Unergatives are different: Two types of transitivity in Samoan

This paper provides arguments in favour of a non-unified treatment of transitive and unergative verbs, based upon the patterning of unergative constructions in the ergative-absolutive language Samoan. Building upon a proposal by Massam (2009; to appear), I argue that unergative subjects in Samoan are merged lower than transitive subjects, the result being a difference in case marking patterns associated with each verb type. In the spirit of much work which has advocated for a split v P structure (e.g., Pylkkännen 2008; Harley 2013; Legate 2014; a.o.), I propose that unergative subjects are merged in the specifier of v P, while transitive subjects are introduced in a higher projection (VoiceP). This proposal is motivated primarily from split case patterning: while Samoan unergative subjects appear with absolutive case, addition of an object to an unergative verb does not yield the typical ergative-absolutive pattern associated with canonical transitives. Instead, a non-ergative case pattern arises, in which the subject is absolutive, and the object is marked with the prenominal marker i . The Samoan unergative + object construction bears similarities with another set of two-place (so-called middle verbs) which exhibit the same abs -i case frame. Despite the absence of ergative case, both unergative + object and middle constructions are syntactically transitive with respect to various language-internal diagnostics. I argue that the case split results from differing case assigning properties of v and Voice: i is analysed as structural accusative case, assigned to the object by v 0 , while ergative subjects are assigned case by Voice 0 . The division of external arguments across two VP-external projections can be captured by expanding Dowty’s (1991) framework of thematic proto-roles, whereby unergative and middle “proto low ” ( v P) agents encompass a subset of the semantic properties of full-fledged transitive “proto- high ” (VoiceP) agents. The additional properties of proto-high agents correspond to additional phrasal structure. these extensively This paper focuses on the behaviour of unergative and other non-ergative subjects in the Polynesian language Samoan, and presents an alternative approach to the widespread existence of ergative languages with non-ergative unergative subjects, which does not rely on the aforementioned transitivity condition. I argue that subjects of unergative verbs in Samoan do not receive ergative case (see 2a) because they are not merged in the same position as transitive (ergative) subjects ; rather, they are merged below the phrasal projection in which ergative case is assigned (see Massam 2009; to appear). I motivate this conclusion on the basis of several parallels between unergative and a type of (seemingly) semi-transitive construction known as the Polynesian “middle”. Samoan middle verbs (e.g., mulimuli ‘follow’ , mana’o ‘want’) are canonically transitive, but do not follow the typical ergative-absolutive pattern; instead, they exhibit the same absolutive – i case frame as the unergative + object in (2c). demonstrates that ergative- and accusative-marked arguments indeed do co-occur. Here, the high-agent causer ( le tama „the boy‟) receives Spec Head ergative case from Voice 0 , while the low-agent causee ( le teine „the girl‟) and object ( le masi „the cookie‟) ar e both assigned structural case: the causee receives absolutive case from T 0 , and the object receives accusative case from v (which is able to assign accusative, by having introduced the external causee). The structure of (56b) is shown below 27 .


Introduction
Unergative verbs (e.g., dance, cry, laugh, dream, work) are often treated as being structurally akin to transitive verbs, other than lacking an internal argument. The widely-adopted Split-VP hypothesis (Chomsky 1995;Kratzer 1996;Marantz 1997;i.a.) posits that transitive and unergative subjects (i.e., external arguments) are base-generated outside of VP, in the specifier of vP, and are both assigned their theta-role (e.g., agent) by v 0 . Internal arguments (i.e., objects and unaccusative or passive subjects) are selected for and thetamarked (e.g., theme) by V 0 itself (1).
In this paper, I argue that the absence of an object is not the only differentiating factorat east not universally. I present novel data from the Polynesian language Samoan, in which ddition of an object to an unergative verb does not transitivize the predicate in the expected anner, insofar as the ergative-absolutive case pattern typical of Samoan transitive onstructions such as (2b) does not surface. The cognate object in (2c) appears with the renominal marker i, which I refer to (for now) under the descriptively neutral label of i-case 1 . All unreferenced data presented in this paper from are from consultation with two native speaker consultants. ther data are drawn from existing syntactic literature. Abbreviations for sources are: CH = Chung 1978;MH = osel and Hovdhaugen 1992;SE = Seiter 1978. Unergative predicates differ from transitives only insofar as they do not contain an overt internal argument, consisting only of an external subject, merged in the specifier of vP. Under Hale and Keyser's (1993) approach, unergatives are simply concealed transitives, in which the object has been lexically incorporated into the verb root. The absence of an overt (i.e., non-incorporated) object renders the verb formally intransitive.
In this paper, I argue that the absence of an object is not the only differentiating factor -at least not universally. I present novel data from the Polynesian language Samoan, in which addition of an object to an unergative verb does not transitivize the predicate in the expected manner, insofar as the ergative-absolutive case pattern typical of Samoan transitive constructions such as (2b) does not surface. The cognate object in (2c) appears with the prenominal marker i, which I refer to ( Ergative is commonly treated as an inherent case, assigned to the external argument by v 0 (e.g., Woolford 1997;Legate 2002;Aldridge 2004;Laka 2006;Coon 2013;a.o.), as in (3).
(3) Inherent Ergative case vP AGENT v' ERGATIVE v VP As Baker and Bobaljik (2017) point out, this analysis predicts that all external arguments in an (inherent) ergative language should appear with ergative case; that is, transitive and unergative subjects. In a great majority of ergative languages (Samoan included), however, unergative subjects appear with absolutive case. Inherent ergative theories (e.g., Legate 2002;Aldridge 2004) account for this by positing a "transitivity condition" on v 0 (or T 0 ), whereby As Baker and Bobaljik (2017) point out, this analysis predicts that all external arguments in an (inherent) ergative language should appear with ergative case; that is, transitive and unergative subjects. In a great majority of ergative languages (Samoan included), however, unergative subjects appear with absolutive case. Inherent ergative theories (e.g., Legate 2002;Aldridge 2004) account for this by positing a "transitivity condition" on v 0 (or T 0 ), whereby ergative case is only available for an external argument when v 0 is transitive (i.e., if an object is also merged, internally to VP), and not when v 0 is intransitive (i.e., when no object is present). The transitivity condition raises an unresolved theoretical concern: by what mechanism(s) exactly is v 0 able to determine whether V 0 selects for an object or not? Precisely how a syntactic head (such as v 0 ) is able to determine the compositional structure of its complement remains unsettled. In a recent paper, Baker and Bobaljik (2017) seize upon this issue as an argument against the treatment of ergative as an inherent case, arguing instead that ergative is a dependent case, assigned configurationally to a higher DP (in some local domain) in the presence of a second, non-oblique DP. Under their account, ergative case fails to appear on unergative subjects due to lack of a local case competitor (i.e., an object).
On the other hand, inherent ergative theories can straightforwardly account for languages in which unergative subjects do receive ergative case (e.g., Basque, Georgian, Hindi, Kashmiri): v 0 in these languages assigns ergative case to all external arguments, regardless of verb transitivity (see also Sheehan 2014). Baker and Bobaljik (2017) account for ergative unergatives by appealing to the presence of a covert cognate object, which acts in these languages as a case competitor for the unergative subject (which is thereby essentially a transitive subject). Assuming the presence of a covert object for these language is not uncontroversial, however: Preminger (2012), for example, argues extensively against the presence of covert objects in Basque unergative constructions.
This paper focuses on the behaviour of unergative and other non-ergative subjects in the Polynesian language Samoan, and presents an alternative approach to the widespread existence of ergative languages with non-ergative unergative subjects, which does not rely on the aforementioned transitivity condition. I argue that subjects of unergative verbs in Samoan do not receive ergative case (see 2a) because they are not merged in the same position as transitive (ergative) subjects; rather, they are merged below the phrasal projection in which ergative case is assigned (see Massam 2009;to appear). I motivate this conclusion on the basis of several parallels between unergative and a type of (seemingly) semitransitive construction known as the Polynesian "middle". Samoan middle verbs (e.g., mulimuli 'follow', mana'o 'want') are canonically transitive, but do not follow the typical ergative-absolutive pattern; instead, they exhibit the same absolutive -i case frame as the unergative + object in (2c).

(4)
Samoan middle Sā mulimuli [le leoleo] [i le au gaoi]. pst follow det police.abs i det robbers 'The police followed the robbers.' The core of this proposal is that Samoan has two external argument positions, and two transitive case frames, differentiated semantically and syntactically on the basis of the degree of effort and volition on the part of the external argument, and/or its resultant effect (upon a theme argument, if present; I remain agnostic as to whether Samoan unergatives may have covert objects). Expanding Dowty's (1991) theory of thematic protoroles, these two types of external arguments correspond to what are termed here as proto -high agents (highly effortful and/or affecting agents or experiencers) and proto -low agents (non-effortful and/or non-affecting agents or experiencers). Low agents (i.e., unergative and middle subjects) are generated in the specifier of vP; high agents are introduced in a higher projection -namely VoiceP -which is responsible for assigning ergative case to the external argument, in a specifier-head relationship (similar to that shown in (3)). Crucially, however, middle and unergative predicates lack VoiceP. I argue for an analysis of Samoan i-case of (2c) and (4) as structural accusative, assigned to an object by v 0 when a proto-low agent is merged in the specifier of vP. Essentially, it is the difference in the position of the subject which gives rise to the ergative vs. nonergative case split. This analysis also demonstrates how inherent-based approaches to ergative case can accommodate non-ergative unergative subjects, despite the concern raised by Baker and Bobaljik.
The structure is as follows: the next section presents the relevant essentials of Samoan syntax, discussing the source of ergative and absolutive case, and the behaviour of unergative predicates. Section 3 deals with the middle construction and the nature of the middle object, and section 4 considers how unergative and middle predicates differ semantically from ergatives. The core syntactic analysis is presented in section 5. Section 6 discusses some wider implications of the proposal, including how languages such as Basque, in which unergative subjects appear with ergative case, may be accounted for. Section 7 concludes.

Clause structure basics
Samoan (Samoic Outlier, Polynesian, Austronesian) is head-initial, and displays a predominantly VSO word order; 2 a T(ense) -A(spect) -M(ood) marker appears clause-initially; (5). Collins (2016) builds a compelling argument -on the basis of coordination, negation, and complex predicate patterns -that verb-initial word order is derived from movement of the VP 4 to a functional projection below TP (labelled FP in (6); its precise identity is undetermined). 5 Crucially, a DP object moves out of the VP -to spec vP -prior to (remnant) VP fronting (I will later posit that this is for interpretative reasons, following Aldridge 2004). Following Massam (2001a) and Longenbaugh and Polinsky (2017) for Niuean, and Collins (2013; for Samoan, I assume that movement of an internal DP targets the lower vP specifier; the external argument is merged in a higher, outer vP specifier, akin to "Tucking in" proposed by Richards (1997;. 6 This structure is shown in (6) (the specifics of v/Voice will be revised in forthcoming sections, however). 2 VOS can also be used where the object is focused (Mosel and Hovdhaugen 1992). I focus only on VSO here, although see footnote 24 for remarks on how VOS may be derived. 3 The determiner le can be interpreted as either definite or indefinite. 4 Although Samoan also allows non-verbal predicates (see Collins 2016 for discussion), this paper focuses only on VP predicates. 5 Various labels have been given to the projection labelled as FP in (6): Massam (2001b) uses the label IP, and situates the TAM marker in a higher CP (reflecting the complementary distribution of TAM and clausal markers in Niuean); Aldridge (2004) labels this node AspP in order to capture parallels between Niuean and Seediq (Formosan). Because nothing crucial in the present paper hinges on the identity of this node, I follow Collins (2016) in using the neutral label of FP. 6 As noted by a reviewer, Tagalog is proposed to differ in this regard; Rackowski (2002), Aldridge (2004), and Rackowski and Richards (2005) argue that object movement targets the outer vP specifier. This captures the VOS word order of Tagalog object shift (see e.g., Rackowski & Richards 2005: 572) and the ability of the absolutive argument, but not the ergative, to undergo A-bar movement ("syntactic ergativity"; see e.g.,

Ergative and absolutive case
Samoan has an ergative-absolutive alignment: transitive subjects are marked with the prenominal ergative marker e (7a), while transitive direct objects (7a), and intransitive subjects (7b, c), appear (mostly) with no case marker (glossed as abs). Ergative is commonly analysed as an inherent case (e.g., Woolford 1997;Legate 2002;Aldridge 2004;see Collins 2013 for the Samoan ergative), assigned to an external argument by v 0 , hand-in-hand with an external (e.g., agent) theta role.

Ergative and absolutive case
Samoan has an ergative-absolutive alignment: transitive subjects are marked with the prenominal ergative marker e (7a), while transitive direct objects (7a), and intransitive subjects (7b, c), appear (mostly) with no case marker (glossed as ABS). Applying the inherent ergative analysis to Samoan, however, requires some modifications. As Koopman (2012) points out, the Samoan ergative is not restricted to subjects that are traditionally regarded as semantic "agents": inanimate DPs (e.g., natural forces, items of machinery) may appear with ergative case, as do experiencers of certain psych-predicates (e.g., lagona 'feel', iloa 'know'; I discuss this in further detail in §5.3). Note that such criteria may not necessarily be viewed as grounds for rejecting the inherent ergative approach entirely: the theta-role tied to ergative case, and assigned by v 0 to the externally-merged DP, may not necessarily be exclusively associated with a narrow set of semantic criteria such as animacy Aldridge 2004: 27-8). The landing site of a moved object is a possible point of parametric variation across languages, which is (to my knowledge) largely unexplored, although Longenbaugh and Polinsky (2017) provide evidence from quantifier float in French that shifted objects also target the lower vP specifier in Romance languages.
Applying the inherent ergative analysis to Samoan, however, requires some modifications. As Koopman (2012) points out, the Samoan ergative is not restricted to subjects that are traditionally regarded as semantic "agents": inanimate DPs (e.g., natural forces, items of machinery) may appear with ergative case, as do experiencers of certain psych-predicates (e.g., lagona 'feel', iloa 'know'; I discuss this in further detail in §5.3). Note that such criteria may not necessarily be viewed as grounds for rejecting the inherent ergative approach entirely: the theta-role tied to ergative case, and assigned by v 0 to the externally-merged DP, may not necessarily be exclusively associated with a narrow set of semantic criteria such as animacy (see also Legate 2012, for related discussion). In view of this, and in anticipation of the present analysis, I adopt a modified version of the inherent hypothesis, treating ergative as being assigned by v 0 to its external argument under a specifier-head Agree relationship (see also Assmann 2015). I refer to this as "Spec-Head ergative", although it does not differ greatly from the inherent approach, other than that it is not necessarily tied to an "agent" theta-role, in the traditional narrow sense. Theta roles are discussed in greater detail in section 4. As for absolutive case, Legate (2002; and Aldridge (2004) identify two types of absolutive systems. In the first type, absolutive case is assigned structurally by (finite) T 0 to the closest case-less DP (much like nominative case): either to an intransitive subject, or a transitive object (recall that the transitive subject has already received ergative case from v 0 , leaving the object as the closest case-less DP). In the second type, absolutive is a morphologically unmarked default for two separate abstract cases: nominative, assigned by T 0 to intransitive subjects, and accusative, assigned by v 0 to transitive objects. Although Samoan absolutive is less marked than ergative, Yu (2015) argues that it does not have null phonological spell-out: it is realized by a high tone on the last mora of the word preceding the absolutive. 7 Furthermore, Mosel and Hovdhaugen (1992: 143) and Yu and Özyildiz (2016)  Since Samoan absolutive is therefore not morphologically unmarked, Koopman (2012) proposes as a null hypothesis that absolutive is assigned by a single source, namely T 0 , and is therefore analogous to nominative case. Conversely, Collins (2013) proposes that Samoan absolutive is a morphological default for abstract nominative and accusative case, and falls under the second category. Because nothing crucial in the analysis presented in this paper hinges upon the choice between these two alternatives, I will for the sake of concreteness assume the single source approach (i.e., absolutive = nominative, following Koopman 2012); see footnote 23 however for discussion of how Collins' (2013) analysis may also be accommodated under my proposal.

Unergatives (and their objects)
Subjects of unergative verbs in Samoan are always absolutive; ergative case never marks an unergative subject (10). Following the Spec-Head ergative approach prima facie, this suggests that they are not merged in the specifier of vP (or at least, not an ergative caseassigning vP).
(10) Sā siva (*e) le teine. pst dance (*erg) det girl.abs 'The girl danced.' Deal (2010) proposes a version of ergative case assignment in which object agreement (with v 0 ) is a necessary condition for ergative case assignment to a subject. This type of approach could potentially capture the Samoan data in (10): the absence of an object of dance means that ergative case assignment necessarily fails. The picture becomes more complex, however, when unergative verbs are transitivized: the addition of an object does not restore ergative case marking on the subject. As shown in (11) though (13), unergatives do not exhibit an erg-abs case configuration even when an object is added; instead, the subject appears as absolutive, and the object is marked with prenominal i. There are two exceptions to this pattern. As for the English unergative verb sing, Samoan has two different verbs, usu and pese. Usu obligatorily follows an erg-abs pattern, but cannot be used without an object, as in (15). Milner's (1966) Samoan dictionary translates usu into English as 'sing (a song)' (Milner 1966: 304). Usu is therefore obligatorily transitive (and may not be a direct translation of English sing). Unlike usu, pese exhibits the same unergative-transitive alternation as English sing. When an object is present, as in (16b), the abs-i frame is obligatory -just as in (11) A consultant reported that abs-i patterning would have the meaning of 'pray to (someone)'. It is therefore possible that an erg-abs pattern arises as in order to avoid ambiguity between the frames [pray + DP] and [pray to + DP]. Because it is likely that pray is more frequently used with the latter frame than with the former, the abs-i case marking in (17b) is blocked, as a means of differentiating [pray + DP] from the more expected [pray to + DP] (i is used as a prepositional marker in Samoan, as I will discuss in depth shortly). It may also be that the English translation of (17b) is not an entirely accurate rendition of the original Samoan.
The exceptions of usu ('sing') and tatalo ('pray') aside, Samoan unergative + object constructions overwhelmingly follow a non-ergative case pattern. This suggests that the presence of an object is not the sole syntactic difference between unergative and transitive predicates in Samoan: when transitivized by the addition of an object, unergative verbs do not conform to the same (ergative-absolutive) case frame observed in transitive sentences such as (5). What underlies this difference? In the next section, I investigate the abs-i case pattern further, by turning to another set of predicates in which it appears: the Samoan "middle".

Samoan "middle" verbs
A number of Samoan verbs show the same non-ergative pattern observed with unergatives, appearing with an absolutive subject and an i-marked object (17). Such verbs are referred to as "middle" verbs in Polynesian literature (e.g., Chung 1978;Seiter 1978). Middle verbs are so-called because they are traditionally regarded by Polynesian linguists as semi-transitive: they require both a subject and an object, but do not exhibit the ergabs case array associated with prototypical transitives. Note that, while the term "middle" is often used by syntacticians to refer to a set of derived intransitive constructions (e.g., The bread cuts easily), I use it in this paper to describe the aforementioned class of verbs, in keeping with Polynesianist tradition.  Middles verbs include verbs of perception, emotion and communication (e.g., want, love, listen to, call, meet with) and a number of other verbs including follow, wait for and visit. Both Chung (1978) and Seiter (1978) point out that one common semantic characteristic of middles which distinguishes them as a class from verbs which follow the erg-abs patterns is that they involve a lower degree of affectedness than regular transitive verbs. They typically describe events that " […] do not affect the direct object immediately" (Chung 1978: 47), such that "[…] their objects are only indirectly affected, if at all" (Seiter 1978(Seiter : 1291. Examples of Samoan transitive (erg-abs) and middle (abs-i) verbs are given below.   Coon (2013), and Coon and Preminger (2017), argue that ergative "splits" (i.e., non-ergative case frames in an ergative language) of the type exemplified in (19) arise from the structural position of the object in the non-ergative pattern, claiming that it is an indirect prepositional object as opposed to a direct object of the verb; the subsequent absence of a direct object renders the verb intransitive. By this analysis, Samoan prenominal i would be considered an oblique, or prepositional case marker (20).
At first glance, treating the middle object as an oblique object seems to be on the right track: the case marker i indeed functions as an oblique marker in Samoan. It is used to mark instrumental or locative DPs (21), or to introduce a ditransitive goal argument (22 i pro 8 pl them some money.abs 'I gave some money to all of them.' As evidenced by Chung (1978) and Seiter (1978) however, Samoan middle objects do not behave syntactically as indirect or oblique objects with respect to three tests for direct objecthood and transitivity. As far as is shown by the behaviour of object quantifier float, pronominal clitics, and Pseudo Noun Incorporation, middle objects in fact behave in the same way as absolutive direct objects. I discuss each of these tests in the following subsections.

Quantifier float
The first piece of evidence to suggest that middle objects are direct, as opposed to indirect objects is based upon the behaviour of the quantifier 'uma ('all'). Seiter (1978) shows that 'uma can be moved from its modifying DP (23a) and attached as a post-verbal clitic; (23b, c, d). This applies unrestrictedly to any non-pronominal subject (23b, c) or direct object (23d). According to Seiter (1978), however, oblique quantifier float is subject to two restrictions which do not apply to direct objects (or subjects). Oblique quantifiers may undergo float only if they are both (i) animate and (ii) the first DP after the verb (hence the grammaticality of 24), but not otherwise, as shown in (25) As Seiter notes, this supports a syntactic treatment of middle objects as direct objects. A full analysis of the restrictions on oblique 'uma float is beyond the scope of the present paper; the relevant point here, is that i-marked objects of middle verbs pattern as direct objects, as opposed to oblique DPs.

Clitic placement
Samoan has a set of pronominal subject clitics, which are morphologically distinct from full pronouns (a full paradigm is given in Mosel & Hovdhaugen 1992: 122, 374). While subjects are typically post-verbal (27a), subject clitics must appear in a preverbal position, without case marking; (27b). Third singular pronominal subjects may cliticize only in transitive clauses (28a); cliticization of a third singular subject in an intransitive clause is ungrammatical 10 -even with a prepositional DP present; (28b). First and second person (1/2) pronouns are widely argued to contain a richer featural composition than third person (3rd) pronouns. For example, Harley and Ritter (2002) among many others -argue that 1/2 pronouns encompass a [+participant] feature which 3rd pronouns lack. Adopting this approach, the data in (28-29) suggest that Samoan subject cliticization is sensitive to [+participant] (and/or indeed any other feature common to 1/2 pronouns but not present in 3rd pronouns), unless a direct object is present. While an analysis of the sensitivity of cliticization to transitivity is a topic for future research, the crucial observation to be made here is that middle objects allow for 3rd subject pronoun clitics (29b), in contrast to oblique objects (28b), which does not. The Samoan middle, therefore, displays the hallmarks of a transitive predicate, as opposed to an intransitive predicate with an oblique object.

Pseudo-Noun Incorporation
Samoan displays VOS order with bare NP objects, which are interpreted semantically as indefinite, non-individuated, and non-affected (30, 31). The subject bears absolutive case.  Massam (2001b) observes this same phenomenon in the related Polynesian language Niuean, proposing that the type of VOS in (30b, 31b) is derived from lack of NP movement out of the VP prior to VP raising (to FP; see §2.1). Because the object does not move out of the VP, it raises along with the verb, thus appearing directly adjacent to it. Massam refers to this as "Pseudo Noun Incorporation" (PNI); note that that the object does not morphologically incorporate (hence "pseudo").
Samoan PNI applies to all direct objects (bar a few lexical exceptions), but not to oblique locatives, 11 as in (32b), or to ditransitive goal arguments (33b) -even though the direct object in ditransitive constructions may incorporate (33c).
pst run house they.abs 'They were running in houses.' The behaviour of middle objects is therefore akin to that of absolutive direct objects, and unlike ditransitive goal objects or locative objects -suggesting that middle i-case is not oblique.

Unergatives again
Before discussing the identity of middle i-case, I return to the unergative + object construction (recall that this displays the same abs-i case pattern as middles). Applying the above tests for transitivity and direct objecthood to the unergative + object construction, we find that it indeed behaves in the same manner as the middle construction: the quantifier 'uma may be floated despite inanimacy of the object (35b), third singular subjects may cliticize (35c), and objects may undergo PNI (35d). Like middles, unergative + object constructions exhibit the syntactic hallmarks of a transitive construction, consisting of a subject and direct object -and much like an erg-abs transitive. This is unsurprising in view of the standard syntactic treatment of unergatives as discussed in section 1, namely that transitive and unergative constructions are alike in consisting of a VP-external subject, but differ only in the presence vs. absence of an internal object. The question to be addressed is why Samoan unergatives -and middles -do not display a typical transitive erg-abs case array. I approach this from two perspectives: first, how middle/unergative and erg-abs verbs differ semantically ( §4) and second, the resulting structural differences which give rise to the two case frames ( §5).

Unifying unergatives and middles
It has been noted by both Chung (1978) and Seiter (1978) that Samoan middle verbs differ semantically from erg-abs verbs on the basis of effect: the object of an erg-abs verb is typically affected, while the object of a middle verb is not. Cruse (1973;discussed by Beavers 2011) proposes a straightforward diagnostic for affectedness: the "what happened to X is Y" test. As exemplified in (36), both middle and unergative verbs fail this test, while erg-abs verbs (typically) pass it.
(36) a. Ergative-absolutive verbs (cf. 19a) What happened to the dog is that the boy hit it. What happened to the ball is that the girl caught it.
b. Middles and unergatives # What happened to the car is that the police followed it. # What happened to the waltz is that the girl danced it. Travis (2010) proposes that affected objects are, in many languages, structurally higher than non-affected objects. In Samoan, however, there is no evidence for a distinction in structural height between absolutive and i-marked objects, which behave alike with respect to the syntactic and morphological diagnostics discussed in section 3. Furthermore, there is considerable evidence that object-affectedness is not the fundamental source of the split between erg-abs and abs-i case frames. First, verbs such as mata'i ('observe') and mātamata ('see'/'look at') follow erg-abs and abs-i patterns (respectively), yet an (absolutive) object of mata'i ('observe') is not any more intuitively 'affected' than a (middle) object of matamata ('see'/'look at'). Second, not all erg-abs verbs pass Cruse's "what happened to X is Y" test (37). Some such verbs, notably tuli ('hunt') and su'e ('search for'), are known to be compatible with objects which do not exist (e.g., 'the meaning of life', 'the city of lost gold'), and therefore cannot possibly be affected by the event. 14 (37) # What happened to the meaning of life is that the boy searched for it. # What happened to the city of lost gold is that the girl hunted for it.
What can be said of these verbs, however, is that they necessarily require a volitional subject; one which is capable of making a certain effort and/or is consciously trying to create a certain situation (e.g., discovery of the meaning of life/the city of lost gold). Thus, the distinction between erg-abs and abs-i case frame verbs is potentially better characterized in terms of properties of the subject, as opposed to those of the object. erg-abs verbs, such as those in (19a) tend to require highly agentive, effortful, and volitional subjects, which very often (but not always) trigger an affect or change of state in an object. Subjects of middle/unergative (abs-i) verbs, on the other hand, tend to be less agentive, effortful and volitional, and are less likely to trigger an effect upon an object. Such generalizations are tendencies only, however, and do not discretely or categorically classify subjects of erg-abs verbs in contrast to abs-i verbs. For example, the aforementioned properties which are characteristic of ergative subjects -agentivity, effort, and volitionality -are also most typically associated with animate subjects. erg-abs verbs in Samoan, however, are compatible with inanimate (ergative) subjects, such as natural forces (38a)  We cannot therefore exhaustively classify ergative-marked subjects as thematic "agents", in any kind of narrow sense, since agents are typically though-of as volitional entities, as opposed to inanimate causers. Equally, we cannot classify ergative subjects as necessarily triggering an affect upon an object, since, while this is true of the bolded subjects in (38), it is not always true of subjects of verbs such as mata'i ('observe'), tuli ('hunt'), and su'e ('search for').
The relative fuzziness of the distinction between subjects of erg-abs and of abs-i verbs can be captured by considering a broader view of theta-roles. I adopt Dowty's (1991) treatment of thematic roles as cluster concepts -which he terms proto-roles -as opposed to discrete categories (e.g., agent vs. causer vs. experiencer). Dowty proposes that two fundamental proto-roles are required in theta-theory: proto-agent (corresponding to the VP-external argument), and proto-patient (corresponding to the VP-internal argument). These proto-roles do not classify arguments exhaustively or definitively, but consist instead of sets of properties, or entailments, given to an argument by its selecting predicate. A proto-agent, for example, tends to be a sentient causer of an event, and to undergo movement relative to another participant in that event, while a proto-patient typically undergoes a change of state, is affected by another participant, and does not undergo movement. Dowty stresses, however, that arguments may possess properties of both protoroles, such that whether an argument is ultimately assigned a proto-agent or proto-patient role is subject to a certain degree of idiosyncrasy and unpredictability. As a case in point, Dowty discusses cross-linguistic variation in the divide between unergative and unaccusative verbs (following Rosen 1981;1984). He contends that "[…] no single sematic criterion -volition, agentivity, presentational meaning -or combination of criteria seems to determine this correctly for all verbs" (Dowty 1991: 606). The intransitive verb sweat, for example, is unergative in Italian, and unaccusative in Choctaw, while die is unergative in Choctaw, and unaccusative in Italian. This type of variation can be captured under the non-discrete approach to theta-roles advocated in Dowty's theory of protoroles.
Adapting this approach to the current discussion, I propose that Samoan (and other languages, potentially, although this is to be demonstrated on a case-by-case basis), distinguishes two types of proto-agent role, which I call proto-high agent and proto-low agent. Proto-high agents -the subjects of erg-abs verbs -have properties as listed in (39a), while Proto-low agents -subjects of abs-i verbs -have properties as listed in (39b). The first two properties -(i) initiating and (ii) experiencing -are common to both protoroles; however, the high agent role encompasses five additional properties: (iii) effect, (iv) causing a change of state, (v) effort, and (vi) volition, and (vii) conclusion of an event. Although low agents are typically non-affecting and do not cause a change of state, they themselves are also not typically affected, nor do they undergo a change in state: these properties are associated with proto-patients (see Dowty 1991).
(39) (a) Proto-high agent (i) initiates an event (ii) experiences an event (iii) triggers an effect upon another entity (iv) brings about a change in state initiates an event (ii) experiences an event (iii) neither affects another entity, nor is physically affected (iv) neither brings about nor undergoes a change in state A crucial component of Dowty's analysis, however, is that a participant does not have to demonstrate every property which characterises their assigned proto-role. As is evident from (38), Samoan proto-high agents may lack properties such as volition -which are generally considered fundamental properties of agenthood -but still demonstrate enough other relevant properties to be mapped to the high agent role. This is evident if we consider once again the Samoan middle verb mulimuli ('follow') and unergative siva ('dance') (cf. 36). The subjects of these predicates encompass all properties of the proto-low agent role in (39), and by extension, the first two properties of proto-high agent role. However, both subjects are also volitional (a-v) and (at least to a certain extent) effortful (a-vi): properties which are associated with the high agent role. For subjects of mulimuli and siva, it is therefore their behaviour with respect properties (iii), (iv), and (vii) -namely they do not trigger an effect or change of state, or conclude an event -which ultimately determines mapping to the proto-low agent role.
However, (iii), (iv), and (vii) do not categorically determine proto-roles for all predicates. The verb 'tuli ('hunt'), for example, requires an ergative (i.e., proto-high agent) subject. As discussed, however, its object need not be affected or undergo any change in state as a result of the hunting event, and the event is not concluded. In this case, properties (iii), (iv), and (vii) would predict that the subject be mapped to the proto-low agent role. Nonetheless, the properties of effort and volition (a-v and a-vi) are more influential in this instance, with the result that the proto-high agent role is assigned. As a further case in point, certain transitive verbs, such as va'ai ('see') can appear either with an erg-abs or a middle (abs-i) case frame. However, there is commonly a difference between the two meanings, such that the subject is interpreted as more effortful when marked ergative (40a) than when absolutive (40b).

(40)
Ergative-middle alternation The contrast between this pair of sentences is not so much one of object affectedness but rather foresight and intent on the part of the subject. As with 'tuli, it is properties of effort and volition (v and vi) which determine the mapping to (ergative) proto-high agent in (40a), and the lack thereof, to (absolutive) proto-low agent in (40b). Mosel and Hovdhaugen (1992: 424) indeed note that "if the […] erg-abs pattern is found with cognitive verbs, the clause mainly expresses the fact that the act of cognition involves the intention or at least a high degree of consciousness of the participant referred to by the ergative noun phrase".
Similarly, a number of abs-i patterning verbs have a derived erg-abs counterpart, formed by addition of a suffix (to be discussed in greater detail in §5.3). Cook (1996: 69) characterizes the derived verb as having "a related and (for some forms) more intense meaning". In (41a), erg-abs patterning alofagia ('care for') is derived from the simple abs-i patterning verb alofa ('love'; 41b). As with va'ai in (40), the ergative subject in (41a) encompasses properties (a-v) and (a-vi). The event of caring for someone denotes an action -thereby involving effort and volition on the part of the subject -while (41b) denotes a state which entails neither (a-v) nor (a-vi). The properties of effort and volition thereby seemingly determine mapping of the external argument to the proto-high role in (41a) and to the proto-low role in (41b).
For another pair of verbs, mana'o ('want'; 42b) and its derived erg-abs counterpart mana'omia ('need'; 42a), however, the contrast between proto-high and -low roles can be characterized in terms of property (ii): while both the ergative subject in (42a) and the absolutive in (42b) are experiencers, the event of needing in (42a) invokes greater intensity than wanting in (42b).  Cook (1996: 69) reports that his consultants describe mana'omia as being 'stronger' than mana'o (hence his translation of the former as need as opposed to want). It is this greater intensity with regards (ii) which seemingly influences the mapping of the external argument to the proto-high agent role in (42a), in contrast with the proto-low role in (42b).
Finally, properties (i) and (vii) concern telicity. In their typologically-based Transitivity Hypothesis, Hopper and Thompson (1980) observe that actions with discernible endpoints (i.e., telic events) are more likely to be grammatically encoded as transitive than those with no discernible endpoint (i.e., atelic events). Samoan ergative-absolutive and middle verbs typically straddle this divide: verbs associated with ergative case tend to denote actions with endpoints, whereas actions described by middle-and unergative-patterning verbs tend to lack endpoints. The verbs listed in (19), along with the verbs discussed in this section are grouped according to this parameter in (43)  Although telicity and case frame do not categorically co-vary, all telic verbs listed in (43) are ergative-absolutive patterning, while the majority of atelic verbs pattern as absolutivei. Adapting this to the present framework of proto-roles, both proto-low and -high agents initiate an event (property i). Only proto-high agents, however, also bring the event to its conclusion (property a-vii), giving rise to telicity. This distinction will be discussed further in section 5.4. To conclude this section, I have proposed that the split between erg-abs and abs-i predicates is best characterized by differences in properties of the subjects of those predicates, and propose that they receive different thematic (proto-)roles. Two externallyassigned proto-role types can be distinguished: highly effortful, and/or volitional, and/or affecting proto-high agents, and non-affecting, non-effortful proto-low agents. The proto-low agent role includes experiencers, and agents which initiate an event but do not necessarily cause any effect, or bring about any change in state. It can be concluded from examining a variety of verbs that the mapping of an external argument to either a proto-high or low agent role is not determined according to any narrow categorical criteria. Nonetheless, the characteristics in (39) are intended to serve as a principled means of predicting the classification of a thematic subject, and concomitantly, its case marking.

Case assignment in Samoan: v and Voice
This section presents the core analysis of the Samoan case split, considering first the source of i-case, and second, how the two different transitive case frames are derived syntactically. I then discuss additional evidence for VoiceP in the Samoan verbal projection, and finally consider how Pseudo Noun Incorporation may be accounted for within this picture.

What is i-case?
Having established that middle and unergative i-marked objects are not prepositional objects, an important clue as to the identity of the i-marker lies in its role elsewhere in the Polynesian family. While i indeed functions as an oblique or prepositional marker in Samoan (and more widely throughout Polynesian), it also serves as a structural accusative case marker in those related Polynesian languages with accusative alignment; e.g., Maori (44a), and Hawaiian (44b). Note from (44b) in particular that Hawaiian i case serves both as an accusative marker (marking the direct object) and oblique marker (marking the indirect object). Indeed, Chung (1978: 49-50) remarks that "the case marking of canonical transitive clauses in accusative languages resembles that of middle clauses in all Polynesian languages". A treatment of Samoan i-case as structural accusative is consistent with its appearance on raised (i.e., non-thematic) arguments. Certain middle predicates such as mana'o ('want') and fia maga'o ('hope') take clausal complements, and allow the embedded subject to While sentences such as those in (45) and (46) are often referred to as "raising-to-object" constructions in Polynesian literature, Massam (1985) argues for related Niuean that the relevant movement does not, in fact, target the object position of the matrix clause. 18 Rather, the raised DP remains in the embedded CP, occupying a CP-peripheral specifier position, labelled as "SPEC2" in (47). Here, it is assigned structural case, from a relevant case-assigning head in the matrix clause, via ECM.
Raising-to-object in Polynesian (Massam 1985) 'Pete hopes that Sarah kicks the ball.' While sentences such as those in (45) and (46) are often referred to as "raising-to-object" constructions in Polynesian literature, Massam (1985) argues for related Niuean that the relevant movement does not, in fact, target the object position of the matrix clause 18 . Rather, the raised DP remains in the embedded CP, occupying a CP-peripheral specifier position, labelled as "SPEC2" in (47). Here, it is assigned structural case, from a relevant case-assigning head in the matrix clause, via ECM.

SPEC
C' Bejar and Massam (1999) propose a theory of Multiple Case Checking in order to account for several cross-linguistic constructions (including Polynesian raising-to-object) in which a DP appears to receive more than once case. Their claim is that, in Multiple Case constructions, the case of a DP is interpretable only if the DP remains in the appropriate checking configuration with its case-assigning head: once the DP moves, case must be reassigned. Applying this to the Samoan raising in (45b) and (46b), the raised DP (le teine in (45) and Sarah in (46)) first receives case in the lower clause. After raising into the specifier of CP, however, the DP is no longer in its original case-checking configuration, and case must therefore be re-checked. Since both (45) and (46) (46)), i-case is available to be assigned to the raised DP.
For the present purposes, the relevant point is that Samoan i-case may be assigned to derived objects, which are not thematically selected by the middle verb; this suggests that a treatment of i-case as structural case (as opposed to lexical or inherent case) is most appropriate. I therefore propose that i-case used to mark the Samoan unergative and middle object is structural accusative case (glossed henceforth as ACC); Samoan hereby has two transitive case case 17 As pointed out by a reviewer, Davies (2005) argues that a related type of construction in Madurese (Indonesian) does not involve raising but is instead an instance of NP prolepsis, in which the purportedly raised argument is in fact merged in the higher clause as a VP adjunct, which is coreferential with a null pronoun in the lower clause. In Polynesian, constructions such as those of (42-3) are typically analysed as involving raising (e.g., Seiter 1980;Massam 1985;et seq.). Indication that a proleptic analysis is not appropriate for Samoan is the impossibility of a pronoun at the gap site, which is licit both in Madurese prolepsis (see Davies 2005: 650), and in English (e.g., 'I believe about Alice that she likes horses'). In Samoan, however, the constriction is judged as completely unacceptable if a pronoun is inserted in the lower clause.
(i) E mana'o Alice i le teine e nofo (*ia) i le fale. prs want Alice.abs i det girl prs stay 3sg obl det house 'Alice wants the girl to stay in the house.' 18 Chen and Fukuda (2015) propose a similar analysis for the Formosan languages Puyuma, Amis, and Seediq, but argue that the "raised" DP in these languages is base-generated in SPEC2, as opposed to moving there. Bejar and Massam (1999) propose a theory of Multiple Case Checking in order to account for several cross-linguistic constructions (including Polynesian raising-to-object) in which a DP appears to receive more than once case. Their claim is that, in Multiple Case constructions, the case of a DP is interpretable only if the DP remains in the appropriate checking configuration with its case-assigning head: once the DP moves, case must be reassigned. Applying this to the Samoan raising in (45b) and (46b), the raised DP (le teine in (45) and Sarah in (46)) first receives case in the lower clause. After raising into the specifier of CP, however, the DP is no longer in its original case-checking configuration, and case must therefore be re-checked. Since both (45) and (46) (46)), i-case is available to be assigned to the raised DP. For the present purposes, the relevant point is that Samoan i-case may be assigned to derived objects, which are not thematically selected by the middle verb; this suggests that a treatment of i-case as structural case (as opposed to lexical or inherent case) is most appropriate. I therefore propose that i-case used to mark the Samoan unergative and middle object is structural accusative case (glossed henceforth as acc); Samoan hereby has two transitive case frames: ergative-absolutive, and absolutive-accusative. This type of analysis is not without precedent: Aldridge (2011) argues that Malagasy (Austronesian) also has two transitive frames, which reflect its intermediate stage on a continuum of reanalysis from a fully ergative -to a fully accusative -aligned language. 19 I turn now to the question of how case is assigned in Samoan, and what governs the split in transitive case patterns.

Case assignment in Samoan
I begin this section with a brief recap of Samoan case markers: in section 2.2, I addressed the nature of ergative and absolutive case. I treat absolutive as a structural case, assigned by T 0 to the closest DP with unvalued case, just as is nominative case (see Koopman 2012). Ergative case, conversely, is assigned in a specifier -head configuration ("Spec-Head ergative"); I will propose in this section that the relevant head here is Voice 0 . Middle/unergative i-case was identified in section 5.1 as structural accusative, and is assigned by v 0 , as will be discussed shortly. A summary is given in Table 1.

Accusative case
At least since Burzio (1986), it has been claimed that assignment of accusative case to an object is dependent upon assignment of a theta role by a predicate to its sub- ject. Conversely, accusative case is unavailable in a sentence which lacks a thematic subject (see Marantz 1991: 14). 20 Since the implementation of the Split-VP hypothesis (Chomsky 1995;Kratzer 1996;Marantz 1997; i.a.), the task of introducing a VP-external thematic subject, and assigning accusative case to a VP-internal object, has been afforded to v 0 , as in (48a). If no external argument is introduced (i.e., in an unaccusative or passive sentence), then the internal argument receives structural case from T 0 (48c); in nominative-accusative languages, this is referred to as "nominative" case. External arguments, whether transitive subjects (48a) or unergative subjects (48b), also receive this same structural case from T 0 -the only difference being that accusative case is not assigned in unergative constructions if no object is present. 21 The model of case assignment in (48) is, I propose, precisely how absolutive and accusative are assigned in Samoan -besides the labelling of nominative as "absolutive". Case licensing in the middle and unergative + object construction proceeds as in (48a); unergative and unaccusative subjects are licensed as shown in (48b, c). The structure of the Samoan middle and unergative + object construction is shown in (49). Recall also from section 2.1 that DPs in Samoan, as in other related Polynesian languages, move to the inner specifier of vP following case assignment, 22 before the VP undergoes remnant movement to a functional projection below TP (Massam 2001b 20 There has been debate, however, regarding the universality of Burzio's Generalization. Goodall (1993), for example, draws attention to several languages in which accusative case appears on the single argument of a passive, despite the purported absence of an external subject. 21 The issue of having an unassigned accusative case feature in unergatives receives no consensus in syntactic theory. I view accusative as a conditional case feature (in the sense of Collins 2016): in other words, it must be assigned if a DP requiring case is present in its c-command domain, but not otherwise (see also Preminger 2011, who claims that failure of a case licensing head to assign a case feature does not result in a crashed derivation). 22 In Samoan, we only have indirect evidence of object movement: namely, the fact that the verb and the object are not adjacent (due to the absence of PNI; see §3.3). Nonetheless, movement of accusative-marked objects is also evident in languages in which VP-fronting does not take place (such that movement is more transparent); for example, Gribanova (2016: 6) shows that bare objects in Uzbek (Turkic; nominativeaccusative; SOV) must remain directly adjacent to the verb, while accusative-marked objects may raise to a higher position (e.g., to the left of an intervening indirect object).

Two case frames
As for the second, more familiar Samoan transitive case frameergative-absolutivetwo questions arise: first, how does it differ structurally from the absolutive-accusative frame? And second, why does it differ? What is the function of a second transitive case pattern? Here, I expand on a proposal by Massam (2009; to appear) for the related language Niuean (see also Polinsky 2016; Tollan 2016). Massam claims that unergative subjects are structurally lower than ergatives. Treating both unergative and middle subjects as PROTO-LOW AGENTS, to the exclusion of ergative subjects (PROTO-HIGH AGENTS), I propose that proto-low agents are introduced (and assigned their proto-role) by v 0 . Ergative subjectsproto-high agentsare introduced by a higher thematic head, namely Voice 0 . The additional phrasal structure of ERG-ABS constructions corresponds to the additional properties encompassed by ergative proto-high agents in (39a) which distinguish them from proto-low agents in (39b), namely iii through vii (either all of these properties, or a subset of them). Recall that properties i and ii characterize both high and low agents. As such, i and ii are contributed by v 0 , while iii through vii are contributed by Voice 0 . This is illustrated in (50) and in Table 2.

Two case frames
As for the second, more familiar Samoan transitive case frame -ergative-absolutive -two questions arise: first, how does it differ structurally from the absolutive-accusative frame? And second, why does it differ? What is the function of a second transitive case pattern? Here, I expand on a proposal by Massam (2009;to appear) for the related language Niuean (see also Polinsky 2016; Tollan 2016). Massam claims that unergative subjects are structurally lower than ergatives. Treating both unergative and middle subjects as proto-low agents, to the exclusion of ergative subjects (proto-high agents), I propose that proto-low agents are introduced (and assigned their proto-role) by v 0 . Ergative subjects -proto-high agents -are introduced by a higher thematic head, namely Voice 0 . The additional phrasal structure of erg-abs constructions corresponds to the additional properties encompassed by ergative proto-high agents in (39a) which distinguish them from proto-low agents in (39b), namely iii through vii (either all of these properties, or a subset of them). Recall that properties i and ii characterize both high and low agents. As such, i and ii are contributed by v 0 , while iii through vii are contributed by Voice 0 . This is illustrated in (50) and in Table 2.

Two case frames
As for the second, more familiar Samoan transitive case frameergative-absolutivetwo questions arise: first, how does it differ structurally from the absolutive-accusative frame? And second, why does it differ? What is the function of a second transitive case pattern? Here, I expand on a proposal by Massam (2009;to appear) for the related language Niuean (see also Polinsky 2016;Tollan 2016). Massam claims that unergative subjects are structurally lower than ergatives. Treating both unergative and middle subjects as PROTO-LOW AGENTS, to the exclusion of ergative subjects (PROTO-HIGH AGENTS), I propose that proto-low agents are introduced (and assigned their proto-role) by v 0 . Ergative subjectsproto-high agentsare introduced by a higher thematic head, namely Voice 0 . The additional phrasal structure of ERG-ABS constructions corresponds to the additional properties encompassed by ergative proto-high agents in (39a) which distinguish them from proto-low agents in (39b), namely iii through vii (either all of these properties, or a subset of them). Recall that properties i and ii characterize both high and low agents. As such, i and ii are contributed by v 0 , while iii through vii are contributed by Voice 0 . This is illustrated in (50) and in Table 2.

(50) External argument positions in Samoan
VoiceP The partition of vP and VoiceP is well-established in the literature: a split v/Voice structure has been adopted by Alexiadou et al. (2006), Pylkkännen (2008), Coon (2013), Harley (2013), and Legate (2014), besides many others. Both Massam (2009;to appear) and the present analysis differ, however, by arguing that both split heads are responsible for introducing (different types of) external arguments. I further demonstrate how the difference between arguments introduced by v 0 and those introduced by Voice 0 may result in split case patterning.

Ergative case and the split v/Voice structure
Case assignment in Samoan high transitives is shown -with relevant movement -in (51). Spec-Head ergative case is assigned by Voice to the external argument in its specifier (i.e., a proto-high agent). The internal argument raises out of VP into the lower specifier of vP (note crucially however that it is not merged in this position), and receives absolutive case from T 0 . The transitive subject does not act as an intervener between T 0 and the internal argument in VP, since it has already been licensed (as ergative) by Voice 0 ; I assume, following Legate (2002;, Aldridge (2004)

28
The partition of vP and VoiceP is well-established in the literature: a split v/Voice structure has been adopted by Alexiadou et al. (2006), Pylkkännen (2008), Coon (2013), Harley (2013), and Legate (2014), besides many others. Both Massam (2009;to appear) and the present analysis differ, however, by arguing that both split heads are responsible for introducing (different types of) external arguments. I further demonstrate how the difference between arguments introduced by v 0 and those introduced by Voice 0 may result in split case patterning.

Ergative case and the split v/Voice structure
Case assignment in Samoan high transitives is shownwith relevant movementin (51). Spec-Head ergative case is assigned by Voice to the external argument in its specifier (i.e., a proto-high agent). The internal argument raises out of VP into the lower specifier of vP (note crucially however that it is not merged in this position), and receives absolutive case from T 0 . The transitive subject does not act as an intervener between T 0 and the internal argument in VP, since it has already been licensed (as ergative) by Voice 0 ; I assume, following Legate (2002;, Aldridge (2004), and Coon et al. (2014), among others, that its case features are now defective 23 .  Crucially, because high external arguments are not merged in vP, v 0 has no accusative case feature, and the object receives absolutive case from T 0 . 24 In other words, accusative case is available for the internal argument only if v 0 introduces a (low) external argument. Voice 0 assigns Spec-Head ergative to a high external argument, and T 0 checks structural absolutive case on any remaining argument. It follows from this that Burzio's (1986) Generalization -namely, that structural accusative case is available only if an external argument is merged -applies extremely locally. That is, accusative case is available on a VP-external thematic head (e.g., v in the current analysis) only if the corresponding external argument is introduced by the same head. The merge of a VP-external argument higher in the thematic domain (i.e., VoiceP) is not sufficient to endow a case feature on a lower thematic head. Note that, with the adoption of proto-roles (Dowty 1991), ergative case can be regarded as inherent ergative (see § 2.2), whereby it is semantically associated properties (a-iii) though (a-vii) in (51), and therefore assigned inherently to proto-high agents. Since nothing in the present discussion hinges upon the choice of terminology, I continue to refer to "Spec-Head" ergative for consistency.

Voice morphology
Some support for the claim that Samoan erg-abs (high) transitives involve more structure than abs-acc (low) transitives comes from a small number of middle verbs, including fesili ('ask'; 52a), and vala'au ('call'; 53a). These verbs may also pattern as erg-abs (52b, 53b), but require the addition of the suffix -C(i)a (in which the initial consonant and high vowel are determined based on phonotactics) to do so. Under the analysis in section 5.2, -C(i)a can be treated as a spell-out of Voice 0 . This provides some evidence that erg-abs predicates involve more structure than abs-acc predi- 24 The derivation in (51) leaves open the question of whether VoiceP constitutes a phase boundary in Samoan.
If we are to assume that VoiceP is universally phasal, then the absolutive object in (51) must move to the phase edge in order to be visible to T 0 . One way of achieving this is to assume that such movement takes place covertly in VSO constructions such as (51), and conversely results in VOS word order when overt, as proposed by Tollan and Clemens (2016) for Tongan.
cates (as opposed to being derived from a different flavor of vP with a different featural composition, in the sense of Folli andHarley 2004, andAldridge 2012). It should be noted, however, that this type of derivation is not fully productive, applying only to a minority of middle verbs. 25 Furthermore, two experiencer verbs in particular -iloa ('know') and lagona ('feel') -which might be expected to pattern as low transitive, appear obligatorily with an erg-abs case array. As pointed out by Koopman (2012), however, both of these predicates are "morphologically complex": both comprise the -C(i)a Voice suffix (see also Milner 1966 for further discussion). It is also worth noting that -C(i)a functions as a passive suffix in nominative-accusative Polynesian languages (e.g., Maori), which is further indication of its status within the family as a Voice marker. I turn now to the distribution of the causative prefix fa'a, which provides further indication that erg-abs transitives involve additional structure. Fa'a can be added to an unaccusative verb to form a causative counterpart, with erg-abs case patterning, as in (54) It can also be added to unergative verbs, with the meaning of "make X do something"; (55). The causer is marked ergative and the causee absolutive. 25 A further issue to be addressed is how to ensure that -Cia is morphologically attached to the main verb before VP movement takes place. This could be achieved via a head lowering operation, as proposed in Collins (2016: 15, fn. 13), in which Voice 0 and v 0 are morphologically concatenated with V 0 prior to VP raising. An alternative approach would involve raising of the -Cia suffix at LF in accordance with the Stray Affix Filer (Lasnik 1995), in much the same way that English tense markers are proposed to attach to the main verb. In English, this is achieved by affix lowering; in Samoan, however, the main VP raises, meaning that a stray verbal affix must also undergo raising. Thanks to a reviewer for bringing this issue to my attention. 26 I can find two exceptions: the middle verbs tago ('take') and mulimuli ('follow') do not allow the fa'a prefix. Notice that fa'a causatives formed from middle and unergative verbs demonstrate a key prediction of the structure proposed in (50), namely that both VP-external argument positions can be occupied simultaneously: (55b) and (56b) both exhibit an ergative causer (taken to occupy spec, VoiceP) and an absolutive causee (taken to occupy spec, vP). Furthermore, the presence of an accusative object in (56b) demonstrates that ergative-and accusative-marked arguments indeed do co-occur. Here, the high-agent causer (le tama 'the boy') receives Spec-Head ergative case from Voice 0 , while the low-agent causee (le teine 'the girl') and object (le masi 'the cookie') are both assigned structural case: the causee receives absolutive case from T 0 , and the object receives accusative case from v (which is able to assign accusative, by having introduced the external causee). The structure of (56b) is shown below. 27 (57 32 can be occupied simultaneously: (55b) and (56b) both exhibit an ergative causer (taken to occupy spec, VoiceP) and an absolutive causee (taken to occupy spec, vP). Furthermore, the presence of an accusative object in (56b) demonstrates that ergative-and accusative-marked arguments indeed do co-occur. Here, the high-agent causer (le tama "the boy") receives Spec-Head ergative case from Voice 0 , while the low-agent causee (le teine "the girl") and object (le masi "the cookie") are both assigned structural case: the causee receives absolutive case from T 0 , and the object receives accusative case from v (which is able to assign accusative, by having introduced the external causee). The structure of (56b) is shown below 27 .

DP [+ERG]
Voice" e le tama Voice vP <fa'a-> CAUS DP [+ABS] vP le teine a/the girl DP i [+ACC] Comrie (1975) characterizes causative marking as increasing the valency of a predicate by one argument. In the examples above, I suggest that fa'a spells out Voice 0 , and introduces an externally-merged high agent. If ERG-ABS pattering predicates already bundle VoiceP, then one crucial prediction is that they should be unable to be causativized by fa'a, because there should be no place in the structure for it (assuming, following Grohmann 2003 The intended meanings in (58b, 69b) must be expressed instead with a biclausal construction, which is translated loosely as 'X compelled Y to do Z'. This supports an analysis in which erg-abs transitives involve more phrasal (here: thematic) structure than abs-acc transitives. 29

Pseudo-Noun Incorporation and unergatives
As discussed earlier, Samoan exhibits VOS word order with bare objects, as in (61b). This is known as Pseudo-Noun Incorporation (Massam 2001b). The subject of a PNI construction is absolutive; the object appears with no determiner or case marking, 30 and is interpreted as indefinite and non-specific (Chung 1978). According to Massam's analysis, PNI objects are NPs (or at least, do not contain D structure), and therefore do not require 28 While I treat fa'a as an instantiation of Voice 0 in the examples presented here, it is likely not exclusive to Voice 0 . Read (2010) and Collins (2016) present examples in which fa'a morphology appears in the absence of an ergative subject, which indicate that fa'a may also spell out v 0 . Fa'a can also be added to nouns and adjectives to form intransitive verbs (see MH: 176-8). I do not attempt a full analysis of fa'a here; the key observation is the potential for unergative and middle verbs, but not erg-abs verbs, to be causativized by fa'a. 29 I cannot find any examples in which fa'a co-occurs with the -(C)ia suffix. 30 Yu (2011) observes that an absolutive high tone is never found on the word preceding an incorporated object, as would be expected if it had absolutive case. 33 known as Pseudo-Noun Incorporation (Massam 2001b). The subject of a PNI construction is absolutive; the object appears with no determiner or case marking 30 , and is interpreted as indefinite and non-specific (Chung 1978). According to Massam"s analysis, PNI objects are NPs (or at least, do not contain D structure), and therefore do not require case-licensing. They fail to raise out of VP prior to VP fronting, thereby appearing adjacent to the verb, as illustrated below. The VOS word order of PNI hereby results from the absence of object movement.  Yu (2011) observes that an absolutive high tone is never found on the word preceding an incorporated object, as would be expected if it had absolutive case.

ABS
While there is no consensus on the motivation for object movement, I follow Aldridge (2004)'s proposal for Tagalog, in which she claims that DP objects must raise in order to obtain a referential interpretation. Raising out of VP -to vP -allows for the object to escape existential closure, which would ensue if it were to remain VP-internal (Diesing 1992). In the context of the current paper, an important observation to be noted is that PNI (i.e., non-referential) objects never appear with proto-high agents (i.e., ergative-marked subjects). 31 This raises questions regarding the precise nature of the relationship between the properties of an agent subject, and the properties of a theme object. In this section, I discuss how this may be dealt with within the current approach; first, why the subject of a PNI predicate is always an (absolutive) proto-low agent as opposed to a proto-high agent, and second, how the relationship between agent and theme is instantiated in the syntax.

PNI and proto-roles
Of the properties of proto-high agents listed in (39), five distinguish them from low agents: effect upon another entity (a-iii), cause of a change in state (a-iv), effort (a-v), volition (a-vi), and conclusion of an event (a-vii). Under the current proto-role approach, PNI subjects are mapped as low agents; thus, they must lack some (or all) of these properties, by contrast with non-PNI subjects. I suggest that the distinction lies with properties (a-iii), (a-iv), and (a-vii): effect, change of state, and event conclusion. While these are situated 31 Woolford (2015) discusses several other languages (e.g., Niuean, Tagalog, Dyirbal) which pattern similarly, whereby the subject is marked ergative only if the object has raised out of VP. In other languages, however, ergative marking is retained even when the object has undergone incorporation (e.g., K'ichee'; see Aissen 2011). as properties of a proto-agent, they are also necessarily influenced by the nature of the theme, as opposed to being determined by the agent alone. Regarding (iii) and (iv), it is not immediately obvious how the agent of a PNI construction, such as le fafine in (61b), should necessarily be less affecting, or cease to cause a change in state, as compared with its non-PNI counterpart (cf. 61a). Rather, the distinction between affectedness in PNI and non-PNI constructions concerns the nature of the relationship between the affecting agent and the affected theme. Hopper and Thompson (1980)'s typologically-motivated Transitivity Hypothesis posits object individuation as a key parameter in influencing clause transitivity; in their terms, individuation refers "both to the distinctness of the theme from the agent, and its distinctness from its own background" (Hopper & Thompson 1980: 253). Specifically, a quantized, referential entity (in the present context, a non-incorporated, full DP object) is individuated, while a nonquantized, non-referential entity (i.e., a PNI object) is not. Thus, in terms of properties (iii) and (iv), an agent cannot trigger an affect or change in state, if the relevant theme is not sufficiently distinct from the agent and/or from the background of the event, such that it has the potential to be affected or undergo a change. Regarding (a-vii), events denoted by PNI predicates lack discernible endpoints; thus, the low agent of a PNI construction initiates the relevant event, but does not bring that event to its conclusion. However, event telicity has long been recognised to depend on the properties of the theme (e.g., Krifka 1989;Dowty 1991;Ramchand 1997;a.o.), such that an event can be concluded only if the theme is quantized. Thus, PNI predicates are necessarily atelic because they do not comprise a quantized theme.
By this reasoning, it the absence of properties (iii), (iv), and (vii) which fundamentally determine the mapping of a PNI subject to the proto-low agent role. On the one hand, no agent can hold these properties unless the theme argument also meets the relevant criteria: it must be a quantized, referential entity. On the other hand, properties (iii), (iv), (vii) are not wholly determined by the theme: the presence of a referential, quantized entity does not entail that the agent will bear these properties, and consequently be mapped to the high-agent role. As evidenced by many examples in earlier sections, proto-low agents do not only appear with predicates in which the object lacks such properties (see, among other examples, the pairs in (40,41), in which the object is the same in both (a) and (b) sentences). I turn now to the issue of how this agent-theme relationship is instantiated in the syntax.

Object raising
To recap from earlier discussion: when a theme object is referential, it is realised as a full DP, and must raise out of the VP. In this case, the subject may be either a proto-low or proto-high agent, depending on a number of parameters. Conversely, pseudo-incorporated objects are semantically non-referential, and non-specific. PNI objects are realized as NPs (i.e., lacking DP structure), and do not raise out of VP (Massam 2001b). Subjects of PNI constructions are always mapped as proto-low agents (merged in vP) bearing absolutive case. Here I discuss the structural instantiation of this.
In the thematic domain, the merge of VoiceP -the thematic projection which introduces a proto-high agent, bearing ergative case -appears to be contingent upon the syntax of the lower argument. Object movement -to the specifier of vP -is required if VoiceP is to be merged. Conversely, object movement does not require the merge of Voice 0 : the object in low transitive (middle and unergative + object) constructions is assumed to have raised out of VP (cf. 49), yet a non-ergative case frame arises, indicating that the agent is mapped to the proto-low role.
To account for this, I draw upon a proposal by Hale and Keyser (2002), aimed at explaining causative-inchoative alternations in terms of asymmetries in lexical structure. Specifically, Hale and Keyser note that certain verbs, such as sink (62a) can undergo transitivity alternations, while others, such as sparkle (62b), cannot.
(62) a. The boat sank. The storm sank the boat.
b. The diamond sparkled. *The sun sparkled the diamond.
Hale and Keyser propose that the asymmetry between (a) and (b) is due to a difference in the structure forming possibilities of the lexical roots sink and sparkle. Verbs such as sink and sparkle minimally consist of a verbal host and a lexical root (which later incorporates with the host via a process known as "conflation"). The difference between sink and sparkle lies in whether the root requires the projection of a specifier: sink does (63a), whereas sparkle does not (63b). Crucially, only if the verbal projection contains a specifier can further verbal structure be merged (labelled here as V 2 ; this is potentially a causative light v head). As such, the V 1 maximal projection in (b) must combine directly with its single DP argument (e.g., the diamond) -no further verbal structure building is possible due to the absence of a specifier in the lower projection (in other words, because the already-present verbal structure is not maximally saturated).
host via a process known as "conflation"). The difference between sink and sparkle lies in whether the root requires the projection of a specifier: sink does (63a), whereas sparkle does not (63b).
Crucially, only if the verbal projection contains a specifier can further verbal structure be merged (labelled here as V2; this is potentially a causative light v head). As such, the V1 maximal projection in (b) must combine directly with its single DP argument (e.g., the diamond)no further verbal structure building is possible due to the absence of a specifier in the lower projection (in other words, because the already-present verbal structure is not maximally saturated). Expanding this principle from the lexical domain to the VP-external event domain 32 , we can explain from a structure building perspective why the high agent-introducing thematic headnamely VoicePis not merged in PNI constructions: vP contains no specifier. The unattested derivation is one in which VoiceP is merged, but the specifier of vP is empty, as in (64).
Only if the specifier of vP is filledeither by a raised internal argument (as in any high transitive) or by a merged low external argument (e.g., as in a causativized unergative such as (55b))can Voice 0 be merged. This is characterized below in (65).
(65) VoiceP may be merged if vP contains a specifier, filled either by a moved internal argument and/or by a merged (low) external argument.
Note, however, that (65) is not a bi-conditional: the presence of a specifier in vP is a prerequisite for the merge of Voice 0 , but not a sufficient factor. VoiceP is not present in low transitives, b.
host via a process known as "conflation"). The difference between sink and sparkle lies in whether the root requires the projection of a specifier: sink does (63a), whereas sparkle does not (63b).
Crucially, only if the verbal projection contains a specifier can further verbal structure be merged (labelled here as V2; this is potentially a causative light v head). As such, the V1 maximal projection in (b) must combine directly with its single DP argument (e.g., the diamond)no further verbal structure building is possible due to the absence of a specifier in the lower projection (in other words, because the already-present verbal structure is not maximally saturated). Hale & Keyser 2002: 2) Expanding this principle from the lexical domain to the VP-external event domain 32 , we can explain from a structure building perspective why the high agent-introducing thematic headnamely VoicePis not merged in PNI constructions: vP contains no specifier. The unattested derivation is one in which VoiceP is merged, but the specifier of vP is empty, as in (64).
Only if the specifier of vP is filledeither by a raised internal argument (as in any high transitive) or by a merged low external argument (e.g., as in a causativized unergative such as (55b))can Voice 0 be merged. This is characterized below in (65).
(65) VoiceP may be merged if vP contains a specifier, filled either by a moved internal argument and/or by a merged (low) external argument.
Note, however, that (65) is not a bi-conditional: the presence of a specifier in vP is a prerequisite for the merge of Voice 0 , but not a sufficient factor. VoiceP is not present in low transitives, Expanding this principle from the lexical domain to the VP-external event domain, 32 we can explain from a structure building perspective why the high agent-introducing thematic head -namely VoiceP -is not merged in PNI constructions: vP contains no specifier. The unattested derivation is one in which VoiceP is merged, but the specifier of vP is empty, as in (64).
Only if the specifier of vP is filled -either by a raised internal argument (as in any high transitive) or by a merged low external argument (e.g., as in a causativized unergative such as (55b)) -can Voice 0 be merged. This is characterized below in (65).
(65) VoiceP may be merged if vP contains a specifier, filled either by a moved internal argument and/or by a merged (low) external argument.
Note, however, that (65) is not a bi-conditional: the presence of a specifier in vP is a prerequisite for the merge of Voice 0 , but not a sufficient factor. VoiceP is not present in low transitives, despite movement of the internal argument (i.e., an unergative or middle object) to spec vP. 33 Essentially, Voice 0 is only merged if the relevant predicate also requires a proto-high agent. The merge of an agent in VoiceP, is hereby ultimately determined by the properties of the agent, but is concurrently dependent on the properties of the lower argument.

Summary
The various clause types and patterns of case assignment are summarized Table 3.
In the final discussion section, I consider the cross-linguistic implications of split v/Voice system I have proposed for Samoan.

Cross-linguistic variation
The main claim of this paper is that Samoan exhibits two external argument positions, which results in two transitive case frames. Crucially, unergative subjects are merged lower than ergative subjects (see also Massam 2009;to appear): the lack of an object in unergatives is not the only factor which distinguishes them from erg-abs transitive frames. In this final section before concluding, I address some of the wider theoretical issues and typological predictions raised by this proposal. What is the potential range of crosslinguistic variation? Are unergative subjects merged lower than transitive subjects in all languages?
First, it should be stressed that Samoan unergative subjects are merged in the same position as some types of transitive subjects, namely the subjects of "middle" predicates. Under this approach, unergatives and some transitive subjects share common properties, while high agent ergative subjects are thematically "peripheral" (in the sense of Kim 2011). There is no a priori reason to maintain that every language should exhibit the same split v/Voice system proposed for Samoan -or, if so, for the two proto-roles to be governed by the same sematic criteria. In contrast to Samoan, for example, Niuean does not allow inanimate ergative agents (Massam 2006), suggesting that inanimate agents cannot be mapped to the proto-high role in Niuean. 33 Similarly, the merge of further verbal structure is not obligatory for a lexical root with a specifier (e.g. sink in (63a)), such that both the causative and inchoative constructions are possible (see Hale & Keyser 2002). Looking beyond Polynesian, different variants of the ergative alignment pattern are observed. In a number of so-called "active" languages (e.g., 66, 67), unergative subjects are marked with ergative case (see Bittner & Hale 1996;Woolford 2015) -exactly as predicted by the standard approach to argument structure shown in (1)  This suggests that active languages do not make the same syntactic distinction between high and low agents that is evident in Samoan: all external arguments are merged in a single v/VoiceP, and assigned ergative case by v/Voice 0 . It is also expected that an alignment may undergo historical change and/or dialect variation arising from splitting or collapsing of vP and VoiceP. This is precisely what we find in Basque: while western dialects (67) exhibit an active alignment, with unergative and transitive subject both marked ergative, eastern dialects exhibit the same type of alignment as Samoan, with unergative an unaccusative subjects both surfacing as absolutive. Since eastern dialects are reported to be more conservative, and western dialects more innovative (Aldai 2008), this change conceivably amounts to a collapsing of vP and VoiceP in the western dialects, while the split v/Voice system persists in eastern dialects. Another possibility for active languages would involve the v/Voice split of the type recently proposed by Harley (2013) and Legate (2014), in which v does not introduce an argument, such that all proto-agents are merged in VoiceP.

On the absence of dedicated unergative morphology
One final typological puzzle remains. The split v/Voice structure proposed in this paper might predict the existence of languages in which unergative subjects are marked differently from both transitive subjects and internal arguments (unaccusative subjects and direct objects). So far, nothing in the theory presented here explicitly rules out a language with a split vP structure in which v 0 and Voice 0 both assign a uniquely marked Spec-Head case, as in (68). (68) Hypothetical Bi-ergative language 39 both transitive subjects and internal arguments (unaccusative subjects and direct objects). So far, nothing in the theory presented here explicitly rules out a language with a split vP structure in which v 0 and Voice 0 both assign a uniquely marked Spec-Head case, as in (68).
(68) Hypothetical Bi-ergative language VoiceP HIGH EXTERNAL Voice' To my knowledge, no such language has ever been reported. Why should such a typological gap arise, if the grammar allows the option to merge transitive and unergative subjects in different positions?
The reason for the absence of (68) may lie partially beyond the realm of the narrow syntax. Du Bois (1987;2017) argues extensively that ergative systems are isomorphic to inherent patterns of human discourse, proposing a "Preferred Argument Structure", in which transitive subjects are more likely to be avoided more than any other core argument (i.e., intransitive subject; transitive object) in discourse. Ergative alignments are, according to Du Bois, a grammatical manifestation of universal discourse preferences: as a (partial) result of being least favoured in sentence production, transitive subjects are most morphologically marked. Intransitive subjects are much more widespread than transitive subjects; Du Bois relatedly observes in a corpus study that intransitive clauses are more prevalent than transitive clauses. Given these preferences, it is unsurprising that intransitive subjects are most commonly unmarked cross-linguistically. Since unergative verbs are (for the most part) intransitive, there would be no usage-based motivation for a language to develop a grammar which would license them in a unique way, especially if they could be licensed by another means (e.g., by T 0 ).
Considering also the properties of proto-low agents as compared with proto-high agents, high agents have several distinctive featural properties which low agents lack (i.e., a-iii through avii in (39)). By contrast with high agents and patients, there is nothing inherently distinctive at all about low agents: of the four properties in (39) which characterize low agents, twoinitiation and experience of an event (b-i and b-ii)are shared with high agents. The other two features (b-iii and b-iv) can be characterized as defaults to denote the lack of properties typical of either high agents (i.e., causing a change in state or affecting another entity) or patients (i.e., undergoing a change in state or being affected by the event). Thus, the absence of languages with a morphological "Case B" follows from the relative semantic unmarkedness of proto-low To my knowledge, no such language has ever been reported. Why should such a typological gap arise, if the grammar allows the option to merge transitive and unergative subjects in different positions?
The reason for the absence of (68) may lie partially beyond the realm of the narrow syntax. Du Bois (1987;2017) argues extensively that ergative systems are isomorphic to inherent patterns of human discourse, proposing a "Preferred Argument Structure", in which transitive subjects are more likely to be avoided more than any other core argument (i.e., intransitive subject; transitive object) in discourse. Ergative alignments are, according to Du Bois, a grammatical manifestation of universal discourse preferences: as a (partial) result of being least favoured in sentence production, transitive subjects are most morphologically marked. Intransitive subjects are much more widespread than transitive subjects; Du Bois relatedly observes in a corpus study that intransitive clauses are more prevalent than transitive clauses. Given these preferences, it is unsurprising that intransitive subjects are most commonly unmarked cross-linguistically. Since unergative verbs are (for the most part) intransitive, there would be no usage-based motivation for a language to develop a grammar which would license them in a unique way, especially if they could be licensed by another means (e.g., by T 0 ).
Considering also the properties of proto-low agents as compared with proto-high agents, high agents have several distinctive featural properties which low agents lack (i.e., a-iii through a-vii in (39)). By contrast with high agents and patients, there is nothing inherently distinctive at all about low agents: of the four properties in (39) which characterize low agents, two -initiation and experience of an event (b-i and b-ii) -are shared with high agents. The other two features (b-iii and b-iv) can be characterized as defaults to denote the lack of properties typical of either high agents (i.e., causing a change in state or affecting another entity) or patients (i.e., undergoing a change in state or being affected by the event). Thus, the absence of languages with a morphological "Case B" follows from the relative semantic unmarkedness of proto-low agents compared with other arguments, which would likely constrain any potential motivation of a grammar to mark them in a distinctive manner.

Conclusion
In this paper, I propose that unergative subjects in Samoan are structurally lower than ergative subjects, building upon an observation that adding an object to an unergative verb does not yield an erg-abs case frame; instead, an abs-i pattern arises. This case pattern is found in another construction: the Samoan "middle". As discussed in section 3, both middle and unergative i-marked objects behave as direct objects per a variety of tests, leading to the conclusion that abs-i middle and unergative + object constructions are syntactically fully transitive. In view of this, I propose that Samoan has two transitive case frames and two external argument positions: a low position (the specifier of vP) and a high position (the specifier of VoiceP). Unergative and middle subjects are introduced by v 0 , while ergative subjects are merged in VoiceP.
Section 4 discussed the various semantic differences between erg-abs-and abs-ipatterning verbs. While absolutive subjects tend only to initiate or experience an event, ergative-marked subjects typically possess several additional properties, notably volition, effortfulness, cause of a change in state, effect upon another participant, and conclusion of the event. However, no single property can categorically determine, for any given verb or event, whether an erg-abs or an abs-i case frame will surface. Following Dowty's (1991) theory of thematic proto-roles, I propose that unergative/middle and ergative subjects constitute two subclasses of what Dowty refers to as proto-agent: proto-low agent (a proto-role assigned by v 0 ) and proto-high agent (a proto-role assigned -along with ergative case -by Voice 0 ). Proto-low agents typically possess properties associated with a low degree of agentivity, such as initiation and/or experience of an event, but they neither affect a participant, nor are themselves directly affected by the event. Proto-high agents possess a number of additional properties: effort, volition, and initiation of an effect or a change in state of another participant. These additional properties given to high agents by their selecting verbs constitute more phrasal structure.
Turning to the syntax of case marking, I proposed in section 5.2 that the contrast between the two case frames arises from differing case assignment properties of v 0 and Voice 0 . Voice 0 assigns ergative case to a proto-high agent in a Spec-Head Agree configuration, and an erg-abs pattern arises. When a proto-low agent is merged in vP, v 0 is endowed with a structural accusative case feature, which is realised upon the internal argument as prenominal i, yielding an abs-acc case frame. When no proto-low agent is merged, however, v 0 is no longer able to assign accusative case, as per Burzio's (1986) Generalization; merging of a proto-high agent in VoiceP does not influence the case assigning properties of v 0 . The ergative-absolutive pattern is hereby derived from additional phrasal structure which is not present in a non-ergative configuration.
More broadly speaking, this paper provides further arguments in favour of a split between vP and VoiceP (e.g., Pylkkännen 2008;Harley 2013;Legate 2014;a.o.), and supports Massam's (2009;to appear) claim that unergative subjects are merged lower than full-fledged transitives. This treatment of unergatives illustrates how lack of ergative case on unergative subjects need not pose a problem for an inherent approach to ergative case (as is claimed by Baker and Bobaljik 2017): in Samoan, ergative case fails to appear on unergative subjects because they are structurally lower than ergatives, rather than because of the absence of an object as a case competitor. Building upon Massam's proposal, I also show how differing case assigning properties of v 0 and Voice 0 may yield two distinct transitive case frames -with ergative case assigned only when VoiceP is present. Abbreviations abs = absolutive; acc = accusative; agt = agent; aux = auxiliary; caus = causative; cl = clitic; det = definite/indefinite determiner; dir = directional; erg = ergative; es = ergative suffix; gen = genitive; nom = nominative; obl = oblique; q = question marker; pass = passive; pers = person; pl = plural; poss = possessive; pro = pronoun; prog = progressive; prs = present; prv = perfective; pst = past; pt = presentative; sg = singular.