Two steps to high absolutive syntax: Austronesian voice and agent focus in Mandar

Abstract

The westernmost languages of the Austronesian family show verbal alternations that are traditionally referred to as a voice system. This paper investigates the syntax of the voice system in Mandar, a language of the South Sulawesi subfamily. It argues that this alternation tracks alternations in argument structure, determines patterns of Case-Licensing in the voice p, and positions a single argument, the pivot, to raise to the highest a-position in the clause. The process that positions the pivot is decomposed into two steps: first, a process of Object Shift that moves definite arguments out of the vp and second, a process that places the pivot in spec,tp as the result of Case-Licensing by t⁰. Evidence for this analysis is drawn from contexts where the external argument undergoes ${\bar{\textsc {a}}}$-extraction and the internal argument is definite. In that context, the language employs a special construction which allows the external argument to be the pivot, allows the internal argument to undergo Object Shift, and provides the means to Case-License it within the v p. I refer to this construction as the Agent Focus and argue that it has a syntax similar to the analogous construction in the Mayan languages of the Q’anjob’alan subfamily.

1 Introduction

The westernmost languages of the Austronesian family, spoken along an arc that stretches from Taiwan to Sulawesi to Madagascar, show morphological alternations on the verb that are traditionally referred to as a voice system. (van der Tuuk 1864; Adriani 1893). These alternations have ramifications that spread through the morphosyntactic system, determining patterns of case-marking, interacting with ${\bar{\textsc {a}}}$-extraction, and singling out one argument- the “pivot”- as the most prominent nominal in the clause (Keenan 1976; Guilfoyle et al. 1992; Kroeger 1993a; Rackowski 2002; Aldridge 2004; Pearson 2005; Erlewine 2018; Nie 2020; Hsieh 2020). The complexity of these systems has inspired a body of research on the nature of the voice morphology, the position of the pivot, and the mechanisms that position that argument in its final position, and these topics continue to be debated in individual languages and at the broader regional level (Payne 1982; De Guzman 1988; Gerdts 1988a; Shibatani 1988; Kroeger 1993a; Richards 1993, 1996, 2000; Paul 2000; Pearson 2001, 2005; Aldridge 2008, 2011, 2012; Liao 2004; Paul and Travis 2006; Ndayiragije 2006; Chang 2011; Chen 2017; Erlewine and Lim 2022)

The goal of this paper is to investigate the syntax of this system in Mandar, a language of the South Sulawesi subfamily of central Indonesia. The languages of this subfamily show a morphological profile that is relatively unique to the island of Sulawesi, and as a result, typological and formal discussions have made few attempts to connect their voice systems to those of the Philippines (Finer 1997; Himmelmann 2005; Blust 2013; Chen and McDonnell 2019). But when we look deeper into the syntax of this system in Mandar, we will find that it can be understood along much the same lines as the systems of that type. I will show that the voice alternation in this language tracks alternations in argument structure, as it does in Philippine-type languages on the ergative tradition of analysis (Payne 1982; De Guzman 1988; Gerdts 1988a; Aldridge 2004; Liao 2004; Paul and Travis 2006; Ndayiragije 2006; Chang 2011). The alternation between the two basic voices- the Agent Voice and the Patient Voice- turns on the requirement for definiteness-related Object Shift (Rackowski 2002; Aldridge 2004) and determines patterns of Case-Licensing in the voice p. Its ultimate result is to position one argument- the pivot- to be Case-Licensed by t⁰ and to raise to spec,tp (Guilfoyle et al. 1992). The resultant analysis is one on which the pivot reaches its final position in the Patient Voice through two distinct steps, and this is shown in the tree in (1). I will refer to this theory of Mandar clause structure as the Two-Step Theory.

On the Two-Step Theory, the syntax of the Mandar voice alternation resembles the transitivity alternations that occur in ergative languages which require the absolutive argument to raise to the highest a-position in the clause: namely, those which show High Absolutive Syntax (Bittner and Hale 1996b, a; Legate 2006; Coon et al. 2014; Brown 2016; Ershova 2019). From this perspective, it is possible to integrate this language into a broader continuum of High-Absolutive ergative systems that stretches across Western Austronesia (Payne 1982; De Guzman 1988; Gerdts 1988a; Aldridge 2004; Liao 2004; Paul and Travis 2006; Ndayiragije 2006; Chang 2011). But pressing further than earlier work in the ergative paradigm, the Two-Step Theory makes it possible to recognize a parallel with certain High Absolutive languages of the Mayan family which has gone unnoticed in previous work in the region. Mandar shows a Subjects-Only Extraction Constraint in the ${\bar{\textsc {a}}}$-domain, and in clauses that launch extraction of the external argument, it is impossible for the internal argument to interact with t⁰. On the Two-Step Theory, this raises a problem when the internal argument is definite: the internal argument must undergo Object Shift and be Case-Licensed without interacting with t⁰, but these requirements cannot be satisfied in the Agent Voice (which does not allow Object Shift) or the Patient Voice (which does not allow the external argument to interact with t⁰ and be extracted). As a result, the Two-Step Theory predicts that the language must resort to a distinct construction, and this prediction is correct. In that context, Mandar employs the verbal form in (2). I will refer to this construction as the Agent Focus and argue that it shows the syntax in (3): it allows the internal argument to undergo Object Shift and be Case-Licensed in the v p.

The Mandar Agent Focus construction in (2) is analogous to a verbal form with the same shape and distribution in many languages of Sulawesi (Martens 1988; Friberg 1991; Mead 1998, 2002; Zobel 2002; Jukes 2006). When properly understood, its properties will provide clear support to the key claims of the Two-Step Analysis in Mandar. But beyond the level of language-internal analysis, this investigation will also deliver a parallel to a similar construction that appears in many High Absolutive languages of the Mayan family: the Agent Focus Construction. In certain Mayan languages that show a clausal syntax on which the absolutive argument is Case-Licensed by t⁰, it has been argued to provide the means to Case-License the internal argument when it cannot interact with that head, along the lines of (3) (Coon et al. 2014).

The remainder of the paper proceeds as follows. In Sect. 2, I provide background information on Mandar, show that the voice system tracks alternations in argument structure, and demonstrate that the pivot raises to an a-position. In Sect. 3, I present the Two-Step Theory. In Sect. 4, I adduce further evidence for this proposal from the contexts where the demands of the systems of Extraction, Object Shift, and Case Licensing conflict, in which the language employs the Agent Focus Construction. In Sect. 5, I conclude.

2 High absolutive syntax in Mandar

Mandar is an Austronesian language of the South Sulawesi subfamily which is spoken by 400,000 people on the south coast of West Sulawesi, Indonesia (Grimes and Grimes 1987). It has been the subject of several publications by the Indonesian Ministry of Education and Culture, including a grammar (Pelenkahu et al. 1983), a description of adverbs (Sikki et al. 1987), and a compilation of poetry (Muthalib and Sangi 1991). Those publications and this paper focus on the prestige variety of the language, which is the dialect spoken in the district of Polewali Mandar (Grimes and Grimes 1987). I will refer to it as Mandar.

The data that appear in the discussion to come are drawn from prior descriptive work, Mandar-languague media, and elicitation with one native speaker of the language, Jupri Talib (JT). This elicitation has been ongoing since the fall of 2018, and it has been conducted using a standard methodology for linguistic fieldwork (Matthewson 2004). Judgments were collected in the following way: a discourse context was set up in Indonesian, and the consultant then was presented with a sentence or pair of sentences in Mandar (attributed to an imagined speaker interacting with an established set of characters). The consultant was then asked, in Mandar, to judge the well-formedness of each sentence in comparison with minimally different examples of similar shape. Each example was then discussed in Mandar, with occasional recourse to Indonesian to clarify the context or cross-check the interpretation. All judgments were checked on minimally two separate occasions, and data are cited from the date of their final elicitation.

2.1 Mandar basics

Mandar has much in common with typical Philippine-type languages (Himmelmann 2005), including a complex voice system, a system of second-position clitics, and a predicate-initial word order (Brodkin 2020, 2021a, b, c, 2022b, a). These properties are shared with the other languages of the South Sulawesi subfamily, and they have been well-described in those languages in a body of descriptive and theoretical work (Campbell 1989; Friberg 1991, 1996; Strømme 1994; Matti 1994; Valkama 1995a, b; Jukes 2006; Lee 2008; Kaufman 2008; Laskowske 2016; Finer 1994, 1997, 1998, 1999; Béjar 1999; Finer and Basri 2020). The structural parallels between these languages and those of the Philippines are partially obscured by a set of morphological innovations common to many languages of Sulawesi, including the loss of case morphology and the rise of agreement, and as a result, they are often left unmentioned or excluded from typological discussion of the languages of Western Austronesia (Himmelmann 2005; Chen and McDonnell 2019). But pursuing the connection between these systems will provide the means to sharpen our investigation of Mandar and inform our broader understanding of Western Austronesian voice systems in turn.

To begin, the language shows a predicate-initial word order. The verb follows negation and aspectual auxiliaries, and in pragmatically neutral contexts, it precedes all arguments and vp-level adjuncts (4a). Across all bivalent clause types, the external argument (ext) canonically precedes the internal argument (int) (4b), yielding a basic word order of vso (more accurately, v-ext-int). Constituent order is flexible in the postverbal domain, and all arguments can be separated from the verb and reordered with each other, as elsewhere in the South Sulawesi Subfamily (Friberg 1996; Finer 1997; Jukes 2006). Nevertheless, non-default orders like vos (verb-int-ext) are prosodically and pragmatically marked.¹

Every finite clause in the language contains a syntactically privileged argument that I will underline and refer to as the pivot. This argument is the most phrase-structurally prominent nominal in the clause. It can originate in any of several thematic positions, much like the arguments identified as pivots in other languages of Western Austronesia (Keenan 1976; Schachter 1976, 1996; De Guzman 1988; Shibatani 1988; Gerdts 1988b; Guilfoyle et al. 1992; Kroeger 1993b; Richards 2000; Rackowski 2002; Rackowski and Richards 2005; Pearson 2005; Legate 2014; Chen 2017; Erlewine et al. 2017). The pivot is not distinguished from other postverbal arguments by its morphological case or linear position, as the language lacks morphological case distinctions and shows a consistent basic word order of v-ext-int. Rather, it is singled out in the system of agreement. Every finite clause contains an agreement enclitic which surfaces in second position, following the leftmost phonological phrase. It spells out the person features of the pivot. An example is shown in (5), where the pivot is a 1sg pronoun and it is indexed by the first-person agreement enclitic a’.

The identity of the pivot is closely tied up with a system of morphological alternations on the verb. In Mandar, finite verbs typically surface in one of two forms when they take both an ext and an int. The first is the Agent Voice (av), which is marked with m-initial prefixes like maN- (6a; note that coda nasals denasalize before all non-nasal segments but ; Pater 1999). The second is the Patient Voice (pv), which is marked with a prefix that spells out the person features of the ext (6b).

The alternation between av and pv determines the pattern of clause-level agreement and, as a result, the identity of the pivot. In clauses whose verbs surface in av, the second-position clitic tracks the ext (6a), and in those clauses, the pivot is the ext. In clauses whose verbs show pv, in contrast, the clitic tracks the int (6b), and in those clauses, the pivot is the int. The resultant interaction between voice, agreement, and pivot status is very similar to the system that has been described in Philippine-type languages, where an analogous voice alternation determines the identity of the argument that receives a particular morphological case and shows the morphosyntactic privileges of pivot status (Schachter 1976; Payne 1982; Cumming and Wouk 1987; De Guzman 1988; Shibatani 1988; Gerdts 1988b; Guilfoyle et al. 1992; Kroeger 1993b; Mithun 1994; Himmelmann 1996, 2005; Richards 2000; Wouk and Ross 2002; Rackowski 2002; Rackowski and Richards 2005; Aldridge 2004; Riesberg 2014; McDonnell 2016; Chen 2017; Erlewine et al. 2017; Erlewine and Lim 2022; Nie 2019, 2020; Hsieh 2020), and similar systems in other languages of Western Austronesia (Chung 1976b; Wechsler and Arka 1998; Paul 2000; Pearson 2001, 2005; Legate 2014; Hemmings 2015, 2016; Erlewine 2018).

The languages of the South Sulawesi subfamily are traditionally described as showing morphological ergativity, as the system of agreement contrasts a paradigm of agreement with the transitive ext (the prefixes that mark pv) with a paradigm of agreement with the transitive int and the sole argument of an intransitive verb (the second-position clitics that index the pivot). In this respect, it can be compared directly to the system of case-marking in Philippine-type languages on the tradition that takes those languages to show morphological ergativity as well (Payne 1982; De Guzman 1988; Gerdts 1988a; Aldridge 2004; Liao 2004; Paul and Travis 2006; Chang 2011; Nie 2020). Adopting this perspective, I take the the second-position clitic that tracks the pivot to spell out absolutive agreement, assume that pv is the basic transitive construction, and take the prefixes that track the non-pivot ext in that voice to spell out ergative agreement, on a par with the ergative analysis of the case morphology of the languages of that type.

As a consequence of this analysis, av is treated as a morphologically intransitive construction which allows the ext to trigger absolutive agreement despite the presence of an int. Following the ergative analysis, I will refer to it as an antipassive. This terminological move obscures the fact that av does not require the int to be demoted (in fact, I will argue that it assigns abstract Accusative Case), but it accounts for the fact that it allows the int to be omitted (7), as is typical of antipassives outside of Western Austronesia (Heaton 2020). And looking ahead, this move will open up a productive line of comparison with the transitivity alternations that occur in ergative languages outside of Western Austronesia (Craig 1979; Ayres 1983; Bittner 1987; Spreng 2006, 2013; Polinsky 2017a; Vigus 2018; Heaton 2020).

Beyond the interaction between voice, agreement, and pivothood, there are two further properties of the Mandar voice system that will be important in the discussion to come. The first involves the system of extraction. In Mandar, the processes of wh-movement, focus-fronting, and relativization are subject to the restriction that they target only the pivot. This is the Mandar-particular instantiation of the Subjects-Only Extraction Constraint that is common to many languages of Western Austronesia (Keenan 1976), and it resembles the Ergative Extraction Constraint that recurs in High Absolutive languages further afield (Larsen and Norman 1979; Campana 1992; Murasugi 1992; Campana 1992; Bittner and Hale 1996b, a; Otsuka 2006; Legate 2008; Coon et al. 2014; Assmann et al. 2015; Brown 2016; Deal 2016; Polinsky 2016, 2017b; Aissen 2017; Ershova 2019).

The following examples show the Subjects-Only Extraction Constraint with wh-movement. In Mandar, there is a process of wh-movement that moves wh-words to the left periphery and triggers the disappearance of absolutive agreement, rendering the clause non-finite. This is a pattern of Anti-Agreement (Ouhalla 1993), and here I assume that it arises as a result of morphological impoverishment linked to ${\bar{\textsc {a}}}$-extraction (Baier 2018; cf.Brodkin 2021a). Example (8a) shows an av clause, where the pivot is the ext. In such a clause, it is possible and obligatory for the ext to undergo wh-movement when it is a wh-word (8b). It is impossible, however, for a non-pivot int to do the same (8c).

The second pattern of interest involves a definiteness effect (Milsark 1974). In bivalent clauses that do not show extraction, the choice of voice morphology is restricted by the definiteness of the int (Lee 2008). The basic alternation is shown in the examples in (9), which form an excerpt from a narrative. In example (9a), the verb ala “catch” has an int which is indefinite: bau “fish.” In that context, the verb must surface in av, which is marked on this verb with the prefix me-. In example (9b), the verb ande “eat” takes an int which is definite: di’o bau o “those fish.” In that context, this verb is forced to surface in pv. The class of internal arguments that trigger this alternation includes pronouns, names, strongly quantified nouns, and all types of definite description, including those which are not explicitly anaphoric but still contextually unique (Russell 1905; Abbott 1999; Roberts 2003): This can be seen in contexts like (9c), which represents an out of the blue comment made in England. There, the verb must surface in pv when the int is the non-anaphoric but contextually unique mara’dia “queen.”

The Definiteness Effect, once again, reflects a broader pattern that is common to many languages of the region: both those of the South Sulawesi subfamily (Campbell 1989; Valkama 1995b; Friberg 1996; Jukes 2006; Laskowske 2016), and many languages of the Philippines as well (Bloomfield 1917; Adams and Manaster-Ramer 1988). This pattern will provide the motivation for a key step in the syntactic analysis of the voice system, and once its analysis is in place, we will return to the way it which it interacts with the Subjects-Only Extraction Constraint to develop an argument for the Two-Step Theory.

2.2 The role of the voice morphology

With the basic shape of the voice system established, we can now turn to the role of the voice morphology. On the ergative analysis, the voice alternations of Philippine-type languages are traditionally understood to track alternations in argument structure (Payne 1982; De Guzman 1988; Gerdts 1988a; Aldridge 2004; Liao 2004; Paul and Travis 2006; Chang 2011; Nie 2020). I will refer to this view as the Argument Structure Analysis. It contrasts with an alternative, the Indexing Analysis, which takes the voice morphology to index properties of the argument that becomes the pivot, such as its abstract Case or its thematic role (Shibatani 1988; Kroeger 1993a; Pearson 2001, 2005; Rackowski 2002; Rackowski and Richards 2005; Chen 2017). The analytical divide between these approaches has ramifications that spread throughout the analysis of the syntax, influencing theories of the structural position of the voice morphology, the position of the pivot, and the extraction restriction. The debate between the two, moreover, cross-cuts the debate on morphological ergativity, and many authors do not employ the terminology of ergativity but adopt proposals that pattern more closely with those of the first camp (Guilfoyle et al. 1992; Legate 2014; Erlewine et al. 2017; Erlewine 2018; Erlewine and Lim 2022; Ting 2022).

The Argument Structure Analysis has been extended to many languages that are not Philippine-type, such as Malagasy (Ndayiragije 2006; Ting 2022) and Indonesian (Aldridge 2008), and here I argue that it provides the right means to understand the voice system of Mandar as well. This move continues the traditional stance of the descriptive literature on the languages of the South Sulawesi subfamily (Campbell 1989; Valkama 1995b; Jukes 2006; Laskowske 2016), and it allows the analysis of Mandar to be integrated into a framework that places the languages of Western Austronesia along a broad continuum of Ergative-to-Nominative alignment (Paul and Travis 2006; Aldridge 2011; Chang 2011).

The first piece of evidence for the Argument Structure Analysis comes from the shape of the voice morphology. Consider a version of the Indexing Analysis which took the Mandar voice morphology to spell out the Abstract Case of the pivot (Rackowski 2002). This analysis would encounter an immediate challenge in the morphology of pv, which is an ergative prefix that spells out the person features of the non-pivot ext. It is not clear how this could be taken to spell out abstract Accusative Case or any other property of the int. This suggests that the Indexing Analysis is on the wrong track.

The second argument for this analysis comes from the distribution of the voice morphology. In Mandar, the presence of voice morphology is sensitive to the argument structure of the verb. There is a morphological split between two types of syntactically intransitive verbs in this language: unaccusative verbs invariably lack voice morphology (10), while unergative verbs typically surface with an av affix like $-$ um $-$ (11).

This split can be easily captured on the Argument Structure Analysis, on which the voice morphology can be positioned in a set of heads that select for those which determine the argument structure of the verb. But it is difficult to explain on an Indexing approach, which would be forced to stipulate that agreement-like morphology was absent from a broad class of verbs that shared a single argument structure.

The third argument for the Argument Structure Analysis comes from the broader voice system of the language. Beyond av and pv, Mandar shows a range of additional voices that demonstrably trigger the introduction or suppression of arguments. There are, for instance, three passive constructions that prohibit the expression of the ext (12). These constructions are each built with a prefix that appears in complementary distribution with the morphology of av and pv: di- (12a), ti- (12b), and ka- (12c).

In the same vein, the language has a Comitative Voice (cv) that is built with the prefix si- (13). This prefix alternates with the morphology of av and pv and creates several different types of predicates, including reciprocals (13) and comitatives (for careful discussion of its function in other languages of the subfamily, see Valkama 1995b and Jukes 2006). In many cases, it plays an applicative function: the predicates that surface in this voice are able to take arguments that would otherwise need to be expressed in pps. This is shown in the examples in (14). The adjective kadeppe’ ‘near’ requires its locative argument to surface in a pp when it takes no voice morphology (14a), but in cv, it allows that argument to surface without the p⁰ (14b).

Finally, the language has a Locative Voice (lv) which is built from two morphological parts: the ergative prefixes that mark pv and an applicative suffix -ang (8). This suffix continues the Proto-Austronesian lv suffix *-en (Zobel 2002), but it does not only appear in contexts in which the pivot is an applied argument (goal), like (15a). Rather, it also appears in contexts where applied arguments are introduced but do not trigger absolutive agreement. For instance, ditransitive verbs take this suffix even when they surface in av, though the resultant clauses show absolutive agreement with the ext (15b). This suggests that -ang is simply an applicative suffix, much like the Indonesian -kan (Chung 1976a).

2.3 The subject position

With the Argument Structure Analysis secure, we can now turn to the syntactic behavior of the pivot. On the basis of patterns like the Subjects-Only Extraction Constraint, it is commonly assumed that the pivot raises to a phrase-structurally prominent position in many languages of Western Austronesia. Despite near-consensus around this view, there is disagreement over the nature of the position to which the pivot moves. Some authors take it to be an a-position, and as a result, treat the pivot as a nominative subject or High Absolutive argument (Guilfoyle et al. 1992; Rackowski 2002; Aldridge 2004; Erlewine and Lim 2022, a.o.). Others take it to be an ${\bar{\textsc {a}}}$-position (Richards 1993, 1996, 2000; Pearson 2001, 2005; a.o.) or a mixed a/${\bar{\textsc {a}}}$-position (Erlewine et al. 2017; cf. Van Urk 2015; Zyman 2018) and as a result, treat the pivot as a type of topic.

Following the a-analysis, I propose that the pivot in Mandar raises to the highest a-position in the clause. This step of movement is covert and has no influence on the word order of the language, but it can be shown to occur with a range of diagnostics which are familiar from other languages of the area.

2.3.1 Condition C

The first argument for a-movement comes from patterns of pronominal coindexation. Following Chomsky 1981 and Reinhart 1983, I assume that Condition C of the Binding Theory rules out coindexation between a pronoun in an a-position and an r-expression that sits in its c-command domain. This sets up two predictions. First, if the pivot moves to an a-position above all other arguments in the clause, then it should be impossible for it to be a pronoun that is coindexed with an r-expression in another argument. This yields a ban on the Mandar analogue of the coindexation in (16):

Second, on the assumption that a-movement does not reconstruct for Condition C, it predicts that it should be possible for a pivot to contain an r-expression that is coindexed with a pronoun in the position of any non-pivot argument in the clause. This yields the possibility for the Mandar analogue of the coindexation in (17).

In Mandar, these predictions are both correct. The pattern of coindexation in (16) is indeed impossible: when the pivot is the int, it cannot be a pronoun that is coindexed with an r-expression in the ext. This is shown in example (18).

In this example, the pivot is the int, which is a null pronoun, and the ext is the dp “Nina and Kaco’s mom.” It is impossible for the r-expressions in the ext, Nina and Kaco’, to be interpreted as coindexed with the null pronominal int. In other words, the pattern of coindexation in (16) is ruled out.

In the same vein, the pattern of coindexation in (17) is grammatical: when the int is the pivot, it can contain an r-expression that is coindexed with a pronoun in the position of the ext. There are several contexts in which this pattern can be seen. For instance, it comes into view in clauses where the int contains a possessor, as in (19).

In this example, the pivot is the int “Nina and Kaco’s mom,” and the ext is the null pronoun “she.” This sentence allows the coindexation in (17): the pronominal ext may be coindexed with the r-expression Nina, which is a possessor of the int.

The same pattern can be seen in contexts where the int contains a relative clause. An example of such a configuration is given in (20).

In this example, the pivot is the int “the book that Nina bought yesterday,” and the ext is the null pronoun “she.” This clause, like that in (19), allows the pronominal ext to be coindexed with the r-expression inside of the pivot. For this pattern of coindexation to be possible, it follows that the pivot must undergo a step of a-movement to a position above all other arguments in the clause.

2.3.2 Variable binding

The second argument for the a-analysis lies in the system of variable binding. Following Reinhart 1983, I assume that variable binding between clausemate arguments is sensitive to c-command: quantified arguments can only bind variables in their c-command domain. On the a-movement analysis, this assumption leads to two predictions. First, it should be possible for the pivot to bind variables in every other argument in the clause, on a par with the arguments that undergo a-movement to the subject position in English (21).

Second, it should be possible for a variable in the pivot to be bound by a quantified argument that c-commands it in its base position. This is because unambiguous cases of a-movement are able to reconstruct for variable binding, as seen with Raising in (22).

Transposed to Mandar clauses in pv, this leads to the following predictions. First, it should be possible to observe the pattern of variable binding in (23a): a quantified int should be able to bind a variable in the ext. Second, it should be possible to observe the pattern of variable binding in (23b): the int should be able to contain a variable that is bound by the ext.

To test these predictions, a brief detour into the system of quantification is required. In Mandar, universal quantification is expressed with the second-position element nasang. This element construes with one postverbal argument, but is linearly separated from it by other second-position clitics and any other arguments or adjuncts that intervene. Its position is shown in example (24): there, it associates with the dp di’o bau o “those fish.”

The quantifier nasang can associate with many types of arguments. Its associate can be the pivot, as in (24), or it can be the ext of a verb in pv, as in (25a). It can also be an oblique argument that surfaces inside of a pp, as in (25b).

With this background, we can now see that the predictions in (23a)-(23b) are correct. First, the language allows the pattern of reconstruction in (23b): when the pivot is the int, it can contain a variable that is bound by a quantified ext. This is shown in (26): the pivot is the int sanaekena ‘her kid,’ and the ext is the quantified nasang...kindo’ ‘every mother.’ The ext can bind the variable in the int, in the same way that experiencers can bind variables in the arguments that raise across them in English (22).²

Second, the language allows the pattern of binding in (23a): when the pivot is the int, it can bind a variable in the ext. This pattern holds without regard to the linear order of arguments in the postverbal domain, and it is shown with several different word orders in (27).

In example (27a), the pivot is the int nasang... passikola “every student” and the ext, which precedes it, is gurunna “her teacher.” The ext contains a variable, and it is possible for the int to bind this variable. Example (27b) shows a similar pattern with the order of the ext and int reversed: here, the pivot is the int nasang... sanaeke “every kid,” and the ext, which follows it, is kindo’na “her mom.” As in (27a), this clause allows the int to bind the variable in the ext.

These patterns of variable binding, then, provide a second argument that the pivot moves to an a-position above all other arguments in the clause.³

3 Two steps to high absolutive syntax

The conclusions of the preceding section guide the analysis of the Mandar voice system in two ways. First, the adoption of the Argument Structure Analysis suggests that the voice alternation must be localized to a set of heads that sit in the thematic domain. Second, the fact that the pivot raises to an a-position suggests that these alternations must trigger or feed movement to such a position. These conclusions are familiar from work on the syntax of other languages of the Western Austronesian region (Aldridge 2004), and they are similar to those that have been drawn in many ergative languages outside of this region which show High Absolutive syntax (Bittner and Hale 1996b, a; Legate 2008, 2017; Coon et al. 2014; Brown 2016; Ershova 2019).

The principal puzzle that surrounds the High Absolutive configuration is one of locality: in contexts where the absolutive argument originates beneath the ext, how can it cross over that argument to reach its high position? This problem is sketched for pv in (28).

The literature has furnished two types of answer to this question. The first involves non-intervention: in contexts where a non-highest argument raises to the position of the pivot, some independent constraint is claimed to render all intervening arguments invisible to the attracting probe (e.g., Activity: Chomsky 2001) (Guilfoyle et al. 1992; Legate 2012, 2014, 2017; Ershova 2019). An alternative analysis involves leapfrogging (Douglas et al. 2017): in contexts where the int is the pivot, it has been argued that it undergoes an intermediate step of a-movement to become the highest dp in the c-command domain of the head which attracts it to its final position (McGinnis 1998; see also Coon et al. 2014). In the Western Austronesian context, an influential line of work adopts the second analysis and formalizes it in terms of Object Shift: the int invariably undergoes a step of leapfrogging movement above the ext when it is definite, and this positions it to become the pivot (Rackowski 2002).

This paper argues for a variant of the non-intervention analysis. In Mandar, I argue that the pivot raises to spec,tp because it enters into a Case-Licensing relationship with t⁰, yielding absolutive agreement. This step of movement draws the pivot over the base position of the ext in pv and lv, and this apparent violation of locality is made possible by the fact that in those voices the ext is Case-Licensed- and thus rendered Inactive- within the voice p. It is prefigured by a step of definiteness-related Object Shift in pv and lv, but this step does not place its target above the base position of the ext. As a result, the pivot reaches its high position in two distinct steps in pv and lv. The resultant syntax of a pv clause is schematized in tree (29). I will refer to this approach as the Two-Step Theory of High Absolutive Syntax.

3.1 The anatomy of the voice system

We can begin to build toward this theory by laying out a syntactic analysis of the voice morphology. So far, we have seen that av and pv appear in complementary distribution with other voices that drive the introduction and suppression of arguments. On the assumption that argument introduction is localized to a syntactic space that is close to the verbal root (Baker 1988), this pattern suggests that the morphology of the voice system is localized to a single string of heads that sits within that syntactic space.

This assumption is one that has much precedent in the literature on the Argument Structure Analysis in other languages of Western Austronesia (Aldridge 2004; Nie 2020; Erlewine and Lim 2022), and in Mandar, it can be supported with a separate line of evidence from the domain of morphology. In this language, many of the affixes in the voice system show a distribution that is lexically idiosyncratic, surfacing exclusively before certain roots or taking different shapes before others. These patterns suggest that the voice morphology sits in a set of heads that are able to select the root or show contextual allomorphy in its presence. And as both selection and contextual allomorphy turn on relationships of syntactic locality (Embick 2010; Bobaljik 2012; Merchant 2019), this provides convergent evidence that it sits close to the root.

This pattern is illustrated first with cv and the involuntary passive. Before certain roots, the prefix si- surfaces as siN- (30a), and before others, ti- surfaces as ta’- (30b).

The same pattern can be seen in av. Beyond maN- and $-$ um $-$, there are eight other prefixes that are recruited to mark this voice in Mandar: ma’-, me-, me’-, ma-, mo-, meN-, mu-, and m- (31; see Valkama 1995b for discussion of this morphology the related Duri). These affixes can be split into two syntactic classes, as the verbs which bear maN-, ma’-, me-, and me’- can take an int (31b) and those which take $-$ um $-$, ma-, mo-, meN-, mu-, and m- cannot (31c). But within each class, it is not possible to predict the affix that any given root will take when placed in av. As a result, a proper analysis of the distribution of these forms must involve some degree of contextual allomorphy or selection between the affix and the root.

Where, then, do these morphemes lie? I assume that the syntactic space around the verb has the shape in (32). It consists of four heads. The most deeply embedded is the verb itself. Above this is a low applicative head, v⁰_appl, which introduces applied arguments in its specifier (Pylkkänen 2008). Above this head sits v⁰, which is the head responsible for three separate roles: the introduction of the ext, the triggering of Object Shift, and the assignment of abstract Accusative Case (Marantz 1997; Collins 2005; Merchant 2013). This head, finally, is selected by voice⁰, which determines the morphological transitivity of the verb. To derive the surface v-ext-int word order, I assume that the verb undergoes head-movement up to voice⁰, placing it before all of its arguments and voice p-level adjuncts. I will refer to the space that contains these heads as the thematic domain.⁴

This decomposition provides the means to capture the network of syntactic relationships that hold between the appearance of particular voice morphemes and alternations in argument structure. Working from the bottom up, I propose that v⁰_appl hosts the two morphemes that introduce applied arguments: the suffix -ang and the cv prefix si-. In the same vein, I argue that v⁰ is the locus of the morphology that appears on verbs that do not take an ext, such as the passive prefixes di-, ti-, and ka-. This move captures the behavior of these morphemes: the suffix -ang and the cv prefix are applicative heads, while the passive prefixes are $v^0$s which do not introduce an ext.

Finally, I take the outermost morphology of av and pv to sit in voice⁰. This move has a number of advantages. To begin, it allows us to link the position of ergative agreement in Mandar to the low postion of ergative agreement outside of Western Austronesia (Wiltschko 2006; Coon 2017). Moreover, it captures the linear position of the av and pv morphology in the verbal complex. This is because the av and pv affixes strictly follow the prefixes that spell out heads in the middle field, such as the future marker na- (33a) and the emphatic negator taN- (33b). This fact of order follows from the mirror principle (Baker 1985) if they sit beneath the middle field.

Pressing further, this analysis provides the means to capture a number of the distributional restrictions that we have seen through selection. The morphology of av and pv sits in complementary distribution with the passive prefixes di-, ti-, and ka- because the voice⁰_av and voice⁰_pv select for v⁰s that introduce an ext. By the same logic, the voice⁰_av and voice⁰_pv do not appear with unaccusative verbs because these are selected by a null unaccusative v⁰. I assume that the passive v⁰ and the unaccusative v⁰ are both selected by a null intransitive voice⁰.

Most importantly, however, the analysis allows us to decompose av and pv in a manner that will be useful for the analysis to come. pv is built from two heads: first, a voice⁰_pv which hosts the ergative prefixes and second, a null v⁰_pv. In the same vein, av is decomposed into two heads as well: a voice⁰_av, which contributes the segment m-, and a v⁰_av, which contributes the remainder of the segmental content of the ten av affixes in the language.⁵ These affixes appear exclusively in av because the v⁰_av is strictly selected by the voice⁰_av. The syntactic split between these affixes can be localized to v⁰: those which allow for the presence of an int (maN-, ma’-, me-, me’-) contain an antipassive v⁰_av which is able to Case-License an int, while those that do not allow for the presence of an int ($-$ um $-$ , ma-, mo-, meN-, mu-, m-) contain an unergative v⁰_av which plays no Case-Licensing role.

The resultant analysis of the morphology at the core of the voice system is tabled in (34). In the discussion to come, I will abstract away from the morphological details of this system to focus on the syntactic properties of two heads: the v⁰_pv, which is null, and the antipassive v⁰_av, which is variably realized in portmanteaux with the voice⁰_av as maN-, ma’-, me-, and me’-. We will return to the morphology of av in Sect. 4.

3.2 Voice and object shift

With this much in place, we can now turn to the syntax of the voice alternation. Our starting point is the Definiteness Effect. As we have seen, the language requires bivalent verbs to surface in pv when the int is definite and av when it is not. Following Rackowski (2002) and Aldridge (2004), I take this alternation to turn on a need for Object Shift. In Mandar, I assume that all definite nominals are subject to a requirement that they move out of the vp (Diesing 1992). The int can only undergo this step of movement in the presence of an [$+$ epp] feature on v⁰, which triggers Object Shift and allows it to satisfy this constraint. I propose that the v⁰_pv and the v⁰_av differ with regard to the presence of this feature. The null v⁰_pv bears the feature [$+$ epp] and allows Object Shift (35a). The v⁰_av lacks the feature [$+$ epp], and as a result, does not allow Object Shift (35b).

This proposal provides the means to capture the relationship between voice and the definiteness of the int through a theoretical paradigm that has been developed around the Definiteness Effect in Tagalog (Rackowski 2002; Aldridge 2004; Rackowski and Richards 2005; Sabbagh 2016). In the terminology of that framework, the antipassive v⁰_avs cannot occur in contexts where the int is definite because they are anti-EPP morphemes: like the Tagalog pag-, they spell out a v⁰ which does not allow the int to undergo Object Shift. The distribution of the v⁰_av and v⁰_pv is governed by the logic of Last Resort. Following Chomsky 2001, I assume that the v⁰s which bear the [$+$ epp] feature are merged only when they will trigger movement that has an effect on semantic interpretation. As such, they only appear in clauses in which a definite argument must raise out of the domain of existential closure. Following Rackowski 2002, I assume further that the v⁰_av, which lacks the feature [$+$ epp], is “the default [ v⁰ ] and is blocked by a competing contextually conditioned Ø-[ v⁰_pv] in the presence of a [+EPP] value.” (Rackowski 2002, 89). This places the two v⁰s in a relationship of complementary distribution that is determined by the presence or absence of an argument that must undergo Object Shift.

Given that the v⁰_av is strictly selected by the voice⁰_av, the result of this analysis is that AV is ruled out in clauses where the int must undergo Object Shift. In that context, rather, the language requires the presence of a v⁰ that allows this process. Under typical circumstances, its response is to employ the v⁰_pv. And as this v⁰ is generally selected by the voice⁰_pv, the use of PV is forced.

3.3 The landing site of object shift

In Mandar, I propose that Object Shift places its target beneath the ext in an inner specifier of v⁰ (see Pearson 1998, Travis 2010 and Paillé 2021 for similar proposals in other languages of the region). As a result, pv clauses show the v p-level syntax in (36).

This proposal breaks from a range of previous analyses which assume that Object Shift places shifted arguments above the base position of the ext, as part of a leapfrogging derivation (Douglas et al. 2017). This class includes the proposals of Rackowski 2002, Aldridge 2004 and Rackowski and Richards 2005, other approaches to Philippine-type voice systems which assume that the pivot raises to the highest a-position within the v p (often for reasons other than definiteness-related Object Shift per se: Legate 2014; Erlewine and Levin 2021; Erlewine and Lim 2022, a.o.), and similar analyses of High Absolutive systems outside of Western Austronesia (Coon et al. 2014; Yuan 2018). The kind of movement that these analyses generally assume is sketched in (37).

The evidence for the proposed process of Low Object Shift in (36) comes from a set of contexts in which the influence of definiteness can be disentangled from that of Case-Licensing. In finite bivalent clauses, these two properties cannot be separated, because in all contexts where the int is definite, it triggers absolutive agreement and raises to become the pivot. As a result, it is impossible to determine whether the int moves above the ext as the result of a process of definiteness-related Object Shift, as in (37), or as the result of a separate interaction with the head that hosts agreement and assigns it abstract Case (as on the proposal of Guilfoyle et al. 1992). But the influence of definiteness can be separated from the influence of Case-Assignment in trivalent constructions, in which it is the goal that triggers absolutive agreement and raises to the position of the pivot. It is possible for the int to be definite in contexts of that type, and under those circumstances, it is clear that it cannot raise above the ext.

The low position of the definite but non-pivot int can be seen in clauses whose verbs surface in lv. In that context, the goal triggers absolutive agreement and raises to the position of the pivot. This step can be shown with the familiar diagnostics for a-movement. For instance, when the goal is quantified, it can bind into the ext (38).

Given that the int does not raise to the position of the pivot in lv, it is possible to observe the influence of definiteness-related Object Shift in isolation. If this process placed the int above the ext, then we would expect the int to be able to bind into the ext, even though it would not raise further to the position of the pivot, as in (39).

In Mandar, this is impossible. This fact can be seen in example (40), which is an lv clause that contains a quantified non-pivot int. That argument, the dp nasang...buku “every book,” is unable to bind the variable in the ext.

This pattern holds despite the fact that universally quantified dps must undergo Object Shift in the language, as evidenced by the fact that they cannot appear in the position of the int in av (41). It suggests that the process of definiteness-related Object Shift does not place the int above the ext, as in the tree in (36).

This fact can be replicated in every context that allows the int to be definite without raising to the position of the pivot. It can also be shown, for instance, in the causative. Mandar has a causative construction that is built with the v⁰_caus pa- (42a). When a causative verb surfaces in pv, it is the causee that raises to the position of the pivot. The movement of that argument can be shown with the familiar set of diagnostics. When it is quantified, for instance, it can bind into the ext (42b).

As the int is not the pivot in this context, it is possible to test the influence of Object Shift in isolation once again. If this process placed the int above the ext, then we would expect the int to be able to bind into the ext, as in (43a). It cannot (43b).

3.4 The role of agreement

The facts of trivalent clauses show that Object Shift does not place the int above the ext. It follows, then, that there must be a second process in the language which attracts this argument to an higher position in the contexts where it is the pivot. The natural candidate for such a process is a step of movement linked to Case-Licensing by a higher functional head. In Mandar, as we have seen, the pivot invariably triggers absolutive agreement. I propose that this agreement arises as the result of a Case-Licensing relationship between the pivot and t⁰. Endowing t⁰ with the feature [$+$ epp], we can derive the observed correlation between pivot status, agreement, and position: the pivot undergoes a step of covert movement to the highest a-position in the clause because it triggers agreement on t⁰ and is Case-Licensed by that head. This analysis is shown in (44).

The analysis in (44) follows the logic of a broader approach which takes the pivot to raise to a high a-position for reasons of Case-Licensing in Philippine-type languages and Malagasy (Guilfoyle et al. 1992; Ting 2022) and in many languages of Western Indonesia (Aldridge 2008; Erlewine 2018). An important component of this analysis is the claim, first made by Hung 1988 for Malagasy, that the head which Case-Licenses the pivot sits above the class of heads that host voice morphology (and on the present proposal, trigger Object Shift).

Within the generative literature on the languages of South Sulawesi, there is disagreement over this point. In Selayarese, Béjar 1999 asserts that the pivot is Case-Licensed by t⁰ and that the voice morphology sits in a lower v⁰, while Finer 1997, 1998 proposes that the pivot is Case-Licensed by an abs⁰ that sits below an erg⁰ that hosts the morphology of pv. In Mandar, it is clear that the first perspective is correct. The key evidence for this view comes from the distribution of absolutive agreement. There are a number of constructions in the language in which this agreement does not occur, and these have the distribution of non-finite clauses in English. Some are selected by c⁰ (45a), others appear bare (45b), and yet others are selected by control verbs (46a) and raising predicates (46b). I assume that these constructions lack absolutive agreement because they do not form finite tps.

While the non-finite constructions in (45)-(46) show no absolutive agreement, they are able to contain the morphology of av (45a) and pv (45b). This fact suggests that the distribution of absolutive agreement is linked to finiteness in a way that the distribution of voice morphology is not. This step, in turn, suggests that the head responsible for Absolutive Case-Licensing sits above the heads that host voice morphology (see Legate 2006 for a similar point on absolutive case morphology in other High Absolutive languages).

With this conclusion secure, we can turn to the identity of the head that hosts absolutive agreement. Two patterns show that it is t⁰. The first involves portmanteau formation with asp⁰. There are three second-position clitics in Mandar that mark aspect: mo ‘already,’ pa ‘yet,’ and da ‘just.’ These clitics surface in their canonical forms in clauses which lack absolutive agreement, such as the existential construction in (47a). But when they appear alongside absolutive agreement, they surface in forms that cannot be derived from the regular phonology of the language. The clitic mo, for instance, surfaces as mi alongside the third-person i and ma’ with the first-person a’ (47b).

The descriptive literature considers clitics like ma’ and mi to be portmanteaux in the other languages of the South Sulawesi subfamily, (Campbell 1989; Friberg 1996; Jukes 2006), and adopting this conclusion for Mandar, we arrive at an argument for the position of absolutive agreement. Like selection and contextual allomorphy, the formation or insertion of portmanteaux is subject to meaningful constraints on locality (Noyer 1992; Williams 2003; Trommer 2010; Bye and Svenonius 2010; Merchant 2015; Svenonius 2016; Woolford 2016). As a result, the presence of portmanteaux that expone absolutive agreement and aspect suggest that the head which hosts absolutive agreement must sit close to asp⁰ in the syntax.

We can pin down the position of this head with a parallel observation from the opposite direction. As in the other languages of the South Sulawesi subfamily, absolutive agreement participates in a pattern of allomorphy in Mandar that is conditioned by c⁰ (Valkama 1995a; Friberg 1996; Jukes 2006). Beneath the complementizer anna’, it surfaces in a distinct irrealis paradigm. This pattern is shown in the song lyric in example (48), in which first-person absolutive agreement is realized as the suffix -u rather than the typical enclitic a’.

This pattern of allomorphy suggests that absolutive agreement is hosted by a head that is syntactically local to c⁰. From this result, we can conclude that this agreement must sit in a head that stands between asp⁰ and c⁰. The natural candidate for a head in this position is t⁰. And endowing this head with the feature [$+$ epp], we arrive at a theory on which the pivot reaches its final position as the result of two distinct steps in pv and lv: one step of object shift that takes it out of the vp but positions it beneath the ext, and then a second step, linked to this pattern of agreement, that places it in an a-position above all other arguments in the clause. Taking this head to assign abstract Absolutive Case, this analysis comes to resemble that which has been proposed for ergative languages which show High Absolutive syntax (Bittner and Hale 1996b, a; Legate 2006; Coon et al. 2014; Brown 2016; Ershova 2019). As a result, I refer to it as the Two-Step Theory of High Absolutive Syntax. It is shown in tree (49).

3.5 Activity and case-licensing

With these pieces in place, we can now turn to the question of locality: if Object Shift does not place the int above the ext, how is it possible for the int to interact with t⁰ and raise to spec,tp in pv? I propose that the answer lies in the notion of Activity (Chomsky 2001). Following much work on classical Case Theory, I assume that all nominals are subject to a requirement for Case-Licensing that can be satisfied in the syntax by the establishment of an Agree relation with a c-commanding functional head (Vergnaud 1977; Chomsky 1980, 1981, 2001; Schütze 1993). Taking morphological agreement as evidence for Case-Licensing (Raposo 1987), we arrive at a picture of Case-Licensing that is familiar from work on High Absolutive languages further afield (Bittner and Hale 1996b, a; Coon et al. 2014): one on which the absolutive argument interacts with t⁰ and raises to spec,tp because it is the highest Active argument in the voice p by the time that t⁰ probes.

In pv, there are two instances of agreement: t⁰ agrees with the int and voice⁰_pv agrees with the ext. In this context, I argue that the int is Case-Licensed by t⁰ and the ext is Case-Licensed by voice⁰_pv. Representing Case-Licensing with a dotted arrow, the resultant state of affairs is shown in tree (50).

With these patterns of Case-Licensing in place, the puzzle of locality can be resolved. As the ext is Case-Licensed within the voice p, it is Inactive by the time that t⁰ is merged. At the derivational stage where t⁰ probes its c-command domain, the highest unlicensed argument in the voice p is the int. Assuming that only unlicensed arguments are visible to Case-Assigning probes, or Active in the sense of Chomsky 2001, it follows that the probe on t⁰ will not be able to target the ext, even though it is highest in the voice p. It follows as a result that this probe will be able to interact with the int and attract it to spec,tp without incurring a violation of Locality (Guilfoyle et al. 1992; Ershova 2019).

Generalizing across the other voice frames in the language, I propose that the same logic guides the choice of the pivot throughout the language. On this view, the pivot is invariably the highest unlicensed argument in the voice p at the derivational stage where t⁰ probes. Given the logic of Activity, the result is the following: no matter its intermediate position in the voice p, the pivot will invariably enter into an agreement relationship with t⁰, be Case-Licensed by that head, and then undergo a step of a-movement to spec,tp.

The following tree shows how this analysis derives the syntax of pv. Starting at the level of the v p, we can see that the int undergoes a step of Object Shift to spec, v p, which is triggered by the feature [$+$ epp] on the v⁰_pv. This process places the int beneath the base position of the ext. To allow this argument to interact with t⁰ and raise to spec,tp, it is important that it remain unlicensed until the derivational stage where t⁰ is merged. As a result, I propose that in pv the int is not Case-Licensed in the v p: in other words, the v⁰_pv is unable to assign structural Accusative Case (Guilfoyle et al. 1992; Bittner and Hale 1996b, a). Moving up to the voice p, the voice⁰_pv bears a phi-probe and searches its c-command domain for the highest unlicensed argument therein. Finding the ext, it triggers ergative agreement and Case-Licenses that argument in-situ. This step of agreement renders the ext Inactive. Finally, once t⁰ is merged, it probes its c-command domain and finds the int as the highest Active argument in the voice p. It triggers absolutive agreement with that argument, Case-Licenses it, and triggers a step of covert movement to spec,tp. The surface word order, which is v-ext-int, arises as the result of a process of head-movement that places the verb in voice⁰.

The following tree shows how the analysis derives the syntax of av. Starting at the level of the v p, we can see that the int does not undergo Object Shift to spec, v p. This is because the v⁰_av does not bear the feature [$+$ epp]. Nevertheless, I propose that the v⁰_av is able to Case-License the int. This move has two effects. First, it allows the ext to interact with t⁰ and raise to spec,tp without yielding a problem of Case-Licensing within the v p. Second, it provides the means to capture the split that we have seen in the system of av: the affixes that allow for the presence of an int contain a v⁰_av which assigns abstract Accusative Case (31). Moving up to the voice p, the voice⁰_av does not bear a phi-probe, and as a result, it is unable to Case-License the ext. This failure to license the ext renders that argument the highest Active argument in the voice p. Once t⁰ is merged, it probes its c-command domain, finds the ext, triggers absolutive agreement with that argument, Case-Licenses it, and triggers a step of covert movement to spec,tp. The word order, again v-ext-int, arises again as the result of head-movement of the verb to voice⁰.

Finally, the tree in (53) shows the syntax of lv. Starting at the level of the v p, we can see that there is a goal that is introduced above the base position of the int by v⁰_appl. I assume that v⁰_appl is able to Case-License the int with structural Accusative Case (Adger and Harbour 2007). The goal then undergoes a step of Object Shift to spec, v p, which is triggered by the feature [$+$ epp] on the v⁰_pv; in clauses where both the goal and the int are definite, I assume that the int undergoes Object Shift as well and tucks in to a lower specifier of v⁰ (Rackowski 2002). The process of Object Shift places the goal beneath the base position of the ext, and as a result, the goal must remain unlicensed until the derivational stage where t⁰ is merged. This is made possible by the fact that the v⁰_pv does not assign it structural Accusative Case. Moving up to the voice p, the voice⁰_pv once again Case-Licenses the ext, rendering it Inactive, and by the time that t⁰ probes, the highest Active argument in the voice p is the goal. t⁰ triggers absolutive agreement with this argument, Case-Licenses it, and triggers a step of covert movement to spec,tp.

On this analysis, the syntactic and distributional differences between av, pv, and lv follow from three syntactic parameters: the possibility for Object Shift, the availability of abstract Accusative Case, and the availability of abstract Ergative Case. These differences, in turn, follow from the featural content of the functional heads that make up the voice system: v⁰_pv can trigger Object Shift while v⁰_av cannot, v⁰_appl and v⁰_av can assign abstract Accusative Case but v⁰_pv cannot, and voice⁰_pv can assign abstract Ergative Case but voice⁰_av cannot. Their effect is to determine the possibility of a definite int and to position a single argument to raise to the position of the pivot, by rendering it the highest unlicensed argument in the voice p. The properties of this analysis are summarized in table (3.5).

4 The agent focus construction

The Two-Step Theory of the Mandar voice system builds on previous approaches of the voice systems of other languages of Western Austronesia, in that it adopts an Argument Structure Analysis of the voice system (Aldridge 2004), takes the pivot to raise to an a-position as the result of a Case-Licensing relationship with t⁰ (Guilfoyle et al. 1992; Rackowski 2002), and links the alternation between av and pv to the need for Object Shift (Rackowski 2002; Aldridge 2004). It is a first success of this theory, then, that it is able to reduce the syntax of the unfamiliar voice system of the language to the sum of a set of analytical pieces developed for languages which are morphologically distinct from those of South Sulawesi. And this success, in turn, reinforces the value of those pieces to our broader understanding of the typology of the voice systems of the region.

The principal value of the Two-Step Theory, however, lies in the solution that it provides to a puzzle that arises in contexts where the demands of Case-Licensing, Object Shift, and Extraction conflict. This is one which emerges in a configuration that has been the subject of some study in other languages of Western Austronesia, and in particular, Tagalog (Adams and Manaster-Ramer 1988; Rackowski 2002; Latrouite 2012; Sabbagh 2016) (see also Erlewine 2016a). As we have seen, Mandar imposes a Subjects-Only Extraction Constraint on wh-movement, focus-fronting, relativization, and a-raising. A common and unsurprising result of this constraint is that the use of pv is ruled out in clauses that launch movement of the ext. When the int is definite in this context, however, the Two-Step Theory yields a set of predictions that are not shared with other approaches to the voice system. First, it predicts that it the int must undergo Object Shift, necessitating the use of a v⁰ that bears the feature [$+$ epp]. Second, it predicts that the int must receive abstract Accusative Case, as it cannot be Case-Licensed by t⁰. On the system developed here, this pair of requirements cannot be resolved simultaneously by either the v⁰_pv or the v⁰_av: the v⁰_pv triggers Object Shift but cannot Case-License the int, while the v⁰_av can Case-License the int but cannot trigger Object Shift. As a result, the Two-Step Theory predicts that the language must employ a construction which recruits the voice⁰_av, to allow the ext to raise to the position of the pivot, but nevertheless employs a distinct v⁰ which can trigger Object Shift and Case-License the int in the v p.

When we turn to the relevant configuration in Mandar, we find that this is exactly what occurs. In clauses where the ext is extracted and the int is definite, the language employs a verbal form whose morphology is superficially similar to that of av. When we look past this morphological parallel, however, we will find that it is syntactically distinct. The shape of this construction is shown in (55).

The appearance of a construction with these properties in this context in Mandar is part of a broader pattern in the Western Austronesian region, and one which extends to High Absolutive systems outside of this area. From a morphological perspective, the construction in (55) has exact parallels in the languages of the South Sulawesi Subfamily and their neighbors, which all employ a verbal form that recruits av-like morphology and shows a pattern of accusative agreement with the int in this context (Martens 1988; Friberg 1991; Mead 1998, 2002; Zobel 2002; Jukes 2006). Abstracting away from the presence of agreement, many Philippine-type languages have also long been recognized to employ a very similar construction in the same context: one which bears av-like morphology but allows for patterns of differential object marking (dom) on the int which are much more constrained in finite matrix av clauses (Adams and Manaster-Ramer 1988; Rackowski 2002; Latrouite 2012; Sabbagh 2016; Erlewine 2016a). Tagalog, for instance, employs a construction in this context which allows the int to be marked with the dom marker sa (56).

Setting our sights further afield, a morphologically similar verbal form has been noted to appear in a very similar context in Mayan languages of the Q’anjob’alan subfamily. This verbal form is known as the Agent Focus, and much like the Mandar construction in (55), it is built with the morphology that appears in clauses where the ext is absolutive and it implicates an exceptional pattern of agreement with the int (Smith-Stark 1978; Larsen and Norman 1979; Dayley 1981; Ordóñez 1995; Stiebels 2006; Coon et al. 2014; Erlewine 2016b; Aissen 2017; Coon et al. 2021). As we will see, the Two-Step Theory predicts- and the facts of the language suggest- that the Mandar construction in (55) shows a syntax which is similar to that which has been proposed for the Agent Focus in a particular strand of research on the languages of that subfamily (Coon et al. 2014, 2021). To underscore this parallel, I will refer to the Mandar form as Agent Focus as well.

4.1 Extraction, object shift, and case-licensing

The starting point of this investigation is the Subjects-Only Extraction Constraint. In Mandar, we have seen that this rules out wh-movement, focus-fronting, and relativization of all arguments except the pivot. The result of this constraint is that it is impossible to extract the pv ext, the av int, the lv ext, the lv int, the applied argument in cv, and most obliques. The constraint is shown with the pv ext in (57).

Following Aldridge (2004), I take this constraint to follow from a restriction that holds over the ${\bar{\textsc {a}}}$-probe that drives these operations: it targets only the closest dp in its c-command domain. This proposal continues the logic of the standard theory of the Ergative Extraction Constraint in High Absolutive languages (Campana 1992; Murasugi 1992; Ordóñez 1995; Bittner and Hale 1996b, a), and it is widely adopted in work in Austronesian and beyond (Coon et al. 2014, 2021; Erlewine and Lim 2022; Branan and Erlewine 2022). Its basic result is schematized in (58).

The result of this restriction is that it is not possible to employ pv in clauses where the ext undergoes extraction. In clauses of that type, rather, we expect that the verb must merge with the voice⁰_av. This is because the raising of the ext to spec,tp is forced by a property of that head. As we have seen, the voice⁰_av fails to Case-License the ext (52). This renders it the highest unlicensed argument within the voice p, and as a result, the argument that is targeted by t⁰ at the derivational stage where it probes.

This constraint raises two theoretical challenges in contexts where the int is definite. In that context, it is obligatory for the int to undergo Object Shift. As a result, we predict that the language must employ a v⁰ that bears the feature [$+$ epp], as in (84).

A priori, we expect this requirement to rule out the use of the v⁰_av in contexts where the ext is extracted and the int is definite. This is because that v⁰ does not bear the feature [$+$ epp ] and, as a result, it is unable to trigger Object Shift. In clauses that contain this v⁰, moreover, there is no way for the int to escape the vp by interacting with higher probes in the clause, such as t⁰. This follows from the pattern of selection that was established in Sect. 3.1: the v⁰_av is invariably selected by the voice⁰_av, and this head does not Case-License the ext. As a result, in clauses that contain the v⁰_av, the ext will always serve as the highest Active argument in the voice p, and in the resultant configuration, it is not possible for a probe outside of the voice p to interact with the int. The resultant problem of Object Shift is shown in (61).

This logic leads us to predict that the v⁰_av cannot be employed in contexts where the ext is extracted and the int is definite. As a possible resolution, we might imagine that the language should turn to the v⁰_pv instead, as this head bears the feature [$+$ epp] and would therefore resolve the requirement of Object Shift. But looking ahead in the derivation, it quickly becomes clear that the result will fare no better. This is obvious in a derivation in which the v⁰_pv is selected by the voice⁰_pv, which fails to set up the locality-compliant configuration in (58). This is because the voice⁰_pv Case-Licenses the ext and renders it Inactive, forcing t⁰ to attract the int to spec,tp. In this configuration, it is impossible for the ${\bar{\textsc {a}}}$-probe on c⁰ to attract the ext. The resultant problem of Locality is shown in (62).

Forcing this derivation into compliance with the Subjects-Only Extraction Constraint, moreover, creates a separate problem in the domain of Case-Licensing. To see why, consider a parallel derivation in which the v⁰_pv is exceptionally selected by the voice⁰_av. This derivation would avoid the problem of locality, as in the presence of the voice⁰_av, the ext would not be Case-Licensed in the voice p and would thus interact with t⁰ and raise to a position from which it could interact with c⁰. But in allowing the ext to interact with t⁰, this derivation would render it impossible for that head to interact with and Case-License the int. In other words, introducing the head that allows extraction of the ext eliminates the canonical pathway to Case-Licensing that rescues the int when it must undergo Object Shift. The resultant problem of Case-Licensing is shown in (63).

4.2 The agent focus construction

The Two-Step Theory, then, delivers a clear prediction about the syntax of clauses in which the ext is extracted and the int is definite. When the ext is to be extracted, the language must employ the voice⁰_av, as this is the head that positions the ext to interact with t⁰, raise to spec,tp, and arrive at the requisite proximity to the probe on c⁰. To allow for the presence of a definite int in that context, moreover, the analysis predicts that the language must employ a v⁰ that does two things. First, it must allow the int to undergo Object Shift to leave the vp in the absence of a relationship with t⁰, unlike the v⁰_av. Second, it must allow the int to be Case-Licensed in the absence of a relationship with t⁰, unlike the v⁰_pv. The v⁰ which is required in this context is one which bears the feature [$+$ epp ] and assigns abstract Accusative Case. Its predicted behavior is presented in (64).

These predictions are borne out. In contexts where the ext is extracted and the int is definite, Mandar employs a distinct verbal form which is shown in (65). This is one which has two special properties: it hosts morphology which is similar to that of av and shows dom with the int in the form of accusative agreement.

The morphological shape of this construction is presented in (66). It is built from two components: a prefix maN- (recall that coda nasals denasalize before all non-nasal segments but ) and an accusative agreement enclitic that follows the verb. Abstracting away from the morphological shape of the prefix, a nearly identical form is employed in the same context in the other languages of the South Sulawesi subfamily (Campbell 1989; Friberg 1991, 1996; Strømme 1994; Matti 1994; Valkama 1995a, b; Jukes 2006; Laskowske 2016; Brodkin 2021b) and in many languages of the Kaili-Pamona and Bungku-Tolaki subfamilies nearby (Martens 1988; Mead 1998, 2002; Zobel 2002).

The af prefix maN- is segmentally identical to one of the ten affixes that mark av in Mandar, and the accusative agreement enclitics are segmentally identical to the absolutive agreement enclitics that sit in t⁰. But despite these phonological similarities, it is clear that the af prefix and the accusative enclitics are morphosyntactically distinct from their homophonous counterparts. The accusative enclitics, to begin, surface in a different linear position from the absolutive enclitics. This fact can be seen in clauses that contain preverbal material, such as clausal negation. The absolutive enclitics strictly follow the leftmost phonological phrase in the intonational phrase, as part of their general second-position behavior (67a), and I assume that they reach this position through a process of phonologically-driven postposing from the left edge of the intonational phrase (Halpern 1995; Harizanov 2014; Bennett et al. 2016; Brodkin 2021c). The accusative enclitics, however, are strictly postverbal (67b). This suggests that they do not sit in t⁰.

Convergent evidence for this analysis comes from the interaction with aspect. As we have seen, the absolutive agreement enclitics form portmanteaux with aspect, and this pattern provides evidence that they sit in t⁰ (47b). The accusative agreement enclitics, however, do not enter into portmanteaux with aspect in the same way. In clauses where prosodic constraints should force the aspectual clitics to follow a verb in af, it is impossible for portmanteaux to appear (68). This fact provides a second argument that the accusative agreement enclitics do not sit in t⁰.

Tightening the screws further, we can observe that the presence of accusative agreement is contingent on the presence of the particular affix maN-. This agreement cannot appear in the presence of any other voice prefix in the language, such as the cv prefix si- or canonical av affixes like ma’-, me-, and me’- (69). As a result, many verbs show a morphological difference between av and af. In contexts where the af form is required, the verbs which canonically take av prefixes like me- surface with the prefix maN- (69c).

This pattern suggests that accusative agreement originates in a head that surfaces in the af portmanteau maN- but not the av affixes ma’-, me-, and me’- (69). The natural candidate for this head is v⁰. As the af portmanteau is homophonous with one of the av prefixes in the language, I propose that it can be syntactically decomposed into a combination of voice⁰ and v⁰ as well. And as the af prefx maN- shares the initial segment m- with the other av prefixes in the language, I assume that it contains the voice⁰_av. This morphological conclusion dovetails with one of the syntactic predictions of the Two-Step Theory that was established above: the af construction must contain the voice⁰_av in order to position the ext to raise to the position of the pivot and then extract. Pressing further, it also allows us to identify the head that is unique in the af construction: it is a v⁰, selected by voice⁰_av, which hosts accusative agreement and contributes all the segmental content of the af prefix maN- beyond the segment m-. I will refer to this v⁰ as the v⁰_af.

4.3 The agent focus v ⁰

With the existence of the v⁰_af established, we can now turn to the syntax of af. As we have seen, the Two-Step Theory predicts that the int must be able to undergo Object Shift and must be Case-Licensed in the v p in the contexts where it is definite and the ext is extracted. I propose that the v⁰_af allows both of these things to occur. On this analysis, this v⁰ bears the feature [$+$ epp ] and assigns structural Accusative Case, as in (70).

Straightforward evidence for accusative Case-Licensing comes from the pattern of agreement. As we have seen, in the af construction the int triggers agreement in the form of the accusative enclitic on the verb (66). This agreement can be linked to the v⁰_af. On the assumption that overt agreement provides evidence for Case-Licensing (Raposo 1987), this pattern suggests that the v⁰_af does Case-License the int.

Pressing further, we can find evidence for the occurrence of Object Shift in the distribution of a prosodic process that targets only constituents in the vp. This is Prosodic Incorporation, a process which targets focused constituents and groups them into a single minimal phonological phrase with the verb. This process has been observed in descriptive work on several other languages of the South Sulawesi subfamily (under the name of ‘Incorporation’; Campbell 1989; Laskowske 2016), and in Mandar, its occurrence can be detected in two ways. First, it forces the deletion of a boundary tone which normally follows the verb and allows the verb to undergo various processes of reduction that are blocked at the right edge of the phonological phrase. Second, it has an effect on the linear position of second position clitics. The absolutive enclitics generally follow the verb in verb-initial clauses, but when a constituent undergoes Prosodic Incorporation, the absolutive clitics follow the sequence of verb and incorporand. This makes it possible to observe the application of this process in (71b), where it targets the adverb dini “here.” I will represent its application with a hyphen.

The process of Prosodic Incorporation can target many types of phrasal constituents in the vp, such as manner adverbs (72a) and locative pps (72b). But its application is subject to a syntactic constraint: it cannot target material outside of the vp. As a result, cannot target tp-level adjuncts (73a) or the non-pivot ext (73b).

This constraint provides the means to diagnose the position of the int in af. In clauses which show av, it is possible for the int to undergo Prosodic Incorporation. This result is expected on the Two-Step Analysis, as the v⁰_av does not trigger Object Shift and thus allows the int to remain in the vp. It can be seen in matrix clauses, such as (74a), and also in embedded contexts. One context of that type is a headless relative construction that surfaces beneath the existential predicate diang (74b).

In af, we find the reverse. In relative clauses whose verbs surface in af, it is impossible for the int to undergo Prosodic Incorporation. This constraint is shown in the examples in (75), which contains another headless relative clause beneath the existential predicate diang. That relative clause contains an int that is a dp and a gap in the position of the ext. The int must follow the accusative agreement enclitic (75a). It cannot surface between the clitic and the verb (75b).

This pattern does not follow from a general constraint against Prosodic Incorporation in af, because it is possible for constituents other than the int to undergo Prosodic Incorporation in the context above. This can be seen in example (76), which contains a similar headless relative that contains a verb in af. In that context, the focused vp-internal directional adverb dai’ can undergo Prosodic Incorporation and split the verb from accusative agreement.

These observations suggest that the int does indeed move out of the vp in the af construction. This result is striking for two reasons. First, it provides explicit positional evidence for a process of Object Shift in the language, therefore supplementing the theoretical logic that led us to postulate this process on the basis of the Definiteness Effect. And second, it provides confirmation for the second prediction of the Two-Step Analysis: when the ext is extracted and the int is definite, the language employs a construction that allows the int to be Case-Licensed and undergo Object Shift.

Connecting these patterns to the v⁰_af, we conclude that this head has the ability to assign abstract Accusative Case and trigger Object Shift, and in this respect, it is distinct from both the v⁰_av and v⁰_pv. The features of the v⁰_af, and the differences between these v⁰s, are tabled in (4.3).

4.4 Agent focus and extraction

With its v p-level syntax in place, we can now see how the af construction provides the means to resolve the tension that arises when the demands of the systems of Extraction, Object Shift, and Case-Licensing conflict. I propose that af clauses like (77) show the syntax in (78). In this construction, the int undergoes a step of Object Shift to spec, v p, on a par with clauses that show pv (51). Unlike those clauses, however, it is the ext which proceeds to interact with t⁰ and raise to spec,tp. As a result, that argument can be extracted in compliance with the Subjects-Only Extraction Constraint.

On this analysis, the af construction provides the means for the ext to undergo ${\bar{\textsc {a}}}$-extraction in a manner that is fully compliant with the Subjects-Only Extraction Constraint. Despite the fact that the int moves out of the vp and triggers a form of agreement that is similar to the agreement in t⁰, I argue that it does not move above the ext at any point in this derivation. This claim finds support from several different patterns in the language, and I will briefly review the most important of those here.

The first is an argument from Condition C. So far, we have seen when the int is the pivot, it cannot be a pronoun that is coindexed with an r-expression in the ext (18). The same is not true of af. This is shown in (79), which is a clause where the verb surfaces in af and the int is a null pronoun. The ext, which is focused, contains a possessor which is an r-expression. Here, it is possible for the int to be coindexed with the r-expression in the ext, despite the fact that this cannot occur when the int is the pivot (19). On the assumption that ${\bar{\textsc {a}}}$-extraction of the ext should reconstruct for Condition C, this suggests that the int does not raise above the position from which the ext is extracted in clauses that show af.

A similar argument can be made from the facts of variable binding. As we have seen, the int can bind variables in the ext when it is the pivot (27). In the af construction, however, it cannot. This pattern is shown in example (80): when the int is quantified in af, it cannot bind a variable inside of an extracted ext.

To this pair of diagnostics we can add a final fact about linear order. Although the pivot does not occupy a consistent linear position in Mandar, it is often able to undergo an operation of leftward preposing that is not available to other arguments in the clause. In clauses that contain a preverbal auxiliary or negator, the pivot can optionally surface after that element and before the verb. This pattern is shown in example (81), which is a clause that contains preverbal negation. The pivot of that clause, which is the demonstrative di’o “that,” surfaces after negation and before the verb. I refer to the process that places the pivot in this position as Pivot Preposing.

Pivot preposing targets only the pivot. This is shown in the pv clause in (82), which contains a preverbal auxiliary. When a dp surfaces after this auxiliary, it can only be interpreted as the int, not the non-pivot ext. As a result, the process yields a diagnostic for pivot status: if an argument cannot be preposed, then it is not the pivot.

Taking this diagnostic to af, we find that it is impossible in that construction for the int to undergo preposing (83). In af clauses that contain a preverbal auxiliary, the int must surface postverbally (83a). It cannot surface after the auxiliary (83b).

What these observations collectively suggest is that the af construction does not allow the int to raise to the position of the pivot. And with this conclusion in tow, we can be reasonably certain that in af clauses we are dealing with the syntax in (78): one on which the int undergoes a step of Object Shift to spec, v p and the ext raises to spec,tp.

From this position, we can now turn to a derivation. Returning to the example in (77), consider the v p in (84). This v p contains two arguments: an ext which is a wh-word and must undergo extraction, and an int which is definite and must undergo Object Shift. Given the requirement for Object Shift, the language must employ a v⁰ that bears the feature [$+$ epp] to allow the step of movement in (84). This rules out the appearance of the v⁰_av and forces the presence of the v⁰_pv or the v⁰_af.

As we have seen, there is no way for the derivation in (84) to converge in the presence of the v⁰_pv. If the v⁰_pv is merged and selected by the voice⁰_pv, we arrive at a problem of Locality, as the voice⁰_pv Case-Licenses the ext, renders it Inactive, and delivers a situation where the int must interact with t⁰ and raise to spec,tp (62). In an analogous derivation where the v⁰_pv is merged and then selected by the voice⁰_av, moreover, we run into a separate problem of Case-Licensing: as the voice⁰_av does not Case-License the ext, that argument must interact with t⁰, and this deprives the int of the typical pathway to Case-Licensing that is available in clauses that contain the v⁰_pv (63).

With this background established, we can now see how the af construction resolves the tension between the systems of Extraction, Object Shift, and Case-Licensing. When the v⁰_af is merged in the derivation in (84), the three problems above do not arise. To begin, the v⁰_af bears the feature [$+$ epp] and therefore resolves the problem of Object Shift, unlike the v⁰_av. Moreover, the v⁰_af is able to assign abstract Accusative Case to the int and therefore resolves the problem of Case-Licensing that emerges in contexts where the int is unable to interact with t⁰, unlike the v⁰_pv (63). Finally, as a result of this property, it is possible for the v⁰_af to be selected by the voice⁰_av, which fails to Case-License the ext and thus allows it to interact with t⁰ and raise to spec,tp, resolving the problem of Locality. The resultant derivation, repeated from (64)-(78), is presented in (85).

4.5 Agent focus as a last resort

These results cast the function of the af construction into sharp relief: it provides the means to trigger Object Shift and Case-License the int in a different way when the use of pv is ruled out by the demands of the system of extraction.

Seen from this perspective, a generalization emerges around the distribution of the v⁰_af: this head is merged exclusively in the set of contexts where the int must undergo Object Shift and cannot be Case- Licensed by any higher head. This description is met plainly in the ${\bar{\textsc {a}}}$-contexts above, where the Subjects-Only Extraction constraint demands that the ext be the pivot in every clause where it extracts. But it is also met in a range of other contexts in the language that do not involve ${\bar{\textsc {a}}}$-extraction, and in those too, the v⁰_af can be found. To provide a single example from the a-domain, this v⁰ appears equally in contexts of Raising. Mandar has a process of Raising to Pivot that targets the pivot of a non-finite clause. An example is shown in (86). Here, the dp iKaco’ surfaces after the non-thematic matrix predicate masae “long (of time).” It is thematically connected to the predicate of the following clause, which is non-finite and contains a gap in the position of the pivot.

I assume that Raising to Pivot shows a similar syntax to Raising to Subject in English. In clauses that show Raising, the highest unlicensed argument of a non-finite clause undergoes a step of movement into a higher clause as the result of an agreement relationship that it establishes with a Case-Licensing head therein. In (86), this Case-Licensing head is the matrix t⁰. Setting word order aside, I take this clause to show the syntax in (87).

Like wh-movement, focus-fronting, and relativization, the process of Raising to Pivot is subject to a strict constraint on locality. It is possible for this process to target the ext of an embedded clause when the embedded verb surfaces in av, as in (88a). But it is impossible for it to target the ext of an embedded clause whose verb surfaces in pv (88b).

Raising to Pivot, therefore, provides the means to set up another configuration in which the use of pv is taken off the table. And when the ext undergoes this process in a clause that contains a definite int, the embedded verb surfaces in af (89).

The embedded clause in (89) shows the syntax in (90), on which the ext serves as the highest unlicensed argument in the voice p and the int undergoes Object Shift and is Case-Licensed in the v p. Both of these processes are triggered by the v⁰_af.

The broader distributional profile of the v⁰_af, then, suggests that it is not a head whose distribution depends in an intimate way on the presence of ${\bar{\textsc {a}}}$-extraction (as has been proposed of the morphosyntactically similar constructions in Mayan; Deal 2016; Coon et al. 2021). Rather, its appearance in contexts like (89) suggests that it has the distribution of a Last Resort (Safir 1993; Bošković 1997; Chomsky 2001). Specifically, I propose that it is a Secondary Licensor, in the terminology of Levin and Massam 1985. On this view, the v⁰_af is a head which is merged exclusively in the contexts in which the int must undergo Object Shift but cannot be licensed in the typical High-Absolutive way of clauses where this occurs. In this respect, it is similar to the heads that are merged to rescue other types of arguments that would otherwise go without Case-Licensing, such as those which assign abstract Ergative Case (Bobaljik 1993), those which surface in contexts of Differential Object Marking (Kalin 2018), and those which repair violations of constraints like the pcc (Rezac 2011).

What is striking about this result is that, on the present analysis, it allows the Mandar v⁰_af to be understood as part of a broader class of Secondary Licensors that are employed in a consistent context in the ergative languages that show High Absolutive Syntax. There is a tradition of work on certain languages of the Mayan family, including those of the Q’anjob’alan subgroup, which takes them to be syntactically similar to Mandar, on the Two-Step Theory, in two respects (Coon et al. 2014, 2021; Royer et al. 2021; Royer 2022, to appear; Brodkin and Royer to appear): first, they require the absolutive argument to raise to the a-position in the clause as the result of a Case-Licensing relationship with t⁰, and second, they impose a locality constraint in the ${\bar{\textsc {a}}}$-domain that prohibits the extraction of the ext in clauses of that type. This perspective opens up the possibility for a common prediction in the contexts where Mandar employs the af construction, and in many languages of the Q’anjob’alan subfamily, this prediction is borne out. In the contexts where the ext is extracted and the int is a certain type of nominal, many languages of that subfamily employ a construction which shows the morphology of clauses in which the ext is absolutive but shows an exceptional pattern of agreement with the int. Coon et al. (2014) argue that this construction shows a syntax that is strikingly similar to that of the Mandar af: in Q’anjob’al, one language of this type, they argue that it provides the means to Case-License the int when it would be otherwise unable to receive abstract Case.

Seen from this perspective, a parallelism emerges between the v⁰_af in Mandar and the head that hosts the af affixes in the languages of the Q’anjob’alan subfamily: in both cases, it provides a source of abstract Accusative Case for the int in the contexts where the canonical pathway of High Absolutive Case-Licensing breaks down. And from this parallel, in turn, a number of deeper symmetries begin to unfold between the voice system of Mandar and the systems of alignment that recur in the languages of that area. It is a final success of the Two-Step Theory, then, that it leads us to connections between languages and constructions that, from different and earlier perspectives may have seemed to be fundamentally unalike.

5 Conclusion

This paper has investigated the clausal syntax of Mandar, a language of South Sulawesi, and shown that it can be readily understood through the combination of a set of theoretical proposals that have been developed in the ergative tradition of analysis of Philippine-type languages. In this language, I have argued that the alternation between av and pv turns on the need for definiteness-related Object Shift out of the vp (Rackowski 2002; Aldridge 2004) and shown that the broader voice system serves to position a single argument to be Case-Licensed by t⁰ and thus raise to the highest a-position in the clause (Guilfoyle et al. 1992). The theory that emerges from the combination of these pieces, the Two-Step Theory of High Absolutive Syntax, makes it possible to understand a broad set of patterns in the language, some of which are familiar from other languages in the region and others which are not. And at the same time, this theory makes it possible to understand a pattern that emerges in contexts where the systems of Extraction, Case-Licensing, and Object Shift conflict, where Mandar employs a construction that allows the int to undergo Object Shift and be Case-Licensed in the v p, in a break from the canonical pattern of High Absolutive Case-Licensing.

Looking beyond this particular language, much work is required to understand the ways in which other languages of this region respond to the particular tension that we have outlined and investigated above. But it is a promising first success of the Two-Step Theory that it provides the means to understand the response that surfaces in Mandar, and a striking result that it is ultimately similar to that which emerges in a similar context in certain High Absolutive languages further afield.