Attributive modification in South American indigenous languages

3.1 Nominal expressions (‘nominal groups’ or non-integral NPs)

This type of construction involves two constituents which form a unit semantically, but the syntactic bond is loose. The latter is evidenced from such characteristics as a flexible relative order, non-contiguity, phrasal morphology appearing on each individual constituent instead of a single phrase-final position, and prosody. Each constituent typically forms a referring expression by itself (see Rijkhoff 2002: 19, 22), but it seems possible to say that one constituent conveys information about the other, thus, in a sense, there is some restriction of reference at play. Obviously, for non-integral NPs one cannot speak of a ‘head’ and a ‘modifier’ sensu stricto. I adopt Rijkhoff’s (2002: 22) term ‘appositional modifier’ as a proxy for a constituent that involves a property concept, and I will use an analogous term, ‘appositional head’, for a constituent which is the head semantically.^[9]

Nominal expressions in the sense outlined above have been reported for the languages of the Cariban family (Meira and Gildea 2009). An ‘appositional modifier’ and an ‘appositional head’ are “pragmatically assumed to refer to the same real-world entity” (Meira and Gildea 2009: 114). Property concepts as in (1) are encoded mainly by nouns and adverbs in this family. In many, if not most, Cariban languages these are syntactically independent constituents (Meira and Gildea 2009: 116). One piece of evidence for their syntactic independence is the lack of obligatory contiguity and constraints on constituent order: a property term can either precede the ‘appositional head’ (4a) or follow it (4b) or be non-contiguous with it (4c). Another piece of evidence that the constituents do not form a syntactic unit is prosody. For example, in Hixkaryana (5) and Macushi (6) there is typically a pause in between (Abbott 1991: 70; Meira and Gildea 2009: 114), indicated in the examples with a comma.

Meira and Gildea (2009: 114) suggest that constraints on constituent order may be emerging in Macushi, where the ‘appositional modifier’ would precede the ‘appositional head’. However, these scholars also note that there is no phonological or morphological evidence for phrasal unity. In an earlier account of Macushi by Abbott (1991: 31, 70, 87–88) no constituent order constraints are observed, and constituents are noted to frequently “occur in a discontinuous sequence” (Abbott 1991: 70). Furthermore, such nominalized constituents usually occur with an intonation break (Abbott 1991: 70), in accordance with Meira and Gildea. Example (6) shows a property term aˈneˈ ‘hot one’, which follows its ‘appositional head’. The constituents are separated by a pause, and each takes a postposition.

Besides the Cariban languages, there are a few other languages in the sample in which attributive modification can be analyzed in terms of nominal expressions. This mainly concerns languages spoken in the Western and South-Western part of the Amazon, which possess rich inventories of classifiers, often of the ‘multifunctional’ type (Krasnoukhova 2012: 194). The classifiers in question combine a number of functions: semantic categorization, the function of derivation, and referential function (Payne 1987; Derbyshire and Payne 1990; Aikhenvald 2000; Grinevald and Seifart 2004; Seifart and Payne 2007; inter alia). Typically, such classifiers occur in ‘multiple environments’ (Aikhenvald 2000), viz. on numerals, nouns, verbs and – most relevant in our case – on roots denoting properties. Doris Payne (1987: 39) suggests that reference tracking may be the primary function of classifiers in these languages. Payne specifies that classifiers “may be used to avoid full noun phrases and expanded noun phrases in particular” (1987: 39). Concretely, when a classifier occurs “in any syntactic context (numeral, demonstrative, attributive modifier, predicate nominal, or another noun), it indicates sufficient features of the referent such that a more specific head noun may not be pragmatically necessary in a given context” (Payne 1987: 39). Let us take Bora for illustration. Example (7) shows a few different uses of the root imi denoting ‘good’. In (7a), ‘good’ is the predicate, evidenced from its high tone on both syllables, and the Ɂ ending (Thiesen and Weber 2012: 225). In (7b) ‘good’ is an adverbial modifier of the predicate ‘to sleep’. In (7c), ‘good’ functions attributively, always taking a classifier (Thiesen and Weber 2012: 224, 226; Seifart 2005). Attributive constructions such as (7c) are analyzed by Thiesen and Weber (2012: 225) as two NPs in apposition (see also Seifart 2005: 150–156).

Property roots combined with a classifier, like those in (7c) and (8), form referential expressions by themselves, with the classifier functioning as the head. Specifically, classifiers in Bora are nominal elements, and some have corresponding free nouns (Seifart 2005; Thiesen and Weber 2012: 182). Classifiers can stack, as shown in (8a), and it is the final classifier in that case that sets “the semantic domain within which the referent is to be found” (Thiesen and Weber 2012: 179). This is an indication that classifiers head their phrases. Just like noun roots, classifiers in Bora can denote and refer to “objects that prototypically can be localized in space and persist over time” (Thiesen and Weber 2012: 184). A difference between noun roots and classifiers is that the former have specific meanings, whereas the latter have less specific meanings (Thiesen and Weber 2012: 184).

Seifart (2005: 151) notes that even when appositional NPs occur adjacent to each other and one NP restricts the reference of the other, they show “a high degree of syntactic independence”. The following pieces of evidence suggest this (see also Thiesen and Weber 2012: 190, 223):

1. Any noun phrase can be omitted.

2. Case markers, the restrictive marker, clitics, or even other words, can intervene between noun phrases in apposition.

3. The order of the noun phrases in apposition is relatively free.

4. Noun phrases in apposition always form separate tonal phrases.

5. There is often a prosodic break (e.g., a pause) between the noun phrases in apposition.

(Seifart 2005: 151).

We see a similar phenomenon in other languages with multifunctional classifiers. Example (9) is from Cubeo. Some attributive structures in Cubeo show signs of constituency (e.g., variable order, although with a certain tendency and oblique case markers occurring only at the phrase boundary) and may thus be analyzed as semi-integral NPs (like Kwaza in Section 3.2).

While there is a good amount of evidence to analyze property roots used attributively in Bora (and probably in some cases in Cubeo) as nominal expressions, i.e., non-integral NPs, one can note two features which, at first sight, seem to contradict this analysis. First, the NP with a property term constitutes a relative clause, formally marked by tone as subordinate (Seifart 2005: 133). However, the fact that a relative clause constituent can stand on its own, and that no further clues of syntactic dependency can be observed (see Seifart 2005: 155) suggests that a constituent involving a property concept is syntactically independent. Second, the use of classifiers on all constituents, as in (8) and (9), can be taken as a form of grammatical agreement and thus syntactic dependency between the constituents. However, a true agreement system typically involves a division of all nouns into rigid classes and shows an obligatory and predictable occurrence of an inflectional marker. As shown by Seifart (2005), classifiers in Bora primarily function to establish co-reference in discourse and not to mark syntactic structures. For example, there is some alternation possible in the choice between so-called general and specific classifiers within one clause, which rather depends on “the degree to which a speaker wishes to specify a referent at a given point in discourse” (Seifart 2005: 172). Furthermore, it is the derivational capacity of classifiers (viz. their capacity to add semantic content) that ensures referent identification and syntactic independence from the referent noun (see Seifart 2005: 173–174).

Table 2 gives a summary of characteristics of nominal expressions discussed in this section.

In Bora and in the Cariban languages, nominal expressions constitute the only available strategy for attributive modification. In Rießler’s (2016) world-wide (although unbalanced) sample, only one language out of 71 is of this type, viz. Cariban Hixkaryana. In Rijkhoff’s (2002: 133, 138) sample, this is the case for two out of 52 languages (Nunggubuyu in Australia and, again, Hixkaryana in South America). While it may be a consequence of the samples used, such a rare occurrence suggests that relying exclusively on nominal expressions is typologically relatively uncommon. Tukanoan Cubeo and Kotiria constitute cases where attributive constructions involve a mix of those with characteristics of nominal expressions and those with some signs of constituency. Furthermore, nominal expressions are also found in languages that do have simple integral NPs for attributive modification (e.g., Matsés, Hup, Southern Aymara).

3.2 Semi-integral NPs

In some languages with multifunctional classifiers, the NPs can display some characteristics of syntactic unithood. Cubeo and Kotiria were already mentioned as possible cases. I will illustrate the category with Kwaza though, as the relevant variation is more evident. Property concepts in Kwaza are encoded by verbal roots, and these must be nominalized to occur attributively. Nominalization can be done either by a semantically neutral classifier (10a) or by a specific classifier (10b) (van der Voort 2004: 181, 2006). The inventory of classifiers ranges from etymologically opaque to etymologically transparent ones, and some classifiers are more semantically specific than others (van der Voort 2004: 137–176). Referent nouns are regularly omitted when specific classifiers are used, as “most classifiers display a predictable meaning in most constructions” (van der Voort 2004: 176). As shown in (10b), the referent (a’xy ‘house’ in this case) can be omitted, since the property root occurs with a classifier that is semantically specific enough. Example (10a), which shows the neutral classifier -’hỹ, can also occur with the specific classifier -’xy for ‘house’, in which case the referent noun can be omitted (Hein van der Voort p.c.). (10c) and (10d) illustrate a full NP construed by another specific classifier: -’mũ, which refers to juice, lemonade or fermented beer or the like. This particular classifier comes from the lexical form e’mũ ‘liquid’ (van der Voort 2004: 157).

Examples (10a) and (10b) each contain two NPs, viz. ‘house’ and ‘new/old house’. A head and a modifying constituent usually occur contiguously; however, there are also cases when a predicate occurs in between (Hein van der Voort p.c.). The order of constituents in attributive constructions is usually ‘head’–‘modifier’, but it can also be reversed (van der Voort 2004: 104, van der Voort 2006: 90). While both constituents are referring expressions on their own, they seem not to exhibit the same amount of syntactic freedom as those in Bora. Based on these characteristics, I treat constructions that involve semantically specific classifiers in Kwaza (as 10b) as ‘semi-integral’ NPs. Constructions that involve semantically neutral classifiers (as 10a) seem to show even stronger characteristics of an integral NP: the fact that the referent noun cannot be omitted suggests that we have a construction with some syntactic dependency and structure. The modifying constituent with a semantically neutral classifier serves for a restriction of reference rather than for reference itself (Hein van der Voort, p.c.). Therefore, I suggest that attributive constructions with property concepts in Kwaza range from those with weaker evidence for one integral NP to those with stronger evidence for it. More generally, it illustrates that a language does not necessarily have one particular type of NP in terms of constituency and could be seen as a system in flux.

Table 3 gives a summary of characteristics of attributively used property roots in Kwaza with either specific or semantically neutral classifiers.

Sample languages with rich of classifier inventories and thus showing a similar profile with respect to attributive modification (Tariana, Yagua, Mako, Kanoê, Cubeo, and Kotiria) are concentrated geographically in the Northwest Amazon (Payne 1987; Derbyshire and Payne 1990; Aikhenvald 1996: 84, 86; Seifart and Payne 2007) and in the Southwest Amazon region (van der Voort 2005). Thus, attributive modification relying on structures such as nominal expressions and semi-integral NPs can be suggested as one of the areal traits.

3.3 Complex NP: relative clause or nominalization

As shown in Table 1, in many SA languages property concepts are encoded by verbal roots. In many of these languages such verbal roots can occur as modifiers when nominalized or used as part of a relative clause. In SA, nominalization is a common strategy for dependent clause marking, including relative clause formation (van Gijn 2014). Kukama-Kukamiria is an example. (11a) shows a property concept (‘be white’) which is encoded by a verbal root taking the nominalizer -n (see also (2) above). (11b) illustrates the same nominalizer for a relative clause formation with an active verb. A nominalized relative clause can occur before or after the head noun (Vallejos 2016: 468–469). However, no pauses or intervening material can occur between the head noun and the nominalized verb (Rosa Vallejos p.c.). Besides, if we were to add a quantifier to (11a), it would appear before the head noun and would extend its scope to the nominalized verb (e.g., ‘a lot of white clay’) (Rosa Vallejos p.c.).

Besides nominalization, some languages use a relative clause construction with a relative clause marker. As a result, attributive modification constructions constitute (minimal) relative clauses. This is shown in (12) for Bororo. The relative clause marker wü is always the final element of a relative clause (Crowell 1979: 109; Nonato 2008: 142).^[11] Relative clauses that involve property terms can either precede or follow the noun they modify, with no difference in meaning; the latter order, however, is noted to be more common (Crowell 1979: 112; Nonato 2008: 144). Example (12a) illustrates the predicative use of the property concept ‘be good’. In (12b) the same property term is part of a relative clause (marked by the relativizer wü) that functions as a modifier of the referent noun ‘fish’. Note that the referent ‘fish’ is plural: the noun itself is in plural form, it is also marked by the plural marker ge following the relative clause, and the transitive verb ‘kill’ takes the plural object prefix. As (12b) also shows, the subject prefix on the verb in the relative clause (viz. on the property root) is always singular, regardless of the number of the referent (Crowell 1979: 109; Nonato 2008: 144).

There is no information about prosody for the relative clauses in Bororo. It is not clear if the occurrence of the relative clause marker can be taken as an indication of syntactic dependency between a relative clause and its head noun. Adjacency of the two constituents seems to be required, just as the presence of the referent noun. I tentatively treat this construction as a complex but integral NP.

Table 4 is a summary of characteristics of property roots in Kukama-Kukamiria and Bororo.

A complex NP construction involving either a relative clause or a nominalization is found in 18 sample languages. In 12 languages of these languages, this is the only strategy. Attributive modification realized either by nominalizations or relative clauses is common crosslinguistically, adhered by languages in which property concepts are encoded by verbs.

3.4 Complex NP: as a possessive construction

Another complex NP construction for attributive modification is formally identical to the one encoding inalienable possession. It is known from typological studies that some languages do not differentiate between modification by property terms and alienable possession (Gil 2013), and in some languages there is no formal difference between modification by property terms and inalienable possession (Nikolaeva and Spencer 2019: 77). In some languages the shared strategy involves a simple juxtaposition of two unmarked constituents. The construction that I am concerned with in this section involves a morpho-syntactically complex NP: the two constituents are linked by a morphological marker.

Consider Wari’. In this language, some property concepts are argued by Everett and Kern (1997) to be encoded by nouns and some by verbs. When properties encoded by nouns are used attributively, the construction is formally identical to the one marking possession (Everett and Kern 1997: 332): the constituent encoding the possessed entity takes a possessive marker (either a suffix or a clitic), which inflects for person, number and gender of the possessor (Everett and Kern 1997: 229–230). It must be mentioned that the occurrence of property concepts as modifiers is much less frequent than their occurrence as predicates (Joshua Birchall, p.c.). If a property term does occur as modifier, the first nominal takes a possessive marker (Everett and Kern 1997: 332). The relative order of the head and the modifier is fixed.

It is of interest to note that the semantic and syntactic dependency between the attributive modifier and the head is reversed here (see also Rießler 2016: 37). Rießler (2016: 38) offers the following analysis of such constructions: “the semantic dependent is ‘upgraded’ to the function of the syntactic head of the phrase, i.e., the possessed”, whereas the semantic head is “syntactically ‘degraded’ to the (dependent) possessor function”. In the next section I show that Hup allows a similar strategy, although in Hup it is not the only strategy for attributive modification.

Table 5 is a summary of characteristics of property roots in Wari’.

In the present sample, this construction is found only in two languages (Wari’ and Hup), suggesting its relative infrequency in South America. While Hup allows also simple NPs, this is not the case for Wari’ (and where the general preference goes to a predicative use of property terms). In Rießler’s (2016: 254) world-wide sample, only two out of 71 languages use the possessive construction for property modification: Saliba (Oceanic) and Wari. This strategy seems pervasive in Oceanic languages though (Rießler 2016: 38, referring to; Ross 1998).

3.5 Simple NP

Constructions that are subsumed under ‘simple NP’ involve two constituents that are morphologically underived and form one syntactic unit. The indications for unity include obligatory contiguity and fixed constituent order, phrasal markers, grammatical agreement between the two constituents, obligatory presence of the head, and finally, prosodic clues.

A simple NP can be found in Cavineña. The language has two types of property terms: (i) a small number of property terms which are used attributively with no morphological modification, and (ii) a large (and open) number of properties that mainly occur predicatively. The latter can be used as modifiers when relativized. An example of a property term of the first type is shown in (14). Both the head noun and the modifier form one phrase, since (a) nothing can intervene between the two constituents (Guillaume 2008: 466) and (b) a phrasal clitic – in this case, the locative marker =ju – occurs on the final constituent of the NP.

Property terms of the first type never occur on their own and obligatorily require a head noun (Guillaume 2008: 466). In fact, the strength of the bond between these constituents even calls for an alternative analysis in terms of compounding (Guillaume 2008: 474). However, Guillaume (2008: 474) prefers to treat attributively used property terms as independent modifiers. The reason is that attributively used properties can have an independent pitch contour, and even a pause is possible between the head noun and the modifying property. Besides, at least some property terms are productive: when combined with a noun the compositional meaning is transparent (Guillaume 2008: 474). In Section 3.6, we return to attributive constructions that constitute compounds (note, however, that non-compositional semantics will not be presupposed as a central feature of such compounds).

It is also the case that property terms can resemble (inalienably) possessed nouns. Take Hup as an example. Property concepts denoting dimension, age, value and color belong to a separate adjective class, although they share some characteristics with both nouns and verbs (Epps 2008: 114, 442). These property terms can modify a noun directly with no additional morphology, forming a clear syntactic unit: the modifier follows the head, phrasal markers (viz. plural and case markers) attach to the final constituent, and phonologically they are treated as “a single lexical unit, with primary stress on the second constituent” (Epps 2008: 328, 326, 187). Example (15) illustrates the point.

However, these property terms must always be preceded by some nominal form. In (15) it is the noun (pídiya ‘battery’). To occur independently property terms have to take the preform tih=, which is a 3rd person singular pronoun (Epps 2008: 327). This preform tih= has a nominalizing function, evidenced by a number of behavioral characteristics (for example, property terms with the preform are negated by a nominal negator instead of a verbal one) (Epps 2008: 332). An attributive construction can also consist of the head noun and a property term preceded by the preform, viz. [N tih=A]^[12]. This construction is illustrated in (16a). Similar to cases discussed in Section 3.4, this construction can show characteristics of one complex but integral NP, evidenced from the presence of a single phrasal marker and the absence of a pause between the constituents (Epps 2008: 332). However, these elements can also behave as a nominal expression (i.e., non-integral NP), similar to cases discussed in Section 3.1. This is seen from the fact that each constituent can take phrasal morphology. The latter is illustrated in (16b), where the object marker -ǎn can occur on the noun (‘man’) when the modifier (‘big’) carries the preform tih= (Epps 2008: 180).

Thus, Hup allows for at least three types of attributive construction with property terms: a simple NP, a complex integral NP, and a non-integral NP. As noted by Epps (2008: 332) speakers judge the constructions [N tih=A] and [N A] to be interchangeable, but there is some semantic difference. The former seems to place the property or quality in focus and indicate a quality that is permanent or inherent. The latter construction may indicate a quality which is either permanent or temporary (Epps 2008: 332). Whereas the preform is not obligatory, some color terms are observed to almost always occur with it, and a few property concepts, such as tuhúp ‘beautiful, new’, and tidó ‘red’, have even undergone phonological reduction (Epps 2008: 333).

The construction [N tih=A] is similar to the ones we saw in Section 3.4 in Wari’. Syntactic clues for headedness are few in Hup, but it is possible to analyze the property term as the head and the preform as the modifier (Epps 2008: 329–330). While the preform is a personal pronoun in origin, synchronically in this context it functions as a nominalizer and a marker of inalienable possession, and also as a linker, signaling the relationship between the head and the modifying constituent (Epps 2008: 333, p.c.). Epps (2008: 328–329) notes a formal resemblance between property terms with tih= and bound (and inalienably possessed) nouns. While synchronically the two constructions show some differences, Epps (2008: 331, 333) suggests that these constructions are very likely to be historically and/or functionally related. In fact, the use of tih= with property terms is likely to be an extension of its use with bound nouns, which probably did emerge in inalienable possession constructions (Patience Epps, p.c.). Epps (2008: 333) also raises the question of a possible further development of the preform into an attributive marker,^[13] which finds support crosslinguistically.

In this light it is of interest to briefly illustrate attributive constructions found in Macro-Ge. First of all, it should be mentioned that studies on Macro-Ge show some disagreement on the analysis of property terms as belonging to the category of verbs or the category of nouns. This is due to the fact that the two categories have a number of features in common (Oliveira 2003: 244; Ribeiro 2012: 212–214; inter alia). Among the features shared between nominal and verbal roots in Macro-Ge is the occurrence of a so-called ‘relational prefix’.^[14] Existing analyses of its function are not uncontested among Macro-Ge specialists. According to Rodrigues (1999: 181), it functions as a linker, i.e., an indicator of “textual contiguity”, marking the head of a nominal, verbal and postpositional phrase. Rodrigues (1999: 181) notes that “[t]his device must be a very old one in the Macro-Ge stock”; in some languages of the family it is a disappearing system with very restricted occurrences. The actual forms of relational prefixes differ among the Macro-Ge languages. Some languages have different allomorphs of the relational prefix, and its occurrence can be phonologically conditioned. For example, in Apinajé, a relational prefix occurs only on vowel-initial roots (Oliveira 2003: 268, 2005: 182–190, 197).^[15] What we find in Macro-Ge is a constructional parallel between attributively used property terms and inalienable possession, and the reverse semantic and syntactic dependency – just like in (13) from Wari’ (and possibly in (16a) in Hup). The head of a phrase always follows its dependents, and it is the head that occurs with a relational prefix (Rodrigues 1999: 181; Oliveira 2003: 269). Example (17) from Canela-Krahô, a Northern Ge language, and (18) from Karajá, a language forming its own branch within the Macro-Ge stock, illustrate constructions with a relational prefix. One could have treated these constructions on a par with possessive constructions in Wari (Section 3.4); however, since a relational prefix functions as a contiguity marker, not limited to possessive constructions, I discuss these constructions in this section.

Discussing relational prefixes in Apinajé, Oliveira (2003: 251–252) mentions that “[i]t is possible that these forms are old third-person morphemes”, and synchronically, if present, “they are a strong and systematic indicator of constituency”. However, since a relational prefix can be absent on some roots, the issue of constituency does seem to arise. We discuss this issue in Section 4.

Table 6 gives a summary of characteristics as discussed for Cavineña, Hup and Canela-Krahô.

Attributive modification having a structure of a simple integral NP is possible in 28 out of 65 sample languages. While in 19 of these languages the simple integral NP seems to be the only or the main strategy, the other nine languages use additional strategies for attributive modification. We can conclude that a construction of a simple NP is far from being a dominant (or typical) strategy for attributive modification in SA languages.

3.6 Attributive modification as compounding

In some SA languages the function of modification is carried out by property terms which morphologically constitute bound roots. An attributive construction has a structure of a complex word or a compound. A definition of compounding from a crosslinguistic perspective remains a challenge, with Lieber and Štekauer (2009) devoting a whole chapter to this problem. Compounds usually involve lexemes (Bauer 2009: 343), but in many SA languages spoken in the Western and South-Western Amazon, it is hard to draw a line between bound lexemes and ‘lexico-grammatical’ elements,^[16] viz. classifiers. Thus, I will use the term ‘compound’ in a loose sense: it will refer to a combination of roots, one of which can have a semi-grammatical status, yielding a morphologically complex noun which constitutes one phonological word. Note that the present term does not presuppose non-compositional semantics. I will now discuss constructions of attributive modification that arguably involve compounding as discussed above.

Consider Arawakan Baure. In Baure, some property concepts denoting dimension, age and value are encoded by nominal lexical roots which cannot occur in isolation (Danielsen 2007: 156). When modifying an obligatorily possessed noun, the two elements form a compound.

Baure is a language with a rich classifier inventory. The borderline between classifiers and lexical roots is fuzzy: these notions are ‘scalar’, depending on the amount of lexical information that an element encodes (Admiraal and Danielsen 2014: 82). Except for a small group of highly grammaticalized classifiers, most classifiers in Baure are noted to be semantically transparent and have a less grammaticalized character, and thus could be treated on a par with lexical roots (Admiraal and Danielsen 2014). Compounding is a productive process in Baure, and, as Danielsen (2007: 134) notes “some of the classifiers have been derived from reduced noun stems, which could mean that compounding was one of the ways in which classifiers evolved in the first place”. Thus, one can analyze a property root occurring with a classifier as a derivation, but it can also be analyzed as a compound, since these do not differ formally (Admiraal and Danielsen 2014: 100).^[17] Thus we get constructions such as one in (20), with a classifier -ar(o) for liquids.

Classifiers frequently occur as ‘head nouns’, in the sense that the resulting complex word is referential and can occur on its own (Danielsen 2007: 169). A referent noun can also be present: Example (21) shows the property concept ‘big’ and the numeral ‘three’, both taking the classifier -wok referring to trees, and a noun ewokoe’ ‘tree’ is also present. However, the fact that all three nominals can be used on their own (Admiraal and Danielsen 2014: 102) suggests that in this usage they can be analyzed as nominal expressions (similar to those discussed in Sections 3.1).

We find a parallel situation in Movima. Haude (2006: 112) argues for an existence of a separate adjective category in Movima, the members of which share many characteristics with nouns. In general, property concepts can occur as modifiers either as part of a relative clause or by forming a compound with a noun they modify semantically (Haude 2006: 115–116). Some property concepts (new, fat, hard, unripe/raw, cf. Haude [2006: 113–114]) seem always to occur as part of a compound. Similar to Baure, this means that a property term can attach to a bound lexical root (22a) or a classifier (22b).^[18] In either case, a property term functions as a semantic modifier, and the lexical root or a classifier as the semantic head.

Presumably, the classifier -ba for round objects has its origin in the noun ba∼ba:kwa ‘fruit’, which decomposes into a reduplication of the bound form ba and a derivational suffix -kwa (Haude 2006: 212, also fn.103). In many examples with the classifier -ba, a lexical noun for ‘fruit’ is not overtly present, but the translations indicate that the classifier is specific enough to evoke the meaning ‘fruit’ (e.g., Haude 2006: 481, 484, 521). Cases like this are borderline between the process of derivation (of a property root by a classifier) and the process of compounding (two lexical roots), and the analysis would presumably vary depending on individual elements involved, as their synchronic status and semantic transparency are likely to vary.

It should be mentioned for Movima that a few property concepts, among which are mere’ ‘big’ and tochi’ ‘small’, can also occur in simple juxtaposition with the noun. In the case of juxtaposition, there is a short pause between the two constituents and the word order is different: a property concept follows the noun (Haude 2006: 115–116). This is the same constituent order as in a relative clause construction, but without a relativizer. Interestingly, these two property concepts can also be used as adverbs to modify verbs (Haude 2006: 116). It is not clear whether different constructions with one and the same property concept entail different interpretations, for instance, in terms of permanent or temporary qualities.

A compounding strategy can also be structurally close to noun incorporation. In the Takanan language Ese Ejja, there is a limited number of property concepts that can be used attributively. A similar situation is also reported for its sister language Cavineña, as discussed earlier. While for Cavineña, a noun and an attributively used property concept are best analyzed as two constituents forming one NP (see Guillaume 2008: 474), in Ese Ejja a noun and an attributively used property concept form a single phonological word (Vuillermet 2012: 324). However, the majority of property concepts in both languages are so-called ‘predicative adjectives’. Morphologically they are marked with a semantically neutral prefix kya- when used in a citation form (Vuillermet 2012: 563). Interestingly, predicative adjectives can incorporate their semantic head nouns. Vuillermet (2012: 539, 556) observes that this strategy is productive in Ese Ejja and can be regarded as a functional approximation of attributive modification. Examples in (23) illustrate this case.

Compounding (property term and the head) is found at least in five of 65 sample languages, i.e., Baure, Trinitario, Ese Ejja, Movima, and Cubeo. In these languages compounding constitutes just one of the strategies (with a preference for predication in the former three languages). Except Cubeo, all other languages are spoken in the Guaporé-Mamoré region, which is shown by Crevels and van der Voort (2008) to be a linguistic area. While compounding is not mentioned among the shared features of the area, I propose that it should be explored further for areal clustering.

Interestingly, this strategy is noted by Dahl (2004: 225) to have received relatively little attention in linguistic literature. Dahl drew attention to a number of languages in which compounding with property terms is a regular strategy (viz. Lakota (Siouan), Chukchi (Chukotko-Kamchatkan), Burmese (Sino-Tibetan) and some traditional dialects of Northern Scandinavia), in some cases with both tighter and looser compounding structures (Southern Ute (Uto-Aztecan), Celtic and Romance languages) (see Dahl 2004: 225–236). Rießler (2016: 246–251) reports this strategy also for Adyghe and Karbardian languages in the Caucasus, Sarcee (Athapaskan) and Cayuga (Iroquoian) in North America, in Itelmen and Siberian Eskimo Yupik. Property-noun compounding is also found in Australian languages, for example, in Gunwinyguan languages from Arnhem Land (Baker and Nordlinger 2008). Furthermore, the strategy is noted to be pervasive in Sub-Saharan Africa (Creissels 2016), where this is a regular strategy for attributive modification at least within the Mande and Gur language families (Creissels 2016). SA languages remained under the radar so far regarding the compounding strategy for attributive modification.

3.7 Attributive modification by means of classifiers

Qualities and properties denoting shapes and physical characteristics are not necessarily rendered by lexical means but can be encoded by classifiers. It has already been noted by Payne (1990: 220) that in Western Amazon languages, which typically have few morphologically underived adjectives, classifiers can approximate the function of attributive modification. In the previous section, we saw a couple of languages in which classifiers on a property term are argued to function as the head of the phrase. However, in some cases classifiers can be argued to function as modifiers. The main piece of evidence is semantic: such classifiers add to the description of certain physical characteristics of a referent noun.

Let’s take Apurinã as an example. Property concepts denoting dimension, shape and consistency can be expressed by verbs in this language. They are only found in their predicative use, never as attributive modifiers. However, Apurinã has a set of classifiers – referred to as ‘classificatory nouns’ in the source material – that can be used either with their literal meaning, viz. semantics restricted to body elements and plant/forest elements, or with a metaphorically extended meaning (Facundes 2000: 342). Classifiers attach to any noun as part of a productive compounding strategy (i.e., involving two lexemes and carrying one primary stress) to describe its referent in terms of a shape or consistency (Facundes 2000: 342–343, 195–196). This way classifiers play the role of attributive modification along with descriptive verbs, which cover a wider range of semantic meanings (Facundes 2000: 180, 342). For example, the form ke with its source meaning ‘stick of, rod of’ generates a description of the referent noun as ‘of a thin and generally flexible consistency’ (Facundes 2000: 342). Another example is the form mata with its literal meaning ‘skin’ (24a). Its target meaning is ‘of a flat surface’. In (24b), mata is used as part of a compound with the noun komeru ‘manioc’. It results in a complex meaning which also contains information on the physical shape of the resulting object, in this case ‘flat manioc bread’.

The occurrence of classifiers in a modifier function is also noted for Mojeño Trinitario by Rose (in prep.). Mojeño Trinitario has few morphologically underived property concepts and even those are rarely used as modifiers, with the predicative use being more common (Rose in prep.). Classifiers and property terms are complementary with regard to semantics (Rose in prep.), in the sense that the former can contribute meanings not expressed by the latter. Classifiers can express meanings mainly related to physical properties. Interestingly, as shown by Rose, a text contained 16 instances of classifiers that qualified a noun, and the same text contained 40 instances of property terms, with only 12 of them carrying out the function of modification. Just as in Apurinã, classifiers in Mojeño Trinitario do not encode a particular physical property, but they add to its interpretation.

Complementarity in terms of meaning that can be encoded through lexical items and classifiers is also found in Tariana. Tariana has no lexical means to describe a form like round, hollow, or curved. These meanings are expressed by classifiers (Aikhenvald 2003: 72). Example (25a) illustrates the attributive function of the classifier -da, whereas in (25b) we see that the same classifier derives a noun with a new meaning. Example (25c) illustrates the use of the classifier -kwema for a flat object adding the semantics of being round.

In Cubeo (Tukanoan) the situation is similar. In Tukanoan languages, property concepts are generally encoded by stative verbs (Morse and Maxwell 1999: 123). This is also the case in Cubeo, although there is a small and closed class of underived lexemes that can be analyzed as adjectives (Morse and Maxwell 1999: 124). However, Tukanoan languages in general possess large inventories of classifiers. Example (26) shows that a combination of lexical material, a classifier, and also diminutive morphology all contribute to an overall picture of a property attributed to a referent noun. The extended translation, which is adopted from the source material, gives an impression of possible meanings.

There is even a language in the sample which, among other concepts, can render color with classifier morphemes. In Mamaindê, property concepts are encoded by verbs that either need to be nominalized to occur attributively, or, most commonly, they occur as predicates taking regular verbal morphology (Eberhard 2009: 364). However, some classifiers express semantic notions typical of property concepts, and thus function as attributive noun modifiers. Mamaindê classifiers constitute a closed class of 24 elements. Relevant for our discussion is that some are semantically related to the shape of solid things (round, flat, sticklike, etc.), the physical state of non-solid entities (liquid, powder, etc.), and also to color (black, tawny) (Eberhard 2009: 330). The classifier which has a meaning ‘spotted’ (27a) is noted to relate to the verb root kalokalo ‘spotted’ (Eberhard 2009: 335, fn. 335). The fact that it is not followed by a nominalizer suggests that it is not a verb here (Eberhard 2009: 335). Eberhard further notes that the other classifiers expressing colors never occur with any verb morphology, and thus cannot be treated as verbs synchronically. Diachronically, however, the classifier -tunni ‘black’ (27b) is likely to originate in the verb root ton ‘to be black’ (David Eberhard p.c.). The capacity of grammaticalized elements – what the Mamaindê classifiers are – to express colors is typologically quite unusual.

In the present sample, we find at least eight languages where classifiers (i.e., lexico-grammatical elements) can approximate the function of property concepts.^[19] Except for Mamaindê, semantically these are limited to the domain of dimension and physical properties.

Finally, it is of interest that a noun for ‘child, offspring’ can be used to express small size. In a number of SA languages, a lexeme for ‘child, offspring’ has taken the role of a diminutive marker. Diminutives originating in nouns for ‘child, offspring’ are well-attested crosslinguistically (see Jurafsky 1996 and Kuteva et al. 2019: 88–90). Kuteva et al. (2019) do not mention any SA language for this grammaticalization path. In the present sample, however, it is found in Aguaruna (Overall 2017: 173), Awa Pit (Curnow 1997: 89), Hup (Epps 2008: 353), Mojeño Trinitario (Rose 2018: 176), and Yurakaré (van Gijn 2006: 171). We can observe a similar grammaticalization from a noun meaning ‘seed, kernel’: it can give rise to a classifier for small objects, and/or carry out the function of a diminutive marker. This is the case in the Arawakan languages Paresí (Brandão 2014: 184), Apurinã (Facundes 2000: 253) and Mojeño Trinitario (Rose 2018: 169),^[20] Barbacoan Tsafiki (Dickinson 2002: 75), and also in the isolate Movima, although in Movima the use is not productive (Haude 2006: 211). Also note that a semantic connection between ‘child, offspring’ and ‘seed, kernel’ can be posited for some languages. This polysemy can be found in some African languages (Dogon, Susu, Baka) discussed in Kuteva et al. (2019: 89). In Paresí, we find a lexeme itsani ‘son/daughter, offspring’ (Brandão 2014: 205, 330), ^[21] and there is a bound noun tse ‘seed.of’ (Brandão 2014: 179). Although it is not mentioned in the source material, it is not unlikely that the two lexemes are cognates. Example (28) from Apurinã illustrates the occurrences of the form -ku ‘seed’. (28a) shows the literal meaning of a bound noun -ku ‘seed / kernel’. (28b) demonstrates its grammaticalized nature: it derives a new noun with a new meaning. (28c) illustrates the form -ku functioning as modifier, attributing a characteristic of being small to a referent noun.

In a related language Paresí, a noun for ‘seed’ is formally different (viz. -tse), but it has comparable usages in the sense that (i) it derives a new meaning, similar to (28b) (Brandão 2014: 184–185), and (ii) it can express small size, in fact functioning as a diminutive marker (Brandão 2014: 187). The support for the latter is that the classifier -tse is the only classifier that can co-occur with another classifier within one word (Brandão 2014: 187). However, whether it is the classifier or the bound noun that has developed into a diminutive marker remains open.