Sound Change and Phonologization

Paul KiparskyStanford University

1 Sound change: some questions that a theory should answerWeinreich, Herzog, and Labov 1968 formulated five defining problems for the theory of lan-

guage change.

(1) a. The constraints problem: What are the general constraints on change, if any, that de-termine possible and impossible changes and directions of change?

b. The transition problem: By what route does language change?c. The embedding problem: How is a given language change embedded in the surround-

ing system of linguistic and social relations? How does a change in one part of thegrammar affect or get constrained by other parts of the grammar?

d. The evaluation problem: How do members of a speech community evaluate a givenchange, and what is the effect of this evaluation on the change?

e. The actuation problem: Why did a given linguistic change occur at the particular timeand place that it did? How and why do changes begin and proceed?

Every approach to historical linguistics must address them, but their theoretical interpretation, andthe range of possible answers to them, are highly theory-dependent. For example, the constraintsproblem in its modern sense arose only with structural linguistics, when it was conceived as thequestion how synchronic universals of linguistic structure and general constraints on phonologi-cal and analogical processes determine the class of possible changes. The transition problem inphonological change came into focus with the structuralist thesis that sound change is gradual,but its effect on the phonological system proceeds through abrupt structural reanalyses. The em-bedding problem spawned much research seeking to show that the phonological system channelssound change into effecting symmetry and economy of phonemic systems and dispersion (maxi-mum distinctiveness) of phoneme realization. The actuation problem, meanwhile, was held to beintractable and all but ignored in theory and practice.

These answers naturally raised new questions, and were themselves in turn challenged, partlyin response to new theoretical approaches in general linguistics. Structuralists generally did notquestion the neogrammarian thesis that sound change is exceptionless. In fact, it provided themwith the key argument for the autonomous phonemic level that they posited, which in turn wasthe principle on they relied in explaining how new phonemes arise from allophones in so-calledsecondary split.

1

Empirical study of ongoing sound change complicated the picture by uncovering cases of ap-parent word-by-word spread of sound changes (lexical diffusion), a process which had alreadybeen championed by Schuchardt in his anti-neogrammarian polemics. Another troubling featureof such changes was that they seemed to be governed by word frequency, a factor which had noplace in the neogrammarian/structuralist scheme of things. And some linguists turned the rela-tion of typology and universals on its head: rather than relying on linguistic universals to explainsound change, they assumed that sound change itself explains typological tendencies and univer-sals. Recent work based on information theory suggests that the actuation problem might not be ashopeless as had been thought (Cohen Priva 2012).

How, then, does sound change happen? How does it change the linguistic system? How is itconstrained by the system? Let us pursue these questions and the answers to them, as they evolvehand in hand with linguistic theory from Saussure through Stratal Optimality Theory.

2 Phonologization in structuralist phonology2.1 The phonologization and non-phonologization problems

If there is a phonemic level of representation, the question arises how and why phonemes origi-nate and are lost. This is the problem of PHONOLOGIZATION and MERGER. The structuralists heldthat allophones are phonologized, i.e. become phonemic, when their conditioning environment ofis eliminated by sound change. The idea was introduced by V. Kiparsky (1932) and Twaddell(1938),1 who both illustrated it with the phonologization of the front rounded umlaut vowels and in German, they argued that the umlaut vowels became distinctive when the i or j that condi-tioned them was reduced or deleted, at which point lexical representations were restructured withthe former allophonic variants as phonemes.

(2) a. Old High German (OHG): Nom.Pl. /huot-i/ heti hats, helmets Dat.Sg. /huot-e/ huote

b. Middle High German: Sound change: heti > hete Restructuring: /huot-i/ > /het-e/

This mechanism, called SECONDARY SPLIT (Hoenigswald 1965), has been standardly assumedto account for phonologization. However, it leaves two questions unanswered. First, when theconditioning environment goes away (as here by reduction of the full vowels to -e or to -@), whydo its effects remain? Why did the front vowels not become back again, why did the frontnessstay, once the influence of /i j/ was removed? (Liberman 1991: 126). Secondly, why does theloss of a conditioning environment not always cause phonologization? Why do the conditionedallophones sometimes just go away?2 Lets call them the PHONOLOGIZATION PROBLEM and theNON-PHONOLOGIZATION PROBLEM, respectively.

1Kiparskys article is an introduction to synchronic and historical structural phonology, written after his stay inPrague where the new theory was just then taking shape. It correctly avoids Twaddells assumption that OHG orthog-raphy was phonemic.

2Why do allophones sometimes remain and other times revert? (King 1971: 4).

2

Both problems are deep ones, and reach the foundations of phonological theory. The phonol-ogization problem has been at the core of phonological theorizing, but has no generally agreed onsolution. The non-phonologization problem has barely begun to be discussed. I begin by reviewingthe two main proposed solutions to the phonologization problem, and suggest a new one based onLexical Phonology and more particularly its OT version (Stratal OT), which I argue also resolvesthe non-phonologization problem.

The best-known proposal for resolving the phonologization problem is due to Saussure andBloomfield. It depends on a paradoxical marriage of synchronic structuralism to diachronic neogram-marianism. The key idea is that a sound change is located outside of the linguistic system that itwill transform.

An alternative that is increasingly beginning to be explored is to enrich the phonology withphonetic information, abandoning the concept of a phoneme as a contrastive entity, and positingthat phonemes-to-be get in some sense phonologized before they become contrastive through theloss of the conditioning factor. This entails a break with structuralism, where contrast is the centralconcept, but turns out to fit well into the generative view of language as a computational system.

After reviewing these solutions in turn, I will argue for a version which combines aspects ofboth, implemented in the Stratal OT framework.

2.2 Saussures FirewallFor Saussure, the basic fact about language is the ARBITRARINESS OF THE SIGN. Among its

consequences are a sharp divide between synchrony and diachrony. This division is not merelymethodological, or practical, or based on conventional boundaries between academic subdisci-plines. It is a conceptually necessary consequence of arbitrariness.

Linguistics must be separated in two. There is an irreconcilable duality, created bythe very nature of things [. . . ] in systems of values. (C3:104, cf. C:79:80).

A historical event such as a sound change qua phonetic mutation, and the restructuring of thephonological system it may cause, are totally different things: a sound change can bring aboutradical discontinuities, or it could have no effect on the system whatever.

There is no inner bond between the initial fact [the phonetic change] and the ef-fect that it may subsequently produce on the whole system [phonology or grammar].(C:87)

The sharp segregation of historical change from synchronic structure (call it SAUSSURESFIREWALL) implies that, although everything in grammar is interrelated as a system, sound changehas no access to that system. Blindly and structure-independently, it alters merely the materialimplementation of speech.3 The abstract synchronic system, characterized by networks of rela-tions and systems of constraints, is affected only indirectly by those alterations. For Saussure,the synchronic constraints in the mind of the speaker and the historical processes that modify the

3A modern version of this idea is eloquently stated in Hale 2007.

3

articulation of speech are not only ontologically, but also formally distinct. Constraints are GEN-ERAL (transparent, or inviolable, in current terminology), whereas processes are ACCIDENTAL andPARTICULAR. Constraints are PRECARIOUS (they could be overturned by the next change), whileprocesses are IMPERATIVE (sound change is exceptionless).

Back to the example of umlaut: the phonetic mutations of umlaut and reduction/syncope bothaltered the physical aspect of speech, but they had very different effects on the system. The struc-tural reflex of umlaut was purely phonetic: it introduced the constraint no back vowels before i.This is an allophonic distribution and has no bearing on the phonological system. Vowel reductionand subsequent syncope had no phonetic repercussions on the umlaut vowels, but an all the moredrastic impact on their phonological status: it caused them to be reanalyzed as distinct phonemes.The new phonetic givens lead to a restructured phonological system with new phonemes //, //and a new constraint no unstressed full vowels. The site of phonemic contrast has shifted onesyllable to the left.

This is a consistent theory of sound change. But the dualist ontology of Saussures Firewall is aheavy price to pay. It essentially makes the constraints problem insoluble, at least for sound change,by excluding all structural explanations of the sort pioneered by Jakobson and since pursued indifferent ways by Martinet, Labov and others, and, still differently, in generative and OT work.In particular, it makes inexplicable the central fact that sound change never subverts phonologicaluniversals including implicational universals and that sound changes are overwhelminglynatural processes.

Another objection to this theory is that it offers no solution to the non-phonologization problem.Why does the predicted secondary split sometimes not happen, and the conditioned allophonesjust disappear? For example, English has a front and back allophones of /k/ and /g/ dependingon whether they are tautosyllabic with a back vowel or a front vowel. So we would expect vowelfronting in various dialects of English (e.g. in calf, goat, cough), and vowel backing (girl, dialec-tally in kit), to produce contrasts between front and back k. The mystery is why this happens sorarely (an exception is the Jamaican English contrast cat [kyat] vs. cot [kat] or [k6t], Wells 1982:569), and why that fact was rarely regarded as a problem for structuralist doctrine. On the contrary,when a linguist proposed that umlaut vowels disappeared in Scandinavian when the triggering frontvowels were syncopated, Benediktssons (1982: 9) objected to it on principle: The principle thatphonetic variants, in consequence of the conditioning factors, may revert to the neutral starting-point [. . . ], though perhaps consistent with generative theory, seems hardly compatible with thoseof structural phonology; at any rate, if it is accepted, the principle of phonemicization is then re-duced to an ad-hoc postulate, of little or no explanatory value. Since this issue falls under theheading of the actuation problem, which was assumed to be insoluble anyway, the failure to ad-dress it was not considered a weakness.

Finally, Saussures Firewall is also a hindrance to resolving the embedding problem, specif-ically that part of it which concerns how sound change is constrained by the linguistic system.A general instance is the PRIMING phenomenon, that languages tend to have stable phonologicaltypologies (e.g. the prosodic system of the Germanic languages, and the tonal systems of Sino-Tibetan languages are enduring characteristics of the families as a whole).

4

Here belongs also another type of rule insertion, also problematic for the structuralist accountof secondary split4 where a sound change interacts with, and is constrained by, existing phonolog-ical processes and constraints in the language. An example is Old English syncope and voicingassimilation. In Old English, obstruents were obligatorily devoiced next to voiceless obstruents.This process was postlexical since it applied across word boundaries (Moulton 2003): compoundssund + corn suntcorn, med + sceatt metsceatt. Assimilation across words in a sentenceis not written but is inferred from indirect evidence (Luick 1921-1940). During the Old Englishperiod, the language undergoes syncope of unstressed vowels in word-final syllables. Syncopenever produces adjacent obstruents of unlike voicing (Campbell 299, Moulton 2003). Such clus-ters are assimilated as part of the sound change, e.g. bidest (> *bidst) > bitst you bide, ky:est*ky:st ky:Tst you inform, kyss-ide *kyssde kyste kissed. Saussures Firewall, however,predicts that syncope should extend the voicing opposition to what was until then a neutralizingassimilation environment, creating a contrast between previously existing assimilated clusters suchas -ts- and new clusters from syncope such as *-ds-. This is certainly a conceivable outcome, butit is not what happened in Old English; there is no reason to believe that clusters such as *-ds-ever existed, even immediately after syncope, as King 1973 noted. Syncope FEEDS the previouslyexisting voicing assimilation rule, so that the outcome is bintst, bitst.

The same difficulty arises with sound changes that are BLOCKED by existing synchronic con-straints. For example, syncope can fail to apply just in those cases where it would create a prohib-ited stress configuration (e.g. a lapse or clash), or a prohibited syllable structure or foot structure.In English, the variable pre-sonorant syncope in words like generative is inapplicable before astressed syllable, as in generate (*genrate), where it would produce back-to-back stresses, whichare disfavored in English. Technically, such conditions on sound changes can be specified as con-ditioning factors, but only at the cost of a loss of the generalization that the conditioning factorsare manifestations of active phonological constraints of the language.

The first two types of problematic cases are the historical analogs of the two types of transparentrule interaction in synchronic phonology: vowel backing in calf BLEEDS k-fronting, and syncopein bidest FEEDS voicing assimilation. The third type of problematic case also involves transparentinteraction, in the sense that sound change avoids creating surface exceptions to a constraint that isoperative in the language.

In short, sound changes can interact transparently with existing processes. Such transparentinteractions can involve feeding or bleeding by the sound change, or blocking of the sound changeby a constraint. Alongside such transparent interactions, sound change can also result in opacity,which in terms of change means phonologization and the creation of new contrasts. Structuralisthistorical phonology has privileged the latter scenario to the point of all but ignoring the well-documented possibility of transparent interaction.

Classical generative grammar was able to unify these cases by generalizing the processualapproach and modeling sound change as the addition of rules, treating such cases descriptivelyas rule insertion. Addition of rules at the end of the phonological component correspondedto neogrammarian sound change; stipulating that they must be added at the end replicates Saus-sures Firewall, and allowing them to be added before the end accounted for the cases where soundchanges interact with existing rules, albeit in a purely descriptive way. It was recognized that rules

4As well as for the theory espoused by Blevins 2004, see Kiparsky 2006.

5

cannot be added anywhere, but saying that they are added to the end of the grammar is arbitrary,and in any case fails to do justice to the cases of rule insertion. But once rule insertion is recog-nized, we have lost the only theoretical understanding of phonologization by secondary split. Post-structuralist theories which relate historical and synchronic phonology have been unable either toreplicate Saussures Firewall without some extrinsic stipulation, or to derive the generalization thatit is intended to capture in some other way.

3 Constraint-based phonology: Optimality theoryThe interaction of sound changes with existing phonological constraints can be dealt with read-

ily by Optimality Theory (Prince, Smolensky, MCarthy). In this theory, underlying forms aremapped into phonetic forms not by step-by-stem conversion but by filtering the possible outputsthrough a ranked set of constraints and selecting the best fit.

(3)

The principles of the theory are summarized in (4).

(4) a. Constraints are VIOLABLE, but violation is MINIMAL.b. There is a UNIVERSAL inventory of constraints (and possibly also language-specific

constraints). A is preferred to B by a constraint C if A contains fewer violations of Cthan B does.

c. Constraints are RANKED on a language-specific basis. Output A is more OPTIMALthan output B if A is preferred to B by a constraint C and there is no higher-rankingconstraint that prefers B to A.

d. PARALLELISM: There are no rules and no stepwise derivation. All possible outputs arecandidates for simultaneous evaluation.

Sound changes can then be regarded as the promotion of markedness constraints. This solvessome of the above problems straightforwardly. First, constraints can both trigger and blockprocesses. Very schematically:

(5) a process P Q is triggered in the context X___ Y if *XPY *Q, a process P Q is blocked in the context X___ Y if *XQY *P.

6

(6) The theory predicts that sound changes can interact with existing processes in two ways.a. They can feedexisting processes (provide inputs to them) just in case they are active

in the postlexical phonology.b. They can be blocked by postlexically active constraints (but not by lexical constraints).

(7) A famous problem in Germanic phonologyGermanic OHG Old Icel.

a. Light root: *wariom werita vara I protectedb. Heavy root: *warmiom warmta verma I warmed

(8) a. The North Germanic distribution is surprising. Why would light syllables not undergoumlaut?

b. Kock (1888) posits two periods of syncope and two periods of umlaut, with the weightrestriction applying at the first round of syncope, at which point only deleting vowelstrigger umlaut.

1. Unstressed i is syncopated after a heavy syllable, but triggers umlaut.2. Then unstressed i is syncopated after a light syllable, and doesnt trigger umlaut.3. Remaining is are not syncopated, and trigger umlaut.

The two umlaut periods would have to be interrupted by a period where umlaut wasturned off, so that there could be no historical connection between them.

(9) Word minimality governs syncope.a. Germanic feet must have at least two moras, and words must contain at least a foot, so

words must have at least two moras.b. Hence there are no [C V] words.c. Because final consonants are weightless, there are no [C VC] words either.d. Therefore, In all early Germanic languages, word-final syncope is restricted to heavy

and polysyllabic stems (Riad 1992). Deletion after a light syllable would lead to asubminimal foot.

(10) a. FOOTBIN: A foot must contain at least two moras.b. NONFINALITY: A word-final consonant is weightless.

(11) Effect of two-mora foot minimum (parentheses mark final weightless consonants)a. No final syncope: [CV.CV] 6 [CV(C)].b. Final syncope OK: [CVV.CV] [CVV(C)].c. Final syncope OK: [CV.CV.CV] [CV.CV(C)].

7

(12) Old Norse (but not West Germanic) prohibits superheavy (3-mora) syllables.*: A syllable rhyme is maximally binary.

(13) a. Syncope OK: [CV.CV.CV] [CVC.CV], [CV.CVC] [CVC(C)]b. No syncope: [CVV.CV.CV] 6 [CVVC.CV], [CVV.CVC] 6 [CVVC(C)]

Because final -C is weightless, word-final codas can then be longer than medial ones byexactly one consonant, with final vowel deletion applying to that extent more freely thanmedial vowel deletion.

(14) runic Old Norseasugsalas Asgsls (compound PN)wanaraas Vandras (compound PN)

(15) a. Heavy stems: *hir.joo > NGmc *hir.o (ON hira)b. Light stems: *ni.joo > NGmc *ni.jo (ON nija).

(16) a. *ltl-er > litler littler (cf. ltell little)b. *mn-r > minn mine (Nom.Sg.Masc.), cf. Fem. mn (remember that final -C does not

count)

(17) a. Kragehul Asugisalas (< *-gs.las).b. Tune worahto (< *worh.to); cf. Tjurk wurte, By orte, Slvesborg urti, where the same

constraint is implemented by deletion, as in literary Old Norse.c. Istaby wulafR, Stentoften wolafR (wulfR would have a three-mora syllable, since -R is

weightless).

(18) a. West Germanic *[tal.jan] > *[tell.jan] (OHG zellen)b. North Germanic *[tal.jan] > *[tel.jan] (ON telja)

(19) Our theory suggests that looking for the environment of syncope is barking up the wrong tree.The right question is: how was syncope constrained by syllable and foot structure onstraints*, FOOTBIN, and NONFINALITY? This way of asking the question immediately solvesthe classical problem.

(20) STAGE 0 (550 A.D.): no syncope.a. Medial syllables:

Wiwila [wwila] (Veblungsnes, ca. 550, KJ 126; or ca. 500, K 172), aluko (?) (Frde,550?, KJ 109), raisido-kA I raised (Ellestad).

8

b. Final syllables: WidugastiR (Sunde, ca. 500, KJ 198), SaligastiR, aliR (Berga, ca. 500,KJ 193), haukouR hawker (hawkeye) (Vanga, ca. 500, KJ 147), erilaR priest (?),wilagaR cunning (Lindholm, early 6th c., KJ 69), Asugisalas, erilaR (Kragehul, early6th c., KJ 64), Harabana, erila, waritu [wartu] I write (Jrsberg, 500-550, KJ 156),irilar, Wiwila [wwila] (Veblungsnes, ca. 550, KJ 126; or ca. 500, K 172), laiigaR(Mgedal, 500-550, KJ 195), hiwiga, -winaR (Arstad, ca. 550, KJ 130), SigimaraR[-mara] (Ellestad).

(21) STAGE 1, 550-600: syncope where * and FOOTBIN permit; final -C weightless.a. Syncope in medial syllables: two-mora maximum allows - V, -VC, but not - VC or

-VCC*satido > sate [satte] (Gummarp, ca. 600, KJ 205; B 141), Nom.Sg. *hrorijaR (>*hrorjaR) > HroreR (By, 550-600, KJ 158; ca. 500, B 176, 3.Sg.Opt. *wtije (>*watje) > wate [wate] wet! (Strm, ca. 600, KJ 110), *wiwijon > wiwjo (Eikeland,ca. 600, KJ 47; ca. 550 on archeological grounds according to B 84), *hawiu > hau[hau] mowing (Strm, ca. 600, KJ 110; 550-600, B 176), *wulfijaR (> *wulfjaR) >*wulafjaR > -wulafiR (Istaby, ca. 625, KJ 218; ca. 590, B 142).

b. Syncope in final syllables: two-mora maximum plus weightless final -C. Allows fi-nal - V, - V(C), -VC(C), disallows - VC(C) or -VCC(C), *laou > lao invitation(Acc.Sg) (Halskov, date?, Krause 1971: 149), 1.Sg.Pres. *fahi(j)u > fahi [fahi] I de-pict (Noleby, late 6th c., KJ 148, B 176; sum, 550-600, KJ 267), *wiwaR > wir[wr] (Eikeland, ca. 600, KJ 47; ca. 550 on archeological grounds according to B 84),*wulfaR (> *wulf R) > -wolafR (Stentoften, older part of the inscription, ca. 600?, KJ209; ca. 590, B 142), *wulfaR (> *wulf R) -wulafR (Istaby, ca. 625, KJ 218, ca. 590, B142), *ftiR (> *ftR) > AfatR after (Istaby).

c. The bimoraic foot minimum blocks syncope in C VCV words (including as parts ofcompounds).alu (Kinneve, late 6th c., KJ 114, Krlin, late 6th c., KJ 105), -ekA I (Ellestad) Hari-,Hau. . . (Istaby, Stentoften).

d. Where glide deletion and epenthesis are inapplicable, the bimoraic syllable maximumblocks medial syncope after heavy syllables.HroraR (By, 550-600, KJ 158), fahide [fahide] depicted (Halskov).

(22) STAGE 2, 600-800: syncope no longer subject to *, but remains subject to FOOTBINand weightlessness of -C.

a. First instances of new extended syncope, producing final - VC(C) and -VCC(C) (tri-moraic syllables, plus weightless -C).*habukaR > *haukaR > -haukR hawk, second part of compound name (Vallentuna,before 650; perhaps ca. 600, B 91), *hrou-waldaR > RhoaltR (Vatn, 700 or earlier, KJ152, B 176), *taitaR > TaitR (Tveito, 7th c. or later, KJ 202, B 176), *brytir > bArutR[brytR] breaks, *lausaR > -lausR loose, *haidiR > haidR brightness (Bjrketorp,

9

ca. 675, KJ 217; ca. 590, B 142), *fiskaR > fiskR fish, *mennaR > manR [mennR]men (Eggja, ca. 700, KJ 227; ca. 650, B 114), *uddaR > UdR [udd], (Roes, ca.750, KJ 235), *woinaR > Uin [uinn], *unninR > unin [unninn] (Ribe, before 750,B 313; ca. 720, Grnvik 1999), *wulfas > -wulfs wolfs (Rvsal, ca. 750, KJ 184;B 334 however dates it to the 9th century, in which case it would belong in the nextgroup, Stage 3 below).

b. Still no syncope in C VCV words (including as parts of compounds).sunu son (Dat.) (Slvesborg, 750-800, N 100; Helns, Rk, 800-850; cf. B 222, 345,291), alu (Setre, 7th c.. KJ 114), kuumut [guumund] (Helns, B 345).

(23) STAGE 3, after 800: full syncope, unconstrained by or FOOTBIN.ut:sta:ni (compound, Acc.Sg.) [ni] offspring, [sta] settlement (Gimsy, early 9th c.,B 324), sun son (Acc.Sg.) (Rnninge, Tryggevlde, ca. 900, Nielsen 2000: 98; Mejlby,Haddeby, 10th c.), ver man (Acc.Sg.) (Tryggevlde, Glavendrup, ca. 900).

(24)The lexical phonology

Input Output AGR(BACK) * * IDENT(Back)tal-i-a tal-i-a * ***

tel-i-a *** * tal-a **

tel-a ** *dom-i-a dom-i-a * ***

dm-i-a *** *dom-a * **dm-a * ** *

At first, the postlexical phonology has the same constraint ranking as the lexical phonology.

(25) IDENT(Back) is promoted, causing umlaut to stop applying postlexically; at this point backvowels are restored in postlexical configurations such as (35), and the umlaut vowels , become quasi-phonemes, potentially but not yet actually contrastive within lexical words.

The postlexical phonology: second phaseInput Output IDENT(Back) AGR(BACK) * *tala talia * ***

telia * *** tala **

tela * **dmia domia * * ***

dmia ***doma * * **dma * **

10

(26) Overt contrasts between umlauted and nonumlauted vowels are created. This happens at thelatest at Stage 2 of syncope, when syncope is generalized by promotion of * over * inthe postlexical phonology (its relative ranking with respect to IDENT(Back) and UMLAUT isimmaterial). From now on it applies even when three-mora syllables result, as in dma.

The postlexical phonology: third phaseInput Output * IDENT(Back) AGR(BACK) *tala talia *** *

telia *** * tala **

tela ** *dmia domia *** * *

dmia ***doma ** * *

dma ** *

Secondly, constraints are universal and constitute a theory of linguistic typology. Therefore,in so far as the constraint set is empirically correct, it is predicted that sound changes must benatural processes. The exceptionlessness of sound changes also follows if we assume that lexicalrepresentations remain invariant under the actual sound change (though they may be restructuredlater as a result of the change).5

However, if we move from structural levels or generative rules to OT constraints, we have lostthe solution to the phonologization problem: we no longer have any theoretical understanding ofhow secondary split is possible.

King (1971) argued that sound changes interact transparently only with phonetic rules thetrivial case of rule insertion, as he called it. His observation has held up well; the non-trivialcases have been fairly convincingly explained away (see most recently Jasanoff 2003). It is fair toassume that the phonetic rules of Kings generalization are a languages postlexical rules. If so,we can rephrase his generalization like this:

(27) a. SECONDARY SPLIT: Sound changes render lexical processes opaque.b. BLOCKING AND RULE INSERTION: Sound changes interact transparently with postlex-

ical processes.

The second, less well explored approach to the phonologization problem assumes that prospec-tive phonemes are already phonologized by the time they become contrastive (Ebeling 1960, Ko-rhonen 1969, Liberman 1991). For example, if the umlaut vowels are already phonemes (or QUASI-PHONEMES, as Korhonen calls them) before the -i- that conditions them is lost, then they wouldnaturally remain unaffected by the latter sound change.

5Labovs recent results confirm the view of sound change as a phonetically driven process that affects all wordsin a phonologically defined set. The close study of them reveals tham to be just as Paul, Leskien, Osthoff, Brugmann,Sauusure and Bloomfield described them. The tantalizing glimpses of lexical peculiarities in the data seem to bestatistical accidents (Labov 2010:285.

11

This approach to phonologization requires a way of specifying when and why non-contrastivefeatures become assigned in the lexical phonology why allophones become quasi-phonemes independently of the post hoc information that they are phonologized when another sound changeoccurs. It has been suggested that features tend to be phonologized if they belong to a feature classwhich is already distinctive (Kiparsky 1988). Though generally consonant with observations aboutpriming effects in sound change, this idea is not precise enough to make predictions about whenphonologization will take place. Another suggestion, made by Janda (2003: 413) in a vigorousplea for early phonologization, is that allophones become quasi-phonemes for reasons having todo with phonetic distance, though he does not say how much distance, and on what dimension, orcite evidence that distance matters at all.

A starting point for a more substantive theory of phonologization might be Jakobsons obser-vation that allophonic properties can become perceptually more salient than the phonemic onesthat condition them (Jakobson, Fant, and Halle 1952). Russian [1] and [i] are allophones of /i/ afterrespectively back and front consonants, yet the allophonic vowel distinction is a more salient cue tothe contrast than the phonemic consonantal one (especially in the case of sibilants because of theirhigh-frequency noise, e.g. /si/ [s1] and /si/ [si]). Related to this perceptual saliency of the vowels,as Jakobson pointed out, is the fact that [1] and [i] are perceived as categorically distinct elements,and even reified in the metalinguistic terms [1kat] to pronounce [1] and [ikat] to pronounce [i]).The vowels [1] and [i] are like two phonemes in that any unrounded high vowel token is assigned toone or the other type; perceptually they are two distinct categories. Other Russian vowels are alsostrongly affected by palatalization: e.g. /a/ is fronted towards [] to varying degree before, after,and most of all between palatalized consonants, but the allophones are apparently not categoricallyperceived as belonging to two types; correspondingly there is no *[kat] to pronounce [].

A plausible hypothesis is that allophones become quasi-phonemes when they become governedby categorical rather than gradient constraints (Flemming 2001), and acquire greater perceptualsalience than their conditioning environments. How are these two properties related to each other,and how we can build a theory of phonologization on them? In structuralist phonology, categorial-ity and saliency is attributed to phonemic representations. Feature specifications at the phonemiclevel are understood as categorical, while allophonic/postlexical feature specifications may be gra-dient. And phonemic representations specify all invariant distinctive features of the language. Butquasi-phonemes are not allowed at the phonemic level because of the rather fundamental propertythat it excludes redundant, predictable feature values from lexical representations.

4 Stratal OT4.1 Secondary split the Stratal OT solution

How are these issues addressed in constraint-based theories such as OT, which eliminate pro-cesses in favor of constraints, and model sound change as the promotion of markedness con-straints? Here it is important to distinguish parallel and stratal versions of OT. As we shall see,the generalizations in (28) are predicted by Stratal OT. According to this theory, constraints in-teract transparently within a level. Promotion of a postlexical constraint will therefore lead tonon-phonologization (blocking and rule insertion) effects. Blocking arises when the promotedpostlexical constraint is dominated within the postlexical phonology by an antagonistic constraint(e.g. syncope by a restriction on syllable structure). Rule insertion (a misnomer in this frame-

12

work, of course) arises when the promoted postlexical constraint winnows away candidates thatwould otherwise emerge as winners by the lower-ranking postlexical constraints.

Phonologization (secondary split), on the other hand, takes place because constraints do notinteract transparently with constraints at earlier levels. It is the masking of a lexical process bya postlexical one, that is, by the sound change qua promoted markedness constraint. From theperspective of Stratal OT, then, the reason why the umlaut vowels became phonologized when thetriggering context was lost is that they were introduced in the lexical phonology.

Two predictions immediately follow: the masked process (the source of the future phoneme)must be lexical, therefore restricted to the word domain, and interacting transparently with all lex-ical constraints, and the masking process (the trigger of the phonemicization) must be postlexical,therefore applicable across words, and and interacting transparently with all postlexical constraints.

Before going into the details, let us emphasize that this solution is not available in parallel OT.As far as I can see, parallel OT actually has no coherent characterization of secondary split, forreasons which are homologous to its failure to deal with opacity. To see why, consider a bare-bonesOT constraint system for the pre- and post-phonologization stage of umlaut.

(28) a. AGREE(FRONT): no back vowels before i, j (the constraint that enforces umlaut).b. IDENT(Hi): underlying high vowels are realized as high.c. *, *: rounded vowels are back.d. REDUCE: no full (unreduced) unstressed vowels.

(29)Stage 1: allophonic umlaut

Input Output AGREE(FRONT) IDENT(Hi) *, * REDUCEuCi uCi * *

Ci * *uCe *Ce * *

uCe uCi * * *Ci * * *

uCeCe *

Ce uCi * * *Ci * * *

uCeCe *

Vowel reduction results from promoting REDUCE over IDENT(Hi). But on any ranking, this undoesumlaut:

13

(30)Stage 2: promotion of REDUCE (wrong!)

Input Output REDUCE AGREE(FRONT) IDENT(Hi) *, uCi uCi * * *

Ci * * uCe *

Ce * *

The bottom line is that Saussures Firewall has no place in constraint-based theories such asOT. This is arguably an advantage because, as noted above, it is stipulative, deprives us of goodanswers to the constrants problem, and runs into several empirical problems.

Stratal OT phonology provides a more articulated theory than parallel OT in that it incorpo-rates Lexical Phonologys stratal organization (level-ordering) to OTs parallelism of constraintinteraction (Booij 1996, 1997, Orgun 1996, Kiparsky 2000; for diachronic phonology, see espe-cially Bermdez-Otero 1999, 2006a, 2006b, Bermdez-Otero and Hogg 2003). Stratal OT doesnot in principle banish predictable feature values from lexical representations. Rather, it claims thatlexical representations are determined by best satisfaction of the lexical phonological constraints.They will include such redundant feature values as those lexical constraints may assign. For thisreason they can accommodate quasi-phonemes.

For Stratal OT, the grammar is a hierarchy of serially related modules, each of which is aparallel constraint system of the classical OT type.

(31) Stem phonology

Word phonology

Postlexical Phonology

As in Lexical Phonology and Morphology, the Stratal OT levels are morphological as well asphonological subsystems, which form a hierarchy of domains: stems, words, phrases. A constraintsystem of level n+1 may differ in ranking from a constraint system of level n by promotion ofconstraints to undominated status. Each is governed by a (parallel) constraint system, but theyinterface serially. The interaction of constraints is determined by the intrinsic relation of the levels.A constraint at level n is visible to a constraint at level m iff n m. Opacity reduces to con-straint masking, and cyclic effects reduce to ordinary faithfulness: bigger constructions inheritthe phonological properties from the smaller constructions they contain, in so far as compatiblewith the applicable constraints.

Postlexical processes may be restricted to certain prosodic domains, of which the smallest isthe CLITIC GROUP, and the larger ones are the PROSODIC PHRASE, the INTONATION GROUP, and

14

perhaps others (Inkelas & Zec 1990). Lexical processes apply to stems (level 1) and prosodicwords (level 2).

If we reconstruct quasi-phonemes in Stratal OT as lexically specified but distributionally pre-dictable phonological segment types, we get an interesting additional prediction. In Stratal OT,lexical representations are specified by the word-level constraint system. This entails that quasi-phonemes are elements whose distribution is governed by or relevant to at least one lexical con-straint, therefore within the domain of a prosodic word. The same elements may of course alsofigure in postlexical constraints.6

That leads directly to a solution for the phonologization problem. Processes become phonolo-gized when they become applicable to the lexical phonology formally, when the constraints thatdrive them are promoted over the antagonistic faithfulness constraints in the lexical constraint sys-tem. At that point their outputs become quasi-phonemes, understood as lexical allophones. Theeffect of this promotion is that they assign categorical feature values, that their distribution is deter-mined by constraint that operate within the word domain, and that in virtue of these very facts theyare perceptually salient in the sense stated above. Other than the fact that real phonemes have anat least partly unpredictable distribution, there is no basic difference between quasi-phonemes andordinary phonemes, on this view.

The promotion of constraint rankings from the postlexical phonology into the lexical phonol-ogy does not mean that those rankings necessarily cease to apply postlexically. The process is, infact, the generalization of new constraint rankings from the postlexical phonology, where they arefirst introduced as sound changes, into the lexical (word-level and unltimately stem-level) phonol-ogy. The cause of this spread of constraint rankings, I conjecture, is a preference of learners forassigning structure as early as possible. That is, there is a bias in acquisition in favor of locatinginformation in the lexicon.

Although the phonologization of a process in this sense is compatible with its continued postlex-ical operation, the next step is typically disappearance of its postlexical reflexes formally, by thepromotion of antagonistic faithfulness constraints in the postlexical phonology. Once this happens,there is unambiguous evidence for the phonologization, in that the process ceases to apply acrossword boundaries, its output is strictly categorical, and it is perceptually salient.

In the final act of this phonologization scenario, the potential contrasting quasi-phonemes be-comes overtly manifested. This can happen either when a sound change (the promotion of a con-straint in the postlexical phonology) renders their conditioning environment opaque (this is so-called secondary split), or when new lexical entries from borrowing or other sources exploit them.On this understanding, the rise of phonological contrasts is analogous to the rise of phonologicalopacity by constraint masking.

Returning to umlaut, we can now offer an analysis of the phonemicization of front roundedvowels. As a sound change, umlaut is the acquisition of the constraint ranking (33) in the postlex-ical phonology. The vowels , (and , if that is the output of umlaut at this point) are in comple-mentary distribution with u, o, a.

6For example, in Russian [i] and [1] play a role in the lexical phonology, but [i] becomes [1] after a velar consonantacross a word boundary within a clitic group or phonological phrase. See Rubach 2000, Blumenfeld 2001, Padgett2003 for discussion of this interesting case.

15

(32)The sound change: postlexical umlaut

Input Output AGREE(FRONT) IDENT(Hi) *, * IDENT(Back)uCi uCi *

Ci * *uCe *Ce * * *

uCe uCi * *Ci * * *

uCeCe * *

Ce uCi * * *Ci * *

uCe *Ce *

In the second phase of the change, the ranking (33) enters the word phonology. At that point, theumlaut vowels become quasi-phonemes, present in lexical representations and constituting inputsto the postlexical phonology. Since lexical umlaut at first applies in a subset of the contexts inwhich postlexical umlaut applies, this is initially a covert change. It becomes overtly detectable atthe latest in the next phase, when back vowels are restored before clitics with -i- in configurationslike (35), while umlaut continues to apply within the phonological word. Formally, this means thatIDENT(Back) is promoted in the postlexical phonology but remains dominated by umlaut in thelexical phonology. The umlaut vowels are not yet overtly contrastive.

In the third phase, another sound change affects the umlaut-triggering i, j in such a way as tocauses the conditioning of umlaut to become opaque. Let us continue to assume that this happensthrough the promotion of REDUCE in the postlexical phonology. Lexical umlaut vowels are unaf-fected, both phonetically and phonologically. The change in the postlexical phonology that masksthe context of umlaut does, however, cause them change from covertly contrastive to overtly con-trastive elements at this point. In principle, they might also become overtly contrastive through theacquisition of any lexical item with an umlaut vowel in a non-umlauting context, whether throughborrowing, onomatopoeia, or word-formation, along the lines of the Russian example cited above).

In Old High German, this final phase of the change is reached when postlexical vowel reduction(by promoted REDUCE), applying to the output of (33), produces contrasts between uCe and Ce:

16

(33)Overt phonologization: postlexical Vowel Reduction

Input Output REDUCE IDENT(Back) UMLAUT IDENT(Hi) *, *Ci uCi * * *

Ci * *uCe * *

Ce * *uCe uCi * * *

Ci * * * * uCe

Ce * *Ce uCi * * * *

Ci * * *uCe *

Ce *

Although the postlexical promotion of REDUCE renders the conditioning of umlaut opaque, thelexical umlaut vowels themselves are retained. They just become overtly contrastive elements atthis point.

Whereas Saussures Firewall prises apart sound change and phonology and fences them offinto separate worlds assigned to distinct fields of inquiry, this alternative explains phonologizationthrough the internal stratification of phonology into a lexical and a postlexical component. And thatstratal organization is independently motivated by rich evidence, including cyclic (paradigmatic)effects and phonological opacity. In fact, secondary split is just the historical counterpart of opacity,and Stratal OT provides the same solution to both.

This theory predicts that any phonologization process will proceed in three overt stages. Allof them can be documented for umlaut in Old High German. In the earliest stage, after the soundchange enters the language, umlaut was postlexical, and hence crossed lexical word boundaries,applying within clitic groups.

(34) a. /mag iz/ meg iz may itb. /drank ih/ drenc ih I drankc. /gab ima/ geb ima gave himd. /girah inan/ gireh inan avenged them

In early OHG, umlaut became a lexical process, and ceased to apply across word boundaries, butwas still transparently conditioned within the lexical word. The umlaut vowels were now quasi-phonemes. In the third stage, they became overtly contrastive as a result of sound changes thatrendered their conditioning environments opaque.

The theory also predicts that our three criteria for quasi-phonemes should be satisfied at thesecond stage. As far as it is possible to tell, this is the case. The first criterion is certainly satisfied,for umlaut at that stage became restricted to applying inside lexical words. The second criterion

17

is also satisfied: umlaut vowels must have been more salient exponents of vowel frontness thantheir triggers, at least in the normal cases where the umlaut vowels are stressed and the context isunstressed. The third, categoriality, is hardest to verify. The vowels , began to be written onlylate, because the Latin alphabet had no letters for them, but the umlaut of a was written e alreadyat the second stage, that is, well before the reduction of -i to -e that (on the structuralist view)caused it to become phonemic. This could be taken as an indication that they were perceived ascategorically distinct from a at stage 2, i.e. prior to the point at which the structuralist theory ofphonologization claims that they became phonemic.

Crucially, Stratal OT departs from Lexical Phonology by giving up structure-preservation(Strata, yes, structure-preservation, no, as the slogan of Roca 2005 has it). To put it anotherway, Stratal OT severs the structuralist link between CONTRASTIVENESS (unpredictable distribu-tion), a structural notion, and DISTINCTIVENESS, a perceptual notion.7 Phonemes are contrastiveand distinctive, allophones are non-contrastive and non-distinctive. The other two combinationsare the surprising ones. Quasi-phonemes are non-contrastive but distinctive that is, they are pre-dictable but perceptually salient. The fourth logically possible case, contrastive but nondistinctiveelements, exists as well. These are NEAR-MERGERS (Labov 1994, Ch. 12), as when a speakerreliably produces near-merged sounds slightly differently, but cannot distinguish between them, inthe speech of other such speakers or in her own speech, e.g. source and sauce in New York. Thefour cases are shown in (36).

(35)contrastive non-contrastive

distinctive phonemes quasi-phonemesnon-distinctive near-mergers allophones

The upshot is that while delinking contrastiveness and distinctiveness in a sense preserves thephoneme as a theoretical construct, it does so only by negating the founding intuition behind it.

4.2 The embedding problem the Stratal OT solutionStratal OT also offers a solution to the empirical problems for Saussures Firewall that we iden-

tified above. It predicts that sound changes will relate transparently to other postlexical processes.This has the three consequences that we cited above as difficulties for Saussures Firewall.

First, when conditioned allophones are created in the postlexical constraint system, they willjust disappear when their conditioning environments are lost, and no secondary split will occur.In other words, sound changes can bleed existing postlexical processes. That is, they can elim-inate some of their former inputs. English velar to palatal assimilation is postlexical, since it isdetermined by the context across word boundaries (e.g. sock it vs. sock us). Stratal OT predictsthat under these circumstances it cannot become phonemic by secondary split. Therefore, vowelfronting and backing sound changes do not result in a contrast between front and back k. Whilequasi-phonemes survive the loss of their conditioning environment, postlexical allophones disap-pear.

7This link was axiomatic at least in post-Bloomfieldian Americal structuralism. Bloomfield himself allowed dis-tinctive sounds to be non-contrastive, for example if they were morphologically predictable, a practice later condemnedas mixing levels. The Prague school distinction between phonetic and allophonic processes might also be seen asimplying the separation of distinctiveness from contrastiveness.

18

The second consequence is that a sound change can feed other existing postlexical processes,i.e. add new inputs to them. Consider a language that has obligatory voicing assimilation of ob-struents within some postlexical domain, such as the phonological phrase or the phonological word(the clitic group). The prediction is that when sound change creates sequences of obstruents in sucha language, voicing assimilation will automatically eliminate them, as in the previously mentionedOld English example bidest (> *bidst) > bitst. The parenthesized intermediate form is a virtualstage which is not pronounced but forms part of the sound change itself.

Third, sound changes can be blocked just in case their output does not conform to a constraintthat holds at the postlexical level.

(36) Quasi-phonemes can condition the distribution of real phonemesA new kind of secondary split: second-order quasi-phonemes.

Dark and light /l/ and the goat split in British and Australian (Simpson 1979, Wells 1982:312,Borowsky 2001).

(37) a. goat [g2t] (or [g2Ut])b. goal [gOU]

(38) Distribution of [2], [2U] before light [l]:a. [OU] before word-level vocalic endings: goalie, holey, wholly, roller, tolling, control-

lable, controlism, Warholiteb. [2] before stem-level vocalic endings: polar, Mongolian, molar, condolencec. [2] in monomorphemic words: holyd. [2] before -l endings: lowly, slowly

(39) a. Stem level: *l] * 2 OU IDENT(Back)b. Word level: *l] * IDENT(Back) 2 OU

(40) The goat split (some British and Australian dialects)

a. goat [g2t] (or [g2Ut])b. goal [gOU]

(41) Distribution of [2] and [OU] before light [l]:a. Word-level vocalic endings: goalie [gOUli:], holey, wholly, roller, tolling, controllable,

controlism, Warholiteb. Stem-level vocalic endings: polar [p2l], Mongolian, molar, condolencec. Monomorphemic: holy [h2li:]

19

d. Consonantal endings: lowly [l2li:], slowly

(42) a. Stem level: *l] * 2 OU IDENT-I/O(Back)b. Word level: *l] * IDENT-I/O(Back) 2 OU

(43)Stem level *l] * *2 *OU IDENT-I/O(Back)Input: goal /gOUl/ (or /gOU/, /g2/, /g2l/)1a. [gOU] * *1b. [gOUl] * *1c. [g2] * * *1d. [g2l] * *Input: goat /gOUt/ (or /g2t/)2a. [gOUt] *2b. [g2t] *

(44)Word level *l] * IDENT-I/O(Back) *2 *OUInput: [gOU] (from stem level)1a. [gOU] * *1b. [gOUl] * *1c. [g2] * * *1d. [g2l] * *Input: [gOU]i:2a. [gOUi:] * *2b. [gOUli:] *2c. [g2i:] * * *2d. [g2li:] *

(45) The base can be bound

a. Disyllabic, front [l]: [stOU.ln"], [swOU.ln

"] (normal pronunciation)

b. Monosyllabic, back []: [stOUn], [swOUn] (marginal option)

4.3 Theoretical consequences(46) Quasi-phonemes vs. allophones

a. Processes restricted to the stem and word domains are stem and word level processes,respectively, i.e. they belong to the lexical phonology. (Corollary: no structure-preservation.)

b. The output of lexical phonology bears categorical feature specifications.

20

c. Quasi-phonemes are lexical allophones (non-distinctive segments introduced in the lex-ical phonology).

d. In sound change, quasi-phonemes survive the loss of their conditioning environment,while postlexical allophones disappear.

(47) More predictions Intrinsically lexical processes (such as assignment of word accent) must become opaque

if sound change masks their conditioning. Hence, segmental changes cant affect suchprosodic structure directly.

Sound changes cant be conditioned by word boundaries (Hock 1999). Secondary split should be controlled by the postlexical phonology. E.g. deletion of

C in V.CV will result in disyllabic V.V iff postlexical hiatus is OK (i.e. if *ONSET isdominated by faithfulness in the postlexical phonology).

Quasi-phonemes (unlike allophones) should interact with lexical constraints. Features which fit into the system are more easily lexicalized, e.g. tones in systems

which are already tonal..

(48) Structuralist doctrine Allophones: structurally predictable, functionally non-distinctive, Phonemes: structurally unpredictable (contrastive), functionally distinctive.

(49) Our conclusion: two kinds of allophones distinct phonological objects (quasi-phonemes): [ i ] and [i], English [l] and [], bid and

bit, bead and beat one phonological object, different phonetic realizations: front and back versions of /k/,

e.g. look in, look out

(50) Two kinds of allomorphs distinct morphological objects: go and went one morphological object, different phonological realizations: -s, -z, - i z

(51) Quasi-phonemes (lexical allophones) undermine the structuralist program because theyshow that predictable elements can be distinctive.

Structuralism was wrong in categorizing processes on the basis of contrastiveness andconditioning. The important criteria seem to be their domain and formal properties(e.g. locality).

(52) Near-mergers The fourth logically possible case, unpredictable but nondistinctive elements,are NEAR-MERGERS (Labov 1994, Ch. 12). A speaker reliably produces near-mergedsounds slightly differently, but cannot distinguish between them, in the speech of other suchspeakers or in her own speech. E.g. source and sauce in NYC.

21

4.4 The life-cycle of phonological processesIn order to account for secondary split and neogrammarian exceptionlessness, we do not have

to stipulate that the promotion of constraints is limited to the postlexical stratum. Constraints canbe reranked at any stratum. Reranking at the word and stem levels simply amounts to another typeof change, namely analogy (including LEXICAL DIFFUSION, the extension of a lexical rule to newitems, Kiparsky 1995).

This can be illustrated with the more recent development of umlaut in German. It has split intoa stem-level and a word-level process. The word-level process applies to word-based formationsmade with inflectional suffixes and some productive derivational suffixes, illustrated here by thecomparative suffix -er. Historically, it is the result of an analogical streamlining of the synchronicumlaut process. It only triggers vowel fronting, and only in a syllable adjacent to the trigger-ing suffix. Stem-level umlaut, on the other hand, preserves the inherited umlaut process with itshistorically accreted complexities. It generates, in addition to vowel fronting, the synchronicallyarbitrary rounding switch of au to oi (spelled u), as in (54b), it applies non-locally across a syl-labic sonorant, as in (54c,d), and it is often blocked in compounds, as in (54e).

(53) Stem Level Word Level (Word+Suffix)a. arm poor rmlich impoverished rmer poorerb. blau blue blulich bluish blauer bluerc. sauber clean subern to clean sauberer cleanerd. schwanger pregnant beschwngern to impregnate schwangerer more pregnante. nah near nchste nearest hautnahste nearest to the skinf. kalt cold klter colder eiskalter more ice-cold

This unifies what superficially look like two distinct umlaut processes in the synchronic phonology.German Umlaut never crosses a syllable. This locality restriction is common to both the word-level and stem-level versions of the process (e.g. Bubi Bubi-lein, not *Bbi-lein little Bubi).Because r is not syllabified at the stem level (as can be shown on independent grounds), umlaut incases like beschwngern actually does not cross a syllable. Final r after -C becomes syllabic at theword level, blocking umlaut in word-based formationa like like schwang[r

"]-r.

BibliographyBERMDEZ-OTERO, RICARDO 1999. Constraint interaction in language change: quantity in En-

glish and Germanic. Ph.D. Dissertation, University of Manchester.BERMDEZ-OTERO, RICARDO, AND RICHARD M. HOGG. 2003. The actuation problem in Op-

timality Theory: Phonologisation, rule inversion, and rule loss. In D. Eric Holt (ed.), Opti-mality Theory and language change. Dordrecht: Kluwer.

BERMDEZ-OTERO, RICARDO 2006a. Diachronic phonology. In Paul de Lacy (ed.), The Cam-bridge handbook of phonology. Cambridge: CUP.http://myweb.tiscali.co.uk/bermudez/handbook.pdf

BERMDEZ-OTERO, RICARDO 2006b. Phonological change in Optimality Theory. In KeithBrown (ed.), Encyclopedia of language and linguistics, 2nd edn., vol. 9, 497-505. Oxford:Elsevier. http://myweb.tiscali.co.uk/bermudez/encyclopedia.pdf

22

BLEVINS, JULIETTE. 2004. Evolutionary Phonology. Cambridge: CUP.BLUMENFELD, LEV. 2001. Russian palatalization in Stratal OT. MS, Stanford University.BOOIJ, GEERT. 1996. Lexical Phonology and the derivational residue. In Durand, Jacques &

Bernard Laks (eds.) Current Trends in Phonology: Models and Methods. European StudiesResearch Institute and University of Salford.

BOOIJ, GEERT. 1997. Non-derivational phonology meets Lexical Phonology. In Iggy Roca (ed.)Derivations and constraints in the lexicon. Oxford: Clarendon Press.

FLEMMING, EDWARD. 2001. Scalar and categorical phenomena in a unified model of phoneticsand phonology. Phonology 18: 7-44.

HALE, MARK. 2007. Historical Linguistics: Theory and Method.HOENIGSWALD, H. 1965. Language change and linguistic reconstruction. Chicago: University

of Chicago Press.HOGG, RICHARD. 1992. A grammar of Old English. Oxford: Blackwell.INKELAS, SHARON, & DRAGA ZEC. (1990). The Phonology-Syntax Connection. Chicago: Uni-

versity of Chicago Press.IVERSON, GREGORY K., AND JOSEPH C. SALMONS. MS. Naturalness and the life cycle of lan-

guage change.JAKOBSON, ROMAN, GUNNAR FANT, AND MORRIS HALLE. 1952. Preliminaries to speech anal-

ysis. Cambridge: MIT Press.JANDA, RICHARD D. 2003. Phonologization as the start of dephoneticization or, on sound

change and its aftermath: of extension, lexicalization, and morphologization. In Joseph,Brian D. and Richard D, Janda (eds.), The handbook of historical linguistics, p. 401-422.Oxford: Blackwell.

JASANOFF, JAY. 2003. Plus a change. . . Lachmanns Law in Latin. Harvard Working Papers inLinguistics 8.

KING, ROBERT D. 1973. Rule insertion. Language 49: 551-578.KIPARSKY, PAUL. 1984. On the Lexical Phonology of Icelandic, in C.-C. Elert, I. Johnson, and

Eva Stangert (eds.), Nordic Prosody III. Stockholm: Almquist & Wiksell.KIPARSKY, PAUL. 1988. Phonological Change. In F. Newmeyer (ed.), The Cambridge Survey of

Linguistics, Vol. 1. Cambridge: CUP.KIPARSKY, PAUL. 1998. Sievers Law as prosodic optimization. In Jay Jasanoff, H. Craig Melchert,

and Lisi Oliver (edd.), Mr Curad: Studies in honor of Calvert Watkins. Innsbruck: Inns-brucker Beitrge zur Sprachwissenschaft.

KIPARSKY, PAUL. 2000. Opacity and cyclicity. The Linguistic Review 17:351-367, 2000.KIPARSKY, PAUL. 2006. Amphichronic linguistics vs. Evolutionary Phonology. Theoretical Lin-

guistics, 32: 217-236.KIPARSKY, PAUL. To appear. Grammaticalization as optimization. In Diane Jonas (ed.), Proceed-

ings of the 8th Diachronic Generative Syntax Conference.KIPARSKY, VALENTIN. 1932. Johdatusta fonologiaan. Virittj 36: 230-250.

23

LIBERMAN, ANATOLY. 1991. Phonologization in Germanic: Umlaut and vowel shifts. In ElmerH. Antonsen and Hans Hock (eds.) Stfcrft. Studies in Germanic linguistics: Select papersfrom the 1st and 2nd symposium on Germanic linguistics. Amsterdam: Benjamins.

PADGETT, JAYE. 2003. Contrast and post-velar fronting in Russian. Natural Language & Lin-guistic Theory 21: 39-87.

ROCA, IGGY. 2005. Strata, yes, structure-preservation, no. In Geerts, Twan, Ivo van Ginneken,and Haike Jacobs (eds.) Romance languages and linguistic theory 2003. Amsterdam: Ben-jamins.

RUBACH, JERZY. 2000. Backness switch in Russian. Phonology 17, 39-64.SCHULTE, MICHAEL. 1998. Grundfragen der Umlautphonemisierung. Berlin: de Gruyter.TWADDELL, W. FREEMAN. 1938. A note on OHG umlaut. Monatshefte fr deutschen Unterricht

30.177-181.

24