3.1.1 Ruling out a narrow-syntax approach to phonologically determined nominal concord

Any model of nominal concord must also account for phonologically determined concord systems like the one found in Guébie. A narrow-syntax approach to phonologically determined concord could take a number of forms.

Here I walk through each of these possible analyses, demonstrating that each of them encounters problems satisfactorily accounting for the Guébie data.

The first option is that phonological features are present in the syntax and available for copying during morphosyntactic agreement processes. This analysis relies on syntactic sensitivity to phonological information, which rejects the fundamental notion of a phonology-free syntax (Pullum & Zwicky Reference Pullum and Zwicky1986, Reference Pullum and Zwicky1988).

Most current theoretical frameworks assume the y-model of grammar (recall 1 and 2), where the syntax takes place before phonological information is present (cf. The Minimalist Program (Chomsky Reference Chomsky1993), Distributed Morphology (Halle & Marantz Reference Halle and Marantz1993, Reference Halle and Marantz1994)).

There are empirical reasons for adopting the Y-model of grammar, where syntactic operations occur before phonological ones. An analysis where phonological features are present during the syntactic module makes pathological predictions; if syntax was sensitive to phonology, we would expect word orders and other syntactic phenomena to be sensitive to phonological features such as segmental properties. Such phenomena are not found across languages. For example, we never see a language where a subject moves to pre-verbal position only if it begins with a voiceless consonant.

A model of grammar which disallows syntactic sensitivity to phonological features, like the Y-model, is more restrictive than one which allows phonology to affect syntactic operations, and makes better predictions about the linguistic systems we see cross-linguistically.

Option two in (12) assumes that all lexical items are arbitrarily indexed for the noun class they belong to: /ɓitə^2.3/ ⇔central vowel class; /sukulu^1.1.3/ ⇔back vowel class. In this case, the fact that the phonological form of the pronoun or concord marker is predictable given the form of the noun is coincidental. While this analysis is feasible, it assumes that all noun class assignments are memorized rather than productive. Additionally, on this analysis we might expect exceptions to the phonological predictability of the Guébie agreement system, or a lack of phonologically predictable extension to loan and nonce words. In certain Bantu languages we find two different /mu-/ noun class prefixes. These prefixes surface on the noun, and for one of the two mu- forms, there is a phonologically identical mu- which surfaces on agreeing elements in the noun phrase (ex: Class 18 in Ganda). However, there are other nouns which take a mu- prefix but trigger phonologically distinct prefixes on agreeing elements (ex: Class 1 in Ganda) (Meeussen Reference Meeussen1967). We never see such non-phonological agreement in the Guébie non-human class system.

A slight variation on this analysis would be to say that each noun in Guébie is composed of a root plus final vowel, where the final vowel expones noun class marking. The class marker would then determine the form of concord marking on pronouns and adjectives. Kaye (Reference Kaye1981) addresses this possible analysis and finds it untenable for Vata, a closely related Kru language. He rules out such an analysis for three reasons: (a) Claiming that each noun is underlyingly without its final vowel would mean that we have consonant-final roots in the language. This would be the only example of a consonant-final root, word, or syllable in Vata. The same reasoning holds in Guébie. (b) It is not possible to manipulate the final vowel of the noun to change the meaning of a word, and the final vowels of noun roots are not associated with any semantic meaning. Thus, there is no reason to consider the final vowels of nouns to be separate morphemes. (c) The final vowel of a noun root is not associated with an independent tone melody; its tone is determined by the lexical tone melody of the noun root. In Vata, there are no examples of toneless morphemes; every morpheme is underlyingly associated with a tone. Thus, analyzing the final vowel as a separate morpheme would be the only example of a toneless morpheme in the language. While the same is not true of Guébie, since other suffixes such as the definite marker lack a tone specification, I follow Kaye’s other arguments in ruling out this possible analysis. An additional reason for rejecting such an analysis is given below.

A particularly problematic set of data for this analysis comes from the fact that it is not always the final vowel of the noun which triggers concord. Recall that for nouns with plural suffixes or definite enclitics, it is the final vowel of the suffix or enclitic which triggers agreement. For example, the presence of the definite enclitic /=a/ always triggers the central vowel pronoun, agreeing with the definite marker rather than the noun itself, Table 14.

Table 14 Definite enclitics trigger central vowel phonological agreement.

If each noun+final vowel were indexed for a particular lexical class, there would be no a priori reason to predict that the definite marker should suppress the noun class agreement triggered by the noun or final vowel itself. We would expect that the class marking on the noun would still determine concord.

The final option in (12) says that phonological agreement results from multiple copies of the noun being present in the syntax, some of which can be reduced phonologically to the final vowel of that noun. Dimitriadis (Reference Dimitriadis1997) proposes a version of this analysis for phonologically determined nominal concord in Bainuk (Atlantic) and Abu’ (Arapesh).

On this multiple-copy analysis, for the Guébie noun phrase ‘new red house’ in (13), we would need to say that there are three copies of the noun present in the syntax, one which surfaces as a full copy, and two at the end of each adjective, which are reduced to the final vowel of the noun.

A Guébie noun phrase like (13) would require three copies of the noun to be present in the syntactic structure, where one of them is fully pronounced and the other two are partially pronouncedFootnote ^[9] . The problem is that there is not supporting evidence, syntactic or morphophonological, for such redundancy in Guébie. If future work were to find that this analysis of multiple syntactic copies of the noun is in fact compatible with the syntactic structure of DPs in Guébie, we would still need to explain why in this particular case only, we see multiple copies of the noun (partially) spelled out, while other instances of multiple syntactic copies in Guébie only result in one overtly spelled out copy.

Additionally, this analysis predicts the existence of some language in which multiple copies of the noun exist and are fully pronounced on the surface. To my knowledge, this pattern is not attested.

The three models discussed here for accounting for phonologically determined concord encounter fatal flaws when faced with modeling the Guébie data.

In Section 3.1.2 I rule out a purely phonological approach, and I turn to the proposed analysis in Section 3.2.

3.1.2 Ruling out a purely phonological account

Instead of accounting for phonologically determined concord in the narrow syntax, one might consider pursuing a purely phonological analysis. This could take the form of long-distance phonological agreement in the Agreement-By-Correspondence (ABC) framework (Rose & Walker Reference Rose and Walker2004). In such a model, the concord marker on the pronoun and/or adjective would be in correspondence with the nominal antecedent, and phonological identity would be required between the two. However, the phonologically agreeing pronoun occurs even when the noun is not pronounced in the discourse, as in (14).

In the context above, the word ‘eggplant’, /trobiə^3.2.2/ has not been uttered aloud; however, the pronoun must surface with the agreeing vowel [a³] and not another third-person singular pronoun vowel, #[ɛ, ʊ, ɔ].

ABC requires agreeing elements to be overt and within the same local domain so that one element can copy features from the other. Because agreement between a noun and pronoun is required in Guébie even when the noun is not present (14), ABC is not enough, at least on its own, to account for the phonological agreement of pronouns with nouns in Guébie. Because Guébie nominal concord is non-local, and the head noun need not be overt within the same utterance or even in the same discourse for agreement to hold, a long-distance phonological agreement analysis will not suffice (Sande Reference Sande2014).

The account proposed in Section 3.2 makes use of morphological agreement mechanisms, in combination with ABC-style constraints plus output-to-output paradigm correspondence in the phonological component to derive the Guébie facts.

3.2.1 Syntactic structure

To determine the syntactic structure of DPs in Guébie, we examine one further set of data. Pronouns in Guébie can occur alone within a noun phrase, Table 15(b), but they can also co-occur with an overt noun, Table 15(c). For most speakersFootnote ^[10] , the definite marker can never co-occur with the pronoun. Unlike pronouns, definite markers cannot surface without an overt noun, Table 15(f). Adjectives and numerals are impossible in noun phrases that contain a pronoun, Table 15(g, h).

Table 15 Distribution of nouns and pronouns (lau_20150617).

Constructions like Table 15 (c), where the noun and pronoun surface together within the same noun phrase, are similar to the ‘we linguists’ construction in English, except that in Guébie they are not restricted to first and second persons. This noun–pronoun construction is independent of topicalization; there is a separate topicalization construction in Guébie which can also result in an overt pronoun following an overt noun. However, in topicalization constructions, the noun must be clause-initial, and there is a pause between the topicalized noun and the pronoun, which surfaces in subject position, immediately before the inflected verb or auxiliary.

In [Noun Pronoun] constructions as in Table 15 (c), the noun and pronoun are phrased together, while in (15) the noun and pronouns are separated by a required pause.

Following Elbourne (Reference Elbourne2001)’s analysis of e-type pronouns, I assume that pronouns take a noun phrase complement which is optionally elided at PF. The pronoun is a D-head in a head-initial DP, and a phrasal projection containing the head noun moves to the specifier of D to result in Noun-Pronoun order. The noun, which has moved to a higher position, is optionally elided (where ellipsis is licensed by the presence of the pronoun), resulting in all and only the two grammatical overt pronoun structures, Table 15(b, c). To account for the fact that the definite marker and pronoun do not co-occur, the two are both analyzed as D heads, where only one element can surface in the D position in a given constructionFootnote ^[11] .

The syntactic structure which serves as the input to the morphological component is as in (16)Footnote ^[12] .

The crucial component of this structure is that the NP surfaces in the specifier position of the pronoun D-head, and can optionally be elided during the phonological component. This structure serves as the output of the syntactic component and the input to morphology.

Along with the structure above, I make one additional assumption about the syntax–morphology interface. Namely, that syntactic structures are made available to morphology and phonology in subparts, where each syntactic phase head triggers spell-out. At each syntactic phase boundary morphological and phonological operations take place (Chomsky Reference Chomsky, Martin, Michaels and Uriagereka2000, Marvin Reference Marvin2002), and DP is a phase (Svenonius Reference Svenonius2004, Cinque Reference Cinque, Aboh, Guasti and Roberts2014), or minimally the head noun and pronoun are spelled out simultaneously.

Synchronic syntactic evidence that DP is a phase in Guébie comes from the fact that extraction from a DP is impossible. For example, it is possible to say (17a), but not (17b, c).

Whenever a non-human pronoun is produced in Guébie, it agrees in phonological features with the noun in the same noun phrase, whether or not that nouns are overtly pronounced. I propose that the concord-triggering noun is always present in the syntax, within the same phase as the pronoun D-head. Then, phonological concord is conditioned by a morphological rather than syntactic agreement relationship between the noun and pronoun. This agreement relationship is founded during the morphological component, via insertion of an Agr(eement) node on the pronoun (Kramer Reference Kramer2010, Norris Reference Norris2014), as discussed in further detail in Section 3.2.2. Due to this agreement relationship, morphosyntactic features of the noun are copied to the pronoun. The phonology, which also applies at phase boundaries, has access to the morphosyntactic features of heads within that phase, and phonological constraints ensure phonological identity between those heads in the DP which agree in specific features, as detailed in Section 3.2.3. Ellipsis of the noun optionally occurs at PF, licensed by overt phonological agreement between the noun and the pronoun.

In Section 3.2.2 I discuss the morphological component, based on DM, and in Section 3.2.3 I provide a formal analysis of phonologically determined nominal concord in Guébie, reliant on morphosyntactic features being maintained through the morphology, available to the phonological component.

3.2.2 The morphological structure

In the proposed model, morphology and phonology apply cyclically to syntactic structures by phase, and each DP is a phase. The morphology takes the syntactic structure as an input. I assume a DM morphological component (Halle & Marantz Reference Halle and Marantz1993, Reference Halle and Marantz1994; Embick & Noyer Reference Embick and Noyer2001). The crucial reasons for choosing a DM-style morphology are late insertion of vocabulary items, spell-out at phase boundaries, and a mechanism for establishing post-syntactic agreement (concord) relationships. Any morphological framework which follows DM in these three respects would serve equally as well in the given account.

Via regular DM agreement mechanisms, an AGR(eement)-node is inserted on D, and the  feature is copied to it from the noun (see Noyer (Reference Noyer1997), Embick & Noyer (Reference Embick and Noyer2001), Kramer (Reference Kramer2010), Norris (Reference Norris2014)).

Node insertion in DM occurs only when the relevant morphological features have no bearing on semantics (cf. Oltra Massuet (Reference Oltra Massuet1999) and Embick (Reference Embick2010) on dissociated morphemes, and Harbour (Reference Harbour2003) on ‘default’ feature insertion). That is, only those terminal nodes which affect the truth value of the sentence are present in the syntax, and others are inserted during the morphological module of grammar, post-syntactically. After AGR-node insertion occurs, feature copying from the head noun to the AGR node applies.

Agreement proceeds as shown in the noun–pronoun construction in (19) for the noun sukulu ‘school’. The vocabulary item sukulu has two features, the feature   because it is a noun, and the feature motivating optional ellipsis at PF, which following Merchant (Reference Merchant2001, Reference Merchant2008) I call   The   feature is discussed further in Section 3.2.3. The noun feature of sukulu has been copied to the AGR node on D. Because the shape of pronouns, whether human or non-human, is always a -ATR vowel (V), I assume that the non-human pronoun vocabulary item is a vowel specified for the phonological feature   but underspecified for other features, specifically   The backness value will be specified via the constraint-based phonologyFootnote ^[13] .

For simplicity, I leave out the syntactic nodes below D in (19). These include positions for adjectives and numerals.

Lexical items are inserted into the structure in (21) via regular DM-style vocabulary insertion rules. For the majority of terminal node feature bundles in Guébie, there is some lexically associated phonological content. This content can be fully specified, as in ‘school’, /sukulu ^1.1.3/, or partially specified, as in third-person non-human pronouns,   There could also be a set of morphosyntactic features for which there is no corresponding phonological content. This is seen elsewhere in Guébie, where the imperfective morpheme triggers a particular phonological process, but is not associated with any underlying phonological content (Sande Reference Sande2017: Chapter 5).

After vocabulary items and AGR nodes are inserted, the morphological structure in (19) is linearized via DM Linearization mechanisms, as laid out by Embick (Reference Embick2010).

Note that in the proposed analysis, the morphological features associated with terminal nodes are preserved through morphology, including Linearization, and are available to the phonology (following Gribanova & Harizanov (Reference Gribanova and Harizanov2015), Winchester (Reference Winchester2016); contra Halle (Reference Halle1990), Bobaljik (Reference Bobaljik2000))Footnote ^[14] . This feature preservation will allow the phonological component to see morphosyntactic agreement relationships.

3.2.3 The phonology

Here I adopt a constraint-based phonological component with paradigm output–output faithfulness (Burzio Reference Burzio1994, Benua Reference Benua1997, Kager, van der Hulst & Zonneveld Reference Kager, van der Hulst and Zonneveld1999). For other recent works combining a DM-style morphology with a constraint-based phonology, see Jenks & Rose Reference Jenks and Rose2015, Kastner Reference Kastner2016; Sande Reference Sande2017, Reference Sande2018.

While I choose to show the phonological component evaluated in parallel, a cyclic approach using strata, like Stratal OT (Kiparsky Reference Kiparsky2000, Reference Kiparsky, Vaux and Nevins2008; Bermúdez-Otero Reference Bermúdez-Otero1999), or serial derivation, like Harmonic Serialism (McCarthy Reference McCarthy and Hirotani2000) would work equally as well as the parallel approach provided hereFootnote ^[15] . Since my point here is not to choose between a parallel, stratal, or serial phonology, but rather to show that a model of grammar where phonology follows syntax and is sensitive to morphosyntactic features can account for phonologically determined nominal concord, I set aside the differences between stratal or serial and parallel approaches and use parallel evaluation for simplicity.

In this model, phonology applies at phase boundaries, and DP is a phase. Thus, the DP, which crucially includes nouns, pronouns, and adjectives, will be evaluated together as a unit by phonological constraints.

The linearized structure provided by the morphological component of grammar serves as the input to phonology. This linearized structure consists of vocabulary items and morphosyntactic features, (21). Note that the phonological features of the pronoun vowel other than [ATR] are not specified in the input to the phonological component. The quality of the pronoun will instead be determined via ranked constraints.

To arrive at the correct output, i.e. [sukulu ʊ] or   based on the linearized input in (21), we need a constraint ensuring identity between the final vowel of the noun and the vowel of the pronoun. This is accomplished with Anchor-R, which anchors agreement to the right edge of a word, (22).

This constraint is only active if the heads in question agree in some morphosyntactic feature. If they do agree morphosyntactically, segments at the right edge of each head will correspond. A constraint ensuring that heads in correspondence are phonologically identical must also be active. I propose the use of an output–output identity constraint Ident-OO which says that heads that agree in the feature   must agree in phonological features.

Ident-OO assigns a violation for each non-identical set of corresponding segments in two corresponding heads. This constraint will only affect nouns and their corresponding pronouns and adjectives, since other heads do not agree morphosyntactically.

The combination of the two constraints in (22, 23) has the result that two heads agreeing in the morphosyntactic feature n within the same spell-out phase will be phonologically identical, starting from the right edge of the word. The optimal candidate violates a single constraint here, namely Dep-feature, which penalizes output features not present in the input. Dep-feature is violated by the optimal candidate because the pronoun vowel has fully specified vowel features in the output, but not in the input. The benefit of Dep-feature is that it rules out candidates like [sukulu sukulu] where the pronoun is identical to the noun in more than just one segment, because [sukulu sukulu] involves insertion of more new features from input to output than does [sukulu ʊ].

The tableau below shows that the presence of Ident-OO rules out a pronoun vowel that does not agree phonologically with the noun, Table 16(d). Anchor-R rules out a pronoun that is phonologically identical to the left edge of the noun, Table 16(c). Dep is necessary to rule out a pronoun that is identical to the entire phonological form of the noun, or even anything more than the final vowel, Table 16(b). Here I mark a single violation of Dep for each segment present in the output that was underspecified or not present in the input. This decision is for simplicity of reading the tableaux, because in fact each candidate below would incur many more Dep-feature violations than marked, namely, one for every phonological feature inserted, rather than one for every consonant/vowel segment inserted.

The justification for the ranking in Table 16 followsFootnote ^[16] .

Table 16 Ident-OO, Anchor-R  $\gg$ Dep.

The combination of the correspondence constraint Anchor-R and the identity constraint Ident-OO function to rule out candidates that fail to agree, as per ABC theory (Hansson Reference Hansson Gunnar2001, Rose & Walker Reference Rose and Walker2004). Candidate (d) is ruled out by the output–output correspondence constraint, Id-OO, since the vowel of the pronoun is not featurally the same as any corresponding segment in the noun. Anchor-R rules out candidate (c), where there are identical corresponding segments in the noun and pronoun, but the corresponding segments are at the left edge of the noun, rather than the right. Dep-feature rules out candidate (b), since more features are inserted between the input and output than in the winning candidate, (a).

While the tableau in Table 16 rules out a number of unwanted candidates, without an additional constraint, the candidate [sukulu] with a null pronoun would beat the optimal candidate because it involves no feature insertion. We must ensure that the pronoun surfaces overtly, despite its lack of fully specified phonological feature information in the input. This can be assured with a RealizeMorpheme constraint, which penalizes an output candidate that does not overtly realize an input morpheme, (25).

While the constraints in Table 17 explain why we get a surface pronoun that is a single segment and agrees with the final segment of the noun, they do not explain why the features of the final vowel of the output noun are identical to the input features. That is, why don’t we have an optimal output candidate [sukule ɛ] from input /sukulu V/? While Dep-feature rules out adding a feature from input to output, it does not rule out changing a feature from input to output. Thus, to rule out feature changing, which violates input–output identity, I propose the use of a highly ranked Ident-IO constraint.

Table 17 Ident-OO, Anchor-R, RealizeMorph.

The full Guébie vowel inventory contains ten vowels, [i, ɪ, e, ɛ, u, ʊ, o, ɔ, ə, a], but there are fewer possible singular non-human pronoun vowels, [ɛ, a, ʊ]. The specified {-ATR} feature on the pronoun vocabulary item limits the possible pronoun vowels to [ɪ, ɛ, ʊ, ɔ, a]. Additional constraints such as PeriphVowel preferring peripheral vowels [ɪ, ʊ, a] and *ɪ dispreferring the output segment [ɪ] in Guébie account for the reduced number of pronoun vowels [ɛ, a, ʊ], compared to the full Guébie vowel inventory aboveFootnote ^[17] . As this is secondary to the point of this section, I leave these constraints out of the tableaux below.

Ranked as in (27), the above constraints lead to the correct output of a [Noun Pronoun] structure, where both the noun and the pronoun are overt. These constraints ensure that the pronoun agrees phonologically with the final vowel of the noun in question.

When a pronoun surfaces without an overt noun, I posit that the noun is present in the syntax but is elided at PF,   ‘it (school)’ (see Merchant (Reference Merchant2001), Lasnik (Reference Lasnik2007)). Constituents that can optionally be elided are marked with a feature   in the syntax (Merchant Reference Merchant2001), and here I propose a model of ellipsis where the phonology has access to the   feature of the noun, just as it has access to other morphosyntactic features, such as the   feature triggering phonological agreement. The option of eliding the noun is then determined via constraints.

The presence of an   feature triggers what I call here an ellipsis paradigm. This paradigm involves two output forms evaluated together, one where ellipsis has occurred, and one where it has not. Both cells of the paradigm are evaluated simultaneously in the same instance of spell-out.

The novel constraint in (28) is an output–output paradigm correspondence constraint (Burzio Reference Burzio1994, Benua Reference Benua1997, Kager et al. Reference Kager, van der Hulst and Zonneveld1999, McCarthy Reference McCarthy, Downing, Hall and Raffelsiefen2005), which ensures that the phrase (or syntactic phase) containing the elided element is as similar to the optimal non-elided output as possible. For example, the elided form   must be faithful to the non-elided [sukulu ʊ].

In an output–output paradigm correspondence model such as this, candidates consist of paradigms, which are evaluated together as a unit. In Table 18 there are both input–output correspondence relationships, as well as output–output paradigmatic correspondence relationships. We see that in Guébie, when DPs containing elided and non-elided nouns are evaluated together in a paradigm, the undominated constraint in (28) together with those constraints in (27) gives the correct output. That agreement can be sensitive to unpronounced material is well known (Merchant Reference Merchant2015: 16), and the proposed constraints show an articulated model of this particular phenomenon.

Table 18 A constraint-based approach to ellipsis.

Every form in Table 18 receives at least one RealizeMorph violation because the noun /sukulu^1.1.3/ is unrealized in the second form of the paradigm.

The proposed analysis forces phonological agreement and provides the option of ellipsis at PF simultaneously via constraints (with regards to the latter, this analysis is similar to Bennett, Elfner & McCloskey (Reference Bennett, Elfner and McCloskey2015)’s analysis of Irish ellipsis at PF). A terminal node which has a morphosyntactic  feature, available to the phonology, can optionally be elided via an ellipsis paradigm at PF, as in Table 18.

By evaluating paradigms of elided and non-elided candidates at each syntactic phase boundary, we predict phonological agreement of elements within a phase (here, DP) that agree in some morphosyntactic feature. Further predictions of the proposed model are discussed in Section 4.

3.3.1 Possessive and emphatic pronouns

Non-human possessive and emphatic pronouns also show phonologically determined concord marking triggered by the head noun. However, unlike subject and object pronouns, the concord-marking vowel in possessive and emphatic pronouns is not rightmost within the word (ex: the front vowel emphatic pronoun ɛɓa^3.2 and possessive pronoun ɛnɛ^2.3). I propose that, just as for nominative and accusative pronouns, the initial vowel in non-human possessive and emphatic pronouns is underspecified for all phonological features except -ATR: / $V_{-\text{ATR}}$ ɓa/, / $V_{-\text{ATR}}$ nɛ/. The ranking of Ident-IO above Anchor-R prevents the rightmost segments of the emphatic and possessive pronouns from undergoing a change to show phonological identity with the final vowel of the noun, since those segments were present in the input. The rightmost manipulable segment in the emphatic and possessive pronouns is the featurally underspecified initial vowel. This segment, then, shows phonological identity with the right edge of the head noun, due to Anchor-R.

3.3.2 Plural and definite-marked nouns

When a non-human noun surfaces as plural or definite, the quality of the plural or definite vowel determines concord. This, too, is accounted for with the phonological constraint ranking above, since the definite marker and plural suffix are part of the same phonological word as the concord-triggering noun, and Anchor-R targets the right edge of a phonological word. Thus, the rightmost vowel within the noun word determines the features of concording vowels, whether that vowel is part of the root, suffix, or enclitic.

3.3.3 Guébie human pronouns

Third-person human pronouns in Guébie do not follow the phonological agreement pattern of non-human nouns. Instead, they predictably take the forms /ɔ/, singular, and /wa/, plural. I repeat the pronoun chart for Guébie from Table 3 in Table 19 below.

Table 19 Human and non-human subject pronouns.

The model described in Section 3 extends to human pronouns in Guébie without modification. We saw that nouns are present in the syntax in the same DP as pronouns, and their features are copied to the pronoun via a morphological AGR node. I claim here that human nouns not only have a   feature which is copied to the pronoun, they also have a   feature (Richards Reference Richards, Bornkessel-Schlesewsky, Malchukov and Richards2015, Van der Wal Reference Van der Wal2015), Table 20.

Table 20 Pronoun features and realization.

The presence of the   feature triggers insertion of a different vocabulary item than for non-human pronouns. In the case of human pronouns, the inserted vocabulary item is fully specified for phonological features. This is exemplified for [ŋudi^3.1 ], ‘man’, in (29).

When AGR-node insertion on D occurs and features are copied from a human noun to the AGR node on the pronoun D,   and   features are copied along with the   feature. These   and   features are absent on non-human nouns.

Then, during vocabulary insertion, the particular bundle of features containing   and   is spelled out as [ɔ], as in (31). That is, the 3rd singular human vocabulary item /ɔ/ is inserted in the context of the features  Similarly, the plural human pronoun [wa] is inserted in the context of the features   This differs from all non-human nouns which are not marked for person or number features in the syntax.

If certain semantic features of the noun such as person and number are copied to the pronoun D via morphological agreement mechanisms and spelled out by a vocabulary item with specified phonological features (   $\rightarrow$ [ɔ]), that vocabulary item is not subject to phonological identity. Instead, a highly ranked constraint ensures faithfulness to the phonological content inserted during vocabulary insertion. This Ident-IO constraint must be ranked higher than the Ident-OO constraints requiring phonological agreement between agreeing elements in the DP, providing evidence for a more nuanced constraint ranking than the one presented in Table 18. We could imagine, then, a language with the same constraints but where input–output faithfulness was low-ranked, where the entire pronoun system would be phonologically determined, including first and second persons. As far as I know, no such language has been described, but the system proposed here predicts that it could exist.

Phonological identity between the pronoun and agreement-controlling noun seems to be a last resort agreement strategy in Guébie. Specifically, phonological identity holds only in those cases where there is no relevant vocabulary item with specified phonological content to insert. This prediction is supported by Corbett (Reference Corbett1991)’s generalization that when semantic and phonological criteria for determining noun class are at odds, semantic features take precedence. I return to the discussion of the relationship between semantically and phonologically conditioned concord in Section 5.

3.3.4 Guébie adjectives

Adjectives in Guébie agree in final vowel with the noun that they modify.

We can derive this agreement in the same way as noun–pronoun agreement. Syntactically, nouns, along with the adjectives that directly modify them, are present in a single syntactic phase (DP). An AGR node is inserted on the adjective by the morphology. Features of the noun (namely, ) are copied to the adjective so that the adjective and noun are in morphosyntactic agreement. The phonology ensures that agreeing heads (the noun and its adjectival modifiers) are phonologically similar, via the same constraints discussed in Section 3.

Much in the same way that pronouns license ellipsis of the agreement-triggering noun (15b, c), adjectives that agree with the head noun license ellipsis of that noun, (34).

Just like optional nominal ellipsis in [Noun Pronoun] constructions, [Noun Adjective] candidates are evaluated in paradigms, with two forms in each paradigm: one where the noun is elided and one where it is overt. A Faith-NoElide constraint ensures output–output paradigm faithfulness so that the adjective agrees phonologically with the noun even when the noun is elided. The relevant constraint ranking is identical to the one shown for noun/pronoun agreement in Table 18.

3.3.5 Interim summary: Analyzing phonological concord in Guébie

A single analysis dependent on phase-based application of morphological and phonological operations explains the phonological concord marking on nominative, accusative, emphatic, and possessive pronouns, as well as adjectives in Guébie; all involve an element (pronoun or adjective) with an optionally elided noun in the same DP, undergoing morphological AGR-node insertion, followed by phonological correspondence and identity.

4.1.1 Krahn

Bing (Reference Bing1987) describes an agreement pattern in ɡbɔbo, a dialect of Krahn (Western Kru) spoken in Liberia and Côte d’Ivoire, that is quite similar to the Guébie pattern. There are nine vowels in the Krahn system, and there are four possible third-person singular pronoun vowels: one for humans and three phonologically determined ones for non-humans. Non-human nouns that end in front vowels take the front vowel pronoun [ɛ], those that end in non-high back vowels take the pronoun vowel [ɔ], and those that end in high back vowels take the pronoun vowel [ʊ], Table 21.

Table 21 Krahn phonological agreement.

Since the Krahn system is so similar to the Guébie one, it requires no extra theoretical tools to account for the data. The proposed model would apply to Krahn just as it does to Guébie, ensuring phonological agreement between the final vowel of the noun and pronoun unless the noun is human, in which case the semantic features win out. The only significant difference is that Bing does not mention any category of element other than pronouns that agrees with the noun in Krahn. If adjectives do not agree phonologically with the nouns they modify, we can assume that adjectives in Krahn do not agree morphologically in features with nouns; thus, no phonological identity is required to hold between them.

4.1.2 Godié

Godié is an Eastern Kru language spoken in Côte d’Ivoire. Just like Guébie and Krahn, there are four possible third-person singular pronoun vowels in Godié: one human vowel and three phonologically determined vowels. However, Godié agreement processes target not only pronouns, but also definite clitics, demonstratives, and adjectives (Marchese Reference Márchese1986b, Reference Márchese1988).

In the Godié example below, the human word ‘man’ triggers the agreement vowel [ɔ] on the adjective and demonstrative following it. The final front vowel of the word ‘animal’ triggers the front agreement vowel [ɛ] on the adjective [kəd-ɛ] that describes the word ‘animal’.

Since demonstratives, definite clitics, pronouns, and adjectives are all within the DP domain, all of them should be equally likely to agree with the noun. I have proposed that the phonology applies by phase, and that DP is a phase, so the phonological analysis applies to any two elements within a DP phase as long as they are in a morphosyntactic agreement relationship. Thus, the difference between the Godié agreement system and the Guébie system is that in Godié demonstratives and definite markers are in morphosyntactic agreement with the noun, while in Guébie they are not. Guébie lacks demonstratives entirely but has a definite clitic /=a/ which surfaces on the noun. Further research is need to determine whether there are any true syntactic differences between Guébie and Godié definite markers which shows that they are in agreement with the noun in Godié but not Guébie. In the proposed model, demonstratives and definite markers in Godié are analyzed as agreeing morphologically with the head noun.

4.1.3 Vata

Vata is an Eastern Kru language spoken in south-central Côte d’Ivoire. The Vata system differs slightly from the phonological agreement systems of other Kru languages discussed thus far. There are ten contrastive vowels in Vata, at five places of articulation with an ATR contrast, /i, ɪ, e, ɛ, u, ɯ, o, ɔ, ə, a/. Rather than three possible non-human pronoun vowels like Guébie, Krahn, and Godié, Vata has five non-human singular pronoun vowels: one for each of the five degrees of height and backness /ɪ, ɛ, ɯ, ɔ, a/ (Kaye Reference Kaye1981).

The elements that show concord marking in Vata include pronouns, relative pronouns, definite markers, and a limited number of adjectives, Table 22.

Table 22 shows one noun ending in a +ATR and one ending in a -ATR vowel for each of the five height/backness distinctions. The pronoun and relative pronoun themselves remain -ATR even when the noun ends in a +ATR value. Only the backness, height, and rounding of the vowel are determined by the final vowel of the noun. Concord between nouns and definite markers is shown in (37).

Table 22 Pronouns in Vata.

We can extend the analysis from Section 3 to Vata agreement with little changeFootnote ^[18] . We only need to rerank certain constraints to get the right output. In Guébie, there is a ten-vowel system in the language which is reduced to three possible agreeing vowels for non-human pronouns. I mentioned in Section 3 that in order to account for the reduced number of possible pronoun vowels in Guébie, [ɛ, a, ʊ] as opposed to the full ten [i, ɪ, e, ɛ, ə, a, u, ʊ, o, ɔ], we would need constraints like PeripheralVowel which prefers the peripheral -ATR vowels /ɪ, a, ʊ/, and *I to prefer /ɛ/ over /ɪ/. In Guébie these constraints must be highly ranked, only crucially out-ranked by Ident-IO. However, in Vata, the same constraints must be very low-ranked, because they play no role in the Vata agreement system. In Vata, for every distinct final vowel on nouns, there is a corresponding pronoun vowel that has the same height, backness, and rounding features. Only the ATR features of the pronoun are pre-specified on the pronoun vowel. Thus, by simply ensuring that Ident-OO outranks PeripheralVowel and other such vowel markedness constraints, we get the correct output for Vata without otherwise changing the analysis for Guébie presented in Section 3.

It is worth noting that the kind of minor typological variation we see between Guébie and Vata is predicted by a constraint-based analysis like the one presented here, where by reranking the proposed constraints we model another predicted system. However, this variation is less obviously expected in a rule-based phonology or a purely syntactic approach to phonologically determined nominal concord.

4.1.4 Summary of Kru phonological agreement

Krahn and Godié, like Guébie, have three possible forms for non-human third-person singular pronounsFootnote ^[19] . The optimal form is the one that agrees with the noun phonologically. In Vata, there are five possible vowels for non-human third-person singular pronouns, where height and backness, as opposed to just backness of the pronoun vowel is determined by the final vowel of the noun, Table 23.

Table 23 Nominal concord across Kru.

We can see from Table 23 that the most likely elements among Kru languages to show concord with the noun are pronouns.

Interestingly, surface word order in Vata differs from that in Guébie, Krahn, and Godié. In Guébie, surface order in noun phrases is either Noun-Def Numeral Adjective, or Noun Adjective Numeral-Def. There are no demonstratives in Guébie. In Godié and Krahn, demonstratives are post-nominal. However, in Vata, demonstratives are pre-nominal, while all other modifiers follow the noun. Future work will determine whether these word order differences are related to the differences in which elements within the noun phrase show concord with the head noun across Kru languages.

The phonologically determined nominal concord systems in Krahn, Godié, and Vata all closely resemble the Guébie system except that a different set of elements agrees with the noun in each language. However, because all of the agreeing elements occur within the DP phase, each system above is predicted by the proposed analysis (cf. Prediction A in 35).