MPhil Seminar: Evaluating OT

1
MPhil Seminar Evaluating OT

• L1/L2 Acquisition and learnability
• 8-3-2007

2
Overview

• Focus on comparison of leading theories (RBP
  OT)
• What each predicts to be possible and impossible
• How these predictions compare to the data
• General acquisition effects
• SLA effects

3
Scientific study of acquisition research
questions
theory
Top-down what does each theory predict to be
possible and impossible?
Bottom-up what are the central acquisition
phenomena that any theory must account for?
data
4
Evaluating theories
Is a theory that makes fewer predictions better?
Venn diagram of outputs for grammar 1 and grammar
2

• Grammar/Theory 1
• Lexicon a
• Rules
• a optionally ? b
• Grammar/Theory 2
• Lexicon a
• Rules
• a optionally ? b
• b optionally ? c

a b c
grammar 1
grammar 2
Loci of predictive difference
box range of conceivable outputs for any theory
5
Some basic learning issues

• All existing OT learning models (except some
  flawed nascent work by Tesar) require being
  handed correct URs, SRs, and CON.
• How can an OT grammar be learned, given the large
  number of constraint rankings that can produce
  the limited data set to which a language learner
  is exposed? Does UG provide a default ranking of
  constraints?
• Prince and Smolensky 203 since constraints cant
  be inferred from surface data (because theyre
  violable), they must be part of UG
• Sherrard 1997 and McMahon 2000 in OT 10 (untied)
  constraints produce 36 million possible grammars,
  posing a learning problem, whereas rules simply
  state observable generalisations.
• McMahon and Calabrese the first step in learning
  the grammar is observing alternations/generalizati
  ons. DP stops there, but OT requires an extra
  step of going from there to a constraint ranking.
  Why is this necessary?

6
Typical OT presentation of the learning task

• http//www.ling.rochester.edu/people/cross/acquisi
  tion-of-ot.pdf
• Prince and Smolensky 1993 All constraints are
  universal, occurring in all languages.
• Languages only vary with respect to their
  lexicons (input forms) and the rankings of the
  constraints.
• Implication If constraints are innate, you don't
  have to learn them.
• Constraints that have to be learned anyway
• morpheme constraints
• Leftmost(um) Place the morpheme um- in the
  leftmost position possible. (PS 1993).
• clitic alignment constraints, à la Anderson 1994.
• parameterisable constraints
• RhythmicTypeIamb/Trochee Feet are
  iambic/trochaic. (PS 1993)
• conjoined constraints?
• Any two simple constraints can be conjoined to
  form a complex constraint. Are all possible
  conjunctions of simple constraints universal? Can
  you conjoin a previously conjoined constraint?

7
Hale and Reiss 1998

• Fact child language production differs from
  comprehension
• Children understand khæt but produce ta
• Smolensky 1996 Optimality Theory (OT) can
  generate this
• Hale and Reiss no it cant, and it shouldnt
• (1) The nature of child phonology
• a. The Strong Identity Hypothesis, which holds
  that child phonology is governed by the same
  principles as adult phonology H/R
• b. The view that child phonology is
  fundamentally distinct from adult phonologythat
  it licenses processes unattested in adult
  language, that it depends on a series of
  developmental stages, and so on S
• (2) The nature of the evidence
• a. Deviations from target formsin childrens as
  well as adults grammarsare to be attributed to
  performance effects, including nonlinguistic
  cognitive and motor processing. H/R
• b. Many deviations from target forms are the
  result of child phonology (i.e., the childs
  phonological competence)grammatical effects for
  which the target language provides no evidence.
  S

8
OT learning algorithms

• M F
• GLA
• CDA

9
Smolenskys scheme

• initial state of grammar M F
• production OT grammar selects most harmonic
  output for a given input
• comprehension OT grammar selects most harmonic
  input for an observed output
• correction algorithm when the comprehension
  evaluation is less harmonic (i.e. there are more
  asterisks) than the production evaluation,
  relevant constraints are demoted to make the
  comprehension parse more harmonic than the
  production parse

10
Smolenskys scheme
11
Problems with S identified by H/R

• cant account for surface neutralization, e.g.
  German /rad/, /rat/ ? rat
• Asudeh 20017 OT grammars seem to be
  non-trivially non-reversible, whether they use
  interpretive parsing or not. This is in conflict
  with the reversibility prized by other
  constraint-based architectures and is tantamount
  to having separate grammars for production and
  comprehension, which is conceptually
  undesirable.
• it isnt possible to have a steady-state grammar
  of Smolenskys type, because the correction
  algorithm will always kick in immediately and
  rerank the constraints
• cant account for Captain Hazelwoods drunken
  speech (misarticulation of liquids, final
  devoicing, deaffrication), which in OT requires
  reranking of constraints (see next slide on
  performance)
• S has severe problems with chain shifts and
  opacity effects.
• S cant account for higher accuracy in direct
  imitation
• i. improved performance under concentration
• S predicts simply better pronunciation of s for
  
• ii. parroting
• S would have to require instantaneous ranking of
  Cree constraints!

12
competence vs. performance

• Johnson, Pisoni, and Bernacki (1990) report on
  the intoxicated speech of the captain of the
  Exxon Valdez around the time of the accident at
  Prince William Sound, Alaska. (669)
• Observed effects
• misarticulation of /r/ and /l/
• final devoicing
• Deaffrication
• Performance problem, not different grammar
• Cf. mimicking Cree sentences doesnt mean that
  one has acquired Cree grammar (671)
• attempts at producing adult phen pen collected
  from a 15-month-old child in a 30-minute period
  (Ferguson 1986)
• mã?, v?, dedn, hIn, mbõ, pHIn,
  tHn?tHn?tHn?, bah, dhau?, buã
• Kids who cant speak yet but can accurately
  identify subject and object of sentence,
  distinguish phonemes, etc. (Hirsh-Pasek and
  Golinkoff 1991)

13
Hale and Reisss C-P model
14
Some problems with OT acquisition theories
identified by Menn 2000

• mushy and non-phonemicized early outputs
• rules that are no longer constraint-based
• U-shaped developmental curves
• the need to differentiate input in the sense of
  the adult form the child is trying to
  approximate from input in the sense of
  abstract strings of concatenated morphemes to be
  parsed

15
The Gradual Learning Algorithm

• Boersma and Hayes 2001
• The GLA is cast within a stochastic version of
  OT that places constraints on a numeric scale
  (Boersma 1998).
• Each time the grammar is used to evaluate
  Input-Output pairings, the ranking values are
  perturbed slightly (by evaluation noise) before
  converting the values to a ranking order. This
  produces variation between the rankings of
  constraints whose values are close to one
  another.
• In the version proposed by Boersma and Hayes
  (2001), the values of all constraints preferring
  the winner are raised, and the values of those
  preferring the loser are lowered, all by an equal
  amount termed the plasticity.
• The result is that the model is much closer to a
  standard stochastic grammar than would at first
  appear phonological templates which are
  abundantly instantiated in the training set end
  up being highly favored by the grammar, and those
  which are poorly instantiated end up disfavored.
  Thus, the rankings of constraints closely track
  the frequency values that would be assigned to
  the same constraints in some stochastic
  grammars. (Pierrehumbert 2003)

16
Problems GLA

• non-convergence (Pater 2005)see later slide
• Pierrehumbert 2003225-6
• requires constraints like k?s to get
  morphophonemic rules like velar softening
• in fact, one needs to postulate a universal set
  of constraints pairing each phoneme with each
  other phoneme (i.e. k?s, k?t, k?p)
• the price of folding unnatural
  morphophonological correspondences into the
  phonological grammar is splitting the
  correspondences into unrelated constraint pairs,
  which are ranked separately.
• In the GLA model, all constraints are at the
  same level and all constraint rankings are
  trained on the same data set. Thus, the
  connection drawn above between the granularity of
  constraints and the size of the effective
  training set does not appear to be available.
• Phonetics, in common with many other physical
  processes, provides examples of skewed e.g.
  gamma distributions relating to physical
  nonlinearities and saturationsIn contrast,
  distributions arising from repeated independent
  decisions (as in coin-flipping, or a forced
  choice experiment) tend to be Gaussian. Since the
  assumption of Gaussian distributions is critical
  to the mathematical tractability of GLA, the
  existence of non-Gaussian distributions appears
  to be problematic.
• The GLA model also does not distinguish effects
  relating to type frequency from effects relating
  to surface, or token, frequency.
• It also provides no way to downweight the
  grammatical impact of extremely frequent words,
  as Bybee (2001) and Bailey and Hahn (2001) show
  to be necessary.

17
The Constraint Demotion Algorithm

• Tesar and Smolensky 1993 version, as reported by
  Crosswhite and Keer

18
Problems CDA

• learning of unnatural patterns
• cant deal with variation (Tesar and Smolensky
  1998, Boersma and Hayes 2001)
• see later slide

19
Learning unnatural patterns

• CDA predicts that natural patterns will be easier
  to learn than unnatural ones, because they
  require fewer departures from UG ranking.
• Support from Paters 2005 study?
• On the other hand, if learning involves
  formulating rules and their difficulty is
  computed on the basis of their formal structure,
  then unnatural patterns should be equally
  learnable.
• Supported in studies by Pycha et al. 2003,
  Morrison 2005, and Seidl Buckley (forthcoming)

20
Variation is a problem for constraint demotion
algorithms

• inconsistent data, such as variation in the
  ambient language, causes RCD to choke
• McCarthy 2002204-5
• Cf. Hayes 2000
• Intra-individual variability in L2 production
• Tropf 1987174multiple renditions of German
  nicht not in a one-hour session with a Spanish
  speaker learning German nIS 31, nI 25, nISt 4,
  nIs 4, nIk 4, nEt 2, nIC 2, nI?t 1, nIZ 1
• Similar findings in L1 acq
• Hayes alternative (strictness bands) predicts
  that only adjacent constraints can be involved in
  optionality
• Problem Pierrehumberts hovacity data

21
Non-convergence for GLA in complex credit
scenarios (Pater 2005)

• The GLA will succeed on the learning problem
  presented in the last section if it is run in
  demotion only mode, in which it only decreases
  the values of loser preferring constraints, and
  does not change the value of winner preferring
  constraints. However, as Boersma (1997, 1998)
  shows, this version of the GLA fails to converge
  in cases involving variation.

22
Variation in CDA

• from Pater 2005

23
Second Language Acquisition
24
OT on SLA

• We assume that second language acquisition
  involves creation of a new grammar, using the
  same resources as first language acquisition
  (though other cognitive strategies may be used as
  well). One major difference, however, is that the
  initial state of second language acquisition is
  the final state of first language acquisition
  (Pater and Tessier 2005)

25
Predictions of canonical OT

• M/F-based learning
• No opacity, derived environment, or avoidance
  effects that dont appear in L1 or L2
• wouldnt make sense to spontaneously invoke
  constraint conjunction, Null Parse, sympathy
  constraints, etc.
• Consistency
• Consistent cross-linguistic treatment of a given
  phenomenon, e.g. resolution of theta
• same constraints as characterize natural
  languages, so D/_ will always be dealt with via
  devoicing, etc.
• No levels
• no level-based effects, since there are no levels
• unnatural processes will not be imported into L2,
  because they are morphologically conditioned
  (according to Lombardi, Steriade, etc.)
• Markedness
• IL effects will result from either hidden UG
  rankings in L1, or from intermediate degrees of
  constraint demotion/promotion NOT from reversion
  to UG rankings when already superseded by L1
  rankings
• markedness-based changes will conform to
  universal markedness hierarchy
• neutralizations will be in direction of unmarked
  member of opposition
• Natural/unmarked patterns will be easier to learn

26
Some central SLP phenomena

• Contra M/F-based learning
• Nonderived environment blocking
• Opacity
• Avoidance
• Contra Consistency
• Optionality and variation
• Convention vs. automaticity
• Final devoicing
• Contra No levels
• Level-based interference
• Unnatural interference
• Contra Markedness
• Cases where IL phenomenon ? NL, TL
• Unnatural patterns not harder to learn

27
Nonderived Environment Blocking

• Eckman and Iverson 1995 et seqq.
• Suppression of s-palatalization in Korean acq of
  English
• Suppression of spirantization in Spanish acq of
  English
• Kiparsky and Menn 198747derived environment
  effect in acq of Greek
• Polish devoicing and raising with loanwords
• snop (not snup) but pagoda ? pagut, toga ? tuk
• Standard OT treatment of NDEB constraint
  conjunction (Lubowicz 1999)
• Smolensky only postulate CC as warranted by PLD
• DEC is problem for
• OT claim that grammars only differ in constraint
  ranking
• OTs rejection of generalizations/rulesDEC in
  SLA is clearly a generalization kicking in, not a
  constraint conjunction spontaneously appearing

28
Opacity in SLP

• Counterfeeding chain shift substitution
• Cho and Lee 2001, Idsardi 2002 on opacity in
  Korean acq of English
• sin ? sjin thin ? sin
• same is found in L1 (Dinnsen, OConnor, and
  Gierut 2001)
• opaque substitution contrast maintenance
  ordering, not sympathy, turbidity, targeted
  constraints, etc.
• Smolensky only postulate constraint conjunction
  as warranted by PLD
• Idsardi 2002 this spontaneous chain shiftdoes
  not reflect properties of their original L1
  grammar, the target L2 grammar, or of Universal
  Grammaronly by employing persistent rules can we
  correctly create the conditions for chain-shift
  persistence of constraints and constraint
  rankings into the L2 does not correctly induce
  the chain-shifting behavior
• Counterbleeding repairs
• Weinberger 1987412Mandarin learners of English
  who apply final epenthesis before final C-cluster
  simplification, e.g. ? aen?
• Counterfeeding and counterbleeding in toy L2 acq
• already discussed earlier in course

29
Avoidance

• speakers sometimes avoid complex L2
  configurations even if their L1 has them
• Laufer and Eliasson 37, Jordens 1977, Kellerman
  1977, 1978, 1986
• Celce-Murcia 1977
• child learning English and French simultaneously
  avoided words containing fricatives in one
  language by using the word from the other
  language, e.g. couteau for knife
• Well-documented in L1 phonological acquisition
  and disorders also
• cleft palate speech, lisp
• Standard OT treatment of avoidance Null Parse
• Wrongly predicts phonologically empty output,
  rather than contentful output that crashes
• see Orgun and Sprouse 1999, Nevins and Vaux 2004
  for further discussion and exemplification

30
Ambiguity and animal wug tests
Gallistel, C. 2003. Conditioning from an
information processing perspective. Behavioural
Processes 61.31234 1-13.
31
Consistency

• Lombardi 2003 repair of L2 ?, ð predictable
  from structure of L1 system
• /s, z/ Japanese, France French, German
• /t, d/ Russian, Quebecois, Hungarian, Sinhalese
• Cf. Ritchie 1968, Nemser 1971, Hancin-Bhatt 1994,
  Weinberger 1997)
• Actual facts intra-lingual/individual variation
  in L1 and L2 acq
• multiple L1 substitutes for unfamiliar/difficult
  L2 sound (Hammarberg 1990)
• Polish replacement of ?, ð by s, z t, d or
  f, v (Gussmann 198431)
• Japanese, German, and Turkish speakers vary in
  what they hear ?, ð as (Hancin-Bhatt 1994)
• Calabrese uses f/v, whereas other Italians use
  t/d (Flege, Munro, and MacKay 1996) (cf. Cockney
  vs. New York Archibald 1998102)
• Unschooled French speakers use t (Berger 1951)
  beginners use f, intermediate learners use t
  (Wenk 1979) Quebecois use t, France uses s
  (Archibald 1998102)
• Korean variation between tense s' and tense
  t' (think ssink ttink) (Oh 2002)
• Austrian learners of English vary between f s
  dental s lenisized ? (Wieden 1997232)
• Cf. L1
• English L1 acq ? ? f in stages IV-VII ð ? d
  or v in stages V-VII (Wenk 1979, Grunwell 1982)
• ? ? f s, e.g. Susies think ? sink fink
  (Vihman and Greenlee 1987)
• Conclusion ambiguity resolved by arbitrary
  choice (convention)

32
Dealing with coda voice

• Overview
• IL final devoicing as TETU?
• L1 vs. L2 strategies for dealing with coda
  voice
• Sources and mechanisms of devoicing

33
Devoicing as TETU?

• From Uffmann 2004
• 2 guiding principles
• Initial state L1 ranking
• L2 learners may also assume default M F
• Ranking for lgs that dont allow codas
• Coda/voi, Coda Faith
• Ranking for lgs that allow contrastive voicing in
  codas
• Faith Coda/voi, Coda
• Demotion of Coda below Faith based on TL
  evidence
• Coda/voi Faith Coda
• NB requires ignoring evidence for voiced codas
• To get Hungarian-English phenomenon (IL devoicing
  despite both NL and TL having voice contrast in
  codas), Uffmann proposes L2 learners assume
  default M F until they receive evidence to the
  contrary
• Problem contravenes OT assumption that L2
  learners start with L1 ranking (Pater 2005, etc.)

34
L1 vs. L2 strategies for coda voice

• L1 claimed to only use devoicing
• The too many solutions problem (Lombardi 1995,
  Kager 1999, Steriade 2001, McCarthy 2002)
• Lombardi MaxOns, Lar, MaxVoi
• Steriade P-map
• McCarthy 2002 targeted constraints can get
  deletion repair, and therefore shouldnt be part
  of OT
• Kiparsky 2004 blocking also used (Konni, Meccan
  Arabic)
• L2 epenthesis and deletion attested
• Deletion
• Chinese Anderson 1983, Xu 2004
• Epenthesis
• Brazilian Portuguese Major 1987 Korean Kang
  2003, Iverson and Lee 2004 Vietnamese Nguyen
  and Ingram 2004 Chinese Eckman 1981, Xu 2004
• Also found in L1 acq (Major 2001)

35
More problems with OT analysis of final devoicing

• L1 prediction doesnt fall out nicely from
  constraint inventory requires conspiracy of
  several constraints
• Coda/voi-less system (Lombardi 2000, Beckman
  2004) misses key articulatory motivation by
  avoiding (cf. Smith et al. 2005)
• Lombardis scheme should hold for NC as well, but
  doesnt (Myers 2002)
• Lombardi cant get word-initial neutralization,
  like in English and Lac Simon
• L1 prediction falsified by L2 data
• Not clear why devoicing should be easier to learn
  than contrast
• L2 devoicing may not be emergent UG effect
• English has devoicing (Haggard 1978, Ladefoged
  1982, Pierrehumbert and Talkin 1992, Smith 1997)
• May be articulatory phonetic problem
• L2 speakersmay not have developed adequate
  voicing control abilities (Smith et al. 2005)
• English-Hungarian case requires reversion to UG
  ranking, not hidden rankings

36
Levels

• Only low-level automatic L1 rules can cause
  interference (Linell 1979, Rubach 1980, 1984,
  Broselow 1984, Singh and Ford 1987, Eckman and
  Iverson 1995)
• Morphologically-conditioned processes do not
  cause interference
• Problem 1 incoherent in monostratal model
• Problem 2 interference from processes treated as
  morphologically-conditioned in OT (McCarthy 1997,
  Steriade 2001, Picard 2002, Lombardi 2003)
• r-insertion in RP pronunciation of L2 French,
  German, Spanish (Wells 1982)
• Jétais déjàr ici, ich bin jar fertig,
  vivar España
• n-insertion in Korean (H. Kim 1999)
• lukenñu?selph look at yourself

37
Cases where IL phenomenon ? NL, TL

• Cases where hidden rankings arent involved
• Idsardis 2002 spontaneous chain shift
• Hungarian-English final devoicing (?)
• Japanese antepenultimate mora accentuation (?)
• Hungarian gemination after stressed vowel (?)
• Cases where novel marked configurations are
  produced, violating markedness hierarchy
• Russian c
• Odd neutralizations in L1 acq
• Child with dental click for all coronal
  continuants (Bedore et al. 1994)
• Child with ingressives for all postvocalic stops
  (Gierut and Champion 2000)
• 48 subject, Jason, produces pfw to represent
  nearly all word-initial liquid clusters, as well
  as initial labial fricatives (Edwards 1996153)
• Amahls word-initial engma and voiceless
  sonorants (Smith 19734)
• child with s in onset, x in coda (Gierut 1993)
• L2 Yuchuns production of ü for i in English
  UG, phonology, etc.
• NB she doesnt just do it after j

38
Conclusions

• Classic OT predictions disconfirmed by
  acquisition data
• contra M/F-based learning NDEB, opacity,
  avoidance
• contra Consistency variation, convention
• contra No levels Level-based interference,
  Unnatural interference
• contra Markedness Cases where IL phenomenon ?
  NL, TL Unnatural patterns not harder to learn
• Universal markedness/UG plays a role in SLA
• final devoicing, cluster simplification
• Not well captured in OT SLA requires reference
  to UG system overridden by L1 and to phenomena
  not in L1/L2/UG-CON
• Conversely, language allows a broader range of
  possibilities than is countenanced in OT
• strategies to deal with devoicing, unnatural
  rules
• Acquisition is generalization-driven and
  potentially variable
• supported by DEC effects, conventionalized
  segmental substitutions, which V to delete in
  hiatus, deneutralization
• Cf. Kiparsky and Menns 1977 model, which
  involves active hypothesis formation and testing
  on the part of the child, and Fey and Gandour
  1979, vs. Stampes and OTs, which are closer to
  behaviorist stimulus-response
• For variation, cf. Macken and Fergusons flexible
  learning model that allows for variation,
  contrary to earlier deterministic models that had
  predictable L1?L2 transfer

39
References

• Bedore, L., J. Leonard, and J. Gandour. 1994. The
  substitution of a click for sibilants A case
  study. Clinical Linguistics Phonetics
  8283-293.
• Broselow, Ellen, S.-I. Chen, C. Wang. 1998. The
  emergence of the unmarked in second language
  phonology.  Studies in Second Language
  Acquisition, 20, 261-280.
• Broselow, Ellen. 2004. Unmarked structures and
  emergent rankings in second language phonology.
  In Lleó, C., Vogel, I. (Eds.), On the
  acquisition of second language phonology Special
  issue. International Journal of Bilingualism
  8.151-65.
• Cebrian, J. 2002. Phonetic Similarity,
  Syllabification and Phonotactic Constraints in
  the Acquisition of a Second Language Contrast.
  Doctoral dissertation, University of Toronto.
• Celce-Murcia, M. 1977. Phonological factors in
  vocabulary acquisition a case study of a
  two-year-old English-French bilingual. Working
  Papers in Bilingualism 1327-41.
• Cho, Mi-Hui and Shinsook Lee. 2001. Phonological
  transparency and opacity in the sound
  substitution of interlanguages. Studies in
  Phonetics, Phonology and Morphology. 7.2. 449-468
• Connelly, Mark (1994) Phonological Markedness
  and Second Language Error Interpretation.
  Doctoral dissertation, Harvard University.
• Davidson, Lisa. 2002.  The effects of hidden
  rankings on the acquisition of consonant
  clusters. In A. James J. Leather (Eds.), New
  Sounds 2000 (pp. 87-96).  Klagenfurt, Austria 
  University of Klagenfurt.
• Dell, G. S., Reed, K. D., Adams D. R. and Meyer,
  A.S. 2000. Speech errors, phonotactic
  constraints, and implicit learning A study of
  the role of experience in language production.
  Journal of Experimental Psychology Learning,
  Memory, and Cognition, 26 1355-1367.
• Dinnsen, Daniel, Kathleen OConnor, and Judith
  Gierut. 2001. The puzzle-puddle-pickle problem
  and the Duke-of-York gambit in acquisition.
  Journal of Linguistics 37.
• Dziubalska-Kolaczyk, Katarzyna. 1988. How do
  Poles perform English tips of the slung? In
  Papers and studies in contrastive linguistics
  22179-86.
• Eckman, Fred Iverson, G. K. (1994),
  Pronunciation difficulties in ESL coda
  consonants in English interlanguage In M. Yavas
  (Ed) First and second language phonology.
  (pp.251-265). San Diego Singular Publishing
  Group.
• Eckman, Fred Iverson, G. K.  (1997).  Structure
  preservation in interlanguage phonology.  In S.
  J. Hannahs, M. Scholten (Eds.), Focus on
  phonological acquisition (pp. 183-207). 
  Amsterdam John Benjamins.
• Eckman, Fred, A. Elreyes, Greg Iverson. 2003.
  Some Principles of Second Language Phonology.
  Second Language Research 19169-208.
• Eckman, Fred. (1987b) The reduction of word-final
  consonants in interlanguage. In. A. James L.
  Leather (Eds.) Sound patterns in second language
  acquisition (pp. 143-162). Dordrecht Foris.
• Gierut, Judith and Annette Hust Champion. 2000.
  Ingressive substitutions typical or atypical
  phonological pattern?
• Grijzenhout, Janet and Bertus van Rooy. 200x.
  Second language phonology acquisition through
  gradual constraint demotion.
• Grunwell, P. 1982. Clinical phonology. London
  Croom Helm.
• Gussmann, Edmund. 1984. Contrastive analysis,
  substantive evidence and the abstractness issue.
  In Theoretical issues in contrastive phonology,
  S. Elliason (ed), 27-36. Heidelberg Julius Groos
  Verlag.

40
References

• Kiparsky, Paul. 1977. On the Acquisition of
  Phonology (with L. Menn). In J. Macnamara (ed.),
  Language Learning and Thought, Academic Press,
  1977. Reprinted in Georgette Ioup and S. W.
  Weinberger (eds.), Interlanguage Phonology the
  Acquisition of a Second Language Sound System,
  Cambridge, Mass. Newbury House, 1987.
• Kiparsky, Paul. 2004. Universals constrain
  change, change results in typological
  generalizations.
• Lombardi, Linda. 2001. Why Place and Voice are
  Different Constraint-Specific Alternations in
  Optimality Theory. In Lombardi, Linda, ed. (2001)
  Segmental Phonology in Optimality Theory
  Constraints and Representations. Cambridge
  Cambridge University Press.
• Lombardi, Linda. 2003. Second language data and
  constraints on manner Explaining substitutions
  for the English interdentals. Second Language
  Research 19225-250.
• Lubowicz, Anna. 1999. Derived environment effects
  in OT. The Proceedings of the Seventeenth West
  Coast Conference on Formal Linguistics, ed. by
  Kimary Shahin, Susan Blake and Eun Sook Kim,
  451-65. Stanford, CA Center Study Language
  Information. Rutgers Optimality Archive 239
• Major, Roy. 1987. A model for interlanguage
  phonology. In G. Ioup S. Weinberger.(Eds.)
  Interlanguage Phonology (pp 101 - 124). Rowley,
  MA Newbury House Publishers.
• Morrison, G. S. (2005, May). Phonetic naturalness
  and phonological learnability. Paper presented at
  The 13th Manchester Phonology Meeting (mfm 13),
  Manchester, UK.
• Nguyen, Thu and John Ingram. 2004. A corpus-based
  analysis of transfer effects and connected speech
  processes in Vietnamese English. Proceedings of
  the 10th Australian International Conference on
  Speech Science Technology, Macquarie
  University.
• Pater, Joe and Anne-Michelle Tessier. 2005.
  Phonotactics and alternations Testing the
  connection with artificial language learning. In
  Kathryn Flack and Shigeto Kawahara (eds.),
  University of Massachusetts Occasional Papers in
  Linguistics 311-16.
• Pierrehumbert 2003, in Bod, Rens, Jennifer Hay
  and Stefanie Jannedy, eds. 2003. Probabilistic
  Linguistics. Cambridge, MA MIT Press.
• Pycha, Anne, Pawel Nowak, Eurie Shin, and Ryan
  Shosted. 2003. Phonological Rule-Learning and Its
  Implications for a Theory of Vowel Harmony. WCCFL
  22 Proceedings, ed. M. Tsujimura and G. Garding,
  pp. 101-114. Somerville, MA Cascadilla Press.
• Rubach, J. 1984. Rule typology and phonological
  interference. In Theoretical issues in
  contrastive phonology, S. Elliason (ed), 37-50.
  Heidelberg, Julius Groos Verlag.
• Seidl, A. and E. Buckley. To appear. On the
  learning of arbitrary phonological rules. In
  Language Learning and Development.
• Singh, Rajendra and Alan Ford. Interphonology and
  phonological theory. In Sound patterns in second
  language acquisition, James and Leather, eds.,
  163-172. Dordrecht Foris.
• Smith, Bruce, Darcie DeMille Amy Roberts, Ann
  Bradlow Tessa Bent. 2005. Devoicing in
  word-final voiced stop targets produced by native
  and nonnative speakers of English.
  http//www.sfu.ca/spchlab/A54.pdf
• Smith, Caroline. 1997. The devoicing of /z/ in
  American English effects of local and prosodic
  context. Journal of Phonetics 25.4471-500.
• Stark, J. 1974. Aphasiological evidence for the
  abstract analysis of the German velar nasal.
  Wiener linguistische Gazette 721-37.
• Tarone, E. 1976. Some influences on interlanguage
  phonology. Working Papers in Bilingualism
  887-111.
• Tarone, E.  1988.  Variation in interlanguage. 
  London Edward Arnold.