5.1 Alternations

The alternations involved in argument-head tone sandhi are complex, dependent upon both the tone of the argument and the tone of the head. Making the situation even more complicated is the fact that the patterns differ slightly depending upon whether the argument is a simplex pronoun, a complex pronoun, or a non-pronominal DP, and, in some tonal combinations, on the syntactic category of the head. I will address (simplex) pronominal arguments first in §5.1.1 then describe how non-pronominal arguments diverge in §5.1.2.

5.2 Domains

The question of domains of sandhi application is not without complication. I will first identify what I consider to be the core domains and sketch out two possible approaches to formally defining them for sandhi application. I will then turn to domains in which sandhi does not occur, showing that these present challenges to both approaches. Finally, I turn to other domains where sandhi occurs, which are likely derived (synchronically or diachronically) from these core domains.

5.2.1 Core domains of application

Argument-head tone sandhi, in its most canonical application, occurs between a head and its preceding internal argument. There are three such domains where sandhi is attested:

1. Possessive DP + inalienable/relational noun

2. Object DP + transitive verb

3. DP + postposition

Examples of possession were given in tables (10) and (13) above; comparison tables for OV and PP environments are given in (20) using pronominal arguments:

(20)	a.	Tone sandhi with pronominal objects of transitive verbs
	b.	Tone sandhi with pronominal complements to postpositions

Verbal and postpositional heads display identical sandhi behavior, but they differ from the inalienable nouns in (10) in that they have an X-toned allomorph after H-toned pronouns, regardless of whether the pronoun is simplex or complex (emphatic). All heads behave identically with non-pronominal arguments, where H-toned allomorphs are never derived.

If we wish to generalize about the domain of sandhi application, we have two possibilities. The first is that the domain is defined with direct reference to the syntax (e.g. Pak 2008). In this approach, we could argue that the core environments in which sandhi occurs share a parallel syntactic structure, shown in the tree in (21):

1. (21)

1. 

In this tree, a relational NP, a VP, and a PP all take an internal DP argument that is the complement to the head. For VPs and PPs, this structure is relatively uncontroversial, but for relational or inalienable nouns, it is more commonly argued that possessors are generated in the specifier of NP rather than as a complement to N (see e.g. Abney 1987; Stowell 1989; Vergnaud and Zubizarreta 1992; Español-Echevarria 1997). However, structures like that in (21) have also been proposed (see e.g. Yoon 1990; Ura 1996). In this approach, a head X plus a DP internal argument would form the domain for sandhi application.

In an indirect reference approach (e.g. Nespor and Vogel 1986; Selkirk 2011, among many others), prosodic constituents are built off syntactic structure but are not necessarily isomorphic with it. Given the small size of sandhi domains, the most likely prosodic constituent would be the phonological phrase (φ).⁸ If we assume a Match Theory approach to prosodic constituent building (Selkirk 2011), we could propose that the left and right edge of an XP match the left and right edge of a phonological phrase. For the tree structure in (21), this would create the following recursive structure:

(22)	[ [ [ N ]φ ]φ X ]φ

In other words, the noun and its NP forms one phonological phrase, the DP layer forms another one, and both of these are embedded within a larger phonological phrase for the whole XP. The domain for sandhi would be any phonological phrase, whether minimal or maximal.

7.1 Sandhi and plural formation

Argument-head tone sandhi is by no means the only morphophonological tone process in Seenku. In some cases, sandhi can interact with these other processes, and these interactions provide us with evidence of timing of process application in the language. In this paper, I will focus specifically on the tone raising found in plural formation.

First, as outlined in McPherson (2017a), the plural in Seenku is formed by a combination of tone raising and vowel fronting (resulting in back-front diphthongs in stems with round vowels, (40c)); I will focus on the tonal effects here. Tone raising is argued to be driven by the affixation of a [+raised] tonal feature, raising X to L and H to S, as shown in (40) (where slashes represent the form of the noun in the lexicon):

S is the highest tone in the language, already specified as [+raised], and thus the plural has no tonal effect (40c). The fact that singular bî (with a HX contour) raises to S in the plural rather HL is one piece of evidence that these nouns are lexically level H-toned, with the X added in the phonological component to satisfy a tonotactic ban on final H in nouns.

The question is: Does the tone raising of plural formation apply before or after tone sandhi? The answer is both. If the argument is plural, its plural form is computed before it triggers sandhi on the head, as shown in (41):

I have shown the lexical heads in the input to sandhi in small caps, representing the fact that in the allomorph selection approach there is no single underlying form that we can put as the input to sandhi. Rather, the correct allomorph of the head is selected based on the form of the preceding argument, and here crucially, it is the raised tone of the plural (S in (41a) and L in (41b)) that serves as the sandhi trigger (selecting S- and L-toned allomorphs, respectively, rather than the X we would expect from their singular tones).

However, if the head is plural, tone raising applies to the output of tone sandhi:

Under the allomorph selection account proposed here, the head does not even have a phonological form that plural tone raising could apply to until the correct sandhi allomorph is selected. In (42a), a Class II noun will surface as X-toned after a H-toned argument, thus yielding the form mó cɛ̏; plural tone raising applies to this X tone, raising it to L and yielding the correct surface form mó cɛ̀. The same process applies in (42b), with H triggering the S-toned allomorph kɔ̰̋n of the Class I noun ‘head’. Plural marking applies to this form, yielding the surface form í kɔ̰̋ɛ̰.¹²

For the sake of argument, let us assume that these heads have abstract underlying forms, cɛ́ ‘hand’ and kɔ̰̏n ‘head’ respectively, as first presented in §5.1. If plural applied to the possessed noun before sandhi took place, it would derive cɛ̋, as in (43):

This S-toned input to sandhi, when combined with a H-toned possessor, would then incorrectly derive *mó cɛ̏:

Similarly, in the case of ‘your heads’ above, the plural would derive L-toned kɔ̰̀ɛ̰ if applied directly to the abstract underlying form kɔ̰̏n; unfortunately, since there are no underlying L-toned heads to which sandhi applies, we cannot say what the output would be in that case. Regardless, the surface form í kɔ̰̋ɛ̰ is consistent with the order Sandhi > Plural.

7.2 Cyclic spell-out

While this may seem to present an ordering paradox (Plural > Sandhi > Plural), it is perfectly consistent with a cyclic model of grammar. However, by cyclic, I do not mean cyclic computation within the morphophonological component (e.g. Stratal Phonology, Bermúdez-Otero 1999; forthcoming; Kiparsky 2000), since even the partial order Sandhi > Plural would require a phrasal process (sandhi) to apply before a word-level process (plural formation), running counter to the spirit of these frameworks. Rather, I use cyclic here to mean systemic cyclicity, where linguistic structure is built up incrementally, passing material through each component of grammar in turn. The specific framework I will adopt here is that of cyclic spell-out (Uriagereka 1999; Chomksy 2000), in which chunks known as “phases” are sent off to receive semantic interpretation and morphophonological form at various stages in the derivation. This systemic cyclicity is illustrated in Figure 2.¹³

Figure 2

Cyclic spell-out of grammar.

Under this model, certain syntactic projections (commonly CP, vP, and DP) project phases, with the head of the phrase (C, v, D) known as phase heads.¹⁴ When a phase head is reached in the derivation, its complement (the computational unit, abbreviated here as CU) is sent to spell out, which consists of Morphology, Phonology and Semantics. The output of the phase is reinserted into the syntactic structure with only its phonological form and semantic interpretation; in other words, it crucially loses all internal syntactic structure, which accounts for a range of syntactic inalterability effects (see e.g. Chomsky 2000 on the Phase Impenetrability Condition).

A growing body of literature shows how phase-based spell-out interacts with and constrains morphophonological process application (e.g. Pak 2008; Newell and Piggott 2014; McPherson 2014; McPherson and Heath 2016; Sande 2017, among many others). While the morphological component in (44) is often implemented in Distributed Morphology (Halle and Marantz 1993; Embick and Noyer 2007), there is nothing inherent about this framework to the cyclic architecture of grammar; phase-based spell-out likewise places no restrictions or demands on the phonological component, which could be rule-based or constraint-based, with ranked or weighted constraints.

As the analysis in §6 laid out (and following on work in McPherson in press, a), I argue that the morphological component in Seenku is firmly rooted in the lexicon, with lexical entries and meta-entries capable of generating novel forms. After syntactic structure is built in the syntactic component, the structure with its abstract lexical roots like √HIT is matched with lexical entries, which could be as simple as the Seenku lexical items for a particular root or as complex as a lexicalized phrase consisting of an argument and a head.¹⁵ The forms pulled from the lexicon are then sent on to the phonology, where regular phonological operations like downstep or register assimilation take place. At the end of the cycle, internal syntactic boundaries are erased, and what remains is a chunk of structure with phonological form and semantic interpretation (essentially a lexical compound) that is then reinserted into the syntax and can be passed through subsequent cycles of spell-out.

Let us consider how these pieces combine to allow plural formation to apply to both the input and output of sandhi. First, I assume that D is a phase head in Seenku, which means that when it is reached, its complement (the NP and any other modifiers) is sent to spell-out.¹⁶ An inalienable possessive phrase consists of the possessor’s DP embedded within the DP of the inalienable noun, as shown in (21). The NP (the complement of the phase head) of the more deeply embedded possessor will be sent to spell-out first and will receive its morphophonological form. This form will then be reinserted into the syntax and will undergo spell-out a second time when the higher DP is reached, triggering the spell-out of the NP headed by the possessed noun. We can illustrate this spell-out process with the phrase bɛ̀ɛ ɲà ‘the pigs’ mother’:

I assume that the plural has its own lexical entry as a suffix consisting of two featural affixes, [+raised] and [+front]. Thus, when the embedded D is reached, the NP dominating PIG_PL is sent to spell-out. There, PIG matches the lexical entry bɛ̏ɛ and the plural matches the lexical entry {[+raised][+front]}; when combined, the form /bɛ̏ɛ [+raised, +front]/ is sent to the phonology, which governs the docking and realization of the floating features (McPherson 2017a). The form [bɛ̀ɛ] is then reinserted into the syntactic structure under NP, which no longer has any internal syntactic structure (i.e. if there had been any word boundaries in a more complex NP, they would be erased at this stage). When the higher D phase head is reached, the NP headed by MOTHER is sent to spell-out, taking with it the spelled-out material [bɛ̀ɛ] as its complement. MOTHER is an inalienable noun, and thus its lexical entry includes not only itself but also the tone of the preceding argument, which determines the allomorph. The L-tone on [bɛ̀ɛ] triggers the use of the L-toned allomorph of MOTHER, ɲà. This form bɛ̀ɛ ɲà is then sent to the phonology, which leaves it unaltered on the surface. Thus, here, plural applies before sandhi by virtue of the fact that the possessor has already passed through Morphology and Phonology on a previous cycle of spell-out.

This architecture likewise explains how plural heads show the application of sandhi before plural formation. Consider the expression mó cɛ̀ ‘my hands’ in (42a) above, with syntactic structure as in (46):

The possessor DP for the 1SG is more deeply embedded and it defines the phase sent to spell out first, where the lexical item mó is selected. This form is then reinserted into the syntax. When the higher DP containing HAND_PL is reached, the NP is sent to spell-out with mó, where the lexical entry for ‘hand’ designates the allomorph cɛ̏ for use after a H-toned argument. This is passed to the phonology with the plural morpheme as /mó cɛ̏ [+raised, +front]/, where [+raised] raises X to L. In other words, there is no opportunity for the plural to act on an “underlying” form of ‘hand’ before sandhi applies, since the very act of selecting an entry from the lexicon requires sandhi to apply.

The crucial facts of this analysis are simply that the grammar is cyclic, and that this cyclicity is not confined to a single component like the morphophonology; if it were, it would create an ordering paradox and also complicate the order of application between word-level and phrase-level processes. The phase-based approach I employ here avoids these issues through systemic cyclicity, allowing multiple passes through each component of grammar as the structure is built up. In principle, any framework with this property would fare equally well. For instance, the data could be accounted for with a morphophonological grammar that works on a linear string from left to right. The leftmost word would not have a sandhi form, since it is not preceded by a sandhi trigger, so it would be evaluated for plural and any other morphophonological processes, then the second word would first be evaluated for its sandhi form based on the preceding tone, before the application of plural, and so on and so forth. This is schematized in (47):

This sort of linear application of morphophonology is reminiscent of recent work in computational phonology implemented using finite state transducers (e.g. Chandlee 2014; Chandlee, Eyraud, and Heinz 2014; Heinz 2018, etc.), and left-to-right linear structure building has been proposed in the realm of morphosyntax as well, e.g. in works like Phillips (1996; 2003). If we assume that each new word in the derivation first retrieves its appropriate allomorph from the lexicon, along with any other morphemes corresponding to its morphosyntactic features (e.g. plural) and then is sent on to phonological computation, then this strictly linear model would work equally well. Since I know of no cases of long distance or non-adjacent application of sandhi, the two models would make the same predictions. Further evidence, perhaps from purely syntactic phenomena, are required to test the predictions of each the linear versus phase-based models.

What is crucial for Seenku argument-head tone sandhi is that the grammar as a whole is cyclic.

7.3 Recursive sandhi domains

The cyclic approach laid out in §7.2 also naturally accounts for the recursive application of sandhi. Sandhi domains can be recursive, through recursive possession and/or having possessive DPs as objects or complements of a postposition. In both of these constructions, more deeply embedded DPs are on the left, with heads stacking out towards the right. Sandhi applies from left-to-right, which in this case matches the depth of embedding (inner to outer):

As with the interaction of the plural and sandhi in §7.1, it is difficult to say what right-to-left application would look like, since the heads require the input of the argument’s tone in order to make an initial allomorph selection. Thus, in (48a), the H-toned pronoun mó triggers the selection of ɲȁ for the Class II noun ‘mother’, which in turn triggers the selection of bȍn for the Class III noun ‘back’. In (48b), the X-toned pronoun ȁ triggers the selection of the H-toned allomorph bón for the same Class III noun ‘back’, which triggers the X-toned allomorph nɛ̏ for the Class II locative postposition. Finally, in (48c), the initial argument-head combination is the same as in (48a), but the X-toned allomorph of ‘mother’ ɲȁ triggers the selection of the L-toned allomorph sà̰ of the Class I verb ‘buy’, which in turn raises the tone of ‘mother’ to L (see §9.2).

If we gave these heads abstract underlying forms, as in the table in (9) earlier in the paper, we would see that right-to-left application of sandhi would result in incorrect surface forms. For instance, in (48b), right-to-left application of sandhi in an abstract underlying sequence /ȁ bőn nɛ́/ would yield incorrect *ȁ bón nɛ̋, while the same order of application in (48c) on the abstract sequence /mó ɲá sȁ̰/ would yield *mó ɲȁ sa̰̋. Left-to-right application both predicts the correct forms and avoids the need for abstract underlying forms.

These results follow directly from either of the cyclic approaches laid out in the last subsection. To take (48a) as an example with the phase-based approach, the possessor mó is the most deeply embedded DP, and hence the first to be sent to spell-out, where /mó/ is selected from the lexicon (or perhaps /ń wó/, which is transformed to [mó] in the phonology). This is reinserted into the syntax, and at the next deepest DP, mó + MOTHER is sent to spell-out, where the X-toned allomorph of ‘mother’ is selected since it is preceded by a H-toned argument; the output of this cycle is [móɲȁ], which is reinserted into the syntax (without its internal boundaries). Finally, the highest DP level is reached, and móɲȁ + BACK is sent to spell-out, where the X-toned allomorph of ‘back’ is selected because of the preceding X-final argument. The linear model predicts the same results; in other words, the application of recursive sandhi could be viewed as either a consequence of the syntactic structure or of the linear string passing through a left-to-right cyclic grammar.

7.4 Complex heads

The recursive domains in the last subsection stacked sandhi domains by having the argument of a head be itself an argument + head combination. In this configuration, the left-to-right directionality mirrored the branching, meaning it could be naturally accounted for by either model of systemic cyclicity (phase-based or linear). However, it is also possible to have complex heads, and in this case, the bracketing would be [ X [ Y Z ] ] instead of [ [ X Y ] Z ]. In what order or direction would sandhi be applied?

The linear left-to-right account would of course predict left-to-right application, regardless of branching structure. While the phase-based account might seem to predict the opposite, recall that the compound noun would be sent to spell-out in the same cycle, together with the previously spelled out argument, leaving the morphological component to select the correct allomorphs corresponding to the syntactic structure in the same pass. Let us consider a concrete example with the compound noun ‘chest’ (lit. ‘liver-front’):

Under cyclic spell-out, the DP containing the 3SG possessor defines the first phase, and 3SG would spell out first, receiving the morphophonological form ȁ. This would be reinserted into the syntax, leaving the following:

When the higher DP is reached, the NP consisting of {ȁ LIVER FRONT} is sent to Morphology, where the lexicon looks to match the structure with lexical entries. As a compound noun, ‘chest’ is likely to be lexicalized rather than computed fresh each time a speaker wishes to utter it. And since the compound is an inalienable noun, as indicated by the syntactic structure, it too is likely to have listed allomorphs for different contexts (brɔ̰̋n-je̋n after S, brɔ̰̀n-jèn elsewhere). In the context of the 3SG ȁ, brɔ̰̀n-jèn would be selected (causing ȁ to raise to à in the phonology; see §9.2 below). Lexical listing of compounds would likewise account for exceptional compounds like kɔ̰̏n-kɔ́ ‘skull’ that resist sandhi, as seen in §5.2.2.

Even if the tonal form of the compound were computed online from individual lexical entries, the result would be the same. For instance, the lexical entry of LIVER would be activated, with the correct allomorph chosen based on the tonal context to the left regardless of what tone follows; this would yield brɔ̰̏n. The second stem is likewise inalienable, and Morphology is unable to select the correct allomorph until it has a tonal context to its left, requiring the leftmost stem to be computed first; once the allomorph brɔ̰̏n is selected, then jèn can be inserted for ‘front’, and this sequence X X L would be L L L in the phonology (§9.2).

Thus, left-to-right application of sandhi with complex heads is consistent with either strictly linear cyclicity or with a phase-based approach, precisely because the lefthand context must be defined before the correct allomorph can be selected. The only difference is that the phase-based model predicts that spelled out material is resubmitted to phonological computation at each cycle, if it is embedded in a phase sent to spell-out, whereas the linear model may have to stipulate this separately in order to allow the later raising of X to L. I return to this point in §9.2.

8.1 Phonological approach

Those who advocate for an allomorph selection approach to Taiwanese tone sandhi state that the system is simply too opaque and phonetically unnatural to be a phonological in nature (see e.g. Moreton 1999 on the non-computability of circular chain shifts in OT), but phonological analyses have been proposed. Hsieh (2005) and Barrie (2006), for instance, use an anti-faithfulness constraint (Alderete 2001) ¬IDENT-IO(T) to motivate the sandhi alternations; in other words, tones should not remain the same as their underlying form in sandhi domains. To avoid all tones neutralizing to some unmarked sandhi tone, contrast preservation (Lubowicz 2003) could be invoked. Set up correctly, this could go so far as to motivate a circular chain shift, but it requires the grammar to evaluate entire scenarios comparing input-output pairs rather than taking forms individually, adding a layer of complexity to the computation of surface forms.

For Seenku, we would additionally need a slew of markedness constraints to force sandhi tones away from unattested sequences, like *X S to force lowering after X, or *S¬S to drive the S spreading pattern. The tone features [upper] and [raised] could likewise be invoked to force agreement between arguments and heads, such as agreement in [raised] between X and H (as opposed to between X and L, or X and S), which could help provide phonetic motivation for some of the avoided sequences. Note, however, that there is no independent motivation for a markedness constraint like *X S, as this sequence is tolerated in other environments.

As shown in McPherson (in pressa), even this complicated phonological grammar, with a whole toolkit of enhancements, cannot distinguish between the attested tone circle (S→H→X→S) and its reverse. In the interest of space, I do not lay out tableaux here, but refer readers to McPherson (in pressa). Even beyond the issue of motivating the correct alternations, this approach would suffer from further problems, including lexical exceptions, the imperviousness of multi-tonal heads, and the ordering between sandhi and plural formation, since now a phonological process (sandhi) would need to precede a morphological one (plural), even within a single cycle. That is, a form like mó tòee ‘my ears’ would need the phonology to choose the sandhi form |tȍo|, to which plural tone raising would apply. To maintain the phonological nature of sandhi, cyclicity would have to be introduced into the morphophonology to allow sandhi to precede plural formation, but this would mean a phrasal process would occur before a word-level process, which is far from the spirit of lexical phonology and morphology (Kiparsky 1982; 2000, etc.).

The allomorph selection approach advocated for here avoids this overly baroque phonological architecture while at the same time naturally accounting for process ordering and exceptions.

8.2 Co-phonologies by phase

A middle ground could be found between this purely lexical approach and a purely phonogical approach. Recent work by Sande (2017; 2018) and Sande and Jenks (in press) take a similar approach to the (morpho)phonology-syntax interface, with the timing and domains of process application defined by cyclic spell-out, and lexically-, morphologically-, and syntactically-conditioned phonology implemented in co-phonology theory (Orgun 1996; Anttila 2002; Inkelas and Zoll 2005). This hybrid framework, Co-phonologies by Phase (CBP), formally implements construction-specific phonology by associating partial rankings of phonological constraints with particular lexical items (or Vocabulary Items, in the Distributed Morphology terms used in the approach). When a phase is sent to spell-out, any partial rankings of lexical items within the phase are percolated up to the top of the phase, and this new ranking is used to compute the phonological form of the spelled out material.

We would first have to decide where to associate the co-phonologies: to the inalienable nouns, post-positions, and verbs themselves? To the arguments? Or to some structural property of the construction, like an irrealis projection for verbs or some sort of inalienable possessive projection for nouns? The benefit of associating the co-phonologies with lexical items is that the entire system, with its circular chain shifts, need not be accounted for in the same phonological grammar. There could be a grammar for underlying X tones, which causes them to raise to L or S, another grammar that polarizes H to X or S, etc. The issue with associating partial rankings with the lexical items themselves, however, is that they may not always find themselves in sandhi domains. This is especially true for regularly alienable nouns that wind up in dative possessive constructions, where they undergo sandhi (§5.2.3). The problem is worse if we associate the co-phonologies with the arguments, which likewise do not always trigger sandhi on a following word. If we associated the co-phonologies with a structural aspect, such as irrealis, we are left with the same complicated phonological grammar as we saw in §8.1.

The other issue with the CBP approach, assuming we could find the correct locus for co-phonologies, is that the partial rankings are inherited by the whole phase and apply to that domain. As we saw in §5.2.2, there are other elements in the spell-out domain of a sandhi undergoer that resist sandhi; adjectives and numerals, for instance, also occur within the DP but they resist sandhi. In other words, sandhi domains are not co-extensive with spell out domains in Seenku. Thus, we need an approach that allows us to target specific heads and not an entire spell out domain, which the allomorph selection approach allows us to do.

8.3 Distributed Morphology

As shown in §7.2, some assumptions of the analysis, such as Syntax before Morphology, are consistent with Distributed Morphology (Halle and Marantz 1993; Embick and Noyer 2007, etc.). Indeed, DM likewise allows for the listing of allomorphs in different subcategorization frames (phonological, morphological, or syntactic). However, in order for allomorph selection to work in DM, every noun, verb, and postposition would need to have listed allomorphs, since the Vocabulary in DM is nothing but a list of entries (Vocabulary Items); in other words, the lexicon in DM is not generative and cannot account for the productive extension of sandhi patterns if listed allomorphs are responsible.

For sandhi to productively extend to novel combinations or novel forms in DM, the tonal alternations would need to be the product of morphophonological rules applying between vocabulary insertion and the regular phonological component. This would require positing abstract underlying forms for sandhi undergoers (and diacritics for exceptional forms that resist sandhi). Taking the idea of “rule” literally, we could posit SPE-style rules like S → H/X__, but if we assume some form of constraint-based phonological grammar (as many do in current phonological theory), this has the consequence of requiring two kinds of phonological computation (rules and constraints). If morphophonological rules are instead sets of ranked or weighted constraints, then we run into the same problems as we saw in the phonological approach in §8.1.

By giving the lexicon greater power to generate novel forms through the use of meta-entries, we avoid both the problem of having to list every allomorph for every possible sandhi undergoer and the issue of having to compute tonal alternations by rules or constraints.

9.1 Productivity: evidence from alienable nouns

Recent studies of Chinese sandhi systems have been exploring the productivity of sandhi through nonce word (wug) testing (see e.g. Hsieh 1970; 1975; 1976; Zhang and Lai 2008; 2010; Zhang, Lai and Sailor 2009; 2011; Chuang, Chang, and Hsieh 2011; Zhang 2016; Zhang and Meng 2016, etc.), with varying results. For instance, studies on Taiwanese Min tone sandhi have shown only limited productivity of the opaque tone circle (Zhang, Lai and Sailor 2011), while in Mandarin, the phonetically natural half-third tone sandhi was found to be more productive than the less natural third tone sandhi (Zhang and Lai 2010). We could argue that the productivity of a system should be directly inverse to its degree of lexicalization.

It would be enlightening to study the productivity of Seenku tone sandhi in this way, but unfortunately the methods do not translate well to the rural field setting where Seenku is spoken. First, Seenku is an unwritten language, so prompts would have to be oral. Unlike in the Chinese languages of these studies, there is often not a citation form that avoids sandhi application that could be presented orally. And finally, I have found wug testing to be a challenge in West African field situations, where speakers are unsure of what they are being asked to do or unwilling to consider the notion of a non-word (“But that’s not a word!”).

The closest we can get to wug testing Seenku argument-head tone sandhi is to apply it to words that do not regularly undergo sandhi, i.e. alienable nouns in the dative possessive context. I take this construction to be an extension of the regular inalienable construction (a sandhi environment) to nouns that would not otherwise have lexicalized allomorphs. In other words, the lexical entry for a word like ‘dog’ would be simply <DOG (n.) jəgȅ >, the independent form, rather than having listed allomorphs.

In the framework proposed here, the lexicon is a network of interrelated constructions, which may activate one another. We have seen the lexical entries for individual lexical items and their allomorphs, but there are likely higher level phrasal constructions for phrases with idiosyncratic properties (semantic idiosyncrasies in the form of idioms, Booij 2010b, or phonological idiosyncrasies, as in Dogon tonosyntax, McPherson 2014). Dative possession is likely to be one of these idiosyncratic constructions, treated like inalienable possession by the morphophonology despite taking alienable nouns (and no reported difference in meaning from the regular genitive alienable construction). I suggest that this construction might activate the paradigms for inalienable nouns with their sandhi alternations, which override the lexical tone of the alienable nouns. This configuration is schematized in (51):

In the phonological component on the left, the lengthened dative postposition lɛɛ is unspecified for tone, since it will select its sandhi allomorph by virtue of being a postposition. The possessed noun, ω_k, is linked in the lexicon to the constructional templates for inalienable nouns, classes I, II, and III.

The one potential snag in assigning an alienable noun to the correct class returns to the question of underlying tones. Recall that inalienable nouns never surface without sandhi, so any underlying form is necessarily abstract. There are two possible solutions to this problem: the first is that there may be enough common alienable nouns lexicalized with sandhi variants (‘thing’, ‘person’, ‘woman’, ‘man’, etc.) that these can provide the clues for class membership. The second is that the networks of sandhi forms in the lexicon extend beyond inalienable nouns and include postpositions and verbs as well, the latter of which provide better clues to an underlying or citation form, since they do not undergo sandhi in realis enviroments.

Returning to the question of productivity, we could use the measure of how many alienable nouns have this pattern extended to them as a rough approximation. More systematic testing is required, but based on data gathered so far, I have found no exceptions among monotonal native Seenku roots (like bɛ̏ɛ ‘pig’, cî /cí/ ‘house’, mənı̋ ‘woman’, etc.); all of these follow the expected pattern of sandhi. The only avoidance I see of this construction is when sandhi would neutralize the distinction between singular and plural (e.g. móò bɛ̀ɛ ‘my pig(s)’), in which case speakers will sometimes revert to the genitive plural where the two are distinct (e.g. mó tȅ bɛ̏ɛ ‘my pig’ vs. mó tȅ bɛ̀ɛ ‘my pigs’). The other area of interest would be loanwords, which would be less likely to have multiple lexicalized allomorphs. While we could design a study asking speakers to produce possessive expressions with loanwords, there would likely be a priming effect, such that once speakers start producing sandhi variants in this context they are likely to continue producing them. More telling would be a corpus study, looking at natural behavior. The current corpus does not contain enough examples of this possessive construction, especially with loanwords, to draw any conclusions. I leave productivity testing to future work.

9.2 Raising of the argument to L

The final question to address is the mechanics of the retroactive change of X-toned arguments (which have already spelled out) to L before an X-toned head. We can say with certainty that this is not the result of a local change, such as a left-docking floating [+raised] feature from the L-toned sandhi variant, since X to L raising effects the entire domain. This is the case even in recursive sandhi domains, where multiple preceding morphemes in a complex argument all raise to L:

There is reason to believe that this is not a lexical/morphological process but rather a more general phonological one sensitive to phonological phrasing. For instance, we find the same X to L raising of possessive arguments when the head is marked for plural, a tone raising process that I have argued to be implemented in the phonology by docking a floating feature:

In (53a), we can see that the sandhi allomorph of breast is X-toned, and the possessor likewise retains its X. When [+raised] docks to this X tone as in (53b), the effect of raising is felt in the possessor as well, despite the fact that [+raised] of the plural usually just affects the final tone (i.e. it is a suffix; see McPherson 2017a). While this looks like a case where perhaps the X tone of the possessor has spread onto the head noun, leaving just a single X autosegment to raise with the addition of [+raised], other sandhi combinations do not involve spreading (H-X, X-H, etc.), so I hesitate to posit phonological spreading in only those cases where the tones are identical.

In fact, evidence suggests that this raising is a more general phonological process in the language, targeting phonological phrases with the underlying sequence X-L. In the two examples in (54), we see a definite article ȁ raising to L before a L-toned plural noun, as well as a singular noun jȕ ‘year’ raising to L before a L-toned numeral:

The definite article ȁ is X-toned in front of all other tones (ȁ bɛ̏ɛ ‘the pig’, ȁ bî ‘the goat’, ȁ mənı̋ ‘the woman’), but with a following L created by plural formation, it raises. In (54b), it is tempting to view the L on jù as being the result of plural formation—the phrase, after all, has a plural meaning. But nouns are not obligatorily plural before cardinal numerals, and further, the plural form of ‘year’ involves vowel fronting: jùi ‘years’. Thus, this L tone is likely phonological in nature, the result of a markedness constraint against the sequence X-L within a phonological phrase. Note that the process must be phrase-bounded, since we do not see raising between e.g. a subject and an intransitive verb (see for example (32a)).

One last point here deserves mention: the fact that previously spelled out possessive DPs can raise from X to L, one or more cycles later, shows that there can be no “phonological impenetrability condition”. In other words, spelled out material can be altered phonologically to meet phonological demands. This is contra previous work, such as Piggott and Newell (2006), Dobler (2008), or Lowenstamm (2010), where the phase impenetrability condition for syntax is extended wholesale to phonology. Rather, I assume, following Michaels (2013), Šurkalović (2013), McPherson (2014), and McPherson and Heath (2016), that the tension between phases apparently blocking the application of morphophonology and cases like Seenku where they do not can be resolved with a violable output-to-output faithfulness constraint IDENT(phase), which languages can rank or weight in different ways. In Seenku, it must be lowly ranked.