In this paper I will argue for the configurational view of the Adjunct Condition, based on new evidence from Balkar (a dialect of Karachay-Balkar, a Turkic language spoken in the Republic of Kabardino-Balkaria, Russia) with some discussion of English. For Balkar, I will use data collected through fieldwork.1 For English, I will reconsider the original data from Truswell (2007) and some new data collected through elicitation.2

The Adjunct Condition was originally proposed by Huang (1982: 497–499, 503–514, see also Paducheva & Zaliznyak 1979), and can be illustrated by the following contrast:

In (1a), a noun phrase which song is extracted from an ing-clause [listening to ___], which serves as the complement of liked, and the sentence is fine. In (1b), the same noun phrase is extracted from the same ing-clause, but the ing-clause is an adjunct, and the sentence is considerably worse.

In both cases the moved constituent is the same (the noun phrase which song), and the clause it is extracted from is the same ([listening to ___]). The only difference is the syntactic position of the clause: complement in (1a) vs. adjunct in (1b).

Since the use of the terms adjunct and complement varies throughout the literature, before discussing the Adjunct Condition in more detail, I will introduce several descriptive definitions that this paper will rely upon.

First, let us call a maximal projection a constituent that does not project its category to its mother, in other words, the constituent that immediately dominates it is a projection of its sister. For example, in both (1a) and (1b) the ing-clause is a maximal projection (it does not project its category). In the trees below, I will notate phrases (non-heads) that are maximal projections as XP, phrases (non-heads) that are not maximal projections as X′ and heads as X. I will also assume that whenever α and β are merged, at least one of them projects.

Second, let us call a modifier or a semantic modifier a special type of maximal projection, as opposed to an argument or a derived specifier. There is a variety of syntactic and semantic heuristics that distinguish modifiers from other types of maximal projections (see Melčuk 1974; Kibrik 1977; Jackendoff 1977; Pollard & Sag 1987; Grimshaw 1990; Ackema 2015: and others).3 Modifiers are optional and usually iterative (that is, they can be stacked if their meaning allows it). They do not satisfy an EPP-feature of their sister, nor do they fill any argument slots, and they are not selected. Their interpretation is independent from that of the main clause. The semantic contribution of a modifier is solely determined by its own meaning, not by the main predicate (modifiers are not assigned a theta-role). Thus, the ing-clause in (1b) is a modifier (it is optional, can be iterated, is not assigned a theta-role and doesn’t occupy an argument slot), while the ing-clause in (1a) is not (it is not optional, cannot be iterated, is assigned a theta-role and occupies an argument slot of the verb like).

Broadly speaking, there are two views on the status of modifiers. According to the semantic view (see e.g. Kibrik 1977; Chomsky 1995; Johnson 2003; Truswell 2007; Bošković 2020), the special properties of modifiers follow solely from their semantics, namely, from the way the meaning of a modifier is composed with that of the main clause. According to the syntactic view (see e.g. Jackendoff 1977; Chomsky 1981; 2004; Stepanov 2007; Hunter 2010; 2015; McFadden & Sundaresan 2019), the special properties of modifiers follow from the fact that they are integrated into the main clause by a special syntactic rule. Stepanov (2007), building on Lebeaux (1991), suggests that all semantic modifiers are always merged late, after all the other operations have been performed (including movement). Chomsky (2004) and Hunter (2010; 2015) propose special structure building operations: Pair Merge and Insert respectively. McFadden & Sundaresan (2019) argue that semantic modifiers select their sister. The choice between the semantic and the syntactic view of modifiers is not crucial for the purposes of this paper, as the proposed theory is compatible with either of them (see section 1.3 below).

Finally, in what follows I will reserve the terms ‘complement’ and ‘adjunct’ to refer exclusively to the structural position of a maximal projection within the main clause (see Johnson 2003). A complement will be understood as any maximal projection (semantic modifier or not) whose sister is a head. Meanwhile an adjunct will be understood as a maximal projection which is a semantic modifier and whose sister is a phrase. Thus, in (1a) the ing-clause is complement, while in (1b) the ing-clause is an adjunct, as is illustrated by the structures in (2).

It should be noted that in the X-bar theory (Jackendoff 1977; Chomsky 1981) it is stipulated that all modifiers are adjuncts. In other words, in X-bar theory by an additional assumption all semantic modifiers merge with a phrase (there is no top left cell in (3)). Meanwhile, in Bare Phrase Structure (Chomsky 1995) there is no such assumption: semantic modifiers can merge either with a head or with a phrase. In what follows, I will crucially assume the latter view: all cells in (3) can be instantiated.

Let us come back to the Adjunct Condition and the contrast in (1). It is obvious that the contrast is due to the status of the ing-clause (that is the only thing that’s different between (1a) and (1b)). There are two respects in which the ing-clause in (1a) is different from the ing-clause in (1b). First, the ing-clause in (1b) is a modifier, while the ing-clause in (1a) is not. Second, the ing-clause in (1b) is merged with a phrase, while the ing-clause in (1a) is merged with a head.

As a result, all the existing accounts of the Adjunct Condition can be classified into two camps: modifier accounts vs. configurational accounts. Modifier accounts (Chomsky 2004; Truswell 2007; Stepanov 2007; Hunter 2010; 2015; Bošković 2020: among a few others) capitalize on the modifier vs. non-modifier distinction. Configurational accounts (Huang 1982; Uriagereka 1999; Nunes & Uriagereka 2000; Johnson 2003; Sheehan 2010) capitalize on the structural distinction: merged with a phrase vs. merged with a head.

Modifier accounts formulate the Adjunct Condition as ‘all semantic modifiers are opaque’ (the top row in (3)) and argue that this is due to modifiers being integrated into the clause via a special syntactic or semantic rule (depending on the theory of modification).

Semantic modifier accounts (Truswell 2007 and Bošković 2020) claim that the relevant special property of modifiers is in how they are incorporated into the main clause at the semantic level. Modifiers are interpreted conjunctively and combine with the matrix clause via Predicate Modification, which is assumed to make them opaque for movement.

Syntactic modifier accounts claim that the relevant special property of modifiers is in how they are incorporated into the main clause at the syntactic level. As mentioned above, according to Stepanov (2007), all modifiers are merged after movement and are hence opaque. Chomsky (2004) and Hunter (2010; 2015)’s special syntactic operations for modifiers (Pair-Merge and Insert) are formulated in such a way as to prevent modifiers from being transparent. McFadden & Sundaresan (2019) argue for a theory of opacity for movement based on selection and Agree. According to them, since modifiers select their sister, they must be opaque.

Configurational accounts, on the other hand, formulate the Adjunct Condition as ‘all maximal projections merged with a phrase are opaque’ (the right column in (3)). This paper argues that this is, in fact, the correct generalization. More precisely, the paper argues for a particular configurational account called the Spell Out Theory or Single Root Derivation, developed in Privoznov (2021) (see also Johnson 2003). The two main empirical claims of the Spell Out Theory are given in (4).

From (4) it follows that all adjuncts, as defined section 1.2, must be spelled out.4 They are maximal projections (do not project their category) and are merged with a phrase. To spell out a constituent means to assign to it its semantic and phonological representation. As a result, a constituent becomes a term, like a lexical item, and, consequently, opaque for extraction. Thus, from (4) it follows that all adjuncts are opaque.

Crucially, it does not follow from (4) that all modifiers are opaque, as modifier accounts predict. The Spell Out Theory predicts that modifiers can be transparent in two cases:

The first case (5a) arises when a modifier is merged low. If it is merged with a head, it is no longer a structural adjunct, it is an un-selected complement and is expected to be transparent. It is still interpreted as a modifier (e.g, via Predicate Modification), in the sense that it does not fill any argument slot of the main predicate and is optional. But structurally it is a complement, because it is a sister to a head (3), and is expected to be transparent.

The second case (5b) arises when a modifier is merged with a phrase, but it is this phrase that is spelled out. In this case the modifier projects its category. That is only possible if the sister of the modifier is a phrase that can be spelled out. Here and below I will assume that for a phrase to be able to be spelled out all of its merge-inducing features (selectional and EPP-features) must be satisfied. In this circumstance either the modifier or its sister can be spelled out, while the other projects its category. An important prediction of this analysis is that the sister of a transparent modifier has to be opaque. No simultaneous extraction out of the modifier and out the main clause should be possible.

In what follows I will show that apparent counterexamples to the Adjunct Condition found in Balkar and English fall precisely under the two categories in (5), which supports the Spell Out Theory.

In sections 2, 3 and 4, I will present a case study of Balkar converbs (un-tensed adjunct clauses). Crucially, Balkar converbs are semantic modifiers. They are optional, can be stacked and do not fill any argument slots of the main predicate. Modifier accounts predict them to always be opaque. However, as I will show below, these clauses can be transparent for scrambling. Furthermore, they are transparent in precisely the two situations described by (5), as the Spell Out Theory predicts. Balkar converbs are a good case study for two reasons. First, the attachment site and the category of the converb can be established independently from extraction. Second, scrambling in Balkar permits moving more than one constituent at once. This makes it possible to test the prediction in (5b).

In section 5, I will briefly consider some known exceptions to the Adjunct Condition in English, originally pointed out by Truswell (2007), in the light of what we find in Balkar. We will see that the analysis proposed for Balkar can be extended to English as well.

Before discussing the Balkar data, I will briefly introduce the Spell Out Theory (Single Root Derivation) and show how it derives the claims in (4). For more details see Privoznov (2021).

According to the Spell Out Theory, syntactic derivation operates with three finite sets: the Lexicon, the Numeration and the Root. The Lexicon and the Numeration contain syntactically primitive elements (heads or spelled out phrases), the Numeration is a subset of the Lexicon. The Root is a singleton set (Single Root Derivation).5 Merge is understood as an operation that takes an element from the Numeration and merges it to the Root.6 Spell Out is understood as an operation that takes the element in the Root, assigns it fixed interpretation and phonological information and stores it back into the Numeration. This derives (4a). Namely, Merge (at least, External Merge) always puts together (i) a spelled out phrase or a head from the Numeration and (ii) the Root.

In addition, the Spell Out Theory assumes the following projection hierarchy: a head > the Root > a spelled out phrase. When Merge creates a set of two elements it is the higher one on the hierarchy that projects. Thus, when a head from the Numeration is merged to the Root, the head projects; meanwhile, when a spelled out phrase from the Numeration is merged to the Root, the Root projects (4b).

In this section, I will introduce converbs (un-tensed clausal adjuncts) in Balkar. We will discuss their size and their syntactic position. But before we proceed, let me make some preliminary observations about the clausal structure in Balkar.

Balkar is a head final (SOV) language with rich verbal morphology:

Given the morphological structure of the verb (6) and following the Mirror Principle (Baker 1985), I will assume the following syntactic structure for a simple clause (here and below I will assume that, like in English, Balkar subjects are base-generated inside the vP and later move to Spec,TP):

There are two assumptions illustrated by the tree in (7) that will be important later. First, the causative morphology (CAUS) is hosted by a v head, which takes another vP (or VP) as its complement. For Balkar causatives this analysis was proposed and developed by Lyutikova et al. (2006). Second, Negation is located low: above the vP, but below light verbs. Here I am following the Mirror Principle, cf. (6) and (7).

Balkar has a variety of clausal modifiers. For the purposes of this case study I will focus on non-finite clauses headed by a verb with the suffix -a7 or -p. These clauses will be referred to as ‘converbs’ or ‘converb clauses’. The term comes from the typological literature (see e.g. Haspelmath & König 1995). In the literature on Turkic languages it is sometimes applied to all non-finite clausal adjuncts (see Kornfilt 1997, Grashchenkov 2015 and Ermolaeva 2016). Here and below I will use the term ‘converb’ specifically for un-tensed clauses formed by the two aforementioned suffixes.

The choice between -a and -p depends on the temporal relation between the converb clause and the main clause. As in many other Turkic languages, in Balkar -p is primarily used for precedence (8a), and -a – for simultaneity (8b).8

As in many other Turkic languages, the converb clause in Balkar usually does not have its own overt subject (Kornfilt 1997: 68). It has a gap in the subject position, presumably, a pro argument that is interpreted as co-referent with the subject of the main clause (possible controllers for pro are discussed in Appendix 2, see the Supplementary file).

There are, however, certain semantic conditions (Grashchenkov 2015 and Ermolaeva 2016) under which a converb may have its own subject, similar to absolute adjuncts in English (Stump 1985). The subject of a converb has zero case marking, which means that either it has no case or that it bears nominative, since nominative is consistently null:

Both -a and -p converbs with pro and -a and -p converbs with an overt subject are semantic modifiers. They are optional, they can be iterated (10), they do not fill any argument slots nor do they satisfy any EPP-features of the main clause, their semantic interpretation is independent from that of the main clause in the sense that they are not assigned a theta-role. In other words, converbs (with or without an overt subject, with -a or -p) are modifiers in all the relevant respects. Thus, modifier accounts predict them to always be opaque for movement, which, as we will see below, is not the case.

Here and below I will assume that -a converbs and -p converbs have the same internal syntactic structure, occupy the same positions in the main clause and behave in the same way with respect to extraction. No data that I have collected contradicts this tentative assumption.

In Balkar most constituents can be moved (i.e., scrambled) away from their base position. In most cases this happens to constituents that are interpreted as given or topical, while focused material remains in-situ, preferably, adjacent to the verb. For example, in (11) the object of the embedded clause Fatimanɨ kitabɨn ‘Fatima’s book’ is scrambled to the left periphery of the main clause.

In what follows I will use long-distance scrambling as a diagnostic for converb transparency for movement. For this reason all the examples that involve extraction below have a clause modified by a converb embedded under an attitude predicate like sun ‘think’.

Scrambling is possible out of both converbs with an overt subject and converbs with pro, as is shown by (12a) and (12b) respectively.

However, there are two crucial differences between the configuration in (12a) and the one in (12b). First, whether a pro-converb can be scrambled out of or not depends on the main verb (13). The main verb cannot be transitive (13a) or unergative (13b). It has to be an unaccusative verb of position (13c) or motion (13d). In what follows I will argue that this is due to different structural positions of converb clauses in the context of different main verbs (see below).

Meanwhile, for the transparency of a converb with an overt subject the identity of the main verb is irrelevant:

Second, scrambling a constituent out of a pro-converb has no effect on the transparency of the main clause. As is schematized in (17), it is possible to scramble a constituent out of the main clause (17a), out of the converb clause (17b), and out of both clauses simultaneously (17c)–(17d).

At the same time, if scrambling applies from a converb with an overt subject, the main clause becomes opaque. It is possible to scramble a constituent out of the main clause (19a), out of the converb clause (19b), but crucially not out of both clauses simultaneously, regardless of the surface order of the scrambled constituents (19c)–(20d).

Crucially, (20) cannot be explained by a construction-specific ban on long-distance scrambling of two constituents at once. In fact, it is possible to scramble two constituents at once from the main clause (21a) or from the converb with an overt subject (21b), but not out of both in parallel (20c)–(20d).

In the remainder of this paper I will argue that converbs with pro confirm the prediction in (5a), and converbs with an overt subject confirm the prediction in (5b).

In what follows I will argue that Balkar converbs come in three varieties.

CP-converbs contain a full CP structure. They can contain a causative morpheme, negation, the light verb tur, a TP-level adverb, an overt subject and an epistemic adverb. They bear a special semantic relation to the main clause, encoded by their silent C.

TP-converbs contain a full TP structure. They can contain a causative morpheme, negation, the light verb tur, a TP-level adverb, but not an overt subject or an epistemic adverb. They lack C and hence do not bear any special semantic relation to the main clause, apart from the temporal one.

vP-converbs are vPs. They can contain a causative morpheme, but not negation, the light verb tur, a TP-level adverb, an overt subject or an epistemic adverb:

CP-converbs are attached at the left periphery of the main clause, they cannot be interpreted in the scope of negation or the subject of the main clause. TP-converbs can be interpreted in the scope of negation and the subject of the main clause. They attach at the T′, the turP or the vP level. vP-converbs have to be attached inside the vP of the main clause. If the main verb does not have its own object and is a verb of motion or position, a vP converb can merge directly with the verb, as its complement:

TP-converbs are always opaque for scrambling. This is as predicted by the Spell Out Theory. They are not complements, and their sister cannot be spelled out,11 so neither of the cases in (5) arises.

Only CP- and vP-converbs can be transparent for scrambling. vP-converbs are only transparent in the context of a verb of motion or position. This means that vP-converbs are only transparent if they are merged as complements. They are still optional, do not fill any argument slots of the main predicate and do not receive a theta-role, so the modifier accounts predict them to be opaque. But they are merged with a head and are transparent, as predicted by (5a). As expected, their transparency does not affect the transparency of the main clause.

Prediction (5b) is confirmed by CP-converbs. It is possible to scramble a constituent out of the main CP (24a), a constituent out of the CP-converb (24b), but not out of both (24c)–(24d). Crucially, it is not the case that scrambling two constituents at the same time is impossible in Balkar in principle. For example, it is possible in the case of a vP-converb (18c). But it is not possible with a constituent inside a CP-converb and a constituent inside its sister:

CP-converbs are modifiers, but because their sister is a also a CP, either the converb or its sister can be spelled out. The spelled out constituent does not project:

The case of CP-converbs will be discussed in section 3. The cases of TP- and vP-converbs will be discussed in section 4.

As was established by Grashchenkov (2015) and Ermolaeva (2016), whether the converb clause may have an overt subject depends on its semantic relation to the main clause.13 With pro the simple temporal relation (precedence or simultaneity) is enough, while with an overt subject there is some additional semantic relation to the main clause, similar to absolute adjuncts in English (Stump 1985). See Appendix 3, the Supplementary file, for more details. I will assume that this semantic relation is encoded by the silent C that embeds the converb clause inside the main one and licenses the overt subject.

In what follows I will assume that Balkar converbs with overt subjects are CPs with a silent complementizer. This complementizer has a double effect: (a) it licenses an overt subject;14 and (b) it is interpreted as ‘with’ (see English translations). Of course, this analysis raises a question of why the silent complementizer has this specific interpretation, but this question will have to be left for the future research.

As was shown in section 2, it is possible to scramble a constituent out of a CP-converb. However, in this respect CP-converbs show two important properties that distinguish them from pro-converbs: (a) main clause opacity; and (b) lack of correlation between the possibility of extraction and the lexical meaning of the main verb.

A CP-converb is only transparent if its sister is opaque. Even though it is possible to scramble a constituent out of the main clause (40a) or out of the CP-converb (40b), it is not possible to scramble out of the main clause and the converb simultaneously (40c)–(40d).

This is what the Spell Out Theory predicts. If one CP modifies another CP, the system may choose to spell out either one of them, but at least one must be spelled out. The non spelled out CP will project its category:

It is also expected that there will be no correlation between the possibility of subextraction and the meaning of the main verb.

It is unclear how modifier approaches to the Adjunct Condition can explain these data. The converb clause is a modifier in all the considered examples. It is optional and does not fill any argument slots. Hence modifier accounts expect it to be invariably opaque, contrary to fact.

Before returning to the discussion of extraction, I will consider certain cases when a vP-converb is attached so low that it becomes the structural complement of the main verb.

It is very common across Turkic languages that in the context of a converb clause certain main verbs lose their lexical semantics and assume the meaning of an aspectual or modal operator, or the meaning that is usually associated with so-called restructuring predicates, like ‘begin’, ‘end’ or ‘allow’. In the literature on Turkic languages this structure is usually called ‘serialization’ or ‘a complex verb construction’ (see Tybykova 1966, Lyutikova et al. 2006, Grashchenkov 2015, among numerous others).

An example of a complex verb construction from Mishar Tatar (a Turkic language spoken predominantly in the Republic of Tatarstan, Russia) is given in (56). Here the verb jat ‘lie.down’ no longer means ‘to lie down’, but functions as an aspectual operator. One could view this in the same light as the use of the -ing form with be or keep to form progressives in English.

In these cases the converb clause is usually analyzed as the complement of the bleached verb, as is illustrated by the tree under (56). The complex verb construction has several properties of Restructuring (see Wurmbrand 2001, and much subsequent work).

For example, if the complex verb construction is to be causativized or passivized, the causative / passive marking sometimes appears both on the bleached verb and on the lexical verb (57). This sentence contains two causative suffixes, but semantically there is only one causation. In fact, adding an extra Causer makes it ungrammatical (see Grashchenkov 2015: 159).

According to Grashchenkov (2015: 154), whether causative or passive morphology appears on the lexical verb, on the bleached verb or on both depends on the language and on the identity of the bleached verb. Double marking seems always to be an option, but there are cases when the causative morphology on the lexical verb is optional, and there are cases when the causative morphology on the bleached verb is optional.

In what follows I am going to assume that one of the suffixes is semantically vacuous and appears only as the result of agreement. For concreteness, let us assume that it is agreement between the v head projected by the lexical verb and the v head projected by the bleached verb, as is schematized in (58). This is similar to Bondarenko’s (2018) account of passive in Buryat Restructuring, building on Wurmbrand & Shimamura (2017).17

According to Grashchenkov (2015: 92–93), main verbs that form complex verb constructions across Turkic languages include verbs of motion (‘go’, ‘come’), verbs of position (‘stand’, ‘put’), transfer of possession (‘give’, ‘take’), perception verbs (‘see’, ‘look’) and copulas (‘be’, ‘stay’).

In Balkar, verbs of motion or position constitute an interesting “intermediate” case. If a motion verb or a verb of position is modified by a vP-converb, this verb still retains its original lexical meaning, as in (59) and (60). This is evident from the fact that the main verb has its own adverbial modifier, i.e. zajau ‘by.foot’ in (59) and šindik-de ‘chair-loc’ in (60).

However, this configuration does show a Restructuring effect, when it comes to derivational morphology. If (59) is to be causativized, the speakers strongly prefer to put causative marking both on the converb and the main verb:18

The same is true for converb clauses modifying a verb of position.

This suggests that the converb clause is, in fact, a structural complement of the main verb. The Restructuring configuration is there, but the higher verb still bears its original lexical meaning (it does not select the converb and does not assign it a theta-role).

Notice that this is not the case with all main verbs, only with verbs of motion or position. In particular, while (61a) is not acceptable with pro controlled by the Causee, (62) is perfectly fine, no causative morphology inside the converb clause is required:19

Why can converbs only merge as complements in the context of a verb of motion or position in Balkar? This relates to a broader question of why only specific classes of verbs across Turkic languages can form complex verb constructions, and to an even broader question of why only certain verbs get grammaticalized into modal/aspectual auxiliaries cross-linguistically. In the paragraphs to follow I will only offer some preliminary speculations on the matter.

It seems that all the verb meanings that participate in complex verb constructions are very basic in some intuitive sense. Verbs of motion and position can describe a very big class of eventualities. An event of Fatima going down the road can mereologically overlap (see Link 1998), that is, have shared subevents and participants, with a very broad class of events, like an event of Fatima singing a song, walking a dog, or talking to me.20

It may be that a vP-converb can only directly merge with a verb head if the event argument of the converb and the event argument of the verb mereologically overlap (cf. Truswell 2007 for English and section 5 for discussion). Take, for example the sentence in (59). Suppose the vP-converb is interpreted as the <s,t>-type predicate in (63a), and that the main verb bar ‘go’ is the <e,st>-type predicate in (63b).

These two predicates can combine via Kratzer’s (1996) Event Identification rule:

This will result in a predicate that describes two overlapping events: an event of Fatima moving and an event of Fatima singing a happy song.

If vP-converbs always have the semantics as in (63a), this would predict that vP-converbs can only describe mereologically overlapping events, which is clearly not the case, as is evident from (62). There the converb is attached within the main vP, below the Causee, so it is a vP-converb. But the main verb is not a verb of motion or position, which can mereologically overlap with a broad class of events. In order to account for these cases, one will have to assume that vP-converbs may also have another meaning, where their event argument only temporally overlaps with the main clause event (not mereologically). It could be something along the following lines:

There are two independent assumptions at play here. First, vP-converbs are ambiguous between an interpretation like (63a) and an interpretation like (66). Second, only those vP-converbs that have an interpretation like (63a) can be base generated as structural complements, only they can combine with the main verb via Event Identification, provided that the main verb describes a class of events that could freely mereologically overlap with other events (namely, if the main verb is a verb of motion or position).

Together, these assumptions predict that vP-converbs can only serve as complements for verbs of motion or verbs of position. This analysis clearly relies on certain assumptions about the lexical semantics, specifically, the lexical semantics of verbs of motion and position. This requires clarification and a more extensive study, which will have to left for the future research. What is important for our present purposes is the fact that only certain verbs (verbs of motion and position in Balkar) can combine with converb clauses as complements.

As was shown in section 2, it is possible to scramble constituents out of converbs with pro. But in this respect, they have two important properties that make them distinct from CP-converbs: (a) the main clause is not opaque; and (b) whether extraction is possible depends on the lexical meaning of the main verb.

A pro-converb is opaque if it modifies a transitive (67a) or an unergative (67b) verb. Extraction out of a pro-converb is marginally acceptable, if it modifies an unaccusative position verb (67c), and is definitely grammatical in the context of a unaccusative motion verb (67d).

In other words, a pro-converb can only be transparent in the context of those intransitive verbs that can attach a vP-converb as a complement. In Balkar these are the verbs that, though not semantically bleached, may show certain restructuring characteristics, as has been shown in the previous section.

One additional argument in favor of the view that pro-converbs are only transparent for scrambling when they are merged as complements comes from iterated converb clauses. As has been shown in section 2.2 above, pro-converbs can be stacked. However, in this case only the converb clause closest to the verb is transparent for scrambling:

These data are easily explained if we assume two things. First, TP-converbs are always opaque. Second, vP-converbs are only transparent when they are structural complements.

This follows from the Spell Out Theory. When vP-converbs are structural complements (in the context of a verb of motion or position), they are expected to be transparent because they are merged with a head. It is also expected that extraction out of a complement converb does not make the main clause opaque and that if converb clauses are iterated, only the lowest one is transparent for extraction.

Meanwhile, TP-converbs are not expected to be transparent. As has been argued in section 4.1.2, TP-converbs are attached above the vP, which means that their sister is always a phrase (includes at least the verb and one or more of its arguments). This means that they cannot be transparent due to being merged with a head. According to the Spell Out Theory, the only other option is for their sister to be spelled out, in which case the converb would project its category. If a TP-converb is merged below the TP-level, its sister is not a TP, in which case the converb projecting its T-category would not fit the selectional restrictions of the next higher head on the clausal spine. If a TP-converb is merged at the TP-level it has to be merged immediately below the subject position, so that its pro can be controlled by the subject, in accordance with the Minimal Distance Principle, as was discussed in section 4.1.2 and Appendix 2, the Supplementary file. But that means that the sister of this TP-converb is a T′ with an unsatisfied EPP-feature and hence cannot be spelled out (see section 1.3).

As an anonymous reviewer points out, another possible attachment site for a TP-converb is at the TP-level, but above the subject position. In this case, the sister of the converb bears the same category as the converb, has all of its merge-inducing features satisfied and, thus, can be spelled out. This predicts TP-converbs to be transparent, contrary to the facts. Note, however, that the subject of the TP-converb in this case does not have an available controller (the TP-converb is attached too high, above all the other DPs in the sentence), which opens two possible analyses. Either TP-converbs cannot be attached at this position (because their PRO-subject has to be bound), which rules out the possibility of their sister being spelled out and TP-converbs being transparent. Or this is, in fact, the case of a converb with an overt subject. If the latter analysis is correct, converbs with overt subjects are, in fact, TPs (their special semantic relation to the main clause and their ability to be modified by epistemic adverbs has to be explained in some other way). With respect to extraction, the analysis remains basically the same: between two TP-sisters either can project, while the other is spelled out and becomes opaque for movement.

Thus, the Spell Out Theory correctly predicts that pro-converbs can only be transparent when they are merged as complements. Importantly, pro-converbs in all of the cases described above are semantic modifiers. They are optional, can be iterated, do not fill any argument slots of the main predicate and do not receive a theta-role. Hence, modifier accounts predict them to be invariably opaque for movement, contrary to fact.

In this section I will briefly consider extraction from ing-clauses in English. The discussion will be limited to ing-clauses for two reasons: (a) they are close in meaning and structure to Balkar converbs; and (b) the possibilities of extraction out of them have already been addressed in the literature, most notably by Truswell (2007).

In what follows I will focus on ing-clauses that serve as modifiers and have a null subject, like whistling Ode to Joy in (69).

Truswell (2007) argues that whether an ing-clause22 can be extracted from depends on its semantic relation to the main verb:

An ing-clause is transparent if and only if its event argument is identified with one of the event arguments in the main clause. This is only possible if the event argument in the main clause is underspecified, which in turn is determined by the lexical semantics of the main verb.

It is not clear how Event Identification should influence possibilities of extraction. However, if availability of Event Identification correlates with the attachment site of the ing-clause, the Spell Out Theory provides a ready explanation for (70). Event Identification is only possible if the ing-clause is a structural complement. In the rest of this section I will go over the cases brought up by Truswell (2007), viewing them from the perspective of the Spell Out Theory and assuming that Event Identification correlates with the attachment site.

According to Truswell (2007), ing-clauses are transparent in the context of three classes of main verbs: (a) unaccusative atelic verbs of motion and position, like walk, (b) some telic unaccusatives, like arrive, and (c) telic transitive verbs with underspecified causing subevent (so-called result verbs), like anger. In what follows I will consider the intransitive cases (a-b) and the transitive ones (c) separately.

According to Truswell (2007), there are two classes of intransitive verbs in whose context an ing-clause can be transparent: atelic verbs of motion and position and some telic unaccusatives.

In both sentences in (71) the ing-clause modifies a verb of position and is transparent for relativization. These examples are judged as grammatical by all the speakers I have consulted.

The same is true for verbs of motion:

These examples fall within Truswell’s generalization. Atelic verbs of motion or position can be construed as mono-eventive predicates. Their event argument is underspecified, so it can be identified with the event argument of the ing-clause.

According to the speakers I consulted, with unergative verbs extraction is slightly worse, although still not completely unacceptable:

As Truswell (2007) argues, the decisive factor here is whether the event argument of the main verb and the event argument of the ing-clause can be identified. That, of course, depends on the lexical semantics of the verbs involved. This situation is very similar to vP-converbs in Balkar, discussed in section 4 above.

Let us assume that some English ing-clauses have the same semantics as some Balkar vP-converbs, that is, they introduce a mereologically overlapping event into the event structure:

This meaning can combine with the matrix verb via Event Identification (65). This is only possible, if the main predicate is vague, like a verb of motion or position. That is, the main verb describes an event that can freely mereologically overlap with a broad class of other events.

Apart from atelic verbs of motion or position, some telic unaccusatives also allow extraction from the ing-clauses that modify them (75). Truswell (2007: 1370) reports extraction with appear as ungrammatical, but the speakers I consulted find it acceptable, though, perhaps, a bit degraded.

Telic predicates can be analyzed as bi-eventive. For example, arrive can be construed as describing two events e₁ and e₂, e₁ being an event of Alex moving, and e₂ being the state of Alex being here,23 where e₁ causes e₂ (see Dowty 1979, Levin & Rappaport-Hovav 1995, Paducheva 2004; 2009, Ramchand 2008, Tatevosov 2015, among numerous others).

Following Truswell (2007), I will assume that in the case of telic predicates (75) the event argument of the ing-clause overlaps with the second event argument of the main clause (the result state). This can also be achieved via Event Identification, if the ing-clause is merged directly with the main verb.

These data are compatible with the Spell Out Theory. Adopting the analysis proposed for Balkar above, we can assume that combining an ing-clause with the main verb via Event Identification is only possible, if the ing-clause is base generated below the lowest argument of the verb and is merged directly to the lowest head inside the VP (either of a mono-eventive atelic verb like walk or a bi-eventive telic verb like arrive):

The tree in (76a) does not match the surface word order. However, it can be remedied by extraposing the ing-clause to a position after the PP. The most crucial assumption here is that for the event argument of the ing-clause and to overlap with one of the event arguments of the main clause the ing-clause must be merged as a complement. This is an assumption I made for Balkar in section 4, but it seems reasonable to extend it to English.24

For Balkar this analysis is independently supported by restructuring effects. Unfortunately, English does not provide similar evidence. However, there are two independent arguments that can be brought up to support this theory.

First, if the verb is elided, a transparent ing-clause has to be elided together with it.25 Consider the pair of sentences in (77). In (77a) the ing-clause is elided together with the main verb walk. In (77b) only the main verb is elided.

What (77) shows is that ing-clause does not have to be elided together with the main verb. Assuming that eliding a head without its complement is not possible, this means that the ing-clause does not have to be the complement of walk.

However, if the ing-clause is extracted from and the main verb is elided, the ing-clause must be elided with it:

In both (78a) and (78b) the ing-clause is extracted from. In (78a) it is elided together with the main verb, while in (78b) it survives ellipsis, which is only possible if the ing-clause is a structural adjunct. All the speakers I consulted perceive a contrast between (78a) and (78b). If the first sentence is acceptable, but degraded, the second sentence is definitely ungrammatical. This suggests that if an ing-clause is extracted from, it must be a complement.

Second, with stacked ing-clauses only the lowest one is transparent for extraction.26 All the speakers I consulted perceive a strong contrast between (79b) and (79c), with the former being consistently judged better. While ing-clauses can be stacked (79a), only one of them, namely, linearly the first one, can be transparent for extraction. This is expected, if we assume that only the first ing-clause can merge as a complement.

Extraction out of ing-clauses that modify transitive verbs seems to be much more marked. According to Truswell (2007), there are two conditions that have to be satisfied for an ing-clause to be transparent in the context of a transitive verb.

First, the main verb has to be an accomplishment specifying result. This is based on the idea that some accomplishments, like draw, specify the manner of a complex eventuality, while others, like drive crazy, specify the result (see Levin & Rappaport-Hovav 1995 and much subsequent work).

Both classes of verbs can be construed as bi-eventive predicates, describing two events e₁ and e₂, where e₁ causes e₂. Manner verbs specify the causing subevent. Result verbs specify the caused subevent. For example, draw describes a pair of events e₁ and e₂ where e₁ is the process of a picture being drawn and e₂ is the state of the picture being complete, and e₁ causes e₂. The verb draw specifies e₁. On the other hand, drive crazy describes a pair of events e₁ and e₂ where e₁ is the event of making someone crazy, and e₂ is the state of someone being crazy, and e₁ causes e₂. The verb drive crazy specifies e₂ and leaves e₁ underspecified.

Extraction from an ing-clause is degraded in the context of a stative verb (80a), a semelfactive verb (80b), or an accomplishment specifying manner (80c). But it is better in the context of an accomplishment specifying result (80d).

Second, the ing-clause itself has to describe the causing subevent. The semantic relation between the ing-clause and the main clause is one of immediate causation. In particular, indirect causation is out:

This is explained if the event argument of the ing-clause is identified with e₁, the causing subevent. What caused John to be angry in (81) has to be him trying to fix the radiator, not any other event associated with it.

This is in accordance with Truswell’s generalization (70): the event argument of the ing-clause is identified with an event argument of the main clause, which is only possible if the event argument of the main clause is underspecified. Because result verbs do not specify the causing subevent, it can be specified by the ing-clause. Consequently, the ing-clause can be transparent in the context of a result verb.

Following Harley (1996), Folli & Harley (2007), Pylkkänen (2008), and Ramchand (2008), among numerous others, I will assume that the causing subevent e₁ is introduced by a high functional head, like v or Voice, that is also responsible for the introduction of the Agent argument (the external argument). Meanwhile the caused subevent e₂ is introduced by some lower head, like V, that is also responsible for the introduction of the Theme argument (the internal argument):

If the event argument of the ing-clause is to be identified with the causing subevent e₁, and if the ing-clause is combined with the main clause via Event Identification, then it has to attach at position A, as represented by (82). This means that the ing-clause is not a structural complement. The Spell Out Theory predicts it to be opaque, contrary to fact.

However, there is some evidence against attachment site A and in favor of attachment site B. The first piece of evidence comes from Condition C. For all the speakers I consulted co-reference between a pronominal internal argument and a full noun phrase inside the ing-clause leads to a condition C violation. All the speakers I consulted perceive a strong contrast between (83a) and (83b), with the latter being consistently judged worse.

The same contrast repeats with extraction:

This suggests that the ing-clause is base generated below the internal argument, that is, at position B, as the structural complement of the main verb.27

Furthermore, like in the case of intransitive verbs, among stacked ing-clauses only the first one can be transparent for extraction:

If the transparent ing-clause is merged as a complement, how does it end up describing the causing subevent? This can be derived if we assume that English ing-clauses may have another reading, where they themselves introduce a causation relation into the event structure (perhaps, by a silent preposition ‘by’):

If a predicate like (86) is base generated as the sister to V, the ∃-bound event argument of the ing-clause ends up describing the causing subevent e₁ due to the semantics of the ing-clause. In other words, the ing-clause itself introduces the causing subevent, which is later identified with the causing subevent introduced by v, simply because there can only be one immediate cause of e₂.

This derives both the semantic relation between the ing-clause and the main verb, and Condition C effects in (83)–(84). Crucially, the ing-clause is attached low, as a sister to the main verb. Hence it becomes transparent, as predicted by the Spell Out Theory.