Typology of partitives

2.2.1 Partitives, quantifier and restrictor

Languages have different means of encoding the true-partitive relation. Grammatical means that are capable of encoding the true-partitive relation are referred to as partitive constructions or simply as partitives in this paper:

Contextual inferences are understood in the narrow sense, excluding anaphora resolution. Thus, partitive pronouns (such as French en ‘thereof’) encode – and do not implicate – the reference to the superset in the same way as, for example, a third-person pronoun is a reduced form of encoding the reference. Moreover, many pronouns may also be used deictically and, on this reading, they should be able to occur in out-of-the-blue contexts with the true-partitive meaning.

In a true-partitive relation (7), the quantifier corresponds to the subset and the restrictor to the superset. While subset and superset are terms confined to the true-partitive relation, the quantifier and the restrictor will be used more generally in this paper for any kind of partitive constructions, including pseudo-partitives (see Section 2.2.4 below).

Crucially, the partitive construction must be available – even if restrictedly – for encoding the true-partitive relation in the language in order to count as a partitive at all. This is also the original use of the term. While this definition covers most of the expressions that have been referred to as partitives in the literature, there are some exceptions. For example, I have not come across any examples of the partitive case in -(r)ik in Basque being used for encoding the true-partitive relation (e.g. in Hualde and Oritz de Urbina 2003, eds.: 124–126) while only the ablative case in Basque (-etatik) has the ability to encode the true-partitive relation. Accordingly, the case in -(r)ik – traditionally referred to as partitive – is not a partitive in the framework of this paper, whereas the Basque ablative case is.

Having said that, the definition in (9) leaves much space for other aspects of partitives. First, this definition allows partitives to have other meanings in addition to the true-partitive relation (see Section 6). As I argue below (Sections 6–7), languages rarely have dedicated grammatical means that are only capable of encoding the true-partitive relation. For example, the Finnish partitive is capable of expressing many more meanings than just the true-partitive relation. In fact, the latter is even much rarer than the other meanings of the partitive case. Thus, the following corollary to the definition in (9) can be formulated:

2.2.2 Implicit expressions

The definition in (9) excludes implicit expression of a true-partitive relation (implicit partitives in von Heusinger and Kornfilt 2017; or covert partitives in de Hoop 2003: 207), i.e., quantifiers, numerals and other expressions that may occur in a context in which the partitive interpretation is likely (henceforth implicit expressions). This is found when the superset may be implicated. For example, in (11), the quantifier phrases some/three/a few flowers do not contain the reference to the superset, and the superset can only be inferred given the first clause as the flowers in the garden:

Therefore, it can be said that implicit expressions do not themselves encode the true-partitive relation. For example, they are unlikely to be interpreted as such in out-of-the-blue contexts with no presuppositional support. Moreover, it would be counter-intuitive to say that three is a grammatical expression to encode the true-partitive relation; nor do some or a few encode the true-partitive relation in English.^[5] Rather, these simply quantify items of a particular kind (henceforth plain quantification). While I acknowledge the tradition of subsuming these under partitives as well (inter alia, Budd 2014; Crowley 1982; Hoeksema 1996a, b; Mosel and Hovdhaugen 1992; Schneider 2010),^[6] I refrain from doing so here in order to have more fine-grained terminological distinctions.

Thus, a typical property of implicit expressions is that they do not encode the reference to the superset (the flowers in the garden in (11)) even not in a reduced or a lexicalized form, but only an implicational reference to the kind (flowers), thus violating the definition in (9).

In contrast to English some, e.g., German davon ‘thereof’ is a partitive:

German davon ‘thereof’ does encode the superset which is resolved either anaphorically, as in (12a-b), or deictically (if (12b) is uttered out of the blue by pointing with a finger to some items). Having said this, indefinite pronouns such as English some may come functionally very close to a partitive. For example, the German pronoun welche ‘some’ (homonymous with the interrogative) may indeed be considered to be a partitive as it patterns very much like davon (Glaser 1992).

To summarize: In order to analyze an expression as a partitive and not as an implicit expression, two conditions have to be met: (i) there must be an example in which this expression encodes the true-partitive relation without contextual inferences and (ii) the reference to the superset must be encoded in this example, i.e., be retrievable deictically, anaphorically or by a full NP.

In contrast to English, some languages do have dedicated indefinite quantifiers that are partitives lexically because they can only be used in proportional quantification but not in plain quantification. For example, Cora (Uto-Aztecan; Mexico) has dedicated indefinite quantifiers to encode the true-partitive relation, cf. héiwa ‘many (non-partitive)’ vs. mwi’iká-ka ‘many-acc (partitive)’ (Casad 1984: 265); Haida (Haida) has dedicated, lexical partitive quantifiers such as t’iij ‘some of’ (Enrico 2003: 771 and elsewhere).^[7]

2.2.3 Generalized partitives

Partitives encoding the true-partitive relation are often binominal, i.e., consisting of two nominals (i.e. nouns, pronouns, pronominal numerals, etc.): the quantifier and the restrictor nominal. However, there are considerable input restrictions on both. The restrictor of a true-partitive is always definite (Partitivity Constraint). The quantifier typically does not admit adjectival quantifiers like every, definite and indefinite articles. Personal pronouns are also typically excluded from occurring in the quantifier position (de Hoop 2003: 180 with references).^[8] More generally, there is a very strong tendency for the quantifier to be indefinite. Thus, many grammars consulted in this study do not even provide examples of partitives with definite quantifiers (except for superlative constructions relying on partitive constructions). What is more, not only is the quantifier most frequently indefinite but it is also most frequently pronominal (i.e., lacking a full noun). The quantifier position is filled predominantly with one, some, none (e.g., some of them), much, many (e.g., much of this stuff) or an indefinite numeral (e.g., five of the students) and very rarely by an indefinite full NP.^[9] The latter is a very rare situation in which the reference to the ontological kind is given in the quantifier NP:

Indeed, a small corpus survey of the Russian partitive iz nix [lit.] ‘from them’ in the oral subcorpus of the Russian National Corpus illustrates this strong bias towards indefiniteness and pronominalization of the quantifier (Table 1). I have analyzed the first 300 hits of this expression. After having excluded all non-partitive instances such as predicates governing the preposition iz ‘from’, 277 instances of the true-partitive relation were left of which ca. 95% had indefinite quantifiers and 0% quantifiers with a full NP.^{[10][11][12][13]}

This frequency disproportion makes sense from the semantic point of view: a definite, full NP contains exhaustive referential information and its referent is immediately identifiable. Any additional referential information such as the reference to the superset is, therefore, superfluous (cf. ^?these students of my students).^[14] By contrast, an indefinite referent often needs additional referential support for discourse cohesion. Moreover, pronominal input is also motivated because the superset already contains the information on the kind and resuming it in the subset would not be efficient (cf. ^?some students of my students).

Crucially, these frequency asymmetries have effects on the grammar of many partitives. Thus, the strong affinity to indefinite, pronominal referents is the main motivation partitives which conventionally leave the quantifier unexpressed, cf. ‘some of his colleagues’ in (14a) vs. ‘[some] of his colleagues’ in (14b).

The frequency asymmetries create the default expectation of the quantifier being an indefinite quantifier used pronominally, which, therefore, need not be expressed. Many languages conventionalize partitives with the quantifier left unexpressed and understood as ‘some/any’.^[15] I refer to this mono-nominal subtype of partitives as to generalized partitives (other terms used in the literature are bare partitives in Hoeksema 1996a: 15; or independent partitives in Seržant 2014a, 2014b, 2015a, 2015b). The development of generalized partitives is a very frequent phenomenon found with partitives: 39% (47/122) of all marked partitives in my database allow for leaving the quantifier unexpressed but understood as ‘some/any’.^[16]

The conventionalized indefinite quantifier ‘some/any’ entails a number of properties of generalized partitives crosslinguistically. First, generalized partitives typically encode low-individuated referents and, accordingly, tend to be formed out of NPs that themselves typically occur as low-individuated, e.g., mass nouns (such as ‘water’) or plurals (e.g. ‘flowers’). As a consequence, generalized partitives tend to encode referents that are episodic and have a low probability of recurring in the following discourse (Helasvuo 1996 for Finnish; Tveite 2004: 150 for Livonian; Seržant 2012 for Ancient Greek; Seržant 2014a: 267 for Lithuanian; Seržant 2014b: 307–309 for North Russian).

Secondly, another consequence of low individuation is that generalized partitives primarily co-occur with those lexical verbs that cross-linguistically tend to demote or leave out their objects most frequently (in terms of type and token frequency), e.g., via antipassives. These are verbs such as ‘eat’, ‘drink’, ‘hear’ or ‘read’ (Næss 2007: 127; Seržant et al. forthc). These verbs typically constitute the core of those verbs that allow for partitive objects. Thus, if partitives are not very productive in a language, then it is these verbs that are most likely to be found with partitives. For example, the partitive use of the ablative case in Mordvin languages (the separative strategy) is not very productive (if compared to the related Finnish, for example). It is these verbs that are found with the partitive ablative objects in Mordvin but not with verbs that do not tend to demote their objects, such as ‘kill’ or ‘open’. In other languages, the latter are primarily found with partitives that have predicate-level functions (Sections 6.2–6.4).

Generalized partitives, once conventionalized, may develop into a category in its own right, often having functional facets that are not found with partitives with an explicit quantifier. For example, generalized partitives may compete with other strategies of argument marking, yielding a special subtype of differential-argument-marking systems (cf. Witzlack-Makarevich and Seržant 2018: 15–16). Notably, in this role, generalized partitives fail to differentiate between A, S and P and mark all three roles alike. For example, the partitive prefix/clitic ʔa’-/aa- in Tlingit (Athabaskan-Eyak-Tlingit) can replace subject and object prefixes (Leer 1991: 123–124; though see Crippen 2012: 327), and similarly the partitive pronouns in Eibela (Bosavi) do not distinguish between all three roles. Armenian (Indo-European) may have partitive-ablative intransitive subjects and transitive objects, thus S=P (Dum-Tragut 2009: 313). Furthermore, a number of Finnic varieties occasionally allow all three roles to be coded by the partitive case, i.e., some A’s and S’s and nearly all P’s (Huumo 2018; Koptjevskaja-Tamm and Wälchli 2001: 658; Lytkin et al. 1975: 108; Seržant 2015a: 396). Accordingly, the following cline of preference obtains from the least likely on the left:^[17]

It is the inherently indefinite meaning of generalized partitives that reduces compatibility of generalized partitives with A and partly with S, since the former but sometimes also the latter have a strong preference for definite inputs (cf. Comrie 1981: 128; Croft 1988; Dahl 2000: 50). Moreover, generalized partitives have a considerable overall effect on the semantic transitivity of the hosting clause (cf. Næss 2004, 2007). To this end, partitives in the P slot come functionally – and sometimes even morphologically (when incorporating a partitive pronoun or a quantifier) – very close to antipassives (cf. e.g., the partitive-like meaning facets of the antipassive in Sochiapan (Otomanguean; Mexico) in Foris 2000: 285). In some languages, the effect on transitivity is even rendered morphologically. For example, in Apma, the partitive marker -te – used in (16b) as a pseudo-partitive (Section 2.2.4) – blocks the transitivity marker -e that is otherwise obligatory in this language, cf. (16a):

To summarize, generalized partitives are a subtype of partitives that may exhibit properties that are not found with the corresponding partitives with an explicit quantifier.

2.2.4 Pseudo-partitives

While the semantic concept of the true-partitive relation necessarily involves two distinct sets or portions, partitives may sometimes also express quantification of just one single set or a portion:

(17) is a construction that morphologically – but not necessarily syntactically^[18] – coincides with a construction that codes the true-partitive relation in English as in (18). However, no true-partitive relation obtains in (17) because there is no clearly defined superset in the restrictor position. Instead, one finds a kind-referring expression tea, which only provides the reference to the ontological kind of substance in the cup but not to a particular referent.

With a kind-referring expression, neither the complement nor the superset can be defined in terms of sets or portions. Accordingly, (17) is semantically a plain measure phrase. However, it is formally somewhat more complex than a plain measure phrase like much tea because it involves the partitive marker of. Measure phrases such as (17) have been referred to as pseudo-partitives following Selkirk (1977), e.g., in Koptjevskaja-Tamm (2001, 2009.

While the true-partitive relation implies proportional quantification, pseudo-partitives denote plain quantification such as amounts (a group of people), measures (a cup of tea) or quantities (a lot of people, a majority of people) of particular kinds (people, tea).^[19] Therefore, pseudo-partitives are sometimes referred to as quantitative partitives (e.g., Ihsane 2013). Accordingly, pseudo-partitives do not encode a relation between two referents but rather encode just one referent that is quantified or measured. Semantically pseudo-partitives pattern with simple measure or quantifier phrases such as many people in English or eine Gruppe Touristen (lit. ‘a group tourists’) ‘a group of tourists’ in German and tend to reduce their original syntactic structure of one NP embedded into another NP towards just one NP (see Selkirk 1977 on English; de Hoop 2003 on Dutch).

In what follows, I make two amendments to the traditional use of the notion pseudo-partitive in the literature. First (i), I argue against treating it as a purely semantic notion with no regard to structural criteria. Second (ii), I argue that pseudo-partitives, like partitives expressing the true-partitive relation, may also occur not only with an explicit quantifier but also without the quantifier, i.e., as generalized pseudo-partitives.

As for (i), there is a tradition to extend the notion of pseudo-partitives to include any kind of measure phrases, including those that have nothing to do with partitives in the language in question. For example, Glas Wein (lit. ‘glass wine’) ‘glass of wine’ in German is a pseudo-partitive according to some researchers (e.g., de Hoop 2003: 192; Koptjevskaja-Tamm 2009, 2001; Tănase-Dogaru 2017) while it is not a pseudo-partitive in the terminology adopted in this paper. The reason is that such an extension of a structural-semantic term into a purely semantic notion unnecessarily overgeneralizes the original term of Selkirk (1977), making it synonymous with the more transparent term measure phrase, or quantifier phrase for that matter. Moreover, this conceptual extension also produces confusion in that a glass of wine in English can no longer be terminologically distinguished from Glas Wein in German despite the fact that these two involve morphologically and syntactically different sorts of constructions.

Selkirk (1977) made the important discovery that a glass of wine in English – although morphologically similar to the expression of the true-partitive relation in English (with its head and dependent NPs) – patterns syntactically (and semantically) like one NP in many respects. It is in this sense that the term pseudo-partitive is justified for the English expression a glass of wine. The latter is indeed a seeming – hence “pseudo” – partitive. By contrast, the German Glas Wein is straightforwardly analyzable as one NP and has no structural, semantic or morphological affinity to true-partitives in German whatsoever. The latter are encoded by means of the preposition von ‘from’ (e.g., viele von ihnen ‘many of them’) or with the genitive case (e.g. einer meiner Studenten ‘one of my (gen) students’). Its only relation to partitives is based on the fact that Glas Wein is a measure phrase that may be translated by a partitive-like expression in some other languages including English.

Another problem is that if pseudo-partitive is taken as a purely semantic notion, then one might misleadingly conclude that, for example, much wine in English is a pseudo-partitive because it is translated with a partitive-like construction into Russian (with the genitive originally stemming from the possessive partitive strategy) or Basque (with the partitive case) and because it is synonymous with the pseudo-partitive a lot of wine in English. This purely semantic use of the term pseudo-partitive is also unfortunate for the description of the diachronic process by which true-partitives first develop into pseudo-partitives and only after that into plain quantifier phrases (Seržant, submitted). I conclude that the purely semantic use of the term pseudo-partitives is ill-advised and, instead, structural affinity to partitives must be part of the definition of pseudo-partitives (19):^[20]

As for (ii), another important aspect of pseudo-partitives is that there are also generalized pseudo-partitives, i.e., pseudo-partitives that occur without an explicit head quantifier. Traditionally, only pseudo-partitives with an explicit, “nouny” quantifier or a measure phrase such as a group of, a bunch of, a cup of, a bottle of, etc. were considered pseudo-partitives (inter alia, Koptjevskaja-Tamm 2001, 2009). However, similar to partitives with explicit quantifiers (21a), pseudo-partitives may also occur as generalized partitives (21b), the only difference being the presence (21a) vs. absence (21b) of the quantifier:

Crucially, neither (21a) nor (21b) represents an instance of the true-partitive relation because the restrictor NP ‘candies’ is indefinite generic; contrast (21) with (22):

I conclude that both (21a) and (21b) are pseudo-partitives according to the definition in (19).

4.2.1 Syntactic types

The first variable is the syntactic type, i.e., whether partitives are encoded NP-internally or pronominally. Note that, if the partitive construction is morphologically marked by zero (Section 4.2.2.5), one can also speak about the third syntactic type of partitive constructions, i.e., juxtaposition.

NP-internal markers are primarily flags (adpositions and case inflection) but also, rarely, possessive indexes. For example, in (35), the ablative case marker -duk is an NP-internal partitive marker:

The second syntactic type, i.e. the pronominal type, comprises two different subtypes, which, however, are grouped together here for practical reasons. Their common property is their position: in contrast to the NP-internal partitive markers, partitive pronominals occur predominantly outside NPs. In most of the cases, these are highly integrated into the verbal complex, sometimes even phonetically, i.e., via cliticization. This is the case in many Bantu languages in which partitives are encoded by means of class markers (classes 16–18) that otherwise index locations:

The pronominal type combines the following subtypes: partitive pronouns, partitive quantifiers and partitive particles. The first two are lexically marked partitives. However, there is an important distinction between the two: only partitive pronouns pronominalize the superset and thus have the meaning ‘of them/of it’, cf. German davon ‘of it/them’. By contrast, partitive quantifiers have the meaning ‘some of’, ‘a few of’ and the superset has to be expressed by an NP. Partitive particles, in turn, behave very much like partitive quantifiers, although they do not originate from quantifiers. I discuss partitive pronouns and quantifiers in more detail.

Partitive pronouns often carry no partitive marking that is found outside the pronominal paradigm in the language, and may be considered as being lexically marked as partitives. For example, there is the partitive pronoun en in French (cf. Ihsane 2013), ne in Italian, er in Dutch and different pronouns in German dialects (cf. Strobel and Glaser 2019) – all can pronominalize only third-person supersets:

Partitive pronouns are also found outside of Europe, in less familiar languages. For example, there is a clitic third-person partitive pronoun -i’ij in Itzaj or aa in Tlingit:

In contrast to French en or Dutch er, which are inherently third person only, some languages have partitive pronouns that may occur in all three persons. In Eibela, there is even a whole set of partitive pronouns inflected for person: 1st ni:jɛː, 2nd gi:jɛ:, 3rd animate i:jɛ: (Aiton 2016: 117). These are evidently derived from the plural forms of the respective persons by adding the marker -:jɛ: (homonymous with a locative marker). Thus, the first person ni-:jɛː can be analyzed as 1pl-loc(=part). These pronouns may only be used in the subject or in the object slot in this language:

Similarly, the partitive-paucal marker -sèŋ attaches to personal pronouns in Wadu Pumi, yielding thus partitive pronouns inflected for all three persons: tə̀-bʉ̀-sèŋ (3-coll-part) ‘one/some of them’, nìŋ-bʉ́-sèŋ (2-coll-part) ‘several of you’, iŋ-bʉ̀-sěŋ (1-coll-part) ‘(five) of us’ (Daudey 2014: 86, 151, 579, 585):

While in Eibela or Wadu Pumi, partitive pronouns are syntactically free in that they need not occur next to the predicate, other languages such as French show the tendency for the partitive pronouns to occur immediately next to the verb and often to cliticize with it.

Partitive quantifiers (‘some of’), by contrast, typically do not pronominalize the superset (the kind with pseudo-partitives); the latter is expressed by an NP, cf. exes. (43)–(46):

Partitive pronouns and partitive quantifiers have often quite different historical sources: while partitive pronouns often develop from deictics like ‘there’ or ‘from there/it’, partitive quantifiers stem from existential quantifiers such as ‘some’ or ‘a few’.

Even though the distinction between partitive pronouns and partitive quantifiers seems to be so far straightforward, in practice, it is often difficult to discern whether the pronominal is inherently a partitive pronoun or a partitive quantifier or a partitive particle that occurs in a clause with object drop as in (47):

Note that the pronominal type is orthogonal to the marking strategies to be discussed immediately below (Section 4.2.2) since partitive quantifiers and pronouns may also reflect different marking strategies such as the separative strategy with the German davon ‘from it/them’ (from the pronominal da(r) and the adposition von ‘from’) or the possessive strategy with the Dutch er ‘of it/them’ (historically from a genitive pronoun) or the unmarked strategy (e.g., tuut ‘some’ in Avava). Having said that, most of the time, it is very difficult to determine the marking strategy of the partitives of this type.

Finally, there is the third subtype partitive particles which is even more difficult to distinguish from partitive quantifiers. Partitive particles are found, for example, in many Bantu languages, cf. the class 18 index =mo in Oshikwanyma:

Crucially for the pronominal type that all its subtypes – partitive pronouns, partitive quantifiers and partitive particles – are used pronominally in that either the quantifier or the restrictor or both are not expressed as arguments.

4.2.2 Marking strategies

The second variable, marking strategy, tags the underlying metaphor for the morphological marking of partitives: the possessive strategy (Section 4.2.2.1), the separative strategy (Section 4.2.2.2), the locative strategy (Section 4.2.2.3), some other, infrequent strategies (Section 4.2.2.4), and finally the unmarked strategy (Section 4.2.2.5) that relies on just juxtaposing the quantifier and the (definite) restrictor NPs.

4.3.1 Synopsis of different syntactic types

Table 4 provides the frequencies of different syntactic types: pronominal, NP-internal and no marking at all.^[29] In order to level out the impact of larger families, the totals of families per type are provided first. Here, one family can only contribute one point per type regardless how many languages of that family attest the type (if a family attests several types, it contributes a point to each of those):

While the pronominal type is much less frequent, the NP-internal marking is certainly the preferred syntactic type:

This preference might even be stronger if one counts the juxtaposition – i.e., the unmarked – type as an NP-internal syntactic type.

4.3.2 Synopsis of the different marking strategies

Before I turn to the quantitative evidence, note that languages sometimes allow for more than one strategy (e.g., the separative von ‘from’, the locative an ‘at’ and unter ‘among’ in German), often with different frequencies and degrees of grammaticalization – something that cannot be captured within the frame of this study.

Furthermore, since my aim is to establish the underlying metaphor and thus the source of partitive markers, the classification is unavoidably somewhat diachronic. Thus, partitive markers that synchronically have different meaning facets in the language are tagged according to the function that the partitive is most probably derived from. For example, the partitive marker de in French functionally matches two of the strategies listed above: the possessive and the separative strategy. However, since the partitive function historically derives from the separative and not from the possessive function as is evidenced by Latin (Carlier and Lamiroy 2014), French is tagged as having the separative strategy (the same holds for the partitive pronoun en ‘thereof’). If the grammar does not provide any diachronic information on the ambiguous marker, the partitive is tagged according to the least abstract meaning of the marker in the given language along the following scale: separative < locative < possessive. This seemed to be a reasonable convention that would only affect very few entries anyway. In some instances, the strategy was tagged as unclear.

Table 5 provides the absolute frequencies of each strategy in the database. These counts show that none of the marked strategies is universally preferred because different strategies are found to occur with comparable frequencies. Having said this, there are some clear preferences. While, e.g., the instrumental strategy is clearly dispreferred (subsumed under other in Table 5), the separative and the unmarked strategy are more frequent than any other strategy. I summarize:

The difference between the locative/possessive and the separative strategy is most probably due to areal pressures that boost up the absolute numbers of my sample, which is biased for Eurasia (Section 5).