Bantu languages in particular have drawn much attention for the complex behaviour of their nominal systems. One major locus of debate revolves around what motivates the categorization of nouns into classes in the systems found in the languages of this large family. All Bantu languages are characterized by a set of prefixes that attach to stems (e.g., mû-rîmi, “farmer”, kî-banga, “machete” [Kîîtharaka]) in addition to prefixal agreement markers which appear on dependent words (e.g., demonstratives, adjectives, possessives) (see e.g., Bleek 1862; Guthrie 1948; Corbett 1991; Carstens 1991; Katamba 2003). These nominal prefixes have sometimes been treated as class markers, i.e., indicating the class of the noun and in turn triggering agreement on dependent words. However, they are in fact very similar to nominal endings like -ette found in French (or -o/-a in Spanish): a given prefix tends to be associated with a given agreement pattern (see e.g., Harris 1991: on why the Spanish -o and -a should not be considered gender markers per se but as predictors of gender). For this reason, we treat the ubiquitous nominal prefixes in Bantu as potential predictors or cues to agreement class and/or gender. Examples of noun class agreement in Kîîtharaka are given in (1), (2) and (3).2 Example 3 illustrates a case, similar to the French exceptions provided above, where the nominal prefix (class 9) and the agreement prefix (class 1) are not aligned. In this case, therefore, the gender of the noun cannot be predicted based on the form of the noun.

Just as in other gender systems, like French, both form (the morphophonological shape of the noun) and meaning have been argued to play a role in Bantu gender assignment systems. Early research argued for a semantically arbitrary system, where class is determined purely based on noun prefixes (e.g., Krapf 1850; Bleek 1862; Meinhof 1906: among others). However, later influential research argued that noun classes are actually associated with semantic regularities (e.g., Guthrie 1967; Richardson 1967; Welmers 1973; Zawawi 1974). This motivated a reconstruction of Proto-Bantu, placing semantics at the centre of its nominal classification system (e.g., Givón 1971b; Givón 1971a; Welmers 1973; Denny & Creider 1986). For example, Denny & Creider (1986) suggested a two-way, semantically-motivated system for count and mass nouns in Proto-Bantu. Count nouns were argued to be broadly categorised on the basis of kind (i.e., animate vs. inanimate) and configuration (i.e., solid objects versus objects with a clear outline, edges, or distinct inside and an outside). Mass nouns were argued to be categorised based on whether they are cohesive or dispersive.

However, as in many complex nominal systems, descriptions of Bantu nouns classes based purely on noun form or on semantics are problematic in several respects. First, similar to the French examples given above, nominal prefixes in Bantu languages do not always deterministically align with agreement patterns. In these cases, researchers typically consider that agreement, not the noun form, indicates its class/gender (see e.g., Herbert 1985; Corbett 1991). That is, the agreement pattern a given noun will take is not always straightforwardly determined from the nominal prefix (see e.g., Msaka 2019: for how the prefix-based system fails to characterise the Chichewa nominal classification system). We will further illustrate this using Kîîtharaka data in the next section. Secondly, there are many nouns in Bantu languages without overt prefixes that nonetheless trigger agreement on nominal dependents. This complicates a purely form-based system (though of course positing something like a null prefix, or a default class is possible). At the same time, the productivity of semantic rules in Bantu nominal systems has been a subject of rigorous debate, with some scholars (e.g., Richardson 1967) arguing that the semantic foundation of the Proto-Bantu system is empirically unverifiable, and others concluding that it is only relevant for animate/human nouns (see, e.g., Crisma et al. 2011; Msaka 2019). Regardless, it is clear that, like for nominal prefixes, the semantic features of nouns do not always perfectly align with the agreement patterns they take.

Typological studies of nominal classification systems have, nonetheless, highlighted the fact that in most complex gender systems, both semantics and morphophonology interact to determine a noun’s gender, i.e., the agreement class it belongs to (see e.g., Corbett 1991; Di Garbo 2014; Corbett & Fedden 2016; Güldemann & Fiedler 2019: and others). As such, a nominal classification system can be considered as having both a set of agreement classes and a set of nominal form classes (Güldemann & Fiedler 2019). Agreement classes are comprised of nouns that share similar agreement across all agreement targets. For instance, the Kîîtharaka noun phrases muntû û-mwe, “one person” and antû ba-îrî, “two people” indicate that the stem -ntû takes agreement û- in singular and ba- in plural. Pairing of nouns on the basis of singular and plural agreement indicates the abstract gender of a noun called target gender in (Corbett 1991: et seq.). In this way, gender can be seen as an abstract feature of a stem, stored in a speaker’s mental lexicon. On the other hand, nominal form classes are based on shared morphophonological attributes of nouns, e.g., nominal prefixes. The Kîîtharaka nouns î-thaga, “metal” and î-rema, “tent” belong to the same form class, while ma-thaga “metals” and ma-rema “tents” belong to another one. Pairing nouns on the basis of singular and plural nominal forms gives rise to what has been called deriflection classes—the morphophonological equivalent of gender (Güldemann & Fiedler 2019).3

Recent research on Bantu nominal classes has used corpus evidence—in particular, lists of nouns along with their nominal form and agreement classes—in order to quantitatively evaluate how both morphophonological and semantic cues together can predict the gender of a noun (e.g., De Schryver & Nabirye 2010; Ngcobo 2010; Taljard & De Schryver 2016; Msaka 2019). However, while these analyses increase the reliability of the data on which theories of nominal classes are based, there has not yet been research applying any specific quantitative theory of productivity to a Bantu gender system.

Therefore, the goal of this paper is first to characterise the nominal classification system of Kîîtharaka, in terms of both gender and deriflection. We will then endeavour to determine whether morphophonological features of nouns (i.e., nominal prefixes) and particular aspects of noun meaning (i.e., noun semantics) are productive cues to gender, i.e., to the agreement patterns that a given noun displays. To do this, we will use the Tolerance Principle, described above, which allows us to evaluate the empirical data we have from a large list of Kîîtharaka nouns, and make predictions about productivity in the face of exceptions in this complex system. We introduce the Kîîtharaka nominal classification system and its particular complexities in the next section.

Kîîtharaka [Bantu, E54] is spoken in Tharaka Nithi County in Kenya by about 215,000 Atharaka people who mainly live in Tharaka North, Tharaka South and Chiakariga Sub-Counties (KNBS 2019: 424).4 The language is relatively understudied, though there is some previous work describing its nominal system (see e.g., Bible Translation & Literacy 1993; wa Mberia 1993). In a brief description, wa Mberia (1993) for example, characterises Kîîtharaka as having 17 “noun classes”, based on the nominal prefix, most of which can be organized as singular/plural (deriflection) pairs as is traditional in the Bantuist literature. Notably, he also outlines a set of potential semantic features, such as cultivated fruits, shrubs and trees, birds, insects, human beings, etc., that characterize the noun system in Kîîtharaka.

As with many semantic accounts of noun class systems in Bantu, this characterisation is largely based on subjective impression or intuition. Further, wa Mberia (1993) notes that there are many exceptions in each class—i.e., nouns which do not share the relevant semantic feature(s). Many of the semantic features he mentions are also invoked for multiple classes. As discussed above, exceptions are common in these types of systems, and that is part of what makes quantitative evaluation using a theory of morphological productivity important. Therefore, we will return to these issues below. Table 1 shows the kind of description of Kîîtharaka given by wa Mberia (1993), with a list of the class numbers for reference. The table includes the nominal prefix associated with each class, and an example noun with agreement (on a numeral) in each class.5 (vowels), prosthetic consonants [y] and [c] are inserted before a stem that starts in a vowel, to satisfy the phonological constraint that two distinct vowels cannot appear at the beginning of a word.

Notably, Table 1 obscures a number of more complex properties of the Kîîtharaka nominal system. First, some classes are treated as distinct despite having identical nominal prefixes—classes 1 and 3, 9 and 10, 15 and 17—unexpected if classes are defined on this basis alone. There is evidence from agreement that the class pairs 1/2 and 3/4 should be distinguished, but this is not done in all cases. For example, classes 2 and 2a are not treated as fully distinct, and many classes which share the same agreement prefix are not collapsed. This reflects the problem of combining the two different notions of gender and deriflection discussed above (as noted by Güldemann & Fiedler 2019). Moreover, what is shown here is the typical patterning of nouns in the system. In fact, there is quite a bit of variation. For example some nouns with the typical class 7 prefix k(g)î-, class 9 prefix n- or with no prefix actually take class 1 (û-) agreement.6 Similarly, nouns in various nominal form (nominal prefix-based) classes e.g., class 1, 7, 9, take the plural agreement prefix ma-, the pattern that is normally used with class 5/6 nouns (see Figure 1, for an illustration of this class mismatch and also Section 2.2 for how such cases are accounted for in our approach).

Second, although it appears that nouns in classes 14 and 15 do not have plural counterparts, this is not entirely correct. For example, the bulk of nouns in class 15 are derived from verbs (specifically infinitives) but there are also some body parts e.g., gûtû “ear”, kûgûrû “leg” which take class 15 agreement in the singular and class 6 ma- agreement in the plural.7 Similarly, class 14 includes mainly abstract nouns but there are some others as well, which either take ma- agreement, if interpreted as collective plural, and otherwise n- plural agreement.

To capture these patterns, we therefore describe Kîîtharaka using both nominal form and agreement classes, which respectively correspond to a deriflection and a gender system (Güldemann & Fiedler 2019). We leave aside here nouns that do not have plural or singular versions either because they are singulare/plurale tantum or because they do not make number distinctions. We treat such nouns as transumerals following Güldemann & Fiedler (2019), hence outside the gender system, together with infinitives. We therefore consider the agreement triggered by such nouns to constitute general agreement classes (GAC) (see Msaka 2019: for a similar treatment of infinitives, locatives and CPs). This results in 8 deriflectional and 8 gender categories for Kîîtharaka, shown in Table 2. Following the established tradition in Bantu (see e.g., Carstens 1991) we label the genders as presented in Table 2. Diagrammatic representations of the deriflection and gender system are shown in Figure 2.

An obvious observation from Figure 2 in reference to Table 1 is that class 17 has disappeared and class 3 is expressed in the same way as class 1—both take the agreement û- in singular. We have only come across one exceptional case—in the context of numeral 1 “one”, the class 3 agreement is articulated with a high tone. For the purpose of comparisons with the set tradition in these analyses, we do not dissolve these classes but position them together to show that they share similar agreement patterns. From this analysis, Kîîtharaka can be characterized as having a “crossed” gender system according to the terminology of Corbett (1991), with variable convergence at classes 6 and 10. Class 6 has elsewhere in the Bantu literature been argued to function as a default plural (see e.g., Bosire 2006; Ström 2012; Fuchs et al. 2018; Fuchs & van der Wal 2022). However, outside of regular class 5 nouns, many of the other nouns that pluralise in this class have a collective interpretation in the plural. Another observation is that the deriflection system lumps the prefix-less nouns together, but these nouns belong to different agreement classes, hence genders. These prefix-less nouns (characterized as belonging to class 1a/2a or 9/10 in the traditional prefix-based system) belong to class 1/2 or 9/10 on the basis of agreement.

A number of these observations point to the fact that, as in other Bantu languages, some of the nominal prefixes of Kîîtharaka can be characterised as multifunctional—they function as the primary prefixes for a particular set of nouns, but other nouns can also take these prefixes (either alone, or in combination with their primary prefix) in order to express particular evaluative meanings. For example, the prefix k(g)î- is typically associated with class 7 (Gender D) nouns. However, a noun from gender A (e.g., mû-ana) can take the prefix (as in kî-ana) to derive a pejorative meaning “ugly child”. In the Kîîtharaka system, class 12/13 can be used to express diminutive meaning, class 5/6 can express augmentative meaning while pejorative meaning is associated with class 7/8. Prefixes used with this more overtly derivational function are sometimes referred to as secondary noun class prefixes (see e.g., Fortune 1970; Dembetembe 1995; Harjula 2006; Déchaine et al. 2014; Di Garbo 2014; Dube et al. 2014; Taraldsen et al. 2018: and others), or as multifunctional morphemes whose particular meaning is dependent on context (as in Di Garbo 2014; Msaka 2019). In many Bantu languages the primary nominal prefixes for the noun classes with loose semantics (i.e., classes which appear to contain nouns from diverse semantic domains), like class 5/6, and 7/8, have apparently been “recycled” to express evaluative meaning (Déchaine et al. 2014; Di Garbo 2014; Msaka 2019). It is however important to note that even in their recycled use, these evaluative prefixes typically stack on the primary prefixes and obligatorily control agreement, thus behaving like the primary counterparts.

An important empirical question is how to accommodate these productively derivational elements within a noun class or gender system in a way that accounts for their unique morphosyntactic properties. Some researchers treat these prefixes as theoretically identical to other prefixes, i.e., they bear noun class features (see e.g., Mufwene 1980; Myers 1987; Carstens 1991; Bresnan & Mchombo 1995; Maho 1999). Others attribute the multifunctionality to different syntactic positions that the two sets of prefixes occupy (see e.g., Déchaine et al. 2014; Fuchs & van der Wal 2022). We will adopt a similar approach here and leave the issue for future research.

To summarize this section, we have here described the major trends in the Kîîtharaka noun class system. The focus here was the relationship between gender (as defined by singular and plural agreement pairs) and deriflection (as defined by singular and plural noun class prefix pairs). We have also highlighted the fact that these two systems do not perfectly match up: there are nouns in particular agreement classes that do not share a nominal prefix (and vice versa). In the next section, we bring in semantic features, which share this property: there are potential correlations between these features and gender, but there are also clearly exceptions. We then assess the degree to which nominal prefixes and semantic features predict Kîîtharaka gender using the Tolerance Principle as an evaluation metric.

As noted above, the Tolerance Principle is a theory of productivity in learning, and thus perhaps the most appropriate corpus from which to derive predictions for Kîîtharaka might be a corpus of child-directed speech (CDS) (Yang 2016; 2018). However, it has been shown that in the absence of such a corpus, common nouns can be sampled from an adult corpus (e.g., the most frequent 1000–2000 words) to represent the kinds of nouns children are likely to encounter in the acquisition process (see e.g., Yang 2018; Kodner 2020). Here we use two adult language corpora: a large extract from the Kîîtharaka translation of the bible, and the Summer Institute of Linguistics (SIL) African Wordlist. While frequency data would in theory be available for the bible sub-corpus, this alone may not represent the kind of nouns children come across. The SIL African wordlist sub-corpus therefore supplements this, comprising the most common nouns in an African context.8 A total of 45,844 word tokens (9,656 word types) was extracted from the Kîîtharaka bible (Bible Translation & Literacy 2019). This text source was chosen because it is the only existing Kîîtharaka corpus. As it was recently translated (published 2019), the bible text can be considered a reasonable source of synchronic language data. The bible text used here includes sections from both the old and new testament to increase type variation across the corpus (see Evans 2007). Chapters that mainly contain a list of proper names (as in Numbers and Chronicles) were excluded. Around 60% of the corpus is from the old testament and the remaining 40% is from the new testament. The second part of our corpus, the SIL African Wordlist, was translated by the first author (see Snider & Roberts 2004). The list contains names of common things like local birds, trees, animals, insects and other phenomena that may have been missing from the bible text.

The bible text was uploaded to lancsbox, a corpus software tool developed by Lancaster University (Brezina et al. 2020). We then extracted a word list which enabled us to manually identify nouns (see Ngcobo 2010 for a similar approach). Further processing of the word list was done so that there was only a single lemma form for each noun retained. Nouns were manually extracted from this word list. This resulted in a total of 901 nouns from the bible text. Of the 1,622 translatable words in the SIL Comparative African Word List (between items 1.1 and 11.1),9 1,426 noun types were not already present in the bible corpus. Together the total noun corpus therefore consists of 2,327 nouns. Each noun was evaluated against a set of semantic features collected from previous work on Kîîtharaka specifically and Bantu more generally (see, Creider 1975; Denny & Creider 1986; wa Mberia 1993).

The complete set of semantic features is shown in Table 3 along with the classes for which these features are expected to be potentially relevant. All the nouns in the corpus were coded for all features, and we tested the predicted productivity of each feature for each agreement class.10 The corpus can be accessed on the Open Science Framework (OSF) using the link: https://rb.gy/ows7r1.

Additionally, morphophonological features—the set of all nominal prefixes in Kîîtharaka, including the null prefix—were also identified and coded for in the same manner as semantic features. The prefixes coded for and their respective expected classes are shown in Table 4. As with the semantic features, all nouns were coded for each prefix, and we tested the productivity of each prefix for each agreement class.

In addition to these features, each noun was coded for agreement class and gender based on the observed patterning of singular/plural agreement classes. For example, there are 423 nouns that take û- (class 1,3) agreement in the singular. 196 of these take ba- (class 2) agreement in the plural, hence are classified as Gender A. 208 take î- (class 4) agreement hence are coded as Gender B. Of the remainder, 11 made no number distinction—coded as transnumeral (TransN) and 1 takes a class 10 prefix and respective agreement, hence placed in Gender D. There were some cases where the same noun lemma occurred with different singular agreement patterns in the corpus (i.e., across different instances in the bible). We dealt with these as follows: in cases where a lemma appeared with two alternative singular agreement prefixes, e.g., nkoma yathi (class 9 agreement) and nkoma athi (class 1 agreement) “(the) devil went”, we coded the noun as in the agreement class matching the nominal class prefix.11 Cases of alternative plural in class 6 were set as noted above.12 As far as we can tell, this kind of variation is not predictable from any obvious contextual feature. However, it may be that alternative agreement patterns reflect speaker evaluation—for example whether they want to treat the referent as more human-like, or not. We return to this in the Discussion below. There were also cases where the same noun lemma had clearly different meanings across contexts, with each meaning consistently corresponding to a different agreement pattern, e.g., kîrundu wa Ngai “the Holy spirit of God” (class 1 agreement) and kîrundu kîa ûrongo “the spirit of deceit” (class 7 agreement). These were treated as two distinct lemmas.

Each of the semantic and morphophonological features coded for was subjected to a quantitative analysis to establish its productivity using the Tolerance Principle (Yang 2016). As mentioned in section 1, the Tolerance Principle is an evaluation metric that provides a way of making predictions about the likely productivity of linguistic rules in the presence of exceptions. The Tolerance Principle makes predictions about how and when learners of a language should be expected to generalise a potential rule. Here, we will use it not to make predictions about when learners will acquire a rule, but as a way of establishing productive rules for gender assignment that speakers of the language may represent. As we have highlighted in the preceding sections, gender systems in Bantu (and beyond) are often characterised by exceptions, with multiple semantic features associated with one class, and even a single semantic feature that transcends several classes. Likewise, the various nominal prefixes (here treated as morphophonological cues to gender assignment) will also be evaluated.

In our case, the relevant variables include the number of nouns, N, with a certain semantic/morphophonological feature across the entire corpus, and the number of exceptions, e, which do not take the particular agreement class or gender for which we are evaluating productivity. Together, these give us the tolerance threshold, θN, for the potential rule, which we evaluate against the exceptions. For example, if there are N nouns with the Narrow feature in the corpus, and we want to evaluate whether the rule Narrow → gender B is productive, the exceptions will be the number of nouns with this feature that take some other agreement pattern, i.e., belong to another gender. Table 5 shows how this potential rule (among others) fares when evaluated under the Tolerance Principle. There are a total of 115 nouns with this semantic feature in the corpus. Out of these, only 32 are gender B (agreement class 3/4) and 83 are exceptional (i.e., are in other agreement classes). The tolerance threshold θN in this case is 24, and therefore the rule is not predicted to be productive. Similarly, Table 5 shows that this same semantic feature is not predicted to be productive for any other gender either.

Table 6 summarizes the results of our analysis of semantic features using the Tolerance Principle. Here we show only the evaluation of features with respect to the gender they were predicted to be relevant for. As expected, all other evaluated rules were predicted to be unproductive based on our analysis. Our results reveal that six of the semantic features tested are predicted to be productive as cues for gender determination in Kîîtharaka: Tree, Augmentative, Pejorative, Diminutive, Manner and Infinitive. All three of the evaluative features are predicted to be productive based on the subset of nouns that occur only in these classes (i.e., for which this is the primary gender) and for nouns that have the relevant class as an alternative (i.e., for this is a derived meaning, or secondary gender).13 We show analyses separating these two types of nouns in Table 7.14

The feature Manner appears unproductive for Gender D (class 7/8)—of 10 nouns, only 1 belongs to this gender. However, there is some regularity observed in that all of the 9 exceptions are transnumerals by virtue of being singulare tantum. Following the mechanism we established in Section 2.1 to account for the non-declensional paradigm, these kinds of nouns are not categorized as Gender D but are cases of GAC. Manner is therefore a productive cue to GAC in the same way as the evaluative features described above. This includes names of languages, e.g., Kîî-ibirania “Hebrew language” which serves as a nominal denoting the manner of “speaking like a Hebrew” and other phrases denoting behavioural attributes e.g., kîî-muntû “like a human”. Finally, and most notably, the feature Human—perhaps the semantic feature most often claimed to be relevant for Bantu—is not predicted to be productive. This is because, despite all nouns in Gender A being Human, the number of human nouns in other genders is above the productivity threshold defined according to the Tolerance Principle. However, the Tolerance Principle allows recursive application with feature conjunction. Intuitively, this can result in the feature being productive, though not in isolation. We return to this in the Discussion below.

Turning to morphophonological features, i.e., nominal prefixes, in contrast to the semantic features tested here, almost all of the morphophonological features tested are predicted to be productive. This is illustrated in Table 8, which shows the evaluations of each nominal prefix with respect to the expected gender. Only the prefix mû(u)- is unproductive, because it is associated (nearly equally) with two genders—A and B. The same applies to the class 10 prefix n- which is associated with two genders— E (class 9/10) and F (class 11/10). The prefixes û- and k(g)û- are productive for GAC, while for ba- there is only one noun bearing the prefix and since no computation is needed for this one observation, the Tolerance Principle by design does not yield any tolerance threshold.

Gender assignment systems are one of the most well-studied features of Bantu languages, but there is longstanding debate as to whether these systems are based on semantics, morphophonology, or both.15 We have argued that part of the reason this debate rages on is that previous work has largely shied away from quantifying the robustness of the features proposed to underlie gender assignment (though cf. Ngcobo 2010; Msaka 2019). Importantly, no previous work on gender assignment in Bantu has engaged with theories of productivity. In this study, we aimed to illustrate how a quantitative approach, couched within a theory of productivity, makes it possible to go beyond intuitions and to deal with exceptional cases in a principled way. We took gender assignment in Kîîtharaka as our test case. Although there are many theories of morphological productivity, here we used the Tolerance Principle (Yang 2016).

Our results suggest that for Kîîtharaka, only a few semantic features are predicted to productively cue agreement class, namely, Augmentative (gender C), Pejorative (gender D), Diminutive (gender G), Tree, and Infinitive (GAC, class 15). On the other hand, morphophonological features—the nominal prefixes—are widely predicted to productively cue gender in Kîîtharaka. In other words, Kîîtharaka speakers are mostly predicted to be able to productively determine the gender of a novel noun (here taken as the agreement pattern it shows) on the basis of a nominal prefix alone. The only apparent exception to this was the prefixes mû(u)- (which is shared by gender A and B) and n- (which appears with plural agreement classes of gender E and F).

In fact, across both parts of our analysis, it is notable that agreement classes 1 and 3 were not predictable from any of the features tested. Perhaps most surprisingly, the feature Human, a target feature for classes 1 and 2 (Gender A), was not predicted to be productive. Human/animacy has been claimed to play a key role in motivating assignment of nouns (to gender A) across the Bantu language family (see e.g., Wald 1975; Contini-Morava 2008; Ngcobo 2010; Crisma et al. 2011: and many others). Likewise, if singular prefixes alone marked agreement class or gender, it would be impossible to predict whether a mû(u)-prefixed noun is in gender A (class 1/2) or gender B (class 3/4). Assuming an explicit theory of productivity forces us to specify what it means for a particular feature to be productive. In this case, that theory requires that nouns with the relevant feature (or features) not be assigned to other classes beyond a certain threshold—clearly the case for humanness and mû(u)- in Kîîtharaka.

However, while it is possible that these features simply are not productive in Kîîtharaka, it is worth considering some other possibilities. First, it could be that our corpus, which is partially extracted from the bible, contains a higher number of certain kinds of nouns than one would find in other texts, or in spontaneous speech (e.g., spirits, God and other super-humans, all assigned the feature Human in our analysis). Indeed some of these nouns are exceptions in our corpus (i.e., they are not in gender A). For instance, kîroria “prophet”, though Human, is in agreement class 7/8 (gender D). An over-representation of such nouns could lead to an increase in the number of exceptions relative to cases that follow the rule, potentially incorrectly predicting the +Human feature to be unproductive for gender A. However, it is not clear how likely this actually is. In fact, a large portion of our corpus—1,426 nouns, more than half of the total—comes from the translated SIL Comparative African Wordlist rather than the Bible text. This is potentially a more balanced representation of commonly used nouns across African languages. Thus, we re-ran our Tolerance Principle analysis with just the subset of nouns from the SIL wordlist. The results were qualitatively similar—e.g., the Human feature was still not productive (see the SIL list only section of the corpus analyses—URL provided in Section 2.2).

A second possibility is that both the semantic feature Human, and the morphophonological feature mû(u)- do productively cue genders A and B, but they require recursive application of the Tolerance Principle. In other words, these features would be productive when re-evaluated based on a subset of the lexicon. In the analysis described above, we treated each semantic and morphophonological feature as individual and independent of any other features. However, both semantics and morphophonology may be used jointly to determine certain aspects of gender assignment. The Kîîtharaka corpus data indeed shows that all human nouns with the prefix mû(u)- are in gender A. Similarly, all non-human nouns with prefix mû(u)- are in gender B. If we treat Human as a binary feature (a possibility for gender features according to Lumsden 1992; Rooryck 1994: and others), the lexicon can be divided into [+Human] nouns and [–Human] nouns. The remaining features, like the morphophonological feature mû(u)-, can then be re-evaluated on each subset. In its original conception, the Tolerance Principle was indeed designed to be applied in such a recursive way. This was motivated by the Maximise Productivity principle, which asserts that learners are actively searching for productive rules (see Yang 2016; 2018). Learners could thus evaluate rules that apply to both the full lexicon, but also those that may apply to only a subset of the lexicon. Recursive application is predicted to affect the time-course of acquisition, but not the eventual productivity of the resulting rule. While recursive rule application also makes the space of possible rules to be tested much bigger, using these particular features immediately jumps out from our data. To explore this, we retested the productivity of the prefix mû(u)- and found that the rule [mû(u)-] → gender A was productive for the subset of the lexicon with the feature [+Human] (N = 166, n = 161, e = 5, θN = 32), and the rule [mû(u)-] → gender B was also productive for the subset of the lexicon with the feature [–Human] (N=231,n=207,e=24 \documentclass[10pt]{article} \usepackage{wasysym} \usepackage[substack]{amsmath} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage[mathscr]{eucal} \usepackage{mathrsfs} \usepackage{pmc} \usepackage[Euler]{upgreek} \pagestyle{empty} \oddsidemargin -1.0in \begin{document} \[ N = 231, n = 207, e = 24 \] \end{document} , θN = 42). Notably, in the same vein, the semantic rule [+Human] → gender A can also be reapplied to a subset of the corpus including nouns with the prefix mû(u)-. For this subset, the rule [+Human] → gender A is productive (N = 165, n = 161, e = 4, θN = 32). In other words, recursive application of rules targeting genders A and B with features mû(u)- and +Human give us additional evidence for productivity.

Finally, it is potentially worth looking more closely at the exceptions to the semantic rule [+Human] → gender A. In the entire corpus, there are 252 human nouns of which 196 nouns are in gender A. The remaining 56 human nouns are distributed across gender C (N = 2), gender D (N = 21), and gender E (N = 21) and a few in genders B and G. While this rule would not be productive with this number of exceptions, the exceptional nouns largely comprise humans with special attributes. For instance, there are those with supernatural abilities e.g., kîroria and ngai (gender D) “a lesser god”. There are also those with properties or behaviours that are deemed socially undesirable, e.g., îrwaya “prostitute” or kîonje “a cripple” (Gender C). Some of these nouns also convey pejorative meanings. For example, the noun kîrundu can mean “spirit of God” (gender A), “demon/evil spirit” (gender D) or “spirit of the dead” (gender D).16 These exceptional nouns are given in Table 9.

These examples reveal the possibility that at least some human nouns are not in gender A in order to convey a specific, mostly negative meaning. This accords with our general finding that the semantic features predicted to be productive in Kîîtharaka are mostly evaluative in nature: Augmentative, Pejorative, and Diminutive. And indeed, it has been argued that evaluation plays a key role in nominal classification in other Bantu languages (Castagneto 2017; Msaka 2019). If these nouns are considered not to have the feature Human, i.e., if they are treated as having a distinct feature that combines Human and Pejorative, this could impact how Kîîtharaka speakers represent these nouns and the classes they belong to. For example, re-evaluating productivity after removing nouns coded as Pejorative renders the rule [Human]→ gender A productive (N = 235, n = 196, e = 39, θN = 43). This is another example of the recursive application of the Tolerance Principle, this time by re-evaluating a semantic feature ([Human]) on a subset of nouns without a particular evaluative meaning ([Pejorative]).17

Regardless of exactly how learners come to form this generalisation, our guess is that the feature [+Human] can be used productively by Kîîtharaka speakers to predict the gender of novel nouns. This feature, and related features referencing animacy (or agency) have, in addition to evaluative features, been argued to be relevant in many Bantu languages see e.g., Ström 2012: 274–281; Di Garbo 2014: 42,148,176; Güldemann 2023. The remaining inherent feature which was found to be productive is Tree. This is somewhat surprising, and is supported by a relatively small set of nouns. Of course, the Tolerance Principle is a measure of predicted rule productivity. The next logical step is to look for evidence of whether the predictions from the present analysis are borne out for speakers of Kîîtharaka. For example, psycholinguistic experiments could test whether Kîîtharaka speakers—either adults or children— use the features that are predicted to be productive based on this analysis, and fail to use those predicted to be unproductive when assigning novel nouns to classes. As noted above, here we have used a corpus of nouns that may not precisely reflect the kinds of nouns young children encounter (though the SIL database may be a reasonable approximation). For this, additional data on child-directed speech in Kîîtharaka would be needed. In addition, while we have proposed at least one case where semantics and morphology likely work together, the question of whether speakers privilege one type of cue over the other (e.g., Karmiloff-Smith 1981; Pérez-Pereira 1991; Gagliardi & Lidz 2014; Lawyer et al. 2024), as well as the time course of acquisition in Kîîtharaka remains open. We leave these issues for future work.

It is also worth noting that the approach we have taken here is to treat meaning and form class markers (prefixes) as potential cues to agreement class, rather than the other way around. In some sense, this gives a special status to noun internal cues (Gagliardi & Lidz 2014), and at the same time assumes both noun form and meaning have equal status in the minds of speakers. However, acquisition data suggests that children learning at least some languages exhibit knowledge of nominal prefixes much earlier than semantic cues (e.g., Demuth 2003: and the references therein) and show sensitivity to the noun form earlier than to noun external syntactic alternations like agreeing determiners or adjectives (see e.g., Karmiloff-Smith 1981). However, this is likely to change at a later acquisition stage. Studies on the acquisition of Romance languages show that from age 10, children begin to make use of both semantic and syntactic distributional cues to determine gender (see, e.g., Karmiloff-Smith 1981; Pérez-Pereira 1991). Both child learners and adult users of a language could, therefore, make use of their knowledge of deriflection to predict agreement, and vice versa. In other words, it could be that gender in fact provides a cue to deriflection. This is argued to be the case in Icelandic, for example, where knowledge of nominal inflection class provides a productive cue to gender in some cases, but in others it is gender that provides the only productive cue to inflection class (Björnsdóttir 2021; 2023). In Kîîtharaka, for the most part, the mapping between deriflection and gender is one-to-one, and the gender of most nouns can be determined from their deriflection class. But this is not always the case. For example, as shown in Figure 1, when a noun lacks overt form class features—i.e., does not have a nominal prefix—it is not clear whether the noun is in gender A or gender E. In this case, learners may derive productive generalisations based on alternative cues, like the semantics of a noun or the agreement pattern it takes, and use these when encountering novel nouns.

At a more general level, while determining the psychological reality of specific cues and how they interact in child learners and adult users of Kîîtharaka are important questions for future research, our findings here highlight that robust predictions can be made about gender assignment in Bantu languages using a data-driven approach combined with an explicit theory of productivity. The same approach can in principle be applied to other Bantu languages. The results may align with ours—which suggest that morphophonological features are generally productive but many semantic features are not—or they may suggest differences across Bantu languages. Either way, the approach has the potential to lead to better understanding of how gender assignment works in this family of languages.