Partitivity and case marking in Turkish and related languages

The paper discusses the conditions for case marking on partitive constructions in direct object position in Turkish and some related languages. We focus on Turkish and then turn to some details of corresponding constructions in some other Turkic languages and in Standard Mongolian. Turkish exhibits Differential Object Marking, which primarily depends on the semantic-pragmatic factor of specificity. Partitive constructions with the ablative for the superset in Turkish come in different forms, depending on how the subset expression is realized: (a) by a lexical noun as head, (b) by the classifier tane ‘ item’, functioning as a “dummy noun”, and (c) by a numeral, quantifier or adjective. Case marking of the direct object is optional for (a), and obligatory for (most instances of) (c). This type of obligatory case marking is dependent on the obligatory marking of the adjective, quantifier or numeral with a default 3rd person singular agreement suffix, which then requires case marking. Construction (b) does not allow for case marking, when the classifier is bare; when the classifier is followed by the default 3rd person singular agreement marking, that marking requires obligatory case morphology, just like in construction (c). We hypothesize that structural case marking can either express the semantic-pragmatic condition of specificity in terms of referential anchoring or it must obey a formal condition, namely the requirement of the agreement suffix to be followed by overt case. The languages we have studied show an interesting micro-variation. They differ (among other properties) with respect to classifiers – in particular, with respect to whether they have [+human] classifiers or not. In addition, one language among the languages under investigation, namely Kirghiz, substitutes the agreement marker in its function as a filler of the partitive’s nominal head by a different marker: a morpheme expressing a set. Here, the agreement marker is used to express specificity, given that its presence is not required for formal reasons. In direct object partitive constructions with subset expressions that are expressed as full noun phrase/lexical noun heads (option (a) above), overt accusative case indicates specificity in most of the investigated languages. In options (b) and (c), the investigated languages provide different patterns when marking the referential status of the partitive heads, thus indicating the variation among these languages with respect to the nominal category feature of the partitive heads involved.


Differential Object Marking
Turkish is a nominative-accusative language with case suffixes.It shows Differential Case Marking for the direct object (with respect to its accusative marking) as well as for the subject in most types of nominalized embedded sentences (with respect to its genitive marking).Differential Object Marking (DOM) follows syntactic, morphological and semantic-pragmatic conditions.A case-marked direct object in the immediate preverbal that case signals specificity, which, according to her view, is based on partitivity.She illustrates this claim by offering examples, which we repeat as (2).(2a) introduces a set of children, out of which the case-marked direct object iki kızı in (2b) selects two girls.In other words, the specific direct object iki kızı is an implicit partitive, and the specificity is explained by the discourse givenness of the set out of which the indefinite direct object selects one element (i.e.here, a subset consisting of two entities).The unmarked direct object iki kız in (2c), however, is not linked to the set of children, i.e. it refers to a set of girls not included in the set of children introduced in (2a): (2) a. Enç (1991: example 16; Enç's translation, our glosses) Oda-m-a birkaç çocuk gir-di.room-1.sg-datseveral child enter-pst 'Several children entered my room.' b.Enç (1991: example 17; Enç's translation, our glosses) İki kız-ı tanı-yor-du-m.two girl-acc know-prog-pst-1.sg'I knew two girls.'c.Enç (1991: example 18; Enç's translation, our glosses) İki kız tanı-yor-du-m.two girl know-prog-pst-1.sg'I knew two girls.'Enç (1991: 10) argues, based on (3), that case marking is obligatory not only for implicit partitives, as in (2b), but for explicit partitives, as well.The numeral ikisini in (3a) exhibits an agreement marker -(s)I(n) as well as the accusative marker -(y)I1 , while the form ikisi without case (but with the same agreement marker) is ungrammatical, as seen in (3b).
(3) a. Enç (1991: example 29a; Enç's translation, our glosses) Ali kadın-lar-dan iki-sin-i tanı-yor-du.Ali woman-pl-abl two-3.sg-accknow-prog-pst 'Ali knew two of the women.' b.Enç (1991: example 29b;our glosses) *Ali kadın-lar-dan iki-si tanı-yor-du.Ali woman-pl-abl two-3.sgknow-prog-pst To summarize, Enç (1991) argues that case expresses specificity and is based on partitivity.She argues that case marking of an indefinite direct object always signals a partitive reading, which has to be interpreted as specific, and that likewise a specific object is partitive and therefore must be overtly marked as accusative.Öztürk (2005) even goes a step further in assuming that overt case is the bearer of referentiality; while we agree with the judgments in (2) -(3), we disagree with both authors on their analyses and show in this paper why neither of these views can be correct, based on a crosslinguistic evaluation of Turkic data as well as of data from some related Altaic languages, with a special focus on Turkish.While Enç's approach was an important step forward in understanding the syntax and semantics of case in Turkish, there are some important modifications to be made.The next section will show that, firstly, the semantics of case in terms of specificity cannot be reduced to partitivity; secondly, the observation that the form in (3b) without case is ungrammatical, while being correct, does not show that each and every partitive construction must have case.In von Heusinger & Kornfilt (2005) we argued that the ungrammaticality of (3b) is due to a formal requirement imposed by the agreement marker -(s) I(n).This leads us to the main question of this article, namely how case marking varies with the type of partitive construction.

Specificity in Turkish as referential anchoring
DOM in Turkish was associated with the semantic-pragmatic notion of specificity from early on (Sezer 1972;Johanson 1977;Erguvanlı 1984;Dede 1986), but without a clear semantic definition of specificity.Enç (1991) was the first to provide a definition of specificity as partitivity: A partitive noun phrase picks out one or more referents from a discourse-given set.Therefore, the referent -even if expressed in an indefinite form -is already introduced in the discourse and shows similar behavior to definite noun phrases: it is presuppositional, it has wide scope etc.Thus partitives can be viewed as one subkind of specific indefinites (see Farkas 1994), but there are other subkinds of specificity such as referential specificity, epistemic specificity and scopal specificity (for an overview, see Farkas 1994;von Heusinger 2011).There is no consensus on one characterization of specificity, but according to one prominent family of approaches the underlying concept of a specific indefinite is the "referential intention" of the speaker.The speaker uses a specific indefinite for a particular discourse referent known to the speaker himself/herself, but unknown to the hearer.In using a specific indefinite the speaker signals that he or she will continue to talk about a particular referent and that the hearer should have a consistent representation of this referent in his or her mind.From this foundational concept, we can derive many properties of specific indefinites: they are referential, they have wide scope, they signal the certainty of the speaker about the identity of the referent, they license discourse anaphora, they show discourse prominence etc.This very general concept is the base for referential, scopal and epistemic specificity, but not for specificity based on partitivity.
After the proposal of Enç (1991) and related proposals in the literature it was shown that partitivity is orthogonal to scopal specificity, as in (4), and epistemic specificity, as in (5). (4) John wants to marry one of Steve's sisters.('He doesn't care which.')(5) One of Steve's sisters cheated on the exam.('We have to find out which.') The partitive indefinite one of Steve's sisters has narrow scope with respect to the intensional verb want in (4) and it has an epistemically non-specific reading in (5).These facts also apply to Turkish and there is an extensive discussion about the particular type of specificity expressed by (structural) case in Turkish (for an overview, see von Heusinger & Kornfilt 2005;Öztürk 2005).In addition to directing the reader to the extensive discussion in Kelepir (2001), Öztürk (2005) and Özge (2011), we would like to add the following argument that goes back to Higginbotham (1987), who argues that typical specific indefinite uses can be accounted for by the concept of "the speaker has in mind": Suppose my friend George says to me, 'I met with a certain student of mine today.'Then I can report the encounter to a third party by saying, 'George said that he met with a certain student of his today,' and the 'specificity' effect is felt, although I am in no position to say which student George met with.(Higginbotham 1987: 64) We find the case suffix in the Turkish equivalents of Higginbotham's two examples.This indicates that specificity cannot be understood as "the speaker knows the referent" or "the speaker has the referent in mind", but rather in a more abstract way: "the referent is referentially anchored to some salient discourse item" (for a more detailed description of this concept see von Heusinger 2002;von Heusinger & Kornfilt 2005).In (6) the specific indefinite bir öğrencimi 'a student of mine + acc' is licensed by the discourse item Ali as a speaker of the sentence, while in (7) it can either be licensed by the subject (Ali) or by the speaker (Osman).( 6) Ali: "Kütüphane-de çok başarılı bir öğrenci-m-i gör-dü-m."Ali library-loc very successful a student-1.sg-accsee-pst-1.sg'Ali: "I saw a very successful student of mine in the library."'(7) Osman: "Ali kütüphane-de çok başarılı bir öğrenci-sin-i gör-müş."Osman Ali library-loc very successful a student-3.sg-accsee-rep.pst'Osman: 'Ali (reportedly) saw a very successful student of his in the library."' In the reminder of this article we will account for specific indefinites by the concept of referentially anchored indefinites, i.e. by indefinites that introduce a particular referent that is known to a salient referential anchor in the discourse.This is very often the speaker, but it can also be the subject of a speech act verb, as in ( 7).This concept is quite flexible, and we will see below that it will be crucial in explaining the data from Sakha.

Partitive constructions
Partitives are constructions of the type NP1 of NP2 (in non-head-final languages) such that the whole construction is indefinite while the superset NP2 must be definite as in (8a).NP1 is the subset and NP2 the superset.Often NP1 consists only of the determiner as in (8b) or of further modifiers, but without the head noun, as in (8c).We distinguish cardinal partitives (8a-c) from measure partitives (8d), fraction partitives (8e) and vague measure partitives (8f) (Hoeksema 1996;de Hoop 2003;Ionin et al. 2006;Koptjevskaja-Tamm 2006).
(8) a. two girls of the children b. two of these eight girls c. two intelligent (ones) of these eight students d. two liters of water / five feet of snow e. three quarters of the cake / the beans / my friends f. a number / lot / bunch of cats / my friends Partitive constructions are semantically characterized by the conditions listed in (9) (Hoeksema 1996;Chierchia 1997;Barker 1998;Zamparelli 1998).
(9) Semantic conditions for partitives (i) NP1 must be indefinite (with certain exceptions) (ii) NP2 must be definite (or specific) (iii) NP2 must be plural (if NP2 is headed by a count noun) (iv) the expressed relation is a part-of relation In English, the part-of relation is encoded in the preposition of, which is of type 〈e,〈e,t〉〉, i.e. it is a normal transitive preposition that takes an entity (e.g. the children) and yields a set (e.g. of the children), which is further restricted by NP1 (e.g. two girls) (Ionin et al. 2006).Some languages can also express the partitive relation via the genitive, as in German.This seems strongly preferred if the lexical noun of NP1 is not realized, as in (10b-c): (10) a. ?zweiMädchen der Kinder two girls the-gen children 'two of the children' b. zwei dieser acht Mädchen two these-gen eight girls 'two of these eight girls' c. zwei kluge dieser acht Studierenden two intelligent these-gen eight students 'two intelligent (ones) of these intelligent eight students' For languages such as English one has to distinguish carefully between the use of of for partitives (11a) and for possession (11b) (Barker 1998: 683) The syntactic structure of partitives has not been very widely addressed in the literature.For English partitives, Ionin et al. (2006) introduce two possible structures, both headed by N, i.e. the partitive's subset (which represents a measure noun or a cardinal), and with the expression of the superset being either a complement PP of the head (for measure nouns or fractions, e.g.'half of these eight apples'), or with the superset expression as an adjunct of the head N (for cardinals, e.g.'two apples of these eight apples'). 2The authors further argue that the semantics of partitives are compatible with either one of these structures.
Turning to explicit partitive constructions in Turkish, we focus our attention on the ablative partitives (some information about different types of partitives in Turkish and a justification of our choice to focus on ablative partitives are given in the next subsection).We analyze this construction as consisting of a DP/KP realizing the superset which is adjoined to another DP which represents the subset, and is the head of the entire construction.Thus, we would have, schematically, the following structure: '(ø) two (pieces/items/∅) of these eight apples' Some traditional grammars such as Lewis (1967;1975), as well as some older literature on Turkish case (e.g.Dede 1981) have viewed the ablative superset expression as an adjunct of the verb, i.e. crucially as not even forming a constituent with the subset expression.We obviously do not share this view and refer the reader to Kornfilt (1984), where the structural status of ablative partitives as one single constituent is argued for. 3 Before turning to a discussion of types of partitive constructions, we would like to offer a brief sketch of the NP/DP structure we are assuming, so as to make the (ablative) partitive 2 For the convenience of the reader, we have italicized the superset expressions in these examples. 3Bošković & Şener (2014) claim that nominal phrases in Turkish are NPs rather than DPs.One reviewer of this article raises this alternative analysis, as well.However, the claims in Bošković & Şener are controversial; we refer the reader to Kornfilt (forthcoming), where arguments are offered against Bošković & Şener's claims.More importantly, for our purposes and claims in this article, the choice between NP and DP for Turkish nominal phrases in general and for partitive phrases in particular is not relevant; we therefore continue to use the DP designation for them.
structure we presented in ( 12) clearer.This is of particular interest with respect to the subset expression.

A sketch of the Turkish DP
Turkish is a head-final language; we take this generalization seriously and thus assume that all projections within the DP/KP are head-final.We assume the existence of a phrasal architecture for the Turkish DP similar to what has been posited, in one way or another, for other languages, too (e.g. by Simpson 2005 for Southeast Asian languages; Watanabe 2006 for Japanese; and by Borer 2004 and others for a variety of languages): a lexical core (NP), which is the complement of a Cl(assifier) Phrase, which in turn is the complement of a Numeral Phrase (which we represent as #P); #P can contain either a numeral or a quantifier in its specifier position, and is headed by a #-head, which can be occupied by the plural morpheme, when present; #P is the complement of DP, which, in turn, is the complement of KP, a Case Phrase: (13) KP < DP < #P < ClP < NP This schema (whereby < reflects relative height among projections, but not directionality with respect to the heads of those projections) is instantiated in the example below (where the DP would be a direct object and thus bear accusative case): I-gen three item rotten grape -1.sg -acc 'my three rotten grapes -acc' Note that the plural morpheme on the head noun doesn't show up when the quantifier is a numeral, as in ( 14); we explain this generalization by positing an anti-redundancy principle for some projections, the #P being one of them: when the specifier is occupied by numerals or certain quantifiers which express plurality, the head cannot overtly express plurality at the same time and thus must take on the unmarked, i.e. singular, shape; cf.Kornfilt (1996).This principle is implemented via a specifier-head (anti-) agreement process and supports our analysis that views the overt numeral as the specifier (rather than the head) of #P, and the plural marker as the head of #P.As also mentioned below, the nominal agreement marker is in D 0 (following Abney 1987), while the case marker is in K 0 .A corresponding example with an overt plural morpheme and a non-numeral quantifier follows: I-gen some rotten grape -pl -1.sg -acc Lit.: 'my some rotten grapes -acc', i.e. 'some of my rotten grapes'4 Please note that the functional heads which we posit in this DP-architecture are all occupied by bound morphemes, whenever such morphemes do show up.The overt quantifiers, numerals, and the classifier, which are all free morphemes, as well as the possessor (also morphologically free) are all in the specifier positions of their respective functional projections, rather than in head positions.We propose that this is an unmarked option for morphology-rich languages such as Turkish.Future research will show to what extent this proposal has cross-linguistic validity.
The hierarchical order we have presented here, as schematized in (13) and illustrated in ( 14) and ( 15), is widely, but not unanimously, assumed in the literature.Simpson (2005) argues, on the basis of syntactic and semantic considerations, that the classifier and the # have to be separate and independent heads.The semantic considerations, obviously of more cross-linguistic relevance, state that individuation and numerical specification are distinct functions.Given that numerical specification partitions the assemblages of individuated items, it makes sense that #P is higher than ClP.We further observe that numeric and quantificational expressions precede the classifier (in the languages studied by Simpson, as well as in Turkish and related languages); this order is obligatory.Thus, the hierarchy proposed here is supported by the facts. 5As for the relative height of KP and DP: Watanabe (2006) assumes the converse order to the one we propose, but he does point out that Bittner & Hale (1996) posit KP < DP.That latter order is one which we assume, as well, given the relative order of case morphemes and of (nominal) agreement (when present): case morphemes follow agreement, as illustrated in ( 14) and ( 15).We assume that nominal agreement morphemes are situated in D; note that this is how Abney (1987) transposes the proposal in Kornfilt (1984) for a nominal AgrP into his DP-based account, with Kornfilt's Agr-head viewed as the D-head of DP.Agreement morphemes follow the plural morpheme, when the latter is present, thus justifying the hierarchy above.
The assumptions that the numeral or quantifier, and the classifier, are specifiers of their respective projections, rather than their heads, together with the hierarchy stated above, result in the correct order of all the relevant morphemes6 , and at the same time obey the typological and phrase-structural characterization of Turkish (and its DP) as head-final.This approach allows us to state co-occurrence restrictions in two familiar ways, depending on the particular projection and the restriction involved; e.g. the ban against cooccurrence between numerals (and certain other quantifiers) and an overt plural marker on N can be stated with respect to the familiar specifier-head relationship, as mentioned earlier; the obligatory order of a numeral or quantifier before a classifier (remarkable in a rather word-order free language such as Turkish) is expressed via the relative height of the #P as being higher than the ClP.
The analysis presented here gives rise to an understanding of interesting phenomena, when the lexical head N of the core NP is not overt; we shall address those in the remainder of the paper, in the various sections where the functional expressions mentioned here are studied.We now turn to a brief presentation of an inventory of explicit partitive constructions in Turkish.

A first typology of partitive constructions
As mentioned earlier, we focus on Turkish partitive constructions of the form NP2.abl NP1, i.e. superset-abl subset in direct object position, as we are interested in the function of the acc case marker in partitive constructions.Partitive constructions in Turkish come in different forms, depending on how the subset expression is realized; we illustrate these options in the example set ( 16), where we provide a context according to which the particular sentences should be interpreted.The subset expression is realized by a lexical noun as head in (16a), by the classifier tane 'item', preceding a lexical noun in (16b), by the classifier tane 'item' by itself in (16c), by the same classifier tane 'item', when it is followed by a "dummy nominal agreement" in (16d), by a numeral, quantifier or adjective likewise followed by a "dummy nominal agreement" in (16e), and by an unrealized subset expression (16f).( 16) Context: My mother always fills a big bowl with different pieces of apples, pears, and bananas.Yesterday evening I was intensively studying the different pieces of fruit, which were 8 apples, 10 pears and 4 bananas, and then… 7 a.Meyve-ler-den üç elma(-*sın)(-yı) ye-di-m.fruit-pl-abl three apple(-*3.sg)(-acc) eat-pst-1.sg'I ate three apples of the (set of) fruits.' ([-specific] with no overt accusative; [+specific] with overt accusative) b.Meyve-ler-den üç tane elma(?-yı) ye-di-m.fruit-pl-abl three item apple(-?acc) eat-pst-1.sg'I ate three apples of the (set of) fruits.' ([-specific]; with acc [+specific]) c.Meyve-ler-den üç tane(*-yi) ye-di-m.fruit-pl-abl three item(-*acc) eat-pst-1.sg'I ate three [non-specific] (entities) of the (set of) fruits.' (This reading obtains for the version without the accusative marker.)Intended reading of the ill-formed version with the accusative marker: 'I ate three [specific] (entities) of the (set of) fruits.' d.Meyve-ler-den üç tane-sin-i ye-di-m.fruit-pl-abl three item-3.sg-acceat-pst-1.sg'I ate three (specific or non-specific entities) 8 of the (set of) fruits.' e. Meyve-ler-den kırmızı-sın-ı / altı-sın-ı / bazı-lar-ın-ı ye-di-m.fruit-pl-abl red-3.sg-acc/ six-3.sg-acc/ some-pl-3.sg-acceat-pst-1.sg'I ate the red (one) / six / some of the fruits (specific for the first example, specific or non-specific for the second and third examples).' 7Two reviewers reported an exhaustive reading for the accusative marked version of (16a), which we presented in an earlier version of the paper without the context preceding the examples.The assumption of those reviewers is that (16a) has an exhaustive reading, i.e. that there were three apples in the bowl and the partitive with case marking refers to the exhaustive set of all three apples.We provided a context that is incompatible with this reading.The context introduces 8 apples out of which 3 are selected.If partitives with case would have an exhaustive reading, examples like (16a) in the given context have to be unacceptable.We tested this with three comparable examples with supersets that had a cardinality clearly higher then the numeral in the subset of the partitive construction.We asked 10 native Turkish speakers (ages 26 to 36) to rate the sentences with and without accusative case marking in the given context on a scale from 1 (unacceptable) to 7 (very good).Sentences with case were rated with 4,1, sentences without case with 4,3.This pilot study confirms our intuition: First, there is no exhaustivity condition on case-marked partitives.Second, there is no difference between acc-marked and unmarked partitives with respect to discourse linking. 8We are indebted to Roberto Zamparelli, one of the editors of this special collection for asking how the specific reading of (16d) comes about, given the lack of specificity of the classifier, and the fact that the "dummy" agreement and the overt accusative are due to formal reasons and thus should not be able to contribute to a specific reading.Our response is that the possible interpretations for the subset expression in (16d) as specific or non-specific shouldn't be understood as a clear ambiguity between two distinct readings, but as an indeterminacy or vagueness.In instances where morphemes such as agr and the overt accusative are present due to formal reasons, the reading with respect to specificity is indeterminate or neutralized.
In this particular example, we can have an interpretation according to which the speaker ate three specific pieces of fruit (either in terms of kinds, i.e. apple, pear, banana, or in terms of items, three particular apples), or it can be interpreted non-specifically, i.e. the speaker ate just a set of three items, say, apples.
f. Meyve-ler-den ye-di-m.fruit-pl-abl eat-pst-1.sg'I ate of the fruits.' (= 'I ate some of the fruits.')Overt accusative case is optional for (16a) and (16b), obligatorily marked for (16d) and (16e), and ungrammatical for (16c) and (16f).The Accusative case marking in (16e) usually depends on the obligatory marking of the adjective (in many instances -we return to this in section 2.4), as well as of the quantifier or numeral with an agreement suffix, which then requires overt case marking.The construction in (16c) does not allow for overt Accusative case, where the classifier is bare, whereas the Accusative is obligatorily marked when the classifier is followed by a default 3 rd person singular agreement marking as in (16d); that marking requires obligatory case morphology.We hypothesize that structural case marking can either express the semantic-pragmatic condition of specificity or it must obey a formal condition, namely the requirement of the agreement suffix to be followed by overt case. 9 We summarize the schematic structures of the examples in ( 16) in Table 1a together with the judgments whether the Accusative is obligatorily, optional or ungrammatical.Table 1b provides information about the grammaticality of the functional elements under discussion.Table 1a: Schematic representation of the ablative partitive construction.
9 Two reviewers informed us about different judgments for the examples in ( 16).Some of their informants did not like any of the ablative partitives while others accepted only some of the constructions.Together with Elif Bamyacı, we constructed a short questionnaire with four different lexicalizations (e.g. three (apples) of the fruits, five (girls) of the students, eight (Mazdas) of the cars, three (cats) of the animals) in the conditions presented in ( 16).The results and their interpretations are reported in the appendix.Our judgments in ( 16) were largely confirmed by this study.We provide full information on the questionnaire and its results in section 5. 10 While the accusative marking is possible on the head N in (16b), it is dispreferred; this marking is fully allowed in this context if N is modified (and thus made more specific or referential), e.g.via adjectives and/ or a relative clause.Turkish can also realize partitives in a construction where the superset receives genitive case.This construction has different properties from the partitive construction with ablatives (see Göksel & Kerslake 2005); it has also a different distribution and is subject to different morphological restrictions: It cannot be used with a full lexical noun, see (17a), which can only express the possessive reading.The (grammatically well-formed) possessive reading 'the fruits own the apples' is pragmatically incoherent, and would be obtainable only if the nominal agreement morpheme as well as the accusative morpheme were both realized.Without those two morphemes, the utterance is ill-formed under any reading.
(17) a. Meyve-ler-in üç elma -sın -ı ye-di-m.fruit-pl-gen three apple -3.sg-acc eat-pst-1.sgWell-formed, but cannot mean: 'I ate three apples of the (set of) fruits.'Genitive partitives always require agreement marking on the subset, which then triggers accusative marking as in (17b-c), thus making the options with both the default third singular agreement marker and the accusative marker the only available ones (for more discussion see von Heusinger & Kornfilt (2005); and more recently Öztürk & Erguvanlı Taylan 2016).( 17) b.Meyve-ler-in üç tane-sin-i ye-di -m.fruit-pl-gen three item-3.sg-acceat-pst-1.sg'I ate three ("items") of the fruits.(lit.'Of the fruits, I ate three "items".')c.Meyve-ler-in altı-sın-ı / bazı-lar-ın -ı ye -di -m.fruit-pl-gen six-3.sg-acc/ some-pl-3.sg-acceat-pst-1.sg'I ate six / some of the fruits.' In this article, we pay particular attention to the ablative partitive construction, mainly because it gives us the best opportunity for cross-Turkic (and even cross-Altaic) comparisons; the genitive partitive receives mixed judgments in some of the languages we investigated, while the ablative partitive appears to be more solidly present in the competence of our native informants. 11 The discussion of the above data illustrating Turkish ablative partitives can be summarized as follows: i) if the subset is a lexical noun no agreement marking is allowed; we observe optionality for case marking, which depends on the semantic-pragmatic function of specificity, as in (16a); if the lexical noun is preceded by the classifier tane meaning '… items of …', the classifier is not case-marked, nor is the lexical noun, unless that noun is modified; cf.footnote 10, and example (16b).
11 Roberto Zamparelli points out the possibility of analyzing these ablative partitives as "among"-partitives, rather than exemplifying a genuine partitive construction.There are two main considerations that argue against such an analysis: 1. Turkish does have an "among"-construction which includes a P-like element, arasından 'from between, from among'.We don't address that construction here.At any rate, the ablative partitives which are studied here do not have this element.2. The superset expression in Turkish ablative partitives can consist not only of a count noun, as in the examples we present in the text, but also of a mass noun; e.g.
(i) Ali şarap -tan iki bardak iç -ti.Ali wine -abl two glass drink -pst 'Ali drank two glasses of the wine.'Clearly, this is not an "among"-construction and cannot be translated as such: '*Ali drank two glasses from among the wine.' ii) if there is no lexical noun in the subset expression, but rather only the classifier tane, Turkish provides two options: The classifier either behaves like a nonspecific lexical noun indicating a portion or quantity of something (such as kilos, Euro etc.), as in (16c), or else it needs the non-alternating third person nominal agreement marker12 and overt Accusative case marker, as in (16d), similar to adjectives and numerals in (16e); iii) if there is no lexical noun, and the subset is expressed just by an adjective, a numeral or quantifier, the non-alternating third person nominal agreement morpheme is obligatory for numerals and quantifiers, and for most adjectives; the agreement necessitates Accusative case marking as in (16e) (for reasons we shall address later on); iv) if the subset is not expressed overtly (i.e.such as in bare or "naked" partitives) no case marking shows up (i.e.neither agreement nor case morphology is possible); there is only a non-specific (mass noun-like) interpretation of the subset, see (16f) 13 .
In this paper, we first elaborate on the Turkish samples given in ( 16) and provide a broad empirical base for the observations given above.We then extend the empirical base to related Turkic languages, namely Azerbaijani, Kirghiz, Uzbek, Sakha and another Altaic language, namely Mongolian.With these comparative data we can show some micro-variation in the structural properties of partitives with a lexical noun, a classifier, a numeral, or an adjective as the subset expression.We will concentrate on the following questions: 1. Do the observations regarding use of partitive constructions for Turkish also hold for the other Turkic and Altaic languages?2. Which expressions need additional marking in order to be licensed as a subset expression? 3. What are the markers that enable such an expression to function as a subset expression? 4.Under which conditions does overt structural case (and, for our current purposes, Accusative marking) signal specificity? 5.In those instances where case marking does not signal specificity, are there other means that take over that function?

Partitives in Turkish
In the discussion above, we have seen explicit partitive expressions with the larger set (or superset) marked with the ablative, and the subset expressed in a number of ways.We saw that one important point emerged from the data: Partitives can have non-specific subsets; more precisely, this is so in examples (16b) and (16c), where the subset expressions are not marked with accusative morphology and are interpreted as non-specific; in (16f), where the subset expression is not overt and is interpreted as a non-specific (and unspecified) noun; and (16d) and (16e) where the option of interpreting the subset expressions as well as the entire partitive construction as non-specific is available, despite the accusative marker that they carry.These facts argue, against Enç (1991), that partitivity and specificity are notions which are, while related, independent from each other.
In addition to these points, we showed that case marking is forced if the agreement marker is used and that in such cases the case marker does not necessarily express specificity.This observation argues against Enç, who assumes that case marking is always associated with specificity (i.e.partitivity), and also against Öztürk (2005), who assumes that case signals referentiality.According to Öztürk, overt accusative case on direct objects instantiates the referentiality of a noun phrase.She thus explains the contrast between a specific direct object (with case, referential) and non-specific direct object (without case, not referential).Contrary to her assumption, however, we show that if direct objects without a lexical head must be marked by the agreement marker and the overt case, then both specific (i.e.referential) and nonspecific (i.e.non-referential) expressions are realized with case, contrary to her assumptions.
In order to elaborate on this perspective, in the following section we focus on explicit partitives with ablative case marking placed on the expression for the larger set in the partitive construction and (i) lexical nouns, (ii) classifiers like tane 'item', (iii) numerals and other quantifiers, and (iv) adjectives referring to a subset of the larger set.

Partitives with a lexical nominal as a subset expression
We start by illustrating Turkish ablative partitives with examples of lexical nouns in the subset expression.In Turkish, there is no difference between human and non-human direct objects with respect to case marking; we therefore use the inanimate direct object elma 'apple' in our examples, covering both animate as well as inanimate objects.In (19a, b) below elma 'apple' is a subset of meyveler 'fruit'. (18) Context: We have a big fruit bowl in our kitchen, which generally has 5-8 apples, 5-8 pears and several bananas.Yesterday evening, I was inspecting the fruit and then (19) a. Meyve-ler-den üç elma ye-di-m.fruit-pl-abl three apple eat-pst-1.sg'I ate three (non-specific) apples from among the fruit.' b.Meyve-ler-den üç elma-yı ye-di-m.fruit-pl-abl three apple-acc eat-pst-1.sg'I ate three (specific) apples of the (set of) fruits.' (= (16a) in its grammatical version.)Example (19a) shows clearly that, contra Enç, partitives can have non-specific subsets, i.e. that specificity can't be synonymous with partitivity.Furthermore, there is no accusative marker on the subset expressions in this example -a marker which is a reliable indicator of specificity in most instances, according to Enç's claims.In contrast, (19b) exhibits accusative marking on the subset expression where the subset expression is interpreted as specific (but not necessarily as exhaustive -see footnote 7).Genitive partitives with lexical nouns as subset expressions, which thus represent properties that are a subset of the larger set's properties, are ill-formed.(For illustration of this generalization, the reader is referred to Kornfilt & von Heusinger 2009.)

Partitives with classifiers as a subset expression (with and without lexical nominal head)
We now turn to partitive expressions whose heads don't consist of lexical nouns but of other categories.We first observe classifiers in that function.The most widely used (and possibly only) classifier in Turkish is tane 'item', which historically derives from a word for 'grain', and which can typically modify14 a lexical head noun in noun phrases, as in (20).This classifier can modify the subset expression in partitives, as in ( 21): (20) Dün üç tane elma ye-di-m.yesterday three item apple eat-pst-1.sg'Yesterday, I ate three apples.'( 21) Meyve-ler-den üç tane elma ye-di-m.fruit -pl-abl three item apple eat-pst-1.sg'I ate three apples out of (the set of) fruits.' In this subsection, we address the possibility of leaving out the head noun 15 of the subset expression in ablative partitives, thus apparently using the classifier as though it were the head noun, as in (22).There are differences between the distribution of the classifier as the head of a partitive subset expression, and the distribution of a regular lexical noun in the same function, as in (23) vs. (19b): Meyve-ler-den üç tane ye-di-m.fruit -pl-abl three item eat-pst-1.sg'I ate three (unspecified pieces of) fruit.' (lit.: 'Of the fruits, I ate three "items".')(23) Meyve-ler-den üç tane(*-yi) ye-di-m.fruit -pl-abl three item(*-acc) eat-pst-1.sgIntended reading: 'I ate three specific (entities out of the set of) fruit.' (lit.: 'Of the fruits, I ate three specific "items".')The ill-formed example in (23) contrasts with (19b), which is well-formed, and with ( 22), which is also well-formed with its classifier head but without the accusative.The illformedness of ( 23) must be due to the presence of the accusative marker on the classifier; we surmise that the bare classifier is inherently non-specific (in the absence of the agreement marker) and thus rejects the overt accusative marker.Here we attempt to explain the patterns we have discussed thus far by formulating the following three hypotheses: (24) Hypothesis 1: An "overt nominal head" requirement for DPs: Nominal phrases (partitives being one type) need to have at least one nominal head (i.e. a head with the [+N] feature) which has to be filled overtly.
(25) Hypothesis 2: The classifier (or a measure noun) can qualify as the highest overt head with the [+N] feature (cf.( 22)), but it is semantically so bleached as a noun that it is semantically non-specific by default; thus it cannot bear accusative marking (directly, i.e. without any additional morphology between it and the accusative marker).(In other words, in such instances, the classifier or the measure noun is not the specifier, but the head of the ClP.) 16 15 We assume that the head-N position is present even if no lexical noun fills it.A similar view is defended for Japanese partitives by Sauerland & Yatsushiro (2004). 16We would like to remain agnostic about the formal means by which this is achieved.Given that the classifier and the measure nouns carry the [+N] feature, these elements can be lexically inserted into the Cl-head position directly.A second possibility would be for these elements to occupy the specifier position throughout (i.e. both when the head position of the NP is filled with a lexical noun, and when that position as well as the head position of ClP remain unfilled), and to transmit their [+N] feature to the head position of the ClP under specifier-head agreement, thus making the requirement proposed under "Hypothesis 1" into a principle which can be violated with respect to overtness, as long as the head does have the relevant [+N] feature, transmitted to it by an overt specifier.For a formal (but not optimality-theoretical) approach where well-formed utterances can arise even under certain violations of principles, see Preminger (2014), where such an approach is proposed for failures concerning (genuine) agreement-related situations; while we are dealing here with a "dummy" rather than genuine agreement morphology, the spirit of the approach can be maintained; in Preminger (2014), some other syntactic phenomena are mentioned where certain operations are posited whose failure is tolerated, i.e. long-distance wh-movement, the interaction of specificity and object shift, and the interaction of the definiteness effect and movement to canonical subject position.
Interestingly, Turkish offers an additional option: As mentioned earlier, a nominal "dummy" agreement element with the default features of third person singular occupies the head-D position of the DP, due to the requirement above, under "Hypothesis 1" in (24) 17 ; the classifier or the measure noun is in the specifier position of the ClP, as we have proposed for ClPs in general.Now, it isn't the head of the ClP, but the D-head which is the highest overt head in the DP with the [+N] feature, as in (26) (= (16d)): (26) Meyve-ler-den üç tane-sin-i ye -di -m.fruit-pl-abl three item-3.sg-acceat-pst-1.sg'I ate three non-specific or specific (pieces of) fruit.' (lit.: 'Of the fruits, I ate three non-specific or specific items.') In contrast with the ill-formed version of ( 23), where the accusative follows the classifier directly, i.e. without the default agreement, this example is perfectly well-formed.Why can/must the accusative follow the "dummy" agreement (when it functions as the nominal head), given that it can't follow the classifier when it is in head position?

Numerals and quantifiers as subset expressions
Numerals as well as certain quantifiers are similar to tane in their ability to stay in-situ; thus they, too, trigger insertion of the "dummy" agreement into the D-head position of the subset expression in partitives.They are different from tane, however, given that they have no [+N] feature; this makes impossible any potential second option of staying bare, which tane does allow, as we saw earlier.Only the last example in (28), with both dummy agreement and the accusative marker, is well-formed.Other quantifiers like bazı 'some' behave just as numerals in this respect: (28) a. *Meyve-ler-den altı / bazı-lar ye-di-m.fruit-pl-abl six / some-pl eat-pst-1.sgb. *Meyve-ler-den altı-yı / bazı-lar-ı ye-di-m.fruit-pl-abl six-acc / some-pl-acc eat-pst-1.sgc. *Meyve-ler-den altı-sı / bazı-lar-ı ye-di-m.fruit-pl-abl six -3.sg / some-pl-3.sgeat-pst-1.sgd.Meyve-ler-den altı-sın-ı / bazı-lar-ın-ı ye-di-m.fruit-pl-abl six-3.sg-acc/ some-pl-3.sg-acceat-pst-1.sg'I ate six / some of the fruits.' 18 17 In an optimality-theoretical approach somewhat similar to the proposal in Preminger (2014), mentioned in the previous footnote, one could say that the alternative with the "dummy" agreement is the basic one, given that it is motivated by Hypothesis 1, and that the Agr-less alternative, leaving the classifier or measure noun in specifier position, with the Cl-head position remaining empty (under inheritance of the [+N] feature), represents a violation of Hypothesis 1, which we said is a violable principle, especially if the alternative obeys a principle of laziness: "Do nothing (overtly)": the elements in question remain in their basic specifier position, and the Cl-head remains empty, i.e. no insertion takes place.(We have added the modification of "overtly", so that feature transmission or inheritance, here between the specifier and the (empty) head would not count as "doing anything".) The ill-formedness of the examples in (28a) is due to the inability of numerals and quantifiers to qualify as a nominal head (or to transmit a nominal feature to the associated head position) -a fact which is not surprising, as briefly mentioned above, since neither group of items has a [+N] feature, and is also due to the requirement in (24), according to which the DP needs to have at least one head position to be filled overtly with a [+N] element (or, in its more permissive version alluded to in footnotes 16 and 17, needs to have a [+N] feature transmitted to it; numerals and quantifiers, lacking such a feature themselves, are unable to transmit it to a head position); we hypothesize that the examples in (28b) are illformed for the same reasons; in other words, in the absence of the dummy agreement filling the head D-position (or of a lexical noun occupying the head position of the NP-core of the DP), and when there is no other nominal element within the DP which occupies a head position (or which transmits a [+N] feature to a head), the result is ill-formed irrespective of the presence of the accusative marker.The ill-formedness of the examples in (28c) is due to the requirement that agreement, even when realized as a "dummy" element, needs the presence of the accusative, due to the agreement's pronominal features.

Adjectives as subset expressions in partitives
Similar, if not fully identical, paradigms are exhibited by adjectives in subset expressions of partitives: The well-formedness of (29b) for a number of speakers is due to the ability of a good number of adjectives to qualify as a lexical noun, and thus to be able to be inserted into the N-head position of the NP directly -cf. the traditional view that there is no or little distinction in Turkish between nouns and adjectives. 18Note that all speakers provide average acceptability scores for (29d), with dummy agreement in D-head position.Crucially, no unacceptability scores have emerged from examples such as (29d), as shown by our study reported in the appendix.In other words, even those speakers who do allow their adjectives to freely occupy an N-position (and thus allow overt accusative to immediately follow it) accept the insertion of dummy nominal agreement into the D 0 -position and may even prefer it as an alternative.As expected, once D 0 is filled with the dummy agr, overt accusative must follow it, as in (29d).

Conditions for accusative case marking on partitives in Turkish
The Turkish data clearly show that Differential Object Marking expresses a semanticpragmatic feature (here specificity) only if case is not formally required.In Turkish, case is formally required by the agreement marker.On the other hand, the presence of the agreement marker itself is required by (i) the genitive introducing the larger set in the genitive partitive construction or (ii) -as we hypothesize -when the agreement marker functions as a "dummy pronoun" contributing pronominal features to a non-lexical nominal head as the subset expression of a partitive construction, as in (28d) and (29d).Table 2 summarizes our findings: For ablative constructions, the case marking placed directly on the lexical noun heads of partitives (e.g. in (19b)) expresses specificity.However, when a lexical noun head is absent, and in the presence of numerals, quantifiers and most adjectives, excluding those behaving like lexical nouns, as the subset expression, the agreement marker expresses a "formal" specificity, i.e. it "promotes" the phrase, conferring to it the status of a referential phrase (and thus requires further marking with overt accusative), without however expressing semantic specificity.For the classifier tane we observe a mixed picture: without either the accusative case suffix and the "dummy" agreement marker, it expresses the semantic feature [-specific], while with the agreement and case markers this feature is neutralized, as we saw earlier.We will see some interesting variation with respect to this picture in the next section.

Partitives in Azerbaijani
For all practical purposes, Azerbaijani is very similar to Turkish with respect to most of the properties we have looked at so far, including DOM, which allows distinguishing between specific and non-specific direct objects.However, it shows an interesting difference with respect to partitive constructions with a classifier head.For a complete picture we first show the relevant data for lexical nouns.

Partitives with lexical nominal subset expressions
Like in Turkish, lexical nouns referring to the subset of a larger set in a partitive construction can vary with respect to the accusative marking realized on the subset expression.In such instances, the accusative does express semantic specificity.
(30) a. Äli büro-ya ušaq-lar-dan iki qız al-acaq.Ali office-dat child-pl-abl two girl hire-fut 'Ali will hire for the office two girls of the children.' (any two girls) 21 b.Äli büro-ya ušaq-lar-dan iki qız-ı al-acaq.Ali office-dat child-pl-abl two girl-acc hire-fut 19 [∅SPEC] allows for either a [+SPEC] or a [-SPEC] interpretation, as determined by the context, given that [∅SPEC] is indeterminate with respect to specificity, due to the fact that the presence of the accusative marker is due to formal rather than to semantic reasons. 20  'Ali will hire for the office two girls of the children.' (certain two girls) (30a) with its lexical head (and without overt accusative) as well as its counterparts with the count classifier presented in (31a) and (32a) below offer additional illustration for non-specific partitives.Thus, just as in Turkish, Azerbaijani shows that partitives can be non-specific, contra Enç.

Classifiers as partitive heads -the importance of the [+human] feature for case marking in Azerbaijani
In addition to the preceding considerations, we can show that similar to Turkish, the corresponding (and cognate) inanimate classifier dənə 'item' can qualify as a high [+N] functional head in the DP projection, by being directly inserted into the head position of the NP (cf.(31a), but it cannot take accusative case marking, as we see in (31b), because it is a semantically bleached noun, even where it is the DP-head.The subset in (31a) can only be interpreted as a non-specific set of two books (compare to the similar Turkish examples ( 22) and ( 23) above).Also as in Turkish, when dənə is in its regular specifier position (i.e. the same position it occupies when the NP-head position is filled with a lexical noun), default agreement is inserted into D 0 and is obligatorily followed by overt accusative (note the contrast between (31c), where there is no overt accusative marker, and (31d), where the marker is present): (31) a. Kitab-lar-dan iki dənə al -dı -m.book-pl-abl two item take -pst -1.sg 'I took two 'units' of the books.' (non-specific) b. *Kitab-lar-dan iki dənə-ni al -dı -m.book-pl-abl two item-acc take -pst -1.sg Intended reading: 'I took two 'units' of the books.' c. *Kitab-lar-dan iki dənə -si al -dı -m.book-pl-abl two item -3.sg take -pst -1.sg Intended reading: 'I took two 'units' of the books.' d.Kitab-lar-dan iki dənə -si -ni al -dı -m.book-pl-abl two item -3.sg-acc take -pst -1.sg 'I took two 'units' of the books.' (specific or non-specific, depending on context) Azerbaijani, in addition to all its similarities to Turkish, exhibits an interesting option missing in Turkish: it has a [+human] classifier: nəfər 'person'.(32a) should be compared to the similar, and similarly well-formed, (31a), where the inanimate (actually, [-human]) classifier, bare of any case morphology, functions as a lexical nominal head and gets a non-specific interpretation.The [+human] classifier is similar in (32a), in that it is not followed by any case morphology and has a non-specific interpretation.In (32b), however, we see that the [+human] classifier can be followed by the accusative morphology, in contrast with the [-human] classifier in (31b): (32) a. Äli qadın -lar-dan iki nəfər čaγır-dı.Ali woman-pl-abl two person call-pst 'Ali called two (persons) of the women.' (non-specific) b.Äli qadın -lar-dan iki nəfər -i čaγır-dı.Ali woman-pl-abl two person-acc call-pst 'Ali called two [specific or non-specific] (persons) of the women.' (compare to the ungrammatical (31b), where the accusative is precluded on the inanimate classifier head) Here we observe parallels between Azerbaijani and Turkish: The version with the accusative, where well-formed, is indeterminate between a specific and a non-specific reading (and thus allows both interpretations), while the version without the accusative is non-specific.The existence of examples such as (32a) further shows that there are clear instances of non-specific partitive subsets.In light of these observations we ask the following questions: 1.Why does (32b) allow a non-specific reading (in addition to a specific reading), despite the accusative on its partitive direct object?2. Why is the use of overt accusative marking with a [+human] classifier subset expression well-formed in (32b), while the corresponding presence of the overt accusative on the [-human]  Hypothesis 4 addresses both questions.The ill-formedness of (31b) is also explained, along similar lines as its counterpart in Turkish: Even where the (non-human) classifier occupies a high [+N]-head position within DP, it is bleached semantically to such an extent that it cannot be [+specific] on its own and therefore cannot be followed by the accusative marker directly.The claim that in Azerbaijani, the feature [+human] is related to accusative marking, at least with respect to classifiers that are in a high [+N]-position, is based on the following observations: nəfər 'person', just like dənə 'item', is a genuine classifier in a regular DP. (34a) with dənə is well formed, just as are (34b) and (34c) with nəfər: Our informant as well as textbooks of Azerbaijani (e.g.Householder & Lotfi 1965: 26) state that both dənə and nəfər are classifiers.However, they differ with respect to the following details: Textbooks typically state that dənə is used for inanimates, and nəfər for humans, while our informant states that dənə is used for countable [-human] entities (i.e.including non-human animates).Our examples in (34) thus further support the traditional classification of both dənə and nəfər as classifiers.Note that in this respect, nəfər is different from kişi 'person' in Turkish, which is a regular lexical noun and as such can be followed by the accusative when it functions as a partitive head as in (35b), despite the fact that it is rather bleached semantically; importantly, in contrast with nəfər, its explicit use as classifier (i.e. when the NP-head is occupied by a lexical noun) is ill-formed (as in 35a):

Adjectives and numerals as subset expressions
The condition stated in (24) for Turkish, imposing an overtly filled nominal head in DPs, and thus in partitives (as well as possessives) exists in Azerbaijani, too, and plays a similar role: In the absence of a nominal head, usually a dummy agreement element (likewise with the default values of third person singular) is inserted into the D-head position, and that element forces obligatory use of the overt accusative marker.Furthermore, just as in Turkish in general, most adjectives do not qualify as a [+N] head within DP.Only the form with both the dummy agreement and the accusative is well-formed in the following examples, as in (36d).The situation is similar with numerals, i.e. they can't occupy a [+N]-head position within the DP, either; a dummy agreement element is needed to fill such a position (presumably the D 0 position, as suggested for Turkish earlier in this paper), due to the requirement in (24), which applies in these instances, as well.Just as in Turkish, an obligatory accusative marker follows this agreement element.Given that the obligatory use of the accusative marker is due to formal reasons, i.e., due to the fact that the pronominal features of agreement prevent the morphological accusative from functioning as an expression of specificity, the relevant examples are indeterminate between a specific and non-specific reading: (37) (Baxmadan) kitab-lar-dan iki -si -ni al -dı-m.without looking book-pl-abl two -3.sg-acc buy-pst-1.sg'(Without looking), I bought two of the books.' (specific or non-specific, with the primary reading of non-specific, due to the manner adverbial baxmadan 'without looking')

Overt accusative not necessarily expressing specificity: Evidence from Azerbaijani
Thus, where the presence of the accusative marker is enforced due to formal reasons such as the presence of overt agreement morphology, the accusative does not (unambiguously) express specificity.This generalization, as well as generalizations similar to the ones just made concerning adjectives and numerals, are valid for the non-specific pronoun biri 'somebody, someone' (also just as in Turkish-cf. von Heusinger & Kornfilt 2005).The morphological complexity of the pronoun shown in (38) presumably stems from the fact that the numeral by itself cannot be an N and therefore needs a dummy agreement element, similar to our observations on numerals in general.This agreement element requires the obligatory use of accusative marking, just as it does elsewhere, e.g. ( 39): (38) bir-i one-3.sg(39) Bu xüsusiyyət-lər-ə malik ol-an bir -i -ni axtar-ır-am.this property-pl-dat owner be-relpart one-3.sg-accsearch-prog-1.sg'I am looking for someone who has these properties.'This example could be part of a job ad.More precisely, the speaker or writer is looking for anybody who has a particular set of characteristics; the primary reading is non-specific, despite the presence of the overt accusative marking.The accusative shows up due to the agreement marker rather than being due to semantic specificity.The fact that the direct object in (39) with its accusative marker is not specific in its primary reading is made clear by its discourse-functional synonymy with (40), whose direct object bears no accusative, and is clearly interpreted as non-specific.Just as (39), this can be a job ad.Only a particular set of properties of any individual to fill the job are important, without reference to a particular person.With a lexical noun no accusative shows up, as in (40), unlike the pronominal with the dummy agreement, as in (39).( 41) differs from (40) only in having overt accusative on the direct object.Here, the speaker is looking for a specific person/librarian. (40) Bu xüsusiyyət-lər-ə malik ol-an bir insan / bir kitabxanačı axtar-ır-am.this property-pl-dat owner be-relpart a person/ a librarian search-prog-1.sg'I am looking for a person/a librarian who has these characteristics.' (41) Bu xüsusiyyət-lər-ə malik ol-an bir insan-ı / bir kitabxanačı-nı this property-pl-dat owner be-relpart a person-acc/a librarian-acc axtar-ır-am.search-prog-1.sg'I am looking for a person/a librarian who has these characteristics.' (specific) The pair ( 42) and ( 43) makes a similar point: This pair illustrates the observation that agreement, here, a "dummy" one, requires the presence of overt accusative; and the lack of it leads to ill-formedness as in (42).This requirement, namely the formal requirement imposed by the (dummy) agreement marker on the accusative morphology, also explains the indeterminacy between a specific and a non-specific reading for (43), which is well formed.

More Altaic variation
In the previous section, we illustrated an interesting difference between Turkish and Azerbaijani with respect to the morpho-syntactic relevance of the feature [human] in Azerbaijani in the context of a human classifier which, when in nominal head position, makes overt accusative possible.Thus, we concluded that the feature [human] overrides the feature [specific] in these instances; when the overt accusative shows up due to the feature [human], albeit optionally, specificity is not expressed at all.In contrast, perhaps due to the absence of a dedicated human classifier in Turkish, we don't find any interaction of the feature [human] with the feature [specific] in corresponding examples -in fact, we have not found syntactic phenomena determined by the feature [human] in Turkish partitives 21 .
In the following subsections, we shall compare additional variation across some Turkic languages and across Khalkha Mongolian, another Altaic language.We shall see that in Uzbek, classifiers cannot occupy a nominal head position in partitives at all, thus making the sequence agreement-accusative obligatory.Kirghiz, another Turkic language, is interesting in exhibiting a special morpheme for 'group, set', which requires the presence of overt accusative when placed in a nominal head position, a placement motivated by the condition in ( 24).The agreement marker, whose presence is thus not required by that condition, is now free to express specificity.Data from Sakha (Yakut) provide further evidence that the underlying concept of specificity is an abstract notion of "referential anchoring" introduced in section 1.3 (see also von Heusinger 2002).Finally, we observe an interesting variation in the position and function of the agreement marker in Khalkha Mongolian, and we contrast its manifestation with that in the Turkic languages described so far.

Partitives in Uzbek
Uzbek resembles Turkish and Azerbaijani with respect to the main structure of the partitive constructions, as well as with respect to case markers and agreement suffixes, but it differs from them in having developed a very sophisticated system of classifiers or 'numeratives' (for more details see Sjoberg 1963;Beckwith 1998: Bodrogligeti 2003;von Heusinger & Klein 2009).Here we restrict our attention to the behavior of these classifiers in partitive constructions.
Uzbek is slightly different as it requires accusative marking for all human indefinite direct objects, neutralizing the specificity DOM may signal, cf.(44) vs. ( 45), whereas it marks non-human indefinite direct objects with accusative morphology depending on their specificity status, cf. ( 46) and ( 47) (Niyazmetova 2009): (44) *Men bir talaba qidir-ayap-man.I a student look+for-prog-1.sgIntended meaning: 'I am looking for a student.' 23 21 We would like to make it clear that this claim has to be somewhat weakened, in view of the observations reported in section 5. Some native speakers of Turkish reject certain combinations of the classifier and a human noun in ablative partitives, while accepting similar constructions, when the classifier is combined with a noun referring to non-human entities. 22One of the anonymous reviewers has informed us that his/her native Uzbek informant found (44) perfectly well-formed under the intended meaning; for that informant, however, (45), with its accusative marking on the direct object, receives only the specific reading.For such speakers, then, Uzbek has the same conditions as those for DOM in Turkish.Unlike in Turkish and Azerbaijani, the accusative is obligatory in partitive constructions with a lexical noun in Uzbek.Uzbek, however, follows the typical DOM conditions for simple noun phrases, see ( 48), similar to the examples presented above: (48) a. *Ali büro-ga-gi bola-lar-dan ikki q'iz tani-y-di.Ali office-loc-link child-pl-abl two girl know-prog-pst (intended meaning: 'Ali knows two girls of the children in the office.')b.Ali büro-ga-gi bola-lar-dan ikki q'iz-ni tani-y-di.Ali office-loc-link child-pl-ABL two girls-acc know-prog-pst 'Ali knows two girls of the children in the office.' In addition to the simple count noun construction as illustrated by ikki q'iz in (48a and b), Uzbek has developed many specialized classifiers or 'numeratives'; the two main classifiers are nafar (originally 'person') for humans and the non-human classifier dona ('item') (see Sjoberg 1963;Beckwith 1998;Bodrogligeti 2003;von Heusinger & Klein 2009).These are clearly counterparts of the Azerbaijani nəfər (the human classifier) and the Turkish tane and the Azerbaijani dənə (the non-human classifier), respectively.They obligatorily take the agreement suffix and therefore also the accusative suffix -all other forms being ungrammatical.The most general classifier is the suffix -ta, which is enclitic to numerals and can be used without [±human] restrictions.In contrast with Turkish and Azerbaijani (see the example in ( 22) for the former and (31a) and (32a) for the latter) these Uzbek classifiers always obligatorily take the agreement suffix and accordingly also the accusative suffix -all other forms being ungrammatical.We take these observations to signal that these classifiers are not able to occupy a [+N]head, and that this leads to the requirement of the agreement marker, as a means of satisfying the requirement in Hypothesis 1: (50) a. Ali hotin-lar-dan ikki nafar-i-ni tanir-di.Ali woman-pl-abl two person-3.sg-accknow-pst 'Ali knew two (of the) women.' (specific and non-specific) b. *Ali hotin-lar-dan ikki nafar-i tanir-di.Ali woman-pl-abl two person-3.sgknow-pst Intended meaning as in (a) c. *Ali hotin-lar-dan ikki nafar tanir-di.Ali woman-pl-abl two person know-pst Intended meaning as in (a) (51) a. Men rasta-da-gi kitob-lar-dan ikki dona-si-ni ol-di-m.1sg shelf-loc-link book-pl-abl two item-3.sg-acctake-pst-1sg 'I took two (of the) books from the shelf.' (specific and non-specific) b. *Men rasta-da -gi kitob-lar-dan ikki dona ol-di-m.1sg shelf-loc-link book-pl-abl two item take-pst-1sg 'I took two (of the) books from the shelf.' Partitive constructions with a numeral in the subset expression need either the classifier nafar or dona (both of them free morphemes), as we saw in ( 50) and ( 51), or the morphologically bound classifier -ta as illustrated in (52), besides the agreement and case marker.This is unexpected as we do have countable noun phrases without a classifier, such as those in ( 45)-( 48).However, for the forms without a lexical noun, we obligatorily have to insert the classifier, the agreement marker and the Accusative suffix, as seen in ( 50)-( 52 Table 3 presents the summary of the micro-variation among the three languages studied so far with respect to ablative partitive constructions.Please note that this table excludes genitive partitive constructions because they always require an agreement marker and accordingly also a morphological accusative case marker on the subset expression across all three languages, as in Table 3; as we stated earlier, this paper does not address genitive partitives. Turkish allows unmarked tane expressing non-specific partitive subsets, while subset expressions with agreement and case marker are neutralized with respect to specificity.Azerbaijani shows the same distribution for the inanimate classifier dənə, while the human classifier nəfər, when it occupies the position of a [+N]-head, requires the accusative, due 23 It appears, then, that morphologically free as well as bound classifiers in Uzbek do not possess a [+N] feature and thus cannot satisfy the condition in (24), necessitating the presence of the dummy Agr in D for the satisfaction of that condition.Please also note that our informant, Dildora Niyazmetova, stated that partitive constructions in Uzbek with adjectives (see (29) for Turkish) are not well-formed.Alternatively, one can use a partitive with a genitive representing the superset.

[-spec] [∅spec]
[+spec]  to its [+human] feature.In turn, the accusative marker, due to its formal requirement, stops expressing exclusive specificity; thus the accusative marker causes indeterminacy with respect to specificity in the relevant examples.Finally, Uzbek shows a pattern which differs further: Here only forms with agreement and overt accusative case are grammatical for both types of classifier, and thus the contrast between a specific and non-specific reading is neutralized.This contrast can, however, be expressed by the presence or absence of the indefinite article (see von Heusinger & Klein 2009).

Partitives in Kirghiz
Similar properties that raise similar questions as discussed above are present in Kirghiz, as well; for example, similar to Turkish and Azerbaijani, in partitives, the superset in the ablative does not need an agreement element on the subset expression in Kirghiz, see (53).For Kirghiz too we focus on ablative partitives; genitive partitives with agreementmarked lexical nouns as subset expressions are ruled out, arguably for similar reasons as in Turkish and Azerbaijani -i.e. the similarity of genitive partitives to genitive possessive phrases, with the resulting tendency of such constructions being interpreted as possessives rather than as partitives.
(53) Ali ofis-ke bal-dar-daneki kiši-ni al-gana Ali office-dat child-pl-abl two person-acc take-nomzn žat-a-t.light vrb-pres-3.sg'Ali will hire two of the children for the office.' The following pair in ( 54) is particularly interesting.Our Kirghiz informant, Kenjegül Kalieva, stresses that in (54a) the speaker took/bought any two books, i.e. that the subset is not specific, despite the accusative morphology.For a specific reading, the utterance must be changed as in (54b), which is somewhat similar to the Uzbek construction in (52a), but has a morpheme -öö 'set' (which we assume to occupy the head position of #P), rather than a classifier (as in Uzbek); also, in contrast with the classifier in Uzbek, the set morpheme can be directly followed by overt accusative, as in (54a) -a fact which we take to mean that the set morpheme has a [+N]-feature, which allows it to satisfy the condition in ( 24): (54) a. Karabastan kitep-ter-den eki-öö-nü al-dɨ -m.without looking book-pl-abl two-set-acc buy-pst-1.sg'I bought a set of two books [non-specific] out of the (larger set of) books, without looking.'b.Karabastan kitep-ter-den eki-öö-sü-nü al-dɨ -m.without looking book-pl-abl two-set-3.sg-accbuy-pst-1.sg'I bought a set of two [specific] books out of the (larger set of) books, without looking.'Our native informant emphasizes that in (54b), the speaker took a set of two specific books.
We see that in (54a), the subset expression is non-specific, despite the accusative.Furthermore, we observe that a subset expression can be non-specific, even though it is (part of) a partitive.Finally, we see that the accusative does not express specificity here (although it does so in general, with regular lexical nouns).What could the reason be?
Note that here, the numeral is clearly not in any head-position in such a way so as to satisfy the condition in (24): it is not expected to be able to do so, given that it has no [+N] feature.Thus, the non-specific reading must be due to a different reason.These observations lend support to the following hypothesis: (55) Hypothesis 5: The "set" suffix -öö (i.e. the morpheme expressing "set") is in head-#P-position, and it has nominal features; these properties enable it to satisfy the requirement in (24).This morpheme, when placed in the head position of #P, requires overt accusative, just as agreement does in other Turkic languages, as well as elsewhere in Kirghiz.
The meaning of this morpheme as 'set' gives it formally specific features, which require the accusative, as in (54a).We assume that the insertion of an element with nominal features into an empty nominal head position takes place as early in the derivation as possible; given that in Kirghiz, the set morpheme precedes the agreement morpheme in those instances where both show up (as in (54b)), it is the former that gets inserted into the #P-head position, while the agreement morpheme is placed into a higher functional projection -i.e. in D 0 , as we have posited for the other Turkic languages we have discussed so far.Again, this is an instance where the presence of the accusative suffix is enforced by formal morpho-syntactic requirements rather than semantic ones, and thus we have lack of genuine specificity, when the set morpheme is immediately followed by accusative morphology 24 .Semantic specificity is expressed here by the agreement morpheme, when that morpheme shows up after the set morpheme; since its presence is not formally required (because another morpheme, i.e. the set morpheme, fulfills the function of occupying a head position and providing a nominal feature for the satisfaction of the condition in (24)), agreement can have a semantic function here, i.e. its pronominal features are not just formally, but also semantically specific.This discussion is summarised in Table 4.

Partitives in Sakha
Sakha is a Turkic language spoken in Siberia; it is also referred to as Yakut in some literature (see Vinokurova 2005).Its partitive constructions are, by and large, similar to those of Turkish and of the other Turkic languages we have presented so far.However, the constraints on the use of overt structural case marking are somewhat different from those in other Turkic languages.We claim that case marking in Sakha expresses specificity, in the sense of "referential anchoring" and that the anchor can only be the subject of the sentence, but not another salient discourse referent or speech act participant (see section 1.3 for discussion). 24There is a question that arises here: Why don't we have the effect of indeterminate specificity, as in all other instances seen so far when markers whose occurrence is due to formal reasons give rise to a neutralization of the plus and minus values of specificity?In other words, why don't we have here the possibility of either a specific or a non-specific reading?Our answer is that the potential specific reading is blocked by the existence of examples such as (54b), where the "dummy" unmarked 3.sg agreement marker dictates a specific reading, thus making the specific reading in (54a) unavailable.

[-spec] [∅spec]
[+spec] We saw among some partitive constructions used as direct objects some instantiations without accusative marking among our Turkish and Azerbaijani examples, and we characterized them as having non-specific subsets (for Turkish see (16a, b, and c) and ( 17); for Azerbaijani see ( 30)).In Sakha, however, such examples with transitive predicates are illformed.This is illustrated by the following examples, where we contrast Turkish and Sakha:  57), where the partitive direct object bears accusative case morphology, are well-formed in both languages, the examples in (56) without accusative are acceptable only in Turkish and not in Sakha. 25Our informant, Nadezhda Vinokurova, explained the reason for rejecting (56b) via a conflict between the lack of subject-relativized specificity of the partitive and the predicate: If Ali knew the (two) persons involved, then the theme of that knowledge, i.e. the two (of the women), are known to him, and as a consequence, the direct object must be overtly marked with accusative morphology.Additionally, we submitted a number of similar examples to our informant, which were also rejected by her, based on what she perceived to be a conflict of the same kind.In Turkish, no such conflict arises with respect to otherwise completely parallel examples: In Turkish, the specificity of the noun phrase can be relative to either the subject or to the speaker of that sentence.If the speaker thinks that the referent of the direct object, while established and known (to him or her) is irrelevant to the conversation, then the accusative morphology can be omitted, without any perceived conflict with a predicate such as "know", "eat", "see" etc., i.e. predicates which presuppose such knowledge of referent on the part of the speaker.This corresponds to the concept of "referential anchoring" introduced as an analysis for specificity in section 1.3.The difference between Turkish, as well as other Turkic languages on the one hand, and Sakha on the other hand is that Turkish and most of the other Turkic languages studied in this article allow various anchors for the specific noun phrase (including explicit partitive constructions with a specific subset), while Sakha allows only the subject of the sentence.Another difference is seen with respect to adjectives, numerals and quantifiers, which appear to have nominal features in We saw that adjectives and numerals occupying the position of a [+N]-head within DP, while acceptable, are not perfect in Sakha.Thus, another option is to fill the DP-head position with a dummy 3 rd person singular nominal agreement morpheme, to satisfy the requirement in (24), i.e. just as in Turkish.At this point, there are parallels between Sakha and Turkish: They both obey the requirement stated in (24), as well as obeying the requirement of the nominal agreement, whose presence is motivated by that requirement, to be followed by accusative morphology, if the partitive construction headed by that dummy agreement is a direct object; this is clearly shown by the ill-formedness of (60b) in Sakha, where the agreement is not followed by overt case, just as in the Turkish (60a).The only way to save the construction in (60b) is to attach overt case morphology, as in (61b), i.e. like in the Turkish example (61a).(We have parallel examples involving adjectives, to which dummy agreement is attached; we do not list them here, due to space restrictions.) We have seen that the condition in ( 24) is operative in all of the five Turkic languages we have discussed here: A nominal phrase, i.e. a DP, must have an overtly filled NP-head position or a higher functional head position filled with a [+N] element; obviously, the NP-head position can be filled by a lexical noun; but if no lexical noun (or an adjective that qualifies as a lexical noun) is available, a dummy 3 rd person singular nominal agreement element must occupy D (with Kirghiz allowing an additional, optional, alternative, with the 'set'-morpheme -öö in the head position of #P, as the highest nominal head within DP); in the latter instance, the direct object partitive construction headed by a dummy agreement element (or, in Kirghiz, by the set morpheme as an additional option) must also bear overt accusative.In addition to these properties shared by all five languages, we have seen differences with respect to whether modifiers such as adjectives, numerals and quantifiers can or cannot have an [+N] feature allowing them to occupy a head position so as to satisfy the requirement in ( 24): (some) adjectives can have an [+N] feature, while numerals and quantifiers cannot.

Partitives in Mongolian
Mongolian DPs, and, in particular, the partitive constructions are quite similar to the Turkic languages discussed in the previous sections.Mongolian also exhibits Differential Object Marking with similar conditions as Turkish: All definite noun phrases are obligatorily marked for accusative when they are direct objects; an indefinite direct object can be so marked if it is specific.Unmarked indefinite direct objects are ambiguous between a specific and non-specific interpretation (Binnick 1979;Guntsetseg 2016).Mongolian has one strikingly different grammatical feature from Turkic languages regarding the partitive constructions: In Turkish, the case marker follows the agreement marker, as in (62a), while in Mongolian the agreement marker follows the case marker, as in (62b), but case has still semantic scope over agreement (see Binnick 1979: 4).Note that the agreement marker n' is not a suffix, but is an enclitic.
(62) a. Turkish ev -in -de house -3.sg -loc 'in his/her home' b.Mongolian ger -t -n' yurt/home -dat -3.sg 'at his/her home' We have argued in earlier sections that the presence of the agreement marker forces case marking in the Turkic languages, since the agreement marker 'formally' marks a definite expression.Here we ask whether it is possible to apply this argument to Mongolian, regardless of the reversed order of the markers.Mongolian has, like Turkish, genitive and ablative partitive constructions.As we did for other languages, here we especially focus on ablative partitive constructions using the examples of Schlechtweg (2011) and Guntsetseg (2016 With ablative partitives we find an alternation between a version without either accusative or agreement, as in (63a), and a version with accusative and optional agreement, as in (63b) and (63d).( 63a) is ambiguous between a specific and a non-specific reading, while (63b) exhibits the accusative (without agreement) and allows only a specific reading.Examples (63a-d) illustrate that partitives with lexical nouns in Mongolian behave as their counterparts in Turkish (( 16)-( 17)), Azerbaijani ((30)-( 31)) and Uzbek (48a and b).However, the example (63d) from Mongolian provides a distinct structure.While Turkic languages would allow for a noun phrase with agreement (and accusative) morphology in such an ablative partitive construction only when the agreement morpheme is a genuine possessive agreement marker rather than a dummy agreement morpheme (i.e. with the reading two (of) his girls among the children), we assume (based on our informant's judgment) that the agreement marker in (63d) is not a possessive marker, but just a "dummy" agreement marker -a situation we do not find in Turkic languages.The option to add the agreement marker n' is only available if the noun has case while accusative marking can also be found without agreement. 26We find a similar optionality for the agreement marker with numerals, as in (64b) versus (64d).It seems that case marking is obligatory for these constructions, but that a sequence of case and agreement morphemes is just as acceptable as the presence of the case marker alone.
bold book-abl two-acc take-pst 'Bold took two of the books.' 26 Binnick (1979: 118)  The pattern we observe in Mongolian is in line with the general picture we developed based on the Turkic languages: Partitive constructions with an ablative as the superset allow for different subset expressions with particular restrictions.Lexical nouns take structural case (as direct objects) and, due to that case marking, express specificity, while most adjectives must take "dummy" agreement; structural case marking is dictated by that agreement morpheme.In such instances, case cannot function as a specificity marker.In Mongolian we find a reversed affix ordering regarding the case and agreement markers, which allows for combinations we do not encounter in Turkic languages.For lexical nouns we have the same alternation between case-marked and unmarked forms, but we also find a case-marked form with an agreement marker, signaling definiteness.Another difference we observed is that in Mongolian, numerals behave more like lexical nouns with respect to their marking in partitive constructions, while in Turkic languages numerals are more like those adjectives which do not have nominal features and thus must be followed by "dummy agreement", when they express a subset within an ablative partitive construction which lacks a lexical noun heading the subset expression.

Acceptability study for ablative partitives across animacy categories
In section 1.6 we discussed the basic partitive constructions with Ablative in Turkish and reported related grammaticality judgments obtained from native informants.Two reviewers informed us about different judgments for some of the examples in ( 16).Some of their informants did not accept any of the ablative partitives, while others accepted only some of the constructions.One reviewer suggested that this variability might arguably stem from "dialectal variation" with respect to partitive constructions.Therefore we conducted a pilot study described below, so as to respond to this impression. 27aterials: We created a short questionnaire with 9 conditions testing different combinations of numeral, classifier, adjective and quantifier with agreement and case underpinning confirm our intuitions about the grammaticality of the examples.The visual inspection of the graphs representing our data did not reveal any pattern that could guide us to split the group of participants into further groups.The narrow standard error bars on both of the above graphs also further indicate that the variation was not high among the participants.Thus the data do not signal dialectal variation.
To summarize, the pilot study clearly supports the judgments presented in section 1.6 for the basic examples of partitive constructions with Ablative in Turkish.The study also provides more differentiated results with respect to the type of constructions, especially  with respect to the presence of a lexical noun as well as the different functions of the classifier "tane" in Turkish.

Summary and conclusions
In this paper we presented a comparative study of various Altaic languages, which all show Differential Object Marking allowing expression of specificity and some of which also show sensitivity to the animacy or to the [+human] feature of the indefinite direct object.Although all the languages under investigation have ablative and genitive partitive constructions, we focused on ablative partitives.We have shown elsewhere (Kornfilt & von Heusinger 2009) that the genitive marking the larger set in a partitive construction requires an agreement marker in the subset expression independent from the type of the subset expression.For ablative constructions, which are the focus of the present study, these languages show a range of different nominal subset expressions (the parameters having to do with whether lexical noun heads are possible, and what categories can qualify as a higher [+N] functional head in the DP's architecture, when a lexical noun is absent in the head position of the core NP).We showed that in all of these languages the nominal agreement marker, when it is present, enforces morphological structural case (i.e., accusative).This also applies to Mongolian, even though the case suffix precedes the agreement suffix, and the agreement marker has more functions than in Turkic, such as signaling definiteness in this language.We saw in passing that the agreement marker is used as a functional element in genitive partitive constructions, to license the genitive on the expression for the larger set.More centrally for our purposes in this paper, we saw that the agreement marker is used in the ablative partitive construction to fill a head position in the DP, when an element with a [+N] feature is needed, to satisfy the condition in (24).The common pattern that has emerged across the languages we have studied here under a comparative approach is that a functional element which signals, in general, a particular function (in our examples, the accusative and the agreement markers expressing specificity in the Turkic languages) stops being a reliable indicator of that function, when its presence is required by formal properties of its syntactic and morphological context; e.g. the agreement marker, when it has to fill an empty D 0 -position (when the NP-head is not occupied), and the accusative marker, when its presence is required by the agreement marker.Additionally, in some cases we found a different functional element taking over the function of formal specificity, e.g. the set marker in Kirghiz.
These comparisons and the very instructive sample from these cognate languages (Turkish, Azerbaijani, Kirghiz, Uzbek, Sakha, and Mongolian) enabled us to see a surprising picture of micro-variation with respect to DOM, the use of case marking, the function of classifiers, the properties of numerals and adjectives, and finally the function of the agreement marker in interaction with the case marker.In light of our contrastive analyses based on this variety of languages we draw the following conclusions for Altaic languages: 1.All investigated languages exhibit Differential Object Marking, which primarily depends on the referentiality scale.Case marking is obligatory for all definite expressions and optional for specific indefinite expressions.Kirghiz and Uzbek make an additional distinction with respect to animacy: All human indefinite direct objects are obligatorily case marked, while non-human direct objects vary depending on their specificity status.2. For ablative partitive constructions, the case marking of the whole partitive follows the exact same pattern with respect to Differential Case Marking.All languages, except Kirghiz and Uzbek, case-mark the direct object depending on its specificity status, while Kirghiz and Uzbek case-mark obligatorily, thus neutralizing the [± specific] marking.Mongolian is the only language among those investigated here that can freely or optionally use the agreement marker for indicating definiteness, which is not possible in the Turkic languages.3. Turkish, Azerbaijani, Sakha, and Uzbek have classifiers (of different types, most often derived from full lexical nouns). 28The languages differ with respect to two parameters: (i) whether the classifier system is organized according to the feature [±human], and (ii) whether the classifiers can be used like full lexical nouns: Turkish exhibits only one very widely used classifier (discounting the measure nouns we mentioned earlier), namely tane 'item', used prescriptively with non-human nouns (and which can, colloquially, also be used with human nouns), while Azerbaijani has at least one classifier for human and one for non-human nouns.Uzbek has the most sophisticated system in this respect, but also does have one very widely used classifier, the suffix -ta.The Turkish classifier can either appear without case (with a nonspecific reading) or with agreement and case (and then with both a specific and non-specific reading).Azerbaijani allows the human classifier nəfər to take case, which is not allowed for the inanimate classifier dənə, just like the general classifier tane in Turkish.In Turkish and Azerbaijani, the classifier can qualify as a functional overt head with the [+N] feature within its DP (and thus within the partitive construction), while Uzbek does not allow for this option.The contrast between human and non-human classifiers in Azerbaijani allows for varying markings: An alternation between the non-human classifier which cannot bear case (without the dummy agreement), and the human classifier which can exhibit structural case marking in its bare form.This last option is not available for Uzbek since classifiers do not qualify as overt nominal heads which satisfy the condition in (24); this option is also not available for Turkish, where we have only one generalized classifier.4.Only Mongolian allows numerals to function as a nominal head in the DP, but this is possible only in a non-specific reading.All other languages studied here need an additional suffix to provide a [+N]-head.In Turkish and Azerbaijani this is the agreement marker, in Kirghiz it is the set-forming suffix, and in Uzbek both the classifier and the agreement marker are needed.In Turkish, Azerbaijani, Uzbek, and Sakha, the contrast [±specific] is neutralized in the sequence of dummy agreement and case, while in Kirghiz the contrast can be expressed by the agreement marker.In all of the languages under investigation, adjectives cannot qualify as a head with the [+N]-feature (with certain exceptions in Turkish, where this option is possible), thus they need the "dummy" agreement in the position of D. In other words, numerals behave in a similar fashion to adjectives in Turkish and Azerbaijani, but they are different in Kirghiz, Uzbek and Mongolian.5.In most of these languages the dummy agreement marker takes on the function of expressing the nominal head and thus forcing structural case assignment (when the entire partitive construction is a direct object).However, there are interesting exceptions, such as Kirghiz, where the set-creating suffix takes on this function, and in Mongolian, where the dummy agreement is obligatory only for adjectives in the subset expression.6. Across all the languages investigated here, specificity can be expressed by (accusative) case marking.However, this semantic function is neutralized once the saw two [of PART the men].b.I met a friend [of GEN John].

Figure 1 :
Figure 1: Mean acceptability for conditions across all animacy levels.Error bars represent ±2 standard errors (SE).

Figure 2 :
Figure 2: Mean acceptability across conditions for the Human and Non-Human categories.Error bars represent ±2 standard errors (SE).

Table 1b :
Schematic representation of the ablative partitive construction, with grammaticality markings.
27)Hypothesis 3: Agreement markers have pronominal features, which make the expression formally (not semantically) specific (cf.proposals, e.g. as in Aissen 2003, placing pronominals high in referential hierarchies for purposes of DOM, i.e.Differential Object Marking).
Heusinger & Kornfilt 2005;Kornfilt 2008een a specific and a non-specific reading, as mentioned earlier.The accusative marker stops being a reliable indicator of semantic specificity in these instances, where it shows up due to formal requirements that necessitate its presence (cf.vonHeusinger & Kornfilt 2005;Kornfilt 2008).
The information in parentheses shown for all of our Azerbaijani examples was volunteered by our informant, Dr. Vügar Sultanzade, and confirmed by Prof. Dr. Saadat Zeynalova.

Table 2 :
Conditions for case marking in partitives in Turkish.

Table 3 :
Partitive construction with ablative for larger set and classifier for subset.von Heusinger and Kornfilt: Partitivity and case marking in Turkish and related languages Art.X, page 25 of 40

Table 4 :
Partitive construction with ablative for larger set and numeral and quantifiers for subset (for numeral 'two': bare in Turkish and Uzbek, followed by -öö, the 'set' morpheme, in Kirghiz).