2.1 The evolution of gender systems and its sociohistorical and environmental correlates

The fact that, in spite of their overwhelming stability, gender systems are likely to change, and sometimes even break down under the influence of language contact is well known in the literature. Trudgill (1999) and McWhorter (2001, 2007) argue that while gender systems are highly stable in situations of native-like language transmission, they can undergo erosion and eventually fade away when adult learners start playing a significant role in the language ecology of a speech community. In a recent crosslinguistic survey, Di Garbo (2020) investigates the evolution of morphosyntactic complexity in the domain of gender marking and its sociohistorical correlates. The study finds that the reduction, loss, and emergence of patterns of gender marking tend to be distributed in areas at the crossroads between gendered and genderless languages and/or families, and that asymmetries in population structure and/or prestige dynamics between populations play an important role in the direction of change. These results support the claim by Nichols (1992, 2003) that both presence and absence of gender are reinforced by geography: languages with gender are expected to neighbor each other, and languages where complete gender loss occurs also neighbor each other or are surrounded by languages without gender.

Even though solidly established, these generalizations have not received statistical validation through the analysis of large crosslinguistic data sets. On the contrary, attempts at investigating the relationship between gender systems and sociohistorical or environmental factors from a quantitative point of view have so far produced only negative evidence. Grammatical gender is not included in the list of morphological features that Lupyan & Dale (2010) demonstrate to be sensitive to the effect of social structures. Dahl (2019) reports that this omission is not due to lack of testing, but to the fact that, when tested for the effect of population data, the gender features included in the World Atlas of Language Structures (WALS) yield inconsistent results. He further replicates the results of Lupyan & Dale (2010) by correlating each of the gender features in WALS with the logarithm of the number of speakers for each language in the sample, which is also based on WALS. Similar findings are reported by Sinnemäki & Di Garbo (2018), who test whether population dynamics, which they model through the combination of the log number of native speakers and the proportion of second language speakers, have any effect on the distribution of number of genders based on a sample of about 300 languages. The study finds no significant effect of population size and population contact on the inventory size of gender distinctions. Finally, when testing the generalization that creole languages should exhibit radical simplification in several domains of morphosyntax as compared to non-creoles, Blasi et al. (2017) find no indication of such an adaptive response in the domain of gender marking on adnominal modifiers nor on personal pronouns. In the languages of their sample, presence/absence of gender marking in these morphosyntactic domains is a sheer reflection of ancestry, that is, of the type of system attested in the lexifier and/or the substratum language.

In this paper, we argue that the discrepancy between qualitative crosslinguistic research and large-scale quantitative analyses that investigated the impact of language contact on the evolution of gender systems is at least in part due to inadequate variable design in studies of the latter type. All studies cited above except for Blasi et al. (2017) use the gender features in WALS (“Number of gender distinctions,” “Systems of gender assignment,” “Sex-based vs. Non-sex-based gender”) as measures of complexity. The three features have proven to be highly relevant to classify the diversity of grammatical gender systems. However, we suspect that they are not a particularly good fit to answer questions concerning the sociolinguistic typology of gender systems, because they are not directly concerned with gender inflections, that is, with what has been argued to be most sensitive to the pressure of social and environmental factors (see also Cysouw [2005] on the prevalence of simplistic categorical measures in typology).

In order to study the sociolinguistic typology of gender systems, and to seek statistical validation of the generalizations mentioned at the beginning of this section, we thus suggest that a shift in focus, from the WALS features to the morphosyntactic encoding of grammatical gender, is needed. While this is what Blasi et al. (2017) partly do in their study by looking at patterns of gender marking on adnominal modifiers and personal pronouns, we contend that a more fine-grained research design is needed. This should be based on comprehensive accounts of domains (adnominal modifiers, predicative expressions, pronouns), targets (e.g., attributive and predicative adjectives, verbs, personal pronouns), and patterns of gender agreement (syntactic, semantic), and on their relation with gender assignment. Under syntactic agreement, gender marking is consistent with the gender a noun is lexically assigned to. Under semantic agreement, gender marking is based on the referential semantics of a noun, which may ultimately clash with its lexical gender assignment. For instance, in German, the noun Mädchen ‘young woman’ is lexically neuter, but semantically denotes a female human being. Agreement patterns with this noun alternate between neuter (syntactic) and feminine (semantic) in such a way that agreement on adnominal modifiers is often in the neuter, while agreement on personal pronouns is often in the feminine (Corbett 1991). It is a known fact that in those cases in which semantic agreement becomes generalized to all agreement targets and to a large class of nouns, this may also lead to a substantial restructuring of gender assignment rules, and of the overall makeup of a gender system (Corbett 1991: Chap. 8).

Here, we consider the expansion of semantic agreement as an instance of increased transparency (in line with Vihman et al. [2018]), and attempt to identify the linguistic and nonlinguistic scenarios under which these semantically transparent patterns of gender marking rise and spread. This approach tallies with a recent suggestion by Kempe & Brooks (2018), who argue that in order to investigate the impact of population dynamics (and second language learning in particular) on the evolution of morphological complexity, studying the emergence of transparency and compositionality is more revealing than, say, counting the sheer number of morphological distinctions that are made within a given morphosyntactic domain.

In addition, we argue that the issue of variable design also applies to the sociohistorical and environmental variables that are fed to the modeling of linguistic adaptation. To date, demographic variables related to population size and population structure have been at the fore of research in this field (see Lupyan & Dale 2010; Bentz & Winter 2013; Sinnemäki & Di Garbo 2018; among others). While we do not a priori object to the validity of these variables in the task of detecting nonlinguistic correlates of linguistic distributions, we advocate that exploratory studies of this kind should as much as possible aim to enhance the ecological validity of the sociolinguistic factors that are featured in the models. This requires an understanding of the language ecology and history of human populations that goes beyond census counts, as well as translating documented characteristics of the nonlinguistic environments under study into well-suited statistical variables. While such a fine-tuned approach to nonlinguistic variable design may be difficult to attain in worldwide typological investigations, it is certainly more feasible in genealogically and areally focused studies, like the one we present here.

2.2 Bantu gender systems: General characteristics

(Narrow) Bantu is a group of over 500 languages belonging to the Bantoid division of the Benue-Congo subfamily, a subgroup of the Atlantic-Congo family (Hammarström et al. 2018; Nurse & Philippson 2003). Bantu languages are spoken in sub-Saharan Africa; their northern border can be approximated by a line drawn from Nigeria in the west across the Central African Republic and the Democratic Republic of the Congo (DRC) to southern Somalia in the east. Except for the Khoe-Kwadi, Tuu, and Kx’a families and other languages previously described as “Khoisan” in the south, most of the language communities between this northern border and the Cape of Good Hope are Bantu (Nurse & Philippson 2003: 1), spoken alongside Cushitic and Nilotic languages. Traditionally, Bantu languages are classified into fifteen zones, which are labeled A through S and are referred to as Guthrie zones after Malcom Guthrie, who introduced them (Guthrie 1967–1971; Maho 2003, 2009).¹ This study focuses on northwestern Bantu languages (henceforth NWB), which we define as languages within Guthrie zones A, B, C, D, and H, and approximately covering southern Cameroon, southern Central African Republic, Equatorial Guinea, Gabon, the Republic of Congo, northern DRC, and northwestern Angola (see Grollemund et al. [2018] and references therein for an overview of multiple definitions of NWB). NWB is unanimously held to be the typologically most diverse area of the Bantu-speaking world, as well as one in which linguistic phenomena that may substantially diverge from those attested in the rest of the family are also encountered. It is in the northern borderlands of the northwestern Bantu area, for instance, that a variety of highly eroded systems of gender marking can be found (Maho 1999; Di Garbo & Verkerk 2022).

A fully grammaticalized gender system has been reconstructed for Proto-Bantu (Meeussen 1967; Maho 1999; Van de Velde 2019), and is in turn conceived of as inherited from Proto-Niger-Congo. Typically, Bantu gender systems consist of two sets of markers: (1) the overt gender markers, which encode gender distinctions on nouns, and (2) the agreement markers, which encode gender distinctions on adnominal modifiers, pronouns, verbs, and so on. Both nominal class forms and agreement markers also encode singular and plural number distinctions. In most descriptions of Bantu gender systems, one gender consists of combinations (or pairings) of singular and plural classes.² Traditionally, reference is made to cognate noun classes via a numbering system where odd numbers stand for singular classes and even numbers for plural classes. For instance, class 1 is the singular human class reconstructed as *mʊ̀- in Proto-Bantu and class 2 is the plural human class reconstructed as *bà- in Proto-Bantu. Pairings of singular and plural classes are labeled using Roman numerals and referred to as genders (e.g., pairing 1/2 is known as gender I). In this paper, we prefer using the notation ‘gender 1/2’ because we want to make reference to the internal structuring of each pairing. The total number of classes and genders varies quite widely across the languages of the family. While up to twenty-four distinct classes are reconstructed for Proto-Bantu (Katamba 2003; Van de Velde 2019), none of the contemporary languages retains all twenty-four classes, and some languages have lost them all. Gender marking in Bantu is almost always prefixal (Katamba 2003: 111).³ Examples of genders 1/2 and 7 from Akoose, spoken in Cameroon, are given in (1).⁴

(1) Gender marking in Akoose (Hedinger 2008: 16–17)

a. Agreement with a Class 1 noun

aw-í	mw-aád	a-’só	aw-é	a-nsɔ́g
CL1-his	CL1-wife	CL1-first	CL1-which	CL1-PST.fat

‘his first wife, who was fat’

b. Agreement with a Class 2 noun

áb-é	b-ǎn	bé-kal-e	me-tóm
CL2-those	CL2-children	CL2-tell-IPFV	CL6-lies

‘those children tell lies’

c. Agreement (across clauses) with a Class 7 noun

A-tédé	e-hid	é	nyam,	á-kɛɛné	ch-ə́	áhîn	tê
CL1-take.PFV	CL7-bone	CL7.AM	CL7.animal	CL1-take.PFV	CL7-PRO	LOC.CL5.bush	in

‘He took the animal bone and took it to the forest.’

Bantu gender systems are non-sex-based and characterized by a combination of semantic and formal, as well as opaque, gender assignment (Katamba 2003). A typical Bantu gender system is built upon the following semantic notions (Denny & Creider 1976; Katamba 2003; Contini-Morava 2000):

• Animacy: nouns that denote humans are typically assigned to gender 1/2, animal nouns are associated with gender 9/10, and plant and tree names with gender 3/4.

• Size: several classes (typically 7, 8, 11, 12, 13, 20, and 21) are associated with the encoding of diminutive and augmentative meanings.

• Infinitive marking is associated with class 15.

• Locative meanings are associated with classes 16, 17, and 18.

Not all Bantu languages closely adhere to these generalizations, but, for the purposes of this paper, we assume that the typical Bantu gender system is non-sex-based and only partly conditioned by the overt expression of animacy distinctions.

2.3 Bantu gender systems: Animacy-based semantic agreement, with a focus on NWB

One of the major sources of variation in Bantu gender systems is the spreading of semantic agreement (Maho 1999; Van de Velde 2019; Wald 1975). The type of semantic agreement that is most widely attested across Bantu is known in the Bantu literature as “animate concord,” which we call here animacy-based semantic agreement or simply animacy-based agreement.⁵ As mentioned above, in Bantu languages, human nouns are typically assigned to gender 1/2. Under animacy-based agreement, the agreement patterns associated with gender 1/2 are also applied to animate nouns from other genders, for instance the many animate nouns from gender 9/10. This pattern of semantic agreement is most often optional and restricted to some agreement targets only. It typically occurs on verbs, in the form of subject agreement, as well as on a variety of pronouns. Conversely, adnominal modifiers are more likely to maintain syntactic agreement, that is, to agree with the lexical gender of a noun. Semantic and syntactic agreement may coexist as relatively stable variants in one and the same language for long periods of time (see example (3) below for an illustration of alternation between syntactic and animacy-based agreement).⁶

In those languages of the family that exhibit highly eroded gender systems, basic animacy contrasts (such as animate vs. inanimate) are often the only surviving type of gender distinction (see Table 1 below for an illustration from the Bantu language Bila). This raises the question of whether the erosion of grammatical gender and the expansion of animacy-based agreement may be in some way diachronically connected to each other. While Maho (1999) considers this as a plausible scenario, the relationship between semantic agreement and the erosion of gender marking has only recently been more closely investigated by Di Garbo & Verkerk (2022) in a sample of 179 NWB languages. The findings of this study constitute the point of departure for the questions we address in the present paper, and will be summarized in the remaining of this section. Di Garbo & Verkerk (2022) looked at the distribution of types of gender agreement (syntactic vs. animacy-based) on fifteen different agreement targets: attributive adjectives, copula-like constructions, demonstrative modifiers, demonstrative pronouns, genitives/connectives, independent third person pronouns, numerals, quantifiers, possessive pronouns, predicative adjectives, question words, reflexive pronouns, relative constructions, verbs, and other targets. This list of agreement targets captures the general morphosyntactic characteristics of (northwestern) Bantu gender systems and is also reflective of the general typological literature on (gender) agreement systems, and, in particular, on the relation between syntactic and semantic agreement. Data collection was coordinated through a questionnaire that is given in Appendix A. In this paper, we use the data collected by Di Garbo & Verkerk (2022) to test hypotheses about sociogeographic correlates of restructuring in the gender agreement systems of NWB languages.

In order to best capture the diversity of the gender systems attested in the languages of the sample and to avoid committing a priori to a specific model of diachronic change, Di Garbo & Verkerk (2022) identify different profiles of gender systems depending on whether gender agreement is syntactic (that is, based on the lexical gender of the noun), animacy-based, or both, or, alternatively, whether gender marking is missing altogether. Importantly, in the terminology introduced in that paper, the label animacy-based agreement is used for any agreement pattern that singles out the degree of animacy of the noun referent, independently of how this matches specific cutoff points along the Animacy Hierarchy (Smith-Stark 1974). Thus, we apply this label both to agreement patterns that feature an animate vs. inanimate contrast, as is the case of subject agreement in Lika (Augustin 2010: 18–19), and to agreement patterns that are based on a human vs. nonhuman contrast, as is the case for the third person pronouns in Nzadi (Crane et al. 2011: 75).

An example of a language with solely syntactic agreement is Bakole (Glottocode: bako1250), spoken in Cameroon. Syntactic agreement on the verb is shown in (2) for the animate nouns for ‘child’ and ‘sheep’.

(2) Syntactic gender agreement in Bakole (Asobo 1989: 89)

a. Subject-verb agreement with an animate Class 1 noun

mw-ánà	à	mádà
CL1-child	CL1	ate

‘The child ate.’

b. Subject-verb agreement with an animate Class 3 noun

mù-ròŋgi	mú	mádà
CL3-sheep	CL3	ate

‘The sheep ate.’

Bila (bila1255), spoken in the DRC, is a language which displays solely animacy-based agreement. The only targets of gender agreement are adjectives, some numerals, and demonstratives, whose inflections are illustrated in Table 1. Inanimate gender markers are always invariant for number, while the markers of the animate genders distinguish between singular and plural.

In languages characterized by the co-existence of syntactic and animacy-based agreement, verbs are the word class where animacy-based agreement most commonly surfaces, whereas other agreement targets may agree syntactically. Such a combination of syntactic and animacy-based agreement is found in Bomboma (bomb1262), spoken in the DRC. Bomboma has syntactic agreement as described for Bakole on a range of different agreement targets. In addition, it has animacy-based agreement on the genitive markers that are used to mark possession both adnominally and pronominally as well as for verbal subject agreement. Example (3) shows that singular nouns denoting animals may take Class 1 subject agreement prefixes, independently of their lexical gender, while plural nouns for animals may not. The same applies to human nouns from Gender 7/10 (Toronzoni 2004: 70).

(3) Syntactic and animacy-based agreement in Bomboma (Toronzoni 2004: 65)

a. Animacy-based subject-verb agreement with a singular animate Class 9/10 noun

N-va	a-wei
CL9-dog	CL1-died

‘The dog is dead.’

b. Syntactic subject-verb agreement with a plural animate Class 9/10 noun

N-va	i-yato	i-wei
CL10-dog	CL10-three	CL10-died

‘Three dogs are dead.’

Di Garbo & Verkerk (2022) discuss three languages where animacy-based agreement also extends to the domain of inanimate nouns, leading to what they call generalized animacy-based semantic agreement. These languages are Lika (lika1243), Mpiemo (mpie1238), and the Bibaka variety of Ukhwejo (ukhw1241). In Lika, generalized animacy-based agreement is obligatory on subject agreement on verbs, which only differentiates between animate and inanimate subjects, while gender agreement on adnominal modifiers is syntactic (Augustin 2010). In Mpiemo and Bibaka Ukhwejo, generalized animacy-based agreement optionally runs through the entire agreement system. Animate nouns may trigger gender 1/2 agreement whereas inanimate nouns trigger agreement in gender 7/8 (Mpiemo, Thornell 2010) or 7 (Bibaka Ukhwejo, Thornell 2012). This can happen with all agreement targets, irrespectively of the lexical gender of a noun. Syntactic agreement is still in use, particularly among the older generations of speakers. Di Garbo & Verkerk (2022) argue that the agreement systems attested in these three languages offer crucial empirical evidence for how the erosion of gender marking may possibly be connected with the expansion of animacy-based agreement. When, as happens in Lika, Mpiemo, and Bibaka Ukhwejo, semantic agreement is generalized to inanimate nouns, this leads to a situation in which at least some agreement targets (such as the verb in Lika) or, optionally, all of them (as in Mpiemo and Bibaka Ukhwejo) only encode two gender values, the animate and inanimate. If this polarization of agreement around animacy contrasts is generalized to all existing agreement targets, and becomes obligatory, a gender systems of the type found in Bila, with only two genders, animate and inanimate, may emerge.

In Di Garbo & Verkerk’s (2022) sample, complete gender loss is attested in Homa (homa1239), Komo (komo1260), Kituba (two varieties, one spoken in the Kwilu-Kwango regions of the DRC [kitu1246] and one in the Lower Congo region of the Republic of Congo [kitu1245]), Polri (pomo1271), and Yansi (yans1239). Homa is a now extinct language and was once spoken in South Sudan. We know next to nothing about it, except that it may retain some fossilized forms of prefixal animacy-based marking on some attributive adjectives (Santandrea 1963). Komo, spoken in the DRC, still has fossilized prefixes on nouns, but no gender agreement. Thomas (1994) notes the different forms these fossilized nominal prefixes have, and hypothesizes that Komo may have borrowed extensively from other Bantu languages with more conservative systems. The two varieties of Kituba are creoles whose system of nominal prefixation is mostly intact, but which do not have any gender agreement left (Mfoutou 2009; Mufwene 1997; Buchanan 1996–1997; Stucky 1978). In Polri, spoken in Cameroon and Congo, and Yansi, spoken in the Kwilu and Mai-Ndombe provinces of the DRC, relics of the now lost gender system are left on nouns, whereas patterns of agreement only express number distinctions (Wega 2012; Mufwene 2006).

View Full Size

Figure 1

Plot of the 179 northwestern Bantu languages in Di Garbo & Verkerk (2022), based on how many of the fifteen agreement targets in each language show animacy-based agreement and how many show syntactic agreement. Points have been jittered so they do not overlap

Citation: Language Dynamics and Change 12, 2 (2022) ; 10.1163/22105832-bja10017

• Download Figure
• Download figure as PowerPoint slide

Although they exhibit substantial variation in their respective patterns of gender marking, the languages of Di Garbo & Verkerk’s (2022) sample cluster into four main typological profiles, as shown in Fig. 1: (1) languages with only syntactic and no animacy-based agreement (in black, 121 languages); (2) languages with both syntactic and animacy-based agreement (in blue, 40 languages); (3) languages with only animacy-based agreement (in orange, 11 languages); and (4) languages devoid of gender (in green, 6 languages).⁷ An overview of the distribution of syntactic and animacy-based agreement in all the languages of the sample is included as Supplementary Information 1, “Distribution of syntactic and animacy-based agreement per language and across target types.”

View Full Size

Figure 2

Polygon map representing the distribution of the four-way typology among northwestern Bantu languages described in the text in the context of neighboring languages and their genealogical affiliations. Language polygons are taken from the World Language Mapping System (version 17, Global Mapping International, 2015, http://worldgeodatasets.com). White areas represent uninhabited land or bodies of water; towards the north, other Atlantic-Congo, Central Sudanic, and Ubangi languages are not drawn.

Citation: Language Dynamics and Change 12, 2 (2022) ; 10.1163/22105832-bja10017

• Download Figure
• Download figure as PowerPoint slide

Di Garbo & Verkerk (2022) make several points regarding diachronic change and the geographical distribution of the four types identified above, which we cannot examine in detail here. Figure 2 illustrates some of these tendencies, by representing the languages and types of gender systems under study in the context of surrounding languages. For instance, the figure shows that there are clear clusters of languages with only syntactic agreement (in purple) and with both syntactic and animacy-based agreement (in dark blue). Languages with only animacy-based gender (in orange) or no gender (in green) are scattered, but primarily found along the northern edge of the reach of the Bantu family, as also noticed by Maho (1999). The most northwestern corner of the area investigated contains zone A and B languages, which are closest to the root of the Bantu family tree (Grollemund et al. 2015). Since these are almost all languages with only syntactic agreement (in purple in Fig. 2), Di Garbo & Verkerk (2022) infer that Proto-Bantu probably also had only syntactic agreement, an assumption we also make here.⁸

To summarize, Bantu gender systems are most often described as complex and conservative nominal classification systems, with several gender distinctions, pervasive patterns of syntactic agreement, and pockets of animacy-based agreement mostly localized in the eastern coastal regions (Contini-Morava 2008). However, recent studies suggest that many NWB languages also show pervasive animacy-based agreement and at least seventeen of them display highly eroded systems of gender marking, which only distinguish between animate and inanimate nouns, or have no gender at all. Building upon the data set and findings of Di Garbo & Verkerk (2022), in this paper we test the hypothesis that these pervasive patterns of restructuring occurred in languages characterized by a history of intense and prolonged contact with non-Bantu populations and/or in creolized Bantu varieties.

3.1 Variable design: Identifying sociohistorical and environmental correlates of gender restructuring

Based on existing literature (see Section 2.1), we can posit two potential factors that might have impacted patterns of gender marking in the NWB context: (1) contact with non-Bantu languages without gender or with non-Bantu-like gender systems (see Nichols [1992, 2003] as well as Di Garbo [2020] on the influence of neighboring languages on the retention or loss of gender systems) and (2) massive non-native language learning resulting in pidginization and creolization (McWhorter 2001, Trudgill 1999). In this section, we summarize how these factors relate to the seventeen most-eroded gender systems of our 179 NWB language sample. As mentioned in Section 2.3, of these seventeen languages, eleven have solely animacy-based gender while six lack gender altogether. Our aim in this section is to use insights from sociolinguistic and ethnographic resources on the languages and population histories of these communities as a basis for variable design.

The seventeen languages are given here grouped by their closest genealogical relations (according to Glottolog; Hammarström et al. 2018):⁹

• Ababuan (Central-Western Bantu): Bera (bera1259), Amba (amba1263), Bila (bila1255), and Komo (komo1260) (Komoic); Homa (homa1239), Bodo (bodo1272), and Kari (kari1306) (Ngbele-Ngenda); and Beeke (beek1238) (Bali-Beeke)

• Likouala-Sangha (Central-Western Bantu): Pande (pand1264) and Mbati (mbat1248)

• West-Coastal Bantu (Central-Western Bantu): Yansi (yans1239) and Nzadi (nzad1234)

• Makaa-Kako (Bantu A–B10–B20–B30): Polri (pomo1271) and Kako (kako1242)

• Two creoles: Lingala (ling1263) and Kituba (two varieties, kitu1245 and kitu1246)

For each of these languages, we consulted grammars and ethnographies, as well as sociolinguistic overviews and survey reports, in order to gather information on the (historical) sociolinguistics of the speaker population.¹⁰ We also gathered data, when there was any to be found, on other remarkable grammatical features of the languages as well as on sources of lexical borrowing. A summary of this literature review is included in Appendix C.

The Ababuan group (McMaster 1988) contains, as far as we found, only languages with solely animacy-based gender or no gender. In the other groups, we have pairs of closely related languages (Pande and Mbati, Polri and Kako) or singleton languages (Yansi, Nzadi) that have closely related sister languages that do display syntactic gender agreement (and sometimes animacy-based agreement). The majority of languages with solely animacy-based gender or no gender in our sample (twelve out of seventeen) are located in the northern Bantu borderlands. These are the Ababuan, Likouala-Sangha, and Makaa-Kako groups listed above. The location of these twelve languages is displayed in Fig. 3, together with the locations of the most relevant Ubangi and Central Sudanic languages and of several Pygmy populations taken from Bahuchet (2012).

View Full Size

Figure 3

Location of the twelve languages with only animacy-based gender marking or no gender at all in relation to surrounding non-Bantu languages, and the Central African rainforest. Language polygons are taken from the World Language Mapping System (version 17, Global Mapping International, 2015, http://worldgeodatasets.com). White areas represent uninhabited land or bodies of water; towards the north, other Central Sudanic and Ubangi languages are not drawn. Locations of Pygmy populations are taken from Bahuchet (2012). The rainforest polygon was taken from Grollemund et al. (2015). The darker shade of some language polygons can be ignored.

Citation: Language Dynamics and Change 12, 2 (2022) ; 10.1163/22105832-bja10017

• Download Figure
• Download figure as PowerPoint slide

Data on the social history of these twelve northern borderlands languages, and those in the east of the sampled area, are not always clear nor extensive. However, where information is available, it points towards contact with various non-Bantu populations both in the ethnographic sense and also in the linguistic sense (loanwords, unusual features that have likely been borrowed). These contacts feature Ubangi and Central Sudanic speakers as well as Pygmy populations.

Ubangi speech communities came into contact with Bantu speakers when the former migrated from the north as a result of Nilo-Saharan migrations even further north (Bostoen & Donzo 2013: 440, and references therein). Historical records point to a history of prolonged contact, characterized by cultural and linguistic assimilation and influence in both directions, also involving bidirectional language shift (Bostoen & Donzo 2013: 440; Vansina 1990: 66). The Kari (Bantu) speakers, for example, are described as being assimilated into the Zande-speaking community (Ubangi) which suggests long-standing bilingualism in Zande. The prestige of Zande as the majority language of the area may have impacted Kari in a substantial way, perhaps even determining the conditions for Kari to fall out of use altogether.¹¹ Ubangi languages either lack gender or, if they do have gender systems, these may differ from those typical of Bantu languages. For instance, in several Ubangi languages, gender marking tends to be suffixal rather than prefixal (Boyd 1989), and in Zande, gender marking is only pronominal and features a combination of sex-based and non-sex-based distinctions (Corbett 1991). A few Ubangi languages exhibit an interesting combination of syntactic and animacy-based agreement, which is partly reminiscent of what we also encounter in some NWB languages (Fedden & Corbett 2017: 32–37; Fiedler et al. 2021 on Mba).

To the east of these northern borderlands, western Bantu farmers came in contact with Central Sudanic speakers, who were traditionally herders or cereal farmers (Vansina 1990). Vansina (1990: 66–67) reports on the dramatic impact that the expansion of speakers of the Mamwu-Balese subgroup had on western Bantu-speaking communities such as the Bira and Komo, which are both part of our sample and exhibit eroded gender systems. These newly arriving Central Sudanic populations imposed their social organization on the area, but, in the end, “fused with the original farmers” (Vansina 1990: 66). Central Sudanic languages typically lack grammatical gender, and compared to other Nilo-Saharan groupings, they tend to display reduced systems of nominal number marking (Dimmendaal 2000). While, in these languages, nominal morphology is generally suffixal, instances of prefixation also occur, which, in some cases, can be explained as borrowings from neighboring Bantu languages (Dimmendaal 2000: 237; Blench 2018). Finally, the languages originally spoken by the Pygmy populations before they shifted to Bantu, Ubangi, and/or Central Sudanic languages are unfortunately unknown (Bostoen & Gunnink forthcoming 2022; Klieman 2003).

Güldemann (2018: 506–507) and Bostoen & Gunnink (forthcoming 2022) suggest that some of the diverging structural features attested in the northern Bantu borderlands arose as the result of substrate effects from the native languages of the Pygmy groups. Güldemann (2018: 506–507) describes how in the central part of the Macro-Sudan belt and the Bantu spread zone, lower altitude and denser forestation are more prone to pattern with restructured gender systems than higher altitude and savanna-like environments. This typological distribution could be connected with the history of early Bantu settlements in this area. The first populations arriving in these areas were not familiar with lowlands rainforest environments. Adapting to this unfamiliar environment required investment in contact with the local forest-based communities (see also Klieman 2003). This scenario must have favored intimate contact between communities, which ultimately led the autochthonous communities to shift to the languages of the newcomers, and created the right context for shift-induced changes to occur. Restructuring of gender marking could be counted as one of these substrate effects. The retention of traditional gender systems in other rainforest Bantu languages does not contradict this hypothesis as speakers of the more conservative languages probably arrived in these areas much later, when familiarity with the environment had already been acquired, and the contact dynamics between rainforest populations were already more asymmetrical and less tight. Considering the genealogical groups listed above, it seems that Ababuan, where we find only languages with heavily restructured gender systems, may fit the Pygmy substrate scenario, although we should note that many of the sources mentioned in Appendix C also remark on contact with Central Sudanic. For the other borderland languages (Pande, Mbati, Polri, and Kako), a more localized scenario involving contact with Ubangi speakers seems more fitting.

Based on this overview, we can distinguish at least two layers of contact-induced cultural adaptations. The first layer concerns interactions between Bantu farmers, newcomers in the area, and local forest-specialists (Pygmies). The second layer concerns interactions between NWB speakers and Ubangi and Central Sudanic populations, who also expanded into the Central African rainforest. As the current discussion and the data in Appendix C show, both layers of contact had profound effects on the linguistic profile of these areas, either in the form of rapid and abrupt language shifts or in the form of prolonged bilingualism and reciprocal assimilation. We can speculate that rapid and abrupt language shift may explain the radically restructured gender systems or the complete loss of gender which we find attested in these twelve northern Bantu borderlands languages.¹² Conversely, prolonged bilingualism could account for the coexistence of syntactic and animacy-based agreement, which is observed both among NWB and Ubangi languages of the area.

As mentioned above, gender restructuring may also result from pidginization and creolization. In our sample, this is clearly associated with Kituba and (Kinshasa) Lingala. As described in Appendix C, the process of pidginization at the origin of these languages in both cases precedes the arrival of the European colonizers and stems from trade practices between local populations around the Congo. What is remarkable in both cases is that while the restructuring of gender marking around animacy-based distinctions is pervasive in the agreement domain, to the effect that no agreement is left in Kituba and only animacy-based agreement is found in Kinshasa Lingala, class prefixation on nouns tends to be maintained. Maho (1999: 140) argues that this type of development is typical for Bantu languages of wider communication.

Finally, it is worth noting that the relationship between contact and gender restructuring is not one-to-one. While northern Bantu borderlands languages that have both syntactic and animacy-based agreement (e.g., Pagibete, spoken on the border of Northern Ngbandi) or only syntactic agreement (e.g., Mpongmpong, spoken slightly south of Polri and Kako, and surrounded by Baka and Ubangi languages (Bouquiaux & Thomas 1994)) can perhaps be explained as later arrivals to the rainforest, there are also instances of restructuring where no clear evidence for language contact exists. This is the case for Yansi (no gender)¹³ and Nzadi (only animacy-based gender, with a human/nonhuman contrast). These two areally close languages are spoken in the southwestern-most part of Fig. 3, on the southern border of the rainforest, far away from attested non-Bantu populations. To the best of our knowledge, there are no clear-cut sociolinguistic clues as to why these languages exhibit heavily restructured profiles from a general Bantu perspective (both Yansi and Nzadi have other peculiar characteristics beyond the gender domain; see Pacchiarotti & Bostoen 2021). While we cannot exclude the possibility of internal change, Bostoen & Gunnink (forthcoming 2022) mention an unknown substrate as a potential cause.

While the overview presented in Appendix C and summarized in the current section is limited in scope, the results of this survey point towards contact with various non-Bantu languages, at different points in time, and to the impact of large numbers of second language learners as relevant factors for the emergence of animacy-based gender and the complete loss of gender. We translate these observations into a set of sociolinguistic and geographical variables, which we aim to model alongside our typological variables, and which we operationalize as follows:

• Sociogeographic variables: Ancestor in rainforest; current rainforest overlap; border with Ubangi/Central Sudanic. We introduce the first two variables here in order to test whether past or current presence in the rainforest predicts a higher incidence of restructuring in gender marking. They are meant to test the hypothesis that restructuring in gender marking is a substrate effect, directly connected with contact between Bantu populations and autochthonous rainforest-specialist populations. The third variable is meant to complement the rainforest variables, and to directly test whether proximity with Ubangi- and/or Central Sudanic-speaking populations predicts a higher incidence of restructuring in the languages of our sample.

• Geographic variables: Latitude; longitude.¹⁴ Since radically restructured languages tend to be located in the northern Bantu borderlands, the geographical position of languages as captured by these two variables may be an important factor in explaining their distribution. Beyond these languages, the presence of animacy-based agreement also has a distinct geographic signature (see Fig. 2).

• Demographic variable: Number of L1 speakers. As per the discussion of Lingala and Kituba above, this variable is relevant to capture processes of language change that might take place in languages of wider communication. By including this variable, we also expand on previous research which looked at whether the sheer number of gender distinctions correlates with community size (see, e.g., Sinnemäki & Di Garbo 2018).

We believe that, taken together, these variables can provide a representative picture of the NWB language ecology as depicted by the sources that we consulted (see Section 5.1 for discussion). While these variables are neither exact nor exhaustive, they appear to shed light on aspects of the interaction between gender systems and the local sociogeographic environment, which we want to test further with the support of quantitative methodologies and based on the full data set of 179 NWB languages. Choosing these variables does not exclude the possibility that other processes, which these variables do not capture, may influence the distribution of types of gender system that we observe in our sample. Given this proviso, we factor in these variables in our quantitative analyses, which we introduce in the next section.

3.2 Variable implementation: Quantifying sociohistorical and environmental correlates of gender restructuring

We investigate the hypothesis that a relationship exists between the type of gender system a NWB language has and the sociohistorical and environmental factors identified in Section 3.1. To assess this relationship, we use regression analyses, where the response variable (the type of gender system) is assessed in terms of the sociolinguistic and sociogeographic variables identified in Section 3.1. Our response variable—that is, the types of gender systems that we observe in our sample—can be operationalized in three different ways in our models:¹⁵

1. As a four-way typology, which we presented in Section 2.3. Languages are classified as having either (1) only syntactic agreement; (2) both syntactic and animacy-based agreement; (3) only animacy-based agreement; or (4) no gender. This translates into a four-way multinomial measure, or three binary measures if the first group is taken as the reference group.

2. As a simplified binary typology. This is obtained by dropping the distinction between types (2), (3), and (4) as given above, that is, by devising a binary measure which contrasts languages with only syntactic agreement with languages that have any other type of gender system (i.e., languages that have various amounts of animacy-based agreement or no gender)

3. As the counts of the number of targets that receive syntactic agreement and of the number of targets that receive animacy-based agreement, that is, as two count measures.

For regression modeling, we exclusively use two of these response measures: the simplified typology (binary measure; only syntactic agreement vs. any other type of gender system) and the counts of how many targets receive syntactic agreement and how many receive animacy-based agreement. The four-way typology given in Section 2.3 classifies the languages of the data set into four groups, some of which (those containing languages with only animacy-based gender or no gender) are very small. We attempted to fit such a response variable at earlier stages of this project, and faced issues related to statistical power because of the skewed distribution of languages across groups (see Supplementary Information 2, Section 7, for a discussion of statistical power). From a typological point of view, choosing where to draw the line in the range from languages with a small amount of animacy-based marking through to those with no gender agreement at all is somewhat artificial, as differences between these types can sometimes be minimal (see Section 2.3). For instance, languages with no syntactic agreement and where animacy-based marking occurs on a single target only minimally differ from languages with no gender agreement. Similarly, languages with both syntactic and animacy-based agreement, where the latter is potentially available for all agreement domains, are in effect not so different from languages with only animacy-based gender. In view of this, we think that if any binary distinction is to be made in our typology, it should be between languages that only have syntactic agreement and all other languages (i.e., languages with any amount of animacy-based agreement or no gender marking at all). We do use the four-way typology for plotting in Sections 2.3 and 4 as we find it more fine-grained and insightful.¹⁶

The data on the sociogeographic variables were collected from the following sources:

• Ancestor in rainforest and current location in rainforest. The data on whether the ancestor language was located in the rainforest are based on the reconstructed route of the Bantu speakers by Grollemund et al. (2015: Fig. 3).¹⁷ Following Grollemund et al.’s analysis, most of the NWB languages in the current sample have an ancestor in the rainforest; only languages in Guthrie zones A30, B60–B70, D20–D50, and H10–H20 do not. The data on current location in the rainforest are based on the ratio of the overlap between each language area and the Central Africa rainforest (ranging from 0 if the language is spoken entirely outside the rainforest to 1 if spoken entirely inside the rainforest). The polygons used for these calculations were taken from Grollemund et al. (2015) and the World Language Mapping System.¹⁸

• Latitude and longitude. These data are almost entirely taken from Glottolog (Hammarström et al. 2018). When data were missing from Glottolog, information was taken from primary sources.

• Border with Ubangi and/or Central Sudanic. The language polygons from the World Language Mapping System and/or identified from the literature were used to identify speaker communities that are in close contact (i.e., touching borders) with Ubangi- and/or Central Sudanic-speaking communities. Data on where these communities are located were taken from Ethnologue (Lewis et al. 2016).

• Number of L1 speakers. These data are taken from Ethnologue (Lewis et al. 2016).¹⁹

Given that we excluded Shiwa (see footnote 7) and two other languages due to missing data on population size (see footnote 19), the final data set contains 176 languages. The relationship between the type of gender system and demographic and geographic factors was assessed using several different models. As we deal with a mixture of data types (binary, continuous, counts), we primarily constructed generalized linear mixed effects models (GLMM s) using the package brms (Bürkner 2017) in R (R Core Team 2017). This package allows the user to fit Bayesian multilevel models in R using the probabilistic programming language Stan (Carpenter et al. 2017). The case for using mixed models in language typology has been made by Cysouw (2010) and Coupe (2018) and is gaining ground (Sinnemäki & Di Garbo 2018; Schmidtke-Bode & Levshina 2018; Schmidtke-Bode 2019; among others). Our immediate reason for choosing Bayesian GLMM s over frequentist statistics was that Bayesian packages in R (brms [Bürkner 2017] and MCMCglmm [Hadfield 2010]) allow for a wider range of models, including a mixture of categorical and continuous fixed effects, as well as random effects that control for genealogical relations by using the full structure of a phylogenetic tree (set). Additional reasons for opting for a Bayesian framework are discussed by McElreath (2020). Other recent typological studies that opt for a Bayesian approach are Erben Johansson et al. (2020), Urban & Moran (2021), and Guzmán Naranjo & Becker (2021).

Since our sample exclusively consists of closely related languages, we employ statistical methods that can account for historical relationships between these languages. We do this by using (a) grouping variables, that is, classifications in terms of big(gish) genealogical language groups; and (b) phylogenetic trees. The grouping variables capture the relatedness of the languages of our sample at 3000–5000 years in the past (Grollemund et al. 2015). Phylogenetic trees capture the entire pedigree of the sampled languages, as they consist of hierarchically nested groups of closely related languages.

The grouping variable that we used for the analyses reported in the main text is the classification of Narrow Bantu by Glottolog (Hammarström et al. 2018).²⁰ It contains six groups: Ababuan (22 languages in our sample), Bantu A–B10–B20–B30 (70), Central-Western Bantu (130), East Bantu (12), Lebonya (6), and Mbam-Bube (15).²¹ Our sample does not include all Central-Western and all East Bantu languages, as many of these are spoken in different Guthrie zones. We included only languages from the NWB area, from Guthrie zones A, B, C, D, and H (see Section 2.3). For more information regarding the use of Glottolog as the main grouping variable and for the phylogenetic trees used for additional analyses, see Supplementary Information 2, Section 3.

5.1 Discussion of methods and findings

This paper explored the hypothesis that gender systems tend to be eroded or even break down under the pressure of language contact (Trudgill 1999; Dahl 2019). NWB languages form an excellent test bed for this hypothesis. This area features languages with traditional gender systems, where only syntactic agreement occurs, and languages with different types of restructured or eroded gender systems, which display various forms of animacy-based agreement, or have no gender at all. We identified three potential factors that may have impacted NWB gender systems: substrate influence from languages that were originally spoken by native Pygmy peoples, who later switched to Bantu (or Ubangi or Central Sudanic) languages as these spread (Klieman 2003; Güldemann 2018; Bostoen & Gunnink forthcoming 2022); contact with Ubangi or Central Sudanic languages without (Bantu-like) gender systems (Vansina 1990; Bouquiaux & Thomas 1994; Wega 2012); and creolization of trade languages which speakers with different language backgrounds came to learn (Fehderau 1966; Samarin 1991; Bokamba 2009; Meeuwis 2013). The first scenario may have applied to the Ababuan languages; the second scenario probably accounts for Polri, Kako, Pande, and Mbati but also for Bodo, Homa, and Kari, and for all northern borderlands Bantu languages to some extent; and the third scenario applies to Kituba and (Kinshasa) Lingala, as well as, presumably, to other Bantu lingua francas that we have not sampled here.

In Section 4, we modeled the three above-mentioned language contact scenarios using six predictors: number of L1 speakers, latitude, longitude, having an ancestor in the Central African rainforest, current location with respect to the Central African rainforest, and proximity to the Ubangi and Central Sudanic language families. We found that five of these six variables are relevant predictors for one or more of the three main models presented (in Figs. 6, 7, and 8), with two of them, the number of L1 speakers and sharing a border with Ubangi or Central Sudanic, as the most robust. The number of L1 speakers and sharing a border with Ubangi or Central Sudanic were found to be relevant predictors across a variety of methods to control for genealogical relatedness (see Supplementary Information 2, Section 3). None of the reported effects “survives” a more rigorous control for genealogical relatedness, the use of random slopes (Supplementary Information 2, Section 7).²⁵

Since the paper deals with exploratory rather than confirmatory analyses, it is useful to consider what insights these findings may offer to more confirmatory studies of the long-debated idea that gender systems become eroded or break down under language contact. We do this by further commenting on the nature of the typological measurements of gender systems that we use in our models, the effect of demographic variables, and the role that heavily restructured languages play in the shaping of our results.

First, the three response variables (the binary typology and the target counts) are inter-correlated, and we should consider them as a relevant set of measures, rather than trying to pick the “best” one. Contrasting languages that have only syntactic agreement with languages that exhibit any kind of animacy-based restructuring in the binary typology, we find that the latter are more likely to share a border with Ubangi or Central Sudanic, as well as to have lower current overlap with the rainforest. This matches what we find when using target counts as the response variables. Languages with animacy-based agreement or no gender at all tend to have fewer or no targets for syntactic agreement and more targets for animacy-based agreement. The number of targets that receive animacy-based agreement can be explained by a set of geographical predictors. This suggests that animacy-based agreement shows the greatest sensitivity to contact effects, either as a result of language shift or convergence with neighboring languages. Our study clearly illustrates that using different types of typological measures of complex morphosyntactic phenomena, such as gender marking, has important consequences for the output of sociolinguistic typological studies.

As far as the effect of demographic variables is concerned, languages devoid of syntactic agreement have higher numbers of L1 speakers than most languages that mark syntactic agreement on any number of targets. However, these “big” languages are clearly outliers in the data set, as most languages of the sample have moderate to small L1 populations, and, as discussed above, the effect of population size disappears (or even changes direction) when these big outliers are excluded. One explanation could be that while the number of L1 speakers (discussed in Section 4 and in Supplementary Information 2, Section 4) might be a relevant factor for worldwide sociolinguistic typological investigations (Lupyan & Dale 2010; Bentz & Winter 2013; Sinnemäki & Di Garbo 2018), this may not necessarily be the case for family- or area-based studies (in Africa), where the demographic profiles of languages may not be normally distributed. In our data set, L1 speaker populations are skewed in both directions. We have many languages spoken by (very) small communities, and a small set of languages that are spoken by (several) millions of speakers. The effect of number of L1 speakers on the type of gender system is different in these subsets of languages, and we would argue that this is a by-product of the family-based approach that we took. We would also suggest that more relevant and fine-tuned variables than a count of L1 speakers can be constructed in order to capture contact dynamics related to incidence of contact and adult second language learning.²⁶

Furthermore, the models excluding languages with only animacy-based gender or no gender (Supplementary Information 2, Section 5) suggest that, overall, the effects that we find significantly rely on this subset of most restructured languages. These observations may suggest that such languages should be actually excluded from future work on the typology and evolution of Bantu gender systems, because they are so markedly different from the rest of the family. We would strongly disagree with this idea for a number of reasons. First, several of the sociogeographic effects remain relevant in the abovementioned models even excluding the languages with only animacy-based gender or no gender, which suggests that these factors are also important for explaining the distribution of languages with more traditional gender systems. Second, in those areas where we find several languages with only animacy-based gender or no gender, there are also several languages that we could not include in the sample due to lack of data (see Fig. 2). We cannot ignore that more instances of heavily restructured gender systems could be found among these. Languages with heavily restructured gender should ultimately not be seen as marginal outliers, or at least not in this part of the Bantu-speaking world. Most importantly, irrespectively of how many they are, these languages are of crucial importance to understand how gender systems change, both inside and outside NWB.

5.2 Prospects for future research

This paper is the first large-scale comparative study to demonstrate a palpable link between the shape and fabric of gender systems and a number of demographic and geographic factors. However, we feel that this is not the end of the story but rather the beginning. The next steps to be taken are those where we measure more exactly the impact that population movements within the Bantu family and contact with non-Bantu populations had on the evolution of Bantu gender systems. We can imagine the following options, all of which (unfortunately) require detailed research that falls outside the scope of this paper:

• Comparative studies of contact with Ubangi and Central Sudanic. This could be implemented through qualitative studies of language contact dynamics on the languages of our sample that border with Ubangi or Central Sudanic languages. Information regarding the gender systems of these languages, if they have any, would be vital to find out how NWB gender systems changed under “pressure” from non-Bantu-like gender systems and nominal morphosyntax. Such a qualitative study would also inspire further quantitative analyses of gender systems, as well as further relevant sociolinguistic and sociogeographic measures.

• Measures of the amount of time spent inside the rainforest. This could be achieved by looking at branch lengths in the trees of Grollemund et al. (2015), Koile et al. (in review), and future phylogenetic investigations of the Bantu language family. A variable of this kind could help to better disentangle the difference between older and recent waves of expansion in the rainforest, which, as argued by Güldemann (2018), is crucial to motivate the presence of both languages with only syntactic agreement and with restructured gender systems in present-day rainforest areas.

• Measures of sociolinguistic dynamics at the macro- and micro-level, including demography, processes of pidginization or creolization, information on register usage, passive understanding versus active usage, age at which a language was learned, and so on. This is a goal which is not achievable based on presently available sociolinguistic data, but ongoing research in the field of sociolinguistic typology (Di Garbo et al. 2021) makes it possible to believe that further advancement in the field could be achieved in the near future.

• Measures of the sociopolitical complexity of speech communities, for instance along the lines of the variable “Jurisdictional hierarchy beyond local community” in D-Place (Kirby et al. 2016). See also David Gil’s ideas (in the LingTyp discussion list, October 2018) about a scale of language-related sociopolitical complexity ranging from national languages, through local varieties of national languages and local languages historically part of a larger political entity, to local languages only recently part of a larger political entity. Such a measure could capture aspects of language use whenever more detailed sociolinguistic information is not available.

• Comparison of linguistic data on the distribution of restructured gender systems with molecular anthropological data. While such analyses are not immediately pursuable given the current level of genetic sampling in the relevant areas of interest (Pakendorf et al. 2011: 59), future advances in the field could make this possible.

A valid alternative to the steps above would be to compare the NWB data with better documented contact scenarios within the Bantu-speaking world. Probably the best candidate for such a comparative endeavor would be the southern Bantu languages, whose history of contact with non-Bantu, “Khoisan”-speaking populations is well documented both from a linguistic and a molecular anthropological point of view, and with the added advantage that the relevant contact languages are still spoken in the area and relatively well described. While extensive work has been done on Khoisan influence on southern Bantu at the phonetic level (Pakendorf et al. 2017), and the population dynamics in the context of which these patterns of language convergence took place have also been reconstructed with the support of molecular anthropological data, not much work has been done in the domain of nominal morphosyntax, let alone gender marking. Nevertheless, existing research points to the possibility that in the southern Bantu languages, contact-induced change in nominal morphosyntax goes in the direction of the weakening, and sometimes loss, of the locative and evaluative (diminutive and augmentative) genders, and, in parallel, of the grammaticalization of suffixes which express these functions in alignment with the marking strategies that are found in neighboring Khoisan languages (Creissels 1999; Güldemann 1999). Whether, in combination with these developments, any form of animacy-based agreement also occurs, and can be motivated as the result of contact with neighboring Khoisan languages, is, at present, unexplored territory.

In addition to areally restricted comparisons, a pan-Bantu study encompassing all sub-areas of the family would be the most natural long-term development of this work, especially as we suspect that animacy-based agreement is more widespread throughout the family than previously thought (Wald 1975; Maho 1999: 135). Through a comprehensive study of gender systems across the entire Bantu family, data on synchronic distributions could be used to test transition probabilities between types of gender systems and patterns of gender agreement. Their interaction with relevant language-external factors beyond those considered for the NWB area could also be studied.

A.1 Gender marking on nouns

• How many singular noun class forms?

• How many plural noun class forms?

• How many number-invariant noun class forms?

• How many singular/plural pairings of noun class forms?

A.2 Gender marking on agreement targets

• How many distinguishable singular agreement classes?

• How many distinguishable plural agreement classes?

• How many number-invariant agreement classes?

• How many paired singular/plural agreement classes?

A.3 What are the word classes that carry syntactic agreement?

The coding for these variables is a “yes/no/unknown” type of coding. Except for the variable “Other,” which is listed at the end, variable names are ordered alphabetically.

• Attributive adjectives: adnominal modifiers encoding property words.

• Copula-like constructions: constructions expressing nominal and/or locative predications.

• Demonstrative modifiers: adnominal modifiers indicating different degrees of spatial distance from the speaker and/or the listener.

• Demonstrative pronouns: pronominal expressions indicating different degrees of spatial distance from the speaker and/or the listener.

• Genitives/connectives: markers that are used to introduce nominal possessors. In Bantu languages, they generally consist of the stem a preceded by a pronominal prefix, which agrees in gender with the possessor. They are also used to encode adjectival types of meanings with modifying nouns encoding properties and/or entities.

• Independent third person pronouns: anaphoric pronouns corresponding to ‘he/she/it’ in English.

• Numerals: adnominal modifiers encoding cardinal numbers. In Bantu languages, ordinal numbers also agree in gender, but they are expressed through genitive constructions with cardinal numbers as modifiers (thus gender agreement is marked on the genitive relator rather than on the numeral as such).

• Quantifiers: adnominal modifiers encoding quantity expressions such as, for instance, ‘some’, ‘all’, ‘many’.

• Possessive pronouns: pronominal prefixes agreeing in gender with the possessee.

• Predicative adjectives: property words used predicatively, i.e., taking TAM inflections.

• Question words: selective interrogative such as ‘how many?’ and ‘which?’, as well as interrogative pronouns (‘who?’, ‘what?’)

• Reflexive pronouns: reflexives in Bantu are usually invariable prefixes, which are part of the set of inflectional markers that a verb can take. However, in some cases, there can be reflexive intensifiers, which are independent words that can sometimes take pronominal markers in agreement with the gender of the noun. This is what we target in our coding.

• Relative pronouns/Relative constructions: independent pronominal expressions functionally comparable to the English ‘who/whom/which/that’. Constructions encoding relative clauses, which do not fall under “relative pronouns.”

• Verbs: lexemes for the encoding of prototypical predicative expressions (actions, states).

• Other targets and/or domains of gender marking: here we include anything that cannot be captured by the features listed above.

A.4 What are the word classes that carry animacy-based agreement?

The coding for these variables is a “yes/no/unknown” type of coding. Except for the variable “Other,” which is listed at the end, variable names are ordered alphabetically.

• Attributive adjectives: adnominal modifiers encoding property words.

• Copula-like constructions: constructions expressing nominal and/or locative predications.

• Demonstrative modifiers: adnominal modifiers indicating different degrees of spatial distance from the speaker and/or the listener.

• Demonstrative pronouns: pronominal expressions indicating different degrees of spatial distance from the speaker and/or the listener.

• Genitives/connectives: markers that are used to introduce nominal possessors. In Bantu languages, they generally consist of the stem a preceded by a pronominal prefix, which agrees in gender with the possessor. They are also used to encode adjectival types of meanings with modifying nouns encoding properties and/or entities.

• Independent third person pronouns: anaphoric pronouns corresponding to ‘he/she/it’ in English.

• Numerals: adnominal modifiers encoding cardinal numbers. In Bantu languages, ordinal numbers also agree in gender, but they are expressed through genitive constructions with cardinal numbers as modifiers (thus gender agreement is marked on the genitive relator rather than on the numeral as such).

• Quantifiers: adnominal modifiers encoding quantity expressions such as, for instance, ‘some’, ‘all’, ‘many’.

• Possessive pronouns: pronominal prefixes agreeing in gender with the possessee.

• Predicative adjectives: property words used predicatively, i.e., taking TAM inflections.

• Question words: selective interrogative such as ‘how many?’ and ‘which?’, as well as interrogative pronouns (‘who?’, ‘what?’)

• Reflexive pronouns: reflexives in Bantu are usually invariable prefixes, which are part of the set of inflectional markers that a verb can take. However, in some cases, there can be reflexive intensifiers, which are independent words that can sometimes take pronominal markers in agreement with the gender of the noun. This is what we target in our coding.

• Relative pronouns/Relative constructions: independent pronominal expressions functionally comparable to the English ‘who/whom/which/that’. Constructions encoding relative clauses, which do not fall under “relative pronouns.”

• Verbs: lexemes for the encoding of prototypical predicative expressions (actions, states).

• Other targets and/or domains of gender marking: here we include anything that cannot be captured by the features listed above.

A.5 Additional questions

• Is animacy-based agreement obligatory outside the NP?

• Is animacy-based agreement obligatory everywhere?

• Does agreement only signal number?

• Do noun class forms only mark number?

• Do noun class forms only mark animacy?

• Do noun class forms mark animacy and number?

• Is there extra marking of animacy on nouns (e.g., animacy markers are juxtaposed to the nominal gender markers)?

• Is there extra marking of plurality on nouns (e.g., in addition to their nominal gender markers, nouns take an additional plural marker which is gender-invariant)?

• Is there extra marking of animacy and number on nouns (e.g., animacy/number markers are juxtaposed to the nominal gender markers)?

• Notes (this is a free text variable where the coder can write in any additional remark on the language which is being described).