of its complement. This requirement can be captured in many different ways. In his now mainstream approach, Merchant (2001) posits a lexical feature [e] (or even a family of features) which instructs PF to block pronunciation of the complement phrase (the semantics of the feature will be given in §4.2). The [e] feature, however, can originate on a fixed number of heads: for English, the list includes at least C, T (or Asp), Voice and D.

In his discussion of English VPE, Merchant (2013c) shows that VPE is the deletion of the complement of Voice; this is schematically represented in (45). Since the Voice head bears the [e] feature and it is thus not located within the ellipsis site, VPE is predicted to tolerate voice mismatches; this will be discussed in more detail in §5.

One more consequence of positing the [e] feature on Voice concerns the possible interpretations of the modifier again. Von Stechow (1996) observes that in German, when wieder ‘again’ modifies a transitive predicate, it is ambiguous between two readings: repetitive and restitutive. The ambiguity, as von Stechow claims, is syntactic in nature and it is due to the ability of the adverb to attach in two different structural positions – as an adjunct of either vP or VP – where it introduces a presupposition that the eventuality denoted by the modified constituent has occurred before. Johnson (2004) demonstrates that the same ambiguity also arises in English, see (46).

That is, the ambiguity is between the two structures: (47), where again is adjoined to the vP that denotes the event of the causation of the state denoted by its complement VP, and (48), in which again modifies the stative VP that denotes the result state.

Johnson (2004) further notes that von Stechow’s approach to the ambiguity of again can be fruitfully implemented to diagnose the constituent that is silenced under ellipsis. It turns out that only repetitive readings are available in English VPE, cf. (49).

The contrast is totally expected if VPE deletes an entire vP. The adverb can survive ellipsis only if it merges with the vP (50), giving rise to the repetitive reading (49a). In turn, if again is adjoined to the VP, as in (51), it must be deleted, thus making it impossible for the restitutive reading (49b) to arise.

Turning now to Muira Dargwa, the again-test can be employed to adjudicate the size of the phrase that vVPE targets. Under the current analysis of CPrs, presented in §2.2, the distinction between eventive and stative components is clear. For instance, the NV hark ‘open’ of the CPr hark agr-arq’ ‘to open’ denotes the result state of being open, while the LV agr-arq’ ‘make’ expresses the causing event. When ʡibrra ‘again’ is used in the context of vVPE, quite surprisingly, both readings are permitted: the repetitive reading in (52a) presupposes the previous existence of the opening event, while the restitutive reading in (52b) has the presupposition that there previously existed the state of the door being open.

Given the availability of both repetitive and restitutive interpretations of ʡibrra ‘again’ in the context of vVPE, it is reasonable to conclude that vVPE in Muira Dargwa is smaller than VPE in English. More concretely, I suggest that vVPE targets the complement of v. This is schematically represented in (53), where the [e] feature is located on the v head.

This lower-than-average size of the VPE has a number of consequences. Here, I address only two of them, postponing discussion of others until §5 (some other languages with ellipses of apparently similar size will be discussed in §6). First, it should become evident that the widely employed adjunct-test is of no value in differentiating vVPE from AE. If, as it is traditionally assumed, low adjuncts are attached to vP/VoiceP, Muira Dargwa vVPE should never be able to target them, as they are attached to a projection outside the ellipsis site. AE, in turn, simply deletes individual arguments and cannot target adjuncts. In both (54) and (55), missing adjuncts are not syntactically present, as they cannot be interpreted in the elliptical clause. So, as expected, they are missing completely, and not deleted by vVPE.

Second, recall the discussion of the structure of CPrs in §2.2. It was suggested that there exist two structural positions where NVs – which correspond to bare roots in the narrow syntax – can be merged. So far, the discussion has been limited to roots that merge with the direct object and describe the end state that is associated with a direct object of a change-of-state event. There are, however, roots that modify the event introduced by v, for example imposing a manner on an activity event, – these are argued to be attached directly to v. Consequently, under this view, the proposed analysis predicts an important ellipsis asymmetry: direct object-adjoined NVs must be targeted by vVPE, while v-adjoined NVs are expected to be exempt from it. The reason for this is that the [e] feature instructs the non-pronunciation of the complement of the head it is adjoined to. So, if the feature is located on the v head, as argued here, vVPE targets the complement of v. Crucially, unlike direct object-adjoined NVs, v-adjoined NVs are never contained within the expected ellipsis site and are predicted to be non-elidable. This is schematically represented in (56).

It turns out that this prediction is borne out partially. The status of vVPE in these contexts is subject to some inter-speaker variation. For two consultants, omitting an NV may result only in somewhat degraded acceptability.⁹ For now, I do not have a good explanation for this, so I leave it for further research (though see footnote 9). It should be emphasized, however, that the unacceptability of these examples is not due to the elision of just the NV – in §3.6 it was demonstrated that, under certain circumstances, it is indeed possible for vVPE to elide only the NV.

4.2 Identity requirement

In the previous subsection, it was established that the head which hosts the [e] feature and licenses ellipsis is v. It was also noted that the [e] feature not only gives PF the instruction to not pronounce the complement of the head it is adjoined to, but it is also used to capture the identity requirement, introduced in (44). Obviously, the identity to an antecedent may be of different nature, see (Merchant 2013b) for an overview of existing approaches. In his seminal work, Merchant (2001) argues for a semantic identity requirement that he calls e-givenness, see (61) for the definition.¹⁰ The [e]-feature, as can be seen in (60), is endowed with the presupposition that the complement must be e-given. Thus, [e] plays a dual role, acting as a mediator between PF and LF.

To illustrate how it all works, the interpretation of CPrs must first be elaborated. As discussed in §2.2, I follow (Marantz 2013) and much subsequent work by taking roots to be syntactic and semantic modifiers describing a particular event. With respect to DPs, I assume that some DPs are interpreted purely as entities, while others can be interpreted as events. A DP can be eventive either by virtue of its inherent meaning or because it is coerced into being eventive, specifically a state. This coercion is the result of the application of the rule (62), see (Marantz 2009a; b).

As for the semantic composition, besides strictly compositional rules, such as Functional Application and Predicate Modification (Heim & Kratzer 1998), the elements can also be combined by the non-compositional rule of Event Identification (Kratzer 1996), defined in (63).

Taking this into account, but nevertheless simplifying somewhat, the denotation of (64) with a change-of-state CPr can be represented as in (65), where P is a variable over eventive predicates, e a variable over events, and x a variable over individuals.

Now we are in a position to examine how the identity requirement is checked under vVPE. Let us look at the baseline example (3), repeated here as (66).

In order for the mutual inclusion conditions to apply, the focus marked elements undergo λ-type shifting operation. In the case of (66), however, there are no such elements. The first part of the definition in (61) requires that the denotation of the target include the denotation of the antecedent. (67) shows that the denotations are identical, so the requirement is trivially satisfied.

The second part of (61) reverses the relationship between the target and antecedent. Now, the denotation of the former must be included in the denotation of the latter. Obviously, as the denotations are identical, this is also the case.

Both parts of the identity requirement are satisfied, and the ellipsis is licit. In these examples, F-closure doesn’t play any significant role, as nothing in the target or antecedent clauses is F-marked. Let us now turn to examples where it has some weight.

Sentences that resemble pseudogapping, discussed in §3.6, pass the identity check in a similar way. Consider (69). In both (70) and (71), the contrastive focus marked internal argument in F-closed clause is abstracted over by λ-type-shifting, thus making the clause a superset of the denotation of the other clause. So, the ellipsis here is also licit.

In this subsection, I have demonstrated that due to the rather small amount of the deleted material, the identity requirement can be easier to satisfy. In the majority of cases, the calculation is trivial. With this, however, there emerge a number of consequences. I now proceed to discuss some of them.

5.1 v and ellipsis: inchoative-causative alternations

In §4.1, it was noted that, according to Merchant (2013c), VPE in English targets the vP, leaving outside the Voice head, which is responsible for hosting the voice features. As (72) shows, this correctly predicts the grammaticality of voice mismatches between the target and antecedent clauses, see (Poppels & Kehler 2019; 2025) and references therein for a more in-depth discussion and important qualifications.

For ellipses that target constituents larger than the vP, voice mismatches result in ungrammaticality; this is illustrated in (73) for sluicing.

Merchant, following a long tradition in literature on argument structure, associates causative semantics with the v head (Chomsky 1995; Hale & Keyser 2002; Pylkkänen 2008), which is immediately below Voice. So, another important prediction of his analysis is the impossibility of transitivity alternations under any type of ellipsis in English, even VPE. Example (74) demonstrates that the prediction is borne out (see, however, §6). Indeed, this observation is not new, but Merchant seems to be the first to give a principled explanation for the ban on such ellipses.

Returning now to Muira Dargwa, if the analysis proposed here is on the right track, vVPE is expected to tolerate transitivity alternations, as the operation does not target the v head, which is responsible for transitivity and eventive semantics. Let us look at examples (75) and (76). In both cases, CPrs denote the same warming up event. However, one important difference between the target and antecedent clauses in these examples is that one of the CPrs introduces an agent argument, while the other one does not. This difference, as argued in §2.2, is determined by which Voice and v heads are selected. In the case of inchoative CPrs, Voice is specified so as not to introduce an external argument, and v is such that it simply denotes a change-of-state event, no external force is presupposed. In turn, causative CPrs are formed with the help of a thematic Voice head that must have a DP in its specifier, and a causative v which implies the presence of an agent. Despite this mismatch, ellipsis in both (75) and (76) is licit. So, the initial prediction that mismatches in the content of v should not play a role in vVPE is borne out.

It should be noted that the internal argument of an inchoative CPr is considered to move from the vP to the surface subject position, leaving a copy inside the ellipsis site. Under the assumption that A-movement reconstructs, as it has been argued to do for reasons of scope (May 1977), the satisfaction of the identity requirement appears to be as trivial as in the example (66), discussed in the previous section. That is, the identity check between the target and its antecedent in (75) and (76) is exactly as in (67) and (68), modulo lexical entries.

5.2 Going up: causativization

Causativization is the only productive process of valency change in Dargwa languages. In all languages of the group, both morphological and periphrastic causative constructions are attested. Here, I will discuss only the former, since the latter have not been thoroughly investigated yet and, it seems, will not affect the current discussion in any meaningful way.

Morphological causative is formed by adding the affix -iq directly to the verbal base. The process of causativization is highly productive – causatives are easily formed from verbs with all types of argument structure: intransitive (both unaccusative and unergative), transitive, ditransitive and affective. The subject of the causative construction is always in the ergative case. The arguments of the causativized verb are expressed exactly the way they are in the non-causative correlate, with the only exception of the external argument of a causativized transitive/ditransitive predicate – its case marking is changed from ergative to inter-lative. Example (77) illustrates how the construction is formed on the transitive base.

Since (Pylkkänen 2008), morphological causative formation is regarded as adding a dedicated causative head on top of either vP or VoiceP. So, it is predicted that the exact type of Muira Dargwa causative is irrelevant, because in both scenarios vVPE is too small to target the relevant fragment of the structure.¹¹ As can be seen in (78), LVs, being outside the ellipsis site, need not match – it is possible to have a causative LV in the antecedent clause and a causativized inchoative LV in the target clause, or vice versa.

Moreover, as was the case with inchoative-causative pairs, discussed in the previous subsection, transitivity mismatches are also tolerated, see (79) and (80). Thus, the proposed analysis makes correct predictions here as well.

5.3 Going down: antipassivization

Dargwa languages are characterized by the presence of an antipassive construction, which is quite rare in Nakh-Daghestanian languages, see (Comrie et al. 2021) for an overview. In this construction, the logical object of a transitive predicate is realized not as a direct object, but as an oblique argument, adjunct, or is not expressed at all (Polinsky 2017). In Dargwa languages, antipassive is quite peculiar in that it does not have any special morphological marking. Semantically, it is limited to imperfective aspect and is used to express some kind of habituality or iterativity. Example (81) shows an ergative construction and its corresponding antipassive correlate. Superficially, (81a) and (81b) differ in case marking of the arguments and agreement on the lexical verb – these are the main morphosyntactic reflexes of the antipassive contruction.

Trying to capture these and some other properties of the antipassive construction, Kalyakin (2024) proposes the structure (82). In this structure, the internal argument is licensed by a special functional head ap (freely Merging Case-assigning head, Yuan 2018), which is located rather low – just above the internal argument. This head is aspectual in nature and is endowed with the ability to assign inherent ergative case. The external argument is licensed by the intransitive Voice head, so it receives structural absolutive case. This derivation is argued to explain the lack of agreement with the internal argument and the fact that the construction is limited to the imperfective aspect.

The semantics of ap have yet to be studied. Though, Kalyakin (2024), following Wharram (2003) and Deal (2008), suggests that the preliminary denotation may be along the lines of (83). As can be observed, the operator mediates the composition of a property-type object with a verb, providing existential closure of the former. This approach correctly predicts obligatory narrow scope and non-specific (de dicto) interpretation of the internal argument.

Given the low structural position of the ap head, it is worth investigating how antipassivization interacts with vVPE. First, witness (84), which demonstrates that the CPr dawk agr-irq’- ‘to repair’ may be used in the antipassive construction.

Turning now to vVPE, ellipsis is infelicitous if an antipassive construction is anteceded by an ergative construction (85), or vice versa. This seems to be expected, as the semantic contribution of the antipassive operator should make the denotations of the two clauses sufficiently different.

This, however, is not the full story. It turns out that vVPE is prohibited even in those cases where both clauses contain antipassive predicates, cf. (86).

No matter how mysterious this situation may seem at first glance, I would like to claim that this is also to be expected. Recall the discussion of unergatives in §2.2 and §4.1. It was suggested that in the case of these predicates, roots impose a manner on an activity event and are attached directly to v. With this, NVs of these CPrs cannot be targeted for deletion by vVPE. In turn, antipassives are usually considered to also denote activity events (Polinsky 2017) – it seems to be the case for Muira Dargwa as well. Given this parallel between unergative and antipassive predicates, I propose that under antipassivization roots that are usually attached to direct objects are coerced into being exceptionally merged as modifiers of the v head. Consequently, as in the case of regular v-attaching NVs, vVPE is unable to target the NV of an antipassive CPr. If this is on the right track, the unacceptability of (85) must be not due to the lack of identity between the antecedent and ellipsis clauses, as initially suggested, but rather because the NVs in antipassives resist non-pronunciation for structural reasons – they are too high to be deleted.¹²

5.4 vVPE and current theories of ellipsis

Now, having explored how vVPE interacts with argument structure alternations, it is worthwhile to discuss what existing theories of ellipsis have to say about the observed non-trivial empirical picture. The analysis proposed above was based on Merchant’s (2001; 2013c) assumptions about the crucial role of the size of parallelism domains. Specifically, it has been argued that ellipsis tolerates mismatches between the ellipsis clause and the antecedent clause, provided that the relevant head is located outside the constituent targeted for deletion. In Muira Dargwa, due to the lower structural position of the [e] feature, vVPE allows for a greater number of mismatches than, for example, VPE in English.

Other approaches that do not appeal to the size of the elided constituent seem to face serious problems in explaining the availability of the indicated mismatches in vVPE. For some theories that posit only semantic identity based on entailment relations (Dalrymple et al. 1991) and for inference-based (Culicover & Jackendoff 2005) theories the puzzle is why transitivity mismatches should, in principle, be allowed, since inchoative and causative clauses are not mutually entailing (the denotation of the former is a proper subset of the denotation of the latter) and do not allow for the relevant inferences.

A similar problem holds for discourse-based theories of mismatches under ellipsis (Kehler 2002). According to Kehler’s (2002) approach, the (un)availability of certain mismatches depends on discourse coherence relations between the clauses that contain the antecedent and elided material. In his discussion of voice mismatches under VPE in English, Kehler shows that two families of discourse relations are relevant: the Resemblance family of relations and the Cause-Effect family. Resemblance relations allow sentences of a discourse to be connected by emphasizing “commonalities and contrasts among corresponding sets of entities and relations” (Kehler 2002: 15). In turn, utterances that stand in a Cause-Effect relation require “a path of implication connecting a pair of propositions” in the two sentences (Kehler 2002: 20).

Kehler argues that when there is a voice mismatch in VPE, sentences where there is a Cause-Effect relation between antecedent and ellipsis site are licit, while sentences where there is a Resemblance relation are illicit. The contrast can be observed in (87).

Despite considerable success in explaining the (un)availability of voice mismatches in VPE, Kehler’s theory faces serious challenges when it comes to explaining the behavior of Muira Dargwa vVPE. According to Kehler, sentence (88) produces ill-formed discourse: the expected Resemblance relation cannot be established because the agent of the causative clause has no correspondence in the inchoative clause. Contrary to expectations, all Muira Dargwa speakers I have consulted find (88) fully acceptable.

The acceptability of vVPE in (88) and the contrast between (87a) and (87b), however, is completely expected for Sailor’s (2014) structural approach to featural mismatches in VPE. Building on (Merchant 2013c), Sailor contends that the most immediate factor dictating the (un)availability of mismatches is syntax, not discourse coherence. Based on this, he considers VPE to be a non-uniform operation: in environments that allow voice mismatches, the Voice head is outside the ellipsis site (low VPE); in environments that prohibit voice mismatches, Voice is inside the ellipsis site (high VPE). Thus, the contrast between (87a) and (87b) boils down to the type of VPE employed: low VPE in (87a) and high VPE in (87b). For this approach, therefore, the acceptability of (88) is anticipated, since v – the head responsible for transitivity and eventive semantics – is outside the ellipsis site, and thus free to differ featurally from its antecedent. So, I suggest that the data from Muira Dargwa vVPE provides novel evidence in support of Sailor’s (2014) structural approach, further expanding the typology of VPE with extra low VPE – a subtype of VPE that targets the complement of v.

Finally, it is important to emphasize that this line of analysis makes a number of predictions about possible featural mismatches in ellipses of different size. Merchant (2013c) claims that there is no language that would be an inverse of English allowing voice mismatches in TP, but not VP, ellipses. We can extend this logic and suggest that it is impossible to imagine an inverse of Muira Dargwa – a language that would not tolerate transitivity mismatches in vVPE, but allow for them in, for example, sluicing. Additionally, it is predicted that if an ellipsis operation allows for transitivity mismatches between the antecedent and ellipsis clauses, it should also allow for voice mismatches. Unfortunately, this prediction is hard to test on Muira Dargwa data, as the language does not have rich voice morphology. So, it is worth identifying some other languages with a similar ellipsis process and investigating the patterns of available featural mismatches.