3.1. Idiom reconstruction in Ngemba

The basic assumption underlying this test is that an idiom keeps its meaning if part of it is extracted, but loses it if what looks to be part of it is base-generated in a different position (see, for example, Scott Reference Scott2021, Georgi & Amaechi Reference Georgi and Amaechi2023, Yip & Ahenkorah Reference Yip and Ahenkorah2023 for similar applications of this movement diagnostic). Applying this test to clause-internal topicalization in Ngemba shows that the idiomatic meaning is lost if part of the idiom gets topicalized. Below, I show this with the idiom ŋwɔ́Ɂ shwò (literally ‘bend mouth’) which, in Ngemba, means ‘to be angry’. shwò, in (17b), can undergo topicalization, in which case the only possible reading is one that does not involve being angry, but rather the state of the mouth (bent).

Comparing the topicalization example in (17b) to focus (18) demonstrates that, unlike topicalization, focus involves movement. This is so because when shwò ‘mouth’ is focused, the idiomatic reading is maintained. Under a movement approach, the moved item can be reconstructed for interpretation (whether this involves interpretation of a silent copy is not important). Importantly, this is not an option for base-generation.

This diagnostic can be extended to DP-internal contexts. Ngemba allows idioms that must contain a possessive pronoun. An example is given in (19).

What is interesting about the idiom in (19) is that in the absence of the possessive pronoun, it is ‘ungrammatical’. In other words, it is ungrammatical to simply say lá ø-kә̀lám ‘cook kelam’ with the idiomatic reading. For the idiomatic interpretation to be possible, the possessive pronoun must be used. Crucially, the possessive pronouns in the two DPs (the subject DP n-dúm ø-à ‘my husband’ and the object DP ø-kә̀lám ø-ɔ́ ‘your kelam’) in (19) can be focused, hence appear in pre-nominal position. When this happens, the idiomatic reading is kept.Footnote ⁹

The fact that the idiomatic reading is maintained when the possessive pronoun surfaces in pre-nominal position provides empirical evidence that it is focus-moved into the focus position, just as is the case for clause-internal focus (18). A silent copy is interpreted in the movement site. I conclude that Poss-focus in Ngemba involves movement of Poss into the left periphery of the DP. There are reasons to think that this extends to the Medumba data. In the section that follows, I present suggestive evidence from tonal reflexes of movement.

3.2. Tonal and segmental reflexes of movement

Ngemba and Medumba instantiate tonal reflexes of clause-internal IS-driven movement (Keupdjio, Reference Keupdjio2020, Fongang, Reference Fongang2025b). In Medumba, as discussed at length in (Keupdjio, Reference Keupdjio2020), the verbal root kέ ‘choose’, for example, has a low tone in non-movement contexts (21).

In ex-situ object wh-questions (22a), focus (22b) and topicalization (22c), the (only) vowel of this verb surfaces with a falling tone.Footnote ¹⁰

Keupdjio (Reference Keupdjio2020) uses this and other diagnostics for movement (e.g. sensitivity to islands (23) and reconstruction effects (24)) to come to the conclusion that these changes reflect IS-driven movement in Medumba.

Ngemba verbs also behave this way and, this has been taken as a diagnostic for movement in the language as well (c.f., Fongang Reference Fongang2025b). In a simple declarative sentence such as (25), the verbal root zhó ‘see’ surfaces with a high tone.

In the object focus example in (26), this vowel surfaces with a falling tone.

What is striking about topicalization is that topic marking does not trigger this type of tonal change (27).

These facts, coupled with two other tests (e.g. the islands in topicalization (28) and focus (29)), demonstrate that topicalization in Ngemba involves base-generation of an XP in the topic position and binding of a resumptive pronoun. Focus, however, involves movement to the left periphery of the Ngemba clause (c.f., Fongang Reference Fongang2025b).

Of the many tests for IS-driven movement that have been proposed in the literature, tonal reflexes can straightforwardly be applied to the DP, under the premise that DPs parallel IPs/TPs in that they also have a CP-like left periphery that can, for example, host focalized DP-internal elements (see, among others, Aboh Reference Aboh2004). In other words, DP-internal material can be moved for focus, topicalization, etc. What is interesting to observe about Ngemba and Medumba is that there is a parallel between the effects of clause-internal movement discussed above and focalized possessive pronouns.

In Medumba and Ngemba, the tone changes of the type in (22) and (26) are not limited to the clausal domain. There is a parallel ‘within’ the DP. As discussed at length in Section 2, possessive pronouns in the two languages follow the head noun in non-focus contexts, as illustrated in (30) for Medumba, and (31) for Ngemba. In the two languages, the first-person possessive pronoun root in N > POSS has a single tone.

The pronouns in (30) and (31) can be focus-marked. When this happens, they precede the head noun on the surface. This is shown in (32), for Medumba and (33), for Ngemba. The examples in (32) and (33) are counterparts of (30) and (31), in POSS > N order.

As (32) and (33) show, the possessive pronoun roots surface with additional tones. Comparing the vowels on verbal roots in the clausal domain in both languages to those on the possessive pronoun roots in (32) and (33) shows that the effects are quasi-similar. In non-movement contexts, the verb roots within the clausal domain, and the possessive pronoun roots within the DP have a single tone. In movement contexts, a new tone is added. I assume that this parallel is not arbitrary, and propose that these effects also track movement from within the DP in these languages (see also Chow et al. Reference Chow, Daniel, Keupdjio, Mac-Deh and Ontkean2015 for a similar conclusion). It follows that tone changes can also track movement from within the DP in Ngemba and Medumba. The conclusion, therefore, is that POSS > N order in Ngemba and Medumba is derived by movement of the possessive pronoun into the DP left periphery. Since the same types of effects are also visible in Shupamem (c.f., Section 2), I suggest that POSS > N also involves movement in the language.Footnote ¹¹

In the section that follows, I propose a derivation of POSS > N and N > POSS in the three languages, and discuss how φ-features (number and gender, in this case) combine in Eastern Grassfields Bantu to spell out noun class.

4.1. Deriving the orders POSS > N and N > POSS

In deriving the relevant orders, I assume the more or less standard DP-structure in (34) and propose that IS-neutral N > POSS is derived by moving nP to SpecDP. This is somewhat in line with the idea that Bantu noun-initial word order can be derived via N-to-D movement (c.f., Carstens Reference Carstens1991, Nkemnji Reference Nkemnji1995, Tamanji Reference Tamanji2006, Carstens Reference Carstens, Everaert and van Riemsdijk2017). Poss is in D, and the functional projection NumP stands in-between D and nP (Ritter, Reference Ritter and Rothstein Susan1991, Preminger, Reference Preminger2020).

(34)

(35)

I adopt the view that DPs have a CP-like left periphery that hosts focused elements (see, among others, Aboh Reference Aboh2004). Specifically, modifier focus implies the presence of a FocusP above DP, as schematized in (37). POSS > N is derived by moving D to the Focus head (38).

(37)

(38)

Having shown how the two relevant orders can be derived, one can turn to how agreement and class exponence operate in the three languages. To do this, I will, in the next section, extend Carstens (Reference Carstens2026)’s grammar of gender in Bantu to Eastern Grassfields Bantu.

4.2. The grammar of gender in Eastern Grassfields Bantu

Certain phenomena in Bantu languages point to the direction that default agreement and/or animacy agreement is formal. One such phenomenon – and which is somewhat related to the Ngemba facts – is ‘animate concord’ (Wald, Reference Wald1975). In Chiyao, for example, the majority of human-referring nominals which, morphologically, fall outside of class 1 (the class that contains human-singular nouns in Proto-Bantu) obligatorily agree in animacy, i.e. they trigger the class 1 agreement marker on the verb. Below are examples from Riedel & Taji (Reference Riedel and Taji2022). Although chi-lundamisi ‘last.initiate’ is morphologically marked for class 7, the agreeing subject marker must be of class 1 (40a). The expected class 7 subject marker is ungrammatical (40b).

Xhosa also instantiates conjunct agreement phenomena that point to the direction that [animacy] features need to be grammaticalized. As discussed at length in Taraldsen et al. (Reference Taraldsen, Medová and Langa2018) and Carstens (Reference Carstens2026), a conjunct of class 1 nouns would trigger regular class 2 (human-plural class in Proto-Bantu) agreement, as expected (41).

The empirical puzzle is that some human-referring nouns that also fall outside of class 1 (morphologically) also trigger class 2 agreement. Example (42a) illustrates this for the combination [3 + 3], and (42b) for [5 + 5]. In (42a), the two conjoined nouns are morphologically marked for class 3, and one would expect the class 4 agreement marker on the verb. This is not what we get, as only human-plural ba- is grammatical. The simplest explanation for this would be that the agreeing verbs in (42) track the default (semantic) features of the conjoined nouns, and not the features of the classes to which they morphologically belong.

In her account of the type of phenomena that precede, Carstens (Reference Carstens2026) proposes a grammar of gender in Bantu, in which she argues for, among other things, the facts that n heads are of different types, and can host interpretable ( _I[gender]) and uninterpretable ( _U[gender]) gender features. This is meant to capture the empirical observation that human-denoting nouns, for example, can morphologically fall into non-human classes, such that noun class is not always semantically predictable, yet agreeing heads can sometimes track the relevant semantic features. A germane example in Xhosa is qhawe ; the root for ‘hero’. Although a human-denoting noun (which would, in most cases, warrant the class 1 marker), it is morphologically marked for class 5 which, semantically, is a non-human class (43a). The semantically regular class 1 prefix is ungrammatical (43b).

According to Carstens (Reference Carstens2026), the fact that agreeing heads can sometimes target the default semantic class features (rather than the formal class features) of certain nouns follows, if nouns are specified for interpretable (core semantic gender), as well as uninterpretable genders. For the nominal root qhawe ‘hero’, the interpretable gender would be [class 1] (Gender A, in Carstens Reference Carstens2026), and the uninterpretable gender would be [class 5] (Gender C).

That agreeing elements can target any of the two instances of gender follows if, she further assumes, n heads stack within Bantu DPs (see also Kramer Reference Kramer2015 and Fuchs & van der Wal Reference Fuchs and van der Wal2022, i.a.). She, following Kramer (Reference Kramer2015), proposes a structure similar to (44), where n_{U [gender]} always stacks on top of n_{I [gender]}.

(44)	nU [gender] stacks on top of nI [gender]

To derive the conjoint asymmetry in Xhosa (c.f., (42)), she argues that the gender and/or number features of conjunct members may clash, thereby forcing deletion of _U[gender]. The result is that the _I[gender] survives and can participate in agreement. Since the data I am interested in do not involve conjuncts, and since she assumes feature clash in conjunct contexts, I refer the reader to Carstens (Reference Carstens2024, Reference Carstens2026) for details of how deletion works, so as to derive the puzzling conjunct data from Xhosa.

The Eastern Grassfields Bantu data that the paper focuses on do not involve two DPs, hence feature clash of the type discussed in Carstens (Reference Carstens2026) is not expected to occur in the relevant constructions. What is crucial in accounting for the Eastern Grassfields Bantu data is that (a) n heads are of different types, and (b) they can stack. The empirical evidence for having _I[gender] and _U[gender] is even more pronounced in the languages under investigation in this paper, since the three of them only have 5–6 morphologically distinct classes. This presupposes, as discussed in the Grassfields Bantu literature (c.f., among others, Voorhoeve Reference Voorhoeve1968, Hyman et al. Reference Hyman, Voeltz and Tchokokam1970, Nkemnji Reference Nkemnji1994, Reference Nkemnji1995, Maho Reference Maho1999, Nchare Reference Nchare2011, Reference Nchare2012, Kouankem Reference Kouankem2012, Déchaine & Keupdjio Reference Déchaine and Keupdjio2016), that a lot of nouns are re-organized into single classes based on morpho-phonological, rather than semantic factors. Class 1, for example, is not semantically restricted to humans in the three languages (c.f., Kouankem Reference Kouankem2012, Nchare Reference Nchare2012, for relevant examples from Medumba and Shupamem, respectively). In Medumba, the non-human nouns ø-b ú ‘dog’ and b^hú ‘hand’ are morphologically marked for class 1. A relevant example in Shupamem is ø-tέt ‘pot’. In Ngemba, examples include ø-ndә̀ ‘house’, ø-kwɔ̀Ɂ ‘chair’, ø-pŭsì ‘cat’ and ø-tɔ̀ŋ ‘ear’. With this in mind, I adopt Carstens (Reference Carstens2026)’s theory, and propose that the internal structure of each Eastern Grassfields Bantu noun contains two types of n: n _I[gender] and n _U[gender]. In Carstens (Reference Carstens1991)’s theory of noun class exponence, [gender A] equates to class 1/2 (class 1 in the singular, and class 2 in the plural), [gender B] to class 3/4, etc.

(45)	Make-up and loci of gender features within Grassfields Bantu DPs

I propose that agreeing elements (in D, for possessive pronouns) always copy the features on the two instances of n, as schematized in (46). As (46) illustrates, the D head copies number features from Num, formal gender features from n2, and semantic gender features from n1.

(46)	Possessive pronoun agreement

Given (46), and in line with the idea that the n head spells out the nominal prefix in Bantu (Fuchs & van der Wal, Reference Fuchs and van der Wal2022), n1 and n2 compete under the Subset Principle (47), and n2 always wins because it hosts more specific gender features.

As such, the nominal prefix is the spell out of n2 in Grassfields Bantu. Number information combines with the formal gender features in n2 via allomorphic-like spell out rules of the type in (48) (c.f., Carstens Reference Carstens2026). (48a), for example, reads ‘[gender A] is spelled-out as the prefix /X-/ in singular contexts ([+sg])’. The same feature is spelled out as a different prefix, namely /Y-/, in plural contexts ([-sg]).

The D head in (46) scans its c-command domain for agreement targets, and copies all the φ-features (gender and number) it sees on its way (c.f., (46)). In the morphology, D, in class 2 contexts, looks like (49) below. Since [gender A] and [animacy] are of the same type and can be put in a superset/subset relation ([animacy] is the superset of [gender A]), [animacy] only participates in morphological realization if its subset [gender A] is absent. The result is that only features that refer to formal classes can be inserted in non-focus contexts.

(49)	Spell out of D in non-focus contexts

In the section that follows, I propose a derivation of the class exponence asymmetry with focus-marked possessive pronouns.

5.1. Ngemba

Ngemba is peculiar in that it instantiates animacy agreement in singular POSS > N order. The relevant data are repeated in (52)–(55).

In deriving the Ngemba puzzle, I propose the impoverishment rule in (56), such that the movement feature [•foc•] always triggers deletion of the feature [gender A, B, C], in singular contexts. (56) reads ‘[gender A, B, C] becomes null (ø) in the context where [•foc•] and [+sg] are present’. The rule in (56) is restricted in the sense that it never applies to non-focused contexts (because [•foc•] is absent) and in plural ([-sg]) contexts. [animacy] survives (c.f., the tree in (51)) because only formal gender features, i.e. [gender A, B, C] are affected by the deletion rule. Recall from (51) that the agreeing head copies both [gender A, B, C] features (on n2) and [animacy] features (on n1). Only [gender A, B, C] features are targeted for deletion, [animacy] is not. As such, it survives, and j- realizes [-animate], and w-; [+animate].

1. (56) Ngemba impoverishment rule

   [gender A, B, C] → ø / [•foc•], [+sg]

The rule, as it is formulated, does not apply to plural nouns, both in focused and non-focused contexts. This explains the fact that regular class markers are used throughout in the plural.

5.2. Medumba

The Medumba puzzle is such that only the class 1 agreement marker is replaced by the single glide j-. The empirical facts are repeated in (57) and (58) below.

I treat Medumba j- as an elsewhere/default prefix and propose the rule in (59), such that [•foc•] triggers deletion of [gender A] and [animacy] in singular contexts.(59) reads ‘[gender A] and [animacy] become null (ø) in the context where [•foc•] and [+sg] are present’. Recall from the end of Section 4.2 that in Carstens (Reference Carstens1991)’s theory of noun class exponence, [gender A] equates to class 1/2 (class 1 in the singular, and class 2 in the plural), [gender B] to class 3/4, etc. Since only class 1 is affected in Medumba, the impoverishment rule only targets [gender A].

1. (59) Medumba impoverishment rule

   [gender A] [animacy] → ø / [•foc•], [+sg]

1. (60) Feature specification of j- in Medumba

   /j/ ⟷ [ ]

When (59) applies, only elsewhere j- can be inserted.Footnote ¹²

5.3. Shupamem

As far as Shupamem is concerned, Section 2 has shown that agreement markers on focus-marked possessive pronoun roots are always replaced by the single glide j-. The empirical details are repeated in (61)–(64) for convenience.

I assume that Shupamem j- is also an elsewhere/default prefix in this language, and propose that [•foc•] triggers deletion of [gender A, B, C] and [animacy] in singular and plural contexts. (65) reads ‘[gender A, B, C] and [animacy] become null (ø) in the context where [•foc•] and [±sg] are present’.

1. (65) Shupamem impoverishment rule

   [gender A, B, C] [animacy] → ø / [•foc•], [±sg]

1. (66) Feature specification of j- in Shupamem

   /j/ ⟷ [ ]

The rule in (65) ensures that regular class markers never show up in modifier focus contexts in the language.

5.4. Why gender, but not number

A few questions remain to be answered. The first relates to the fact that all the impoverishment rules that I have so far proposed target gender, but not number. As a quick answer to this question, I propose that φ-features, i.e. gender and number in the case at hand, are hierarchically organized in Eastern Grassfields Bantu, such that impoverishment triggered by a movement feature cannot get to number without first deleting all gender features. The relevant hierarchy – also predicted to be at work in Baier (Reference Baier2018)’s system – is given in (67) below.

1. (67) Feature impoverishment hierarchy in Eastern Grassfields Bantu

   [gender] ≺ [number]

The second question relates to why impoverishment has to target [gender A, B, C], but not [animacy] in Ngemba. I propose that this is because [gender A, B, C] is somewhat a subset of [animacy]. Deleting [animacy] automatically deletes [gender A, B, C] under this view, which is not what we expect. Deleting [A, B, C], however, does not automatically delete [animacy]. Note that one can also think of this in terms of a feature hierarchy.Footnote ¹³

The third question has to do with why the targets of deletion are different in the three languages. As a quick reminder, [gender A, B, C], leaving out number features, are deleted in Ngemba. In Medumba, only [gender A], i.e. class 1, is deleted ([gender B, C] and number features are not affected). In Shupamem, full gender and number features are deleted. My view on this is that these differences reflect stages in an incremental process of gender and/or number deletion in possessive pronoun focus contexts within the Eastern Grassfields Bantu language family. Languages like Yemba (c.f., Fongang Reference Fongang2025a), which do not instantiate the kind of class asymmetry that Ngemba, Medumba and Shupamem do, would be classified in the group of languages in which such deletion processes are yet to take place. Medumba would be at the very early stage of deletion, as only class 1 is affected. Ngemba would be much more advanced than Medumba, as all singular classes are affected. Shupamem would be at the very last stage, as all the relevant features are deleted. Descriptive and typological investigations of Eastern Grassfields Bantu languages are, in this regard, warranted to, on the one hand, check whether other differences exist and, on the other hand, identify and group other languages under the different stages of deletion.