phonology seminar 2007 opacity by yaofenji


									MPhil Seminar: Evaluating OT                                                         25/1/07
Lecture 2: Opacity

Today’s topics
• Worked example illustrating predictive powers of RBP and OT: Turkish VH
• Opacity

1. Comparison of OT and RBP wrt process: Vowel Harmony and Locality
A classic vowel harmony system: Turkish
(1) Turkish vowels
                 [-back]              [+back]
           [-round] [+round] [-round] [+round]
[+high]        i         y         μ          u
[-high]        E         ø          A         ç

(2)   tHAnμ-mA-dμkH-…AR-μm-μz ‘those we do not know’ (know-neg-ppl-pl-1st-pl)
      sEv-mE-dicH-lQR-im-iz ‘those we do not like’ (like-neg-ppl-pl-1st-pl)

Harmonic alternations in Turkish orthography
(3)  nom. sg.        gen. sg.        nom. pl.       gen. pl.      gloss
     ip              ipin            ipler          iplerin       rope
     kız             kızın           kızlar         kızların      girl
     yüz             yüzün           yüzler         yüzlerin      100
     pul             pulun           pullar         pulların      stamp
     el              elin            eller          ellerin       hand
     sap             sapın           saplar         sapların      stem
     köy             köyün           köyler         köylerin      village
     son             sonun           sonlar         sonların      end

The same as they’re actually pronounced
(4)   nom. sg.        gen. sg.      nom. pl.        gen. pl.      gloss
      ipH             i»pHin        ipH»lQR         ipHlQ»Rin     rope
      kHμz            kHμ»zμn       kHμz»…AR        kHμz…A»Rμn    girl
      jyz             jy»zyn        jyz»lQR         jyzlQ»Rin     100
      pHu…            pHu»…un       pHu…»:AR        pHu…:A»Rμn    stamp
      El              E»lin         El»:QR          El:Q»Rin      hand
      sApH            sA»pHμn       sApH»…AR        sApH…A»Rμn    stem
      cHøj            cHø»jyn       cHøj»lQR        cHøjlQ»Rin    village
      son             so»nun        son»…AR         son…A»Rμn     end

(5) NB glides aren’t affected by harmony and are transparent to it (koyun ‘sheep’, not *koyin
    or *koÁün)

(6) Analysis
a. RBP                                                    b. OT
                     UR                                                    UR
                      ⇓                                                     :
            explicit set of rules                                  black box/OOPS/GEN
                      ⇓                                                     :
                     SR                                                     SR

RBP analysis
(7)   tH A n μ m A d μ kH … A R μ m μ z
      | | | | | | | | | | | | | | | |
      X X X X X X X X X X X X X X X X

(8)           X             X       (spread contrastive [back] specifications, iterative L→R)

OT analysis (Baković, etc.)
(9) i. AGREE[color]: all segments in a word must agree in their specifications for [back] and
       ii. *C[+back]: no d[+back], etc.
       iii. *O: no [-high, +round] vowels in non-initial σ
       iv. DEP: All elements in the output must have a correspondent in the input.
       v. MAX: All elements in the input must have a correspondent in the output.

                          *O            *C[+back]       AGREE[color]         DEP           MAX

                                                            ***               *
   [sonon]                 *!                               ***               *
  [sonμn]                                                  ****!              *
   [sunun]                                                  ***               *                 *!
                                          *!*                               ****

Note that nowhere in the evaluation is the process by which the output candidate was
produced examined.

1.1. More on predictions of OT and RBP wrt Vowel Harmony and Locality
(10) Languages that can be derived simply in OT, but not in RBP
a.     Turkish′:     same vowel inventory as Turkish
                     [-back] dorsals don’t participate in harmony
                     vAkHJμtH ‘time’ (form in Turkish: /vAkHJtH/ → [vakHJitH])
       Turkish′′:    same vowel inventory as Turkish
                     no alternations except A ~ e
                     (i) tHAni-mA-dikH-lAR-im-iz
                     (ii)    root: pakitoçü ‘grokker’
                             suffix: -DA ‘locative’
                             ‘in the grokker’: pakitoçüda

(11)         RBP analysis
a.           v A kHJ V tH
             | |   | | |
             X X X X X
                 | ×
               [+bk] [-bk]

a.           p a k i t o ç ü D A
             |  | | | | | | | | |
             X X X X X X X X X X
                |   ×
                    |   ×
                        |   ×
                   [+bk]   [-bk]    [+bk]   [-bk]

(12)         OT analysis
          i. AGREE[color]: all segments in a word must agree with the first vowel in their
               specifications for [back] and [round].
          ii. *C[+back]: no d[+back], etc.
          iii. *O: no [-high, +round] vowels in non-initial σ.
          iv. DEP: All elements in the output must have a correspondent in the input.
          v. MAX: All elements in the input must have a correspondent in the output.

       | |                   AGREE[color]           DEP
       [+b] [-b]
       |    | |
     [+b] [-b] [+b]                 *               **
     [vAkHJitH]                    **!               *
        |   \/
      [+b] [-b]

Problem: if we exposed a child (or adult, for that matter) to either of the languages in (10),
would they be able to learn it? Again, we have the problem of what is attested/plausible vs.
what is computationally possible/learnable.

2. Opacity
Main points:
    Opacity plays a central role in phonological systems.
    OT does not provide an insightful account for opacity.
    RBP provides an insightful account for opacity.
    To the extent that we want a phonological theory to address the fundamental issues of
    phonology, and opacity is perhaps the most fundamental issue in phonology, we must
    prefer RBP over OT.

Key theoretical issues:
   • existence of abstract intermediate stages of computation

1. Introduction to opacity
1.1. History of abstractness and opacity
(13) Opacity (Kiparsky 1973:79)
A phonological rule P of the form A → B / C _ D is opaque if there are surface structures
with any of the following characteristics:
        a.     instances of A in the environment C _ D
        b.     instances of B derived by P that occur in environments other than C _ D

Hooper's 1976 True Generalization Condition (Kager 1999:3)

1.2. Predictions
RBP: rampant opacity
OT: limited opacity of a certain type

3. Types of opacity
(14) 2 types: counterfeeding, counterbleeding
       CF: Opposite ordering of the two processes would be feeding (underapplication)
       CB: Opposite ordering of the two processes would be bleeding (overapplication)
    • Which is (1a), and which is (1b)?
    • NB violable constraints are inherently opaque, but this opacity is of a different sort
       (Prince and Smolensky 1993:203; McCarthy 1997)

3.1. Counterfeeding
counterfeeding 1 and 2 (McCarthy 1997)
(15) Counterfeeding type 1: focus counterfed (no chain shift)
       UR                   ABC
       D→E/A_               doesn't apply
       B→D/_C               ADC

OT can get this type (McCarthy 1997:1):

 (16) OT version: *AD >> Faith(D→E); *BC >> Faith(B→D)
    /ABC/        *AD        Faith(D→E)       *BC       Faith(B→D)               Faith(B→E)
  [ADC] op         *                                         *
[AEC] trans                                                                           *
   [ABC]                                       *

To get opaque [ADC]: Faith(B→E) >> *AD, Faith(B→D)
To get transparent [ADC]: Faith(B→E) dominated by either *AD or Faith(B→D)

(17)   Arabic (Kiparsky 2000)
       a. stress falls on heavy penults
       b. epenthesis before unsyllabified consonants: /fihm/ → [fíhim] 'understanding'
       c. interaction: /fihm-na/ → [fíhimna] 'our understanding'

(18)   Counterfeeding type 2: environment counterfed
       UR                   ABC
       B→D/_E               doesn't apply
       C→E/_#               ABE

(19)   OT version: *BE >> Faith(B→D); *C# >> Faith(C→E)
         /ABC/           *BE       Faith(B→D)        *C#                   Faith(C→E)
        [ABE] op           *                                                     *
       [ADE] trans                       *                                       *
         [ABC]                                         *

Second candidate incurs lesser constraint violations than first candidate. Hence there is no
ranking of the as-yet unranked constraints that will yield the first candidate as the output. OT
allows only transparent interaction in such cases.

(20) Isthmus Nahuat (Kager 1999:374ff.)
a. Apocope               delete unstressed word-final vowel
                         »tami ~ »tam 'it ends'
b. Approximant devoicing devoice {l w j} word-finally
                         »tajo˘l 8 'shelled corn'
c. Interaction           SikA»kIli ~ SikA»kIl     'put it in it'

(21)   the problem in classical OT
        /SikA»kIli/         FINAL-C        MAX-IO      *VOICEDCODA       IDENT-IO(VOICE)
         SikA»kIli              *!
         SikA»kIl                              *             *!
           SikA»kIl•                           *                                 *

3.2. Counterbleeding
(22) Counterbleeding
       UR                ABC#
       B→D/_C            ADC#
       C→E/_#            ADE#
(23) OT version: *BC >> Faith(B→D); *C# >> Faith(C→E)
         /ABC/         *BC       Faith(B→D)        *C#                     Faith(C→E)
        [ADE] op                      *!                                         *
       [ABE] trans                                                               *
         [ADC]                        *!            *
         [ABC]          *!                          *

Second candidate incurs subset of the violations of the first candidate. Hence there is no
ranking of the as-yet unranked constraints that will yield the first candidate as the output. OT
allows only transparent interaction in such cases.

(24)   2 relevant types of counterbleeding:
       a. relevant conditions present in UR
       b. relevant conditions only present in intermediate representation

(25)   Uyghur Umlaut and Vowel Harmony (Hahn 1991:51, 85)
       a. Umlaut: raises low vowels to mid in open syllable before i
               /baS-m/ → [beSim] 'my head'
       b. Vowel Harmony
               /ät-DA/ → [ättä] ‘in flesh’
               /at-DA/ → [atta] ‘in name’
       c. Interaction: umlaut counterbleeds VH
               /ät-I-DA/ → [etidä] ‘in its flesh’
               /at-I-DA/ → [etida] ‘in its name’

(26)   Tiberian (Masoretic) Hebrew
       a. epenthesis into final clusters
               /melk/ → [melex] 'king'
               //erts/ → [/erets] 'land'
       b. /-deletion in coda
               /qara// → [qa:ra:] 'he called'
       c. interaction: counterbleeding (Epenthesis >> /-deletion)
               /deS// → [deSe] 'tender grass' (not *[deS])

(27)   Counterbleeding interactions are rampant, e.g.
              epenthesis vs. r-deletion
                      fire [fAj´] (McCarthy thinks the schwa preserves part of the r)
              vowel lengthening and devoicing
              dialectal dremp = dreamed (drEmt → drEmpt → drEmp)

4. OT treatments of opacity
Many attempts to deal with counterbleeding, cyclicity, ordering (see below).
• Classic OT can’t get opacity effects (McCarthy 2002). It also can’t get chain shifts in
    child language acquisition (McCarthy 2002:42-3)
• OT has to use many different strategies to get opacity effects (esp. counterbleeding,
    cyclicity) [NB several authors are forced to invoke two or more of these at once]:
    • Boersma and Hayes 2001:55 try to avoid opacity in Ilokano /t/ → [/], /// → Ø in coda
        by saying that the t-debuccalization isn’t really neutralizing, but instead produces
        something intermediate between t and glottal stop. They then hedge and say that if it
        actually does become glottal stop, they can invoke any number of opacity tricks in

4.1. Parse-Fill approach
Prince and Smolensky 1993

4.1.1. The proposal
    • Structurally encode properties of input in output.
    • Deleted elements present in output, but unparsed.
    • Inserted elements are empty nodes only.
    • Constraints are sensitive to unparsed and empty elements.

(28)   works for Barrow Inupiaq case (McCarthy 1997:4)
       a. palatalization after i: /savig+lu/ → [savig¥u] ''
       b. absolute neutralization: /ˆ/ → [i]
       c. Interaction: Palatalization applies before neutralization
               /kamˆk+lu/ → [kamiklu] '' (not *[kamik¥u])

4.1.2. Problems
Doesn't work when intermediate level (in serial terms) is crucial (McCarthy 1997:4, Idsardi
1998:59, Kager 1999:379), as in (26)
    In deSe, e is inserted for prosodic reasons, which aren't met if / is unparsed (M 1997)
    Because unpronounced elements are not removed in the phonology, no coherent theory of
    adjacency relations can be supplied (Idsardi 1998:59)
    Also problematic because epenthetic Vs can harmonize (Idsardi 1998:60)

Also requires absolute neutralization, which many/most phonologists prefer to avoid

4.2. Two-level approach
Koskenniemi 1983, McCarthy 1995 (remarks on phonological opacity in OT), Orgun,
Steriade 2003

4.2.1. The proposal
Allow well-formedness constraints to refer to input
tries to deal with e.g. Hebrew spirantization: malake: → malxe:
         k → x if it or its input correspondent is post-vocalic

4.2.2. Problems
    Same as Parse/Fill approach: can't work when intermediate level (in serial terms) is
    Two-level constraints function as rules, combining a structural condition and a repair
    (Kager 1999:381)
    A theory allowing 2-level constraints can stipulate any relation between input and output,
    and is equivalent in this respect to RBP (Kager 1999:381)
    Coupling structural conditions and repairs undermines OT's solution to the conspiracy
    problem (Kager 1999:381)

4.4. Output-Output constraints
McCarthy 1995, Benua 1995
Meghan Sumner (LSA 2001) uses OO correspondence to get the Polish and Slovak facts
discussed by Rubach

4.4.1. The proposal
Idea:          Opacity derives from faithfulness to another form in the paradigm
Application: Hebrew spirantization: /malake:/ → [malxe:]
               k → x from OO faith to related m´la˘xi˘m
Variants: Base Identity, Paradigm Uniformity, No URs

4.4.2. Problems
    Won't work in cases where no form in the paradigm shows the desired phonological
    process (e.g. deSe; McCarthy 1997:5)
    Peperkamp 1997 and Glot article points out that “I argue that this approach is too
    restricted in two ways (cf. Peperkamp 1997). On the one hand, paradigmatic identity
    effects can arise in derived words the morphological base of which is not an occurring
    same thing: attempts to get DE effects via base identity can’t work with e.g. Skt RUKI
    (Kiparsky 1973:63), where it is the base morpheme (which never appears on its own) that
    controls the alternations (cf. also bound morpheme-conditioned phonology in Swiss
    German, Kiparsky 1973:69)

        Kiparsky attack on Kager's 1999 O-O analysis of Arabic:

(29)    Kager's definition of the base relation (Kiparsky 2000:9)
        B is the base to A iff:
                A contains a subset of the gramm feats of B and:
                B is a free-standing output form, ie a word.

doesn't work for Philadelphia tensing
        Kager: tensing of Q to E before tautosyllabic fricatives and front nasals
               as in nom. sg. [pEs] from pQs
               transfers to inflected forms, even if trigger is in onset
                       pl. and genitive [pEs´z]
               problem 1: nom. sg. pass is not Base of the plural or the genitive
               problem 2: by (16b), the stem pass- is not a Base either
               Thus, (16) does not allow transfer of Tensing to inflected form
               Basic problem: OO tries to do m/p interface without a theory of morphology

Kager's O-O account of Arabic:

(30)     a. /fihim/ 'understood'
            i.   fíhim il-wálad    'he understood the boy'
            ii. fíhim              'he understood'
            iii. fhím-na           'we understood' (transparent HV-Del and stress)
            iv. fihím-na           'he understood us' (why no HV-Del?)
       b. /fihm/ 'understanding'
            i.   fíhm il-wálad     'the boy's understanding'
            ii. fíhim              'understanding'
            iii. fíhim-na          'our understanding' (why antepenult stress?)

(31) a. HEADMAX-BA: Every segment in the base prosodic head has a correspondent in the
        affixed form.
     b. NO [I]: /i/ is not allowed in light syllables.
     c. MAX-(IO): Every segment in the input has a correspondent in the output.
     d. a >> b >> c

(18a) accounts for (17aiii) vs. (17aiv) as follows:
       base of fihím-na 'he understood us' is fihím 'he understood us'
       fhím-na has no base: fíhim 'he understood' fails (16a); fíhim- fails (16b)
       (18a) therefore blocks i-deletion in fihím-na but not in fhím-na:

(32)    OT analysis
     Palestinian               HEADMAX-BA                   NO [I]                                     MAX-(IO)
Input: /fihim, -na/, Base: [fíhim]
     1a. fi.hí                                              *
     1b. fhí                    *                                                                            *
Input: /fihim, -na/, Base: none
     2a. fi.hí                                              *                                                *
     2b. fhí                                                                                                 *

Problem: in LPMOT the contrast between subj and obj endings is a matter of stems vs.
words; in OO it's free vs bound forms (Kiparsky 2000:11).
These make different empirical predictions.
OO deals incorrectly with 'cow' in the Tripoli dialect, where post-tonic non-final a deletes:

(33)   a.      bá/ar 'cattle'
       b.      bá/r-a 'a cow' (a-deletion applies)
       c.      bá/ar-i 'my cattle' (a-deletion doesn't apply)

Kager: no base can be found for bá/r-a 'a cow'
       Wrong: the base is bá/ar, which is a collective, not a plural
       Kiparsky: -a is stem level, -i is word-level

Kager's Align R analysis doesn't work:
(34) Kager's analysis of the Arabic facts (Kiparsky 2000:16)


                                                                                                  NO [I]


                            Input: /fihim l-walad/, Base: [fí.him, wá.lad, il.wá.lad]
                            a. *[fí.him.][li.wá.lad]                              *                                         *       **
                            b. correct [fí.hi.m][il.wá.lad]     *                 *                                                 **
                            c. *[fíhm.][li.wá.lad]                    *           *                                        **        *
                            d. *[fíh.m][il.wá.lad]              *                                                           *       **

The correct form (34b) is thrown out by Align-R, but demoting Align-R would result in
massive misgeneration elsewhere, and would not even help in this case, predicting (21d).

4.8. Sympathy
basic problem so far: opacity that refers to intermediate stages
classic case: deSe (though there are many more)

4.8.1. Classic Sympathy
Sympathy treatment of the same facts (Kiparsky 2000:4):
    sympathy evades the invisibility to stress problem
    sympathy constraints require faith to a sympathy candidate (marked by κ)
    sympathetic candidate is optimal candidate that also obeys a designated faith constraint,
    the selector constraint (marked by Θ)

Sympathy constraint: ☼IDENT-STRESS (requires faith to stress of optimal candidate that has
no epenthesis)
Selector constraint: ♣DEP-(I/O)-V
*COMPLEX bans tautosyllabic consonant clusters
The ranking in (22) has the same effect as ordering stress before epenthesis

(35)    Sympathy treatment of Arabic (McCarthy 1997, 1999, Kager 1999)
     Opaque stress            *COMPLEX           ☼IDENT-STRESS                ♣DEP-(I/O)-V
Input: fihm-na
1a.    fhímna                                           *                           *
1b. fíhimna                                                                         *
1c. ☼ fíhmna                      *
Input: katab-t
2a. kátabit                                             *                           *
2b. katábit                                                                         *
2c. ☼ katábt                      *

The optimal candidates that satisfy ♣DEP-(I/O)-V are (1c) and (2c).
The constraints in (22) then select the optimal candidate that has the same stress as these
candidates (as required by ☼IDENT-STRESS), which gives (1b) and (2b).
Sympathy thereby gives a unified analysis for:
   the fact that epenthetic vowels are themselves unstressed
   the fact that epenthetic vowels are “skipped” by stress assignment

4.8.2. Problems
Sympathy constraint has to be invisible to selection of flower candidate (Kager 1999:391)
Sympathy creates chaos in systems with multiple opacities (Idsardi, Kiparsky)
Sympathy constraints otherwise unmotivated (Kiparsky 2000:1)
OT still has to employ conspiracies
• stress shift in Russian
• Hebrew epenthesis and spirantization
• r-deletion
Halle and Idsardi 1997:

   If the schwa in fire → [faj´] is a vocalization of the underlying r, then we predict that
   spar should surface as *[spar´], with vocalization of the r followed by gliding spawned
   from the a.
   In order to avoid opacity problems McCarthy is forced to attribute the r-deletion in spar
   and fire to two different causes: a MaxIO violation in spar and an Ident(r→Ø) violation
   in fire.
   In RBP on the other hand it is easy to account for both deletions in a unified way; both
   simply result from a rule deleting r in syllable coda, without any unmotivated stipulations
   required to avoid theory-internal problems.

Sympathy misses the fact that epenthesis (like all postlexical processes) is invisible to all
word phonology (Kiparsky 2000:5); cf. (23):
The data in (23) [closed syllables are shortened even though postlexical epenthesis opens
them] require a separate sympathy constraint (referring to the same selector constraint) to
“borrow” the opaque shortening in /»Saaf-t/ -> [»Sifit] ‘I saw’ from the failed candidate *»Sift.

(36)   a.      /»Saaf-at/     [»Saafat]       ‘she saw’      (transparent retention of length)
       b.      /»Saaf-t/      [»Sifit]        ‘I saw’        (opaque shortening)

Opaque shortening              *COMPLEX               ☼IDENT-STRESS              ♣DEP-(I/O)-V
Input: »Saaf-t
a. »Saafit                                                    *                         *
b. »Sifit                                                                               *
c. ☼ »Sift                          *

"Once we look at entire phonological systems, not just toy examples of a few interacting
constraints, sympathy results in very serious loss of generalization" (Kiparsky) [cf. also
    • sympathy predicts non-occurring types of constraint interactions, eg mutual non-
    • can't characterize certain real types of constraint interactions [exx?]
    • incompatible with Richness of the Base [cf also Ito and Mester]
    • can't distinguish lexical and postlexical epenthetic vowels (Kiparsky 2000:6)
    • Kiparsky 14-LPMOT and RBP get transitivity of opacity, but Sympathy doesn't:
    • if A is opaque wrt B and B wrt C then A is opaque wrt C [ex. needed!]

4.9. Multistratal OT
"Otherwise unneeded powerful new types of faithfulness constraints, such as O/O, paradigm
uniformity, and sympathy constraints ...have turned out to compromise the OT program very
severely" (Kiparsky 2000:1)
**OT requires at least two now constraint types devised expressly to circumvent the
consequences of straightforward parallel OT, in order to account for opacity and
paradigmatic effects

4.9.2. Levels in OT
Kiparsky’s LPM-OT; Rubach in LI
Kiparsky's reanalysis of the Arabic facts:
    3 phonological levels: Stem, Word, Phrase
    each can have a different constraint ranking
    morphology and phonology are interleaved: affixes added in different strata
"Unlike ordering theories and Sympathy theory, LPMOT relates morphology to phonology in
such a way that level differences motivated by phonological opacity predict morphological
consequences (eg affix ordering) and vice versa"

4.9.3. Problems
    permits implausible ranking inconsistencies between strata (Benua, Kager 1999:385)
    McCarthy 1997:11-allows Duke of York derivations
            McCarthy argues in Duke of York paper that counterbleeding Duke of York cases
            don't exist, but there seem to be many including Karaim CH and Korean
    traditional arguments against interleaving phonology and morphology
        affix ordering
        Halle on Russian
        Why phonological strata shouldn't include affixation (Halle ms.)
        evidence for more than three levels (e.g. Idsardi on Hebrew)

4.11. no URs
Burzio, Hualde, Flemming 1995, Hayes 1998
   • Chomsky 1966:18-19: “the distinction between deep and surface structure emerges
       from even the most superficial examination of real linguistic material”
   • Chomsky 1966:17--surface ambiguity (eg neutralization) motivates URs (cf. Reiss

4.12. More recent OT patches
   • Comparative markedness (McCarthy 2002; problems from pp. 32-36)
          o predicts that all related processes should stand in the same counterfeeding
              relationship with the process they interact with, whereas Stratal OT doesn’t
              predict this (the latter is probably right)
          o sympathy, classic OT, and CM can’t deal with opacity of allophonic
              processes, eg nasal harmony in Sea Dayak, because of what ROTB requires
          o CM allows circular chain shifts (and violations of harmonic ascent in general)
          o CM and classic OT can’t get voice inversion in Luo (Alderete 1998)
   • Candidate chains (McCarthy 2006): concedes ordering

5. Summary and conclusions
Most OT theories of opacity have problems with the deşe type
Sympathy and LPMOT can deal with this class, but each has problems of its own
RBP treatment of opacity is significantly more "elegant":
       predicts exactly the attested types
       deals with them straightforwardly and in a unified way
       OT analyses lose generalizations, create other problems
Opacity is a fundamental issue in phonology; we must therefore prefer RBP


To top