1 Introduction

This squib investigates how animacy is encoded in the grammar of Bantu languages. In particular, we explore how the animacy hierarchy, as formalized, for example, by Hammerly (2023), is directly encoded through the noun class system. A variety of syntactic phenomena from across the Bantu family suggest that animacy plays an active role. At the same time, the noun class system, which governs a wide variety of morphosyntactic processes, does not straightforwardly map onto animacy distinctions. These seemingly misaligned taxonomies raise a puzzle: what features are syntactically present on nouns in Bantu, and how are they realized morphologically? Drawing on recent work on noun class stacking (Fuchs & van der Wal 2022) and final vowel morphology (Mletshe 2019; Msaka 2019; Msaka & Biberauer 2024), we suggest that noun class morphology is essentially a hybrid system: a core semantic distinction between human, non-human animate, and inanimate nouns is present within all nouns, encoded by the final vowel. These basic distinctions are typically realized by class 1/2 for humans, class 9/10 for non-human animates, and class 7/8 for inanimates. Other noun classes can be derived via additional nominal structure stacked on top of this core. The existence of these core noun classes gives a pathway to account for various hierarchy effects and agreement phenomena observed across Bantu languages.

2 A theory of noun class

Our account centers around the idea that there are certain types of nouns classes, dubbed containment-type noun class (Hammerly 2023), which differ from other types of noun classes in their relation to agreement hierarchy effects. Containment type noun classes form part of a prominence hierarchy, which ranks speech-act participants and various other categories related to humanness, animacy, agentivity, concreteness, and more. An example hierarchy is given in (1).

Prominence scales have been used to describe grammatical patterns related to agreement, word order, and case-marking across languages of the world (e.g. Lockwood & Macaulay 2012). As we show in section three, some Bantu languages show animacy hierarchy effects for phenomena like agreement and concord; Hyman & Duranti (1982), for example, proposed the following animacy hierarchy as one of multiple factors determining the likelihood of object agreement:

Much work within Minimalism has sought to derive prominence-hierarchy effects from more fundamental principles of the grammar such as features. This paper builds on recent work by Hammerly (2020; 2021; 2023), who argues that features are further decomposed into sets of morphosyntactic primitives.¹ Primitives syntactically encode fundamental ontological distinctions made between different types of entities, including the author of a linguistic act (I), the addressee (U), non-interlocutor humans (O’s), animals (A’s), natural forces (F’s), concrete things (R’s), and abstract things (S’s). Morphosyntactic features such as [Author], [Participant], or [Animate] are made up of subsets of primitives.

Critically, with containment type noun class features (as well as person features) these sets of primitives stand in proper subset/superset relationships: for example, [Participant] is defined by the set {I, U}, which is a superset of the set {I} that defines [Author]. Therefore any syntactic operation that targets a category with a broader feature necessarily targets all features in a subset relation, capturing hierarchy effects (Hammerly 2020; 2021). The containment relations between features relevant to deriving the categories in (2) is schematized in (3), with the widest category that captures the entire ontology being ɸ.

These features combine to give rise to different nominal categories by way of bivalent feature values, specified as positive (+) or negative (–). At the limit, each primitive defines its own morphosyntactic category. This eventuality is schematized by the contrastive hierarchy in (4), which shows the logically possible feature-value combinations, excluding contradictory combinations (e.g. [–Participant] and [+Author]) and redundant combinations (e.g. [–Participant] and [–Author]).

We assume that primitives, features, and therefore the contrastive hierarchy, are universal: the human language faculty affords the potential to make the set of distinctions in (4).² However, not every language makes use of all possible features. For example, a language that lacks the [Indiv] and [Agent] features is one in which there is no grammatical distinction between abstract things, concrete things, and natural forces – conflating, in the sense of McGinnis (2005), the ontological elements S, R, and F into a category we might call generic inanimate, as schematized in the left-side Venn diagram in (5). If a language additionally lacks the [Hum] feature, it will conflate the human and animal categories, creating a generic animate that includes the ontological elements A and O, as schematized in the right-side Venn diagram in (5).

The account also predicts certain patterns of conflation should never be observed—for example, it is not possible to create a category that encompasses non-interlocutor humans (O) with natural forces (F) to the exclusion of animal beings (A). This is because the only features that contains both O and F also contain at least A (namely, [Agent] and [Indiv]), and there is no feature that contains A to the exclusion of O.

In the coming sections we propose that core noun classification in the languages of the Bantu family are encoded with containment-type features, with other secondary noun classes stacking on top and often obscuring this core system. This deeper pattern of core noun class is reflected in a variety of agreement-related phenomena, and morphophonologically encoded in the system of final vowels.

3 The logic of animacy in Bantu grammar

A salient feature of Bantu languages is the grammaticalized division of nouns into differnt noun classes. As we will see, it is initially difficult to map divisions between noun classes onto something like the prominence hierarchy discussed above. However, when we move away from the specifics of noun class, we find phenomena throughout Bantu languages that suggest a predictable encoding of prominence distinctions.

4 Encoding the animacy hierarchy: core noun classes in Bantu

We have now seen a number of phenomena that suggest that animacy should be encoded in the morphosyntax of Bantu nouns. Animacy-sensitive syntactic phenomena, including anti-agreement, animacy agreement, and object doubling, motivate the need for the syntax to access the animacy properties of Bantu nouns. In turn, the link between animacy-based categories such as human, animate, and inanimate and particular numbered noun classes is motivated by the patterns of convergence: all nouns must underlyingly fall into classes 1/2, 7/8, or 9/10 in order for such convergences to arise. Taken together, we can advance a hypothesis for the relationship between the animacy features [±Animate] and [±Human] and traditional Bantu noun classes, which forms the basis for our proposed analysis.

This basic hypothesis finds a number of antecedents in the literature. For example, Taraldsen et al. (2018) note that “The semantic content of N₁ [class 1] may be just ‘human’ and that of N₇ [class 7] just ‘thing.”’ (p. 1379) and “the N inside the class 9 prefix might denote something like living things” (p. 1376, footnote 53). Similar views go back further, as noted by Katamba (2004: 18): “They [Creider (1975); Denny & Creider (1986)] divide the classes into two basic sets, with partly overlapping morphology. One set includes classes 1/2, 7/8 as well as 9/10 which contain nouns that indicate ‘kinds of entities’ (i.e. people, tools and animals, respectively). The other set contains nouns that indicate spatial and shape configurations such as roundness, length and size”. Like all noun class systems, there are idiosyncrasies: for example, as previously noted, some humans fall into Class 7/8 or 9/10 rather than the prototypical class 1/2, and as such do not converge to show class 2 agreement under coordination. However, the presence of an idiosyncratic fringe does not negate the fact that there is a semantic core. And more importantly, these idiosyncrasies are still consistent with the hypothesis that all nouns fall into one of these three core classes, even if the particular class is not perfectly predictable.

Furthermore, as noted in Section 2, we predict certain patterns of conflation and/or syncretism between the three core classes should be possible. For example, a language may fail to distinguish (either entirely, or just in certain morphophonological corners) between non-human animates (canonically class 9/10) and inanimates (canonically class 7/8) by being insensitive to the [±Animate] feature, leaving just [±Human] to make the cut. This is the case for Xhosa, where SM8 and SM10 have the same morphophonological form (zi-), made possible by the fact that both share [–Human]. Similarly, we expect there to be cases where languages fail to distinguish human animates (canonically class 1/2) from non-human animates (again, canonically class 9/10), creating what we previously termed a “generic animate” pattern. This is the case, for example, in Swahili (e.g. example (13)), where both humans and animals converge to class 1/2 under animacy override. This is made possible by both categories sharing [+Animate].

With this hypothesis in hand, the next subsection builds a particular proposal by identifying a potential morphosyntactic locus for core noun class: the final vowels.

5 Obscuring the core: Stacking

Our way into a solution to these issues builds on Fuchs & van der Wal’s (2022) analysis of derived locatives, diminutive, and augmentative nouns in Bantu. There is variation across Bantu in whether derivational prefixes related to these functions replace the original class prefix (not shown), or stack, as shown with the pair of Shona nouns in (24):

Fuchs & van der Wal (2022) argue that, even in many cases where stacking is not apparent in the overt morphophonological expression of a given noun, these nouns can underlyingly have a “stacked” structure, where multiple ns combine to derive a nominal form (see also Kramer 2015: 208). This is schematized in the tree in (25) with the example from (24b).

Building on this idea, we propose that nouns in Bantu contain stacked nominalizing heads: first a core noun class projection that houses the features [±Animate] and [±Human], which is then modified by a secondary noun class, which as shown in (26) is the realization of a second nP.¹² This is analogous to the structure for diminuatives in (25), where two nPs are stacked on after the other.

While the stacked diminutives in (25) have two layers of nP each headed by a class prefix, we propose that the innermost n head is in fact the suffixal final vowel;¹³ any prefixes express additional layers of structure. This approach builds on Fuchs & van der Wal (2022), who suggest in their footnote 7 that final vowels may represent the lowest layer of a layered n structure, and Mletshe (2019: 24, figure 2), who treats the final vowel as an affix that derives a noun stem out of a verb root, and assumes that the prefixal class marker adds an additional N layer on top of the stem. As we will discuss in more detail below, we assume that nearly every Bantu nominal involves a minimally stacked structure.¹⁴

Interestingly, we are unaware of cases where an overt class 1 prefix stacks inside a secondary class—and note that Fuchs & van der Wal (2022) observe a similar gap in Swahili and Rangi augmentatives and diminutives stacking: while augmentative and diminutive class prefixes generally stack on top of inner noun class, they simply replaced class 1/2 prefixes. Investigating whether there is a true gap for class 1/2 prefixes stacking beneath other class prefixes is a promising line of future research.

Let us now consider a few key example cases given in (27), all with the same root khohlel ‘cough’, but differing in the specification of final vowel and class prefix. The form in (27a) represents the “typical” case, where the animate final vowel and class 1 prefix align, (27b) shows a misalignment such that the animate final vowel combines with a class 7 prefix, and finally (27c) which, like (27a), shows an alignment—but this time between the inanimate final vowel and the class 7 prefix.

When the animacy of the final vowel and noun class align as in (27a) and (27c), we assume that there is a second n stacked over n_coreP that lacks its own class feature and is spelled out with the appropriate allomorph depending on the feature specification of n_core. That is, class 1 in the case of [+Animate, +Human], class 9 in for [+Animate, –Human], or class 7 for [–Animate, –Human]. A main reason to assume that these basic cases involve stacking—rather than, say, a discontinuous morpheme realized as the prefix + final vowel—is that there are cases in which prefixal n is systematically absent. Notably, as Table 1 illustrates, class 1a across the Bantu languages lacks a noun class prefix but otherwise behaves like a class 1 noun; similarly Msaka & Biberauer (2024) shows that some deverbal nouns can be formed without a class prefix. We therefore take the presence of a prefix to correspond to a stacked n head above n_core, and the absence of a prefix to reflect an absence of the higher n projection.¹⁵

With examples like (27b), where the final vowel and class prefix are misaligned—in this case the final vowel appearing in the animate form and the prefix surfacing in the canonically inanimate class 7—we propose that the stacked n bears its own class feature.¹⁶ The distinct features on each subsequent n are spelled out as expected. This is schematized with the tree in (28), where we follow Visser (2008) in assuming that “augment” of the prefix is the realization of D in Xhosa.¹⁷

The same approach can apply to all possible misalignments between core noun class (as expressed by a final vowel, found in the patterns of agreement, or deduced semantically) and prefixal noun class: nouns falling into class 3/4 or 5/6 based on the secondary class of the prefix always contain n_core with features [+Animate, +Human] if human, [±Animate, –Human] if non-human. This means that all nouns in these classes, at least underlyingly, have a stacked structure with distinct feature sets. We can then explain the pervasive animacy hierarchy effects across Bantu languages, which occur even with nouns that fall outside of classes 1/2, 7/8, and 9/10: all nouns are specified for [±Animate, ±Human], so these features can be available in the morphosyntactic computation of agreement and other phenomena.

6 What remains

The primary proposal we defended in this squib is that, in order to capture pervasive animacy-hierarchy effects, all nouns in Bantu languages must be specified for core noun class, resulting in every noun falling into the category human, non-human animate, or inanimate. The hierarchy is encoded by the features [±Animate, ±Human]; in cases where these features align with noun class, they are expressed by a predictable final vowel combining with class 1/2, 7/8, and 9/10 class prefixes and agreement, (contra Carstens 2026: who rejects a grammaticalized role for the animacy hierarchy). We further proposed that these features are encoded on the nominalizing head n. Cases of misalignment between a noun’s core feature specification versus class as expressed by the class prefix are due to the stacking of multiple nPs with distinct features, in line with the proposal of Fuchs & van der Wal (2022) for Bantu diminutives, augmentatives, and locatives. Evidence for the universal specification of animacy features comes from a wide range of animacy-sensitive effects in agreement, including anti-agreement, animacy override, and agreement with conjoined subjects. The underlying logic was this: if there are morphosyntactic effects of animacy, then there must be a morphosyntactic representation for animacy.

We see this proposal as a foundation for pushing ahead and testing theories of noun class in Bantu languages. More details are needed in a number of areas, and we recount two of them here and give a sense of direction.

First, how are other classes represented? We proposed a specific feature representation for classes 1/2, 7/8, and 9/10 with the features [±Animate, ±Human]. Classes 11 and beyond are likely associated with particular features as well, such as those that encode the diminutive, augmentative, or locative. However, the encoding of classes 3/4 and 5/6 are less clear. One possibility, based on the fact that 5/6 is associated with liquids and other miscellaneous things (i.e. abstract things), and 3/4 with trees/plants (i.e. concrete things), is that these categories could be derived by adding in more core noun class features as schematized in (4). The synchronic shakiness of these semantic associations could be due to semantic bleaching, with the system starting out as strongly based in the core semantics, with idiosyncratic cases slowly increasing over time.

Indeed, this process of bleaching is discussed at some length in Corbett (1991: 72–74) and references therein in the context of loanword adaptation in Bantu. Loanwords are overwhelmingly assigned to a noun class on the basis of the morphophonological properties of the borrowed words. While there are also documented semantic effects, most commonly, loanwords are assigned to the class within a language that happens to lack a prefix in the singular. This allows speakers to adapt words without changing the morphological properties of the noun (i.e., there is no class prefix to be found on the borrowed word, so it is put in the class that lacks a prefix). However, loanwords can be borrowed into other classes if their initial sound segments happen to line up with one of the existing noun class prefixes. Corbett suggests that borrowings are one of the possible catalysts for the semantic bleaching of noun classes, where the system becomes less semantically based and more phonologically based as loanwords enter a given class.

Second, how does the computation of the various agreement phenomena access core noun class, and ignore the noun class as expressed by the class prefix? For the phenomena we have discussed, two different mechanisms are likely at play. For something like animacy agreement, where a probe is sensitive to properties of the core noun class that are obscured by stacking, some sort of persistent probing mechanism (e.g. Béjar & Rezac 2009; Deal 2015; Hammerly 2021) that allows a probe to continue to search for a more highly specified version of some feature (such as [Human]) after encountering a less specified version (such as [Animate]), could capture the range of patterns we find across Bantu. Differences in the sensitivity of phi-probes—both within and across languages—yield the variation in attested effects. For example, in a language like Zulu, where phi-agreement for subjects and objects always tracks secondary noun class, a simple “flat” phi-probe would find the closest phi-features of any type, on the outer n, as in (29a). Swahili, by contrast, would have a more articulated phi-probe; the presence of a [+Animate] feature on the probe could cause the probe to search past the outer n to find a [+Animate] on the n_core, as in (29b).¹⁸ Nyaturu, which has agreement with subjects from all noun classes, but only animate objects, might have a subject phi-probe as in (29a) and an object probe as in (29b).

Patterns with coordinate structures likely have at least a partially different source; here, the internal structure of the coordination yields a single set of features—sensitive to the core animacy distinctions—that is visible to a higher probe. Much therefore depends on the process of resolving feature conflicts between the two conjoined nouns, which goes beyond the scope of the current paper (though see Adamson & Anagnostopoulou (2025) for a recent of account of gender resolution in Greek, and Carstens (2026) for an account of Xhosa).¹⁹

Abbreviations

aa alternative agreement, anim animate, cl noun class, dj disjoint marker, fv final vowel, inan inanimate, nfv nominalizing final vowel, om object marker, pl = plural, sg = singular, sm subject marker, vfv verbalizing final vowel.

Noun class is indicated by the corresponding number; local persons are glossed with 1sg/pl or 2sg/pl. Other abbreviations follow the Leipzig glossing rules.

Acknowledgements

We would like to thank audiences at the University of Göttingen and in the Fall 2022 edition of LING 521 at UBC for feedback.

Competing interests

The authors have no competing interests to declare.