This article shows that accusative case and object agreement are not closely related in Amharic, a language in which both are morphologically overt. This suggests that it is wrong to generalize Chomsky’s claim that agreement and case are both manifestations of the same Agree relation from subjects to objects across the board. I show that object agreements in Amharic are the true manifestations of Agree (not pronominal clitics), whereas accusative case is assigned independently, to the lower of two nominals in the same domain. The relationship between case and agreement can thus vary parametrically.
In Chomsky 2000, 2001, and much related minimalist work, case and agreement are seen as two sides of the same coin: both are morphological manifestations of feature sharing created by the application of a single syntactic process, called Agree. As a result, the NP that a functional head H agrees with in Ф-features will be the same as the NP that bears the case value associated with H. There is rather clear motivation for this view from the relationship between agreement on T(ense)/I(nfl) and nominative case in Indo-European (IE) languages and in quite a few others. For example, the English verb agrees with the subject and the subject bears nominative case in (1a), whereas in (1b) the verb does not agree with the subject, and the subject does not bear nominative case. Similarly, in Icelandic, the T associated with a finite verb agrees with a nominative subject, as in (2a), but it does not agree with a dative subject, as in (2b); rather, when the subject is dative, the finite verb agrees ( partially) with the nominative object.
That she walks to work each day is good for her health.
For her to walk to work each day would be good for her health.
So there is some firm empirical grounding for relating agreement on T and nominative case in a tight theoretical fashion. In Baker 2008:chap. 5, I argue that this relationship is parameterized, not universal, but even so it holds for more than 50% of the languages I studied.
The question to be reconsidered here is how this extends to object agreement and accusative case marking. Chomsky’s own view, widely accepted in the literature, is that it carries over directly. It is assumed that there is some other functional head F—typically identified as active/ transitive v—that is distinct from T but is subject to the same principles. This other head F enters into an Agree relation with the NP closest to it, the object, and the result is that the Ф-features of the object appear on F and accusative case appears on the NP. But there is a superficial difference: object agreement is not realized in overt morphology in English and most other IE languages. Thus, one cannot point to clear facts like (1) and (2) to ground this extension—or to challenge it. The parallel treatment is conceptually attractive, but its empirical support is comparatively slight.
There are, however, languages outside of IE that do have both overt object agreement and accusative case morphology. It is interesting, then, to see if one does observe in these languages the close relationship between object agreement and accusative case that much current theory would expect—if one finds analogues of (1) and (2) for objects and object agreement. With this in mind, I focus here on Amharic, a Semitic language spoken in Ethiopia. I show that in fact, object agreement and accusative case do not correspond very closely in this language; rather, they diverge in systematic ways (section 2). I then show that object agreement does have the core properties one would expect of a manifestation of Agree (section 3)—arguing against the plausible view that the morphemes in question are pronominal clitics rather than true agreement. This implies that one needs to identify a source for accusative case in Amharic other than Agree. I propose (in section 4) that the Amharic accusative is instead a dependent case in the sense of Marantz 1991. I conclude that the standard view holds only parametrically.
Mismatches between Object Agreement and Accusative Case
There are some simple and salient situations in which case marking and agreement do seem to be related in Amharic. For example, in (3a) the indefinite object ‘dog’ does not bear accusative case, and it does not trigger object agreement on the verb either. In contrast, in (3b) the definite object ‘dog’ both bears accusative case and triggers object agreement. This could be taken as prima facie evidence that there is a close relationship between the two phenomena.
But there can fail to be a relationship between the two as well. Indeed, it is not hard to show that the relationship between case and agreement breaks down in both directions—as recognized independently by, for example, Kramer (2010).
The simplest mismatch between case and agreement is that a definite object bearing accusative case is also possible in clauses where the verb does not agree with it. Descriptively speaking, object agreement is optional with a definite object: (4) is grammatical, as well as (3b).
i∫∫a-w- in j-aj-al.
Lemma dog-def-acc 3mS-see-aux(3mS)
‘Lemma sees the dog.’
(also WL:181, 186, MA05:299, Kramer 2010:7)
This optionality of object agreement is a general feature of Amharic, noted by all observers.
Some classes of nominal show the pattern in (4) more rigidly: they require accusative case, but object agreement with them is forbidden. This tends to be true for DPs that are not fully/ canonically referential. For example, direct objects that are quantified must be marked accusative, but they cannot appear with object agreement.
Interrogative DPs and reflexive anaphors must also be accusative, but cannot be agreed with.
A third class of accusative nominals that does not correspond to object agreement is theme arguments in double object constructions. In the presence of a goal or source argument, the theme argument of a triadic verb cannot be agreed with. The verb can only bear one instance of object agreement, and that must usually agree with the goal or source. Nevertheless, the theme argument may be marked accusative, as well as the goal/source argument.1
Morphological causative constructions behave similarly (see Amberber 2002:46–47).
Conversely, there are also situations in which the verb shows object agreement with a DP that does not bear morphological accusative case. One case in point again concerns double object constructions. If the higher object is a goal (but not a source, as in (9)), then it can bear the dative case marker/preposition lә rather than the accusative suffix -n. Nevertheless, the verb shows object agreement with the goal (and not the theme) in this case frame too.
dat-Aster child-def-acc show-1sS-3fO
‘I showed Aster the child.’
(also WL:191, 423, MA05:301, Kramer 2010:7)
Something similar happens in Amharic’s (somewhat peculiar) applicative construction. The prepositions lә and bә can appear suffixed on the verb, while their double still appears in construction with the relevant nominal. When this happens, the verb shows object agreement with the NP inside the PP headed by the doubled P. That NP cannot have accusative case. (12) is such an example in which there is object agreement with an NP marked by bә ‘with’.2
It is even possible for a verb to show object agreement with an NP in nominative (unmarked) case. This happens with verbal predicates that have an experiencer or possessor argument. Such arguments must trigger object agreement on the verb, not subject agreement. The experiencer or possessor may nevertheless bear nominative case. (Accusative case is also a less common option with the verb ‘hunger’ in (13), although not with allә ‘exist, have’ in (14).)
‘Almaz is hungry.’
(also WL:435, MA05:308, Kramer 2010:9n8)
‘Aster has a dog.’ (lit. ‘A dog exists to/for Aster.’)
(also WL:439–440, MA02:23, Kramer 2010:9n8)
In conclusion, then, under the right circumstances it is possible to have object agreement with DPs bearing a range of morphological markings in Amharic. At the same time, it is possible to have an accusative DP that the verb does not agree with. So accusative case and object agreement are not morphological manifestations of the same abstract relation in Amharic.
One could try to handle these facts strictly within the standard Chomskyan theory by using the idea that there is some abstractness in how Agree relationships in the syntax are realized at PF. For example, it is conceivable that the lower agreeing head F enters into an Agree relation with both accusative arguments in (8), but only one of those agreements is realized overtly. Conversely, one might say that the goal in (10) and the experiencer in (13) get accusative case from F, but they also get another case from some other head, and only that other case is realized overtly at PF. Or one could say that there is some other functional head (v?), distinct from F (= AgrO?), that agrees with all and only the accusative nominals, but that that head is not spelled out overtly.3 It is hard to entirely rule out all such analyses, given that some mismatch between syntactic Agree relations and their overt manifestations is possible in all standard accounts. Nevertheless, these alternatives seem quite ad hoc, and in general one hopes that the syntax and morphology of agreement will not diverge too much. It is one thing to say that the syntactic reflexes of Agree are not fully realized at PF in some languages (like English); it is another to say that the abstract case and agreement relationships are systematically different from those that are realized at PF—which is what one would have to say about Amharic. Instead, I suggest that we take the mismatches seriously and see what they might tell us about the syntax of case and agreement.
Object Agreement Is the Result of Agree
If object agreement and accusative case are not closely related, it follows that only one of them (at most) can be the result of Chomsky’s Agree. The question arises, then, which one? I claim that visible object agreement is in fact the morphological manifestation of Agree in Amharic. This is, however, debatable. Kramer (2010), for example, argues that the morphemes that we can neutrally call “object markers” (OMs) are really pronominal clitics, and there are some good reasons to think this. If that is right, then true object agreement in Amharic is purely abstract after all. It then becomes possible again to say that object agreement tracks accusative case closely, but one just doesn’t see it, as in the Chomskyan account of English. So some defense is in order for my analysis of OMs as true object agreement, manifestations of Agree. While it is not possible to consider carefully every conceivable analysis of pronominal clitics in an article like this, I show that the claim that OMs are manifestations of Agree in Amharic explains several substantive facts about those OMs, including ways that they differ from familiar and much-discussed instances of pronominal clitics, such as those in the Romance languages (see, e.g., Kayne 1975, 1989, Rizzi 1982, Sportiche 1997).
The first relevant fact is that only one OM is possible on the verb in Amharic. We have already seen that Amharic has roughly the usual range of double object verbs. In principle, both objects of such a verb could be pronouns. And indeed in Romance languages both pronouns can cliticize to the verb (e.g., Spanish Se lo di ‘I gave it to him/her’, lit. ‘him/her it I.gave’). But this is not possible in Amharic; rather, at most one OM can appear on the Amharic verb, as shown in (15).
The correct form would have a single OM that expresses one of the two arguments, while the other is simply not represented on the verb, as in (11). The pronominal clitic view would need to resort to some kind of clitic cluster simplification process for this, whereby one OM is deleted before or after another one at PF—an added stipulation.4 In contrast, the agreement analysis has a very simple account of why object marking is unique: there is only one F head in the functional architecture of the clause ( just as there is only one T head), and it agrees only once.5
Second, and most significantly, the theory of Agree implies that F should be required to agree only with the highest argument inside VP. We already saw evidence that this is true, in (8)–(11). In ditransitive verbs with a goal or source argument, my consultant allows an OM for the goal or source argument but not for the theme argument. This holds both when the nontheme argument is marked by the dative morpheme lә and when it is accusative (although not, perhaps, in certain examples in which lә is arguably a preposition in syntax).6 Another example is (16).
‘Lemma told Almaz the story/his story.’
(also Kramer 2010:8n6)
Similar effects are found in morphological causatives. A prediction that emerges naturally out of the agreement theory is thus supported. In contrast, there is again no immediate explanation of these facts in terms of pronoun cliticization. For example, in Romance languages it is possible to have a direct object clitic in the presence of an indirect object (e.g., French Marie l’a donné a‘ Jean ‘Marie gave it to Jean’, lit. ‘Marie it-has given to Jean’). Amharic is different from Romance in this respect, in a way that supports an analysis in terms of Agree.
Perhaps the best syntactic reason for taking OMs in Amharic to be pronominal clitics is the fact that they are optional with definite objects ((3)–(4)) and impossible with nonreferential objects ((5)–(7)). This is different from canonical instances of agreement (such as agreement on T with a subject), which we expect to be obligatory whenever structurally possible, and not to be influenced by the referential qualities of the NP. Indeed, in Amharic itself it is perfectly possible for T (or Aspect) to show agreement with a quantified or indefinite subject (Leslau 1995:68, 122, 151). In contrast, both conditions recall clitic left-dislocation (CLLD) in Romance, as in French Jean, je l’aime (‘Jean, I love him’, lit. ‘Jean I him-love’), which is an option (not a requirement) for definite NPs and which is impossible for indefinite and quantified NPs (Cinque 1990).
But one simple reason to be wary of analogizing examples like (3b) to CLLD is the fact that the putatively dislocated NP need not be dislocated very far. It is perfectly normal for an object doubled by an OM to be internal to the clause in Amharic. For example, the agreed-with object can follow both the in-situ subject and a low VP-type adverb, as in (19).
So if this is CLLD, it is a funny sort of string-vacuous dislocation, quite different from familiar CLLD as we know it in Italian or French.7
Also significant is the fact that there are structures in Amharic in which an OM is obligatory and insensitive to definiteness/referentiality in exactly the way we expect ordinary agreement to be. In particular, this is the situation with the experiencer arguments of certain nonagentive verbs, such as ‘hunger’ and ‘have’, as shown in (13) and (14). These are the “differential subject marking” constructions discussed by Amberber (2005), called “impersonal” constructions by Leslau (1995). Such verbs may have a theme argument as well as an experiencer or possessor argument, but they have no true agent. The set of such verbs includes simple experiencer verbs (‘worry’, ‘astonish’, ‘hurt’, etc.), experiencer-theme verbs (‘die on’, ‘lose’, etc.), impersonal verbs with an optional experiencer (‘be cold’, ‘be hot’), and existential-possessive verbs (‘exist, have’) (MA02:20–23, MA05:307–309). What these predicates have in common grammatically is that the experiencer argument is optionally marked accusative and obligatorily triggers object marking on the verb (MA05:304).
In this respect, experiencer arguments are different from theme objects, which must be marked accusative (if definite) and optionally trigger object marking on the verb, as seen in (3)–(4). Furthermore, with this class of verbs, object marking can cross-reference nonreferential NPs. (21) shows OMs doubling a bare plural NP, a quantified NP, an interrogative NP, and a negative polarity item.
woman-pl grace exist-(3mS)-3pO
‘Women have grace.’
‘Who is sick?’
every person money lose.3mS-3mO
‘Everyone lost some money.’
who-foc person money neg-get.lost.3mS-3mO-foc contrast (5b))
‘No one lost money.’
With these verbs, then, OMs do behave the way that we expect true agreement to behave. I therefore conclude that they are agreement. But the OM on experiencer verbs is the same formal entity as the OM registering the theme of an ordinary transitive verb.8 I thus conclude that the OM is always agreement on a functional head in Amharic, rejecting any pronominal cliticization analysis that implies that overt objects are dislocated.
On the basis of this range of data, we can sketch an account of OMs in Amharic in terms of Agree as follows. We normally expect that a relevant functional head must undergo Agree if there is an NP accessible to it. Suppose we take that to be true for object agreement in Amharic too. But agreement is not always possible: it is governed by various restrictions, including locality restrictions. In particular, within Chomsky’s (2000, 2001) theory, there must be no intervening NP, and agreement cannot take place across a phase boundary (the Phase Impenetrability Condition). With this in mind, I suggest that the VP counts as a phase-like domain in Amharic, into which F cannot agree.9 Experiencer arguments are automatically within the agreement domain of F: since they bear a thematic role distinct from theme, they are initially generated outside VP, say (for concreteness) in Spec,ApplP. However, theme arguments are not automatically in the agreement domain of F. They are initially inside VP; and if they remain there, F will be unable to agree with them.10 However, certain kinds of nominals have the option of undergoing object shift, a short NP-movement that carries them out of the VP. Object shift in (for example) Germanic languages is known to be related to the definiteness and quantificational properties of the object, according to Diesing (1992) and many others. Let us then suppose that a similar kind of object shift happens in Amharic. When it does, it puts the theme within range for agreement with F—in the same general region of the clause that experiencers inhabit automatically by base-generation. Finally, certain types of NPs are barred from undergoing object shift for semantic reasons ( possibly different ones), so they can never be agreed with.11 The crucial structures are compared in (22), using English words for ease of reference.12
A complication with this view is that even indefinite NPs can apparently move out of the VP in some sense, in that they, like definite NPs, can precede an adverb or a goal phrase.
However, movement of the definite NP can feed object agreement, as in (23b), but movement of the indefinite NP cannot; see (23a). This holds true even though indefinite NPs can trigger agreement if they are base-generated outside the VP in the first place (e.g., (21a)), proving that they have the Ф-features necessary to undergo agreement. I interpret this as showing that indefinite objects undergo a different type of movement than definite ones do: definite NPs can undergo object shift, a type of A-movement, whereas indefinite NPs can only undergo a type of scrambling that counts as Ā-movement. We can then add that F can only see NPs in A-positions to agree with them. Some independent evidence for this hypothesis comes from (24), which shows that an indefinite theme cannot bind a pronoun in a goal argument it has moved over. This is a kind of Weak Crossover violation, showing that the theme undergoes Ā-movement, not A-movement (data from Mengistu Amberber, pers. comm.).
We thus have some converging evidence that a type of movement that cannot create new binding relationships also cannot feed agreement.13
Finally, I should say something about what F, the functional head that agrees with the object, really is. The standard candidate is active transitive v. However, that seems to be the wrong choice in Amharic. It is true that one (and only one) OM is possible in transitive clauses in Amharic. But it is also possible to have one (and only one) OM on unaccusative verbs. The experiencer-theme verbs in (21c–d) are cases in point: they show object agreement with the experiencer, but they should not have an active transitive v. (25) is another such example.
Similarly, object agreement with the goal is possible on verbs in passive voice in Amharic.
Amharic is different in this respect from many languages, including Chicheŵa, Mohawk, and Mapudungun, which do not allow any object agreement on passive or reflexive-middle verbs. Therefore, I claim that F is not the same as transitive v in Amharic. Rather, it is some functional head that is potentially present in any clause. This could be Aspect, or a dedicated AgrO head—or even v, as long as all vs undergo agreement in Amharic, not just active transitive ones. Given the morphological independence of object agreement from any other observable lexical or morphological category in Amharic (Kramer 2010),14 perhaps AgrO is the most natural choice. (I return to a fuller analysis of examples like (25) and (26) below.)
In summary, object agreement has the properties one expects if there is a functional head that undergoes Agree in Amharic. First, it agrees downward, with an object or applied object, but not with the subject in Spec,vP. Second, it is unique. Third, it agrees only with the highest NP in the greater verb phrase (the intervention effect). Fourth, it is subject to phase-like conditions, not agreeing into PP (usually) or transitive VP. The one aspect of Chomsky’s theory of Agree that does not apply to object agreement in Amharic is the activity condition, given that F can agree with an NP that has a case feature valued as dative or oblique (see (10)–(12)). But as already shown in Baker 2008:chap. 5, this condition is parameterized, holding only in one class of languages. I conclude that object marking is indeed a manifestation of Agree in Amharic.
Whence Accusative Case?
This conclusion raises another theoretically important issue. If accusative case does not depend on object agreement, and object agreement is a manifestation of Agree, then accusative case is not a manifestation of Agree. Where then does accusative marking come from in Amharic?
An alternative made available by the literature is that accusative case is a dependent case in the sense of Marantz 1991. Baker and Vinokurova (2010) argue in detail that this is true for accusative case in Sakha, a Turkic language that has subject agreement but no (overt) object agreement. Subject agreement in Sakha is closely related to nominative case in the way familiar from IE. But accusative case, in addition to not being paired with visible agreement, appears in some surprising contexts: in passive clauses with a covert agent, in agentive nominalizations, and in subject-raising constructions even when the matrix verb is not transitive. Baker and Vinokurova use data like these to argue for the principle of accusative case marking in (27) for Sakha. This is a fairly straightforward updating of Marantz’s proposal.
(27) If there are two distinct argumental nominals X and Y in the same phase such that X c-commands Y, then value the case feature of Y as accusative unless X has already been marked for case. (Baker and Vinokurova 2010:595)
Baker and Vinokurova argue that (27) applies in the syntax, immediately when a relevant constituent is built (e.g., prior to scrambling)—in contrast to Marantz (1991), who claimed that dependent case marking happens in PF. That point is not particularly crucial here, however.
If a rule like (27) is the source of accusative case in Sakha, a language without object agreement, it could also be the source of accusative case in Amharic, a language that has object agreement that is independent of accusative case marking. I claim that this is the true situation. For Amharic, I modify (27) slightly, as in (28).
(28) If there are two distinct argumental nominals X and Y in the same clause such that X c-commands Y, then value the case feature of Y as accusative unless X has already been marked for case.
The rule in (28) works straightforwardly in simple examples. If a clause is intransitive, having only one NP argument, (28) does not apply and the sole argument of the verb is left to be nominative. If the clause is monotransitive, then accusative case is assigned to the lower NP, the object, and not to the higher one, as in (3b) and (4). If the clause is ditransitive, having a subject and two internal arguments (and if no dative or oblique case is assigned), then (28) can apply twice, marking both internal arguments as accusative, as in (8) and (9).
The next question to consider is why there is no accusative case marking on the indefinite direct object in an example like (3a). Baker and Vinokurova (2010) consider a superficially similar fact in Sakha and account for it by saying that VP is a phase in Sakha and accusative case assignment is sensitive to phase boundaries. Indefinite NPs in Sakha do not move out of VP; hence, there is only one NP in the VP phase and one NP in the CP phase, and (27) never applies. But this does not seem to be the correct approach for Amharic, because (4)–(7) show that accusative case is possible on definite NPs in Amharic even if they have not moved out of VP, as diagnosed by the fact that object agreement with the NP does not happen. This is my rationale for changing (27) to (28); (28) states that case assignment in Amharic is sensitive to clause boundaries but not to VP or vP boundaries ( phases). Case assignment in Amharic thus depends only on whether there are two nominals in the same clause, not on whether one is in VP or not. The fact that Agree is sensitive to the VP boundary in Amharic but case assignment is not helps to account for the fact that object agreement and accusative case are not closely correlated in Amharic when the NP is definite, as shown in (3b) versus (4).15
Rather than depending on the position of the object in the clause, the crucial factor underlying the kind of differential object marking seen in (3) in Amharic is the internal structure of the object nominal. Accusative case in Amharic is always marked on an object if the object is a proper name, an NP that bears the definite suffix -u, a wh-phrase, a pronoun or demonstrative, a possessed NP, an NP together with a quantifier like ‘every’ or ‘all’, or an NP interpreted generically (Amberber 2005:300; see also Leslau 1995). In contrast, accusative case is not marked on a bare indefinite NP (with or without adjectival modifier), an NP with the indefinite article-like element and ‘one’, or an NP with a weak quantifier like sost ‘three’ or b
izu ‘many’. The generalization seems to be simply that accusative case is marked on DPs but not on NPs. The explanation, I assume, has to do purely with the realization of case morphology. The accusative case assignment rule in (28) applies to all nominals in the syntax in Amharic, but the feature [+acc] is spelled out morphologically as -n at PF only on heads of category D. On heads of category N, [+acc] is spelled out as ∅, not different from a noun bearing the feature [+nom]. In this respect, Amharic is rather like German, where morphological case distinctions have been lost on N but not on D; see also Legate 2008 for discussion of the homophony at PF of nominative and accusative on some categories but not others in various Australian languages. On this analysis, then, bare NPs do not manifest accusative because they have no D, and they do not trigger object agreement because they are barred from undergoing (A-movement-type) object shift out of VP on semantic grounds. This accounts for the largely illusory correlation between not being marked accusative and not undergoing object agreement seen in (3).
Using (28) rather than Chomsky’s Agree to assign accusative case in Amharic pays one more significant dividend. This concerns case assignment in passive and unaccusative clauses. On the standard Chomskyan view, one does not expect to find (structural) accusative case assigned in such clauses, because accusative case is assigned by active v, and passive and unaccusative clauses do not contain that sort of v: they either contain a nonactive v, or no v at all. In contrast, the dependent case view in (28) expects accusative case to be assigned in passive and unaccusative clauses if and only if the clause has two NP arguments—both of which would be internal arguments (nonagents), by hypothesis. In fact, the second prediction is correct. Accusative case is impossible on the one remaining argument in the passive of a monotransitive verb in Amharic (see (29)), or in the simplest unaccusative clauses (see (30)).
This is expected on either account. But if an unaccusative verb has an experiencer argument as well as a theme argument, accusative case is automatically possible on that argument, as in (31).
Amharic differs in this respect from (for example) Korean, where in the passive of a double object construction, both internal arguments can shift to nominative case. This is additional support for the idea that accusative case is not the result of Agree in Amharic—at least not Agree with transitive v.16
Not yet explained by (28) is the intriguing matter of which of the two internal arguments of the verb gets accusative case. The issue is far from trivial, because different predicates behave differently: in (31) and (32) it is the affected argument/goal that receives accusative case, not the theme, whereas in (33) it is the theme argument that receives accusative. There are also differences among these examples with respect to agreement: in (31) and (32) the affected argument triggers object agreement, whereas in (33) the affected argument triggers subject agreement. To add to the mystery, accusative case marking is optional in (31) and (32), but required in (33). Although space precludes my giving a detailed analysis of these matters here, I sketch where I believe the answers to lie; see Baker, to appear a, for fuller discussion.
The pattern in (33) is the simpler one, about which nothing special needs to be said. We know that the source argument is higher in the VP than the theme argument, because it rather than the theme is the target of object agreement in active sentences (see (9)). It is thus expected that in the passive version, when there is no overt agent generated in Spec,vP, the source argument is the closest target for T to agree with as well. Hence, subject agreement is with the source in (33). Similarly, if either NP moves to Spec,TP to satisfy the EPP feature of T, it will be the source, since it is the higher one. The theme NP thus is lower than the source NP at every point of the derivation, both before and after any NP-movement. Hence, it is the theme argument that is assigned accusative case by (28). Finally, F and T never agree with the same argument in Amharic, and F cannot agree with the theme in (33) because the source intervenes between them. Hence, no object agreement is manifest in this structure. The derivation is summarized in (34).
The structures for which something more needs to be said are (31) and (32). Here too, we have reason to think that the goal/affectee argument is structurally higher than the theme, because object agreement favors the goal/affectee in transitive sentences (see (16)). This is also the more normal situation crosslinguistically, and it can be supported by tests involving bound variable anaphora (Baker, to appear a). However, the goal for some reason resists being the target of agreement from T in passive or unaccusative clauses, when there is no agent for T to agree with. To account for this, I propose that the goal/affectee argument (but not the source argument) is embedded inside a PP headed by a null P—a claim that has been made for goal and experiencer arguments in other languages by Landau (2007:502, 2010) (see also Baker 1997 and references cited there). Following Landau (2007), I assume that this PP superstructure makes it impossible for the goal argument to satisfy the EPP property of T, since the EPP cannot be satisfied by a phrase with a null head. I further assume that T also cannot enter into an Agree relation with the goal argument, because T’s agreement properties and its EPP feature are interrelated: T cannot agree with NP X if something other than X satisfies its EPP property. (This is an Amharic-specific property, found also in Bantu languages (Baker 2003, Carstens 2005) but not in English, Icelandic, and so on.) This implies, then, that T cannot agree with the goal argument. F, however, can agree with the goal since F does not have an EPP feature to worry about. The result is that the goal or affectee argument triggers object agreement but not subject agreement in passive and unaccusative clauses, just as in an active clause. T then agrees either with the theme argument (if the goal doesn’t count as a defective intervener) or with nothing (if it does), and the EPP feature of T is satisfied either by moving the theme to Spec,TP or by inserting a null expletive in Spec,TP. I assume that both options exist side by side in Amharic, as in (35) and (36).17
Finally, (28) applies to assign accusative case in these structures. In (35), it does not in fact apply: the goal is not c-commanded by any argumental NP (but only by the expletive subject), and neither is the theme, given that the goal is embedded inside a PP. In contrast, in (36) the goal argument is c-commanded by the theme argument once the theme argument has moved to Spec,TP; hence, (28) applies to assign the goal accusative case. Given that both structures are possible, we derive the result that in these constructions, the goal is sometimes assigned accusative case whereas the theme never is.18
Although this sketch of an analysis raises further questions, it goes far enough to show that the details can plausibly be filled in. The upshot is that (28) is adequate to explain the distribution of accusative case in Amharic, even when there is no active v to assign case in the Chomskyan way, and when accusative case and object agreement do not necessarily line up.
In this article, I argued that accusative case and object agreement cannot be two realizations of the same abstract Agree relation in Amharic. This is clear from the fact that there are systematic situations in which an NP bears accusative case but does not trigger object agreement, and others in which an NP triggers object agreement but does not bear accusative case. I went on to argue that, of the two, it is object agreement that has the properties one expects of an Agree relationship: it is unique, probes downward, is subject to the intervention condition, and respects a phase-like domain condition. This leaves accusative case to be the morphological realization of some other relation. I showed that it is an instance of dependent case assignment, in the sense of Marantz 1991 and Baker and Vinokurova 2010.
How far-reaching is this result? Is the nondependence of accusative case on agreement an idiosyncratic property of Amharic, or is it a parametric property of a broad class of languages, or is it a universal property of all natural human languages? Clearly, a full answer requires serious research on many other languages and thus goes far beyond the scope of this article. However, we can guess that it is at least a parametric property, not a totally idiosyncratic one, since a very similar rule of accusative case marking is valid for the unrelated (Turkic) language Sakha. This also recalls a result of my (2008) study of parameterization in the syntax of agreement. There I found that agreement is closely related to case in one broad class of languages, but not in another (the Case-Dependence of Agreement Parameter (CDAP)). One might reasonably conjecture, then, that rules of dependent case marking are the source of structural case marking in languages where the CDAP is set to “no,” a parametric property.
If that is so, then we should also expect to find languages in which both object agreement and accusative case are overt and where the two do line up nicely—where an NP bears accusative case if and only if it triggers overt object agreement on the verb. In fact, such languages do not seem to be abundant. But one possible case might be the Australian language Mangarayi (Merlan 1982): this is the one fairly clear instance of a CDAP = “yes” language that has both overt object agreement and overt accusative case marking found in my 2008 survey. Another case could be Nez Perce, under Deal’s (2010) analysis (not mine (2008)), where accusative case is found on the object if and only if the verb agrees with that object.
For now, then, the best bet seems to be that accusative case assignment is parameterized, happening via agreement in some languages and via a rule of dependent case assignment in others. And, as we seek to clarify the typological picture further, we can add Amharic to the small but growing list of languages for which there is good evidence that accusative case assignment is independent of object agreement.
Much of the material reported here was developed in the context of a field methods class taught at Rutgers University in the spring of 2010. I thank the other participants of that class—Akin Akinlabi, Will Bennett, Carlo Linares, and Teresa Torres Bustamente—for help and discussion, and especially to Yetnayet (“Mimi”) Lemma for cheerfully sharing her native-speaker judgments with us. Most of the data reported here was collected directly by us in this class, but similar data can be found in works by real Amharicists such as Wolf Leslau and Mengistu Amberber, in the references cited. Special thanks are due to Ruth Kramer and Mengistu Amberber, for their comments on this work and ongoing discussions of Amharic syntax and morphology, and to two anonymous reviewers. Any errors of fact or interpretation are my responsibility.
Abbreviations used include acc, accusative; aux, auxiliary; dat, dative; def, definite; fem, feminine; foc, focus; ger, gerund; neg, negative; nom, nominative; pass, passive; pf, perfective; pl, plural. Agreement markers are glossed with a complex symbol consisting of a number expressing the person of the agreed-with nominal (1, 2, or 3), a lowercase letter expressing the gender/number of the agreed-with nominal (m, f, or p), and an uppercase letter expressing the type of agreement (S, O, or P(ossessor)). Abbreviations for sources of Amharic data are MA02, Amberber 2002; MA05, Amberber 2005; WL, Leslau 1995.
1 Nominals referring to females are feminine in Amharic, whereas nominals referring to males or inanimate objects like ‘book’ are almost always masculine. Aster and Almaz are female names; Lәmma is a male name. For details of the gender system of Amharic, see Kramer 2009.
2 My consultant strongly disprefers examples like (12), however. She favors an alternative version in which the preposition bә is omitted, and ‘broom’ is marked accusative (also given in Amberber 2002:56). One possible way of thinking about examples like (12) is that -bb- on the verb is a manifestation of F agreeing with the PP as a whole. This unusual agreement with PP then makes the PP transparent to further agreement, so that F can then agree with the object of P—just as it does with the object of a null P in my analysis in (35) and (36) below. I cannot pursue this or any other analysis of this construction here, however.
Kramer (2010) reports an instance of a verb agreeing with the object of kә ‘from’, which does not seem to be a case marker, nor part of an applicative construction. But such examples seem uncommon, and my consultant does not always accept them, so I do not analyze them here.
3 I thank an anonymous reviewer who raised (but did not recommend) this possibility.
4 Note that there is no obvious paradigmatic or phonological motivation for this gap in Amharic. The OM forms are transparent enough that one can easily predict what a cluster of them would look like, and such clusters are not phonologically ill-formed in any way.
5 It is not impossible for a single functional head to agree more than once. However, this property is parameterized: see, for example, Baker 2008:99–102 and Nevins 2010. My agreement view thus does not force the OM to be unique in Amharic in any deep way. But functional heads that agree only once are known to be common, so there is a ready account within these terms.
6 This is my interpretation of an otherwise somewhat anomalous example in WL:191. It is notable that the relevant example of agreement with the theme in the presence of a goal has theme-goal-verb order, not goal-theme-verb order. This is a possible further sign that the goal is a PP generated lower in the structure than the theme in this example. I have also collected two examples of this kind—but only with the verbs ‘send’ and ‘introduce’, which are not really double object verbs in Amharic, instead taking only an NP-PP-V frame (cf. (18)). No source I am aware of reports agreement with the theme when the goal is marked accusative, and this is never accepted in my data (in eight attempts). It is not surprising that these results are sharper, because accusative case cannot be alternatively analyzed as a P as dative can. The fact that the OM in (16) represents the goal and not the theme cannot be explained entirely in terms of animacy, because the same asymmetry shows up in an example like (11), where both theme and goal are animate. This example is judged ungrammatical with the verb form assajj-әhw-әt ‘show-1sS-3mO’; see also (8).
7 Some varieties of Spanish do allow clitic-doubled objects to stay inside the core clause, but “clitics” in such varieties have often been analyzed as agreement rather than as moved pronouns.
8 It is conceivable that one could deny this, holding that the OM is agreement on experiencer verbs but that it is a cliticized pronoun on transitive verbs. But this hybrid position would come at a significant cost, given that the same stock of morphemes spells out both, both appear in the same place in the morphological structure of the verb, and both count for the generalization that a verb can bear only one OM in Amharic. There is thus no clear empirical reason for treating them differently apart from the data at hand, and there are some reasons not to.
9 I call it a domain rather than a phase, however, because it seems to be both specific to agreement and crosslinguistically variable. As discussed in section 4, case assignment is not limited by VP domains in Amharic, and VP constituents can have different effects on case and agreement in different languages (cf. Baker, to appear b). This is different from Chomsky’s notion of a phase, which should limit all syntactic processes and be relevant to all languages.
10 Further questions arise here. Is F absent in structures where the verb does not agree with an object? Or can F simply fail to agree if there is no accessible NP? Or is there a null default agreement that is used in such cases, distinct from 3mO agreement (∅ instead of -w/-t)? Any answer may require some theoretical elaboration. I tentatively adopt the third option. This is contrary to the proposal of Kramer (2010), who applies Preminger’s (2009a) criterion to Amharic. From the fact that an OM can simply be missing on a transitive verb in Amharic, Kramer infers that the OM is a clitic, not an agreement morpheme, reasoning that if it were an agreement morpheme, a default realization would be expected—probably -w ‘3mO’ given that -ә ‘3mS’ is the default form for subject agreement in Amharic. I take this consideration to be relevant, but not decisive. It is known that some languages allow default subject agreement to be morphologically distinct from third person subject agreement (e.g., Ukrainian). It is also common for the least marked category in a paradigm to have a zero exponent. Combining these two observations, nothing rules out languages in which default object agreement is different from third person object agreement, and happens to be phonologically null. I tentatively assume Amharic is such a language.
It is an open question whether Preminger’s criterion as formulated will ever identify object marking as agreement rather than clitics, and this seems suspicious. Here is a conjecture about why. All languages have verbal forms that register the subject but not the object, because all languages (I assume) have a robust class of intransitive verbs. In contrast, many languages have no finite verbal forms in which the subject is not marked one way or another, because most languages have few or no zero-argument verbs. It seems plausible (to me) that this asymmetry causes children to conclude that there is a null default form for object agreement in many languages, whereas null default subject agreement is rarer (unless ordinary third person agreement is also null). If so, Amharic (under my analysis) is not as strange as it might seem.
11 For definite versus indefinite NPs, we can adopt the standard Diesing (1992) mapping hypothesis that says that the final position of the NP determines whether it is inside the domain of existential closure or not. Perhaps quantified NPs cannot undergo object shift because this would create an environment where the variable bound by the quantifier after quantifier raising would be in a nonargument position (cf. Rizzi 1986). For reflexive anaphors, I tentatively assume that the body part noun ras needs to adjoin to the verb at LF so as to create a reflexively marked predicate, as in Reinhart and Reuland 1991:291–292 and subsequent work by Reuland. This head movement would be blocked if the reflexive NP were to move out of the VP.
Note that examples like (19) suggest that either the landing site of object shift is quite low in Amharic, or the position of adverbs is relatively high (or both).
12 The question arises as to where goal and source arguments of verbs like ‘give’ and ‘rob’ are generated. Are they like experiencers in being outside VP, so always accessible to agreement with F, or are they like theme arguments in being inside VP, not accessible to agreement with F unless a semantically relevant movement takes place? In fact, my data on this are mixed, and speakers are somewhat uncertain about the judgments. For example, Mengistu Amberber ( pers. comm.) reports that an OM seems necessary with a referential source argument in (i), but sounds strange with a quantified source argument. This pattern is intermediate between that of experiencer arguments (where the OM is always needed) and that of theme arguments (where the OM is always optional, sometimes judged unacceptable). I tentatively assume that goals and sources can be generated either in Spec,ApplP or inside VP in Amharic, and this structural ambiguity makes speakers’ judgments variable and uncertain. But more thorough study of this matter would be welcome.
in / mann- in gәnzәb-u-n sәrrәk’-ә-(#at).
Lemma Aster-acc who-acc money-def-acc rob-3mS-(3fO)
‘Lemma robbed Aster of the money.’/‘Who did Lemma rob of the money?’
(OM -at is needed with source Aster, is judged marginal or unacceptable with source ‘who’)
13 According to my data, an indefinite theme argument can precede a goal in dative case in simpler examples where there is no issue of binding a pronoun. However, one cannot construct a direct comparison with this example, showing that shifted definite NPs both trigger object agreement and bind pronouns, simply because a definite NP is not the sort of expression that is subject to weak crossover. A ( perhaps simpler) alternative to the view in the text would be that definite NPs can move out of VP in the syntax, but indefinite NPs can do so only at PF—too late in the derivation to feed Agree in the syntax or to create binding relations for LF.
14Kramer (2010) uses this as another argument that OMs are clitics in Amharic, because they do not vary with tense or any other verbal category the way that subject agreement often does—a type of consideration also raised explicitly by Nevins (2010). But like Preminger’s (2009a) criterion, this (in Nevins’s version, anyway) seems more applicable to subject markers than to object markers. If subject agreement is agreement on T, it stands to reason that its realization will often depend on the specific T used. But if object agreement is usually agreement not on T but on some other head, it makes sense that its realization will not depend (as much) on that of T. If anything, object marking should vary with different overt choices of transitive v (or AgrO)—and there happen to be few or no morphologically distinct choices of v to condition variation.
15 An anonymous reviewer understandably would like a more principled account of why agreement seems to be sensitive to VP domains but case marking is not. So would I. But no deeper account is within my grasp right now. Whether dependent case assignment is sensitive to VPs or not seems to be a point of raw variation across languages, accounting in part for the fact that some languages have differential object marking (like Sakha) and others do not (Amharic); see Baker, to appear b, for other relevant languages. Baker, to appear b, also shows that the same parameter is relevant to languages with ergative case. I have fewer data on agreement, but what I have suggests that the impact of VP for Agree varies too, and varies independently of case. For example, VP restricts both case and agreement in Nez Perce, VP restricts case but not agreement in Hindi, VP restricts agreement but not case in Amharic, and VP restricts neither case or agreement in Mangarayi. That is probably not the last word on the matter, but to take it further will need both careful empirical work and careful theoretical reflection.
16 We saw above that object agreement is also possible in passive and unaccusative clauses in Amharic. So these facts show that accusative case does not depend on active v in this language, but it could conceivably be that accusative case depends on object agreement with another head, called F above (tentatively AgrO). But the problem would still remain that object agreement and accusative case are not tightly correlated. For example, agreement on F with the theme is impossible in (33) (by the intervention effect) even though it bears accusative case.
17Preminger (2009b) discusses a very similar range of possibilities in Hebrew (another Semitic language). In Hebrew, the optionality of moving the theme to subject position, and its consequences for agreement, are easier to see because Hebrew’s head-initial word order makes it obvious whether something has moved to Spec,TP or not. See Baker, to appear a, for more on how the various options of case, agreement, and word order correlate with each other.
18 This account extends to the optional use of accusative on the experiencer argument of seemingly monadic psych predicates like the one in (13) if one assumes that these have a null cognate argument in the theme position as well as an experiencer argument. In fact, Mengistu Amberber ( pers. comm.) points out that this cognate argument can be realized overtly for many of the psych predicates. For example, alongside (13), it is possible to say Rab rab-ә-лл (‘I’m hungry’, lit. ‘hunger hungered me’); see also Leslau 1995:435. This example has the same analyses as (31), and I assume that (13) is the same with the noun ‘hunger’ pro-dropped.