Glossa Psycholinguistics

1. Introduction

Grammatical gender and noun class systems categorise nouns into classes, reflected by agreement on dependents words, such as demonstratives, adjectives and numerals (see, e.g., Aikhenvald, 2006; Corbett, 1991; Hockett, 1958; Katamba & Stonham, 2006, amongst others). In these systems, nouns that trigger the same agreement patterns are sometimes referred to as sharing agreement class, or gender. Nouns that themselves share morphophonological properties are in the same form or deriflection class (Güldemann & Fiedler, 2019).¹ In some languages, these largely overlap, but in others, they can diverge. In such cases, not all nouns which share some morphophonological feature (e.g., a prefix or suffix) will necessarily have the same gender. For example, in French, most nouns that end in –ette trigger feminine agreement on their dependents, but there are some (e.g., squelette ‘skeleton’) which trigger masculine agreement. In Bantu languages where gender is, in principle, determined by singular-plural agreement class pairs, most nouns are marked with a nominal prefix. Those which share a prefix are in the same form class, but as in French, not necessarily the same gender. Table 1 below provides a summary description of the noun classification system of the Bantu language Kîîtharaka. This table shows, for example, that some nouns with the prefix mû- are in agreement classes 1/2 (gender A), but others with the same mû- prefix (hence, same form class), are in classes 3/4 (gender B).²

In addition, there are exceptional cases when the nominal prefix is not the one typically expected given the class marked by the agreement prefix. For example, in (1) and (2), a noun with a prefix that is in form class 5 and 9, respectively, triggers agreement class 1.

At the same time, as in all noun class systems (Corbett, 1991), there is evidence in Kîîtharaka that at least some nouns in a given gender category share aspects of meaning. For instance, most human nouns are in gender A (see Table 1 for an example). As with morphophonology, though, there are exceptions (e.g., see (3)).

This lack of perfect correspondence between the form of the noun, or its meaning, and the agreement it triggers highlights a longstanding question in research on gender and noun class systems: what exactly determines the agreement class of a noun? That is, what productive rules do speakers of a language form regarding the features of a noun that might predict its gender or class? While it is often argued that gender is stored as a lexical feature of the noun (e.g., see V. M. Carstens, 1991, for claims about Bantu in particular), there is, nevertheless, robust evidence that learners and speakers make use of such cues productively, e.g., to determine the gender of novel nouns (or loan words) (e.g., Demuth, 1992; Gagliardi & Lidz, 2014; Karmiloff-Smith, 1981; Pérez-Pereira, 1991, and others.). Morphophonological features, other phonological properties of nouns and noun semantics have all been argued to function as potential cues to class (see e.g., Corbett, 1991; Kihm, n.d.; Kramer, 2015; Mathieu et al., 2018; Ralli, 2002, amongst others). However, establishing precisely what features productively determine the agreement a noun takes in a given language is not trivial, particularly given potentially high levels of exceptions (see, e.g., Amidu, 1997; Castagneto, 2017; Corbett, 1991; Kanampiu et al., To appear; Rice, 2006), and the possibility of a default class into which novel nouns can be classified regardless of form or meaning (e.g., see Dingemanse, 2006; Richardson, 1967).

This leaves us with fundamental theoretical and empirical questions, like how to determine which cues are productive for class assignment amidst such exceptions, and whether speakers of a language use these cues in a consistent way or not. Traditional approaches to Bantu nominal classes, for instance, have heavily borrowed from the so-called Proto-Bantu schema, an arguably semantically-centred system that characterises the noun classes in Bantu based on semantic features like Human or Animate and more abstract ones like Paired-body parts, Liquid masses, Cohesive mass and so on (see, for instance, Table 2). While some of these features may be relevant for some classes (e.g., human or animates for class 1/2), semantic regularity in most of the classes is imperfect at best – typically, there are either nouns that cut across diverse semantic domains or some that share features with nouns in different classes, e.g., Diminutives, Augmentatives and Inanimates in Table 2.

Most importantly, the semantically-based characterizations of Bantu are often not founded on robust analysis of synchronic empirical data. Rather, in many cases, these are descriptive approaches either based on (non-systematically picked) examples, or on native-speaker intuitions alone (see, e.g., Msaka, 2019; Worsley, 1954, for similar views and other examples). Recently, quantitative corpus-based approaches have contributed to a better understanding of cues to gender in individual Bantu languages. For example, corpus studies on Zulu and Sesotho have found that nouns with certain semantic features are indeed more frequent in some classes than others (see, e.g., Demuth, 2000; Ngcobo, 2010, and others). This suggests these features may be productively used by speakers of these languages as cues to class. Studies on Kîîtharaka have used quantitative data, including formal measures of morphological productivity (specifically, the Tolerance Principle, Yang, 2016) to show that semantic features appear to be productive for only a few classes (e.g., class 1/2) (e.g., Kanampiu et al., To appear) (see Section 2 for more details on the Tolerance Principle). By contrast, (morpho)phonology appears to be the most relevant cue for the remaining classes in both Kîîtharaka, and other Bantu languages Demuth (see e.g., 2000, for Sesotho); Msaka (2019, for Chichewa); Kanampiu et al. (To appear, for Kîîtharaka). While corpus data helps make empirically informed predictions about which rules may (substantially) motivate the classification of nouns, an important question that remains to be answered is whether this reflects speakers’ own language knowledge. In other words, it leaves open whether observation of human behaviour would align with corpus data.

A complementary approach is to use behavioral experiments with native speakers to assess production of gender agreement based on targeted semantic or morphophonological features. If speakers indeed make use of such information – i.e., the form of the noun or meaning – as cues to noun class, systematic patterns of behaviour in how they respond to given stimuli should be observed. Adults and even children have been found to use both noun internal and syntactic distributional information (i.e., agreeing dependents) in making decisions about grammatical categories. For example, participants can predict the likely category or class of the word, given evidence about its syntactic context or agreeing dependents (in both natural and artificial language experiments: see e.g., Gagliardi & Lidz, 2014; Redington et al., 1998; Reeder et al., 2013). As mentioned above, in the absence of syntactic distributional information, speakers (both children and adults) can and do predict the noun class or gender of novel nouns from semantic or morphophonological information (again, in both natural and artificial language experiments, see, e.g., Frigo & McDonald, 1998; Gagliardi & Lidz, 2014; Karmiloff-Smith, 1981). Additionally, evidence suggests that different sources of information may not be used in the same way. Several experimental studies have found that, unlike adults, children acquiring nominal classes have a preference for phonological cues, even in the presence of more statistically reliable semantic cues (see, e.g., Bates et al., 1995; Culbertson et al., 2019; Gagliardi & Lidz, 2014; Karmiloff-Smith, 1981; Pérez-Pereira, 1991).

In Bantu, similar studies have found that children at certain ages have access to phonological restrictions characterizing the use of nominal prefixes (e.g., Demuth, 1992; Demuth & Ellis, 2010), suggesting that phonology is an early-available cue (see Demuth, 2000). For instance, in investigating acquisition of noun class prefixes in Sesotho, Demuth & Ellis (2010) found that children’s early use (and omission) of noun class prefixes was determined by phonological features of nouns.

More recently, Lawyer et al. (2024) used behavioral experiments to investigate whether speakers of Kinyarwanda are sensitive to semantics or morphophonology when classifying nouns. The study used a triadic method (following Burton & Kirk, 1976) in which participants are shown triads of three nouns and asked to pick the odd one out. In each triad, two nouns shared either a semantic feature, a morphophonological feature, or both, with the remaining noun being different from the others. For instance, take the triad iki-bwana ‘puppy’, in-gurube ‘pig’ and umu-kindo ‘palm tree’. Here, iki-bwana ‘puppy’ and in-gurube ‘pig’ share similar semantic features (i.e., they are animals and, specifically, mammals), while umu-kindo ‘palm tree’ is a plant. Likewise, in the triad umu-kubuzo ‘broom’, umu-shushwe ‘rat’, and in-dimu ‘lemon’, all bear different semantic properties, but the first two share a prefix. The assumption is that speakers should be able to pick out either or both of these differences, if indeed they pay attention to such features when making noun class decisions. The authors find a significant preference for semantics overall. In other words, most participants were successful at picking out semantically distinct words in triads, and used this kind of information even in trials where morphophonology could also be used.

Beyond the few studies mentioned above, there is relatively little work exploring how children and adult speakers of Bantu languages use semantic information to classify nouns, and no systematic experimental work testing the use of morphophonological cues in adult speakers. Here, we conduct two experiments investigating how adult speakers of Kîîtharaka make use of semantic and morphophonological features to predict the agreement class of novel nouns. Before reporting these experiments, we first briefly describe some key features of the noun class system of Kîîtharaka.

2. Previous work on Kîîtharaka noun classes

Kîîtharaka is an Eastern Central Kenya Bantu language spoken mainly by Atharaka people in Tharaka-Nithi County.³ The language is mutually intelligible with other closely related languages in the Meru cluster of Bantu languages, though there are varying degrees of lexical and mophophonological differences (see, e.g., Kanana, 2011). According to descriptive accounts, the language has 17 noun classes (defined using the traditional class pairs, based on nominal prefixes), often argued to be based on abstract semantic features (as in BTL, 1993; waMberia, 1993). However, in a more recent analysis, Kanampiu et al. (To appear) showed that, as in many other Bantu languages, nominal prefixes do not perfectly predict agreement classes. In the terminology of Corbett (1991), this system exhibits crossed relations between form class and agreement class. Following Güldemann & Fiedler (2019), Kanampiu et al. (To appear) establish 8 form class pairs, or deriflection classes (i.e., groups of nouns sharing morphophonological features in the singular and plural) and 8 agreement class pairs, or genders (i.e., groups of nouns triggering the same singular/plural agreement).⁴ In a recent study, Kanampiu et al. (To appear) applied the Tolerance Principle – a theory of morphological productivity (Yang, 2016, MS) – to a corpus of Kîîtharaka nouns to evaluate the productivity of selected semantic and morphophonological cues predicted to motivate nominal classification in Kîîtharaka. The Tolerance Principle is a theory of how and when learners form generalizations. It assumes that learners can store both productive rules and also a list of exceptional forms not determined by productive rules. To evaluate a potential rule, a learner considers whether the items in the lexicon that follow that rule meet a productivity threshold, taking into account the potential number of forms that would need to be listed as exceptional. The threshold, θN, is calculated using the formula N/lnN, where (N) is the number of items that follow the rule. A rule is productive if the number of exceptions (e) does not exceed θN. While the Tolerance Principle is primarily intended as a model of learning, it has also been applied to adult corpora to make predictions about productive rules (see, e.g., Björnsdóttir, 2023; Kodner, 2020; Yang, 2016, and others).

Kanampiu et al. (To appear) found that of twenty-three tested semantic features, only seven of them – Human, Trees, Augmentative, Pejorative, Manner, Diminutive and Infinitive – were productive. On the other hand, all morphophonological features tested were found to be productive; with the exception of the nasal prefix n-, all other singular and plural prefixes were productive, according to the Tolerance Principle. This study, therefore, suggests that (i) despite the existence of exceptions, there are specific semantic and morphophological cues which should, in principle, be productive for speakers; however, (ii) morphophonology is likely to be a more robust cue to the agreement class (or gender) a noun belongs to than semantics.

It is worth noting that evaluative semantic features – Diminutive, Augmentative, and Pejorative – can function derivationally in Kîîtharaka. In other words, while some nouns are always in the same evaluative gender class (due to an inherent property, like their size or social stigma), it is also possible to use the evaluative prefixes and corresponding agreement on nouns that are typically in other classes. Such derivation usage is in order to express or highlight one of these evaluative features (for additional discussion of these classes in Bantu and other languages, see Di Garbo, 2013). We will return to this below. However, the key question we address in this article is whether the predictions of productivity reported in Kanampiu et al. (To appear), in fact, align with what speakers are actually sensitive to when making noun class decisions. In other words, we ask whether, when shown a novel object with a particular semantic feature, or a novel noun with a particular prefix, speakers provide agreeing forms in a systematic way, and in line with the predictions of Kanampiu et al. (To appear). If they produce systematic responses for a given feature, then we assume this feature is productive (see 3.4 for details on how we establish what counts as systematic). We, therefore, conducted a series of psycholinguistic experiments to test whether speakers are sensitive to the semantic and morphophonological cues investigated in this previous research.

We report two experiments with native Kîîtharaka speakers. These experiments follow previous research which makes use of artificial lexical items to test knowledge of nominal classes in adults and children (e.g., Bates et al., 1995; Berko, 1958; Frigo & McDonald, 1998; Gagliardi & Lidz, 2014; Karmiloff-Smith, 1981; Pérez-Pereira, 1991).⁵ Our experiments sought to investigate whether adult speakers of Kîîtharaka are sensitive to particular kinds of cues – i.e., the morphophonological features of the noun, and the meaning of the noun – when producing syntactic agreement on adnominal modifiers. As we set out in Section 1, we take agreement as indicative of gender assignment. If speakers consistently use a particular semantic or morphophonological feature when providing nominal agreement, then novel nouns with that feature should be assigned the same agreement, hence, belonging to the same agreement class. Although assignment to a single agreement class does not necessarily indicate the gender of a noun (i.e., gender is a consistent pairing of agreement classes), we, nevertheless, make that simplifying assumption here.⁶ On the other hand, if speakers do not consistently use a particular feature when classifying nouns, no such regularity should be observed, or nouns will be assigned to a default gender. In the latter case, this is likely to be gender E (i.e., class 9/10 agreement). This is the agreement class associated with most loan nouns denoting non-humans, more particularly, those whose noun-initial (morpho)phonological make-up does not correspond to a possible prefix in Kîîtharaka. This kind of behavior is also found in many other Bantu languages, and we expect the same here (see, e.g., Demuth, 2000; Demuth & Weschler, 2012; Diercks, 2012; Rose & Demuth, 2006).

3. Experiment 1: Semantic cues

Experiment 1 tested whether Kîîtharaka speakers are sensitive to semantic cues depicted in images corresponding to novel nouns. A total of nine semantic features were investigated, following previous research in Kîîtharaka (Kanampiu et al., To appear) and related Bantu languages more generally. The set of features is shown in Table 3. These features target 6 genders in Kîîtharaka, corresponding to seven different singular agreement classes: 1, 3, 5, 7, 11, 12.⁷ As noted in Section 2, these particular semantic features were investigated in a corpus of Kîîtharaka (see Kanampiu et al., To appear), and only a subset of them were predicted to be productive based on the Tolerance Principle: the evaluative features Diminutive, Augmentative, and Pejorative; and Human. Here we test whether these predictions are, in fact, borne out for Kîîtharaka speakers tasked with producing agreeing modifiers of novel nouns. Our experimental protocol received ethical approval from the Linguistics and English Language departmental Ethics Committee.

3.2 Materials

We used novel nouns paired with novel images, in a task similar to the classic wug test paradigm (Berko, 1958). There were 27 images total, three for each of the nine target semantic features.⁸

Several notes are in order about the images and the features they test.

First, it is worth noting here that some of the features we test can be interpreted in more than one way. For example, something Narrow could also be Extended (see, e.g., Figure 1). Similarly, a given picture necessarily invokes more than one semantic feature. We attempted to increase the likelihood that specific images were associated with specific target features by piloting the stimuli images (for further discussion of this issue, see 3.5).

Figure 1: An example trial depicting the Narrow feature. Participants were required to type the completed phrase, and then press îthi mbere ‘proceed’ to move to the next trial.

Second, images depicting evaluative features included a standard juxtaposed with an altered version of the same image (see, e.g., Sagna, 2012, for a similar manipulation of the objects to depict size). A red arrow pointed to the altered version corresponding to the novel noun label, as shown in Figure 2. For example, to depict Augmentative, two images of a novel object were next to each other, one larger than the other, and an arrow pointed to the larger one. To depict Diminutive, two novel objects of different sizes appeared, as in the Augmentative, but in this case, the arrow pointed at the smaller image. To depict Pejorative, a novel object was shown next to a malformed or disfigured version of the same type of object, with an arrow pointing at the latter. As mentioned above, evaluative features in Kîîtharaka are productive for nouns that have those features inherently, but can also be used derivationally. We opted to depict evaluative features in a way that is more similar to derivational usage (i.e., the evaluative meaning was depicted in relation to the neutral meaning), due to the difficulty of ensuring that novel nouns in isolation would convey the desired properties. For example, inherent size is difficult to convey without additional context, and pejorative meanings are, in fact, often socially determined, or based on familiarity with normative concepts.⁹

Figure 2: Example test trials depicting evaluative features for (a) Augmentative, (b) Diminutive, and (c) Pejorative.

For the feature Human, we used novel (unfamiliar) human professions. Based on piloting, in these trials we included a brief description of the profession, in order to make clear to participants that this was what the novel noun corresponded to (rather than, e.g., a particular feature of the dress or costume). An example human trial is shown in Figure 3 below. The complete set of novel object stimuli is provided in Appendix A (Figure 10).

Figure 3: An example test trial depicting human profession. The profession is indicated in red below the image as ‘one who plays a Scottish flute’.

The linguistic stimuli were two-word phrases including a novel noun stem and the Kîîtharaka numeral meaning ‘one’. Both appeared preceded by an underscore to indicate a missing prefix. As described below, participants’ task was to provide these prefixes, and thus indicate what class they took an item to be in. As noted above, we treat the agreement prefixes as the main indicators of class. Nominal prefixes are absent, because they would otherwise provide another potential cue to class (tested in Experiment 2). However, it is rare for a noun in Kîîtharaka not to have a nominal prefix; therefore, we asked participants to provide this in their response, as well. Importantly, without either prefix, the linguistic stimuli provide no cues to class. The only hint that speakers have available is the semantic cue provided by the corresponding image.¹⁰ The numeral ‘one’ was chosen to ensure that the respondents always gave singular prefixes, as number and noun class are confounded in Kîîtharaka (see Table 1).

There were 30 novel noun stems (see Table 4). All stems were bisyllabic nonce words that conformed to Kîîtharaka phonotactics. Stems were randomly paired with an image by-participant.

Table 4: A list of novel noun stems used as novel Kîîtharaka nouns in Experiments 1 and 2.

Novel noun stems
-tondi	-timû	-tiinki	-pori	-tago	-pûrû
-tebe	-thiiri	-kiithû	-pikû	-coki	-teerû
-pengi	-kithe	-kirû	-ciira	-kitho	-tuubo
-thindû	-tiirû	-ciibo	-thego	-theo	-kage
-kori	-tori	-taagi	-taki	-pegi	-coi

3.4 Results

Figure 4 shows the average proportion of responses that aligned with the expected singular agreement class number for each semantic feature tested, as outlined in Table 3. The dotted line indicates what we consider here to be chance-level performance, calculated based on the singular agreement class prefixes participants provided, which correspond to the expected classes for the 6 semantic features we tested. With 6 different expected classes, chance is 1/7 = 0.17.¹² Figure 4 suggests that some of the features tested may have been used productively, while others were not. Interestingly, it is also notable that a sizeable proportion of participants used class 7 prefixes (k(g)î- – k(g)î-) with augmentative trials, instead of the predicted class 5 (î- – rî-). Additionally, across all features, participants tended to produce a relatively high proportion of class 9 (n/Ø- – î-) responses (e.g., more than 40% for the Artefact, Extended, Fruit, Narrow and Wavy features). We come back to these patterns in 3.5, but see Figure 11 in Appendix B for an overview of this variation in participant responses.

Figure 4: Average proportion of responses aligned with the predicted class (see Table 3) for each semantic feature. Error bars show standard error on by-participant means.

To further analyse these data, we evaluated whether accuracy for each of these semantic features was significantly above chance. Since Accuracy was a binary feature (correct vs incorrect for the expected and unexpected responses, respectively), we used a binomial test, with the probability of success set to 1/7. We consider accuracy significantly above this level, according to the binomial test, to indicate a productive feature. The results, provided in Table 5, show significantly above chance accuracy for the features Human (p < 0.0001), Fruit (p < 0.0251), Pejorative (p < 0.0001) and Diminutive (p < 0.0001). All the other features, including Augmentative, had accuracy rates that were not significantly above chance (p > 0.05).

Table 5: Results of binomial tests of semantic features in Experiment 1. Upper and lower bounds for 95% confidence intervals around the estimate are also given.

Feature	Estimate	P-Value	Lower bound	Upper bound
Human	0.7778	<0.0001	0.6717	0.8627
Extended	0.1358	0.7367	0.0698	0.23
Augmentative	0.1975	0.1545	0.1173	0.3009
Fruit	0.2346	0.0251	0.1475	0.3418
Artefact	0.1728	0.4272	0.0978	0.273
Pejorative	0.4074	<0.0001	0.2995	0.5223
Narrow	0.1481	0.8737	0.079	0.2445
Wavy	0.1235	0.7510	0.0608	0.2153
Diminutive	0.6049	<0.0001	0.4901	0.7119

Even among features that were productive according to the binomial test, Figure 4 suggests that participants were more sensitive to some of the features than others. Most obviously, the feature Human elicited numerically higher response rates compared to other features. To verify this in a way that also accounts for by-participant variation, we fitted a logistic mixed-effects regression model with Accuracy as the dependent variable and Feature as the predictor variable, along with by-subject and by-item random intercepts, using the glmer function of the lme4 package (Bates, 2010) in R (R Core Team, 2023). The model uses treatment (dummy) contrast coding, with Human as the reference level. Therefore, this model compares mean accuracy on Human feature trials to mean accuracy for each other feature tested. The results of the model are shown in Table 6.

Table 6: Results of a model comparing accuracy on the Human feature (reference level) to accuracy on each other feature tested.

Feature	Estimate	SE	Z-value	P-value
(Intercept)	1.3201	0.2928	4.508	<0.0001***
Extended	–3.2479	0.4351	–7.465	<0.0001***
Augmentative	–2.7832	0.4021	–6.922	<0.0001***
Fruit	–2.5752	0.3931	–6.551	<0.0001***
Artefact	–2.9655	0.4142	–7.160	<0.0001***
Pejorative	–1.7078	0.3636	–4.696	<0.0001***
Narrow	–3.1707	0.4328	–7.327	<0.0001***
Wavy	–3.3825	0.4509	–7.502	<0.0001***
Diminutive	–0.8655	0.3621	–2.390	0.0168*

The results show a significant negative effect of all features in relation to Human. This confirms that the feature Human is indeed more productive than the other features tested. We discuss the implications of these results below.

Additionally, it is possible that frequency plays a role in determining participants’ sensitivity to these cues. If learners more often come across nouns representing certain classes than others, they may be more likely to recruit novel nouns into those classes (see e.g., Gagliardi & Lidz, 2014, for a similar method of analysis). We, therefore, established class weights as baseline probabilities for Kîîtharaka, based on the corpus used by Kanampiu et al. (To appear), publicly available at: https://rb.gy/ows7r1. The distribution of nouns based on type frequency in this corpus is shown in Table 7.¹³

Table 7: Number of nouns in each target gender, based on the Kîîtharaka corpus used by (Kanampiu et al., To appear).

Gender	Agreement Class	Noun types
A	1/2	196
B	3/4	208
C	5/6	158
D	7/8	237
E	9/10	379
F	11/10	108
G	12/13	77
Total	–	1363

To establish whether the probability of a correct response for each feature was higher than would be expected, on the type frequency of the target class, we conducted additional binomial tests. In this set of tests, the probability of success was calculated using the number of nouns in the target class divided by the total number of nouns for the classes tested. The results are as shown in Table 8.

Table 8: Results of binomial tests of semantic features in Experiment 1, with the baseline class probabilities from the corpus as the probability of success. Upper and lower bounds for 95% confidence intervals around the estimate are also given.

Feature	Estimate	P-value	Lower bound	Upper bound
Human	0.7778	<0.0001	0.6717	0.8627
Extended	0.1358	0.7592	0.0698	0.23
Augmentative	0.642	<0.0001	0.1173	0.3009
Fruit	0.2346	0.0025	0.1475	0.3418
Artefact	0.1728	1.0000	0.0978	0.273
Pejorative	0.6543	<0.0001	0.2995	0.5223
Narrow	0.1481	0.0355	0.079	0.2445
Wavy	0.1235	0.1466	0.0608	0.2153
Diminutive	0.6049	<0.0001	0.4901	0.7119

These results largely align with the previous ones: the features Human, Diminutive, Augmentative, Pejorative, Fruit were found to be significantly above chance. However, here, the feature Narrow is also significantly above chance. We return to this difference in 3.5 below.

We also conducted one additional analysis aimed at establishing the level of divergence between the distribution of nouns to respective target classes in the experiment, where semantic features were manipulated, and across all nouns in the corpus data sets, regardless of their features. To do this, we calculated Jensen-Shannon (JS) divergence between these distributions, following Gagliardi & Lidz (2014).¹⁴ This test provides an index between 0 and 1, where 0 is no divergence, and 1 is total divergence. In our case, this index provides additional information about the relationship between participants’ responses to specific features in the experiment and the baseline distribution of nouns across classes, as evident from the corpus. In principle, smaller divergence scores would suggest that participants’ responses are not strongly influenced by a given feature, i.e., that their responses are distributed simply according to the type frequency of classes. Higher divergence scores would, instead, suggest that participants are sensitive to this feature, i.e., their responses diverge from what would be expected, on type frequency. The JS scores for all features are shown below (see Table 9).

Table 9: Jensen-Shannon divergence scores reflecting the difference between the distribution of nouns across target classes for a given feature in the experiment, and the distribution of nouns across all classes in the corpus (as a baseline distribution). Higher scores suggest more divergence from the baseline, and thus more sensitivity to the feature in determining the class of participants’ response.

Feature	Agreement class	JS score
Human	1	0.1835
Extended	3	0.0588
Augmentative	5	0.2559
Fruit	5	0.0772
Artefact	7	0.0568
Pejorative	7	0.1469
Narrow	11	0.0632
Wavy	11	0.0697
Diminutive	12	0.2364

Notably, the features Human, Diminutive, Augmentative and Pejorative exhibit the highest divergence. This aligns with the previous analyses reported above, and suggests that these features are the ones which are used most robustly by participants when determining the agreement class of novel nouns.

4. Experiment 2: Morphophonological cues

In contrast to semantic features, Kanampiu et al. (To appear) predict that all nominal prefixes are productive morphophonological cues to class in Kîîtharaka. In Experiment 2, we test these features. Specifically, we test whether Kîîtharaka speakers use nominal prefixes on a novel noun stem (i.e., a beginning segment or syllable) to predict the form of an agreeing dependent, in the absence of semantic information. We tested both singular and plural prefixes, repeated in Table 11 below. As in Experiment 1, we use the agreement prefix provided by participants as an indicator of the class participants assign a novel noun to. If participants use noun beginnings in this way, we predict that they will consistently assign the predicted agreement pattern to nominal dependents. Our experimental protocol received ethical approval from the Linguistics and English Language departmental Ethics Committee.

4.2 Materials

As in Experiment 1, the linguistic stimuli in Experiment 2 were two-word phrases including a novel noun and the Kîîtharaka numeral ‘one’ with its agreement prefix missing. The set of novel noun stems was the same as in Experiment 1. Critically, here we manipulated the Kîîtharaka prefixes affixed to these nonce stems. Table 11 shows the tested prefixes alongside their agreement classes. Participants’ task was, again, to provide the missing agreement prefixes, and thus indicate what class they took an item to be in. No visual stimuli were used in this experiment, and, therefore, no semantic cues to the meaning of the novel nouns were provided.

Novel stems were randomly paired with prefixes, using nine different lists. In each list, a given stem was paired with a different one of the nine singular or plural prefixes. Participants were randomly assigned to one of the nine lists. For example, one participant would see the prefix mû- with the nonce stem kithe, and another would see the same prefix with the stem tondi.¹⁷ It is also important to note that, as can be seen in Table 11, the class 1 and 3 prefixes are orthographically the same. This means that in the absence of semantic cues, as in this experiment, we expect the participants to classify mû- prefixed nouns into one of the two classes, and we considered either to be correct.

4.2.1 Procedure

Experiment 2 was built and administered using Qualtrics. Participants received a link to the online experiment by email and accessed it using their mobile phones. Participants were gathered in a hall in groups of 10 and and given initial instructions verbally before accessing the experiment. Each participant was seated at a desk, spaced apart from other participants. Participants were instructed that they were going to learn new Kîîtharaka words. The experiment had three experimental blocks. In the first singular-singular block, we tested whether speakers use a singular Kîîtharaka prefix to predict singular agreement. In the second plural-plural block, we tested whether participants use a plural prefix to predict plural agreement. In the third singular-plural block, we tested whether participants use a singular prefix to predict plural agreement. Each block had a training and a testing phase.

In the training phase preceding each block of testing, participants saw an actual Kîîtharaka noun (including its prefix). Under the noun, they saw a carrier sentence starting with ‘I have seen a ____’.¹⁸ Participants were required to complete this sentence using the noun presented above and a numeral. In the singular-singular block, participants saw a familiar Kîîtharaka noun with a singular prefix, and the numeral was ‘one’ in each trial. Similarly, for each trial in the test phase, they saw a novel noun with a singular prefix and the numeral ‘one’ (see Figure 5a).

Figure 5: Sample trial in Experiment 2, showing (a) training trials with native class 5 prefixed noun îkombe and (b) test trials with novel class 9 prefixed noun nkiirû.

In the plural-plural training block, participants saw familiar nouns with a plural prefix, and the numeral was ‘three’ (i.e., __thatû, as in Figure 5b).¹⁹ Likewise, for each trial in the test phase, they saw a novel noun with a plural prefix, and the numeral was ‘three’.²⁰

In the singular-plural block, in each training trial participants saw a familiar noun with a singular prefix, but in this case, the numeral was ‘three’. They, therefore, had to convert the noun to plural before using it to complete the sentence. Likewise, for each test trial, they saw a novel noun with a singular prefix and the numeral was ‘three’. They had to convert the noun into plural before filling in the missing plural agreement prefix on the numeral.

Notably, the position of the the object noun phrase and the missing agreement prefix on the numeral were indicated using underscores (e.g., Nkwona ____ _mwe ‘I have seen __one ____’). As in Experiment 1, there was a lag time (5 seconds) between when the noun appeared, and when the sentence and text box appeared in all trials. The order in which stimuli within a block appeared was randomized for each participant. Feedback was provided after each training trial, but there was no feedback in the test phase.

4.2.2 Results

Figure 6 and Figure 7 show participants’ accuracy at producing the expected agreement prefixes in the singular-singular and plural-plural blocks, respectively. In both these blocks, accuracy is, overall, very high. In the singular-singular block, where participants saw a novel noun with a singular Kîîtharaka prefix, they produced singular prefixes and respective agreements largely as predicted. Performance was perfect for the prefixes kî- (class 7), n- (class 9) and rû- (class 11). Recall that class 1 and 3 prefixes are homophonous. For this reason, the two classes were collapsed, and a response was treated as correct if it had class 1 or 3 agreement. Notably, however, an overwhelming majority of participants chose î- (class 3) agreement for the stimulus prefix mû-, and not û- (class 1). This suggests that participants assumed these novel nouns depicted objects rather than humans. We return to this point below. A similar pattern was also observed in the plural-plural block, where participants saw a novel noun with a plural Kîîtharaka prefix and had to produce plural agreement. Here, accuracy was perfect or near perfect for the prefixes ma- (class 6), mi (class 4), n- (class 10), and tû- (class 13).

Figure 6: Average proportion of expected (correct) class agreement responses for Experiment 2 in the singular-singular block. Note: (i) There are no error bars, because only one prefix per class per participant was tested; (ii) Chance level is higher for the prefix mû- because it had, higher probability of success – participants could provide class 1/2 or 3/4 agreement (see footnote 16).

Figure 7: Average proportion of expected (correct) class agreement responses for Experiment 2 in the plural-plural block. Note: (i) There are no error bars because only one prefix per class per participant was tested; (ii) Chance level is higher for the prefix ma- and n-, because it had higher probability of success – participants could provide class 5/6 or 14/6 agreement for ma-, and 9/10 or 11/10 for n-.

Figure 8 shows participants’ accuracy at producing the expected agreement prefixes in the singular-plural block. Recall that here, participants saw a singular noun and were required to convert the novel noun to plural and produce a plural agreement prefix. Accuracy rates are markedly lower for this block. In most trials where the noun had a class 1/3 prefix mû-, participants produced class 4 (mî/î) responses. In most trials where the noun had a nasal prefix, participants provided class 10 plural agreement, as expected. For several other stimulus prefixes, responses were nearly evenly divided between the expected response and class 6 agreement (e.g., the null prefix, and class 7 prefix k(g)î-; see Figure 13, Figure 14 and Figure 15 in Appendix B for an overview of incorrect/unexpected responses that participants gave in each of the three blocks). We return to the use of class 6 plural agreement below.

Figure 8: Average proportion of expected (correct) class agreement responses for Experiment 2 in the singular-plural block. Note: (i) There are no error bars, because only one prefix per class per participant was tested; (ii) Chance level is higher for the prefix mû- and û- because they had higher probability of success – participants could produce class 2 (a-ba) or 4 (mî-î) agreement for the former, and class 14 (û-bû), 6 (ma-ma) or 10 (n-i) for the latter.

As in Experiment 1, we evaluated whether accuracy for each of these morphophonological features (in this case, a stimulus prefix) was significantly above chance. Again, Accuracy was a binary feature (correct vs incorrect for the expected and unexpected responses, respectively), so we used a binomial test, with the probability of success corresponding to the probability of a correct response for the given prefix. As in Experiment 1, we consider features whose accuracy is significantly above chance level to be productive. The results, as shown in Table 12, Table 13 and Table 14, indicate that accuracy for all the morphophonological features is significantly above chance level across the three blocks, hence, productive.

Table 12: Results of binomial tests for morphophonological features in the singular–singular block of Experiment 2. Lower and upper bounds for 95% confidence intervals around the estimate are also given.

Stimuli prefix	Estimate	P-value	Lower bound	Upper bound
mû-	0.8833	<0.0001	0.7743	0.9518
î-	0.8	<0.0001	0.6143	0.9229
k(g)î-	1	<0.0001	0.8843	1
n-	1	<0.0001	0.8843	1
ø-	0.9	<0.0001	0.7347	0.9789
rû-	1	<0.0001	0.8843	1
k(g)a-	0.9667	<0.0001	0.8278	0.9992
û-	0.9	<0.0001	0.7347	0.9789

Table 13: Results of binomial tests for morphophonological features in the plural–plural block of Experiment 2. Lower and upper bounds for 95% confidence intervals around the estimate are also given.

Stimuli prefix	Estimate	P-value	Lower bound	Upper bound
a-	0.7667	<0.0001	0.5772	0.9007
mî-	0.9697	<0.0001	0.8424	0.9992
ma-	0.9825	<0.0001	0.9061	0.9996
i-	0.8667	<0.0001	0.6928	0.9624
n-	0.9833	<0.0001	0.9106	0.9996
ø-	0.7667	<0.0001	0.5772	0.9007
tû-	0.9667	<0.0001	0.8278	0.9992

Table 14: Results of binomial tests for morphophonological features in the singular-plural block of Experiment 2. Lower and upper bounds for 95% confidence intervals around the estimate are also given.

Stimuli prefix	Estimate	P-value	Lower bound	Upper bound
mû-	0.7667	<0.0001	0.6396	0.8662
î-	0.4333	<0.0001	0.2546	0.6257
k(g)î-	0.3	0.0010	0.1473	0.494
n-	0.5667	<0.0001	0.3743	0.7454
ø-	0.4	<0.0001	0.2266	0.594
rû-	0.4	<0.0001	0.2266	0.594
k(g)a-	0.3	0.0010	0.1473	0.494
û-	0.6	0.0002	0.406	0.7734

To test whether the participants were, overall, better in one block than the other, we fitted a logistic mixed-effects regression model with Accuracy as the dependent variable and Block as the predictor variable, along with linguistic stimuli (ling_stimuli) as random intercepts, using the glmer function of the lme4 package.²¹ The model uses treatment (dummy) contrast coding, with singular-singular (block 1) as the reference level. Therefore, this model compares accuracy in block 1 trials to accuracy for each of the other two blocks.²² The results of the model are shown in Table 15.

Table 15: Results of a model comparing accuracy in singular-singular block 1 (reference level) to accuracy in plural-plural block 2 and singular-plural block 3 of Experiment 2.

Block	Estimate	SE	Z-value	P-value
(Intercept)	2.9145	0.3112	9.365	<0.0001 ***
pl-pl	–0.1709	0.3850	–0.444	0.657
Sing-pl	–2.9873	0.3700	–8.074	<0.0001 ***

The results show no significant difference between singular-singular and plural-plural blocks. However, we did observe a significant negative effect of singular-plural block 3 (β = –2.9873, SE = 0.3700, z = –8.074, p < 0.0001). This confirms that indeed, there is a significant difference between the participants’ sensitivity to singular prefixes as a cue to singular class agreement and the same as a cue to plural class agreement. In the next section, we discuss the implications of these observations.

Since response accuracy in the singular-plural block was lower than in the other blocks, we also conducted an additional set of binomial tests, taking into account type frequency of classes (again, based on Kanampiu et al., To appear) for this block. The distribution of nouns across the respective target classes is shown in Table 16.²³

Table 16: Number of nouns in each plural target gender, based on Kîîtharaka corpus used by Kanampiu et al. (To appear).

Gender	prefix	Agreement Class	Noun types
A	a-	2	188
B	mî-	4	208
C	ma-	6	143
D	i-	8	237
E	n-	10	377
E	ø	10	69
F	n-	10	107
G	tû-	13	77
Total	–	–	1406

The results were largely in line with the previous test, indicating significantly above-chance accuracy levels for most of the features as shown in Table 17.

Table 17: Results of binomial tests of morphophonological features for singular-plural block 3 in Experiment 2, with the baseline class probabilities from the corpus as the probability of success. Upper and lower bounds for 95% confidence intervals around the estimate are also given. The prefix mû- appears in duplicate, to reflect the responses for the alternative two classes (2 and 4), both of which were correct.

Feature	Estimate	P-value	Lower bound	Upper bound
mû-	0.05	0.0572	0.0104	0.1392
mû-	0.7167	<0.0001	0.5856	0.8255
î-	0.4333	<0.0001	0.2546	0.6257
k(g)î-		0.0817	0.1473	0.494
n-	0.5667	<0.001	0.3743	0.7454
ø	0.4	<0.0001	0.2266	0.594
rû-	0.4	<0.0001	0.2266	0.594
k(g)a-	0.3	<0.0001	0.1473	0.494

Finally, the Jensen-Shannon (JS) divergence scores for the singular-plural block 3 are shown in Table 18 below). These are all relatively high (e.g., they are similar to the JS scores for the most robustly productive semantic features in Experiment 1).

Table 18: Jensen-Shannon divergence scores reflecting the difference between the distribution of nouns across various plural, in target classes (singular-plural block) in the experiment and the number of plural nouns in the corpus as baseline probability.

Feature	Agreement class	JS score
mû-	2, 4	0.2207
î-	6	0.1687
k(g)î-	8	0.1600
n-	10	0.1428
ø-	10	0.1438
rû-	10	0.1444
k(g)a-	13	0.1468

5. Comparison between semantic and morphophonological experiments

Based on the results of Experiments 1 and 2, our findings suggest a sharp difference in accuracy based on the semantic features compared to the morphophonological features tested. This is summarized in Figure 9, which shows average accuracy across the productive semantic features and block 1 of the morphophonological experiment.

Figure 9: Average accuracy across the productive semantic and morphophonological features in Experiments 1 and 2, respectively. Chance levels vary by experiment (0.16 and 0.11, respectively).

Considering that semantics and morphophonology are the main factors which interact to determine class and or gender more generally (see e.g., Aikhenvald, 2012; Corbett, 1991; Katamba & Stonham, 2006, and others), this comparison suggests that in Kîîtharaka, morphophonology may generally be the preferred cue type. However, this claim can not be confirmed here since these two cue types were tested separately. We leave this to future research.

To compare these feature types statistically, we collapsed data from the productive features of Experiment 1 (Human, Augmentative, Pejorative, and Diminutive), and from all morphophonological features in block 1.²⁵ We then fit a logistic mixed-effects regression model with accuracy as a binary dependent variable and experiment as the predictor variable, along with by-participant and by-item random intercepts. The predictor variable was treatment (dummy) coded, with Experiment 1 as the reference level. The model reveals a statistically significant difference between Experiment 1 and 2 (β = 2.8190, SE = 0.3663, z = 7.696, p < 0.0001). This confirms that participants are significantly more sensitive to morphophonology than semantics overall, even when considering only those semantic features which were used productively.

6. General discussion

In Bantu languages, both nominal prefixes and the semantics of the noun (both inherent, e.g., human, and evaluative, e.g., diminutive) are considered to determine the agreement class and therefore gender of the noun. However, recent quantitative studies in Bantu languages have shown that both of these cues may have varying degrees of relevance in particular Bantu languages (e.g., Kanampiu et al., To appear; Lawyer et al., 2024; Msaka, 2019). This study aimed to provide behavioral evidence to establish whether speakers of Kîîtharaka, an Eastern-central Bantu language, make use of both the nominal prefixes and specific semantic features of the noun in classifying novel nouns. The study involved two experiments: Experiment 1 tested whether speakers were sensitive to a set of inherent and evaluative semantic features. Experiment 2 tested for sensitivity to nominal prefixes. In both of these experiments, participants were presented with novel stimuli and required to produce class agreement on a Kîîtharaka numeral. In Experiment 1, they saw unfamiliar images highlighting one particular semantic feature accompanied by a prefix-less novel Kîîtharaka stem, along with its prefix-less numeral ‘one’, and were required to produce the missing prefixes. In Experiment 2, they saw a prefixed novel Kîîtharaka noun, along with its prefix-less numeral, in a carrier sentence and were required to produce the missing prefix.

The results show that some of the tested cues were consistently used to predict class agreement, while others were not. First, all nominal prefixes in Kîîtharaka were used productively by participants to predict class agreement on a nominal dependent in our experiment. This implies that in Kîîtharaka, the first segment of the noun, i.e., the nominal prefix, is a robust cue to the gender of a noun. Second, semantic features were, in general, less likely to be used productively to predict class agreement. Indeed, according to our findings, most features tested, such as Extended, Artefact, Narrow, etc., failed to elicit consistent agreement classes. The exceptions were Human, Fruit and the evaluative features Augmentative, Pejorative, and Diminutive, which were all found to be used productively by speakers (though to different degrees). As we argued in 3.5, the evaluative feature Augmentative is likely ambiguously expressed in Kîîtharaka, i.e., it is productively associated with two class prefixes – î- and k(g)î-. This feature was found to be robustly productive across analyses, once this pattern of usage was accounted for. Overall, then, while some semantic features are used as cues to class and gender, these features are not as robust as nominal prefixes in Kîîtharaka. Below, we discuss the implications of these results, along with a few unexpected findings.

The results of the two experiments reported here largely corroborate the findings of a corpus-based study by Kanampiu et al. (To appear), which used the Tolerance Principle to evaluate the productivity of the two types of cues in determining class and gender agreement in Kîîtharaka. While the Tolerance Principle (like other measures of productivity) can generate predictions, experimental results like these provide behavioural evidence that speakers of Kîîtharaka can and do use particular features productively. Our results suggest that many inherent semantic features are not represented by speakers as cues for classifying new nouns in the language. If such features were, at some point, relevant in the diachrony of the language, or the parent Proto-Bantu, most of these features have since lost that relevance. If this is true for Kîîtharaka, then it may be the case for many other present-day Bantu languages. Interestingly, comparing the results of Kîîtharaka to those of Kinyarwanda, as in Lawyer et al. (2024), where participants were found to be more sensitive to semantics, suggests the possibility that Bantu languages may exploit these types of cues differently.²⁶ Studying nominal classification systems of individual languages in an empirically-grounded way is likely to reveal interesting results.

While the alignment between the corpus data and our results suggests the differences between semantics and morphophonology in Kîîtharaka are real, it is also worth revisiting a methodological consideration. It could also be the case that some of the semantic features tested appeared less productive or unproductive in the experiment because of how the participants interpreted the image stimuli. Image stimuli necessarily fail to isolate individual semantic features, in contrast to morphophonological features. Participants may, therefore, have been influenced by those features. For example, as noted in 3.5, some Narrow images could also be perceived as Extended. Second, it could be that the stimuli themselves varied in how much they instantiated the target feature. For instance, some particular images might have been more successful in depicting target meanings than others for a given feature (see Figure 12). For example, the target feature Fruit, may have been more successful at eliciting this feature compared to other features depicted by (sometimes complex) artefacts. While we did our best to select the most appropriate images (e.g., by piloting), there is likely to always be an asymmetry when comparing these two classes of features. Therefore, the possibility remains that the productivity of semantic features in the experiment is under-estimated.

We also noted some unexpected patterns in Experiment 2 that are worth additional discussion. First, we saw that participants tended to associate nouns prefixed with an ambiguous 1/2 or 3/4 class prefix with gender B (class 3/4), rather than gender A (class 1/2). This is despite the fact that in the corpus analyses (see Kanampiu et al., To appear) this prefix is productive as a cue to both gender A and B. As mentioned above, this likely indicates that in the absence of semantics, participants assume that new words refer to objects, rather than humans.

It is also worth returning to the types of responses participants consistently provided when they did not provide the expected responses based on class. First, the majority of such responses were class 9/10 (gender E). This is what we predicted, since this gender is typically considered to be a default, across Bantu languages. However, we also saw minimal but consistent alternative assignment to other classes, mainly class 6 (gender C plural agreement class) and gender B (class 3/4). While further research is needed to understand whether these responses indeed reflect something non-accidental, we have suggested that these may also reflect alternative default classification. Loans or novel nouns that are featurally un/under-specified are typically admitted to one class – mostly class 9/10 in Bantu (see, e.g., Gunnink et al., 2015; Mous, 2019, and others). However, gender B is also a large, heterogenous class, and Kîîtharaka speakers may thus assume novel words are likely to fall into it, particularly when no semantic feature is present, or the semantic features are not productive enough to determine class. We have suggested that class 6 responses may have been used in Experiment 2, in particular, as a default plural class, since it is used to form collective plurals in Kîîtharaka. However, this, and the general trouble our participants experienced deriving plural forms from singulars in Experiment 2 deserves further study.

While we have found here that morphophonological features appear generally more productive than semantic features in Kîîtharaka, future research should also explore more explicitly their relative importance. For example, previous research has used experimental methods to show that when semantics and (morpho)phonology features are both present but provide conflicting cues, adult speakers tend to rely on semantics (e.g., Demuth & Ellis, 2010; Gagliardi & Lidz, 2014; Karmiloff-Smith, 1981; Pérez-Pereira, 1991). Whether this is the case for Kîîtharaka as well remains to be seen.

7. Conclusion

What determines productivity in noun class systems across languages is a longstanding question. While descriptive research suggests that both semantics and morphophonology are relevant, behavioral evidence of productivity in native speakers is missing for many languages. In this study, we set out to establish whether speakers of the Bantu language Kîîtharaka are sensitive to the semantics and the form of the noun (i.e., the nominal prefix) when providing agreement on the nominal dependents. We found speakers are robustly sensitive to a small subset of semantic features: Human, and to a lesser extent, Fruit, Narrow, and the evaluative ones Augmentative, Pejorative and Diminutive. By contrast, speakers were robustly sensitive to all morphophonological cues we tested. These experimental results provide psychological evidence for these cues, and suggest that if others may have been historically relevant (i.e., in Proto-Bantu), their relevance has waned for Kîîtharaka and potentially many other Bantu languages. Our study, therefore, shows that the behaviour of novel nouns can provide insights into productive morphophonological processes in a language, and thus highlights the importance of empirically grounded research in this area.

Appendix

A. Novel images used in the semantic experiment

Figure 10: Novel images used in Experiment 1 with the respective features they tested.

B Cumulative bargraphs showing errors made by participants in classification of nouns in both experiments

Figure 11: Cumulative responses in agreement classes by given stimuli for the semantic experiment.

Figure 12: Cumulative responses in agreement classes by given image stimuli for the semantic experiment.

Figure 13: Cumulative responses in agreement classes by given stimuli for morphophonological experiment (singular-singular block).

Figure 14: Cumulative responses in agreement classes by given stimuli for morphophonological experiment (plural-plural block).

Figure 15: Cumulative responses in agreement classes by given stimuli for morphophonological experiment (singular-plural block).

Data accessibility statement

All data and analyses are available online at: https://rb.gy/vcnhxs.

Ethics and consent

The ethics consent for this study was granted by the Linguistics and English Language departmental Ethics Committee (Ref. No 188-2122/4).

Acknowledgements

Many thanks to Annie Holtz for reading and providing feedback to the earlier versions of this paper.

Competing interests

The authors have no competing interests to declare.

Authors’ contribution

Patrick Kanampiu: conceptualization, methodology, investigation, data collection, writing – original draft, data curation, statistical analyses, review and editing.

Alexander Martin: conceptualization, methodology, investigation, writing – review and editing, project supervision.

Jennifer Culbertson: conceptualization, methodology, investigation, statistical analyses, writing – review and editing, project supervision.

ORCIDs

Patrick Kanampiu – 0000-0002-6625-8827

Alexander Martin – 0000-0003-1241-2435

Jennifer Culbertson – 0000-0002-1737-6296