2.1 Where Accusative Cannot Appear

While accusative may appear on a wide variety of subjects, there are three types which may not take this case: matrix subjects, the embedded subjects of clausal subjects, and the subjects of relative clauses.

The example in (11a) shows that matrix subjects of active verbs may not be accusative marked. Likewise, the embedded subjects of clausal subjects may not receive accusative, as shown in (11b) and (11c) with both inanimate and animate embedded subjects. The embedded subjects of relative clauses may not be accusative-marked either, as shown in (11d). Any analysis of how accusative case appears on the subjects of embedded complement and adjoined clauses should also rule these structures out.

With this empirical background in place, let us first turn to a recent account of the phenomenon, provided by Fong (2019).

2.2 Fong (2019) and Accusative Subjects of Clausal Complements

In her 2019 paper Fong proposes that accusative case is assigned across a clausal boundary by the matrix verb to those subjects which have raised to the specifier position of a CP complement. Fong posits that the C head in Mongolian may bear ϕ-features, attracting an embedded subject to its specifier, where it is visible to the matrix v for Agree. Fong shows that accusative subjects are higher than their nominative counterparts, arguing that the position they have moved into must be the specifier position of the embedded CP. From there, matrix v may assign accusative across the clausal boundary, if it is able to.

As seen previously, accusative subjects also appear in nominalised complement clauses, as well as various adjoined subordinate clauses, in addition to those embedded under a CP. Adjoined subordinate clauses of either the PP or converbial type are not generated within the C-command domain of a structural-case assigning v (shown in sections 2.3 and 4.3), and so some version of a Long Distance Agreement analysis would be required to maintain matrix v as the assigner of accusative on embedded subjects in these clauses. More seriously, there is good evidence that the appearance of accusative subjects is not contingent upon the ability of matrix v to assign structural case. Importantly, without matrix v available as a case assigner, we are left wanting for another explanation for accusative embedded subjects.

2.3 Matrix v is not the accusative case assigner

In a basic sense, whenever an embedded subject appears with what is ostensibly objective case – here, accusative – it is important to investigate whatever connection there may be between this embedded subject and the matrix clause, where such case could be assigned. In the literature on Exceptional Case Marking, there are several ways of establishing this connection. First, the embedded subject could potentially raise into the matrix VP (“raising to object”). Otherwise, matrix v could assign accusative case across a clausal boundary given certain conditions. Both types of analysis rely on the ability of the matrix verb to assign accusative (see e.g. Chomsky (1986)). For any material canonically accessible to case-assigning v arguments against an Agree-based account must demonstrate that accusative is available in circumstances where we would not expect v to be able to assign case, such as in passives or intransitives. This is generally the line of argumentation pursued in Baker & Vinokurova (2010), and is also the primary line I pursue here. I will ultimately show that accusative in Mongolian is available independently from the matrix v’s ability to assign case, refuting both a raising-to-object analysis and the cross-clausal Agree account.

In Fong’s (2019) account discussed above, accusative may be assigned across the clause boundary by v. Fong’s argument that matrix v is the accusative case-assigner for embedded subjects rests on two pieces of data. First, consider (16):

The argument from (16) is that if the matrix verb is unable to assign accusative case as a passive, then the fact that accusative is apparently banned indicates that the embedded subject receives case from the matrix verb. However, (16) might be ruled out independently by the fact that Mongolian speakers statistically disprefer (but not ban) accusative on embedded subjects unless they are linearly close to matrix subjects (Guntsetseg 2016: 147), and since there is no other overt nominal in (16), speakers prefer the version without the accusative.

Data like (16) is complicated by several factors. First, the traditional passive -gd is marginal in modern spoken Mongolian (Janhunen (2012: 246); Kullman & Tserenpil (1996: 124)); not every speaker I consulted accepted data like (16).⁵ However, for those speakers⁶ that did accept (16), the accusative case remains available for both embedded subjects in clauses with gež (17a), as well as nominalised clauses that serve as the derived subject (17b), in both cases just so long as there is another argument in the sentence.

This pattern is unexpected if the matrix verb is the case assigner: a traditional view of the passive should prevent it from assigning accusative, and yet we see in (17) that accusative remains available, if other conditions are met.

Fong (2019: 9) also cites data like (18) where acc is banned on the embedded subjects of clausal subjects. The assertion is that they are out of the C-command domain of v, and cannot receive case, thus affirming that v must be the acc assigner.

While this is one plausible explanation of why (18) bans accusative case on the embedded subject of a matrix clausal subject, it is not the only one. I will propose that these examples are independently ruled out because there is no higher argument to C-command the embedded subject, as required under a configurational view. This is discussed further in section 4.1.

An even more forceful argument for the fact that the accusative may appear independently of structural-accusative-assigning verbs comes from sentences which lack them entirely. Accusative is available for embedded subjects even when the clause contains only intransitive verbs, as in (19) or in (20). This would be surprising under an Agree-based view of accusative assignment.

It is also possible to demonstrate that accusative is available on the subjects of embedded clauses which are not C-commanded by v. Specifically, in adjoined clauses such as converbials, the embedded subject may receive accusative even though it is demonstrably outside of the vP domain. This point is demonstrated in section 4.3 using adverbial scope tests and converbial clauses. The same point can also be supported with binding tests.

In (21), an R-expression in a converbial clause known to host accusative subjects (see 8), is co-indexed with a pronominal goal in the matrix predicate. Converbial clauses are transparent for binding (section 3.2), and so if this converbial clause were generated below the position of the indirect object in (21), we would expect a condition-C violation, which we do not observe. Independently, observe that an accusative-marked direct object appears following the indirect object. In section 3.1 below, I demonstrate that the IO-dat>DO-acc order is derived, and accusative is assigned at the upper edge of the vP. The fact that we see no condition-C violation between the indirect object and the R-expression within the converbial clause therefore requires that this converbial clause is generated above where the direct object receives acc, and therefore crucially, outside of the C-command domain of v. Now, the availability of accusative subjects in these converbial clauses demands an independent explanation.

Finally, sentences may also appear with two accusative arguments:

In (22), both the direct object of the matrix verb as well as an embedded subject receive accusative. Under a standard Agree-based account, there would only be a single case-assigner in (22), which (excluding multiple-agree) should not be able to enter into an agree relationship with two DPs. Alternatively a Dependent Case analysis is able to handle such a mismatch between the number of accusative-marked nominals and available case-assigners in a sentence.

3.1 Accusative in Direct Objects

In Mongolian, as in Sakha, direct objects may either appear with accusative, or be unmarked:

The examples in (29) show a definite object obligatorily receiving acc-marking (29b), a bare indefinite refusing acc (29c), and a simple indefinite variably taking acc-marking. Guntsetseg (2016) points out that a number of other factors aside from definiteness are involved here, including DP-type, incorporation, modification, and specificity, which together condition the relative ‘referentiality’ of the objects.

In rule (26b), B&V state that only an NP (here, an object) which appears in the same phase as another higher, unmarked C-commanding argument (here, the subject) may be accusative marked. This requires objects to shift⁷ to a site at least as high as the lower phase edge – a phenomenon which is well known to occur in DOM languages to specific objects (Diesing (1992); Diesing & Jelinek (1995); López (2012); i.a.). While demonstrating that objects have shifted in this way does not preclude analyses other than DCT, it is a pre-requisite for an analysis along the lines of B&V’s, and thus must be empirically demonstrated in Mongolian.

One method to show that objects have moved is to insert adverbs of various heights in the sentence, and inspect the linear position of marked and unmarked objects. This is what B&V do for Sakha, where they show that when objects appear linearly before certain adverbs, they are obligatorily accusative-marked, and when they appear in the immediately pre-verbal position, they may only receive accusative case with contrastive focus. The same facts hold of Mongolian, where a direct object appearing linearly above an adverb in (30) must be accusative-marked:

Ditransitive and causative sentences provide further evidence for object movement. The DO in these constructions may appear either before or after the dative-marked IO. The S>IO>DO>V order is the default order, the other is marked (Guntsetseg 2016: 24). We can see this in (31) where the DO of ‘introduce’ may appear on either side of the IO:

In the IO>DO order, the direct object may be unmarked:

However, in DO>IO order, the direct object must be accusative, and must be interpreted as specific:

The DO>IO order is a derived order: Gong (2022: 7) demonstrates that short (A–) scrambling is responsible for deriving the marked word orders. The obligatory accusative-marking that appears in the DO>IO order therefore correlates this case with structural height.

The structure in (34) below illustrates our interpretation of the configuration for accusative marking in direct objects under B&V’s theory, as it should apply to the Mongolian data discussed above. Here I assume an articulated verbal domain, as Mongolian exhibits overt Voice in addition to other verbalizing morphology. Ramifications for the division of the verbal domain on the phase boundaries are discussed in section 4.3.⁸

So, we have seen that movement of direct objects is correlated with accusative case assignment. However, a caveat from ditransitives: there exist sentences with the canonical IO>DO order, wherein the direct object – despite being below the indirect object – is accusative marked:

In (35), how do we know that accusative is not assigned to the direct object in its base position, invalidating our correlation between movement and accusative marking? One must demonstrate that word orders as in (35) are generated by further movement of the indirect object over the direct object. To do this, we can apply the same adverbial tests as above.

In (36) and (37) below, although the indirect and direct objects appear in the canonical/base order, they both appear linearly before an agentive adverb, and the DO must be accusative-marked. That is to say, the DO has moved above an adverbial that delineates the lower phase, is obligatorily accusative-marked, and then the canonical IO>DO order is derived by further movement of the IO.

Next let us consider causative sentences. Causees generally receive instrumental case, and direct objects are variably marked.

As in ditransitives, the direct object of transitive causatives may move above the causee. Accusative marking is then obligatory:

Again, his is a structurally higher position into which direct objects move, in which accusative becomes obligatory.

Having established that accusative is assigned to objects which have moved out of their base position, we can also pinpoint the location that this case is assigned. In her (2022) paper (also arguing a DC analysis of accusative in Mongolian) Gong demonstrates that accusative is assigned in a position below the subject, but above upper-VP material (such as indirect objects). The example in (40a) demonstrates⁹ that when a DO-acc (here scrambled to the upper edge of the clause) takes an indirect object binder, there is an available position above the indirect object, but below the subject into which it can reconstruct (that is, there is no obligatory reconstruction into its base position inside the VP), bleeding Condition C.¹⁰ This is contrasted with the ungrammatical example in (40b), where the direct object has a subject binder, and exhibits obligatory reconstruction, triggering a Condition C violation.

This situation and our interpretation of the structure appears in (41). Some details, such as the exact landing site of the shifted object above the lower phase, are discussed in section 4.3.

In these configurations, under B&V’s theory, accusative case is assigned only to those objects which raise into the same phase as the matrix subject.

Now, let us turn to the issue of embedded accusative subjects, a topic which Baker and Vinokurova characterise as “perhaps the most spectacular evidence for the dependent case account of the accusative…” (Baker & Vinokurova 2010: 615).

3.2 Embedded accusative subjects

Sakha and Mongolian are alike in permitting accusative embedded subjects. These subjects appear in a variety of clauses, but perhaps the most straightforward cases are those of complement clauses. The analysis goes as follows: a complement clause moves out of the lower phase in much the same way that direct objects do, landing in the same phase as the as-yet-unmarked subject argument, which C-commands it. If the moved complement clause itself is nominal, it will now receive accusative case by the application of rule (26b). Inside this clause, the embedded subject may raise to the upper edge, where it becomes visible to the matrix case calculations. In this position, it may also receive accusative case. Any “optionality” in the appearance of accusative on the subjects is the result of optional movement to the left edge of the embedded clause. This is illustrated below:

In the case of adjunct clauses, the solution is similar: an embedded subject must appear at the upper edge of the subordinated clause in order to be visible to the case algorithm in the matrix clause. The major difference here is that adjunct clauses do not move from a complement position inside the VP but are instead merged higher, inside the upper phase of the clause, C-commanded by the matrix subject from its position in specTP. This requires that the case algorithm again be triggered once the subject has arrived in specTP; this is discussed in further detail in section 3.3, and 4.4, and the appearance of accusative subjects in adjuncts of differing heights is crucial to the discussion in section 4.3, as well. Figure (43) outlines the configuration for accusative assignment in adjoined subordinate clauses:

Let us examine several instances of embedded accusative subjects to provide explicit examples of the mechanisms described above. In complement clauses there are two relevant movements to consider: the movement of the containing clause, as well as movement of the embedded subject to the upper edge of the containing clause. This generates four possible parses for a sentence with the following word order:

In two of these parses, the complement clause has moved upward, and in two it hasn’t. In section 2.2.1 I correlated movement of embedded subjects in CPs with accusative case on the embedded subject. Here, I will do the same for nominalised complement clauses, and additionally take up the question of clausal movement for both types of complement clauses.

Let us begin discussion of the data with nominalised complement clauses. Examples of these clauses with accusative subjects are found in (45) and (46).

It is possible to demonstrate that the accusative subjects of nominal complement clauses have moved to the upper edge through the set of binding and adverbial tests familiar to us from the discussion of CP complements in section 2.2.1. In (47) we see that if the embedded subject (with refl.poss bound by the matrix subject) appears at the upper edge of the embedded clause, it must be accusative marked.

Let us consider the test from the opposite direction: the presence of accusative case indicates that a subject must be at the upper edge of the containing clause. Recall from section 2.2.1 that finite complement clauses constituted a binding domain separate from their matrix host clauses. Only embedded subjects with accusative case could be licitly bound by the matrix subject; that is, accusative on the embedded subject is correlated with movement to a position in the upper edge of the embedded clause, diagnosed by its inclusion in the upper binding domain. This test can be replicated in nominalised complement clauses. In (48), the upper edge of the complement clause is delimited by an adverb that unambiguously modifies the embedded clause;¹¹ in this configuration only the appearance of accusative case allows this embedded subject to be bound by the matrix subject.

This demonstrates movement of the embedded subject within nominal complement clauses. Let’s now consider the second relevant movement: that of the clause itself into a higher position in the matrix clause. Accusative marking of embedded subjects in nominal complement clauses can be linked to the movement of the containing clauses by demonstrating that in these clausal complements, the accusative is unavailable on the embedded subject unless the containing clause itself is also marked: in (49), the complement clause is unmarked, and the accusative is impossible on the embedded subject ‘Tuyaa’.

In 49, the lack of marking on the complement clause should indicate that the clause itself has not moved into the upper domain, considering what we have already learned about direct object DP’s and their marking in section 3.1.

Let us now consider CP complement clauses. A sentence like (50) theoretically has two relevant parses, focussing on clausal movement: one where the complement clause remains low within the vP, and one where it has moved out.

As CP’s are not case-marked, we cannot examine the same evidence as in (49) for clausal movement. However, it is possible to disambiguate moved from unmoved structures by the placement of adverbs in the matrix clause, as done for direct objects.

Obligatory accusative marking of the embedded subject is observed when the complement appears above certain adverbials outside of the lower phase of the matrix clause. In (51) an adverbial is added to (50). Now, the CP complement with an obligatorily-marked embedded accusative subject appears to the left of this matrix adverbial.

Guntsetseg (2016: 168) has demonstrated that linear adjacency of two subjects is a major factor in conditioning accusative-marking of the embedded subject – yet optionality is still usually observed (see 50). We can account for the obligatory marking in (51) by arguing that the presence of the adverbial makes parses without clausal movement unavailable, and the speaker is only left with a sentence where they must, for functional communicative reasons, differentiate between two unambiguously adjacent subjects by moving and marking the second, embedded one.

Now contrast (51) with (52) from Fong (2019: 3): in the latter, the complement clause appears below a matrix adverbial, and the embedded subject only optionally receives accusative. Like the direct objects examined in section 3.1, obligatory accusative marking is associated with movement.¹²

Now we turn to subjects of adjoined clauses, i.e. converbial clauses (53), or adverbial subordinate clauses under postpositions (54).

With complement clauses, the nom/acc alternation on embedded subjects is accounted for by upwards movement of the containing clauses, as well as movement of their subjects to the upper edge of these clauses. Here, accusative subjects move to a position where they are visible to the case competition domain of the matrix clause, evidenced in part by their ability to be bound by matrix material. This suggests that both anaphor binding and Dependent Case could be correlated, although this correlation becomes less straightforward when we consider adjoined subordinate clauses.

Unlike complement clauses, adjuncts in Mongolian are transparent for binding purposes; they are part of the same binding domain as the rest of the matrix clause. Consider (55a), where the embedded pronominal subject creates a principle-B violation regardless of whether it appears in the nominative or the accusative. Likewise, in (55b) we see that the principle-A obeying possessive-reflexive suffix takes the matrix subject as an antecedent even if the embedded subject is unmarked.

If binding domains and case competition domains are necessarily phasal in Mongolian, one might like to conclude that these adjuncts are not in fact phasal. We might then expect embedded subjects of these clauses to be universally accusative-marked, which is also not what is observed. A simple solution to account for the nom/acc alternation on their subjects asserts that these adjuncts have multiple possible merge positions, both above and below the mid-clause phase edge.

The assumptions that drive this conclusion are not all straightforward, however. First, matrix material may bind anaphors within domains that should be considered phasal: in (56), the matrix subject is the antecedent for the embedded reflposs marker on the embedded subject of an object-gap relative clause modifying a noun that is itself in a locative phrase.

Phase theory (Chomsky (2001; 2008) etc.) would predict at least one phase boundary in (56), either at the level of the adjoined locative phrase, the DP level (ger-), or at the level of the relative clause itself; this complicates the assertion that binding is phase-bound in Mongolian. Likewise, a simple clause forms a single binding domain, yet two phases. If Mongolian binding is not always phase-bound (or even clause-bound), then we are not compelled to assert that adjuncts are non-phasal simply because they are transparent for binding.

What’s more, it is also possible to demonstrate that accusative subjects of PP-subordinated clauses appear higher than nominative ones, just as in complement clauses. Binding tests are unavailable here, but we can employ adverbials to delineate the edge of the subordinate clause.

In (57) it is possible to disambiguate the height of the embedded subject via the scope of an adverbial which may modify either the matrix or subordinated clause. In (57a), where the embedded subject is nominative, the sentence may describe two situations: if the current year is 2022, I will visit Mongolia in 2023; or alternatively, you will go to Mongolia in 2023 and I will visit some time in the further future (perhaps 2024). The former reading is associated with a matrix position for the adverbial ‘next year’, while in the second reading ‘next year’ modifies the subordinated clause only. However, in (57b), where the subject is accusative, the adverbial may not be construed as modifying the embedded clause, i.e. its position delineates the edge of the embedded clause.¹³

As for converbial adjunct constructions, the picture is once again complicated by the fact that converbs are temporally dependent on their hosts, and resist independent temporal modification of the type employed for PP adjuncts. These clauses also form a single binding domain with their hosts, so binding tests also yield no clear results. Without explicit evidence of subject movement, we may instead rely on the possibility raised for non-phasal subordinate clauses: they adjoin at differing heights. Just as direct objects only receive accusative if they have moved to the upper phase, so too adjuncts must adjoin sufficiently high for the case algorithm to operate.

It is possible, in fact, to find evidence that converb clauses adjoin at different heights: we can test for correlations between case-marking on the embedded subjects of converbials, and their readings relative to matrix adverbials (see also section 4.3 for discussion).

The converbial clause in (58) has two scopes relative to žarimdan ‘often’; there is a licit low scope reading like It is often the case that the teacher would drink tea until a student arrives, and a high scope reading, which speakers report as illogical: Until a student arrives, it is the case that the teacher often drinks tea. When presented the sentence in (58) independently with and without an accusative embedded subject, speakers report that the presence of the accusative forces the semantically odd reading (i.e. high-scope), indicating a high adjunction site for the converbial clause, which I take to be the upper phase.

I therefore leave open the possibility that converbial clauses are non-phasal, while accounting for the nom/acc alternation on their subjects. On the other hand, one need not conclude that PP’s are non-phasal, as it is possible to demonstrate movement of their subjects to the upper edge of the embedded domain as in complement clauses.

Having accounted for accusative case in objects and subjects through DCT, something must now be said about nominative and genitive cases.

3.3 On Nominative (and Genitive)

B&V’s approach to case is in fact hybrid nature: while they argue for a configurational account of the accusative and dative, they retain an Agree-based procedure for assigning nominative and genitive:

The basic issue for B&V concerning nominative and genitive assignment is whether they can be treated as fully unmarked cases, or whether they are assigned under Agree (following 59). Their argument for an Agree-based analysis relies on the unavailability of nominative in contexts where one would expect the spellout of default case – specifically inside PP’s where Dependent Case rules could not apply, and in tenseless participial clauses. In both of these environments, they argue (Baker & Vinokurova 2010: 626), lexical case is assigned instead, as neither dependent case rules nor φ-agreement can go through. Crucially, nominative (and genitive in DP’s with possessor agreement) is only available when φ-agreement with T or D is possible, requiring a rule as in (3.3).

An alternative view is one where the Case Filter is not active, and whichever nominals reach spell-out without a marked case, will receive the appropriate unmarked case for their domain: genitive within the DP, and nominative elsewhere. A variation of this alternative is proposed by Levin & Preminger (2015), who argue that a fully dependent model is possible for languages like Sakha, just so long as the causal direction between Agree and case licensing is inverted (the former being possible only after a nominal has been case-marked), permitting the case algorithm to assign unmarked to case to whichever nominals arrive at spell-out without having received case in the course of the syntactic derivation.

There are two important differences between Sakha and Mongolian that concern this issue, however: Mongolian does not exhibit overt φ-agreement, and the Mongolian genitive is morphologically visible in all canonically-expected environments.¹⁴ For these reasons, it is not possible to replicate Baker & Vinokurova’s evidence for the Agree-based rule. Let us instead consider a version of the case assignment algorithm in line with Levin and Preminger’s (2015) proposal, wherein the Case Filter is inactive.

Under this analysis, subjects in a standard clause would undergo movement to TP, and then the DC algorithm would apply. At this point, the subject is still caseless, and would receive nominative as the unmarked case for this domain. This ordering is also a departure from B&V’s original formulation: they ask of the reader to note that it ‘does not matter for this whether the subject moves to SpecTP or not’ (Baker & Vinokurova 2010: 601). As we will see in sections 4.3 and 4.4, Mongolian provides evidence that the subject must be able to trigger the case algorithm from its position in specTP, in order to ensure accusative can appear on the subjects of high adjoined clauses, and what’s more, that the algorithm must not simply apply indiscriminately whenever an appropriate configuration is generated; section 4.4 will constrain this behaviour by positing that the case algorithm is triggered from certain positions in the clause. This has a series of further consequences in the formulation of the procedure for case assignment, one of which being that we cannot assign nominative case to a subject immediately upon its arrival in SpecTP, if we are to maintain the caveat for rule (26b) that the upper (C-commanding) nominal must be caseless. This is more easily achieved through Levin and Preminger’s (2015) proposal, although it does not rule out B&V’s rule (59), just so long as the rules are strictly ordered.

As for genitive, it turns out that case alternations in participial clauses provide additional support for Levin and Preminger’s (2015) view. As Aravind (2021), Guntsetseg (2016) and others have described, the subject of nominalised participial clauses may appear in the genitive case.

In (60) we see that in a nominalised complement clause, the subject may appear in the genitive, in addition to accusative and nominative. This case is also available in contexts where accusative subjects are impossible, including relative clauses and clausal subjects. The crucial ingredient is the nominal nature of the clause.

Aravind (2021) argues that in these nominalised clauses the subject may move upwards cyclically: first, out of the embedded vP and into the nominalising nP layer (specifically, SpecnP), or secondly, if it is to receive the accusative, upwards into the specifier of the maximal DP layer. This behaviour is predictable under an account of the genitive as unmarked case within the nominal domain: if the embedded subject does not move into SpecDP, it remains in SpecnP, where there is no configuration appropriate to trigger DC marking, leaving it caseless until spell-out, where it will receive default case for the nominal domain: genitive. If no movement into nP occurs, then the subject will receive the unmarked case of its base domain: Nominative.

Of course, without evidence of overt agreement or the lexical-genitive data discussed by Baker & Vinokurova, it is hard to fully dismiss the Agree-based analysis for nominative and genitive cases. For further discussion of the status of these cases in Mongolian, I refer to recent work by Lim (2023), who suggests that a fully-configurational account is warranted, given nominative licensing in defective TP’s. For the purposes of this investigation however, I will tentatively assume that nominative and genitive are indeed unmarked case, and leave the topic for further research.

4.1 Where accusative subjects are banned

Recall as discussed in section (2.1) that accusative subjects are not possible in clausal subjects or in relative clauses. Let’s consider how these exclusions can be accounted for under a DCT analysis. First, consider clausal subjects (61):

The clausal subject “Hasa not coming” is clearly not C-commanded by any higher argument. Even if the embedded subject of this clause were to move to the upper edge, there is nothing to trigger the application of rule (26b) – repeated for convenience in (62) below:

Under an Agree-based analysis, these clausal subjects would be too high for matrix-v’s probe. However, this analysis would also rule out accusative subjects of higher adjunct clauses such as converbs. Instead, a more straightforward explanation is that accusative may only be assigned when the DP is C-commanded by some higher argument.

Another place one might expect to never see accusative is on a subject that has no possibility of moving to the edge of its embedded phase. That is to say, if the embedded argument is not visible to the case-assigning algorithm of the matrix clause, it will never receive accusative case regardless of whether it is C-commanded by a higher argument. This is exactly the case of relative clauses, whose subjects may not be accusative-marked:

As noted by Aravind (2021: 385), relative clauses in Mongolian are islands for several types of movement. There is therefore no clear mechanism to trigger movement of the subject out of the relative clause and further up to the edge of the head-noun’s maximal DP layer, where it could be visible to the matrix case algorithm. They should not, therefore, ever receive accusative.

This situation is visualised below in figure (65), following roughly the proposal by Aravind (2021: 380). Note that relative clauses in Mongolian are nominalised – this is indicated by the obligatory verbal noun suffixes on their verbs. While the embedded subject of such a clause could move to the upper edge of this phase (as observed in complement clauses), the entire nominal structure for the head noun that the relative clause modifies still remains between it and other matrix material.

The lack of accusative in relative clauses is exactly what would be predicted under DCT, where an embedded nominal that cannot move high enough to be visible to the matrix cause algorithm should not be marked.¹⁵

4.2 Voice Alternations and the Accusative

Causative constructions in Mongolian are interesting for a discussion of case assignment, as these permit additional arguments in monoclausal constructions. Mongolian causatives are part of a rich set of voice alternations in the language.

One important use of the causative in Modern Mongolian is actually the formation of passive-like constructions. In fact, the use of the causative marker is eclipsing that of the traditional passive marker, which is now used almost exclusively in writing. Example (66a) shows a sentence with ‘normal’ passive morphology, while (66b) is the same sentence, however this time with causative morphology.

When examining these causative-qua-passive constructions, we notice a curious fact: if the patient in a causative construction appears in the accusative, as seen in (67), only a true causative interpretation is available. This is in contrast with (66b) above where only a passive-like interpretation is available with an un-marked patient.¹⁶ (67) forms a minimal pair with (66b) immediately above.

There are several possible analyses of this construction, all of which propose an external argument which is in a position to C-command the patient in the true causative constructions only. It is conceivable that this external argument omitted at PF, (as Mongolian is a subject-drop language) – i.e. after the point at which it triggers the DC algorithm to mark accusative on the patient in (67). Alternatively, the passive versions could be related to adversity causatives (discussed e.g. by Pylkkänen (2008: 89–95) and references therein), where the unmarked subject in (66b) is a maleficiary. There is either no higher syntactic argument to trigger Dependent Case (see e.g. Alexiadou & Anagnostopoulou & Schäfer (2015: 29) who analyse the causative/malefactive alternation in these types of constructions as resulting from the type of Voice head), or alternatively a covert causer is merged in an oblique position where it is unable to C-command the patient-subject (which remains unmarked). The exact syntax of these causative-qua-passive constructions will have to await further research, but the fact that a distinct semantic contrast is associated with the lack of accusative on a patientive subject is intriguing.

There is an interesting parallel to this in the passive constructions in Sakha, as well: Baker & Vinokurova (2010) note that in Sakha the derived subjects of passives may in fact receive accusative case. This is perfectly compatible with a DCT analysis of the accusative, so long as one can argue for an implicit external argument in passives (as e.g. Pylkkänen (2008) does) which C-commands the patient argument.

4.3 Moving the Phase Boundary

In section 3.1 above I discussed how the placement of adverbials in a clause could serve to demonstrate that direct objects had moved out of their base position when receiving accusative marking. While in Sakha direct objects become obligatorily marked if they appear higher than manner adverbials (see: Baker & Vinokurova (2010: 602)), this is not exactly the case in Mongolian:

It has been accepted for some time that adverbs of different types occupy different heights in the syntax (Jackendoff (1972); Cinque (1999) etc.). For example manner adverbials are treated as adjoining lower within the verbal domain, while agent-oriented adverbials modify a projection as high as VoiceP. In Mongolian, we see that it is only objects to the left of agent-oriented adverbs that are obligatorily accusative:

This of course means that the boundary between phases in Mongolian must be slightly higher than in Sakha. For B&V, the edge of the phase was the VP; I have evidence that in Mongolian, at least the vP is the phase boundary. This possibility was mentioned briefly by Baker & Vinokurova (2010: 601), who remark in a footnote that there is considerable disagreement in the literature as to the boundary of the lower phase, and some such as Fox & Pesetsky (2004) argue that individual languages may vary in this respect.

A concern with moving the upper boundary of the lower phase to the edge of the vP,¹⁷ however, is that the subject in its first-merge position now resides in the same phase as the direct (and indirect) object, and ought to trigger accusative case without requiring movement of the object, levelling all marking variation in the language. This solution as it stands clearly overgenerates. In the same footnote, however, B&V note that one could instead implement a more ‘orthodox’ view of phase theory, following Chomsky’s (2000; 2001) view that upon arriving at a phase edge, it is the immediately-lower phrase that is sent to the interfaces: so, when the vP edge is reached, the VP is shipped off. This, however, would require, as Baker and Vinokurova note, a more strict use of the A/A-bar distinction to ensure that phase-internal scrambling does not affect the required configurations. This more nuanced implementation is not adopted or spelled out in any detail, but we will see in section 4.4, that one must take it more seriously, in order to account for further timing paradoxes.

One further consequence of adopting vP as the lower phase is that if the subject of a subordinated clause receives accusative, we must be able to demonstrate that this clause appears outside of the vP. One possible place to examine this issue in Mongolian is in converbs. The converbs of Mongolian are often divided between “specialised” and “general” converbs (Ujiyediin 1998). So far, we have seen the -magč and -saar converb suffixes, both of which are counted among the ‘specialised’ converbs, contributing specific semantic content about the ordering of events. There is precedence in the literature on converbs in Turkic languages (see e.g. Privoznov (2022)) arguing for attachment at various levels – including TP and VP – and one might expect the same in Mongolian.

Readings related to adverbs of frequency in the matrix clause can disambiguate the height of different converb classes. I will refer to two readings for these converbs relative to matrix adverbs: high readings, where the event described by the converbial is not modified or otherwise within the scope of a matrix adverbial, and low readings, where the converbial clause appears along with the matrix predicate, within the scope of the adverbial. General converbs in Mongolian only produce low-scope readings. That is, in (70), a reading like “Sudu often stands, [sometimes] while smiling [sometimes while doing something else]” is unavailable: the adverb of frequency must scope over both events of standing and smiling, together.

This is crucially in contrast to the “specialised” converbs which strongly prefer high readings. In (71a), the preferred reading is ‘If I stand on the bus, I sometimes fall’, not ‘Sometimes, I stand on the bus and fall.’ In (71b), the adverb of frequency likewise only applies to the matrix predicate.

Importantly, the unavailability of high-scope readings for the general converbs places them syntactically below such adverbs as “often”, while high scope readings for specialised converbs point to a structurally higher position in the clause.

Specialised converbs like -magč and -saar (and -xleer below) are perfectly happy with accusative subjects (72):

By the analysis forwarded here, these converbs would adjoin above the lower phase of the clause. Specialised converb clauses may also never appear to the right of unmarked and un-scrambled matrix objects, and unlike general converbs, they may appear sentence-initially. This and the scope facts above suggest a higher adjunction position in the clause – outside of the lower phase.¹⁸

The distinction between these two types of converbial clauses can be visualised as follows:

4.4 Timing and Rule Application

So far, I have mentioned two crucial areas of difference between Baker & Vinokurova’s formulation of Dependent Case and the approach I adopt here: first, in section 3.3, I discussed the possibility that nominative and genitive case are unmarked, or default case, assigned at the point of spell-out. Secondly, I argued that the phase edge for Mongolian is the very upper edge of the verbal domain (i.e. Voice or v) instead of VP. Some consequences of these two deviations include: the possibility that even indirect objects may be marked with the accusative, as they would appear in the same phase as the subject in its base position, as well as the necessity that this subject be capable of triggering the case algorithm from a position in specTP. In both of these situations, the critical issue is the timing of rule application relative to the derivation. I will propose in this section a solution to these issues, in that Baker and Vinokurova’s rules, rather than applying “… immediately when the configurations they describe are first created…” (Baker & Vinokurova 2010: 604), only apply when triggered from specific positions.

Let us elucidate a conflict in timing by discussing a paradox in dative vs. accusative assignment noted by B&V, wherein their rules would seemingly overgenerate with regards to the marking of indirect objects.

In ditransitives, rule (26a) applies iff an indirect object (e.g. a goal of a ditransitive) is merged in the VP in a position C-commanding the direct object. If the direct object remains in-situ, its immediately-C-commanding nominal is now case-marked, and accusative is not assigned in this base position (rule 26b). However, if the direct object moves out of the lower phase (and across the indirect object), it is free to be marked with the accusative if the other conditions of rule (26b) are met. Crucially, these rules must be applied in exactly this order, otherwise one or the other case would never be marked under normal circumstances.

In those languages where the canonical order of ditransitives is S-IO-DO-V, both objects are often marked – that is, the direct object has moved out of the lower phase to receive accusative, yet the dative-marked indirect object appears linearly to the left, and therefore structurally higher as well. In such a sentence, we therefore have a configuration where the dative indirect object is now a potential target for the application of rule (26b): it is in the higher phase, and possibly C-commanded by a unmarked subject nominal. In such cases, why do we not see two accusative-marked objects?

Baker & Vinokurova (2010) consider these configurations, like the one in (74), and argue that the solution is to permit multiple case assignment, or, case-stacking.

B&V contend that in cases like (74), both dat and acc are in fact assigned; the speaker however spells out the innermost case. Similar sentences are easily constructed in Mongolian,¹⁹ where the goal remains dative marked despite appearing in the upper phase (observe the placement of the adverbial sanaataigaar):

It is impossible to tell from sentences like (74) or (75) alone whether the case stacking analysis works. Let us then look to other places where dative appears. Mongolian expresses need predicates with dative-marked experiencers. Consider (76), where the the first-person pronoun appears in the dative.

There are several ways to explain the appearance of the dative in (76). The most straightforward analysis under a Dependent Case approach would have both arguments merged within the vP, the higher of which being the experiencer, which is dative marked as per rule (26a). However, when these clauses are embedded, this experiencer may be marked with the accusative (77):

Under B&V’s case stacking approach, one would not expect to see overt accusative marking on embedded experiencers. This nominal would receive dative first in the embedded domain, and then upon moving to the upper edge of this clause it would be accessible to the cycle of case assignment rules in the matrix clause, where it could receive accusative. Just as in the ditransitive examples, accusative should be “stacked” on dative, which should nevertheless be spelled out, as the innermost case. The appearance of accusative on an originally-dative argument is surprising: if both structures involve case-stacking, why would we spell out the innermost case in ditransitives, but the outermost with embedded experiencers?

An initial solution would be to permit this scrambling in ditransitives to occur after all iterations of the Dependent Case algorithm have applied. However, as Baker & Vinokurova have formulated their approach, these rules must apply whenever an appropriate configuration is generated, all the way up until the end of narrow syntax. One cannot, however, stipulate that these sorts of scrambling occur post-syntactically, as they do appear to have interpretive ramifications: as Guntsetseg (2016: 96) notes, scrambling of this type can affect relative quantifier scope and thus must have some basis in the syntax.²⁰ Yet: the basic intuition that the application of the DC algorithm must occur at specific points in the derivation (i.e. after some movements, but before others) remains.

We now have conflicting demands on timing in our case assignment algorithm:

1. In the upper phase, scrambling in ditransitives must not trigger accusative-marking of dative indirect objects, but we must still be able to overwrite dative on other types of arguments

2. As the phase edge is now the upper edge of the VoiceP, we must not allow the subject to trigger accusative marking in its base position, otherwise we will overgenerate in terms of accusative assignment to arguments in the lower phase (see section 4.3)

3. The subject must remain unmarked until after it arrives in SpecTP, otherwise accusative marking will not be triggered²¹ on the subjects of high adjuncts (see section 4.3 on converbs and section 3.3 on nominative)

As discussed in section 4.3, B&V mention – but choose not to implement – a version of their theory wherein the phase edge is the upper edge of the verbal domain, at vP. They remark that this would require greater use of the A/A-bar distinction in order to prevent scrambling from overgenerating case marking, although they do not elaborate further. I will now spell out a version of this solution in more detail and suggest it as a possibility for Mongolian.

First, we can prevent the subject from triggering accusative in the lower phase in two ways. We could argue that by merging in the specifier of the maximal projection of the lower phase, it does not count for case calculations in the lower domain.²² Alternatively, we could further restrict the application of rule (26b) to the upper phase. Both solutions are not without issue. If the subject is base-merged at the very upper edge of the vP domain, then this leaves no room below it for the object to move into such that it is both outside the lower phase, and still C-commanded by the subject. Adverbial data from the previous section does suggest that direct objects must move rather high to obligatorily receive accusative; however, if we argue that they move into a position above the base position of the subject, we must avoid creating a configuration wherein the direct object triggers rule (26b), assigning accusative to the subject.

Instead, we can constrain the timing of the Dependent Case algorithm by arguing that its application is triggered when a nominal merges (first or subsequent) in specific positions creating the requisite configurations – so far we have seen that those positions must be at minimum the indirect object position (to trigger dative assignment), and SpecTP (to trigger accusative assignment on raised objects, and the subjects of adjoined clauses). One possible way to unify these positions notes that they are the specifiers of phrases that appear as complements of phase heads.

Let us outline specifically the conflicts I resolve with this approach. The first theoretical demand I identified above requires that a dative argument which moves to a position in the upper phase, but below a subject in specTP, will not trigger the (26b) rule. If this rule is only triggered once (when the subject moves to specTP), any scrambling such as would move the dative argument upwards, should never trigger accusative marking, so long as this movement occurs after the movement that brought the subject to specTP. Secondly, if the base-merge position of the subject (specVoiceP) is not among those positions which are tied to triggering the dependent case rules, then it does not matter that the subject falls within or on the edge of the lower phase if the boundary is as high as I have proposed: it should not trigger invariable accusative marking on any objects in the lower phase. As for the third concern, we now trigger the (26b) rule exactly when the subject arrives in TP, above all adjuncts like the converbs examined in section 4.3.

The two positions I hypothesise to trigger DC assignment are a subset of the A-positions in the clause, but we cannot use this natural class as the trigger for dependent case rules: if all A positions could trigger the algorithm, one would also expect the base-merge position of the subject to do so as well, which is exactly the situation we want to avoid. Perhaps then this solution is not exactly what Baker and Vinokurova had envisioned in their footnote; although if we are to take Dependent Case seriously, the correlation between A-positions and case becomes more tenuous regardless. This is ultimately the position that Gong (2022) arrives at through investigation of reconstruction: positions where case may be assigned are more relevant, be they A-positions or not. Instead the relevant conditions are more concerned with the timing of phase spell-out.

In any case, some constraint of the Dependent Case algorithm is necessary, and a theoretically-simple way of achieving this is to trigger it from specific positions – perhaps defined by the language in question, or common among many. The exact implementation of this, however, I leave open for further research.