Phonotactics and knowledge of relative similarity

1
Phonotactics and knowledge of relative similarity

• D.Steriade, MIT

2
Phonotactics?

• System of segmental/prosodic contrasts
• E.g., is there T? D?
• Their contextual distribution
• E.g., is there T? D in /_?

3
The 2 phonotactic questions

• Trigger what factor causes loss of contrast
• E.g. voice? voice/_?
• Process how does contrast loss come about?
• E.g. what happened to (potential) Ds in/_?
• did they devoice, nasalize, turn to glides,
  delete, get a V after, merge with preceding
  segment, allow their voicing to float to better
  positions?

4
1st message today

• Cant understand triggers unless we understand
  processes
• Cant understand processes unless we understand
  similarity relations.
• Cant understand effect of similarity on grammar
  unless we understand the inhibiting effect of
  lexical knowledge

5
2nd message

• Knowledge transfer
• learners transfer knowledge from one domain
  (phonetics, perceptual similarity) to another
  (phonotactic process)

6
Basic issues

• Is there knowledge of grammar?
• Its precise nature?
• Source of universal laws?
• Relevance to study of competence?
• Links between grammar and lexicon?
• Learning

7
Knowledge of phonotactics

•  perception (mis)guided by phonotactic knowledge.
  
• (Pitt 1998, Pitt McQueen 1998 Moreton 2002
• Dupoux et al. 1999)
•  production limited by L1 phonotactics
• (L2 lit Eckman 1978 Broselow et al. 1995 ).

8
Basic issues

• Is there knowledge of grammar?
• Its precise nature?
• Source of universal laws?
• Relevance to study of competence?
• Links between grammar and lexicon?
• Learning

9
Form of phonotactic knowledge

• A result from OT
• Phonotactic systems can be factored into general
  constraints, ready for cross-linguistic
  comparison,
• if the constraints are ranked and violable

10
The Nonfinality example (adapted from Prince
Smolensky 1993)

• Latin no final stress
• Except that monosyllables are stressed.
• Have stress gtgt Nonfinality
• Cairene no final stress
• Except for monosyllables and extraheavy finals
  (CVVC, CVCC)
• Have stress, Stressless extraheavy gtgt
  Nonfinality
• Guptas Hindi no final stress
• Except for monosyllables and the heaviest
  syllable of the word, if final.
• Have stress, Stressless Heavy gtgt Nonfinality

11
Its an ecumenical result

• Open Q if learners factor out their phonotactic
  knowledge into general, ranked and violable
  constraints.
• Established result this factorization yields a
  far clearer view of phonotactic typology than all
  previous ones.
• (cf. Archangeli and Pulleyblank 1987 for attempt
  to characterize typology by breaking down rules
  into elementary operations)

12
Basic issues

• Is there knowledge of grammar?
• Its precise nature?
• Source of universal laws?
• Relevance to study of competence?
• Links between grammar and lexicon?
• Learning

13
Why these laws?

• Right context laws
• If T ? TÓ /_ (? V) then T? TÓ /_ V
• If p ? t ? k /_ (? V) then p ? t ?k /_ V
• Left context law
• If T ? ÓT / (? V)_ then T ? ÓT /V_
• If Ê ? t / (? V)_ then Ê ? t /V_
• (Steriade 1995, 1999)

14
Context affects perceptibility

• Cues are context dependent.
• And (sometimes) asymmetrically distributed
• left context essential in T ? ÓT , Ê ? t
• right context essential in others.
• Scale of optimal perceptibility for some contrast
  implicational scale of licensing positions for
  that contrast
• Crosswhite 1998, Flemming 1995, Hamilton 1994,
  Jun 1995, Kirchner 1999, Kochetov 1999-2002,
  Silverman 1995, Steriade 1994-1999, Zhang 2000,
  
• Result phonotactic laws have identifiable
  sources in speech perception and production.
• general line of thought Ohala 1990, Lindblom
  1990, others

15
Basic issues

• Is there knowledge of grammar?
• Its precise nature?
• Source of universal laws?
• Relevance to study of competence?
• Links between grammar and lexicon?
• Learning

16
Knowledge of the general laws?

• Or just the manifestations to which learners
• are overtly exposed?
• Background
• Jakobson 1941, Prince Smolensky 1993
• Result (and burning issue)
• Preference for unmarked (e.g. mp vs. np) before
• knowledge of language specific phonotactics
• Jusczyk, Smolensky, Allocco 2002

17
Basic issues

• Is there knowledge of grammar?
• Its precise nature?
• Source of universal laws?
• Relevance to study of competence?
• Links between grammar and lexicon?
• Learning

18
Does the phonotactic grammar emerge from the
lexicon?

• Labphon 5, Coleman Pierrehumbert 1998, Frisch,
  Large Pisoni 2000,
• Bailey Hahn 2001
•  Knowledge of lexical patterns not attributable
  to general laws
• Ernestus Baayen 2002, Pierrehumbert 2002
•  Knowledge of phonotactic preferences not
  reflected in
• lexical patterns
• Moreton 2002. Also Shinohara 1997, Fleischhacker
  2000, Davidson 2002,
• Shademan 2002.

19
Learning (Tesar Smolensky 2000, Prince Tesar
1999, Hayes 1999)

• Results 1st learning models that
• (a) extend beyond systems of non-interactive
  parameters (Dresher Kaye 1990)
• (b) do not depend on a fixed learning path
  planted with learning cues (Dresher 1999)
• Models build on the assumption of violability
  and (re)-ranking

20
What I doIntersection of 3 basic issues
 Source of phonotactic knowledge
hidden rankings of correspondence conditions
Davidson 2002
 Knowledge of grammar vs. knowledge of lexicon
lexicon-based vs. hidden constraint
hierarchies
 Nature of phonotactic knowledge
context-sensitive Markedness context-free
Correspondence vs. context-free M
context-sensitive C?
21
Hidden rankings intro

• The phenomenon in general
• loanword adapters ( others phonotactic
  freelancers) converge on solutions to phonotactic
  violation, without prompting from native sound
  system.
• Significance
• Any choice of solution to phonotactic violation
  reveals
• implicit knowledge of a correspondence ranking .

22
Example (based on Cantonese, cf. Silverman 1992,
Phonology cf. also Mandarin, cf. Broselow et al.
1995 SSLA )

• What the lexicon tells the learner
• no word final D/TÓ obstruent tab, tapÓ
• What it doesnt how to fix a deviant input
• delete bad coda? ta (MAX C)
• add V? tabi, tapÓi (DEP V)
• relocate bad feature? dap, tÓap (Linearity)
• remove coda voicing/asp! tap (Ident voice/asp)
• But the learner knows this anyway.
• Hidden ranking
• MAX C, DEP V, Linearity gtgt Ident voice/asp

23
Hidden rankings in cluster resolution part 1

• Language disallows CC onset/CC coda. Coda
  restrictions
• Native system lacks relevant alternations
  learner cant tell the fate of bad syllables
• Dual pattern of preservation
• Strident Cs preserved as such, in all contexts
• Non-stridents lost or modified, depending on
  context

24
Cantonese (Silverman 1992)
  Phonotactics CC onset/coda and fricative
in coda.
 Phonotactic solutions to deviant inputs
Post-V all consonants preserved, some modified
? Strident codas induce epenthesis /bus/
-gt pasi, not pat
? Nonstrident fricative codas become stops
/shaft/ -gt sap, not safi(t),
Ident (strident) gtgt DEP gtgt Ident (cont)
25
Cantonese(cont)
 Non-V adjacent context (? (//V))
? Stridents induce epenthesis /tips/
-gt tÓipsi, not tip /stamp/ -gt sitam, not
tam

• Non-stridents deleted
• /bend/ -gt pen, not penti
• /post/ -gt posi, not posit

 MAX (strident (?(//V)) gtgtDEP gtgt MAX C (?
(//V))
MAX (strident (?(//V)) gtgt Contiguity gtgt MAX C
(?(//V))
26
Cantonese (end)

• Next to vocoid (V, glide or liquid)
• All Cs preserved, phonotactics satisfied via
  epenthesis
• /fluke/ -fuluk, not fuk, luk (contrast /bend/
  -gt pen)
• A further hidden ranking
• MAX (C//(Vocoid)) gtgt DEP, Contig gtgt MAX (C
  (?(//Vocoid))
• fluk-gt fuluk bend -gt pen

27
Similar dual pattern in
Loan adaptation into Hausa Newman 2000,
Dtschang Bird 1999 Seleyarese Broselow
1997 Jahai Burenhult 2001 Sranan Alber Plag
1999 others
28
Hidden rankings part 2 anaptyxis vs. prothesis
 Fleischhacker (2000 UCLA MA, 2003 UCLA
diss) also Broselow (1992) Zuraw 2002
 Dual pattern of CC onset avoidance
? Stop-sonorant clusters anaptyxis Egyptian
Arabic plastic -gt bilastik
? Other CC clusters, esp. s-stop
prothesis Egyptian Arabic study -gt istadi 
? s-stop-sonorant clusters prothesis and
anaptyxis Egyptian Arabic street -gt /istirit
29
The law (degenerate version of Fleischhackers)

• Anaptyxis in s-Stop implies anaptyxis in
  Stop-sonorant
• Only anaptyxis Japanese, Punjabi
• Only prothesis Iraqi (but different pattern in
  sCC)
• Both sites, with anaptyxis limited to
  stop-sonorant
• Egyptian, Amharic, Farsi, Kazakh, Sinhalese,
  Armenian, Wolof,

30
Not just phonotactically motivated V-insertion

• Pierrehumbert Nairs (1995 LgSp) language game
• bNk -gt btNk
• S-stop clusters preserved intact
• skb -gt sktb
• Obstruent-liquid clusters tend to split
• pln -gt ptln
• cf. Fleischhacker 2001 for discussion and Zuraw
  2002 on parallel pattern in Tagalog

31
Relativized contiguity

• General solution
• Contiguity s-stop gtgt Contiguity
  stop-sonorant
• Anaptyxis ATB (svTV, TvRV)
• C/_V gtgt Contig. s-stop gtgt Contig.stop-sonorant
  Prothesis ATB (vsTV, vTRV)
• gtgt Contig. stop-sonorant gtgt C/_V
• Anaptyxis in TvRV, prothesis in vsTV
• Contig. s-stop gtgt C/_V gtgt Contig.
  stop-sonorant

32
Source of the hidden rankings?

• Relative similarity judgments D (x-y) lt D (z-w)
• Choice of final devoicing (over
  C-delete,epenthesis)
• ????D(T-D/_) lt D(C-Ø), D(V-Ø), Steriade 2002,
  below
• Choice of anaptyxis over prothesis in stop-son.
• D (TR-TvR) lt D (TR-vTR) ) Fleischhacker 2000
• Choice of prothesis over anaptyxis in s-stop
• ????D (sT-vsT) lt D (sT-svT) Fleischhacker 2000
• Choice of C-preservation by context
• D(C-Ø(? (//V)) lt D(C-Ø(//V))

33
P-map

• Set of relative perceptual similarity judgments.
• Rooted in phonetic knowledge (Kingston Diehl
  Lg 1994)
•  Similarity rankings provide a tool for
  inferring
• (a) the form of correspondence constraints
• (b) their rankings
• E.g, if learner knows ?(b-d) gt ?(m-n), he infers
  that
• (a) Ident place/ oral C ? Ident place in nasal C
  
• (b) Ident place/ oral C gtgt Ident place in nasal
  C
• And conversely, if he believes ?(b-d) ?(m-n),
  then he is free to posit a single constraint
  Ident place or 2 constraints but fail to rank
  them
• Wilson 2000 alternative way of building
  similarity relations into phonology

34
Expectations of universality?

• Some similarity rankings should be constant
  across languages
•  if based on inherent asymmetries in cue
  distribution between contexts e.g. C//V vs.
  C/(? (//V))
• No reason to expect ATB universality
• (a) VNT vs. VND in Romanian (gradient
  post-nasal voicing)
•  vs. VNT V)T vs. VND VND in English
• (b) stress diff. in Spanish vs. French
  (Dupoux et al. 1999)

35
The real expectation

• Judgments of relative similarity should correlate
  with choices of phonotactic repair.
• And, if the similarity judgment is
  cross-linguistically constant, then choice of
  repair strategy should be too.

36
Sources of similarity data

• Overt judgments
• (Mohr Wang 1968 Singh 1970 Magen 1998
  Fleischhacker 2000)
• Confusion --in noise, in quiet (Miller Nicely
  1956)
• Speeded discrimination tasks (Seo 2001)
• Similarity judgments implicit in choice of
•  half-rhymes (time-nine Zwicky 1976 CLS,
  Steriade Zhang 2001)
•  imperfect puns (shrubs to gardener
  eucalyptus! Zwicky archive Fleischhacker
  to appear)

37
Evidence for choice of repair?

• Phonotactic systems lexically manifest
  alternations
• Phonotactic free-lancing (on-line) adaptation
• Correlate these choices with similarity
  rankings
• Results partially diverge
• greater uniformity of choice in free-lancing.
• better fit with similarity ranking in
  free-lancing.
• Phonology is unnatural. (Anderson LI 1989)
• Lexically entrenched phonotactic systems are
  unnatural.

38
NC 9 no nasal voiceless (Pater 1995)
 All but (g) are attested in phonotactic systems.
 Only (a) is robustly attested in free-lancing.
39
Satisfying NC 9 in Bantu OshiKwanyama
Steinbergs 1983 SAL

• (a) Native lexicon
•  Roots Nasal followed by voiced C only kombo
  goat , no kompo
•  PrefixRoot   Merger
• oku-pota be rude, on-pote -gt omote good
  for nothing
• (b) Loans
•  Roots Postnasal voicing
• stamp -gt sitamba, print -gt pelenda, ink -gt
  o-iNga
•  PrefixRoot   Postnasal voicing
• papier (Afrikaans) -gt om-bapila, kErk
  (Afr.) -gt oN-geleka

40
Similar split in

• Lumasaaba (Brown 1968)
•  NC induces C deletion in old alternations
•  Young speakers substitute Post N voicing
  alternations, via dialect borrowing.
• Cephalonian Greek loans from Romance
• Mazateco loans from Spanish

41
Verify cross-process rankings

• If correspondence rankings derive from P-Map,
  they should be same across distinct phonological
  processes, in all languages (if same similarity
  rankings obtain).
• Ranking for final devoicing
• MAX C gtgt Ident voice
• This ranking contradicted by some NC systems
• (e.g. C-deletion Ident voice gtgt MAX C)
• But confirmed by all free-lance solutions to NC
• MAX C gtgt Ident voice

42
One can infer, then

• Free-lance solutions to NC are based on fixed
• correspondence rankings with a constant source.
  
• But systems of alternation are affected by
  additional forces.
• Telescoped series of sound changes?

43
Historical source of NC 9 induced merger

• Lynch (1975, OcLx) reconstruction of
  Proto-Oceanic
• I Prefix (na-ka)
• II V-loss (nka)
• III Assimilation (Nka)
• IV Post NasalVoicing (Nga)
• V Extension of nasal phase (Na)
• End result ka -Na rather than ka - ga
  alternations
• Perhaps same scenario in Bantu etc.

44
Burning question to Kie Zuraw

• What interaction of grammar and lexicon can
  generate dissimilar alternants through successive
  sound changes?

45
Another burning question

• How does learner reconcile the conflicting
  correspondence hierarchies?
• Established lexical stock NC merger
• Ident voice gtgt Uniformity
• Loans Post-N voicing
• Uniformity gtgt Ident voice

46
Narrow lexical override

• Technically constraint indexing (Fukazawa
  1998ROA)
• Ident voice (I-native O (list))gtgt Uniformity
  gtgt Ident voice (I-O)
• Lexical evidence is narrowly construed
• as bearing only on the the phonology of lexical
  classes where the evidence originates
• Similarity evidence is broadly construed
• as bearing potentially on all phonological
  patterns

47
Half Rhymes (HR) as evidence for similarity
ranking

• Fact some HRs are more frequent than others
• (time-nine vs. fab-glad
• raised-days vs. raised-raids)
• H1 Frequent HRs are closer to identity
• H2 similarity judgments determining HR choice
  those determining choice of repair strategy
• (H3 HR choice is governed by a linguistic system
  of ranked correspondence constraints.)

48
Initial questions

• Does the incidence of feature mismatch in the
  rhyme domain (RD) depend on context?
• E.g. Cab-Cap vs. Cába-Cápa
• Does it increase in contexts of reduced
  perceptibility?

49
Romanian HR corpus

•  A translation corpus 6 rhymed translation
  texts, mostly from Russian, 1956-1971.
• Totals 693 SRs/ 9791 rhyming pairs. SR
  frequency from a high of 18 to a low of .006
•  A poetry corpus 2 native poets, 1950-1961.
• Totals (for 2) 167 SR/6050 rhyming pairs. SR
  frequencies 0.58 and 10 respectively
•  A poets private rhyming dictionary
• Mihai Eminescu (cca 1880) Dict7ionar de rime

50
Uniformity of preference?

• Do poets (if contemporary, same dialect) share a
  hierarchy of HR preferences?
• Yes certain HR types occur in all texts.
• NT-ND (skimb-timp)
• uNC-ÈNC (skund-rÈnd)
• Sparse HR users concentrate on shared core set.
• Liberal HR users augment it with additional
  types.
• Relative frequency of HR types in any given text
  mirrors, in general, their position on shared
  hierarchy
• TV-DV implies NT-ND in any text
• TV-DV less frequent than NT-ND

51
Voicing HR by context in V.Teodorescus
translation of Maiakovsky
of total rhymes whose RDs contain obstruents N
71
52
HRs and phonotactic processes

• NTV-NDV HRs (únde-múnte) postnasal voicing?
• gradient reduction in the duration voiceless
  interval after nasals Steriade Zhang 2001
• T-D (nas-ekstáz) final devoicing?
• variable realization of final T-D contrast
  Steriade Zhang 2001
• NT-ND (unt-fund) combination of post-N voicing
  and final devoicing??
• no two correlates of the contrast (voicing
  during closure/release quality) are reduced in
  N_

53
Correspondence, not markedness, is relevant to HR
selection

• NC cant explain why NT-ND HRs are more
  frequent than T-D HRs.
• ?D (T-D/(?N)_) gt D (T-D/N_)
• Ident voice /(?N)_ gtgt Ident voice /N_
• If this ranking operates in phonology, NC is
  not needed avoice is enough
• D cant explain why T-D HRs are more frequent
  than TV-DV HRs.
• D (T-D/_V) gt D T-D/_(?V))
• Under appropriate correspondence ranking, D is
  not needed avoice suffices here too.

54
Other recurrent processes in common HR types
Process HR example As in
But not Romanian
55
Summary of proposal

• Knowledge of cue distribution projects a
  hierarchy of context-sensitive similarity
  relations.
• Knowledge of similarity projects correspondence
  constraints and their rankings.
• Knowledge of correspondence identifies
•  optimal repair strategy
• (subject to possible lexical override)
•  tolerably dissimilar pairs in rhyming
•  tolerably dissimilar pairs in paradigm
  structure
• (Steriade 1999, Zhang 2000, Garrett 2003)

56
1st message today

• Cant understand triggers unless we understand
  processes
• Cant understand processes unless we understand
  similarity relations.
• Cant understand effect of similarity on grammar
  unless we understand the inhibiting effect of
  lexical knowledge

57
2nd message

• Knowledge transfer
• learners transfer knowledge from one domain
  (phonetics, perceptual similarity) to another
  (phonotactic process)