Kazutaka Kurisu (Kobe College)

Subtractive Morphology as Evidence for Parallelism One of the recent major issues in phonology is the comparison of parallelism and serialism in Optimality Theory (OT; Prince & Smolensky 2004), where serialism refers to Harmonic Serialism (HS; McCarthy 2007, Pruitt 2010, Kimper 2012) and OT-CC (McCarthy 2007, Wolf 2008). In this paper, I defend parallel OT, drawing evidence from subtractive morphology involved in Lardil nominative formation. Lardil nominatives are formed via final vowel deletion of a stem, as illustrated in (1a). Consonants are additionally elided when the result of final vowel deletion breaches the coda condition in the language (i.e., non-place-linked non-coronals are prohibited in codas), as shown in (1b). Finally, no deletion occurs when the stem ends in a coronal consonant, as exemplified in (1c). (1)

a. b. c.

Stems jalulu majara wulunka muŋkumuŋku kethar jaraman

Nominatives jalul majar wulun muŋkumu kethar jaraman

Gloss flame rainbow fruit wooden axe river horse

These examples can be easily handled with parallel OT. The analysis presented below slightly modifies Kurisu’s (2001) analysis. The relevant constraints are stated in (2). The REALIZE MORPHEME constraint (RM), being ranked above MAXNom., is the driving force of subtractive morphology in Lardil nominatives. (2)

a. b. c. d.

REALIZE MORPHEME: MAXNom.: CODACOND: ANCHOR-R(C,Stem):

Every morpheme must receive overt phonological exponence. No segment may be deleted in the nominative formation. Non-place-linked non-coronals are prohibited in codas. The stem-final consonant must be the rightmost segment in output.

My parallel analysis is provided below. The tableaux in (3)-(5) demonstrate that all the examples in (1) are captured with parallel OT. (3)

/jalulu/Nom. ja.lu.lu ja.lul ja.lu

CODACOND

a. b. ☞ c.

/wulunka/Nom. wu.lun.ka wu.lunk wu.lun wu.lu

CODACOND

a. b. c. ☞ d.

/kethar/Nom. ket.har ket.ha

CODACOND

a. ☞ b.

(4)

(5)

ANCHOR-R

ANCHOR-R

*!

ANCHOR-R *!

RM *!

MAXNom.

RM *!

MAXNom.

RM *

MAXNom.

* **!

* ** ***!

*

The ranking of CODACOND » ANCHOR-R(C,Stem) is confirmed by the examples in (6). Phonological final consonant elision occurs to satisfy CODACOND, so CODACOND outranks ANCHOR-R(C,Stem). (6)

Stems waŋalk kurpuɽuŋ

Nominatives waŋal kurpuɽu

Gloss boomerang lancewood

Turning to serialism, HS and OT-CC do not succeed. The two theories work alike in all crucial respects, so I discuss only HS. In HS with the same constraint ranking as above, a problem arises in (1b). Taking /wulunka/→[wulun], the most natural initial step is deletion of final /a/. The final vowel deletion leads to a fatal violation of CODACOND, as illustrated in (7). This means the convergence of the derivation on the first pass. As a consequence, no deletion is expected to occur. (7) a. ☞ b. L

/wulunka/Nom. wu.lun.ka wu.lunk

CODACOND *!

ANCHOR-R

RM *

MAXNom. *

Reranking of CODACOND and RM is not a solution. Two ramifications exist. First, if both CODACOND and ANCHOR-R(C,Stem) are ranked below RM, final consonant deletion is erroneously obtained in (1c). Second, if only CODACOND is demoted below RM, ANCHOR-R(C,Stem) outranks RM and CODACOND. No final consonant deletion is expected in this case, so the nominative forms in (6) would surface with a final non-coronal consonant, contrary to fact. McCarthy (2008) proposes that consonants be deleted via debuccalization. This proposal is not helpful here. [wu.lun.Ka] ([K]=placeless [k]) incurs gratuitous violation of HAVEPLACE and MAX[PLACE], so it is harmonically bounded by faithful [wu.lun.ka]. Again, the derivation converges on the first pass. The failure of HS and OT-CC does not depend on RM as the driving force of vowel deletion. ¬MAX-V (Alderete 1999) and FREE-V (Prince & Smolensky 2004) are conceivable alternatives, but neither salvages serialism. First, ¬MAX-V plays no role in Lardil nominatives. Anti-faithfulness constraints take effect only in OO-mapping (Alderete 1999), a necessary assumption to inhibit rampant phonological anti-faithfulness effects. In Lardil, stems cannot stand as independent words, so OO-correspondence is not invoked. Second, FREE-V, which prohibits words ending a vowel, does not make HS succeed. This analysis fails in /muŋkumuŋku/→[muŋkumu]. FREE-V needs to outrank CODACOND such that violation of CODACOND created by vowel deletion is tolerated in the initial step, as shown in (8). The output on the first pass (i.e., [muŋkumuŋk]) is the input on the second pass. As demonstrated in (9), further deletion of /k/ is blocked because (9b) still ends in a non-coronal. The derivation converges, resulting in incorrect [muŋkumuŋk]. (8)

/muŋkumuŋku/Nom. muŋ.ku.muŋ.ku muŋ.ku.muŋk

FREE-V *!

CODACOND

a. b. ☞

/muŋkumuŋk/Nom. muŋ.ku.muŋk muŋ.ku.muŋ

FREE-V

a. ☞ b. L

CODACOND * *

(9)

ANCHOR-R

*

MAXNom. *

ANCHOR-R

MAXNom.

*!

*

(9b) is made more harmonic than (9a) if *COMPLEXCODA is ranked over ANCHOR-R(C,Stem), but this fix is not successful. On the third pass, final [ŋ] is deleted to satisfy CODACOND. A problem arises on the fourth pass. /muŋkumu/ serves as the input. As illustrated in (10), final /u/ undergoes deletion to respect undominated FREE-V. Too many segments are wrongly deleted. This problem arises since the final vowel of /muŋkumu/ (not the final vowel of /muŋkumuŋku/) is seen as the stem-final segment at this point. (10)

a. L b. ☞

/muŋkumu/Nom. muŋ.ku.mu muŋ.kum

FREE-V *!

CODACOND *

ANCHOR-R

MAXNom. *

Kimper (2009) develops a HS approach to subtractive morphology appealing to subcategorization. The idea is that a designated prosodic constituent undergoes delinking via a series of morphological processes. This idea is reminiscent of positive prosodic circumscription (McCarthy & Prince 1990). In (1a, b), a final mora is subcategorized and delinked. This analysis does not invoke segmental deletion, but it encounters the same problem as the FREE-V account. For /muŋkumuŋku/, the final vowel is delinked if the constraint that requires phonological implementation of the morphological subcategorization is ranked high, yielding [muŋkumuŋk]. The final consonant cluster must be deleted phonologically, so subcategorization plays no role at this point. It follows that the problem mentioned about the FREE-V analysis arises here too. McCarthy (2007) argues that different predictions of HS and parallel OT in pseudo-Lardil favor HS. Suppose that FINAL-C (ban on word-final vowels) and CODACOND (ban on obstruent codas) outrank MAX. In parallel OT, /palasanataka/ is mapped onto [palasan]. In HS, [palasanataka] or [palasanatak] emerges depending on the ranking of CODACOND and FINAL-C. McCarthy argues that parallel OT is disfavored since massive deletion like /palasanataka/→[palasan] is unattested. This argument is invalid since HS also emits an unattested pattern. Given FINAL-C » CODACOND » MAX, /pataka/ and /patak/ surface as [patak] while /patan/ surfaces faithfully: CODACOND is satisfied only when the word ends in a sonorant consonant. To the best of my knowledge, no documented language imposes different restrictions on codas depending on the underlying final segment. If parallel OT is undesirable on the basis of pseudo-Lardil, so is HS. Kavitskaya & Staroverov (2010) discuss phonological phenomena where two processes are mutually in feeding and counterfeeding relations. They claim that such cases are accommodated in OT-CC with slight revisions of McCarthy (2007). They consider Lardil nominatives as one such case because vowel deletion feeds consonant deletion (1b), but not vice versa (1c). This is misconception because Lardil nominative formation is morphologically conditioned, unlike the cases of Kavitskaya & Staroverov’s interest. In conclusion, Lardil nominatives present empirical evidence for parallelism. They are not handled with HS and OT-CC. Examples of subtractive morphology abound across languages, so the problem with the serial theories is not language-particular. It is rather a general problem with HS and OT-CC.