Labial palatalization in Zulu: Dissimilation without the OCP Wm. G. Bennett [email protected] 0. Introduction (1) Zulu1 has a process of ‘palatalization’ that changes labial stops & nasals to (alveo)-palatals; it is triggered by the passive suffix /w/, as shown in (2)–(3).2 (2) a.

(3) a.

uku-ɡub-a

b.

uku-ɡud͡ʒ-w-a

INF-dig-IND

INF-dig-PASS-IND

‘to dig’

‘to be dug’

uku-khumul-a

b.

uku-khuɲul-w-a

INF-undress-IND

INF-undress-PASS-IND

‘to undress’

‘to be undressed’

(4) This seems to be a case of dissimilation: the labial [b] in (2a) changes to nonlabial [d͡ʒ] in (2b) because it is incompatible with the labial [w] in the suffix

Outline & goals of this talk

(5) Present the data & generalizations about Zulu labial palatalization (6) Show how the theory of surface correspondence (Walker 2000, Hansson 2001, Rose & Walker 2004) can produce dissimilation (without the OCP) (7) Propose a correspondence-based analysis of the dissimilation pattern in Zulu (8) Show how the surface correspondence analysis can capture the Zulu facts better than an analysis based on the OCP 1

Zulu also known as isiZulu; Bantu, group S.40 (Nguni); spoken in South Africa. Data given here is from Doke (1927/1984), Khumalo (1987), Beckman (1993), and Poulos & Msimang (1998). Examples from Khumalo (1987), except where otherwise noted. Abbreviations: INF = ‘infinitive’, IND = ‘indicative (mood)’, PASS = ‘passive’, CAUS = ‘causative’, APPL = ‘applicative’, EMPH = ‘emphatic’ 2

* Ideas and proposals in this talk have benefited substantially from discussion and feedback from a number of individuals; I would like to thank in particular Paul de Lacy, John McCarthy, Shigeto Kawahara, Alan Prince, Jeremy Perkins, P. Houghton, Aaron Braver, Toni Cook, Zukile Jama, and participants of the HUMDRUM 2008 workshop at Rutgers.

18th Manchester Phonology Meeting May 20, 2010

Preview of the proposed analysis

(9) A constraint CORR-[lab] demands surface correspondence between labials (10) A constraint CC-EDGE forbids surface correspondence across morpheme edges (11) The ranking CORR-[lab], CC-EDGE » IDENT-[lab] makes labials dissimilate to avoid having to correspond across the boundary between the root & passive suffix /w/

1. The phenomenon: long-distance labial dissimilation (12) Labials in Zulu dissimilate to (alveo)-palatals before the passive suffix /w/, but preserve their other features (nasality, laryngeal features, etc.). (13) Inputs & outputs of palatalization: (Doke 1927/1984, Poulos & Msimang 1998)3 [p(ʼ)] →

[t͡ʃ]

[ph] →

[ʃ]

[mp’] →

[nt͡ʃʼ]

[m] →

[ɲ]

[ɓ]

→

[tʃʼ]

[b]

→

[d͡ʒ]

[mb] →

[nd͡ʒ]

(14) Only non-continuants are palatalized; fricatives (and suffix /w/) remain labial:4 (15) a.

‖ov-a

b.

i-ja-‖ov-w-a

knead-IND

3rd.SG.CL5-PRES-knead-PASS-IND

‘knead!’

‘it is being kneaded’

(*ija‖oʒwa)

3

Khumalo (1987) argues that the only laryngeal features in Zulu are [±aspirated] and [±depressed]; under that model, there is no featural distinction between voiceless stops and ejectives ([p] vs. [p’]). The same mappings in (13) occur in two other contexts as well: before the diminutive suffix –ana, and before the locative suffix –ini. However, only root-final Cs palatalize in those contexts. I leave this issue aside here, and focus on the non-local pattern found in passives. 4 Examples in (15) from Beckman (1993). 2

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

(16) Only labials behave this way; the passive suffix does not trigger palatalization of coronals or velars (as in 17-18): (17) a.

(18) a.

kheth-a

b.

i-ja-kheth-w-a

(*ijakheʃwa)

pick.out-IND

3rd.SG.CL5-PRES-pick.out-PASS-IND

'pick out!'

'it is being dipped'

bekʼ-a

b.

i-ja-bek’-w-a

(*ijabet͡ʃʼwa)

INF-watch.over-IND

3rd.SG.CL5-PRES-watch.over-PASS-IND

'watch over!'

'it is being watched over'

(19) The fact that only labials palatalize suggests that this is a type of dissimilation (and not, e.g., the result of a floating feature from the /w/ suffix) (20) The palatalization triggered by the passive suffix [w] can cross a wide range of intervening material, as seen in (21)-(23): (intervening material underlined)5 (21) a.

(22) a.

(23) a.

uku-ᶢǀoboz-a

b.

i-ja-ᶢǀod͡ʒoz-w-a

(intervening C & V)

INF-dip-IND

3rd.SG.CL5-PRES-dip-PASS-IND

'to dip'

'it is being dipped'

uku-lum-isis-a

b.

uku-luɲ-isis-w-a

INF-bite-EMPH-IND

INF-bite-EMPH-PASS-IND

'to bite hard'

'to be bitten hard'

uku-phapham-is-a

b.

uku-phaʃaɲ-is-w-a

INF-wake-CAUS-IND

INF-wake-CAUS- PASS-IND

‘to awaken’

‘to be awakened’

(intervening affix)

(intervening undergoer)

(24) There is no apparent restriction on how many segments can intervene, and intervening material does not block labial palatalization (as in 23b, e.g.).

5

Examples in (17), (18), (21) from Beckman (1993); (22) from Doke (1927/1984); (23) from Poulos & Msimang (1998). Beckman (1993) gives one example of a root with 2 labials that follows a different pattern; this could reflect a dialect difference, or intra-speaker variation; Khumalo (1987) also notes there are a few lexical exceptions, for which labials do not palatalize. 3

18th Manchester Phonology Meeting May 20, 2010

(25) Root-initial labials are never palatalized by /w/; they remain labial:6 (26) a.

(27) a.

uku-ɓon-a

b.

uku-ɓon-w-a

INF-see-IND

INF-see-PASS-IND

‘to see’

‘to be seen’

uku-mb-a

b.

uku-mb-i-w-a

INF-dig-IND

INF-dig-PASS-IND

‘to dig’

‘to be dug’

(*ukud͡ʒonwa)

(*ukund͡ʒiwa)

(28) Zulu allows roots with multiple labials. In passives of these roots, initial labials are faithful, but any other labials in the root are palatalized: (29) a.

(30) a.

(31) a.

(32) a.

(33) a.

uku-bamb-a

b.

uku-band͡ʒ-w-a

INF-catch-IND

INF-catch-PASS-IND

‘to catch’

‘to be caught’

uku-bem-a

b.

uku-beɲ-w-a

INF-smoke-IND

INF-smoke-PASS-IND

‘to smoke’

‘to be smoked’

uku-ɓoph-el-a

b.

uku-ɓoʃ-el-w-a

INF-tie-APPL-IND

INF-tie-APPL-PASS-IND

‘to tie for’

‘to be tied for’

uku-popol-a

b.

uku-pot͡ʃol-w-a

INF-examine-IND

INF-examine-PASS-IND

‘to examine’

‘to be examined’

uku-phapham-is-a

b.

uku-phaʃaɲ-is-w-a

INF-wake-CAUS-IND

INF-wake-CAUS- PASS-IND

‘to awaken’

‘to be awakened’

(repeated from 20)

6

Examples in (26-27) from Khumalo (1987). The passive form in (27b) has an epenthetic [i], which appears with short (i.e. mono-syllabic) roots: e.g. uku-ɮ-a, ‘to eat’ ~ uku-ɮ-i-w-a, ‘to be eaten’. 4

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

Summary of Zulu generalizations:

(34) The Zulu passive suffix /-w-/ causes any & all non-initial, non-continuant labials in a root to surface unfaithfully, as palatals (35) This labial dissimilation applies non-locally (e.g. as in 32, 33), and is not sensitive to any material between the suffix [w] and labials in the root (36) Root-initial labials (and labial continuants) are unaffected; they remain faithfully [labial], even when other labials in the word dissimilate

2. How surface correspondence produces dissimilation (37) Surface correspondence has been used to analyze consonant harmony patterns (Walker 2000, Hansson 2001, Rose & Walker 2004), since it effectively captures several properties typical of consonant harmony patterns: a. similarity between triggers and undergoers b. non-locality between the segments involved c. insensitivity to interveners/blocking (38) Long-distance [place] dissimilation frequently shares the same properties (Shaw 1991, Hansson 2001). Zulu labial dissimilation shares these properties too.

How Surface Correspondence works for consonant harmony:

(39) A family of CORR constraints require correspondence between segments which have some feature(s) in common.7 e.g.: CORR-[+son]: ‘assign one violation for each pair of [+sonorant] segments in the output that are not in surface correspondence’

7

I follow the model of CORR-[F] constraints proposed by Rose & Walker (2004), where correspondence constraints are defined in terms of shared features, rather than in terms of divergence (cf. Hansson 2001) 5

18th Manchester Phonology Meeting May 20, 2010

(40) A second family of constraints, CC-IDENT-[F], require segments in a correspondence relationship to have the same specification for some feature8 e.g.: CC-IDENT-[nas]: ‘assign one violation for each pair of segments that are in surface correspondence, but have different values of [±nasal]’ (41) Ranking CORR & CC-IDENT-[F] over IDENT can produce assimilatory harmony: the optimal candidate violates (IO)-IDENT to make the right correspondence relation (42) Hypothetical example: Nasal harmony among sonorants /kalana/ a. → kanxanxa b.

kalxanxa

c.

kalxanya

CORR-[+son] CC-IDENT-[nas] IDENT-[nasal] (0)

(0)

(1)

W (0~1)

L (1~0)

W (0~1)

L (1~0)

(43) RECAP: Harmonizing sacrifices IDENT violations to satisfy CORR & CC-IDENT

Surface Correspondence also produces consonant DIS-harmony:

(44) Under certain conditions, the best way to obtain an acceptable surface correspondence relation is not to correspond. (45) Correspondence isn’t the only way to satisfy CORR: (first noted by Walker 2000) a. CORR-[F] demands correspondence between segments that share feature [F] b. An offending [+F] segment in the input can avoid CORR-[F] violations by shedding the feature [F], escaping the correspondence requirement entirely c. If correspondence is ruled out by higher-ranked constraints, then this repair strategy can be optimal 8

Note on terminology: I use CC-IDENT instead of IDENT-[F]-CC (used by Walker 2000, Hansson 2001, Rose & Walker 2004). This is purely for notational convenience. 6

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

(46) Hypothetical example: a constraint against correspondence (‘DON’TCORR’) forces dissimilation instead of harmony /kalana/ a. → kadxanya

DON’TCORR CORR-[+son] IDENT-[nasal] IDENT-[son] (0)

b.

kanxanxa

W (0~1)

c.

kalxanxa

W (0~1)

d.

kalxanya

(0)

(0)

(1)

W (0~1)

L (1~0) L (1~0)

W (0~1)

L (1~0)

(47) Undominated DON’TCORR rules out the harmonizing candidate (b), as well as candidate (c), which corresponds without agreeing for [±nasal] (48) The candidate (d), which has no surface correspondence, satisfies DON’TCORR, but violates CORR-[+son] because it has two non-corresponding sonorants (49) Candidate (a) wins by dissimilating: one [+son] segment becomes [–son] a. There is no correspondence, so DON’TCORR is satisfied b. Because there is only one sonorant, no correspondence is required: CORR-[+son] is also satisfied; only low-ranked IDENT-[sonorant] is violated (50) RECAP: if correspondence between two Cs is both required (by CORR-[F]), and prohibited (by a constraint against that correspondence), then dissimilation for [F] is the only way to satisfy the surface correspondence constraints.

3. Proposed analysis of Zulu (51) Generalizations about Zulu labial palatalization: (summarized from section 1) a. Labials in the root dissimilate to palatals when the suffix /w/ is present b. Root-initial labials do not dissimilate (nor do continuants) 7

18th Manchester Phonology Meeting May 20, 2010

(52) A surface correspondence analysis of dissimilation requires two pieces: a. A constraint that requires correspondence between the interacting segments b. A constraint that prohibits correspondence in the contexts where dissimilation occurs Constraints:

A constraint to require correspondence: (53) CORR-[lab]: ‘assign one violation to each pair of distinct [labial] consonants in the output that are not in surface correspondence with each other’ a. For Zulu, the interacting segments are all (and only) labials underlyingly. b. So, the relevant CORR constraint needs to be one that demands correspondence between [labial] Cs (those in the root, and suffixal /w/). A constraint to prohibit correspondence (and force dissimilation): (54) CC-EDGE-(morph): ‘assign one violation for each pair of corresponding segments that belong to different morphemes’ a. This constraint prohibits surface correspondence across morpheme edges b. CC-EDGE is not unprecedented: Ito & Mester (1994) propose similar constraints against interactions across edges (CRISPEDGE) c. The difference: CRISPEDGE constraints penalize autosegmental association across edges; CC-EDGE prohibits surface correspondence across edges Regular Input-Output Faithfulness: (55) IDENT-[lab]: ‘assign one violation for each [labial] segment in the input whose correspondent in the output is not realized as [labial]’ 8

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

Specialized Input-Output Faithfulness: (56) IDENT-RootInitial-[lab]: 'assign one violation for each root-initial labial in the input whose output is not labial' (57) IDENT-continuant-[lab]: ‘assign one violation for each labial continuant in the input whose output is not labial’ a. Root-initial labials are exceptionally faithful; they don’t dissimilate b. Continuants are also faithful: /w/ in the passive suffix always surfaces as [w] (never [j]), and labio-dental fricatives don’t dissimilate Ranking:

(58) CORR-[lab], CC-EDGE » IDENT-[lab] a. If both of the surface correspondence constraints dominate IDENT-[lab], then faithfulness for [labial] may be sacrificed when it allows for a better surface correspondence configuration b. A string with two labials in different morphemes must violate either CORR[lab] (for missing correspondence), or CC-EDGE (for crossing an edge) c. Dissimilation can satisfy both constraints: no cross-edge correspondence is needed if one of the labials changes to a non-labial (59) This ranking forces labials to dissimilate to non-labial in passives: / ɡub-w-a / a. → ɡud͡ʒx-wy-a b.

ɡubx-wx-a

c.

ɡubx-wy-a

CORR-[lab] CC-EDGE (0)

(0)

IDENT-[lab] (1)

W (0~1) L (1~0) W (0~1)

L (1~0)

(60) Correspondence within roots is permitted; it incurs no CC-EDGE violations 9

18th Manchester Phonology Meeting May 20, 2010

(61) So, labials in non-passives (without /-w/) don’t need to dissimilate: / bem-a /

CORR-[lab] CC-EDGE IDENT-[lab]

a. → bxemx-a b. bxemy-a

(0)

(0)

(0)

W (0~1)

c. bxeɲy-a

W (0~1)

(62) This ranking also predicts that intervening material will not block dissimilation: a. Neither CORR-[lab] nor CC-EDGE refer to the intervening phonological material; together, they assess only correspondence relations among labials. b. Consequently, intervening material will be inert relative to dissimilation: it isn’t required to correspond, and only incurs extra CC-EDGE violations (63) Correspondence relations of any interveners do not affect the outcome: / lum-isis-w-a / a.→ luɲw-isxisy-wz-a

CORR-[lab] CC-EDGE (0)

(0)

IDENT-[lab] (1)

b. lumx-isyisy-wx-a

W (0~1) L (1~0)

c. lumx-isxisx-wx-a

W (0~6) L (1~0)

d. lumx-isxisy-wy-a

W (0~1)

W (0~3) L (1~0)

(64) IDENT-RootInitial-[lab], IDENT-continuant-[lab] » CC-EDGE a. If the specialized faithfulness constraints dominate CC-EDGE, then crossmorpheme correspondence will be tolerated for these special contexts only9

9

I assume here for ease of exposition that faithfully-mapped root-initial labials are in correspondence with the [w] of the passive suffix. This is not crucial; if specialized faithfulness dominates only CORR-[lab] instead, the resulting ranking picks outputs with the same segments, but no correspondence instead. 10

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

(65) This captures the generalization that root-initial labials are faithful: / ɓon-w-a /

IDENT-Init-[lab], CORR-[lab] CC-EDGE IDENT-[lab] IDENT-cont-[lab]

a. → ɓxon-wx-a b.

ɓxon-wy-a

c.

t͡ʃʼxon-wy-a

(0)

(0)

(1)

W (0~1)

L (1~0)

W (0~1)

(0)

L (1~0) W (0~1)

(66) But, any other (non-initial) labials in the root are still forced to dissimilate when suffixal /w/ is present: / bem-w-a /

IDENT-Init-[lab], CORR-[lab] CC-EDGE

IDENT-[lab]

IDENT-cont-[lab] a. → bxeɲy-wx-a

(0)

(0)

(1)

(1)

b.

bxemx-wx-a

W (1~2) L (1~0)

c.

bxemx-wy-a

W (0~2)

L (1~0)

L (1~0)

d.

bxemy-wx-a

W (0~2)

e (1~1)

L (1~0)

e.

bxemx-jy-a

W (0~1) (cont)

f.

t͡ʃʼxeɲy-wz-a

W (0~1) (init)

e (1~1) L (1~0)

W (0~2)

Section summary:

(67) Key constraints a. CORR-[lab] demands surface correspondence between all labials b. CC-EDGE forbids correspondence across morpheme edges (68) CC-EDGE, CORR-[lab] » IDENT-[lab] a. Under this ranking, segments can give up their [labial] feature to improve on surface correspondence 11

18th Manchester Phonology Meeting May 20, 2010

b. When there are labials in two different morphemes, they necessarily violate one of the surface correspondence constraints (CORR-[lab] or CC-EDGE) c. Labial dissimilation obviates the correspondence requirement: it satisfies both CORR-[lab] and CC-EDGE at the expense of IDENT-[lab] d. When only one morpheme has labials (e.g. a root without the /w/ suffix), correspondence doesn’t violate CC-EDGE: labials can co-occur within a root (69) Root-initial faithfulness follows from IDENT-RootInitial-[lab] being ranked over the surface correspondence constraints

4. What about the OCP? (70) Most analyses of dissimilation appeal to a constraint against multiple occurrences of the same feature: the OCP (Yip 1988, Myers 1997, Fukazawa 1999) (71) Can an OCP-based analysis explain the data without needing the extra mechanism of surface correspondence?

An alternative OCP analysis:

(72) In Optimality Theory, the OCP is normally treated as a family of constraints (Suzuki 1998; see also Ito & Mester 1996 & 1998, Alderete 1997) (73) The OCP constraint relevant for our Zulu pattern would be OCP-[labial]: (74) OCP-[labial]: ‘assign one violation for each pair of adjacent [labial] features’ a. The OCP is traditionally understood to hold at the melodic level: it assesses features adjacent on their own tier b. For Zulu, this is crucial: dissimilation can apply long-distance, so the constraint responsible for it must apply to non-adjacent segments (even those separated by intervening consonants with other [Place] features) 12

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

(75) BUT: this analysis makes the wrong prediction for words with root-initial labials a. Root-initial labials surface faithfully; this entails the presence of a labial feature associated with them in the output b. If the OCP applies only at the melodic level, it cannot assess violations for multiple segments linked to the same feature c. This makes the OCP unable to distinguish between a candidate that dissimilates and one that remains re-links to a different [labial] feature. (76) Thus, OCP predicts no dissimilation when a root-initial labial is present: LAB LAB LAB

IDENT-Initial-[lab] OCP-[labial] IDENT-[labial]

/bem–w–a/ a.

☹ b.

LAB LAB LAB

(0)

(1)

(1)

e (1~1)

L (1~0)

W (1~2)

L (1~0)

L (1~0)

W (1~2)

be ɲ–w–a LAB LAB LAB

→! b e m – w – a c.

LAB LAB LAB

bem–w–a d.

LAB LAB LAB

W (0~1)

d͡ʒ e ɲ – w – a

(77) RECAP: the OCP does not penalize root-internal labials when a root-initial labial surfaces faithfully; therefore, it can’t force dissimilation in that context.

13

18th Manchester Phonology Meeting May 20, 2010

An alternative alternative: segment-level OCP

(78) The problem in (76) is that the OCP can’t assess segments directly because it only holds at the melodic level (79) One way to give the OCP this ability is to have it hold at the segmental level (along the lines of Alderete 1997, Ito & Mester 1996 & 1998, Suzuki 1998) (80) OCP’-[labial]: ‘assign one violation for each co-occurrence of [labial] segments within the stem’ (81) This revised OCP constraint favors dissimilation of root-internal labials, even when a root-initial labial is present: / bem-w-a / a. → beɲ-w-a b.

bem-w-a

c.

d͡ʒeɲ-w-a

IDENT-Initial-[lab] OCP-[labial] IDENT-[labial] (0)

W (0~1)

(1)

(1)

W (1~2)

L (1~0)

L (1~0)

W (1~2)

(82) BUT: the segmental OCP analysis predicts all labials to trigger dissimilation. a. The segment-level version of the OCP favors dissimilation in (81) because it prohibits any co-occurrence of labials, in general b. As such, this OCP assigns violations even for labial co-occurrence in a root: it predicts (wrongly) that roots never contain more than one labial (83) Segment-level OCP predicts dissimilation root-internally: (actual form (b) loses) / bem-a / a. → beɲ-a

IDENT-Initial-[lab] OCP-[labial] IDENT-[labial] (0)

b. ☹ bem-a c.

d͡ʒeɲ-a

W (0~1)

(0)

(1)

W (0~1)

L (1~0) W (1~2) 14

Wm. G. Bennett / [email protected] Labial dissimilation in Zulu without the OCP

(84) RECAP: the OCP is either too broad or too narrow to account for labial dissimilation in Zulu a. If the OCP holds over [labial] autosegments (and permits multiple linking), then it cannot compel dissimilation of only the non-initial labials in a root b. If the OCP holds over all [labial] segments, it predicts dissimilation to occur root-internally, not just in the context of the suffixal /w/

5. Summary & Conclusions (85) Zulu exhibits a pattern of long-distance labial palatalization, which is triggered by a suffix /w/ (86) This pattern can be understood as a case of dissimilation of [labial] In this talk, I have shown:

(87) That this pattern can be explained using surface correspondence (Walker 2000, Hansson 2001, Rose & Walker 2004) a. The constraint CORR-[lab] forces labials to correspond b. The constraint CC-EDGE forbids correspondence across morpheme edges c. Labials in different morphemes can only satisfy both constraints if one of them dissimilates to non-labial; then, no correspondence is needed (88) That an analysis based on the OCP is problematic: the OCP would need to both permit and prohibit labial co-occurrence within roots (89) The surface correspondence analysis solves this problem: labial co-occurrence is prohibited only when it requires correspondence across a morpheme edge (90) This analysis requires minimal added assumptions: the theory of surface correspondence is independently supported by cases of long-distance assimilation (i.e. consonant harmony). 15

18th Manchester Phonology Meeting May 20, 2010

References Alderete, J. (1997). Dissimilation as local conjunction. In Proceedings of North East Linguistics Society (NELS) 27, pp. 17-31, Amherst, MA: GLSA. ROA-175. Beckman, J. N. (1993). Feature organization and the strong domain hypothesis in Zulu [labial] phonology. In University of Massachusetts Occasional Papers, vol. 16. University of Massachusetts. Doke, C. M. (1984). Textbook of Zulu Grammar. Maskew Miller Longman, Cape Town. (First published 1927). Fukazawa, H. (1999). Theoretical implications of OCP effects on features in Optimality Theory. Doctoral Dissertation, U-Maryland at College Park, 1999. ROA-307. Hansson, G. Ó. (2001). Theoretical and typological issues in consonant harmony. PhD thesis, University of California, Berkeley. Itô, J. and Mester, A. (1994). Reflections on CodaCond and Alignment. In: Phonology at Santa Cruz, vol. 3, pp. 27-46. ROA-141. Itô, J. and Mester, A. (1996). Rendaku I: Constraint conjunction and the OCP. ROA-144. Itô, J. and Mester, A. (1998). Markedness and word structure: OCP effects in Japanese. ROA255. Khumalo, J. S. M. (1987). An autosegmental account of Zulu phonology. PhD thesis, University of Witwatersrand. Myers, S. (1997). OCP effects in optimality theory. NLLT, 15:847–892. Poulos, G. and Msimang, C. T. (1998). A linguistic analysis of Zulu. Via Afrika, Cape Town. Rose, S. and Walker, R. (2004). A typology of consonant agreement as correspondence. Language, 80:475–531. Shaw, P. (1991). Consonant harmony systems: The special status of coronal harmony. In Paradis, C. and Prunet, J.-F., editors, The Special Status of Coronals: Internal and External Evidence, pages 125–157. Academic Press. Suzuki, K. (1998). A Typological Investigation of Dissimilation. PhD thesis, University of Arizona. Walker, R. (2000). Long-distance consonantal identity effects. In Proceedings of the West Coast Conference on Formal Linguistics (WCCFL) 19, pp. 532–545. Yip, M. (1988). The obligatory contour principle and phonological rules: A loss of identity. Linguistic Inquiry, 19:65–100.

16