2.1 Boundaries as disruptors of word prosody

In the theory of prosodic domains of Selkirk (Reference Selkirk, Aronoff and Kean1980, Reference Selkirk and Fretheim1981) and Nespor & Vogel (Reference Nespor and Vogel1986), among others, morphological and syntactic structure project prosodic structure, which affects phonological rule application. For example, a rule might insert a Prosodic Word (PWd) boundary at the beginning of every lexical word, resulting in schematic prosodic structures such as in (1). We assume that one Prosodic Word can dominate another (Ito & Mester Reference Ito, Mester, Honma, Okazaki, Tabata and Tanaka2003).

1. (1)

   

Adaptations of this approach in Optimality Theory, such as Peperkamp (Reference Peperkamp1997), use Align constraints (McCarthy & Prince Reference McCarthy and Prince1993) to derive these prosodic structures. For example, Align(LexWd, L; PWd, L) requires the left boundary of every lexical word to coincide with the left boundary of some Prosodic Word.

The Prosodic Word is generally assumed to be the domain of footing. Therefore, PWd boundaries disrupt the footing pattern. For the structures in (1), a disruption will occur between prefix and stem, and inside a compound. A stem and suffix, on the other hand, will have the same footing as a monomorphemic word. It is also possible for morphological boundaries to project feet directly, instead of via PWd structure, as assumed in Kager (Reference Kager1997) for Sibutu Sama. Finally, some authors have argued for multiple domain types that are approximately word-sized, including Rice (Reference Rice1992), who argues that three labels are required to account for the domains of various phonological processes found in Slave (Rice Reference Rice1989) and some other Athapaskan languages.

An alternative to prosodic domains is offered in Lexical Phonology, which cyclically interleaves affixation and stress rules (Siegel Reference Siegel1974, Kiparsky Reference Kiparsky and Ingemann1983). For the example in (1), the footing difference would be due to the order suffixation, footing, prefixation, compounding. This interleaved approach allows further distinctions between, say, early and late suffixes. Hargus (Reference Hargus1988), for example, argues for a rich system of levels in Sekani (Tsek'ene). A related approach to interleaving uses optimality-theoretic output–output correspondence constraints (Kenstowicz Reference Kenstowicz1995, Benua Reference Benua1997, Burzio Reference Burzio, Dekkers, Leeuw and de Weijer2000, among others): different affixes are subject to higher- or lower-ranking faithfulness to the base's prosody.

A third major approach to explaining morphological effects on prosody has been the use of different boundary types. Chomsky & Halle (Reference Chomsky and Halle1968) (who do not use feet) treat English -(i)an as having a + boundary, and -ism a # boundary. If prefixes have a # boundary, and suffixes a + boundary, then stress rules can ignore + but be sensitive to #. Selkirk (Reference Selkirk, Aronoff and Kean1980) argues that this theory is excessively powerful, because it does not enforce a hierarchical relationship between + and #.

Prosodic domains, morphology/phonology interleaving, output–output correspondence and boundary types are not inherently in conflict – Inkelas (Reference Inkelas1989), for example, develops a theory that employs the first two – but these approaches usually provide competing explanations of the same facts. §10.2 compares our account of Samoan word prosody using prosodic domains to the alternatives.

2.2 Affix-size differences

As will be seen in §5.2 and §5.3, in Samoan the size of a suffix is crucial for whether it disrupts word prosody. Reports of a prosodic distinction between mono- and disyllabic suffixes or enclitics abound, although the distinction does not always imply that either type of suffixation is treated differently from a monomorphemic word. For example, Anderson & Otsuka (Reference Anderson and Otsuka2006) treat the monosyllabic demonstratives of Tongan (Churchward Reference Churchward1953) as enclitics, because they shift stress to the right ([ˈmoko] ‘gecko’, [moˈko-ni] ‘this gecko’), and the disyllabic demonstratives as separate words ([ˌmokoˈeni] ‘this gecko’). However, the stress data are also consistent with right-to-left footing of the noun+demonstrative sequence: [mo(ˈko-ni)], [(ˌmoko)-(ˈeni)] – in both cases, the stress pattern is the same as in a monomorphemic word. Suffixation in Fijian follows the same pattern (Scott Reference Scott1948: 744, Dixon Reference Dixon1988: 24–31). Similarly, in Rotuman (Austronesian; closely related to Polynesian), Hale et al. (Reference Hale, Kissock and Reiss1998) leave open the question of whether disyllabic suffixes form a separate domain of footing or merely a separate foot, as they would in the absence of a morpheme boundary. McCarthy (Reference McCarthy2000) analyses the distinction between monosyllabic and disyllabic suffixes in Rotuman in PWd terms by having an Align constraint that prefers the stem and suffix to form separate Prosodic Words, overridden by the higher-ranked requirement of foot binarity, so that monosyllabic suffixes join the stem's Prosodic Word.

In several Australian languages, however, there is a clearer stress distinction between mono- and disyllabic suffixes, such as in Warlpiri (Nash Reference Nash1980), Yidiny (Dixon Reference Dixon1977a, Reference Dixonb), Diyari (Austin Reference Austin1981), Dyirbal (Dixon Reference Dixon1972) and Ngalakgan (Baker & Harvey Reference Baker and Harvey2003, Baker Reference Baker2008). In Ngalakgan a disyllabic suffix (underlined in (2)) restarts the left-to-right footing pattern. As shown in (2), the stress patterns of suffixed words are different from those of monomorphemes, and also different from what would be expected if every suffix initiated a new footing domain.

1. (2)

   

Most accounts employed to explain such affix-size differences are domain-based. As Baker (Reference Baker and Mushin2005) discusses, merely interleaving suffixation and stress assignment wrongly predicts that after [(ˈʈotojʔ)-ki] is suffixed with [-ppulu], there is enough material to form a new foot, left-to-right as usual, yielding *[(ˈʈotojʔ)-(ˌki-ppu)lu]. He instead uses domains: every morpheme prefers to be a separate foot, but not at the expense of creating a subminimal foot. When two monosyllabic suffixes are adjacent, Parse-σ (Prince & Smolensky Reference Prince and Smolensky1993) breaks the tie and foots the two syllables together.

McCarthy & Prince (Reference McCarthy and Prince1994) also use domains to treat similar facts in Diyari. In Diyari, a monosyllabic suffix can't initiate a foot even if another monosyllabic suffix follows: [(ˈmada)-la-ntu] ‘hill-characteristic-proprietive’. McCarthy & Prince adopt the footings of Poser's (Reference Poser1989) rule-based analysis, but use Align constraints to establish the domain of footing (1994: A6). The end of every stem closes off a Prosodic Word, [[[mada]_PWd-la]_PWd-ntu]_PWd, preventing footing of subminimal [-la] or [-ntu]. This has similar effects to Poser's (Reference Poser1989) analysis, which foots each affix independently, then removes monosyllabic feet. Alderete (Reference Alderete and Parker2009) extends the Align-based analysis to Dyirbal, Warlpiri and Pintupi, giving further arguments for the recursive PWd structure.

Crowhurst (Reference Crowhurst1994) analyses Diyari in terms of direct morpheme-to-foot alignment constraints. Align(Morpheme, L; Ft, L) instructs every morpheme to initiate a foot, though not at the cost of subminimal feet (*[(ˈmada)-(ˌla)-(ˌntu)]) or of a foot that contains a morpheme boundary (*[(ˈmada)-(ˌla-ntu)]). Crowhurst shows that Dyirbal (Dixon Reference Dixon1972), by contrast, allows a foot to span a morpheme boundary, so two successive monosyllabic suffixes are footed together. (Kager Reference Kager1997 analyses Dyirbal in terms of PWd structure: the root ends with a PWd boundary, but there are no PWd boundaries within the sequence of suffixes, so the suffixes form a single footing domain.) Crowhurst further contrasts Gooniyandi (McGregor Reference McGregor1990), where a root plus any number of monosyllabic suffixes behaves as though monomorphemic (whereas a disyllabic suffix always initiates a new foot). Kenstowicz (Reference Kenstowicz, Miglio and Morén1997b) analyses Diyari, Dyirbal and other related languages in terms of violable paradigm uniformity: *[(ˈmada)-(ˌla-ntu)] is ruled out in Diyari because the grammar prefers the suffix to have the same (unstressed) realisation everywhere.

In summary, accounts of affix-size differences have largely used Align constraints either to place foot boundaries at morpheme boundaries directly or to establish the Prosodic Word as a domain of footing; the size differences themselves emerge from interactions with minimum foot-size requirements. Some of the data are not explainable with cyclicity alone, though the literature generally does not fully explore all the analytic possibilities.

Affix-size differences in Samoan will be treated in §5–§8. The basic stress data are compatible with multiple analyses, but we will conclude from words with two suffixes that cyclicity alone is inadequate and Align constraints are needed.

3.1 Language background

All data in this paper are from the tautala lelei register of speech (which preserves more segmental contrasts than the other register, tautala leaga), described in most other work on Samoan. The phoneme inventory is shown in (3). The symbols have their IPA (IPA Handbook 1999) values, except that /t/ is usually heavily affricated ([tˢ]). The phonemes in parentheses are restricted to loanwords.

1. (3)

   

Samoan phonotactics require every consonant to be followed by a vowel, yielding syllables like [(C)V], [(C)Vː] and possibly [(C)VV], but not *[(C)VC]. We assume that [(C)Vː] and [(C)VV] are bimoraic. (We generally write long vowels as [Vː], but are agnostic as to whether they should be treated as single segments or sequences of identical segments. §5 considers both possibilities for input structure.)

3.2 Data collection

All transcriptions in this paper are of our consultants’ speech. Data were collected in one- to two-hour sessions from September 2007 to November 2008 with one main consultant, age 19, who was born and raised in Upolu and had moved to the Los Angeles area four years previously. Elicitation items were often found using Mosel & Hovdhaugen (Reference Mosel and Hovdhaugen1992) or Milner (Reference Milner1993). The consultant was given the Samoan written form and asked to confirm that he was familiar with the word, to pronounce the word a few times and to check the gloss. All words were elicited in isolation, where the stress patterns were the clearest to our ears (see §4.2 and §4.5), though some morphologically complex words, such as inflected verbs, were first elicited in sentences.

Critical items were also elicited and recorded from six consultants in Apia, Samoa in November 2011, and from our primary consultant, again in Los Angeles. Recordings were made with a head-mounted microphone (Shure SM10A). The consultants in Samoa included four men and two women, ranging in age from 18 to 39 (mean 27), from the capital city of Apia, and other areas of Upolu and Savai'i. All of them spoke primarily Samoan in daily life and were literate in Samoan, but also spoke English as a second language with some fluency. English was used as the contact language. In these sessions, each Samoan word was presented to the consultant on a slide on the computer screen, usually with a picture representing the meaning. For some morphologically complex words, such as verbs with the ergative /-a/ suffix in the Appendix, we sometimes presented both a base and derived form, and elicited the word in a sentence frame.

The consultant was asked to confirm familiarity with the word, and was recorded pronouncing the words and sentences. To elicit judgements about stress patterns other than the one originally pronounced by the consultant, the investigator uttered a pronunciation with an alternate stress pattern and asked if the alternate pronunciation was acceptable. The consultant was then asked to pronounce it him/herself and to confirm that it indeed sounded acceptable/unacceptable. Consultants often volunteered pronunciations with alternate stress patterns. We did not ask consultants to abstract stress patterns away from the pronunciation of the word, such as by hand-tapping on the stressed syllable.

4.1 Primary stress in monomorphemes

The basic primary stress pattern of Samoan is simple: a moraic trochee at the right edge of the word (Homer Reference Homer2007a). Stress is on the final vowel if long, and otherwise on the penultimate. Examples are given in (4), with parentheses enclosing presumed feet.Footnote ¹

1. (4)

   

Samoan primary stress is similar to Fijian (Schütz Reference Schütz and Schütz1978, Reference Schütz1985, Dixon Reference Dixon1988), where, in Hayes' (Reference Hayes1995) analysis, a word ends with a moraic trochee. In both Fijian and Samoan, there are no surface forms with a short final vowel and a long penultimate vowel (e.g. *[ˈmaːnu]) – see §6. We will see a systematic set of exceptions to the basic pattern in §8.

4.2 Phonetic realisation of primary stress

Three primary acoustic features appear to be involved in the realisation of Samoan primary stress: amplitude, duration and fundamental frequency (F0). Amplitude and duration were not investigated systematically; all else being equal, stressed syllables seem louder than unstressed, and the consonant after the stressed vowel may be lengthened (see note 1). We have investigated F0 systematically elsewhere (Orfitelli & Yu Reference Orfitelli and Yu2009): F0 rises on the stressed mora, shown in the pitch tracks in Fig. 1. We label this rise as an LH pitch accent, without notating tonal association (i.e. L*+H or L+H*), since we have no evidence for a difference in stability of tonal alignment of the L or H target, nor for a meaning contrast based on tonal alignment (Pierrehumbert & Steele Reference Pierrehumbert and Steele1989, Arvaniti et al. Reference Arvaniti, Ladd, Mennen, Broe and Pierrehumbert2000). The shaded boxes in Fig. 1 highlight the stressed mora (and preceding consonant) and its F0 rise, which is followed by an utterance-final fall. The F0 rise of a stressed long vowel, e.g. [la(ˈvaː)], is over its first half.

Figure 1

Pitch tracks for items in (4): (a) [la(ˈvaː)] ‘energised’; (b) [le(ˈlei)] ‘good’; (c) [(ˈmanu)] ‘bird’; (d) [ma(ˈnoŋi)] ‘to smell good’.

4.3 Analysis of basic primary stress

Our analysis of primary stress is straightforward: a bimoraic foot is required at the end of the word. This can be captured in Optimality Theory with Prince & Smolensky's constraints Edgemost, FootBinarity and RhythmType (1993: 35, 50, 56). The definitions we assume are given in (5).

1. (5)

   

(The constraints Edgemost-L and RhythmType=Iamb are ranked too low to be active.)

(6) illustrates the pattern with a tableau for /iŋoa/ ‘name’.

1. (6)

   

While the data presented thus far support don't support any ranking arguments for these three constraints, we will see in §8 that Edgemost-R is dominated.

4.4 Secondary stress in monomorphemic loans

It's difficult to find convincingly monomorphemic native words with four or more moras. Some examples are given in (7), but, as is typical, they seem to be reduplicated (if not productively), contain a sequence that can be an affix, such as [maː-], or possibly have undergone vowel lengthening (as suggested by the related forms given for ‘gentle, slow’ and ‘laugh’, and the variation for ‘travel by sea’).

1. (7)

   

Long monomorphemes tend to be English loans. In five-mora loans where neither of the first two vowels is epenthetic (as compared to the English original), secondary stress falls on the first mora (8). In Prince's (Reference Prince1983) terms, Samoan displays an initial dactyl effect. In (8), vowels with stressed English correspondents are underlined, and those with no English correspondent are bold. Even when the first vowel does not correspond to a stressed English vowel (‘democracy’, ‘Olympics’), it bears stress. The sole exception is ‘philosophy’ (no underlining, because final [ia] makes the source language unclear).

1. (8)

   

Unsurprisingly, when the second vowel is epenthetic, secondary stress remains on the first vowel.

1. (9)

   

When the first vowel is epenthetic, stress usually falls on the second, as shown in (10).

1. (10)

   

Similar avoidance of stress on epenthetic vowels in loans has been discussed for Fijian (Schütz Reference Schütz and Schütz1978, Reference Schütz1999, Hayes Reference Hayes1995, Kenstowicz Reference Kenstowicz2007) and Selayarese (Broselow Reference Broselow1999, Reference Broselow, Vaux and Nevins2008). When the footing requires it, it is possible for epenthetic vowels to bear secondary or even primary stress (11).

1. (11)

   

4.5 Phonetic correlates of secondary stress

Phonetic correlates of secondary stress are similar to those for primary stress, but less stable. In a sentence context, there is often no pitch event marking secondary stress, only interpolation between surrounding events. In citation form, a pitch rise (LH) on the secondary-stressed syllable is often smaller than the pitch rise for the primary stress. In Fig. 2 we show two pitch tracks for ‘computer’ (with considerable peak delay for the first rise). On the left, pitch rises are similar on both stressed moras; on the right, the primary-stress rise is sharper.

Figure 2

Variation in relative size of pitch rise between secondary and primary stress in two realisations of [(ˌkomi)pi(ˈuta)] ‘computer’. There is considerable peak delay for the first rise, indicated by ‘<’.

In Fig. 3 one can see the difference between initial and peninitial secondary stress, while Fig. 4 shows two tokens of [(ˌmaː)(ˌloː)(ˈloː)]. In Fig. 4a, we see little pitch rise for either of the putative secondary stresses; in Fig. 4b we see a clear rise for the first stress, but not the second.

Figure 3

Initial vs. peninitial secondary stress: (a) [(ˌmili)o(ˈnea)] ‘millionaire’; (b) [pu(ˌɹini)(ˈsese)] ‘princess’.

Figure 4

Examples of minimal pitch rises for secondary stresses in [(ˌmaː)(ˌloː)(ˈloː)] ‘rest’. In (a) there is little pitch rise for either of the putative secondary stresses (indicated by ‘?’; in (b) there is a clear rise for the first stress.

Our transcriptions of secondary stress are typically based on eliciting multiple tokens of a word in citation form until we were confident of the transcription. In some of the items below, we note that we were unsure of secondary stress – it may well have been present phonologically, but its phonetic realisation was not clear enough for us to confidently locate it. In some cases, we include foot boundaries predicted by our analysis, but no secondary stress mark.

4.6 Analysis of secondary stress

Although the evidence for the initial dactyl effect is limited to loans, we will tentatively assume that Samoan prefers words to begin with a foot, using the constraint definitions in (12), from McCarthy & Prince (Reference McCarthy and Prince1993b).

1. (12)

   

This is illustrated in (13) for /temokalasi/ (cf. English [diˈmɒkɹəsi]).

1. (13)

   

We lack data on longer words to confirm how medial feet align – the only good example is [(ˌʔafa)(ˌkni)si(ˈtna)] in (9).

As mentioned above, stress is avoided on epenthetic vowels ([pa(ˌlani)(ˈkeke)] ‘blanket’ vs. *[(ˌpala)ni(ˈkeke)]), unless the alternative increases the number of unparsed syllables ([(ˌsika)(ˈlamu)] vs. *[sika(ˈlamu)] ‘scrum’, [(ˌpiː)(ˈniki)] vs. *[(ˈpiː)niki] ‘pink’). If we take this to be a synchronic pattern, we have the ranking Parse-σ≫Don'tStressEpenthetic≫Align-L(PWd, Ft). We leave open whether Don'tStressEpenthetic is synchronically active or applies only to initial loan adaptations.Footnote ³

Stress in compounds

In compounds of two roots, the first ends with a foot, indicating a PWd boundary between the two roots. The examples in (14) illustrate how two compounds with the same number of syllables can have different stress patterns.

1. (14)

   

We adopt Peperkamp's (Reference Peperkamp1997) Align-based approach, and propose that Samoan has high-ranking Align(Morpheme, L; PWd, L) (Align-L(Morpheme, PWd)): every morpheme must initiate a prosodic word. As we will see starting in (20), this constraint is violable, and when we turn to disyllabic suffixes, we will see that the constraint should indeed refer to morphemes in general rather than only to lexical words. Because Edgemost-R and Align-L(PWd, Ft) refer to the Prosodic Word, Align-L(Morpheme, PWd) indirectly determines footing. The tableau in (15) illustrates this for /alofi+vae/. We leave open the question of whether the two Prosodic Words in (a) combine to form a larger Prosodic Word, or attach directly to the next level up, such as a Phonological Phrase.Footnote ⁴

1. (15)

   

We have not found prosodic differences among subtypes of compounds. As Mosel & Hovdhaugen (Reference Mosel and Hovdhaugen1992: 240) discuss, it is difficult to distinguish between phrases and compounds without a clear sign of compoundhood (e.g. opaque meaning, a bound root or a compounding affix such as /-ŋaː/). Compounds can be noun+noun or adjective+noun (14), or noun+verb ([ˈave-pasi] (drive+bus) ‘drive a bus, bus driver’).

5.2 Stress with suffixes

5.2.1 Stress with monomoraic suffixes

There are several monomoraic suffixes in Samoan, and they attract stress to the right (16): the domain that must end in a moraic trochee includes these suffixes. In longer items we see that secondary stress can be added when suffixation brings a word up to four moras, but secondary stress does not move after suffixation, whether because of faithfulness to the unsuffixed form or due to preference for an initial foot.Footnote ⁵

1. (16)

   

The pitch tracks in Fig. 5 clearly illustrate the stress shift, for ‘work’. The unsuffixed form has a single pitch rise on [lu], and the suffixed word has an added initial rise for the secondary stress, and a larger rise for the primary stress, now on [e].

Figure 5

Pitch tracks for stress shift under suffixation: (a) [ŋa(ˈlue)] ‘work (vb)’; (b) [(ˌŋalu)(ˈe-ŋa)] ‘work (n)’.

In (17) we see that the denominal suffix /-a/ (‘abounding in N, having N’) behaves in the same way.

1. (17)

   

A homophonous suffix /-a/, with the same stress behaviour, is shown in (18a). The suffix is one of the many forms of a family of verb suffixes often called ergative (productive /-a/ and /-ina/; unproductive /-Cia/, where C can be any consonant; /-na/ (18b)). Chung (Reference Chung1978) and Homer (Reference Homer2007b) find that these occur when an ergative subject moves over the verb, and possibly in other contexts.

1. (18)

   

The final monomoraic suffix we examine is the fairly unproductive /-Ci/, where C can be any consonant, including zero. When accompanied by /fe-/, it usually creates plural verb forms. Alone, its semantic contribution is unpredictable. Adding /-Ci/ also shifts stress to the right.

1. (19)

   

We conclude that these suffixes belong to the same Prosodic Word as the stem, and so are footed together with it. This violates our constraint Align-L(Morpheme, PWd), which would require the suffix to initiate a new Prosodic Word. We assume however, that, as a requirement of either Gen or a high-ranking constraint, every Prosodic Word must contain a foot. And feet, by FootBinarity, must be bimoraic. Therefore, FootBinarity must outrank the Align constraint, as shown in (20) for /teʔe+na/ ‘reject’. We assume that strict layering is favoured by violable constraints (Ito & Mester Reference Ito, Mester, Honma, Okazaki, Tabata and Tanaka2003), Selkirk's (Reference Selkirk and Beckman1995) Non-recursivity (in this case, a PWd should not dominate another PWd) and Exhaustivity (in this case, a syllable shouldn't be left to attach directly to the Phonological Phrase, but should instead belong to a Prosodic Word).

1. (20)

   

5.3 Stress with prefixes

Samoan has few productive prefixes. There is one bimoraic prefix, the extremely common /faʔa-/. We usually hear its stress as [ˌfaʔa-], and never *[faˌʔa-], though this is hard to verify acoustically because both syllables tend to be short and the glottal stop disrupts the pitch; /faʔa-/ is often realised as simply [faː]. Examples are given in (26). The prefix /faʔa-/ is typically described as, roughly, causative.

1. (26)

   

Our constraint grammar so far predicts that /faʔa-/ forms its own Prosodic Word, because Align-L(Morpheme, PWd) requires /faʔa-/ and the root morpheme to each initiate a Prosodic Word. The data in (26) are consistent with /faʔa-/ forming its own Prosodic Word, though they are also what we expect if the whole word is a single Prosodic Word, because of the initial dactyl effect.

There is also a monomoraic prefix /fe-/, used in plural forms of many verbs and illustrated in (27). It is usually accompanied by a suffix.

1. (27)

   

The grammar predicts that /fe-/ should be adjoined to the root's Prosodic Word ([fe-[root]_PWd ]_PWd), because it is not big enough to form a Prosodic Word of its own. This means that /fe-/ should not receive secondary stress, even if the root is longer than two syllables. We found very few examples where the root was long enough for /fe-/ to potentially be stressed, and were unable to confidently transcribe secondary stress.

1. (28)

   

The prefixes /fe-/ and /faʔa-/ can combine. The grammar predicts that /fe-/ is adjoined to /faʔa-/'s Prosodic Word ([fe-[faʔa]_PWd ]_PWd [root]_PWd), yielding the secondary stress pattern [fe-(ˌfaʔa)-…]. Though we have few examples, they match this prediction, as shown in (29), although the pattern is also consistent with the two prefixes forming a single Prosodic Word. If the prefixes were inside the root's Prosodic Word, we would expect *[(ˌfe-fa)ʔa-…]. The stress in [fe-faʔa] is thus weak evidence for a PWd boundary between [faʔa] and the root; we will see stronger evidence in §7 (49) and §8.2 (64).

1. (29)

   

We also have a few examples of /faʔa-/ followed by /fe-/. The grammar predicts a separate Prosodic Word for /faʔa-/, with /fe-/ adjoined to the root's Prosodic Word: [faʔa]_PWd [fe-[root]_PWd ]_PWd. The predicted stress pattern is [(ˌfaʔa)-fe…]. As shown in (30), this is what we find, not *[fa(ˌʔa-fe-)…], which we would expect if the two prefixes form a single Prosodic Word.

1. (30)

   

We return to the prosodification of prefixes in §8.2.

5.4 Stress with reduplication

Reduplication in Samoan occurs mostly in verbs. There are two types of reduplication: CV reduplication, typically for plural agreement, and two-mora reduplication, typically with a pluractional, frequentative or intensified meaning. Both types of reduplication can occur with the same root.

CV reduplication doubles the initial CV of the primary-stressed foot.Footnote ⁸ As shown in (31), the resulting stress pattern looks the same as for a monomorpheme, suggesting that the reduplicant is integrated into the root's Prosodic Word. (In §6, however, we examine some more complex cases.) The CV reduplicant often sounds shorter than a typical unstressed syllable, as Mosel & Hovdhaugen (Reference Mosel and Hovdhaugen1992: 34) note, but still counts as moraic, since it allows a preceding mora to bear secondary stress.

1. (31)

   

Two-mora reduplication typically copies the primary-stressed foot. In this case, we perceive no particular shortening, so the choice of which copy to designate as the reduplicant is arbitrary. Typical examples are shown in (32a), along with an example, (32b), in which a non-final foot is copied (this is less common, and unpredictable). When the root is trimoraic (32c), secondary stress does not fall on the initial syllable, as would be expected under an initial-dactyl preference (*[(ˌsa<ʔe)u>(ˈʔeu]), but instead falls on the reduplicant ([sa<(ˌʔeu)>(ˈʔeu)]), suggesting a PWd break after the reduplicant, though there are few examples.Footnote ⁹

1. (32)

   

When a monomoraic suffix is added, the footing of the second copy changes, but that of the first copy does not, as illustrated in (33). This is again consistent with a PWd boundary between the two copies, though also with a preference for initial dactyls (we found no examples where the root was trimoraic), or partial faithfulness to the unsuffixed form. Further evidence that the reduplicant forms its own Prosodic Word is discussed in §6.Footnote ¹⁰

1. (33)

   

5.5 Summary of stress in complex words

In this section we have shown that suffixes shift primary stress to the right; in the case of monomoraic suffixes, this means that the stem and suffix must be in the same Prosodic Word. We found evidence for a PWd boundary between prefix and stem, between certain prefixes, between reduplicant and base, and between stems in a compound. We found evidence against a PWd boundary between a stem and a monomoraic suffix, with the status of bimoraic suffixes remaining open so far.

We now turn to other diagnostics of footing domains, which will give us additional evidence about the prosodic structure of morphologically complex words.

8.1 Basic diphthong formation

Certain VV sequences disrupt the normal stress pattern, as in [ˈmaile]. Unlike in Fijian (Schütz Reference Schütz and Schütz1978, Reference Schütz1985, Dixon Reference Dixon1988, Hayes Reference Hayes1995), Samoan sequences like the [ai] in [ˈmaile] do not sound shortened.

The stress disruption occurs when the first vowel is non-high and the second is high. Of the twenty possible sequences of non-identical vowels, diphthong formation occurs (in monomorphemic words) for [ai au ei ou], and not for [ae ao ea eo oa ia ie iu ua uo].Footnote ¹⁹

1. (51)

   

We illustrate near-minimal sets with pitch tracks in Fig. 6. In the words on the left (e.g. [ˈmaile]), the rise is over the first half of the VV sequence, whereas on the right (e.g. [maˈela]), the rise is over the second half.

Figure 6

Pitch tracks for (a) and (b) near-minimal pairs of diphthong-forming and non-diphthong-forming sequences ((a) [(ˈmai)le] ‘dog’ and [ma(ˈela)] ‘hollow’; (b) [(ˈmau)li] ‘heart’ and [ŋa(ˈolo)] ‘rattle, abound’); (c) two non-diphthong-forming sequences ([le(ˈaŋa)] ‘bad’ and [mo(ˈana)] ‘fish (sp.)’).

We propose a penalty on prominence mismatches: if the first vowel in a VV sequence is more prominent in sonority, it should also be more prominent metrically (e.g. Hayes Reference Hayes1995, Anttila Reference Anttila, Hinskens, van Hout and Wetzels1997, Kenstowicz Reference Kenstowicz1997a, Crosswhite Reference Crosswhite2001). Our constraint is limited to cases where the second vowel is high.Footnote ²¹

1. (52)

   

Tableau (53) shows the default pattern, when *aˈI is not relevant.

1. (53)

   

As shown in (54), *aˈI can force a violation of Edgemost-R (or FootBinarity, if we take (e) to be the winning candidate). The constraint *CompressedDiphthong rules out shortening the /ai/ sequence to count as monomoraic. We proposed in §6 that Uniformity is violable in Samoan, allowing the rich-base input /tu₁u₂si/ to map to fused [tu_1,2si], so it is not fusion itself that rules out candidate (g), but rather the fusion of non-identical vowels, in violation of Ident(Vfeatures) (which abbreviates Ident[high], Ident[front], etc.). For underlying long vowels, /tuːsi/, a right-aligned binary foot can be achieved by violating only Don'tShorten. But for /maile/, higher-ranked faithfulness constraints are at stake: *CompressedDiphthong, Ident(Vfeatures) and MaxV. The result is a faithful outcome, with a non-aligned foot.

1. (54)

   

In words ending in /…VːVCV/, primary stress is on the penultimate mora as usual, as in (55).

1. (55)

   

There is no need to violate Edgemost-R, because both vowels are stressed, satisfying *aˈI (56).

1. (56)

   

8.2 VV sequences at morpheme boundaries

We have seen that a monomoraic suffix is included in the Prosodic Word and normally shifts stress (§5.2.1). When a monomoraic suffix is added to a stem ending in [ai], [au] or [ei], stress fails to shift, as shown in (57)–(59). This is expected under the PWd structures developed above: a stem ending in [ai], [au] or [ei] followed by a monomoraic suffix simply behaves like a monomorphemic word (cf. [ˈmaile]), because it forms a single Prosodic Word.

1. (57)

   

There is one example with ergative /-na/, and one with a different suffix /-na/; both behave as expected.

1. (58)

   

We have two examples for the suffix /-Ci/. The stress is as in a monomorpheme, shifting to penultimate for /ao/ but not for /au/.

1. (59)

   

As expected, when a stem's penult appears as long under suffixation, there is no diphthongisation, because both vowels in the VV sequence now bear stress, and *aˈI is not violated.

1. (60)

   

In §7 we argued that the ergative suffix /-ina/ forms a separate Prosodic Word. Diphthongisation offers further support: just as the sequence /i+i/ does not shorten when the second /i/ belongs to /-ina/, there is also no diphthongisation in /a+i/ or /e+i/ when the /i/ belongs to /-ina/.

1. (61)

   

Diphthongisation is also blocked when /-ina/ attaches after another suffix that ends in /a/.

1. (62)

   

The pitch track in Fig. 7b shows a clear pitch rise on the [i] of [-ina], instead of on the preceding [a].

Figure 7

Pitch tracks for (a) [le(ˈaŋa)] ‘bad’ and (b) [(ˌfaʔa)-le(ˌaŋa)-(ˈina)] ‘destroy’, showing pitch rise on [i] in the separately footed suffix /-ina/, instead of on the preceding [a].

The lack of diphthongisation with /-ina/ is expected under our analysis, so long as *aˈI is ranked lower than the Align constraints governing PWd formation, as shown in the tableau in (63). (We assume that feet can't straddle PWd boundaries, either as a property of Gen or because of a high-ranking constraint.)

1. (63)

   

In §5.3 and §7 we argued that a prefix is not part of the stem's Prosodic Word – it forms its own Prosodic Word or is adjoined to the stem's. Here we see stronger support for this prosodification: diphthong formation is blocked across a productive prefix–stem boundary.Footnote ²² This is illustrated in (64) for [au ai ei].

1. (64)

   

Once again, if we assume that a foot can't straddle a PWd boundary, this is expected.

Diphthong formation is blocked across compound boundaries, consistent with the evidence in §5.1 and §7 for a PWd boundary between the two stems.

1. (65)

   

There is one morphological pattern that behaves a little differently. When the C in semi-productive /-Ci/ is zero, the preceding vowel is non-high and another suffix follows, the conditions for diphthongisation are in place. We find that stress does fall on the antepenultimate vowel in these cases, just as if the string were monomorphemic.

1. (66)

   

This is to be expected if the stem+suffix+suffix forms a single Prosodic Word. However, our grammar predicts that the two suffixes should form their own Prosodic Word, to satisfy Align-L(Morpheme, PWd) at least partially (see (23)). We conclude that only productive morphemes are subject to Align-L(Morpheme, PWd), perhaps because the grammar doesn't treat unproductive affixes such as /-Ci/ as real morphemes. (See the end of §8.3 for a VVV example.)

8.3 Final VVV sequences

In words ending V₁V₂V₃, when V₃ is not /a/, the diphthongisation behaviour of V₁V₂ is as before: stress is penultimate as usual, unless V₁ is non-high and V₂ is high.

1. (67)

   

However, if V₃ is /a/, we see diphthongisation even when V₁ is low and V₂ is mid.

1. (68)

   

To account for the [ˈaea] and [ˈaoa] patterns, we propose an additional prominence-matching constraint. While *aˈI (repeated in (69a)) penalises a prominence mismatch of a certain size, *ValleyAsPeak (69b) penalises a mismatch of any size, if it occurs on both sides of the vowel in question (*ValleyAsPeak is therefore applicable only to VVV sequences).

1. (69)

   

The tableaux in (70) illustrate how diphthongisation of low–mid sequences is possible only if the sequence is /VVa/.

1. (70)

   

The sonority–stress mismatch constraints in (69) resemble those in previous analyses of sonority-driven stress, such as Kenstowicz's (Reference Kenstowicz1997a: 162) *Peak/i,u and de Lacy's (Reference de Lacy2002: 15) *Δ_Foot/{ə,i/u} (e.g. ‘don't stress a high vowel’) or the SonorityPeak constraint of Clements (Reference Clements and Roca1997) and Dell & Elmedlaoui (Reference Dell and Elmedlaoui2002) (‘a sonority peak within a syllabification domain must be a nucleus’). Neither of these approaches entirely captures the Samoan data. If ranked correctly in relation to Edgemost-R, ‘don't stress [i u]’ and ‘don't stress [e o]’ can capture [(ˈmai)le] and [ma(ˈela)], but some additional mechanism is needed to account for the difference between [maˈela], where mid-vowel stress is tolerated, and [ˈmaea], where it is not. And while a stress version of SonorityPeak could work for sequences like /aia/ and /aea/, in /maile/ the /i/ is not a sonority minimum, because it is followed by even less sonorous /l/. An additional mechanism would be needed to prefer [ˈmaile] over [maˈile].

What do /VVa/ sequences tell us about PWd structure? The first relevant case occurs when the final /a/ is a suffix. In §5.2.1 we saw that the ergative /-a/ and denominal /-a/ suffixes both shift stress rightward, putting them in the same Prosodic Word as the stem. Thus, we expect /VV-a/ to behave the same as monomorphemic /VVa/.

Elicitations with our primary consultant yielded variable data, so we elicited and recorded representative words from six additional consultants in Samoa. Full results are shown in the Appendix: for each word, we show how many consultants had each pattern. The data confirm that stress is consistently antepenultimate for tautomorphemic /aia aua aea aoa/. When the final /-a/ is the ergative suffix, we see variation for /aua aia/, and for /aea aoa/ variation with a tendency towards penultimate stress. When final /-a/ is the denominal suffix, we see a strong tendency towards penultimate stress for /aua aia aea aoa/.

(71) schematises the /VVa/ results, with parentheses round a less common variant.Footnote ²³

1. (71)

   

The suffixed cases with penultimate stress are very surprising. If the ergative and denominal suffixes are treated as part of the root's prosodic domain, stress should be the same as in monomorphemes – antepenultimate. If they are treated as outside the domain, stress should remain where it was in the unsuffixed form – yielding, again, antepenultimate stress: [ˈaV-a]. The puzzle persists if we employ ordering, as in Lexical Phonology and Morphology: if the suffixes are added before stress is assigned, stress should be the same as in monomorphemes. If the suffixes are added after stress is assigned, stress should remain on the root's penult: either way, [ˈaV-a]. Likewise, output–output faithfulness to the unsuffixed form can't explain the [aˈV-a] pattern.

The vast majority of data that a Samoan learner encounters for these suffixes is from stems that end in CV (and even among stems that end in VV, most don't end in the crucial sequences, e.g. /ai/). We have assumed that the grammar explains the stress shift in these cases by putting the root and suffix in the same Prosodic Word. But learners could also learn stress shift as a (redundant) property of each suffix, conflicting with stress constraints only for the crucial VV-final roots.

We sketch a way to capture these suffix-specific stress shifts, using a proposal from Buckley's (Reference Buckley1998) analysis of similarly problematic stress data in distantly related Manam. Buckley appeals to low-ranked AlignHD, requiring the end of any Prosodic Word to be the end of the main-stressed foot, and high-ranked AlignHD^{suf}, which applies only to the ends of suffixes. For Samoan, we need constraints that are even more specific, referring to just these two suffixes, although we do not need to refer to the main-stressed foot specifically.

1. (72)

   

The difference between ergative and denominal /-a/ is in the conflict between each suffix's Align constraint and the prominence-matching constraints *aˈI and *ValleyAsPeak: denominal /-a/'s alignment requirement is ranked high, but ergative /-a/'s is optionally outranked by *aˈi, so that we see variation for /ai-a_erg, au-a_erg/, but hardly ever for /ai-a_denom, au-a_denom/. We have no explanation for why the difference between the two morphemes should be in this direction rather than the opposite.

We saw at the end of §8.2 that the semi-productive suffix /-Ci/ does not initiate a Prosodic Word, even when another suffix follows, allowing FootBinarity to be satisfied. The unproductive /-ia/ ergative suffix provides a similar example for VVV sequences. We were able to gather several examples with /a/- or /o/-final stems. As shown in (73), the /…a-ia/ data, with one variable exception (‘shine on’), indicate that /-ia/ is included in the stem's Prosodic Word, because it does not bear its own stress, but rather stress falls on the antepenult. This is in contrast with the behaviour observed for productive /-ina/, which does bear stress even after stem-final /a/. Once again, it seems that unproductive suffixes are not subject to Align-L(Morpheme, PWd). (See the Appendix for full elicitation results.)

1. (73)

   

Although we have no other data on the diphthongisability of /oi/, the /-ia/ data in (73) suggest that /oi/ does not pattern with /ai/.

To summarise this section, we have seen that prominence mismatches cause stress to deviate from its normal penultimate position, but not at the cost of a foot straddling a PWd boundary. Under this analysis, we have additional evidence that the productive ergative suffix /-ina/ forms its own Prosodic Word, and for a PWd boundary inside a compound. We also saw that a strong prominence mismatch is required to shift stress in /…VVCV#/ sequences, but even a weak prominence mismatch shifts stress in /…VVV#/. The ergative and denominal /-a/ suffixes seem to have developed an idiosyncratic pre-stressing requirement, which we capture by requiring them to be foot-final; this requirement is normally redundant, since it doesn't conflict with the stress pattern otherwise expected, but in /…VV-a#/ cases the suffix-specific constraints prevail.

10.1 Summary

This paper has covered several aspects of the word-level prosody of Samoan. The Hasse diagram in (78), made with OTSoft (Hayes et al. Reference Hayes, Tesar and Zuraw2003), shows the constraint rankings required in our analysis.

1. (78)

   

In (78), solid lines indicate crucial constraint rankings, for example, FootBinarity≫Align-L(Morpheme, PWd). The jagged line between Don'tShortenˈVː_BA and NoBreaking indicates a variable ranking: if Don'tShortenˈVː_BA≫NoBreaking, final long vowels break when suffixation makes them penultimate, and if NoBreaking≫Don'tShortenˈVː_BA, they shorten (see §6). The line between Edgemost-R and NoBreaking is dashed to indicate that although that ranking is required for the breaking variant, it is irrelevant for the shortening variant; the rankings Edgemost-R≫Don'tShorten and Edgemost-R≫Uniformity are required for either variant, either directly or by transitivity.

The two jagged lines connected to Align_erg-a in (78) also indicate variable ranking. If Align_erg-a≫*aˈI, *ValleyAsPeak, then the ergative suffix /-a/ is foot-final, and thus preceded by stress: [fa(ˈi-a)], [va(ˈe-a)] (see §8.3). Under the less frequent ranking *ValleyAsPeak≫Align_erg-a, regardless of the ranking of *aˈI, words with the ergative suffix /-a/ are stressed the same as monomorphemes: [(ˈfai)-a] ‘do’, [(ˈvae)-a] ‘kick’. And if *aˈI≫Align_erg-a≫*ValleyAsPeak, then stress in these words depends on the height of the stem-final vowel: [(ˈfai)-a], [va(ˈe-a)].

We have seen that in monomorphemic words there is a final moraic trochee bearing primary stress, and possibly a preference for a word-initial trochee bearing secondary stress if the word is long enough. Long vowels or sequences of identical vowels are not allowed in penultimate position if the final vowel is short, and we see alternations reflecting trochaic shortening. Stem-final long vowels also shorten or break when suffixation makes them penultimate. When monomoraic reduplication would create a long vowel in the penult, different morphology is used.

We have also proposed that the beginning of a lexical word projects the beginning of a Prosodic Word. Based on stress, monomoraic suffixes are included in the domain of footing, or Prosodic Word – that is, they shift stress to the right. The diagnostic that the penult of the Prosodic Word is the position where a long vowel is not allowed also supports the inclusion of monomoraic suffixes in the domain of footing, and suggests that there is a PWd boundary between the following: (i) stems and /-ina/, (ii) prefixes and stems, (iii) stems in a compound and (iv) a bimoraic reduplicant and its base.

Finally, certain vowel sequences disrupt the normal stress pattern. We propose that a prominence-matching constraint prohibits an unstressed non-high vowel followed by a stressed high vowel: *aˈI. Such sequences are allowed across PWd boundaries, however, because a foot can't straddle a PWd boundary. Notably, the evidence from these sequences supports the PWd boundaries proposed. Another prominence-matching constraint applies to three-vowel sequences: there, even if the middle vowel is only mid, rather than high, it can't be stressed if surrounded by unstressed, low vowels (i.e. [ˈaea ˈaoa], not *[aˈea aˈoa]): *ValleyAsPeak.

10.2 Comparison to other approaches

In §2 we reviewed existing approaches to phenomena similar to those observed in Samoan. The approach we have pursued throughout the paper uses Align constraints to impose prosodic-word boundaries, with the Prosodic Word as the domain of footing. Here we consider how other approaches fare.

A close cousin to the approach taken here is one in which Align constraints place feet directly at morpheme boundaries, without the Prosodic Word as an intermediary (Crowhurst Reference Crowhurst1994). The main challenge we see for applying this approach to Samoan involves words with monomoraic suffixes. In /lolo+fi/ (19), for example, as shown in (79), the incorrect footing *[(ˈlolo)-fi] is if anything preferred by direct morpheme–foot alignment constraints, because at least the stem morpheme begins with a foot. Some constraint is needed to prefer [lo(ˈlo-fi)], such as a preference for the word to end in a foot – in which case we still need a Prosodic Word, though perhaps a simpler one.

1. (79)

   

On the other hand, an advantage of morpheme–foot alignment is that morpheme-specific Align constraints can capture the idiosyncratic behaviour of certain morphemes: we analysed the pre-stressing behaviour of denominal /-a/ and, to a lesser extent, ergative /-a/ (§8.3) with constraints like Align_denom-a, which requires a foot boundary after denominal /-a/ – an even more specific version of Buckley's (Reference Buckley1998) AlignHd^{suf}.

Output–output faithfulness can handle some aspects of the Samoan data. Some form of output–output faithfulness is needed to deal with the different treatment of underlying penultimate long vowels in unsuffixed vs. suffixed words, as discussed in §6: in unsuffixed forms a long penult obligatorily shortens, whereas in suffixed forms there is variation between shortening and breaking, which we attributed to variable faithfulness to the unsuffixed form ([pe(ˈle-ŋa)]∼[(ˌpele)(ˈe-ŋa)], from [pe(ˈleː)]). Moreover, the lack of secondary stress shift under suffixation (§5.2) could be attributed to faithfulness to the unsuffixed form. The lack of diphthongisation (§8.2) and vowel coalescence (§7) before /-ina/ could also have been attributed to high-ranking faithfulness of /-ina/-suffixed forms to their base words: non-diphthongised [(ˌloka)-(ˈina)] is more faithful to its base [(ˈloka)] than is *[lo(ˈka-i)na]; non-coalesced [(ˌkiki)-(ˈina)] is more faithful to [(ˈkiki)] than is *[ki(k-ˈina)] or *[ki(ˈki-na)]. Similarly, the lack of length alternation with /-ina/ and other bimoraic suffixes (except /-Caʔi/) (§6) could have been explained by subjecting these suffixations to stronger output–output faithfulness constraints on vowel length. However, the fact that suffixes divide into two faithfulness groups almost entirely along prosodic lines (monomoraic suffixes and /-Caʔi/ vs. other bimoraic suffixes) would be accidental.

Interleaving affixation and phonological operations, as in Lexical Phonology, can also handle some of the data. Suppose that /loka/ is prosodified before being suffixed, as [(ˈloka)]. If faithfulness to existing footing outranks *aˈI, suffixing /-ina/ will just add a foot, [(ˌloka)-(ˈina)]. Monomoraic suffixes can still shift stress if Edgemost-R outranks faithfulness to existing feet: [(ˈlolo)], after being suffixed with /-fi/, changes to [lo(ˈlo-fi)]. The problem that this approach faces for Samoan is similar to the one that Baker (Reference Baker and Mushin2005) points out for Ngalakgan: suffixing [(ˈmau)-a] ‘get’ with /-ina/ should provide enough unfooted material to yield *[(ˌmau)-(ˈa-i)na], but the result is instead [(ˌmau)-a-(ˈina)] (62), with a new foot initiated by /-ina/, in violation of *aˈI. The failed derivation in (80) illustrates this.

1. (80)

   

To deal with trochaic shortening in Lexical Phonology, it would be necessary to group suffixes into levels. Those that can bleed trochaic shortening (monomoraic suffixes and /-Caʔi/) would be attached before shortening, and those that counterbleed it (the other bimoraic suffixes) would be attached afterward, and the fact that the partitioning of suffixes into levels is almost entirely determined by suffix size would be accidental.

1. (81)

   

Neither Lexical Phonology nor output–output faithfulness provides a natural solution to the problem of prestressed denominal /-a/ (and to a lesser extent ergative /-a/) (§8.3). Suffixed [va(ˈi-a)] is not faithful to its base or earlier derivational stage [(ˈvai)]; nor does ordering suffixation before prosodification help, since the result should be the same prosody as a monomorpheme, *[(ˈvai)-a]. As in our Align approach, an additional stipulation is required for these suffixes.

Finally, we consider an approach based on boundary symbols. As with output–output faithfulness or Lexical Phonology, the lexical entries of affixes must specify which of two groups they belong to. The monomoraic suffixes and /-Caʔi/ would be preceded by the boundary +, and the other bimoraic suffixes by #. Footing rules or constraints would ignore + boundaries. For example, /#loka#ina#/ would be footed as [#(ˌloka)#(ˈina)#], with a foot preceding each non-initial #, but /#lolo+fi#/ as [#lo(ˈlo+fi)#], with a single foot preceding the single non-initial #. All prefixes would bear #, to prevent diphthongisation or vowel coalescence: [#fe#(ˈita)#] (assuming that footing rules/constraints prevent a foot from taking material from both sides of a # boundary).

Given that a difference in prosodic behaviour between monomoraic and bimoraic affixes exists in many other language (see §2.2 for examples), we believe it is an advantage of the two Align approaches that this difference is captured in the grammar rather than being an accidental property of affixes’ lexical entries, as in the output–output, Lexical Phonology and boundary-symbol approaches.