Adjunct modification is known to be influenced by voice alternations. Various studies have explored by-phrases, subject-oriented adverbials, and pure instrument PPs (e.g., Keyser & Roeper 1984; Roeper 1987; Baker et al. 1989; Marelj 2004; Bruening 2013). In addition to these three types, the paper also discusses internal-argument-oriented comitatives. While actives are incompatible with (a) by-phrases, they can be modified by (b) subject-oriented adverbials, (c) pure instrument PPs, and (d) internal-argument-oriented comitatives (see (3)). In contrast, passives can be modified by all four types of adjuncts (see (4)). Middles allow all except by-phrases and subject-oriented adverbials (see (5)). Finally, inchoatives are restricted to internal-argument-oriented comitatives (see (6)).

Let’s examine how the patterns in (3–6) can be handled within the standard view of argument structure. Pylkkänen (2002) decomposes v into Voice and v (her Cause) (see also Alexiadou et al. 2006; Harley 2013; Legate 2014, among many others). v introduces the causing subevent, and Voice introduces the external argument. The lexical category V is typically assumed to license the internal argument in its complement. Consider the sentence John broke the glass. Its semantic representation based on the standard analysis is given in (7).

Adjuncts have frequently been used to test the presence of syntactic projections. The underlying logic is that each of a rigidly ordered sequence of heads is responsible for licensing one specific class of adjuncts (e.g., Alexiadou 1997; Cinque 1999). While agreeing with the general idea behind these analyses, I adopt Ernst’s (2002) more semantics-oriented approach, which derives most of the constraints on adjunct licensing from lexicosemantic properties and compositional rules. As a result, the one-to-one correspondence between a head and a class of adjuncts is not stipulated as a syntactic constraint but is rather a natural consequence of the semantics. I will now outline common approaches to the semantics of adjuncts and show that the structure in (7) is not sufficient to explain the adjunct distribution patterns in (3–6).

There are two main approaches to analyzing by-phrases. Collins (2005) posits that by realizes Voice, and that the arguments in by-phrases are base-generated in Spec, vP. In contrast, Bruening (2013) argues that by-phrases identify the Agent (which he calls Initiator) and are licensed by Voice′. Crucially, both approaches consider a projection of Voice to be the licenser of by-phrases. Since Collins’ syntactic analysis lacks a clear semantic interpretation, I adopt Bruening’s (2013) approach. Thus, by-John in (4a) (The glass was broken by John) can be represented as (8a) according to Bruening. Combining this with Voice′ through Functional Application yields (8b).

Subject-oriented adverbials are usually treated as two-place predicates that take an event and an individual (see, e.g., Ernst 2002). For example, in (4b) (The glass was broken deliberately), the adverb deliberately can be represented as (9a). With the additional lexical restriction that its two arguments must be a causing subevent and an Agent, it can only appear when there is Voice or Voice′, with which it combines through Predicate Conjunction (see Wood 2015: 23), as shown in (9b–c).

Yamada (2010) distinguishes three types of comitatives.3 For now, we are only interested in his Type 2 comitatives (although his Type 3 comitatives will be mentioned in section 3.1). Based on his semantics for Type 2 comitatives, with the plates in (4d) (The glass was broken (along) with the plates) can be represented as (10a). Depending on whether it combines with Voice′ or V, it can be interpreted as either external-argument-oriented or internal-argument-oriented. In the latter case, with the plates can combine with V through Functional Application as in (10b).

Franco & Manzini (2017) propose an alternative semantics for comitatives, arguing that with represents a zonal inclusion relation (i.e., part-whole/possession relation) between two arguments: either between two individuals or between an event and an individual. When used in comitatives, it represents the latter, roughly meaning that the individual argument of with participates in the event argument of this two-place predicate. (11a) presents a semantics consistent with their idea for with the plates. Again, depending on whether it combines with Voice′/VoiceP or V/VP, the comitative can have an external-argument-oriented or internal-argument-oriented reading. Thus, the internal-argument-oriented comitative in (4d) can either combine with V through Event Identification (11b) or with VP through Predicate Conjunction (11c).

I adopt a hybrid semantics from Yamada (2010) and Franco & Manzini (2017) for with (12), which analyzes it as a three-place predicate that takes an event and two individuals, following Yamada (2010), while extending Franco & Manzini’s (2017) zonal inclusion relation With. Crucially, this semantics of with ensures that internal-argument-oriented comitatives are licensed by V.

Pure instrument PPs are the most controversial among the adjuncts discussed here. Bruening (2013) essentially analyzes them as being licensed by Voice′, similar to external-argument-oriented comitatives. With a hammer in (4c) (The glass was broken with a hammer) has the semantics (13a), which identifies the Agent using the hammer as the Agent of the breaking event. Combining it with Voice′, we get (13b).

In contrast, Franco & Manzini (2017) classify pure instrument PPs with comitatives. Essentially, with a hammer has the same representation as the comitative in (11a). Assuming that pure instruments correspond to high Appls (Pylkkänen 2002), sandwiched between vP and VP, they argue that pure instrument PPs should also take VPs as their complements. The result of combining them is (14b). However, Franco & Manzini (2017) suggest that a with-PP is interpreted as a pure instrument PP only if the individual argument is inanimate, and VP is caused (v is present).

(5) and (6) show that by-phrases and subject-oriented adverbials cannot appear in middles or inchoatives, suggesting the absence of Voice′ in these constructions. In contrast, passives allow both by-phrases and subject-oriented adverbials (see (4)), indicating the presence of Voice′. Actives allow subject-oriented adverbials but not by-phrases (see (3)). Thus, actives project Voice′, and Spec, VoiceP is occupied. This yields a three-way split. However, a further split between middles and inchoatives is dictated by the distribution of pure instrument PPs: while middles permit them, inchoatives do not. Bruening’s (2013) semantics for pure instrument PPs (13a) does not explain this. Since pure instrument PPs are licensed by Voice′, they should pattern with subject-oriented adverbials. However, middles suggest the contrary: they are compatible with the former but not with the latter. Franco & Manzini’s (2017) semantics for pure instrument PPs (14a) seems to fare better with the data. Under this analysis, pure instrument PPs differ from by-phrases and subject-oriented adverbials in not requiring the presence of Voice′ and differ from internal-argument-oriented adverbials in requiring the presence of v. However, according to Franco & Manzini (2017), pure instruments must be manipulated by a volitional being, which presupposes Spec, VoiceP, eliminating the difference between pure instrument PPs and subject-oriented adverbials. Therefore, the three-layer structure of Voice–v–V is not fine-grained enough to make the needed distinction.

I have demonstrated that the standard analysis is insufficient to explain the structural contrast between middles and inchoatives. To address this issue, I suggest adding a projection, ApplP, to the structure in (7), with Appl′ licensing pure instrument PPs.4 The proposed structure, with some additional modifications to the one in (7), is as follows. (In this section, I focus on the syntax of the proposed analysis, leaving its semantics for section 2.4, where I will also discuss what is omitted in (15).)

Under the present analysis, only Causers are introduced directly as external arguments by Voice. This is different from the standard analysis, which assumes that Voice introduces both Agents and Causers. Appl is added to introduce Affectees. Causers and Affectees are atomic and primitive, while Agents are composite and consist of both Causers and Affectees.

I propose that Agents are derived through movement from Spec, ApplP to Spec, VoiceP.5 This involves movement into a θ-position, which has been extensively argued to be possible in the literature (see, e.g., Bošković 1994; Lasnik 1995; Hornstein 1999). In GB theories, D-structure is the linguistic level where the Projection Principle and the θ-Criterion apply. These two conditions work synergically to ensure that each θ-position projected by a predicate is filled by an argument at D-structure, making it impossible to move an argument into a θ-position. However, with the elimination of D-structure in the Minimalist Program, movement into θ-positions is no longer prohibited (for empirical arguments that such movement is indeed possible see, e.g., Bošković 1994).6

Let’s return now to the adjunct distribution patterns in (3–6). In actives, every head in the four-layer structure of Voice–v_init–Appl–Trans is projected. I argue that pure instrument PPs are licensed by Appl′, as will be further justified in section 3. Since Affectees presuppose Appl′ and are a subpart of Agents, the presence of Agents is sufficient to license pure instrument PPs (see (16a)).

In both passives and middles, Appl′ is still present, allowing pure instrument PPs to be used in these constructions (see (16b–c)). These PPs modify the implicit arguments in Spec, ApplP, which are represented as PRO in the labelled brackets.7

In inchoatives, neither VoiceP nor ApplP is projected, hence the incompatibility with by-phrases, subject-oriented adverbials, and pure instrument PPs, as shown in (16d).

The difference in adjunct distribution between middles and inchoatives can then be attributed to the presence or absence of Appl′. Middles project Appl′, allowing pure instrument PPs, while inchoatives do not project Appl′ and are therefore incompatible with pure instrument PPs.

One could try to deal with this paradigm under the standard analysis by postulating a defective Voice′. This would involve assuming that Voice′ licenses both subject-oriented adverbials and pure instrument PPs, but its defective counterpart only licenses pure instrument PPs. If we stipulate that middles project the defective Voice′, but inchoatives do not, we could capture the adjunct distribution difference between middles and inchoatives. However, this approach would introduce new machinery into the system, namely the defective Voice′, whereas the proposed analysis does not—Appl is independently needed. Moreover, semantically speaking, it is unclear what it would mean for Voice′ to be defective and what would prevent a defective Voice′ from licensing subject-oriented adverbials. Licensing pure instruments by two different heads also leads to a less unified analysis, as it requires postulating two distinct mechanisms for licensing the same class of adjuncts. In contrast, the proposed analysis establishes a one-to-one correspondence between each head and each class of adjuncts, providing a straightforward account of the adjunct distribution patterns in change-of-state verbs without encountering the issues faced by the alternative approaches. Therefore, the proposed analysis should be preferred.

The structure proposed in (15) is built on Ramchand’s (2008) first phase syntax, which breaks down an event into three subevents: initiation, process, and result. For the purposes of this discussion, I focus on the first two types, omitting the result subevent, which is largely irrelevant.8 Initiation refers to a state, while process indicates an internal change. Since initiation causally implicates process, initiation is interpreted as a causing subevent, while process is interpreted as a caused subevent. Each of these subevents heads a projection, as shown in (17). Ramchand’s theory defines θ-roles as participants in different subevents. DP₂ is the participant in the process and bears the Theme θ-role (her undergoer). DP₁ is the participant in the initiation and bears the Causer θ-role (her initiator). Both the Theme and the Causer are primitive.

The structure in (15) differs from Ramchand’s in two main aspects. First, unlike Ramchand’s system where subevents and their participants are introduced simultaneously by init and proc, the proposed structure, following Pylkkänen (2002) and others, introduces them with two different heads: v_init introduces the initiation subevent, and Voice introduces its participant Causer, while v_proc introduces the process subevent, and Trans introduces its (most prominent) participant Theme.9^,10 The merit of postulating two independent heads is that it allows for certain adjuncts to be licensed by a subevent, rather than by an argument introducer (e.g., cause PPs discussed in fn. 9).

Second, whereas Ramchand’s theory posits only one argument introducer for process, the proposed structure has two argument introducers for it (Trans and Appl). In addition to the two primitive θ-roles Causer and Theme, I suggest that Affectee is also a primitive θ-role, defined as the secondary Theme, namely the second participant in the process subevent. This θ-role minimally differs from the Theme in being less prominent. This difference is illustrated by the Italian example in (18). ‘The father’ is a Theme and is directly involved in the dying event, hence a core participant. ‘Maria’ is an Affectee and is less directly involved in the dying event because she is only negatively affected by the event but is not a core participant in it. As a less prominent participant, an Affectee is only licensed in the presence of a Theme and can always be omitted (see (18b)).11 Further semantic properties of Affectee will be discussed in section 5.3.

I summarize the definitions of Causer, Affectee, Theme, and Agent in (19).

To provide a more tangible illustration of this discussion, I present a sample derivation of the sentence John broke the glass below. → stands for the relation of causal implication. The predicate breaking is decomposed into breaking′, breaking″, and breaking′″ and modifies the sum of e, e₂, and e₃. Resultee is the participant in the result subevent. The free variable e in the representation of VoiceP stands for the initiation subevent and will be existentially closed at a higher position. e₂ and e₃ stand for the process and result, respectively.

Under the proposed semantics, by-phrases and subject-oriented adverbials are licensed by Voice′ as before (21a–b) (but note that under the proposed analysis, by-phrases saturate the Causer rather than the Agent argument); internal-argument-oriented comitatives are licensed by Trans′, the counterpart of V (21c); pure instrument PPs are basically Affectee-oriented comitatives (I explicitly argue for this in section 3) and are licensed by Appl′ (21d).

Comitatives establish a relation of accompaniment between two participants in a subevent (Stolz et al. 2006: 17–18). They come in various types depending on the arguments they modify. I will illustrate three types of comitatives using examples from double object constructions. The external-argument-oriented comitative is exemplified in (22a), where the most salient reading is “I and a social worker gave poor John some food.” Here, the comitative with a social worker modifies the external argument I. The applied-argument-oriented comitative is exemplified in (22b), which can be interpreted as “I gave poor John and other homeless people some food.” The comitative with other homeless people modifies the applied argument poor John. Finally, (22c) demonstrates another instance of internal-argument-oriented comitatives. The comitative with a bottle of water modifies the internal argument some food. (22c) entails that “I gave poor John some food and a bottle of water.”

The above indicates that comitatives possess the property that when the argument contained in a comitative is conjoined with the argument it modifies, the resultant sentence is entailed by the original one. In other words, comitatives are linked to (i.e., have the potential to bear) the θ-roles of the arguments they modify.

However, sentences like (23a) complicate this picture. This is because (23a) is ambiguous and can either mean “Pat and Chris built a sand castle together” (a case of Yamada’s Type 2 comitatives) or “only Pat did the building and Chris was just a bystander” (a case of Yamada’s Type 3 comitatives). The first reading entails (23b), but the second does not. This indicates that the comitative with Chris, in the second reading of (23a), cannot be linked to the θ-role of the external argument. Therefore, a question arises as to which θ-role it is linked to. There is no simple answer to this question under the standard approach to argument structure discussed in section 2.2.

In contrast, the decomposition analysis of Agents does not face this issue. Since the subject Pat is an Agent, it has two sub-arguments: a Causer and an Affectee. If the comitative with Chris modifies both the Causer and the Affectee, as in (24a), we get the “Chris as a co-builder” reading, and (23a) entails (23b). If the comitative only modifies the Affectee, as in (24b), only the “Chris as a bystander” reading is available, and (23a) does not entail (23b). Therefore, (23a) provides a potential argument for the decomposition analysis of Agents.

The semantics of (24a–b) and their entailments are given below, where ⊕ stands for the sum operation.

I now turn to pure instruments, which correspond to Kamp & Rossdeutscher’s (1994) concept of Instruments, as opposed to their Instrument Causers. Pure instruments are devices that are designed for a specific task and must be manipulated by a volitional being. They can be viewed as a special type of comitatives, given that crosslinguistically, accompaniment is often expressed by morphological markers related to instruments (see, e.g., Lakoff & Johnson 1980; Luraghi 2001). In fact, Franco & Manzini (2017) treat comitatives and instruments as two contextual interpretations of the same thing, as with in both cases expresses a zonal inclusion relation between an event and an individual.

Assuming pure instruments are a type of comitatives, we can expect them to exhibit the same entailment pattern as comitatives, which provides a way to determine which θ-role pure instruments are linked to. In (26), the verb eat takes an Agent subject. However, when we conjoin the pure instrument the fork and the Agent subject John, the resultant sentence (26b) is semantically deviant and is not entailed by (26a). This suggests that pure instruments cannot be linked to the Agent θ-role.

In (27), the verb break takes a Causer subject. The modifier falling forces the instrument subject into a Causer reading. The infelicity of (27a) suggests that Causers alone do not license pure instrument PPs. Furthermore, conjoining the Causer subject and the pure instrument is also semantically deviant (27b), which further confirms that pure instruments cannot be linked to the Causer θ-role either. The failure of entailment observed in (26) and (27) is similar to the case in (23).

Under the standard analysis, it is unclear how pure instrument PPs, which cannot be linked to the Agent θ-role, can modify the Agent subject in (26a) and (3c) and be allowed in (4c) and (5c), where no overt external argument is present.

The proposed analysis addresses this issue as follows. Since pure instruments are a type of comitatives, they also specify a relation between an event and an individual (cf. (12) and (21d)). I propose that pure instrument PPs can only modify Affectees and are licensed by Appl′. In (3c), the pure instrument PP modifies the Affectee sub-argument of the Agent subject John (see (28a)). In (4c) and (5c), although no Agent is overtly present, an implicit Affectee is introduced in Spec, ApplP. As a result, pure instruments PPs are still allowed in these sentences (see (28b) and (28c)). In summary, under the proposed analysis, a with-PP is interpreted as a pure instrument PP only if it modifies an Affectee, the argument within it is manipulable by a volitional being, and v_init is present.

Why did we discuss pure instrument PPs at all if they are a type of comitatives? The answer is that pure instrument PPs have much stricter interpretations than comitatives. While a pure instrument PP can only modify the argument introduced by the newly added head in (15), comitatives can modify any arguments. Thus, only pure instrument PPs can serve as a reliable test for the presence of the head in question. So far, we have assumed the head to be Appl, but we have yet to address why this would be so. In this section, I will provide evidence that Appl licenses pure instrument PPs, which supports the claim that Appl is the head divorced from Voice.

One piece of evidence to this effect comes from middles. These constructions are typically assumed to take implicit arguments (see, e.g., Stroik 1992; Hoekstra & Roberts 1993). However, the implicit arguments of middles cannot be Agents, as middles do not allow agentive adverbs such as deliberately (29a) or rationale clauses (29b).

Stroik (1992) observes that middles can introduce an additional argument overtly through for-phrases (see (30)). Hoekstra & Roberts (1993) analyze such arguments as Experiencers, the most common flavor of Affectees (see Bowers 2010). Based on this observation, it seems plausible to assume that the implicit arguments in middles are Affectees. It follows that Appl, which introduces Affectees, is projected in middles.

In contrast, inchoatives have no implicit arguments, which means they don’t project Appl. Therefore, the different ability of middles and inchoatives to license pure instrument PPs can be attributed to the presence or absence of Appl. This supports the hypothesis that Appl′ licenses pure instrument PPs and Appl should be equated with the newly added head discussed previously.

Further evidence comes from Korean inchoatives. Kim (2011b) notes that Korean has two types of inchoatives: zero inchoatives and I-inchoatives.13 Zero inchoatives do not allow by-phrases and pure instrument PPs (see (31)), while I-inchoatives permit pure instrument PPs but not by-phrases (see (32)).

Interestingly, the morpheme -I in Korean also appears in morphological causatives (see (33)). According to Kim (2011b: 115), -I is the morphological realization of (high) Appl, which introduces the causee ‘child’ and assigns the inherent dative Case to it.14

If Kim’s analysis is on the right track, then the minimal contrast in compatibility with pure instrument PPs between zero inchoatives and I-inchoatives can be easily captured by linking the -I morpheme with Appl. When -I is absent, as in zero inchoatives, there is no Appl (or Appl′) projected, and pure instrument PPs are not allowed (see (34a)). In contrast, when -I is present, as in I-inchoatives, Appl (as well as Appl′) is projected, and pure instrument PPs are permitted (see (34b)).

Aspectual si constructions typically involve consumption verbs and exhibit an Animacy Restriction on their subjects. (35) is an example from Italian, with translations and bold emphasis added by me. The verb mangiare ‘eat’ can take an animate subject like ‘the groom’ (see (35a)). However, when it takes an inanimate subject like ‘the sea,’ as in (35b), the sentence is semantically infelicitous. However, Folli & Harley (2005) observe that the deviance of (35b) can be remedied by adding a reflexive particle, referred to as aspectual si in the literature, as in (35c). The question then arises: how does this repair the unacceptability of (35b)?

I assume with Folli & Harley (2005) that consumption verbs require Agent subjects. Under the decomposition analysis, Agents are composed of Causers and Affectees. I posit that the Animacy Restriction is placed on the Affectee sub-arguments of consumption verbs, meaning that these verbs necessitate animate Affectees. As a result, Agents with Affectees as their sub-arguments must also be animate, as illustrated in (36a). Schäfer (2008) proposes that reflexive particles can function as expletive arguments, in the sense that they have no thematic interpretation and reduce the valency of a predicate (also see Wood 2015). That is, they serve only as structure holders without contributing semantic content. I argue that aspectual si has exactly the same function and is inserted in Spec, ApplP. Since it expletivizes Affectees, the Animacy Restriction on them is nullified. Consequently, the remained sub-arguments comprise only Causers, which need not be animate, as shown in (36b).16^,17

In support of analyzing aspectual si as an expletive in Spec, ApplP, I present evidence from MacDonald (2016), who observes that aspectual si (se in Spanish) syntactically parallels indirect object reflexives. Since indirect objects are commonly analyzed as applied arguments (see, e.g., Marantz 1993; Bruening 2010), it follows that aspectual si is also an applied argument (i.e., an Affectee).18

In Spanish leísta dialects, the dative preposition a is used to mark animate direct objects (see (37a)). When a pronominal form is used to refer to an animate direct object, it must be realized as le (which is homophonous with dative le) instead of lo (the accusative counterpart) (see (37b)).

Leísta Spanish does not allow le to co-occur with a dative indirect object clitic. (38) shows that the combination me le ‘me.io him.do’ is marked as ungrammatical. However, this violation can be resolved by replacing le with lo, a process known as the le→lo repair.

It is noteworthy that both aspectual se and indirect object reflexives induce ungrammaticality when combined with le. In (39a), se is an indirect object reflexive, and the animate direct object can only be expressed as lo rather than le. In (39b), se is aspectual, but the same morphological constraint is observed. Crucially, when the reflexive particle se refers to a direct object, as in (39c), no problem arises. In other words, the morphological constraint is violated only when se is an Affectee, not when it is a Theme. Thus, (39) suggests that aspectual se resides in Spec, ApplP.19

I have argued that consumption verbs impose an Animacy Restriction on Affectees. That the same restriction is found with Agent subjects of consumption verbs indicates that Affectees are a sub-part of Agents, supporting the current decomposition analysis of Agents. In the following section, I will show that ditransitive alternations (in English) exhibit a similar Animacy Restriction, which can be easily accounted for under the proposed analysis.

In English, many ditransitive verbs alternate between double object constructions (DOCs) and prepositional dative constructions (PDCs). These constructions display interesting contrasts in the Animacy Restriction (see, e.g., Green 1974; Oehrle 1976). In DOCs, the subjects can be inanimate, as in (40a), but the indirect objects must be animate, as seen in (41a) in contrast to (41c), while in PDCs, the subjects must be animate, as seen in (40b) in contrast to (40c), but the objects of the preposition can be inanimate, as in (41b).20 The arguments in bold indicate the source of ungrammaticality.

I suggest that the DOC (40a) and the PDC (41b) have the structures in (42). In the DOC in (42a), the subject is a Causer and is introduced by Voice, the indirect object is an Affectee in Spec, ApplP, and the direct object is a Theme in Spec, TransP. In the PDC in (42b), the subject is an Agent, which, according to the decomposition analysis of Agents, is derived by moving from Spec, ApplP to Spec, VoiceP. The direct object is a Theme in Spec, TransP, while the object of the preposition is introduced by P in its specifier.21 (I that assume a preposition like to is a morphological marker on the argument in the specifier of a covert P head.)

I suggest that the Animacy Restriction applies to Affectees in both DOCs and PDCs, akin to the restriction imposed by consumption verbs in aspectual si constructions. In DOCs, the Causer subjects can be either animate or not, while the Affectee indirect objects must be animate. In PDCs, the Agent subjects encompass both Causers and Affectees, and thus inherit the Animacy Restriction from Affectees. Notably, objects of to are insensitive to animacy. This explains the contrast in the Animacy Restriction in these two constructions.22 Ditransitive alternations show that Causers and Affectees can be expressed either as a single argument via movement, as in PDCs, or as two separate (non-coindexed) arguments, as in DOCs, a possibility offered by the decomposition analysis of Agents.

Aspectual si constructions and ditransitive alternations exhibit complex Animacy Restriction patterns. The standard analysis of Agents requires different structures and imposes the restriction on different arguments for each pattern. In contrast, the decomposition analysis of Agents enables this restriction to apply consistently to Affectees, allowing a unified explanation of this restriction in these seemingly unrelated constructions.

Agents alternate with Affectees mainly through expletivization or causativization. Let’s start with the subcase of expletivization. For example, in Spanish, transitive sentences typically follow the Nom-Acc alignment, where the predicate agrees with the nominative subject, as shown in (43a). However, unintentional causer constructions, exemplified by (43b), follow the Dat-Nom alignment, with the predicate agreeing with the nominative argument (see, e.g., Cuervo 2003; Kallulli 2006; Schäfer 2008; Wood 2013). This pattern is semantically distinguished from the ordinary case by the level of agentivity. Specifically, (43a) is more agentive than (43b), as evidenced by the incompatibility of (43b) with agentive adverbials like on purpose and rationale phrases.

I propose that the nominative argument ‘the dry-cleaner’ in (43a) is an Agent, while the dative argument ‘to the dry-cleaner’ in (43b) is an Affectee, following Cuervo (2003). Additionally, I argue that the reflexive particle se is another instance of an expletive argument (cf. section 4.1). However, in this case, it is inserted in Spec, VoiceP. In other words, (43b) is derived from (43a) by expletivizing the Causer. The two sentences can be represented as follows:

(45) presents a similar phenomenon in Georgian. (45a) is an ordinary agentive sentence, while (45b) is an involuntary state construction (see, e.g., Kallulli 2006; Rivero 2009). They display the same Nom-Acc and Dat-Nom Case alignment alternation and sensitivity to agentivity observed in (43).

What is worth noting here is the use of the non-active morpheme -in in (45b), which Nash (2020) argues to be a deagentivizer. I suggest that the morpheme -in expresses an expletive head, which deactivates the thematic property of Voice, namely its ability to introduce an argument. The analysis of (45) thus parallels that of (43). For comparison, I provide the structural representation of (45b) below.25

Let’s now turn to the phenomenon of causativization, which is the mirror image of expletivization. An example of this is the Samoan ergative-middle alternations (Tollan 2018). (47a) involves a Polynesian middle verb,26 which is formally transitive, but does not follow the canonical Ergative-Absolutive (Erg-Abs) pattern. Instead, it exhibits Absolutive-i-Case27 (Abs-i-Case) alignment, which is associated with low agentivity. However, when the ergative suffix -C(i)a is added to a middle verb, as in (47b), the Case alignment shifts to the Erg-Abs frame, and the semantics of the verb also changes from low agentivity to high agentivity.28

I assume that the absolutive argument in (47a) is an Affectee, while the ergative argument of (47b) is an Agent. The suffix -C(i)a can be analyzed as the realization of Voice (or Voice plus v_init), introducing a Causer. Therefore, I treat (47b) as the result of adding Voice (or the complex [Voice, v_init]) to the structure of (47a). The subject of (47b) is both a Causer and an Affectee (i.e., an Agent). Assuming that -C(i)a realizes Voice, the structures of (47a) and (47b) can be represented as below.

Similar alternations involving causativization are also found in Tatar, which employs the causative morpheme -t to increase volitionality of a predicate (Kittilä 2009). This meaning shift is associated with changes in argument coding. Specifically, a sentence shows the Dat-Nom alignment with the unintentional reading (49a) and the Nom-Acc alignment with the volitional reading (49b). The analysis of (49) parallels that of (47).

The above discussion demonstrates that Agent-Affectee alternations are widely attested. There are two subcases: in the first one, Agents become Affectees when Causers are expletivized; in the second, Affectees become Agents when Causers are added through causativization. This suggests a derivational relationship between Agents, Causers, and Affectees. This is exactly what the decomposition analysis of Agents predicts, but not the standard analysis. The present analysis is thus preferable.

The two analyses under consideration also make different predictions regarding causativizing agentive predicates, which are predicates with Agent subjects. Nie (2022) identifies two types of causatives: bi-eventive causatives and mono-eventive causatives.29 In bi-eventive causatives, a causing event30 e is added to another event e′. The two events stand in a cause relation. The participant α introduced by the causing event e is understood as the causer,31 while the participant β introduced by the caused event e′ is understood as the causee (see (50a)). I use θ to denote whatever θ-role β bears, which may vary depending on whether the causative construction is bi-eventive or mono-eventive. In mono-eventive causatives, both the causer α and the causee β are participants of the event e (see (50b)). Crucially, Nie (2022) proposes that Causer (her Cause) and θ must be unrelated since it is not allowed for one event to take two Causers, as illustrated in (50c).32

Tollan (2018) observes that the Samoan causative prefix fa’a cannot be added to Erg-Abs predicates, which are highly agentive and take Agent subjects (see (51)). This observation suggests that neither (50a) nor (50b) applies to fa’a causatives. The fact that (50a) is not applicable implies that this prefix does not introduce another event. However, why cannot (50b) represent the semantics of (51b), making it grammatical?

The standard analysis assumes that Causers and Agents are unrelated, making (50b) a possible representation of (51b) if θ in (50b) stands for Agent. It is then expected that causitivizing agentive predicates should be possible, contrary to fact. In contrast, the decomposition analysis posits that Agents include Causers as their sub-arguments. As a result, the semantic representation of (51b) under this analysis would lead to the clearly banned scenario in (50c), where one event has two Causers (see (52)).

Interestingly, Tollan (2018) observes that fa’a can nevertheless be added to middle verbs (see (53)).

Recall from section 5.1 that middle verbs are low in agentivity and have been argued to take Affectee subjects. (53b) then has the partial semantic representation in (54), which instantiates (50b). The contrast between (51b) and (53b) can thus be straightforwardly accounted for by the decomposition analysis, in contrast to the standard analysis.

Georgian morphological causatives exhibit a similar pattern to Samoan. The causative affix -a does not allow direct embedding of agentive predicates (Nash 2020). However, unlike Samoan, Georgian allows causativizing agentive predicates by deagentivizing them with the -in affix, which, as argued before, expletivizes the Voice of the embedded agentive verbs. Moreover, the Case alignment of the embedded structures is reminiscent of lowly agentive predicates, with causees in dative and objects in nominative. I argue that the causee ‘the girl’ in (55) is an Affectee rather than an Agent.

The three adverbials used in (55) are not ambiguous regarding the scope. (55) can only mean “Keti did this intelligently/intentionally/with pleasure, but not the girl.” This indicates that Georgian morphological causatives are mono-eventive and do not distinguish between causing events and caused events. If Causers and Agents were two independent arguments, as assumed in the standard analysis, we would expect the causative affix -a to attach directly to agentive predicates, but this is not the case. Therefore, the fact that agentive predicates must be deagentivized whenever causativized strongly argues for the proposed analysis. Since mono-eventive causatives do not allow the stacking of Causers, the embedded agentive predicates cannot retain their Agent arguments, which include Causer sub-arguments. Through deagentivization, Voice becomes semantically null, and Agents, losing their Causer sub-arguments, are turned into Affectees, which are possible according to (50b). The decomposition analysis thus makes more accurate predictions regarding restrictions on causativizing agentive predicates. Also note that Samoan and Georgian morphological causatives instantiate additional cases where Causers and Affectees are expressed by two separate (non-coindexed) arguments, just like DOCs in English.

As briefly discussed in section 5.1, agentivity plays a crucial role in determining Agent-Affectee alternations. Agent subject constructions (43a, 45a, 47b, 49b) were argued to be more agentive than their Affectee subject counterparts (43b, 45b, 47a, 49a). This section will explore the syntactic encoding of agentivity and its connections with Agents, Causers, and Affectees.

There is evidence that Affectee subject constructions exhibit low agentivity. Oblique subject constructions (OSCs), such as unintentional causer constructions and involuntary state constructions, have been analyzed as having Affectee subjects (e.g., Cuervo 2003). These constructions share the property of low agentivity. Drawing on Old Norse-Icelandic, Latin, Ancient Greek, Old Russian, and old Lithuanian, Barðdal et al. (2012) classify OSCs into two major groups: experience-based predicates and happenstance predicates. The former group subsumes verbs of emotion, bodily states, cognition, attitudes, and perception, while the latter includes verbs of gain, success, happening, hindrance, ontological states, and speaking. In addition, there are three types that do not fit in either group and are listed individually: verbs of modality (obligation), evidentiality, and possessives. All the predicates that take oblique subjects express low volitionality and lack of control, indicating low agentivity.

The difference in the degree of agentivity between Agent subject constructions and Affectee subject constructions can be explained as follows. The former includes an additional head, Voice (or the complex [Voice, v_init]), which is absent in the latter. It is thus plausible to attribute the source of agentivity to Voice (or [Voice, v_init]). As a result, Agents signal high agentivity. Affectees and Themes, on the other hand, signal low agentivity.

The low agentivity of Affectees and Themes gives an account of their diverse interpretations. Affectees can be interpreted as broadly as instruments, locations, experiencers, beneficiaries, goals, sources, etc. Themes can also be interpreted with these roles, but classifying their other interpretations is often challenging. For example, an eatee and a hitee share little in common except being highly affected by the events. In fact, any non-agentive argument can function as an Affectee or Theme. Defining Affectees and Themes as participants in the process subevent avoids associating a specific meaning with either θ-role, explaining their wide range of interpretations.

It should be noted that a few studies have also argued for the existence of two external argument introducers, one being highly agentive and the other lowly agentive. For instance, Kim (2011b) proposes that Voice is highly agentive, while high Appl right below Voice is lowly agentive. Massam (2009) and Tollan (2018) suggest that the external argument of transitive verbs is introduced by Voice, while the sole argument of unergatives is introduced by a lower v (at least in Niuean and Samoan). Tollan (2018) further suggests that the specifiers of the two heads differ with respect to agentivity, with Spec, VoiceP reserved for a highly agentive argument (her proto-high Agent) and Spec, vP for a lowly agentive argument (her proto-low Agent). Kim’s Voice and high Appl, and Tollan’s Voice and v, correspond to Voice and Appl as defined here.33

Massam (2009) and Tollan (2018) argue that the sole argument of unergatives in Niuean and Samoan is lowly agentive and is introduced by v, which corresponds to Appl as defined here. However, the sole argument of unergatives in Germanic and Romance languages does not seem to differ in agentivity from the external argument of transitives and is typically analyzed as an Agent. It is often assumed that all unergatives have an internal argument, whether overt or implicit (e.g., Hale & Keyser 1993). Under this analysis of unergatives, the external argument of unergatives (X in (56a)) is introduced in Spec, VoiceP, while their implicit internal argument (Y in (56a)) is introduced in the complement of V. This analysis can be easily adapted into the proposed theory (see (56b)): the sole argument of unergatives is first introduced in Spec, ApplP and then moves to Spec, VoiceP, while the implicit internal argument is generated in Spec, TransP.

However, the assumption of an implicit internal argument for unergatives is controversial. For instance, Preminger (2012) argues against the existence of implicit objects in Basque unergatives. This poses a problem for the proposed analysis. Since Agents take Affectees as one of their sub-arguments, and Affectees, being secondary Themes, presuppose the presence of a Theme, how can we capture the intuition that the sole argument of unergatives is an Agent without assuming an implicit internal argument (i.e., a Theme)?

I suggest that the sole argument of unergatives is also composite but is made up of a Causer and a Theme rather than a Causer and an Affectee. Given that Affectees and Themes are both participants in the process and only differ in their prominence, as discussed in section 2.4, it is expected that the combination of a Causer and a Theme also results in an Agent reading. Consequently, unergatives and causativized unaccusatives have the same structure (see (57)), but differ in how their arguments are realized. In unergatives, the Causer and the Theme are realized by a single argument, whereas in causativized unaccusatives, they are realized by two separate arguments.

Certain intransitive verbs show mixed properties regarding unaccusativity tests. They can behave like unergatives on certain occasions and like unaccusatives on others. This phenomenon is known as unergative/unaccusative alternations. One commonly used test for unaccusativity is auxiliary selection (see, e.g., Burzio 1986). In Italian, unergatives select the auxiliary avere ‘have,’ while unaccusatives select the auxiliary essere ‘be.’ When a predicate is atelic and takes a locative PP and a for X time adverbial, as in (58a), it behaves like an unergative. On the other hand, when a predicate is telic and takes a directed motion PP and an in X time adverbial, as in (58b), it behaves like an unaccusative (see, e.g., Levin & Rappaport Hovav 1995; Sorace 2000).

Commonly assumed analyses of unergatives and unaccusatives would represent them as in (59) (depending on the theory, VP in (59a) may or may not project). Since the two structures in (59) have nothing overlapping, the standard analysis has difficulty in capturing the alternation in (58).

The proposed analysis represents unergatives and unaccusatives as in (60). The structure in (60b) is a subpart of (60a). When the Causer (together with v_init) is expletivized,34 the resultant structure is identical to that of an unaccusative. This expletivization operation has been employed to account for aspectual si constructions (in section 4.1) and Agent-Affectee alternations (in section 5.1) and does not introduce any new machinery to the theory. The decomposition analysis thus gives a rather simple account of the unergative/unaccusative alternations and unifies them with the two other types of argument structure alternations discussed before.

Further evidence for analyzing the sole argument of unergatives as Causer + Theme comes from causativized unergatives. The proposed theory predicts that when an unergative is causativized, the Causer argument should be dropped, leaving the Theme as the only argument. This mirrors the behavior of a transitive verb’s Agent argument, which becomes an Affectee when causativized, as discussed for Georgian in section 5.2. This is indeed borne out in languages such as Hebrew, Hindi-Urdu, Turkish, and Sason Arabic, as discussed in Arad (1998) and (2023). Due to space constraints, I provide just one argument. In Hindi-Urdu, reduced relatives require the presence of an internal argument (specifically a Theme + Resultee). While reduced relatives cannot be formed directly from unergatives (61a), they can be formed from causativized unergatives (61b). I suggest that in (61b), not only is the Causer argument removed, stranding the Theme argument, but a Resultee is also added.35^,36