2.1 Rotuman phase alternations

The alternation between ‘complete’ and ‘incomplete’ phases (Churchward Reference Churchward1940, Besnier Reference Besnier1987, Blevins Reference Blevins1994, Hale & Kissock Reference Hale and Kissock1998, McCarthy Reference McCarthy2000) involves metathesis (1a), as well as deletion (1b) and umlaut (1c). I will analyse all three patterns as split and fuse, which in the cases of (b) and (c) are then followed by further fusion if the sonority of the final vowel is not greater than the sonority of the penultimate vowel.

1. (1)

   

While differences between these two types of phases are considered to be morphological or syntactic by Churchward (Reference Churchward1940) and McCarthy (Reference McCarthy1995), among others, McCarthy's (Reference McCarthy2000) analysis of phase alternation in Rotuman assumes that the alternations are phonological phenomena (also see Hale & Kissock Reference Hale and Kissock1998). Complete phase forms are generally found preceding monosyllabic suffixes or clitics, while incomplete phase forms are used before polysyllabic suffixes or clitics. Although it is intriguing to contemplate the incorporation of these differences in suffix length in the analysis, for the sake of simplicity I examine the words in the two phases in isolation. I also follow Hale & Kissock (Reference Hale and Kissock1998) and McCarthy (Reference McCarthy2000) in assuming that both phases are derived from an underlying representation that is identical in relevant respects to the surface form in the complete phase.

Since Rotuman stress is trochaic, and the incomplete phase always ends in a stressed syllable, stressed syllables in incomplete phase words are always heavy, so as to satisfy the Stress-to-Weight principle (SWP; Prince 1990). I assume that coda consonants are moraic (see McCarthy Reference McCarthy2000 for arguments for assigning a mora to codas), making CVC syllables heavy.

If the last two vowels in a word have rising sonority, the last CV in the word undergoes metathesis. When the two vowels do not rise in sonority, either the last vowel does not surface in the output (apparent deletion), or the two vowels coalesce (umlaut), depending on the vowels involved. The umlaut cases are observed only with sequences of back + front vowels.

2.2 Proposal

I propose that all of the alternation patterns result from split and fuse operations driven by the constraints in (2a) and (b), SWP and Final Stress (Inkelas Reference Inkelas, Kager, van der Hulst and Zonneveld1999). Having these markedness constraints dominate Integrity and Uniformity in (c) and (d) allows an intermediate stage with two identical vowels straddling the stressed vowel, although this form does not surface as the final output.

1. (2)

   

The derivations presented here assume that stress is first assigned by high-ranked prosodic constraints (not shown in this section). With the constraint SWP undominated, the first step of the derivation (after stress assignment) outputs the intermediate winner /ˈpu₁re₂/ → [ˈpu₁e₂re₂] through vowel splitting, as shown in (3). This is an operation which some have previously seen as copying. However, to make it clear that the constraint violated by the operation is Integrity, not Dep, I will use the term splitting. I assume Max is also undominated in this language, prohibiting deletion of a segment (apparent ‘deletion’ cases in the incomplete phase will be discussed below). Vowel splitting violates Integrity (McCarthy & Prince Reference McCarthy and Prince1995), which prohibits an input segment from having multiple corresponding segments in the output. I assume that, although Integrity bans coindexed output segments that have a single corresponding segment in the input, the second appearance of the segment does not violate general Dep, since the segment has an input correspondent (Staroverov Reference Staroverov2014). Although (3b) violates FinalStress, because the stressed syllable /ˈpue/ is not the final syllable of the word, it still emerges as the winner at this stage, since the stressed syllable is heavy and the SWP requirement is met without deletion of the final V.Footnote ³

1. (3)

   

In the next step, in (4), [ˈpu₁e₂re₂] → [ˈpu₁e₂r], the faithful candidate [ˈpu₁e₂re₂] loses, due to its violation of FinalStress. The fusion of the coindexed vowels does not violate Max, since each input segment coindexed with ‘2’ does have a corresponding segment in the output (for correspondence involving a copied segment, see Stanton & Zukoff Reference Stanton and Zukoff2016). I assume here that the winning candidate created by fusion violates Uniformity, since the two coindexed segments in the input have one corresponding segment in the output (McCarthy & Prince Reference McCarthy and Prince1995).

1. (4)

   

McCarthy (Reference McCarthy2016: 62–63) questions the need for indexation in Harmonic Serialism. It is generally assumed that the operation at each stage can only ‘see’ the input at the current stage of the derivation. Utilising indices throughout the HS derivation diverges from common practice in standard HS (see McCarthy Reference McCarthy2016). However, maintaining faithfulness to the UR throughout the derivation in HS has been proposed previously (e.g. Hauser et al. Reference Hauser, Hughto and Somerday2015). Along with the perspective that additional faithfulness might still be required in HS, I postulate that split (copied) vowels have a special status, and are marked as copies of each other. This assumption is at the crux of the analysis here.

Apparent deletion cases such as /ˈra₁ko₂/ → [ˈra_1,2k] can also be analysed as split and fuse, followed by further vowel fusion. The proposed derivational path is shown in (5), which employs an adapted version of the HS tableaux presented in Pruitt (Reference Pruitt2012). The second column shows the derivation from the UR to the final output, with the arrow in each row representing the winner at each step. Only constraints crucial to the derivation are given in (5). LightDiph bans a VV sequence of non-rising sonority in a closed syllable (McCarthy Reference McCarthy2000). Although fusing two vowels together violates Uniformity, the constraint ranking derives [ˈra_1,2k] with a fused vowel as the most harmonic output.

1. (5)

   

Support for the fusion analysis comes from umlaut cases such as /ˈho₁ti₂/ → [ˈhø_1,2t]. The steps shown in (6) proceed through the same stages as the derivation of /ˈra₁ko₂/ to [ˈra_1,2k] in (5). The only difference is the change in vowel quality observed in the final output.

1. (6)

   

The three phase-alternation patterns have been analysed in the literature as involving three different processes (metathesis, deletion and umlaut), although some analyses involve copy and deletion (Chandlee Reference Chandlee2014). What is novel in the current analysis is the unification of metathesis, deletion and umlaut as all undergoing the same split and fuse operations.

While there is a strong appeal to such unification, the final vowel fusion in (6) (/o₁i₂/ → [ø_1,2]) requires further investigation. While Fallon (Reference Fallon1998) suggests that fusion is a primitive operation, others consider it to be the result of assimilation followed by deletion (e.g. Aoki Reference Aoki1974). As we have seen, high-ranked Max eliminates the possibility of analysing fusion as assimilation and deletion. However, analysing fusion as an operation is problematic in HS, since fusing two segments into one violates more than one Ident constraint. As McCarthy (Reference McCarthy, Blaho, Bye and Krämer2007b) notes, if we assume that Gen can only perform one unfaithful mapping at a time, a fused vowel cannot appear as a candidate in a HS derivation. Although detailed analysis of vowel fusion in HS requires further investigation, it is possible to view fusion in Rotuman /o₁i₂/ → [ø_1,2] as a gradual series of assimilation operations, followed by fusion of identical vowels, as in (7). In this approach, only one faithfulness violation is introduced at each step. Since vowel fusion in split and fuse affects a coindexed identical vowel, limiting it to identical vowels ties together the two types of vowel fusion explored in this section (one being part of metathesis and the other part of vowel coalescence). How fusion should be analysed in the HS framework has not been examined rigorously in the literature (McCarthy Reference McCarthy2016: n. 8). It has therefore been tacitly assumed that fusion is an operation that violates Uniformity, an assumption that I adopt for the most part in this paper. However, a series of assimilations followed by fusion of an identical segment is another way to incorporate the fusion of two segments in HS.

1. (7)

   

Another potential concern regarding the present analysis is the use of coindexation in the HS framework. It might be argued that coindexation requires keeping each step in memory. However, the current analysis requires output coindexation, rather than input–output indexation. As noted in McCarthy (Reference McCarthy2016: 62), faithfulness constraints can look back to the input of the most recent application of Gen; in order to create candidates, the most recent input still has to be referred to. Therefore, coindexation simply indicates that the segments involved had the same input at the previous stage. The correspondence between coindexed elements is important at the splitting stage (see Rose & Walker Reference Rose and Walker2004, Stanton & Zukoff Reference Stanton and Zukoff2018). These coindexed elements are then evaluated at the next step, but there is no need for the system to remember the input/output correspondences at each step. (8) shows the schematic representation of coindexation. This is almost identical to the representation in Stanton & Zukoff (Reference Stanton and Zukoff2018), even though their analysis is carried out in parallel OT. At step n, one input vowel, Vx, is split into the two coindexed output vowels. Since the two output vowels maintain their correspondence with the input, there is no Dep violation. Similarly, at step n+1, both of the input vowels correspond to the output vowel, and thus, since the Vx in the output has a correspondent segment in the input, general Max is not violated.Footnote ⁴

1. (8)

   

2.3 Parallel OT analyses

In contrast to the analysis outlined above, McCarthy's (Reference McCarthy2000) parallel OT analysis of Rotuman phase alternation depends on the ranking LightDiph, FinalStress ⪢ Max ⪢ Linearity to account for both metathesis and deletion cases. A summary tableau, adapted from McCarthy (Reference McCarthy2000: 175), is shown in (9).

1. (9)

   

However, as pointed out in Carpenter (Reference Carpenter, Carpenter, Coetzee and de Lacy2002), this constraint ranking does not rule out a problematic candidate such as [ˈro₂a₁k], with long-distance metathesis, as shown in (10). The intended winner, [ˈra₁k], loses to candidate (e). Despite the two violations of Linearity, [ˈro₂a₁k] does not violate the markedness constraints or Max.

1. (10)

   

Consideration of candidates such as (e) leads Carpenter (Reference Carpenter, Carpenter, Coetzee and de Lacy2002) to conclude that Linearity alone cannot block reordering of segments that span a certain distance. She proposes a family of Adjacency (domain) constraints, i.e. faithfulness constraints that require ‘segments or features to maintain, in the output, the immediate proximity found in the input’ (2002: 5). This specified domain treats various metathesis types differently. For instance, Carpenter utilises I-Adjacency(σ), a constraint which requires that segments within an input syllable remain adjacent to each other in the output, to account for the Rotuman metathesis patterns, successfully eliminating any non-adjacent metathesis candidate that does not surface in this language. The non-adjacent metathesis form in (10e) thus incurs two violation marks in (11e), since the adjacency between /r/ and /a/ and between /k/ and /o/ in the input is not maintained in the output (Carpenter assumes that the alternate form [ˈra₁k] in the incomplete phase is derived from the complete form [ˈra₁.ko₂]).Footnote ⁵

1. (11)

   

The parallel OT analysis works by adopting a family of faithfulness constraints, without which it would be difficult to block unattested patterns. In the HS split and fuse analysis proposed here, however, splitting is only harmonically improved after stress assignment, in order to satisfy SWP. Splitting a vowel such that the two identical vowels straddle the stressed vowel (ˈV₁V₂ → V₂ˈV₁V₂), as in /ˈra₁ko₂/ → [ro₂ˈa₁ko₂] → [ro₂ˈa₁k] does not improve the weight of the stressed syllable, since [o₂ˈa₁] is not a well-formed heavy syllable. This type of vowel-over-vowel splitting is correctly blocked by the proposed ranking for Rotuman, but it may be universally prohibited, since it is cross-linguistically unattested.

The analysis of the Rotuman phase alternations provided in this section has shown that transposition is not needed as an operation in HS, and seemingly different alternation patterns all result from sequential split and fuse operations, as long as we allow coindexation. Unlike McCarthy's (Reference McCarthy2000) parallel OT analysis, in which Max has to be dominated by LightDiph, the ranking is crucially reversed in the HS analysis proposed here. Furthermore, a long-distance metathesis candidate such as *[ro₂a₁k], which necessitates the proliferation of a family of adjacency constraints in parallel OT, does not surface in the split and fuse HS analysis, because of the gradualness of Gen.

3.1 Kwara'ae metathesis

Kwara'ae exhibits multiple metathesis in which sets of adjacent segments are transposed (e.g. /ketalaku/ → [ˈkeat.ˌlauk] ‘my height’; Heinz Reference Heinz2004, Reference Heinz2005a, b). The Kwara'ae pattern is unique, in that attested metathesised words contain up to three metathesis loci (e.g. /raɁeraɁenaɁa/ → [ˈraeɁ.ˌraeɁ.ˌnaaɁ] ‘incline, slope’). A parallel OT analysis of Kwara'ae metathesis is offered by Heinz (Reference Heinz2005a), with a family of Contiguity constraints. As I will show in this section, the attested patterns can be correctly derived by assuming split and fuse operations in HS.

Kwara'ae uses two different speech registers, ‘citation’ and ‘normal’. Forms in the two registers exhibit different phonotactics and stress patterns. In the normal register, vowel hiatus is prohibited, while a tautosyllabic vowel sequence is common. Onsets are required for non-word-initial syllables. The initial syllable of a word in the normal register always has primary stress, while the syllable containing the penultimate mora has secondary stress. The stress pattern observed is generally trochaic, but heavy–light disyllabic words are not found. The phonotactic differences between the two registers can be observed in the examples in (12), from Heinz (Reference Heinz2004).

1. (12)

   

I will follow Heinz (Reference Heinz2004) in assuming that citation and normal forms are derived from the same UR, and that this UR is generally identical to a stressless version of the citation form. Normal forms displaying metathesis are shown in (13).Footnote ⁶

1. (13)

   

3.2 Proposal

I demonstrate that, like the alternations in Rotuman, Kwara'ae metathesis involves split and fuse. The analysis of Kwara'ae metathesis as involving similar multiple operations is not new (see e.g. Blevins & Garrett Reference Blevins and Garrett1998, Chandlee Reference Chandlee2014). However, the complex multiple metathesis patterns have not been analysed in the Harmonic Serialism framework. I propose that the Kwara'ae metathesis patterns also result from split and fuse, but unlike in Rotuman, the patterns are driven by SWP and FootForm constraints.Footnote ⁷ The analysis is motivated by the following factors: (i) syllables that undergo apparent metathesis in the normal form are almost always heavy, (ii) feet are trochaic (Norquest Reference Norquest2001) and (iii) heavy–light disyllabic words do not occur.

Trochaic stress assignment takes place before splitting, since without stress assignment, there is no motivation for splitting. Subsequently, the unstressed vowel is split into two halves, with one appearing in the stressed syllable. The two split halves fuse back in the stressed syllable at the next step, as illustrated in (14).

1. (14)

   

The derivation of the Kwara'ae pattern becomes possible with the definition of the prosodic constraint *HL in (15a) and an Integrity constraint in (15b), *V_αV_α, which militates against the creation of two identical adjacent segments by splitting.

1. (15)

   

The derivational steps are given in (16).Footnote ⁸

1. (16)

   

The motivation for *V_αV_α comes from the following observations. All content words in in Kwara'ae are minimally CVV. The underlying CV sequence surfaces as a lengthened vowel CV_xV_x. However, a stressed vowel is not lengthened to make the syllable heavy when there is another vowel in the adjacent syllable that can split (/ˈsa₁lo₂/ → [ˈsa₁o₂lo₂], not *[ˈsa₁a₁lo₂]). I assume that *V_αV_α is responsible for this, and that violating it is only permitted if there are no other vowels to be split so as to satisfy word-minimality (similar patterns are observed in Rotuman). In other words, as a repair to satisfy SWP, splitting a vowel across a consonant is preferred to lengthening the vowel (i.e. splitting to create V_αV_α).

3.3 Kwara'ae multiple metathesis

The split and fuse analysis provided above also applies to more complex Kwara'ae cases. The unique aspect of the apparent metathesis patterns in Kwara'ae is that longer words can contain up to three loci of apparent CV transpositions, as shown in (17) (from Heinz Reference Heinz2004).

1. (17)

   

As stated above, I assume that stress assignment precedes split and fuse because of the higher ranking of the prosodic constraints in (18), which assign foot structure. This ranking will parse a bimoraic trochaic foot at the right edge, and then assign a main stress at the left edge. I propose that the apparent metathesis occurs in each foot, both to satisfy SWP and to avoid HL feet.

1. (18)

   

The derivations in (19) show pre-metathesis footed intermediate winners derived by the prosodic constraints. In the case of (c) /ketalaku/ → [(ˈketa)(ˌlaku)] and (d) /daroɁanida/ → [(ˈdaro)Ɂa(ˌnida)], each footed (ˈCVCV) undergoes split and fuse, one by one. The correct surface form is derived, regardless of which foot undergoes the process first.Footnote ⁹

1. (19)

   

Restricting the splitting domain to be foot-internal is crucial for deriving patterns such as (b) and (e). Since there are two stressed syllables, [ˈli] and [ˌma], without a splitting domain restriction, the second vowel could be split leftward to make the first stressed syllable heavy, as in *[ˈli₁a₂.ˌma₂.ku₃]. This is correctly blocked if we assume that a constraint prohibiting splitting across foot boundaries is active. Although such constraints may be unusual, there are numerous cases in which prosodic domains restrict the application of phonological processes. For instance, Jurgec (Reference Jurgec2011) notes that assimilation is in some cases restricted to particular prosodic domains. It is therefore reasonable to assume that an active constraint such as (20) restricts the splitting domain on the basis of a relevant prosodic boundary.Footnote ¹⁰

1. (20)

   

Along with *HL, this constraint ensures that the surface form is correctly derived, as shown in (21). SWP presumably motivates splitting of the unstressed vowel, with one of the split vowels becoming part of the stressed syllable, but splitting of the second syllable /a/ is blocked by the foot boundary. Similarly, vowel splitting that would create two identical adjacent vowels (i.e. lengthening) is also blocked by *V_αV_α.

1. (21)

   

Kwara'ae's multiple metathesis patterns are thus correctly derived in HS with split and fuse and the restriction of the splitting domain to the foot. Assuming a prosodic domain effect on splitting distance also ensures that long-distance metathesis (via long-distance splitting) does not arise. Consider the case of /li₁ma₂ku₃/ → [(ˈli₁)(ˌma₂ku₃)] → [(ˈli₁) (ˌma₂u₃ku₃)]; the split vowel /u₃/ cannot be split again into the first stressed syllable *[(ˈli₁u₃)(ˌma₂u₃ku₃)], since this would involve crossing a foot boundary.

3.4 Kwara'ae locality restrictions in parallel OT

Restrictions on the direction of a transposition and on the linear distance of the segments involved in metathesis have generally been handled using a family of constraints that imposes restrictions on which segments can be transposed. As discussed in §2.3, unattested long-distance metathesis candidates in Rotuman need to be blocked by a family of adjacency constraints in Carpenter (Reference Carpenter, Carpenter, Coetzee and de Lacy2002). Similarly, some patterns in Kwara'ae, such as the one in (21), are problematic, and require the postulation of additional constraints in parallel OT. This section summarises such cases. By comparing the HS analysis with the version of parallel OT presented here, I argue that the non-metathesis analysis is to be preferred, because it does not require an additional family of constraints to restrict the domain of transpositions, and the directionality is captured through active constraints in the language.

In Heinz's (Reference Heinz2004, Reference Heinz2005a, Reference Heinzb) parallel OT analysis of Kwara'ae metathesis, in which SWP outranks Linearity, a derivation such as /li₁ma₂u₃mu₄lu₅/ → [ˈli₁.ma₂u₃.ˌmu₄u₅l] is problematic, since this ranking would select [ˈli₁a₂.mu₃.ˌmu₄u₅l] as the optimal candidate, with both of the two stressed syllables [ˈli] and [ˌmu] made heavy by two transpositions, as in (22) (Heinz Reference Heinz2005a: 202).

1. (22)

   

This problematic candidate motivates a family of Contiguity constraints. Observing that the VV sequence in Kwara'ae is never disturbed, while CV sequences are, Heinz (Reference Heinz2005a) argues that V-V MaxContig, which reinforces preservation of the contiguity between adjacent vowels, is ranked above SWP and the general Contiguity constraint, preventing the problematic candidate from surfacing. Candidate (23a) violates V-V MaxContig, because the /a-u/ contiguous string in the input is deleted in the output. The correct output form in (23b) is derived, although contiguous input strings are deleted and new ones created, thus violating general MaxContig and DepContig (Heinz Reference Heinz2005a: 205).

1. (23)

   

In comparison, the non-metathesis HS analysis does not require a family of additional constraints, since prosodic constraints that restrict the splitting domain take care of the undesirable candidates. The derivational path is given in (24): both *V_α)FtV_α and *V_αV_α crucially dominate SWP.

1. (24)

   

Moreover, the foot-based split and fuse approach makes different predictions from the parallel OT analysis. One of the important generalisations in Kwara'ae multiple metathesis is that a vowel is always transposed leftwards (Blevins & Garrett Reference Blevins and Garrett1998, Heinz Reference Heinz2004). This fact is thought to be related to the cross-linguistic relationship between stress and CV/VC metathesis patterns (Heinz Reference Heinz2004). Heinz argues that in trochaic languages such as Kwara'ae vowels move to the left, whereas in iambic languages they move in the opposite direction. In fact, Blevins & Garrett's (Reference Blevins and Garrett1998: 538) analysis of diachronic change in Paman shows the pattern *V₁CV₂ > *V₁CV₁V₂ > *V₁CV₁V₂ > *CV₁V₂, in which the V₁ is copied rightwards into the stressed syllable.

The parallel OT transposition analysis with Linearity alone cannot explain this directionality straightforwardly, since there is no restriction on which segments can be transposed. To account for this generalisation, Heinz (Reference Heinz2005a) speculates that these two kinds of languages differ in terms of the types of contiguous sequence that are allowed in the output: Kwara'ae and similar languages prohibit the creation of new CV transpositions. However, the underlying CV transitions do not need to display the same contiguity in the output. Although a family of contiguity constraints in Heinz (Reference Heinz2005a) derives correct metathesis patterns, it does not capture the generalisation about the locus of metathesis following from foot structure. For instance, in the case of the trimoraic word /fikuda/ ‘to gather them together’, a form such as [ˈhikˌdua] is not a possible surface form, even though it satisfies SWP and matches the possible prosodic pattern of the language (the correct form is [ˈhiˌkuad], which Heinz Reference Heinz2005a derives with contiguity constraints). The locus of the metathesis is most clearly determined by prosodic boundaries: it is the second CV of a disyllabic foot. This generalisation cannot be straightforwardly captured in a non-foot-based parallel OT analysis.

On the other hand, the analysis presented here correctly predicts that the vowel ‘movement’ (split and fuse) into the stressed syllable is leftward. This is because the default stress pattern is trochaic, and the SWP-motivated splitting of a vowel is restricted to the foot. This further correctly predicts that the opposite pattern would result in an iambic language (cf. Heinz Reference Heinz2005b).

The variation that exists in Kwara'ae normal forms provides another reason to prefer the HS analysis, without the transposition operation. In the normal form, a word such as [ˈmauɁ] ‘fear’ can also be realised as [ˈmauɁu̥], with a devoiced vowel.Footnote ¹¹ In a parallel OT analysis of Kwara'ae, vowel splitting, violating Integrity, is independently required to deal with the optional forms (/maɁu/ → [ˈmauɁ] or [ˈmauɁu̥]). Although the optional form is believed to be related to the metathesised forms (Heinz Reference Heinz2004 calls it ‘partial metathesis’), Heinz derives the optionality using output–output correspondence between normal (metathesised) and optional forms in parallel OT. On the other hand, in the HS split and fuse analysis, the output can be derived by the same mechanism with partially ranked constraints (or in Stochastic OT; Boersma & Hayes Reference Boersma and Hayes2001).

The variation between [ˈmauɁ] and [ˈmauɁu̥] can be accounted for by incorporating Max-μ (Stanton & Zukoff Reference Stanton and Zukoff2018), which militates against the deletion of input moras, assuming that Max-μ and *HL are unranked. In the optional case, Max-μ is ranked above *HL, while in other cases the ranking is reversed. The final vowel is devoiced if not fused, in order to enhance the foot structure.

In this section, I have shown that a split and fuse analysis can be readily applied to Kwara'ae metathesis patterns. The analysis here accounts for multiple metathesis patterns involving various FootForm constraints that seem to be well supported in the language. Unlike the parallel OT analysis of Kwara'ae metathesis, the HS analysis with a restriction on the splitting domain can block problematic candidates without the need for a family of new constraints.

If the gradualness of Gen is responsible for blocking long-distance patterns, are there any advantages to analysing metathesis as split and fuse instead of transposition? Or is a HS transposition analysis correspondingly capable of deriving the correct patterns? The next section shows how Linearity is problematic even in a Harmonic Serialism framework.

5.1 Balangao CC metathesis

Balangao, an Austronesian language spoken in the Philippines, exhibits synchronic CC metathesis. In Balangao, complex onsets are prohibited. Complex codas are allowed only when an affix is attached. Thus CC sequences generally appear word-medially. Upon affixation, a vowel in an open or unstressed syllable is often deleted (e.g. /Ɂi-hɨˈgɨp/ → [Ɂihˈgɨp]).Footnote ¹³ When this vowel deletion creates a glottal–plosive sequence, it surfaces as plosive–glottal, as shown in (27) (based on Shetler Reference Shetler1976, Hume Reference Hume2004 and Canfield Reference Canfield2015).

1. (27)

   

While it might be wondered whether it is possible to account for these patterns using split and fuse, this does not seem to be helpful for most cases of CC metathesis, because the intermediate stage with a split segment does not usually improve harmony when C₁C₂ becomes C₂C₁C₂ or C₁C₂C₁. Instead, I propose that Balangao's apparent CC metathesis results from fusion followed by splitting (fuse and split).

The consonant–consonant fusion operation proposed here is a phonological process in which adjacent segments and their features are analysed as being associated to a single segment. The concept is similar to Bradfield's (Reference Bradfield2014) ‘concurrency’, and defines a fused segment as a non-linear complex segment, e.g. /h₁g₂/ → [g͡h_1,2] or [h͡g_1,2]. Crucially, the complex segment created in this step is unordered. However, it has also been suggested that a complex segment with a plosive maintains the ordering (Steriade Reference Steriade, Huffman and Krakow1993). Under this view, the ordering in [g͡h] will determine whether the complex segment is a preaspirated or postaspirated plosive. Even on the assumption that the surface complex segment is ordered, it is possible to create a derivational path that does not involve transposition (e.g. /h₁g₂/ → [ʰ₁g₂] → [h͡g_1,2] → [gʰ_1,2]).

Following Canfield (Reference Canfield2015), I assume that the glottal–plosive sequence violates a constraint such as *GlottalPlosive. In HS, avoidance of such a sequence can be achieved by a series of operations that affect intermediate stages while improving harmony, but the direct effect of these operations is not necessarily obvious in the surface form.

What is crucial in the ‘metathesised’ surface form is that all of the input segments are present in the output. This can be achieved by fusing two adjacent segments to create a complex segment at Step 1 (28a) and then ‘unpacking’ them at the next step. Because of *CompSeg (Padgett Reference Padgett1995), which penalises a complex segment, the candidates will not converge at the fused stage.Footnote ¹⁴ At Step 2, the ‘unpacked’ candidate [Ɂig₂h₁ɨp] wins (and the derivation converges at the following step).

1. (28)

   

Various assumptions are central to my analysis. As discussed in the split and fuse analyses, indexation must be allowed. If the use of indexation throughout the derivation is justified for the CV metathesis cases, then its utilisation for CC metathesis is not unexpected. Fusion violates Uniformity, and an Integrity violation is then incurred by splitting: the opposite of the ordering seen in split and fuse. It might be argued that fusion would lead to the loss of some feature. However, the complex segments that are considered here do not involve such examples. Therefore, I assume all features are preserved in the process of creating a complex segment in which the segments are phonologically unordered. For instance, in the derivation in (28), the post-vowel-deletion intermediate input, /Ɂih₁g₂ɨp/, maps to [Ɂig͡h_1,2ɨp] to create a complex segment [g͡h_1,2].

5.2 Comparison to a parallel OT analysis

How does the non-metathesis analysis in HS differ from a parallel OT analysis? Canfield (Reference Canfield2015) provides a parallel OT analysis of Balangao CC metathesis. She models the variable vowel deletion with a markedness constraint DeleteVowelWhenAffixed (DVWA), and contextualises Linearity constraints based on the observation that the metathesis of glottal and plosive segments occurs only when they originate in the same morpheme. Since the general Linearity constraint does not differentiate between morpheme-internal and morpheme-external reordering, Canfield (Reference Canfield2015: 223–224) formulates the separate morpheme-related Linearity constraints in (29b), utilising McCarthy & Prince's (Reference McCarthy and Prince1995) definition of correspondence in (29a).

1. (29)

   

The constraints in (29b) allow Canfield (Reference Canfield2015) to derive outputs with and without metathesis. While metathesis applies in (30a), it is blocked in (b) by Linearity-outside-Morpheme (adapted from Canfield Reference Canfield2015: 89).

1. (30)

   

The analysis of Balangao proposed below incorporates fusion and splitting, and does not require a family of Linearity constraints. The blocking case of /Ɂumalih-to/ → [Ɂu.ma.lih.to] can be achieved simply by ranking CrispEdge (Itô & Mester Reference Itô, Mester, Merchant, Padgett and Walker1994) in (31) above *GlottalPlosive, as in (32).

1. (31)

   

2. (32)

   

The fact that transposition only occurs with segments originating in the same morpheme is problematic for a parallel OT analysis, necessitating the generation of various Linearity constraints. As we have seen in CV/VC metathesis cases, this again shows that in a parallel OT analysis, Linearity alone cannot capture the metathesis patterns. Each language discussed in this paper requires a different set of faithfulness constraints to account for its metathesis pattern, and to restrict the contiguity/adjacency of the segments. However, in the HS non-transposition analysis, the differences between morpheme-internal and morpheme-external metathesis patterns are accounted for by a constraint that has been independently proposed for other phonological processes in different languages.

5.3 Why fusion and splitting?

One of the reasons for wanting to choose the fuse and split analysis over other operations stems from an observation about the segment types involved in CC metathesis. Hume (Reference Hume2004: 219) cites Blevins & Garrett's (Reference Blevins and Garrett1998) observation that ‘glottals, liquids, and glides are commonly involved in metathesis’. Incorporating Blevins & Garrett's insight, Hume speculates that those segments generally have acoustically longer cues, which can ‘extend over a domain which may encompass adjacent sounds [and] can result in the overlap of important phonetic cues, potentially creating ambiguity concerning the onset and offset of the sounds involved’. According to Hume's Ambiguity/Attestation model, these ambiguous cues are subject to reinterpretation, and surface in the order which is more common in these languages. As noted in Blevins & Garrett (Reference Blevins and Garrett1998), if CC metathesis is observed particularly for segments that create ambiguity, it is not unreasonable to speculate that these overlappable segments are analysed as an unordered complex segment at some point in the derivation. While such complex segments are marked, and never surface in the output, they can still exert their effect at intermediate steps.

Analysing CC metathesis patterns such as the one in Balangao as fuse and split also allows us to postulate that all metathesis involves the general operations of fusion and splitting, although the order of application can differ, depending on the language.

A final advantage of the fusion and splitting analysis is that we see consonant fusion and splitting as independent operations. Consonant fusion has been observed in languages such as Zoque /jt/ → [tʲ] (Sagey Reference Sagey1986) and Korean /h+k/ → [kʰ] (Kim Reference Kim1977). Although rarer, consonant splitting cases have been reported, as preaspiration (e.g. Icelandic /vakʰna/ → [vahkna]; Heimisdóttir Reference Heimisdóttir2014), and as breaking (e.g. Slovak /ʒeːn/ → [ʒien] ‘women (gen)’; Rubach Reference Rubach1993).

The fuse and split analysis can be extended to various other metathesis cases, such as Highland East Cushitic (Hughto & Davis Reference Hughto and Davis2014), Bayso (Garrett & Blevins Reference Garrett, Blevins, Hanson and Inkelas2009) and possibly Faroese and Lithuanian (Hume & Seo Reference Hume and Seo2004, Kenstowicz Reference Kenstowicz1972). In both Faroese and Lithuanian, a stem-final velar stop and a coronal fricative are transposed when a certain consonant is suffixed (e.g. /sk-k/ → [ksk]), as shown in (33) (examples from Hume & Seo Reference Hume and Seo2004).

1. (33)

   

The parallel OT analysis of these patterns is explored in depth by Hume & Seo, who utilise the contextual constraint Avoid X/Y and the Sonority Principle to account for the patterns. For instance, the constraint motivating the apparent metathesis is *stop/C_C(obs), which penalises sequences such as /sk-t/ (Hume & Seo Reference Hume and Seo2004: 43).

The fuse and split analysis of these cases is straightforward. To avoid a plosive–plosive sequence in /s₁k₂-k/, the fricative–plosive /s͡k_1,2/ is analysed as a complex segment, arguably making the sequence better than a simple plosive–plosive sequence. The intermediate candidate derived through this operation does not violate the contextual constraint *stop/C_C(obs). Unpacking the ‘fused’ segment will avoid violating *CompSeg, as shown in (34). The fuse and split analysis of metathesis suggests that segments that have been argued to form single complex segments are potential candidates for fusion.

1. (34)

   

‘Complex segment’, as discussed in this paper, is loosely defined, to allow for the inclusion of various segment types that have been argued to form clusters or be analysed as single segments. This includes (i) the complex segments predicted to be possible by van de Weijer (Reference Weijer1996), (ii) sC clusters that have been argued to be complex segments (Yamane Reference Yamane, van de Weijer and Nishihara2001) and (iii) various secondary articulations, including palatalisation and aspiration. Further investigation is required to determine which of these can be considered as fusible segments. However, it is clear from various arguments in the literature that certain adjacent segments are more closely tied together, and in some cases have been analysed as single segments (e.g. sC sequences in loanword adaptation; Fleischhacker Reference Fleischhacker2001, Zuraw Reference Zuraw2007, Broselow Reference Broselow, Hsiao and Wee2015). The segment types involved in most of the attested synchronic CC metathesis (see Appendix A) seem to match the range of fusible segment described above, which leads me to predict that most CC metathesis can be analysed as fuse and split.

However, there are cases in which a fuse and split analysis seems difficult to maintain. In modern Hebrew, /t/ in the prefix /hit-/ and a stem-initial sibilant fricative become transposed, as in (35) (from Bat-El Reference Bat-El1988: 480 and Bolozky Reference Bolozky1980). This sibilant fricative + /t/ sequence is permitted within a morpheme and with other prefixes, as in /hi-ʦis/ → [hiʦis] ‘it fermented’ and /t-saper/ → [ʦaper] ‘you (masc.sg) will tell’ (Yanagawa Reference Yanagawa, Solé, Recasens and Romero2003).

1. (35)

   

Examples such as /hit-ʦalem/ → [hiʦtalem] present problems for the fuse and split analysis. The root-initial affricate /ʦ/ metathesises with the prefix /t/. Even if we assume that the intermediate candidate winner is derived from fusion (/hit-ʦalem/ → [hiʦalem]), creating a single affricate, it seems unlikely that an affricate can be split to create both an affricate and a plosive.

As shown above, CC metathesis cases are analysable as processes involving fusion and splitting. Although the fuse and split analysis for CC metathesis requires refinement and further supporting evidence, it gives us an opportunity to understand the phenomenon as resulting from a series of fusion and splitting operations, in the opposite order from those involved in CV metathesis. Problematic cases such as Hebrew seem to involve a sibilant. The current fuse and split analysis isolates a specific type of metathesis, suggesting that metathesis involving sibilants might need to be analysed as undergoing other types of sequential operations.Footnote ¹⁵ By eliminating transposition as an operation, CC metathesis in HS does not require specific faithfulness constraints (e.g. a family of Linearity constraints).