2.1 Numeral classifiers

Numeral classifiers¹ are usually contiguous to numerals in expressions of quantity or, more generally, found to occur in the context of quantification (Grinevald 2000:63). In this study, we define a numeral classifier as any morpheme that, independent of its morphosyntactic status, is linearly adjacent to a numeral (or an equivalent quantifier) when it occurs, and that functions as an attribute of a head noun, together with the numeral. A numeral classifier tends to have mutual dependencies (e.g., collocational, morphosyntactic, phonological, or semantic) with both the numeral and the head noun. In the following example of Mandarin Chinese, for instance, the noun shū ‘book’ can only be combined with the numeral classifier běn. Furthermore, the numeral liǎng replaces the general numeral èr ‘two’ in numeral classifier constructions. If běn is not preceded by a numeral it has a set of different meanings.

(1) Mandarin (Sino-Tibetan)

liǎng	běn	shū
two.CLF	CLF	book

‘two books’

Numeral classifiers are generally divided into mensural and sortal subtypes. Mensural classifiers aid in partitioning an uncountable noun (e.g., Modern German ein Glas Bier ‘a glass of beer’).² Sortal classifiers, by contrast, are not limited to uncountable nouns, and have been defined variously as:

• A member of a paradigm that forms a binary phrase with a numeral, which in turn forms a binary phrase with a counted noun (Lehmann 2000)

• A grammatical element that occurs with nouns (regardless of their degree of countability) in construction with numerals (Gil 2013)

• An expression that indicates a unit of counting or measure (Doetjes 2012)

Unlike mensural classifiers, sortal classifiers cannot be modified by an adjective: expressions like ein kaltes Glas Bier ‘a cold glass of beer’ have no equivalent among sortal classifiers; sortal classifiers generally cannot co-occur with mensural classifiers, e.g., Maithili das(*-ṭā) kap cāy ‘ten (*CLF) cup tea’ (Burghart 1992:I:117).

A third type of numeral classifier designates groups, similar to English a flock of birds (e.g., Beckwith 1998:131–133). Such classifiers indicate a set larger than one, including a pair. Sortal and mensural numeral classifiers do not indicate any number on their own, but are differentiated by the type of noun they occur with. Most languages exhibit numeral classifiers that refer to measures and groups. This study exclusively focuses on sortal numeral classifiers, which are much less common cross-linguistically and exhibit a very specific geographic distribution.

While a numeral classifier can provide an index to inherent semantic properties of the head noun (its use is often dependent on these properties) and can also carry pragmatic meaning, it does not achieve semantic modification in the way that adnominal elements such as adjectives do.³ Furthermore, anaphoric use is common, in which case the head noun is excluded.

2.2 Optionality of plural marking

In a given language, individual nouns with plural reference may differ in terms of whether plural number must, cannot, or may be morphologically marked on them. Several Indo-Iranian languages, particularly older ones, have rigorous rules requiring that all semantically plural nouns take plural marking. In some Indo-Iranian languages, phonological and morphological change has resulted in paradigms where morphological plural cannot be marked on some noun types in some cases, i.e., where singular and plural forms are formally identical. In the remaining Indo-Iranian languages, plural marking on some noun types is optional, though it is rarely the case that optional plural marking is allowed on all noun types and pronouns; it is usually required on first person pronouns, at the very least, and tends to be sensitive to referential scales such as the animacy hierarchy (Silverstein 1976). Noun types which take optional plural marking in plural referential contexts are said to exhibit transnumerality or optional plural marking (cf. Corbett 2000:9–19), and it is generally the case that certain kind-denoting nouns can be partitioned into entities via strategies other than plural marking, rendering plural marking on such nouns as optional at best, if even allowed.

In the World Atlas of Language Structures, Haspelmath (2013) establishes three degrees of plural optionality (impossible, optional and obligatory), cross-classified against animacy types. In this coding scheme, a language is coded as having obligatory number marking even when the distinction between singular and plural is neutralized in the context of numerals and other quantity words. Since here we are interested in the interaction between numerals, classifiers and number marking we opt for a coding scheme that keeps these three dimensions distinct. Specifically, we code any variation in nominal number marking as optional marking, regardless of whether there is a concomitant numeral (or numeral and classifier) in the same noun phrase. For present purposes we also gloss over distinctions in animacy contexts, or any other semantic or pragmatic dimension that might regulate the appearance of specific number markers.

2.3 The relationship between numeral classifiers and number marking

The best-known formulation of the observation that languages with numeral classifiers tend to have no, or at best optional, plural marking on at least some noun types (in constructions with numerals as well as those without them) comes from Greenberg (1972) and Sanches & Slobin (1973). This hypothesis, known as the Greenberg-Sanches-Slobin (henceforth GSS) generalization (permutations of the names may vary from work to work), is borne out by a large number of languages. The statistical support for the hypothesis notwithstanding (Tang and Her 2019), there are quite a few exceptions (e.g., Aikhenvald 2000:100–101) and this is also true of South Asia.

As noted in (4), for example, Nepali requires plural marking in the presence of classifiers in at least some circumstances. Similar patterns hold for the Dravidian language Kurux and the Sino-Tibetan language Belhare:

(16) Kurux (Dravidian)

saːt	-jhan	kuke	-xadd	-ar
seven	-CLF	girl	-child	-PL

‘seven daughters’ (Tirkey 2017:389, shortened)

(17) Belhare (Kiranti)

sip	-paŋ	maʔi	-chi
two	CLF	person	NON.SING[ABS]

‘two people’ (Bickel 2003:563)

Still, the GSS generalization appears to represent a dominant statistical tendency, and a number of proposals have been put forth to explain why plural marking is optional in many languages with numeral classifiers.

A prominent theory proposes that numeral classifiers help or are needed to enumerate, individuate or partition kind-denoting nouns, i.e., nouns like WATER or RICE that involve non-individuated and uncountable reference. The theory furthermore proposes that languages differ in their proportion of kind-denoting nouns (Quine 1960, Silverstein 1976, Lucy 1992, Croft 1994, Krifka 1995). The statistical version of the GSS generalization follows from these two proposals: languages with more kind-denoting nouns are expected to be more likely to use numeral classifiers in the service of enumeration; furthermore, since kind-denoting nouns are inherently uncountable, number marking is expected to be absent or at best optional on them. Most versions of this theory assume that the proportion of kind-denoting nouns is constrained by a referential scale, e.g., with human-denoting nouns being less likely to be kind-denoting than, say, food-denoting types (Lucy 1992). Theories differ, however, as to whether the variation involves ontological or merely lexical aspects. The ontological view argues that noun types differ cross-linguistically in their ontological entailments: unlike count nouns, kind-denoting nouns designate masses and material without attention to shape and form, and this has ramifications for cognitive domains beyond language (Cassirer 1923, Lucy 1992, Imai & Gentner 1997). Under the lexical view, noun types vary cross-linguistically according to whether the distinction between kind and entity is specified in the lexicon or whether it is lexically ambiguous (Bisang 2002, 2017).

An alternative theory derives an absolute version of the Greenberg-Sanches-Slobin generalization from universal structural configurations. Thus, generative accounts hold that classifiers and plural markers occupy the same structural position (Borer 2005). This predicts the incompatibility of numeral classifiers and plural marking. When they do co-occur nevertheless, as in Nepali, Kurux or Belhare, the relevant markers are predicted to differ from those in other languages, either because of different formal properties (e.g., the classifiers might not be real classifiers) or independent surface phenomena (e.g., the number marker appears where it does for phonological, not syntactic reasons).¹⁰

At the same time, other theories reject the GSS generalizations and derive the presence of numeral classifiers from properties that are not related to number marking: Aikhenvald & Dixon (1998) derive the probability of numeral classifiers from a general typological variable of reference classification. Gil (2013) sees them as an arbitrary conventionalization, possibly related to less configurational noun phrases (Gil 1987). Lehmann (2010) argues that in some languages they are simply necessary to give a numeral the status of a full word. Under all these theories, the distributions of numeral classifiers and optional number marking reflect independent historical contingencies, especially effects of language contact.

In what follows we assess the GSS generalization empirically, probing the evidence for or against a diachronic correlation between numeral classifiers and optional number marking. While overall the Indo-Iranian data seem to be in line with the correlation, a number of observations challenge it:

• Optional plural marking exists in a number of Indo-Iranian languages, likely a diachronic consequence of the loss of final syllable nuclei, and many of these show no sign of developing numeral classifiers. More generally, optionality of plural marking is not a strong predictor for the development of numeral classifiers in Indo-European. Some IE languages (e.g., Hittite) have optional plural on nouns with numerals only, and nowhere else; in some IE languages (e.g., Breton), singular number is even compulsory with numerals. None of these languages, however, developed classifiers.

• Modern Persian requires plural marking in certain referential contexts, in which case it can co-occur with numeral classifiers, flying in the face of the apparent incompatibility of overt plural marking and classifier use.

• A not insignificant number of Indo-Iranian languages, such as Nepali, Kumzari, Yaghnobi, Pashto, and Sinhala, have numeral classifiers and obligatory plural marking.

These observations raise serious questions as to whether the Indo-Iranian patterns owe to development of general number and a subsequent need to partition kind-denoting nouns. In view of this, we turn to statistical modeling to assess the hypothesis. From the predictions enumerated above, we define two versions of the GSS generalization that can be tested using a phylogenetic model. The MUTATIONAL GSS generalization predicts that languages develop numeral classifiers with higher frequency in the presence of optional plural marking than in the presence of obligatory plural marking, since classifiers aid in partitioning kind-denoting nouns. The SELECTIONAL GSS generalization holds there is something inherently dispreferred about the structural combination of classifiers and obligatory plural marking, and that this co-occurrence will be overall less diachronically stable than that of classifiers and optional plural marking. The selectional/mutational dichotomy we have established here can be compared to the distinction between source-oriented and result-oriented explanations drawn by a number of scholars (e.g., Cristofaro 2012; see Schmidtke-Bode 2019 for a survey). We describe our operationalization of these hypotheses in further detail below.

3.1 Data

The phylogenetic comparative methodology that we employ involves two key ingredients: a tree sample of 1000 phylogenies of the languages under investigation, and a featural representation of each language of interest.

3.1.2 Data coding

We surveyed grammars of 65 ancient, medieval and modern Indo-Iranian languages. The data collected consist of feature codings for each language according to two variables of interest, concerning (1) the presence of numeral classifiers and (2) optionality of and restrictions on plural marking. These variables show the following attested values:

1. Numeral classifiers

   1. Classifiers present: +CLF

   2. Classifiers absent: −CLF

2. Plural marking

   1. Optional plural marking in at least some contexts: −OBL.PL

   2. Obligatory plural marking in all contexts: +OBL.PL

View Full Size

Figure 1

Sample of 1000 Indo-Iranian phylogenetic trees; tip colors represent languages’ states; for languages with numeral classifiers, * indicates the presence of classifiers based on inherited matter, while † indicates the presence of classifiers based on borrowed matter (but not necessarily borrowed as classifiers per se). A maximum clade credibility (MCC) tree can be found in Figure 7. The scale represents age measured in years before present.

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

Combinations of values for these variables in our sample can be seen in Figures 1 and 2. To ensure that our inference procedure is tractable and meaningful, we keep the number of levels for each variable low. We treat languages with morphological restrictions on plural marking as having obligatory plural marking, since languages with restricted plural marking tend to mark plural number on nouns to the extent possible, whereas languages with optional plural marking choose not to mark plural number in contexts where it is possible (languages with morphologically restricted plural marking are denoted by the state MORPH.PL in the Appendix). This leaves us with a single binary feature ±OBL.PL. Furthermore, we treat all four attested combinations of values of ±CLF and ±OBL.PL as a single feature with multiple values; e.g., Bengali has the value (+CLF, −OBL.PL), as it contains contexts where plural marking is optional. This pared-down representation of the typological state space facilitates ease of phylogenetic inference; we discuss ways of incorporating more fine-grained information in the conclusion section of this paper.

View Full Size

Figure 2

Approximate locations of languages in sample, based on closest glottocode matches; ±CLF stands for presence/absence of sortal numeral classifiers, ±OBL.PL for presence/absence of obligatory plural on nouns in a language

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

3.2 Model and inference

3.2.1 CTM models of character evolution

We model changes between different feature states via a common phylogenetic comparative method, the continuous-time Markov (CTM) process of character (i.e., feature) evolution. Under such a model, transitions between different states (i.e., feature variants) take place at non-negative evolutionary RATES, the inverse of which represents the average time the system spends in a given state. Rates between different states can be found in the off-diagonal cells of the instantaneous rate matrix ; diagonal cells of the matrix take values such that rows sum to zero.¹² We place prior distributions over the rates in such that transitions occur over realistic time intervals, and infer posterior distributions of each rate, as defined below:

View Expanded

represents the observed linguistic data; represents the sample of trees. The probability of the data given a tree and set of rates, , can be efficiently computed via the Pruning Algorithm (Felsenstein 2004:251–255). Once the posterior distributions of the rates are inferred, they can be used to reconstruct the probability of a given character state at internal nodes of the tree (i.e., nodes where no data are observed). Posterior rates can also be used to carry out STOCHASTIC CHARACTER MAPPING (SCM; Nielsen 2002, Huelsenbeck et al. 2003, Bollback 2006), an iterative process which samples locations on branches of the phylogeny where changes between states have the highest posterior probability of occurring.

3.2.2 Phylogenetic hypothesis testing

The literature on the relationship between classifier presence and optionality of plural marking surveyed above makes the prediction that certain pathways of diachronic development will be highly disfavored, if not impossible. The mutational interpretation of the GSS generalization predicts that classifiers will be gained more frequently if the previous state is (−CLF, −OBL.PL) than if the previous state is (−CLF, +OBL.PL). What we term the selectional interpretation predicts that if the state (+CLF, +OBL.PL) is synchronically dispreferred, then the rate at which languages abandon this state will be higher than the rate at which they abandon the state (+CLF, −OBL.PL).

In linguistics, phylogenetic comparative methods provide a means of testing for associations between pairs of linguistic features as they evolve in a tree over time. A standard way of testing for correlated evolution between two discrete binary features, such as ±CLF and ±OBL.PL is Pagel’s 1994 Discrete model, which assesses the relative model fit of a dependent model, which constrains evolutionary rates in a manner thought to be compatible with correlated patterns of evolution, against a null, independent model, which models two independent character histories for each of the features in question (Pagel & Meade 2006, Dunn et al. 2011). In the Bayesian context, a common practice is to carry out this assessment using Bayes Factors (i.e., the ratio of marginal likelihoods for each model). In this paper, we largely depart from this framework for the reasons listed below.

First, the Discrete model tells us whether there is support for interdependent evolution, but suppresses most of the dynamics of change over the tree, including directionality of change. Since directionality is built into our hypothesis, we prefer to observe rates from a single model in order to determine whether the classifiers develop more frequently in the presence of optional plural marking—not simply whether a change in one feature correlates with a change in the other feature. Second, the Discrete model has been shown to exhibit problematic behavior under certain circumstances (Maddison & Fitzjohn 2014); in particular, scenarios in which features undergo relatively infrequent changes over the tree can be prone to false detection of the presence of correlated evolution, though this is not a problem for all datasets. Third, while Bayes Factors have traditionally been viewed as a lean way of comparing nested models, statistical science is gradually moving away from their use for the purpose of hypothesis testing in favor of alternative approaches, for both technical and conceptual reasons (Gelman & Shalizi 2013).

Given these concerns, the bulk of our analysis makes use of the computationally simpler approach of carrying out hypothesis testing within a single model (cf. Kruschke 2011) and allowing the most plausible evolutionary story to fall out of these results. Our model involves transition rates between all possible combinations of (±CLF, ±OBL.PL) as long as these rates reflect a single state change, i.e., we rule out the possibility of an instantaneous change from, say, (-CLF, -OBL.PL) to (+CLF, +OBL.PL).¹³ The model is schematized in Figure 3. In our analyses, we observe whether credible ranges of posterior values (or quantities derived from these values) are compatible with the hypotheses explored in this paper. Credible intervals (CI s) are widely used in Bayesian statistics, but there is no general consensus on the cutoff that should be employed (McElreath 2020:54 ff.). We use a 95 % CI, which is conservative in the sense that it leads us to rule out a hypothesis only if it is supported by fewer than 5 % of posterior samples. Since most of our hypotheses are one-sided (i.e., is one change type more frequent than another?), we base our assessments on the upper 95 % range of posterior values.

View Full Size

Figure 3

Character states and allowed state transitions of the model used in this paper

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

4.1 Mutational GSS generalization

The mutational GSS generalization predicts that classifiers are gained more frequently in the presence of optional plural marking than in the presence of obligatory plural marking. We quantify this difference by comparing the posterior rates for the transition q((−CLF, −OBL.PL) → (+CLF, −OBL.PL)) with those for the transition q((−CLF, +OBL.PL) → (+CLF, +OBL.PL)). Figure 5 gives these rates, as well as the difference between their posterior distributions (i.e., q((−CLF, +OBL.PL) → (+CLF, +OBL.PL)) subtracted from q((−CLF, −OBL.PL) → (+CLF, −OBL.PL)) for each sample in the posterior trace); positive values indicate a higher preference for classifier gain in the presence of optional plural marking. 98.4 % of posterior samples show a positive difference between the two rates. This value, above our 95 % CI, indicates substantial support for the GSS generalization in its diachronic form; the development of classifiers in Indo-Iranian languages appears to have been strongly influenced by the presence of optional plural marking. The ratio of the transition rate q((−CLF, −OBL.PL) → (+CLF, −OBL.PL)) to the transition rate q((−CLF, +OBL.PL) → (+CLF, +OBL.PL)) is greater than one in the majority of posterior samples; this ratio is greater than 1 in 98.4 % of samples, greater than 2 in 91.4 % of samples, greater than 3 in 81.2 % of samples, and greater than 4 in 70 % of samples (this ratio is greater than 1.58 in the upper 95 % of samples). Despite our conservative cutoff for credible ranges, it is still the case that a small proportion of samples go against the GSS hypothesis. For this reason, we investigate branch-specific developments in detail by carrying out Stochastic Character Mapping (SCM) in §4.3, which allows us to draw inferences regarding the most probable trajectory leading to the development of classifiers on each branch where they emerge in the tree.

4.2 Selectional GSS generalization

The selectional GSS generalization predicts that the state (+CLF, +OBL.PL) is synchronically dispreferred, and we expect this typological state to have a higher exit rate than the state (+CLF, −OBL.PL). The exit rate of a state can be computed by summing over all transition rates away from the state in question. Hence, the exit rates for the relevant states are the following:

View Full Size

Figure 5

Left: posterior distributions of rates for classifier gain given obligatory plural marking versus classifier gain given optional plural marking. Right: Difference between rates of classifier gain in the presence of optional versus obligatory plural marking; values greater than zero indicate that classifiers are gained more frequently in the presence of optional plural marking as opposed to obligatory plural marking.

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

Figure 6 reveals that the exit rate for (+CLF, +OBL.PL) is higher than the exit rate for (+CLF, −OBL.PL), but not substantially so; the difference in rates is greater than zero in only 75.6 % of samples. The fact that 24.4 % of samples are incompatible with the selectional GSS generalization means that we cannot rule out the null hypothesis that (+CLF, +OBL.PL) and (+CLF, −OBL.PL) are roughly equal in their stability, according to our criterion of a 95 % CI. This shows that there is nothing inherently dispreferred about the co-occurrence of classifiers and obligatory plural marking; this state of affairs may arise less frequently through diachronic change than the state (+CLF, −OBL.PL): the more frequent trajectory (−CLF, −OBL.PL) → (+CLF, −OBL.PL) may reflect a more general bias or pressure towards unitization, while the trajectory (−CLF, +OBL.PL) → (+CLF, +OBL.PL) may occur for sociolinguistic and contact-based reasons.

View Full Size

Figure 6

Left: posterior distributions of exit rates for the states (+CLF, −OBL.PL) and (+CLF, +OBL.PL). Right: Difference between exit rate for (+CLF, +OBL.PL) and exit rate for (+CLF, −OBL.PL); values greater than zero indicate that (+CLF, +OBL.PL) is abandoned at a higher rate

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

Taken together with the results of the mutational GSS generalization, these results indicate that any apparent bias against the state (+CLF, +OBL.PL) that can be detected synchronically across the Indo-Iranian languages (and perhaps beyond) appears to have emerged from diachronic preferences towards the development of certain structures in specific contexts rather than structural constraints on co-occurrence.

4.3 Character histories

View Full Size

Figure 7

Density map aggregating the most probable character histories over a Maximum Clade Credibility (MCC) tree constructed from the tree sample

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

View Full Size

Figure 8

Maximum A Posteriori (MAP) character history over a Maximum Clade Credibility (MCC) tree constructed from the tree sample. The numbers in each branch report the posterior support for the MAP transition history.

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

We carry out stochastic character mapping using the SIMMAP method (Bollback 2006) as implemented in the R package Phytools (Revell 2012). We simulate character histories on a maximum clade credibility (MCC) tree constructed from our tree sample over 1000 iterations, drawing from the posterior sample of transition rates. The standard way for visualizing the aggregation of these histories is to use a density map, which represents the probability of a state in continuous space over the tree using a color gradient. Visualization can be a challenge for more than two states, since colors can become muddy in regions where uncertainty over the state value of the character is high (Figure 7). For this reason, we supplement this visualization by an alternative one that is based on maximum a posteriori (MAP) estimates (Figure 8). We derive the MAP estimates by tabulating for each branch the counts for each type of transition history (ignoring the actual waiting times between transitions) and taking the most frequent transition history.

The figures reveal striking differences in diachronic behavior between Indo-Aryan and Iranian. In Indo-Aryan, classifiers emerge only three times (on branches ancestral to Sinhala, Agia Varvara Romani, and several Eastern Indo-Aryan languages), and their development is not preceded by a long period of optional plural marking. In contrast, an overwhelming number of cases of classifier development in Iranian are preceded by lengthy periods of optional plural marking.

To ensure that these patterns (and specifically, this difference across the two subgroups) are not simply an artifact of the topology of the MCC tree, we carry out SCM on 1000 trees drawn from the tree sample, tabulating the number of times classifiers are gained in the presence of optional plural marking versus obligatory plural marking within Indo-Aryan and Iranian. The results are summarized in Figure 9 and they indicate that for Iranian languages, classifiers develop more frequently in the presence of optional plural marking than in Indo-Aryan languages.

View Full Size

Figure 9

Number of gains of numeral classifiers in the presence of optional versus obligatory plural marking for Indo-Aryan (left) and Iranian (right) over 1000 iterations of SCM carried out using trees from the tree sample.

Citation: Language Dynamics and Change 11, 2 (2021) ; 10.1163/22105832-bja10013

• Download Figure
• Download figure as PowerPoint slide

5.1 Indo-Aryan

As observed above, the MAP tree shows three clear instances of Indo-Aryan classifier development: in Agia Varvara Romani, Sinhala and Eastern Indo-Aryan. In Agia Varvara Romani and Sinhala, classifier development was most likely not preceded by a period of optional plural marking, and in the case of Eastern Indo-Aryan, the period of optional plural marking was relatively short (and perhaps even an artifact of the character model we used, which only permits transitions involving a single change of state). These patterns favor language contact as a more plausible explanation for the presence of classifiers in Agia Varvara Romani and Sinhala, and perhaps this is even the case in Eastern Indo-Aryan due to the relatively rapid succession of cascading state changes.

Agia Varvara Romani shows clear evidence for matter borrowing of the classifier tane from Turkish (see above), possibly due to the prestige of the latter language. The same element is also found in other Romani dialects that are in contact with Turkish (Matras 2002:204). In contrast, the Sinhala classifier consists of inherited phonological material, and therefore does not point explicitly to contact in the form of matter borrowing. Our quantitative results point to a period of obligatory plural marking preceding the emergence of classifiers (MAP support = .53). Incidentally, the Sinhala classifier construction is very similar to that of surrounding Dravidian languages, especially Tamil. Sinhala denaa still has its meaning ‘people’ in other contexts (Chandralal 2010). In Modern Literary Tamil (Lehmann 1993:112–114), numerals above one have a so-called pronominalized form with a suffix -ar. In Modern Spoken Tamil (Schiffmann 1999:132–135), pronominalized numerals higher than one add the noun peeru ‘name’ instead, which can also mean ‘person’. In both languages the newly formed numeral can be used attributively or pronominally. If used attributively, it can precede or follow the head noun. In Literary Tamil, they have a marked genitive reading instead if preposed. In Spoken Tamil, they have a specific or definite reading if postposed. The similarities with Sinhala are striking. In all three languages there is [N [Num Clf]] word order as at least one possibility and there is a connection with animacy, non-animate or non-human entities being unmarked. Some Dravidian languages on the mainland have similar classifier-like constructions for humans. In Telugu, for instance, numerals above eight combine with the word mandi ‘persons’, e.g. padi-mandi ‘ten persons’ (Krishnamurti & Gwynn 1985:106–109). This makes pattern borrowing from adjacent Dravidian languages such as Tamil a likely source of the Sinhala classifier construction.

For Eastern Indo-Aryan there is some evidence that medieval varieties had optional plural marking. Mukherjee (1963:23) states that Old Bengali lacks morphological number but has a number of optional periphrastic means of expressing plurality. It is not clear whether this apparent optionality of plural marking co-existed with classifier use, as there is good reason to believe that classifier use was suppressed in literary registers, our only source of data on languages of this sort (cf. Barz & Diller 1985). Regardless of these specific scenarios, our quantitative results suggest that even if these languages went through a period of optional number marking, it was a relatively short period (Figure 8). That such changes took place relatively rapidly suggests a historical development best compatible with widespread language shift.

Support for this interpretation comes from the fact that Eastern Indo-Aryan languages with numeral classifiers are spoken in the vicinity of Dravidian (e.g., Kurux), Austroasiatic (especially Munda, e.g. Kharia), and Sino-Tibetan (especially Tibeto-Burman, e.g., Jero) languages, many of which also exhibit numeral classifiers. Striking parallels with the behavior of East Indo-Aryan classifiers can be found in the Kradai language Khamti. In Khamti, [N Clf Num] word order has an indefinite reading with the numeral ‘one’, while [N Num Clf] word order is definite. Interestingly, Assamese, Bengali, and Oriya have a connection of word order with definiteness, where [N Num Clf] word order is definite while [Num Clf N] order is indefinite.

(19) Khamti (Kradai)

kuun4maau	koo1	leeung3
bachelor	CLF	one

‘a bachelor’ (Inglis 2007:8, shortened)

(20)

kuun4	saam	koo1
person	three	CLF

‘three people’ (Inglis 2007:11, shortened)

(21) Bengali

chɔ	-ṭa	boi
six	-CLF	book

‘six books’ (David 2015:136)

(22)

boi	chɔ	-ṭa
book	six	-CLF

‘the six books’ (David 2015:137)

While further parallels and matches in morphosyntactic pattern are needed in order to make a conclusive case for a shift from a language specifically like Khamti to Eastern Indo-Aryan, the striking differences from other Indo-Aryan languages, as well as the dynamics of change shown in the evolutionary scenario that we infer, lend support to the idea that the typological profile of Eastern Indo-Aryan languages is due to contact rather than diachronic trends realized elsewhere in Indo-Iranian.

5.2 Iranian

In contrast to the situation in Indo-Aryan, Iranian languages appear to have developed classifiers more frequently; classifiers emerge on nearly a dozen independent branches within Iranian. The majority of these developments are preceded by extended stages with optional plural marking; ultimately, the state of affairs characterized by the GSS generalization appears to have been more active in Iranian than in Indo-Aryan. Our results are also fully in line with the historical record. For example, the development of full-fledged numeral classifiers in Persian was preceded by a prolonged period of optional plural marking. Our evolutionary model suggests that this was the case in several additional lineages. While these findings support the GSS generalization, contact is likely to have played a role as well, especially since Iranian languages are spoken near Turkic languages with classifiers.

Turkic and Iranian languages came increasingly into contact in post-Sasanian times. However, numeral classifiers are likely to be a relatively late development, post-dating the onset of the Turkic expansion from about the 5th century CE (Yunusbayev et al. 2015). The Old Turkic (ca. 7th to 11th century) corpus contains no sortal numeral classifiers (Erdal 2004:226), while the later literary language Chagatay (ca. 14th to 19th century) in Central Asia exhibits certain classifiers, such as baš ‘head’ (Bodrogligeti 2001:155), and classifier use in modern Turkic languages is widespread, found among different branches of Turkic, including Kipchak (e.g., Tatar, Chen Zongzhen & Yi Liqian 1986:70), Oghuz (e.g., Turkmen, Clark 1998:169), and Karluk (e.g., Uzbek, Beckwith 1998). It is not clear at present whether this distribution reflects a widespread late diffusion throughout Turkic, the effects of some sort of drift or slant-like tendency, or the inheritance of a chronologically deep feature. Complicating matters, Turkic classifiers are often Iranian loans; only in restricted occurrences do Iranian languages borrow Turkic classifiers (e.g., Sariqoli ← Uyghur). Further research is needed to address the question of whether Turkic classifiers are due to Iranian influence, Iranian classifiers are due to Turkic influence, or both groups developed classifiers as a response to loosening restrictions on plural marking under the influence of widespread multilingualism. Importantly, all of these competing scenarios attribute an important role to Turkic/Iranian contact history.

In some cases, the presence of numeral classifiers in Iranian languages points to influence from other Iranian languages. The numeral classifiers found in Wakhi, Yaghnobi and certain Pamir languages are identifiable as Tajik. Tajik’s importance as a lingua franca in the region makes borrowing into Iranian particularly plausible. Pashto has likely borrowed the classifier tana (m.)/teni (f.) from a Turkic language, though it has been in the language long enough to be integrated into the gender system and participate in certain phonological processes.

In other cases, the role of contact is less clear. Kumzari, located in Oman and separated from other Iranian languages by the Persian Gulf, exhibits the classifiers -ta and -kas (for human beings), formally identical to the Persian classifier tā and the Persian word kas ‘person’. Kumzari is phylogenetically very close to Old, Middle, and modern Persian (Skjærvø 1989), though it is not clear that it is a descendant of Old or Middle Persian. Given the relative isolation of Kumzari, it is possible that it developed these classifiers in parallel with Persian due to parallel drift, but contact with another Iranian language is also a possibility, since there has been longstanding migration to Oman from the other side of the Persian Gulf (Barth 1983).¹⁵

The remainder of Iranian languages with classifiers (e.g., Mazandarani, Taleshi, and Gilaki) share a core group of classifiers that are formally near-identical to Persian ones (e.g., tā), but as in the case of Kumzari, it is difficult to determine on the basis of sound change whether these forms are inherited or borrowed; at the same time, the dominance of Persian over these languages is well established (Borjian 2009), lending circumstantial evidence to the notion that they are Persian loans. Some Northwest Iranian languages have classifiers not found in Persian, e.g., Taleshi gəla (Paul 2011, Stilo 2018), perhaps cognate with Judeo-Tati gile ‘time, instance’ (Authier 2012:310).

A.1 Agia Varvara Romani [vlax1238] (+clf,+obl.pl)

Singular forms are always morphologically distinct from their plural counterparts, usually via the addition of a plural suffix or alternation of the final vowel (Igla 1996:23 ff.). When a numeral greater than one modifies a noun, the noun is morphologically plural (p. 45). When the counted item consists of indefinite objects, then -tane ( < Turkish tane ‘piece, part’) appears next to the numeral; anaphoric use of this classifier is obligatory (p. 45).

A.2 Assamese [assa1263] (+clf,-obl.pl)

Assamese has a large inventory of sortal classifiers. Plural marking is optional (Borah 2012, Chowdhary 2012).

A.3 Avestan [aves1237] (-clf,+obl.pl)

Plural number is consistently marked, given rich agreement morphology (Hoffmann & Forssman 2004).

A.4 Awadhi [awad1243] (+clf,-obl.pl)

Classifiers are present; information regarding plural marking is difficult to extract from Saksena 1971:115 ff., but it appears to be optional.

A.5 Bactrian [bact1239] (-clf,-obl.pl)

In late Bactrian, case and number distinctions have been neutralized due to the loss of distinctions between final vowels, resulting in “an unmarked form without ending … which may be used with either sg. or pl. reference, and a marked pl. form” (Sims-Williams 2007:40).

A.6 Bagri [bagr1243] (-clf,morph.pl)

There are at least three declensional classes, in one of which the plural and singular direct forms are identical. The distinction between animacy and inanimacy, rather than ending in -i versus other segments, is made by the author, who explicitly states that the suffix -ã is optional on animate nouns. This is almost akin to a mixture of a system like that of Hindi, where plural cannot be marked on some noun case forms, and a system where plural marking is truly optional.

A.7 Bakhtiyari [bakh1245] (+clf,-obl.pl)

Classifiers are present, and plural marking is variable (Anonby & Asadi 2014).

A.8 Bengali [beng1280] (+clf,-obl.pl)

Bengali has a large repertoire of numeral classifiers (David 2015:135). Classifiers are obligatory with non-numeric quantifiers and lower numbers; optional with numbers ending in ‘hundred’, ‘thousand’, ‘lakh’, etc. (p. 142). Numeral classifiers cannot cooccur with nouns denoting a countable unit, e.g., units of weight, currency, time, except in certain emphatic contexts (p. 142). Plural marking is non-obligatory (p. 76).

A.9 Bhojpuri [bhoj1244] (+clf,-obl.pl)

Numeral classifiers are present, and plural marking is non-obligatory (Tiwari 1960:120, 228, 230)

A.10 Dari [dari1249] (+clf,-obl.pl)

Classifier use is common, but not obligatory; plural marking is optional; it is not clear if plural marking can co-occur with classifier use as in Standard Modern Persian (Kiseleva 1985:74–75; Ioannesjan 1999:58–59).

A.11 Dhivehi [dhiv1236] (-clf,-obl.pl)

For nonhuman nouns, plural marking is optional when plurality is clear from context (Gnandesikan 2017:59).

A.12 Domari [doma1258] (-clf,-obl.pl)

Plural marking on enumerated nouns interacts significantly with whether the numbers are inherited Indo-Aryan forms or borrowed from Arabic (see Matras 2012:97, 188 ff.). Plural number is optionally marked on nouns modified by 2–3 (inherited numbers), obligatory on nouns modified by 4–10 (Arabic numbers), and optionally marked on nouns from 11 upward (Arabic numbers).

A.13 Dumaki [doma1260] (-clf,+obl.pl)

For virtually all nouns, the singular form is distinguishable from the plural form (Lorimer 1939:24 ff.); this is achieved via suffixation or a stem alternation.

A.14 Gilaki [gila1241] (+clf,-obl.pl)

Rastorgueva et al. (2012) list several classifiers. Classifier use is optional when enumerating nouns, but appears to be quite common and obligatory in anaphoric use. In all examples given of counted nouns, there is no overt plural marking on the head noun (regardless of whether a classifier is present). Plural can otherwise be marked by means of certain suffixes.

A.15 Gujarati [guja1252] (-clf,-obl.pl)

Cardona (1965:66–67) refers to the plural marker -o as “optional,” and it seems to largely be omitted when nouns are modified by a number greater than 1; so-called “variable” nouns display a special “dependent stem form” when they are semantically plural, regardless of the presence of the suffix -o.

A.16 Hindi [hind1269] (-clf,morph.pl)

Hindi shows four declensional classes; the details of number marking are different for each one. Plural marking is not morphologically possible on C-final masculine nouns in direct case (Oberlies 2005).

A.17 Iron Ossetic [osse1243] (-clf,+obl.pl)

Plural number is marked on nouns by means of the suffix -t- (Thordarson 2008 [2009]:117). In most contexts (except for contexts of enumeration, see below), use of -t- appears to be obligatory. When a noun is enumerated by a numeral greater than one, the noun is marked by the suffix -i (Digor), identical to the genitive suffix. According to Thordarson (2008 [2009]:132), this suffix continues the Old Iranian plural suffix *-ah. Nouns enumerated by numbers greater than one are always marked in a way that renders them distinct from singular nouns.

A.18 Ishkashimi [ishk1246] (+clf,-obl.pl)

Ishkashimi contains at least three classifiers; nouns modified by a numeral greater than one can appear in singular or plural form (Paxalina 1959:50).

A.19 Judeo-Tati [jude1256] (-clf,-obl.pl)

Plural is marked on nouns with the suffix -ho (Authier 2012:79). Overt plural marking on enumerated nouns seems virtually non-existent.

A.20 Kalam Kohistani [indu1241] (-clf,+obl.pl)

Kalam Kohistani achieves plural marking on a number of nouns via a vowel fronting process, which also is found in oblique forms of nouns, and appears to mark plural consistently on nouns (Baart & Sagar 2004:21). The word khur ‘foot’ may show variability in plural marking, but it is not clear from the data given.

A.21 Kalasha [kala1372] (-clf,-obl.pl)

Plural marking is optional (Petersen 2015:35–36).

A.22 Kashmiri [kash1277] (-clf,morph.pl)

According to Wali & Koul (1996:190 ff.): “plurals are formed from singular stems by vowel change, palatalization and suffixation. A few nouns stay invariant. Masculine plurals are formed differently than the feminine plurals.” Mass nouns, most body parts, and borrowed English nouns use the same forms in both the singular and the plural. Masculine nouns do not change for plurality if they have certain phonotactic properties or are borrowed from Hindi/Urdu and English with a final consonant.

A.23 Khotanese Saka [khot1251] (-clf,+obl.pl)

Plural number is consistently marked, given rich agreement morphology (Emmerick 1989).

A.24 Khowar [khow1242] (-clf,-obl.pl)

Plural marking appears to be optional on the basis of examples provided in Endresen & Kristiansen (1981).

A.25 Khwarezmian [khwa1238] (-clf,+obl.pl)

Plural is consistently marked (Durkin-Meisterernst 2009).

A.26 Konkani [konk1267] (-clf,morph.pl)

Certain noun categories have identical singular and plural endings in the direct case; otherwise, plural is consistently marked (Almeida 1989:126 ff.)

A.27 Kumzari [kumz1235] (+clf,+obl.pl)

From Thomas’s 1930 description, plural marking appears to be obligatory. Kumzari numerals are nearly identical to their Modern Persian cognates; however, from seven upwards, the Kumzari numerals all end in -tā, which is analyzed as a suffix. For human beings, a suffix -kay attaches to the number one yek(kay); for two onward, the suffix -kas is used. According to a newer description, the numeral classifier -tā or -ta in Kumzari can also occur on numerals below seven (van der Wal Anonby 2015:47)

A.28 Luri [luri1257] (-clf,-obl.pl)

According to MacKinnon (2003), plural marking is the same as in Modern Persian. No information regarding numeral classifiers is provided.

A.29 Maithili [mait1250] (+clf,-obl.pl)

Maithili has at least two classifiers (Burghart 1992:v. 1, 117). The suffix -sab(h) is an optional plural marker; when added to nouns that are inherently plural (e.g., vegetables), takes on the meaning “X and such things.” Some other suffixes exist for reference to persons, used in formal speech (Burghart 1992:v. 1, 50–51).

A.30 Marathi [mara1378] (-clf,morph.pl)

Plural number must be marked on semantically plural nouns, except where morphologically impossible, e.g., masculine kinship terms, certain loanwords (Pandharipande 1997:366–367). Emeneau (1956:11) claims that Marathi has a classifier jaṇ/jaṇị̄ (f.) that appears “when nouns denoting persons are numerated by numerals higher than four (and optionally for two to four).” Lambert (1943:243) says the following: “When the numerals refer to persons, special forms are used instead of don, tin, car; to other numerals the word zən (m. zən, cf.fem. zəna(n); f. zəni, cr.fem. zəni(n)) is usually added. This word is often added also to the special forms of don, tin, car.” No examples are given. Katenina (1963:50) gives examples of the special forms doghe, tighe, ćaughe, as well as the forms ȷ́aṇ (m.) and ȷ́aṇī (f.) ‘people’ which show the latter form as a head noun, but never in close apposition with another (head) noun. The interaction between the numerals and ȷ́aṇ(ī) is striking; however, other grammars gloss these special forms simply as ‘both’, ‘the three’, and ‘the four’ respectively (Dhongde & Wali 2009:59). These forms appear in Old Marathi as substantivized numerals, e.g., he tighe bhāu ‘these three were brothers’ (Tulpule 1963, apud Southworth 1962:425).

A.31 Marwari [marw1260] (-clf,morph.pl)

There are at least three declensional classes, in one of which the plural and singular direct forms are identical. Plural number is marked on plural nouns, where morphologically possible (Gusain 2004:20, 29).

A.32 Mazandarani [maza1291] (+clf,-obl.pl)

Classifiers are present, and plural marking is optional (Nawata 1984:9–10).

A.33 Mewati [mewa1249] (-clf,morph.pl)

There are at least three declensional classes, in one of which the plural and singular direct forms are identical. Plural number is marked on plural nouns, where morphologically possible (Gusain 2003:20, 29).

A.34 Middle Persian [pahl1241] (-clf,-obl.pl)

Middle Persian can mark plural with the suffixes -hā and -ān, but plural is frequently unmarked on plural nouns (Skjærvø 2009:223).

A.35 Modern Persian [midd1350] (+clf,-obl.pl)

Modern Persian has several numeral classifiers, the most basic and widespread of which is tā, optionally used with numbers larger than one (Windfuhr & Perry 2009:478). Plural marking is optional, but the noun being modified can be marked for plural number if it has specific reference (Mahootian 1997:195). Classifiers are obligatory in anaphoric use.

A.36 Nepali [nepa1254] (+clf,-obl.pl)

Nepali contains several numeral classifiers (Acharya 1991:100); plural number is marked with -haru; according to Acharya this marking is optional (pp. 98–99). From Acharya’s examples, -haru can can co-occur with numeral classifiers (p. 100). According to Bhim Lal Gautam (p.c.), -haru is obligatory with human nouns; however, non-human nouns cannot co-occur with overt plural marking and a classifier.

A.37 Old East Rajasthani [dhun1238] (-clf,morph.pl)

There are at least three declensional classes, in one of which the plural and singular direct forms are identical. Plural number is marked on plural nouns, where morphologically possible (Metzger 2003).

A.38 Old Persian [oldp1254] (-clf,+obl.pl)

Plural number is consistently marked, given rich agreement morphology (Kent 1951).

A.39 Oriya [oriy1255] (+clf,-obl.pl)

Oriya has several classifiers; plural marking is optional (Neukom & Patnaik 2003).

A.40 Ormuri [ormu1247] (+clf,-obl.pl)

According to Kieffer (2003:133), classifiers are used as they are in Dari.

A.41 Pali [pali1273] (-clf,+obl.pl)

Pali generally maintains a clear morphological distinction between singular and plural; although the nominative singular and plural of ā-stems fell together due to regular sound change, a secondary plural suffix came into use in order to distinguish between the two numbers (Oberlies 2001:150–151)

A.42 Palula [phal1254] (-clf,+obl.pl)

Plural is consistently marked on count nouns with one of five suffixes, which are accompanied in some cases by stem alternations (Liljegren 2016:103–104).

A.43 Panjabi [panj1256] (-clf,morph.pl)

There are at least three declensional classes, in one of which the plural and singular direct forms are identical. Plural number is marked on plural nouns, where morphologically possible (Bhatia 1993:214–215).

A.44 Parachi [para1299] (-clf,-obl.pl)

Little information about interactions between numeral modification and number marking can be found in Kieffer (2009). According to Morgenstierne (1929:50), plural marking is optional, and rare when a numeral modifies the noun. No information regarding classifiers is given.

A.45 Parthian [part1239] (-clf,-obl.pl)

Durkin-Meisterernst (2014:272) describes a scenario for all of Middle West Iranian whereby plural marking is optional on all enumerated nouns, but animate nouns are more often marked for plural number than inanimates.

A.46 Pashto [pash1269] (+clf,+obl.pl)

Plural number appears to be consistently marked on Pashto nouns (Penzl 1955:45 ff.); most paradigms are relatively complex and contain stem alternations. Numeral classifiers are possible, and co-occur with nouns marked for plural number, possibly with gender agreement (p. 82).

A.47 Prakrit [maha1305] (-clf,+obl.pl)

Plural number is consistently marked, given rich agreement morphology (Woolner 1928).

A.48 Rakhshani Baluchi [west2368] (-clf,morph.pl)

Nouns can be marked for indefiniteness and singularity via the suffix -e; otherwise, there is no morphological distinction between singular and plural (Barker 1969:3 ff.).

A.49 Sangesari [sang1315] (-clf,morph.pl)

Classifiers do not exist in Sangesari. According to Azami & Windfuhr (1972:70 ff.), plural is consistently marked on oblique nouns with the suffix -uon, but rarely on direct nouns, except for a restricted set of items.

A.50 Sanskrit [sans1269] (-clf,+obl.pl)

Plural number is consistently marked, given rich agreement morphology (Macdonell 1910).

A.51 Sariqoli [sari1246] (+clf,-obl.pl)

Classifiers are present, and plural marking is optional (Paxalina 1971).

A.52 Saurashtran [saur1248] (-clf,morph.pl)

Plural number is marked on Saurashtran nouns by means of two suffixes, -nu and -lu ( < Telugu) (Ucida 1979:45–46). When a numeral greater than one modifies a noun, the noun is plural, but certain non-human nouns (e.g., days, years) do not take plural form.

A.53 Shina [shin1264] (-clf,morph.pl)

Plural appears to be consistently marked on count nouns (Schmidt et al. 2008).

A.54 Shughni [shug1248] (-clf,-obl.pl)

Plural marking is optional, but classifiers do not appear to be present (Zarubin 1960).

A.55 Sindhi [sind1272] (-clf,morph.pl)

There are multiple declensional classes, in one of which the plural and singular direct forms are identical. Irregular plurals can be found for kinship terms. Arabic words often have distinctive plurals borrowed from the source language. Plural number is marked on plural nouns, where morphologically possible (Egorova 1966:27–28).

A.56 Sinhala [sinh1246] (+clf,+obl.pl)

When animate nouns are modified by a numeral, the form of the numeral used is different from that which is used when an inanimate noun is modified by a numeral (Chandralal 2010:60). Plural marking is obligatory, when applicable.

A.57 Siraiki [sera1259] (-clf,morph.pl)

Certain noun categories have identical singular and plural endings in the direct case; otherwise, plural is consistently marked (Shackle 1976)

A.58 Sogdian [sogd1245] (-clf,-obl.pl)

Sogdian heavy stem nouns occasionally show a form that is identical to the singular in plural contexts; this behavior displays variation (Yoshida 2009:313).

A.59 Sorani Kurdish [cent1972] (-clf,-obl.pl)

According to Blau (1980:45–46), the simple form of a noun can have either a singular or plural reading. In general, plural number is marked with the suffix -an.

A.60 South Tati [esht1238] (-clf,-obl.pl)

According to Yar-shater (1969:78), “nouns modified by a numeral higher than one, or by an expression denoting plurality, are generally expressed in the plural in Chali [a particular dialect].” Occasionally, however, the singular is used. In the other dialects, normally the singular is used for enumerated nouns. Plural marking otherwise seems to be the norm.

A.61 Taleshi [taly1247] (+clf,-obl.pl)

Taleshi has a number of numeral classifiers, use of which is non-obligatory (Paul 2011:181–182); additionally, “any noun following a numeral phrase is generally in the singular.” Elsewhere, plural number is marked with a suffix that varies from dialect to dialect.

A.62 Torwali [torw1241] (-clf,morph.pl)

According to Grierson (1929:34), if a noun ends in a vowel, it can take a plural-marking suffix -e; otherwise, the singular and plural forms are identical.

A.63 Wakhi [wakh1245] (+clf,-obl.pl)

Native Wakhi numeral forms are in competition with Tajik numeral forms (Grjunberg & Steblin-Kamenskij 1988:89–90). The word for twenty (bist) is borrowed from Tajik, but numerals 20–30 can combine bist with either Wakhi or Tajik forms in the digits place. A number of classifiers are borrowed from Tajik. Unlike the situation in Yaghnobi, these classifiers can be used with both Wakhi and borrowed Tajik numerals. Plural is marked with the suffix -iš(t) (p. 19), but this marking appears to be optional.

A.64 Yaghnobi [yagn1238] (+clf,+obl.pl)

According to Xromov (1972:21–22), when numbers two and upward combine with nouns, the noun is found in the oblique singular form; if measure terms are used, the measure term is marked for oblique singular. The numerative ta, borrowed from Tajik, can be used, but only with Tajik numerals. Outside of the context of enumeration, -t/d is consistently used to mark plural number on nouns.

A.65 Zazaki [diml1238] (+clf,-obl.pl)

According to Paul (1998:19 ff.), a morphologically singular noun can be used in a generic sense, but nouns denoting a plurality, definite or indefinite, take the plural ending. Plural marking is non-obligatory. An apparent sortal classifier teney co-occurs with nouns marked both for singular and plural.