3.1. Instrument Phrases

Instrumentals are diagnostics for an external argument layer (i.e. a Voice layer). They tend to occur in the same environments as by-phrases (Fillmore Reference Fillmore, Bach and Harms1968, Bruening Reference Bruening2013, Alexiadou et al. Reference Alexiadou, Anagnostopoulou and Schäfer2015). For instance, in (16b) the instrument reading for ‘with hammers’ is not available in the anticausative/unaccusative, whereas it is available in the passive, (16a).

Instrumentals are also grammatical in the MC, which points to the presence of the embedded agent, (17).

These instrument phrases can modify the actions of the implicit agent. This suggests a Voice layer representing such an agent. Note that depending on the felicity of the context, instrument phrases are more saliently ambiguous with respect to whether they refer to the agent of causation or the implicit embedded agent, as in (18).

The main point is that instrumental adverbs can modify the embedded, caused event.

3.2. Agent-oriented Adverbs

Agent-oriented adverbs in SA provide another testing ground regarding the availability of an agent in the embedded verb phrase (Ernst Reference Ernst2001, Matsuoka Reference Matsuoka2013, i.a.). As such, these adverbs are compatible with passives, but not unaccusatives/anticausatives, as represented in (19).

These adverbs can modify the action of the embedded agent, as seen in (20).

Depending on the context, these adverbs may also be ambiguous in terms of whether they modify the action of the matrix agent or the embedded agent. Consider (21).

3.3. Agent-oriented comitatives

Agent-oriented comitatives indicate that the agent had help from the comitative in performing the event. They pattern with instrument phrases and agent-oriented adverbs in picking out a Voice layer (Bruening Reference Bruening2013, Alexiadou et al. Reference Alexiadou, Anagnostopoulou and Schäfer2015). As such, the comitative reading that is available in (22a) is lost with unaccusatives, as in (22b).

Turning to the MC, we see that the comitative reading is also available in this construction.

The ambiguity of modification regarding the embedded or matrix clause is observed with comitatives as well.

Thus far, we have seen that instrumentals, agent-oriented adverbs and comitatives point to the presence of a thematic Voice layer in the embedded event.

3.4. Verbs without embedded external argument

Another diagnostic to demonstrate that the embedded verb phrase is a VoiceP, and not just a VP, involves a constraint on the embedded external argument: as seen thus far, the embedded verb phrase may be transitive or unergative, (25).

On the other hand, unaccusative verbs are disallowed.

In addition to the impossibility of unaccusative predicates in the embedded verb phrase, the restriction on stative verbs is another indication of the presence of a Voice layer. Stative predicates are not embeddable under the ‘make’ causative, as illustrated in (27).Footnote ⁸

Following Harley (Reference Harley1995), Folli & Harley (Reference Folli and Harley2007: 215), I assume that stative verbs such as ‘have’, ‘fear’ do not take an external argument, similar to unaccusative predicates. I interpret this as the absence of a (thematic) Voice layer with such predicates, and hence their incompatibility with the MC. Note that in Sason, such verbs are not passivizable.

Note that in addition to ‘fear’, other subject experiencer verbs such as ‘enjoy’, ‘understand’ are also ungrammatical in this construction.Footnote ¹⁰

Furthermore, instrumentals mapped as external arguments are not felicitous. This is illustrated in (30).

On the other hand, it should be noted that although the embedded agent of the MC is most saliently interpreted as human, there is no such a restriction. As such, an animal interpretation is also possible given with a salient context.

Verbs such as ‘rot’, ‘stink’ are not embeddable under the MC.

I interpret this asymmetry as an argument that the complement of ‘make’ must contain a thematic VoiceP.

3.5. Agentive by-phrases

By-phrases are also grammatical with some restrictions on its acceptability. Most felicitous examples are when the DP embedded under by is indefinite, as in (33a). Note that it is introduced in the same way external arguments are in canonical passives (cf. 53), i.e. with the preposition mι ‘by, from’. Definite by-phrases lead to ungrammaticality, as in (33b).Footnote ¹²

Indefinite by-phrases further improve when they are “heavy” (heaviness reminiscent of ‘heavy-NP shift’), as illustrated in (34).

The best examples are with instances in which an impersonal or generic interpretation is available.

The possibility of by-phrases as in (33–35) suggests that there is a Voice layer, introducing an agent role that can be modified/identified by the by-phrases.

To summarize, we have thus far seen five diagnostics arguing for the presence of a thematic Voice layer: (i) the availability of instrument phrases modifying the embedded agent, (ii) the agent-oriented adverbs associated with the embedded agent, (iii) agent-oriented comitatives, (iv) the requirement for an external argument in the embedded event, thus the absence of unaccusative and stative predicates, and (v) the acceptability of certain by-phrases.

These diagnostics show that the MC in SA differs from similar constructions in embedding a structure up to VoiceP. Next, I investigate the size of complement clause of ‘make’. I argue that ‘make’ embeds VoiceP, but lacks higher projections. In this respect, it also differs from other causative constructions which have full clausal complements.

3.6. The size of the embedded clause

Clitic Left Dislocation (CLLD) demonstrates that no full CP is available in embedded clause. Direct or indirect object arguments in Arabic may normally be CLLD-ed to a left-peripheral position in the CP domain, be it matrix or embedded CP, (Benmamoun Reference Benmamoun2000, Aoun et al. Reference Aoun, Benmamoun and Choueiri2010), as in (36) (The CLLD-ed DP and the resumptive pronoun related to it are italicized).Footnote ¹³

However, (37) shows that such objects may not be CLLD-ed to the right of ‘make’, which indicates that the complement is not a full CP.

Moreover, neither the finite complementizer le nor the subjunctive complementizer te/ta are possible in SA MCs.

Negation is also disallowed on the infinitive, as shown in (39).

The exact position of negation in the clause structure, i.e. whether it is located above TP or between TP and AspP, is not central to the discussion in this paper since the main point is to test its availability in the MC regardless of its exact position in the extended projection. In the literature on Arabic negation, some researchers have argued that negation is location between TP and AspP (e.g. Benmamoun Reference Benmamoun2000, Aoun et al. Reference Aoun, Benmamoun and Choueiri2010), whereas others claim that it is above TP (e.g. Soltan Reference Soltan2007). An initial investigation of negation in SA suggests that the latter approach might be the correct one for SA negation, as evidenced by the relative position of negative morpheme with respect to the auxiliary, as in (40) (I leave the morpheme kι unglossed since at this point its exact status is unclear to me. Though see Akkuş Reference Akkuş2015 in which I treat it as a low tense marker).

The examples in (41) show that the embedded clause cannot have distinct temporal modification, thus point to the absence of bi-clausality, and more precisely the absence of a TP layer (e.g. Wurmbrand Reference Wurmbrand2001 et seq, Landau Reference Landau2004, Legate Reference Legate2014).

As expected, it is not possible to have a single temporal modifier in the embedded structure. Consider (42) with the adverb ‘tomorrow’, which cannot be associated with the matrix clause since the matrix clause is in the past tense, and fails to be associated with the embedded clause as well. The sentence becomes grammatical if the matrix clause was in the future tense.

With respect to the realization of aspect, as a common trait of Semitic languages, it is not easy to pinpoint a distinct aspect morpheme. To this end, for aspect in SA, building on the previous literature (e.g. Benmamoun Reference Benmamoun2000 for Arabic; Kramer Reference Kramer2014 for Amharic), Akkuş (Reference Akkuş2015: 17) suggests that “aspect is morphologically encoded by the position and phonological realization of the agreement marking on the verb”. According to this suggestion, we can track the aspect on the basis of the agreement marking on the verbal root, as such the absence of it can be interpreted as the lack of an Asp on the infinitivals. Relatedly, in SA the passive prefix is sensitive to aspect, as in (43), and is the portmanteau realization of Aspect+Voice heads. Thus, we can take it to be the fusion of Aspect and Voice heads, with agreement added onto it though the exact implementation is not crucial.

The impossibility of the passive prefix on the infinitivals also indicates the lack of the aspect projection.

To summarize, the discussion thus far indicates that ‘make’ embeds an agentive VoiceP, but lacks CP, TP, NegP and AspP.Footnote ¹⁴ Given this background, the next section establishes that this thematic VoiceP manifests an active-passive alternation.

However, before proceeding with that discussion, I investigate another alternative, i.e. ‘make’ embeds a nominalized complement (which is what Folli & Harley Reference Folli and Harley2007 suggests for faire par in Romance causatives). I contend that this analysis cannot carry over to the MC in SA.

3.7. An alternative: ‘make’ embeds a nominalized complement

I use two diagnostics to demonstrate that ‘make’ does not take a nominal complement, but a verbal one. First, Folli & Harley (Reference Folli and Harley2007: 19) draw the generalization that if a v takes a nominal complement (including for faire infinitif vs faire par (Kayne Reference Kayne1975)), it requires an animate agentive external argument - thus disallows causers such as rage, generosity or famine. Therefore, they are disallowed in FP fare (but not in FI fare, which embeds a regular vP). This is illustrated in (45a) for Italian and in (45b) for French (examples from Folli & Harley Reference Folli and Harley2007: 27–28).

Applying this generalization to SA MCs, (46) shows that ‘make’ does not take a nominal complement. Causers such as earthquake and fear are licit in this construction.

Note also the contrast between (47a) and (47b). In (47a), ‘wash’ is used in an argument position, is nominalized, requires the use of the preposition le ‘of’,Footnote ¹⁵ and has a different morphological form. This gerundial form is identical to cognate objects in the language. This form is not possible for the ‘make’ causative construction.

These two tests show that ‘make’ does not embed a nominalized complement.

4.1. (Impersonal) Passive

One reason to think that the embedded construction shows an active-passive alternation comes from the passivization of the matrix ‘make’. When the embedded clause contains a by-phrase, it behaves like a canonical passive, in that the embedded verb does not license the object, instead behaves as licensed by the matrix ‘make’. Therefore, when ‘make’ is passivized,Footnote ¹⁸ the embedded theme raises to grammatical subject position and shows verbal agreement. Consider (48a-48b). Raising is not possible without a by-phrase, as illustrated in (48c).

We can represent (48b) as (49), in which the embedded clause has a by-phrase, indicating its passive nature, and the embedded theme raises to the grammatical subject (as such it manifests subject-verb agreement) when the matrix ‘make’ is also passivized.

Without a by-phrase, the embedded clause behaves like a canonical active, with the embedded object behaving as though licensed by embedded verb.Footnote ¹⁹

The embedded theme remains a grammatical object even when ‘make’ is passivized (see Section 5). Passivization of ‘make’, when the embedded clause lacks a by-phrase, results in an impersonal passive. The embedded theme does not raise to the subject position. No argument is associated with the grammatical subject position, as such ‘make’ is realized with the default 3sg masculine agreement. Consider (50)–(51).

Crucially, under the active embedded analysis, this is expected. It turns out without a by-phrase, the embedded clause behaves as active, even when the matrix ‘make’ is passivized.

SA does indeed independently have impersonal passives, as in (52).

To summarize, the passivization of the matrix ‘make’ allows us to demonstrate the presence of two possible embedded structures, one active and one passive, despite the absence of a morphological reflex of this alternation. Without a by-phrase, the embedded structure behaves as active, as such the embedded theme is licensed by the embedded verb, and remains a grammatical object even when ‘make’ is passivized (see Section 5). Yet, the presence of a by-phrase necessarily leads to a passive clause, in that the embedded verb cannot license the embedded theme, which seems to be licensed by the matrix ‘make’. Accordingly, when ‘make’ is also passivized, the theme raises to become the grammatical subject.

Some remarks are in order regarding the obligatoriness of by-phrases for a passive structure in the MC, since this contrasts with garden-variety clausal passives in which by-phrases are optional, (53).

Another instance the obligatoriness of a by-phrase is reported by Ingason (Reference Ingason2016) for Icelandic, where in the caused-experiencer construction, the causing event can be expressed as a by-phrase adjunct, but this adjunct cannot be omitted. This is illustrated in (54).

Again in Icelandic, by-phrases in short passives are never obligatory, as illustrated in (55).

As rightly noted by a reviewer, given the generalization regarding the obligatoriness of the by-phrase, we are faced with four different constructions in SA: ‘regular active’, ‘implicit active’, ‘regular passive’ and ‘embedded passive’, in the reviewer’s terms. Noting from the start that any explanation must remain at a speculative level at this point, I believe in fact at the conceptual level, it is possible to categorize the four constructions in a way that makes a more symmetric division for the denotations of passive and active. Semantically, the passive is traditionally associated with two semantic denotations, one for the version with implicit agent, and the other one with the by-phrase (see e.g. Bruening Reference Bruening2013). On the other hand, the agent in the active is associated with a single syntactic and semantic denotation, i.e. the presence of a projected argument in Spec,VoiceP. Although it still does not explain the obligatoriness of by-phrases, we can say that in fact the SA facts allow us to have two semantic denotations for the agent in the active as well (which is what this paper essentially argues for; see Section 6.2), thus making it parallel to the passive. As such, the active could have a denotation for a version with a specifier and another denotation for the version without a specifier (cf. Wood & Sigurðsson Reference Wood and Sigurðsson2014, Wood Reference Wood2015).Footnote ²⁰

It should also be noted that it may not be a coincidence that both in SA and Icelandic, the obligatoriness of the by-phrase is observed in causative constructions. Ingason (Reference Ingason2016) attributes the obligatoriness of the by-phrase to the causative semantics in the Icelandic construction that he examines, yet does not elaborate on it. The different behavior from matrix passives, as also noted by the same reviewer, also supports the view that the obligatoriness may be tied to the causation somehow. It could be that the selection requirements of ‘make’, and the size of the complement clause (and possibly other factors) play a role in this obligatoriness. Thus, one could try to resort to either a semantic or syntactic explanation, but since both remain stipulative in the absence of sufficient crosslinguistic data, I must leave this significant point an open question.Footnote ²¹ As mentioned by Ingason, it is also worth exploring if by-phrases are syntactic arguments rather than adjuncts. The “smuggling” analysis of the passive by Collins (Reference Collins2005) is one such proposal. This analysis would still need to account for why some by-phrases are obligatory and not others. It is also possible that languages tend to have more constructions with obligatory by-phrases, apart from the Icelandic and Sason Arabic examples, yet these might have gone unnoticed thus far. Further studies on less studied Voice constructions in general, rather than short passives, might reveal a pattern and give us enough empirical base to draw a theoretical claim.

4.2. Sluicing

Another reason to think that the embedded complement manifests an active-passive alternation comes from ‘sluicing’. While VP ellipsis may in some cases allow voice mismatching, sluicing does not (Merchant Reference Merchant2013), as shown in (56).

Sason Arabic is no exception to this generalization. VP ellipsis allows voice mismatch, as indicated in (57), whereas sluicing disallows voice mismatch, which is shown in (58).

Turning to the MC, we observe that the embedded verb patterns as active for sluicing in the absence of a by-phrase, as in (59).

It is indeed possible to have different interpretations depending on whether sluicing targets the main clause or the embedded clause, as shown in (60). In (60a), the remnant “who” indicates that the sluice is active and it can target the caused event “build” in the complement of “make”, diagnosing “build” as active. In (60b), the remnant “by who” indicates that the sluice is passive, and it can only target the matrix clause, an impersonal passive, not the caused event “build”, again diagnosing “build” as active.

When a ‘by’-phrase is present, the embedded verb now patterns as passive for sluicing.

Thus, sluicing provides another strong piece of evidence that the embedded VoiceP in the MC shows an active-passive alternation.

4.3. Nonpassivizable idioms

SA has a class of non-passivizable idioms, as in (63). These idioms are another test for active-passive alternation (cf. Kayne Reference Kayne1975, Folli & Harley Reference Folli and Harley2007). This also demonstrates that the restriction against unaccusatives is not semantic. The predicate ‘getting lucky’ is not agentive, but the verb ‘break’ has an external argument.Footnote ²²

These idioms may occur in the MC only in the absence of by-phrases.

Idioms of this sort contrast with passivizable idioms, illustrated in (65).

Unlike non-passivizable idioms, which require the absence of by-phrases, such idioms impose no restriction, (66).

Following Harley & Stone (Reference Harley, Stone, Folli, Sevdali and Truswell2013), Harley (Reference Harley2017), I assume that non-passivizable idioms are impossible in (64b) since they require an active Voice to be present. On the other hand, the passivizable idioms are possible with or without a by-phrase, since the former option indicates an active Voice, whereas the latter a passive Voice.

4.4. Interim Summary

All of the evidence combined supports the proposal that syntactically, the embedded construction exhibits an active-passive alternation in SA MCs.

First, we find evidence that there is at least a thematic Voice layer (the availability of by-phrases, instrument phrases, agent-oriented adverbs, external argument restriction, etc.). Then we find a variety of indications that there is an active-passive Voice projection in the complement of ‘make’. In addition to the passive being marked via obligatory by-phrases, several diagnostics converge on this alternation: (i) the fact that ‘make’ can be passivized, and that the object is not promoted unless the embedded clause itself is passivized (ii) sluicing, and (iii) nonpassivizable idioms. As such, this construction varies from similar constructions from other language, some of which mentioned above.

In the next section, I demonstrate that the embedded theme remains as the grammatical object in the active even when ‘make’ is passivized and the embedded verb is active.

5.1. Definiteness effect

The first piece of evidence comes from the definiteness effect. In SA, indefinite subjects tend to occur postverbally with the possibility of appearing preverbally as well. On the other hand, definite subjects are strongly preferred for many speakers in preverbal position. Consider (67).Footnote ²⁴

The low theme in the passive also exhibits definiteness effects, as in (68).

However, the low embedded theme in the complement of ‘make’ does not, as in (69), regardless of whether ‘make’ is active or passive.

This shows that the embedded theme shows properties of a grammatical object of a transitive construction rather than those of a derived subject.

5.2. Clitic Left-Dislocation (CLLD)

The second argument that indicates that the embedded theme is the grammatical object comes from the fact that it can undergo CLLD.

As discussed in Section 3.6, direct or indirect object arguments in Arabic varieties may be CLLD-ed. However, the grammatical subject cannot be CLLD-ed. This holds both for the thematic subjects, (71a), as well as the underlying objects raised to become the subject via passivization, (71b).

When the matrix verb is passive, and there is a by-phrase associated with the embedded verb, the theme cannot undergo CLLD.

In (74), lack of agreement on the main verb indicates that it is an impersonal passive. The embedded clause is active, indicated by the lack of by-phrases. Given that no DP is linked to the grammatical subject position in impersonal passives, the DP associated with the embedded theme clitic is located in the CP domain.

These diagnostics demonstrate that the embedded theme functions as the grammatical object, licensed by the embedded verb in the embedded active VoiceP. Assuming the standard view about active (transitive) clauses (cf. Burzio (Reference Burzio1986), Chomsky (Reference Chomsky1995), i.a.), the straightforward expectation is that the embedded agent is projected despite being implicit. The next section deals with the syntactic and semantic status of the null implicit agent.

6.1. The implicit embedded agent and licensing

The diagnostics in Section 3 identify the presence of a thematic Voice, but do not necessarily entail the syntactic projection of such arguments. For instance, Bhatt & Pancheva (Reference Bhatt and Pancheva2006, Reference Bhatt, Pancheva, Everaert and van Riemsdijk2017) conclude that in some cases implicit arguments seem to be syntactically active, but that there is no good evidence to suggest that they are syntactically projected (see also Williams Reference Williams1985, Roeper Reference Roeper1987, Landau Reference Landau2010, Legate Reference Legate2014, Collins Reference Collins2018a, Reference Collinsb, Šereikaitė Reference Šereikaitė2020 for the discussion of implicit arguments).

This section investigates (i) anaphor binding, (ii) depictive licensing and (iii) scope in the MC and discusses their implications for the status of the embedded agent.

6.1.1. Reflexives, reciprocals and depictives

Sason Arabic exhibits a contrast between active and passive clauses in terms of the binding of reflexives. (75) shows that reflexives need a projected binder, thus are not licensed in passives.

Example (76) demonstrates that the reflexive is not licensed by the embedded agent in the active complement of ‘make’ causatives, either.

Reciprocals pattern like reflexives in needing a projected binder; therefore, they are also not licensed in passives. Consider (77).

The reciprocal is also not licensed by the embedded agent in the active complement of ‘make’ causatives. This is illustrated in (78).

Depictives also require projection of their licensors in SA: accordingly, they are not allowed in passives, as shown in (79).

Similar to reflexives and reciprocals, depictives also are not licensed by the embedded agent in the active complement of the MC.

6.1.2. Scope

The scopal interaction with negation provides another piece of evidence that the null embedded agent is not solely a PF-matter, but that it has interpretive results. I should note that this test is not perfect since, as shown in Section 3.6, negation cannot be in the embedded clause. Still, if there were a null projected agent we might expect it to be able to raise (e.g. QR at LF) above negation). Therefore, it is still worth discussing.

Example (81) demonstrates that in cases where the embedded agent is null, it necessarily takes scope under negation.

The infelicitousness of the sentence in (82) confirms the scopal judgments in (81). This is because whenever the first sentence is uttered truthfully, there is no person that will sing in the wedding, thus the follow-up sentence conflicts with the meaning of the antecedent discourse.

So far, the evidence points towards a view that the embedded agent is not projected despite the active VoiceP, thus SA MCs resemble the active existentials in Lithuanian. Šereikaitė (Reference Šereikaitė2020) demonstrates that in active existentials, illustrated in (83), the voice is active, the theme remains a grammatical object, and yet the embedded agent is not projected (see Šereikaitė Reference Šereikaitė2018 for the diagnostics).

Šereikaitė (Reference Šereikaitė2020) argues that the presence of a thematic Voice head is sufficient to count as active and license accusative case of the grammatical object in certain root clauses in Lithuanian. The embedded agent is not projected, but existentially closed at the Voice level, thus the label active existential.

The next section, however, shows that unlike the situation in Lithuanian active existentials, in SA MCs, pronouns can be licensed by the null embedded agent. Therefore, the analysis for active existentials cannot carry over to SA. Instead the next section posits that the agent is realized as a free variable, rather than being subject to existential closure.

6.2. Proposal: Embedded agent as a free variable

Similar to reflexives, reciprocals and depictives, pronouns cannot be anchored to the implicit agent of passives. In other words, canonical passives do not allow the pronoun to be bound by the implicit agent, thus force a disjoint-reference interpretation, as in (84).

Implicit arguments of passives cannot antecede pronouns occurring subsequently in the clause or in a subsequent clause (see e.g. Koenig & Mauner Reference Koenig and Mauner2000).Footnote ²⁵ Thus, the implicit agents of passives, marked as e in (85) (e.g. by a builder in (85a) or by a mother in (85b)) are not licit antecedents.

Pronouns corresponding to the embedded agent, however, are possible in the MC, as illustrated in (86).

Table 1

Binding properties in Voice types

The patterns are summarized in Table 1: in the active anaphors, depictives and pronouns are licensed, whereas in the passive none is. On the other hand, the ‘make’ causative exhibits a mixed behavior: while the anaphors and depictives are not licensed, pronouns are.

Examples in (86) suggest that a potential analysis along the lines of Šereikaitė (Reference Šereikaitė2018), in which the embedded agent is existentially closed, like in the passive, cannot carry over to SA since that analysis would rule out the licensing of pronouns as well. Accordingly, I posit (following Heim Reference Heim1982) that the embedded agent is present as a free variable, generated in the Voice head itself,Footnote ²⁶ and bound by Voice-level or text-level Existential Closure. This approach treats indefinites as non-quantificational, and as such the indefinite is like a free variable x_i with no quantificational force of its own. The indefinite gets bound in one of two ways: (i) either by being under the scope of an (unselective) quantifier in the sentence, e.g. if-clauses, adverbs, negation, or (ii) in their absence by an operation of existential closure, which puts an implicit unselective $ \exists $ on texts. This view, i.e. the dynamic analysis of indefinites, has been suggested to account for instances of e.g. (87) as well as ‘cross-sentential anaphora’ and ‘donkey sentences’, illustrated in (88) and (89), respectively.

Examples (90) - (92) demonstrate that the same considerations apply to the MC. We see that the free variable in the ‘make’ causative can antecede a following pronoun. This constitutes a contrast with the implicit passive agent in (85). I take this contrast as corroborating evidence that the implicit arguments in the causative construction differ from the implicit agents of passives. Whereas the latter requires an accommodation process in order to serve as an antecedent, the former does not.

The embedded agent in the MC patterns like an overt indefinite, thus semantically equivalent to it. They both can be under the scope of an unselective quantifier, and behave identically for discourse anaphora and ‘donkey sentences’. The two, however, differ in terms of anaphor licensing: whereas an overt indefinite can bind an anaphor (or license a secondary predicate; cf. (79a) versus (80a)), the embedded agent of the MC may not. Consider (93).

Accordingly, I propose that the implicit embedded agent is present not as a DP, but as a free variable, x_i, on the Voice head, as in (94), and can be bound by Voice level or discourse-level Existential Closure (Heim Reference Heim1982).Footnote ²⁷ Existential Closure at the Voice-level explains cases in which the free variable takes scope under negation. The external θ-role is introduced by a functional head Voice; the active matrix Voice selects for a DP specifier; we encode this selection using the feature [• D•] (see Müller Reference Müller2010 deriving). The semantic derivation for a tree that includes an embedded active Voice head is provided in (95).

The core idea then is that pronouns (more precisely the variable they introduce) can be licensed by virtue of being co-indexed with another variable (in this case, the agent variable on Voice head) and being bound by the same operator.Footnote ²⁸ This can be roughly sketched as (96), based on Heim (Reference Heim1982).Footnote ²⁹

Note that the pronoun variable that is not co-indexed gets reference from context.

In the case of anaphors, although they are also semantically variables, they are subject to a further restriction. They are not licensed by $ \exists $ (same restriction applies to depictives as well), and require a fully projected licenser.

On the other hand, I analyse the passive as a subtype of the Voice head itself (following Chomsky Reference Chomsky, Martin, Michaels and Uriagereka2000, Legate Reference Legate2014), although not much hinges on this choice, thus a PassiveP along the lines of Bruening Reference Bruening2013 might work (although as an anonymous reviewer rightly points out, a Pass head should not be possible if Pass requires an Aspect head directly above it). Since Voice only introduces the external $ \theta $-role (Kratzer Reference Kratzer, Rooryck and Zaring1996), only the external $ \theta $-role can be affected by the passive. Syntactically, the Voice_pass head introduces the external $ \theta $-role, but does not syntactically project this argument into its specifier. It is therefore compatible with a ‘by’-phrase adjunct, which optionally adjoins to VoiceP to specify the thematic subject.

Semantically, the passive needs to allow the external $ \theta $-role to be satisfied by the ‘by’-phrase, when present, and to otherwise be interpreted existentially (e.g. Bach Reference Bach1980, Keenan Reference Keenan and Shopen1985, Williams Reference Williams1987, Parsons Reference Parsons1990, Bruening Reference Bruening2013, Alexiadou et al. Reference Alexiadou, Anagnostopoulou and Schäfer2015, Reed Reference Reed2018). I therefore take it that Voice_pass has two associated semantic denotations (see also Bruening Reference Bruening2013, Alexiadou et al. Reference Alexiadou, Anagnostopoulou and Schäfer2015, Legate Reference Legate2014, Legate & Akkuş Reference Legate and Akkuş2017, Legate et al. Reference Legate, Akkuş, Šereikaitė and Ringe2020). In the first one, which does not combine with a ‘by’-phrase, the initiator is existentially bound on the Voice_pass head. Note that $ \exists $ is necessarily internal to Voice⁰, thus cannot bind elements outside Voice, as shown in (97).

In the second semantic denotation, Voice_pass leaves the initiator position open, i.e. $ \lambda e.\lambda x. $agent(x,e), to be accessed by the ‘by’-phrase. On the other hand, in the passive complements of ‘make’ in SA, only the second denotation is available, as such the initiator is saturated by the ‘by’-phrase.

6.3. An alternative analysis: $ \phi $P

In this section, I look at an alternative analysis of the implicit agent, in which it is realized as a $ \phi $P (Legate Reference Legate2014) in Spec,VoiceP. A reviewer suggests that this analysis could be tested in relation to the control facts if they are available in the MC as well as serving an explanation for the (abstract) accusative case on the embedded theme. It could also explain why the embedded theme does not raise to become the grammatical subject in examples such as (48c), repeated here as (98), in which presumably the embedded theme argument illicitly raises over the higher embedded implicit agent.

On the basis of recent work, I first contend that neither implicit control nor accusative case on the embedded theme require the syntactic projection of an implicit argument. Afterwards I suggest that Case licensing straightforwardly explains (98). This is followed by implementation of two arguments Sigurðsson & Wood (Reference Sigurðsson and Wood2020) use to argue for the presence of $ \phi $P in Icelandic indirect causatives, and demonstrate that they do not carry over to the MC in SA.

Let us start with the control test. It is indeed possible to embed both attitude and non-attitude control verbs under ‘make’, as shown in (99). As a background, similar to Hebrew, control is into a finite complement in SA. The obligatory de se reading and absence of the complementizer are two of the hallmarks of control.Footnote ³⁰

The question is whether being a controller necessarily entails syntactic projection. Some researchers have responded positively to this question. For instance, Landau (Reference Landau2010) in his classification of implicit arguments suggests that both SIAs and WIAs, which also include the passive agent, are visible as controller, therefore they should be syntactically projected. However, this conclusion has been challenged (e.g. Bhatt & Pancheva Reference Bhatt and Pancheva2006, Reference Bhatt, Pancheva, Everaert and van Riemsdijk2017) and argued against (Pitteroff & Schäfer Reference Pitteroff and Schafer2019). For instance, similar to the tests discussed earlier (e.g. licensing of by-phrases, agent-oriented modifiers), Bhatt & Pancheva (Reference Bhatt and Pancheva2006, Reference Bhatt, Pancheva, Everaert and van Riemsdijk2017) argue that the ‘control into purpose clauses’ test also does not necessitate the syntactic projection of an implicit external argument, including that of passives. Note also that, Landau’s (Reference Landau2010) main argument for the syntactic projection of implicit arguments (compared to leaving them syntactically unrepresented) is based on the observation that implicit arguments can function as the controller in partial-control contexts and the assumption that this type of control can only be treated in the syntax. However, as discussed in Pitteroff & Schäfer (Reference Pitteroff and Schafer2019), this argument contains two potential confounds: first, Landau develops his argument on the basis of implicit experiencers of adjectives. But since, for principled reasons, this argument cannot be applied to implicit agents of passives, the question of whether implicit agents of passives also license partial control remains open. Secondly, Pearson (Reference Pearson2013, Reference Pearson2016) has recently shown that a purely semantic analysis of partial control is, in fact, possible. As such, the possibility of control does not alone entail the syntactic projection of the controller.Footnote ³¹

The presence of the (abstract) accusative case on the embedded theme has also been argued to be another argument for a potential $ \phi $P. In fact, this is one of the reasons Legate (Reference Legate2014) postulates $ \phi $P, as such $ \phi $P makes it possible for the assignment of accusative case, by a thematic Voice. However, as mentioned above, Šereikaitė (Reference Šereikaitė2018, Reference Šereikaitė2020) demonstrates that a thematic Voice head is sufficient for the assignment of accusative case and thus the assignment of accusative case by Voice may vary independently from the selection of a specifier, be it a DP or $ \phi $P. Moreover, the necessity of a specifier position for case assignment of another DP (cf. Burzio’s (Reference Burzio1986) Generalization or its alternative versions, e.g. Legate (Reference Legate2014) or Dependent Case Theory Marantz (Reference Marantz, Westphal, Ao and Chao1991), McFadden (Reference McFadden2004), Preminger (Reference Preminger2014)) is a theory-internal postulation and a technical device to save Burzio’s (Reference Burzio1986) Generalization. As such, alternative models of case assignment have been argued for. For instance, Clem (Reference Clem2019), Akkuş (Reference Akkuş2020b) propose that ergative case is the result of agreement with multiple heads. Given these considerations, arguments for the existence of $ \phi $P in any language are not conclusive.

Instead, the ungrammaticality of (98) straightforwardly follows from Case licensing. As shown earlier, the embedded theme functions as a grammatical object, and it is already Case-licensed in the embedded active structure, as such it does not need to move.

I further implement two arguments Sigurðsson & Wood (Reference Sigurðsson and Wood2020) use for Icelandic indirect causatives towards the projection of the causee as a $ \phi $P. They argue that the impossibility of VoiceP modifiers like af kappi ‘enthusiastically’, (100a), viljandi ‘intentionally’ (as well as the lack of control into purpose clauses) as an argument for the reduced agency of the external argument, which is realized as a $ \phi $P rather than a DP. Applying this test to the MC in SA, we see that the implicit agent in SA differs from Icelandic and is compatible with such adverbs, as shown in (100b) (see also subsection 3.2).

Another argument Sigurðsson & Wood (Reference Sigurðsson and Wood2020) use for the reduced agency of the implicit causee in Icelandic causatives concerns the possibility of ‘choice’ on the part of the causee with respect to the performing an embedded event. They look at coordination of two vPs with a disjunction, and note that in causatives with an overt DP causee, as in (101a), the causee has the choice to choose between which one of the tasks to do, whereas with the implicit causee, this is not possible, (101b).

Applying this diagnostic to the MC in SA, we see that the implicit agent behaves unlike Icelandic, in that s/he has a choice which task to perform. Consider (102), in which the causee has a choice as to which task to do, as such the sentence can have a continuation of something like “… and from what I heard he decided to cut the grass”.

These comparisons with the Icelandic causatives indicate that the implicit causee in SA cannot be a $ \phi $P.Footnote ³²

Moreover, if the embedded agent is projected as a $ \phi $P and raises to grammatical subject position (i.e. Spec,TP) in impersonal passives (cf. 50-51), as suggested by an anonymous reviewer, it is at least expected that it should be able to trigger matrix agreement (Legate Reference Legate2014, p.c.). However, it always surfaces with default agreement in SA. Consider (103) in which the embedded subject is necessarily female, but feminine agreement on matrix ‘make’ is still ruled out in impersonal passives.

Note also that raising of $ \phi $P presupposes the availability of Spec,TP in impersonal passives, which is not an innocuous assumption. Pitteroff & Schäfer (Reference Pitteroff and Schafer2019) argue that in languages with impersonal passives, (i) Spec,TP is either filled by an overt expletive or (ii) Spec,TP is not projected. Sason Arabic, being a language that shows impersonal passives of unergatives, should then follow one of these two options. Given the absence of an expletive in the language, I assume that Spec,TP is not projected in impersonal passives.