Li6 Phonology and Morphology

1
Li6 Phonology and Morphology

• Rules

2
Lecture plan

• Key point of tension between symbolic
  rationalists and numerical reductionists
• Do humans extract generalisations from the data
  in their perceptual worlds?
• Put differently, is the mind a Turing Machine or
  a recurrent switching network?
• Evidence for rules
• What form rules take
• Degree of specificity
• Formalism

3
Turing machine vs switch network

• memory -
• rules/algorithms/generalisations -

4
Arguments for Turing machine (or against
connectionism)

• Gallistel 2006
• Dead reckoning
• Bee dances
• Temporal learning in conditioning experiments
• truly random control (Rescorla 1968)
• Blocking (Kamin 1969)
• Minsky and Papert 1969 on 2-layer networks
• Exclusive OR (thanks to Marc)
• Cant correctly indicate at its output neuron (or
  neurons) whether there are an even or an odd
  number of neurons firing in its input layer
• Berent et al. 2006 on plurals in English
  compounds
• Vaux
• MSCs, as well see later

5
UR?SR mappings and rules

• We saw in lecture 1 that humans store both
  abstract underlying representations (URs) and
  more concrete surface representations (SRs)
• How does one get from one type of representation
  to the other?
• Hypothesis 1 each is simply memorized
• Hypothesis 2 UR?SR mappings encoded in
  associative/connectionist network
• Hypothesis 3a All URs are transformed into SRs
  (and perhaps vice versa) by an ordered series of
  rules
• Hypothesis 3b Only regular UR?SR mappings
  involve rules

Why favor this one?
6
Animals extract generalisations
7
Generalisation by animals
Gallistel, C. 2003. Conditioning from an
information processing perspective. Behavioural
Processes 61.31234 1-13.
8
Generalisation by infants

• Marcus et al 1999
• Question
• Do infants extract linguistic generalisations,
  and in what form?
• Method
• 16 infants randomly assigned to one of two
  groups, each familiarized with 2-minute speech
  sample
• ABA group 3 reps of each of 16 3-word sentences
  from ABA grammar (ga ti ga, li na li, etc.)
• ABB group same with ABB grammar (ga ti ti, etc.)
• After habituation, testing on sentences of 3
  novel nonce words
• test sentences varied as to whether they were
  consistent or inconsistent with the grammar of
  the habituation sentences.
• Because none of the test words appeared in the
  habituation phase, infants could not distinguish
  the test sentences based on transitional
  probabilities, and because the test sentences
  were the same length and were generated by a
  computer, the infant could not distinguish them
  based on statistical properties such as number of
  syllables or prosody.
• Results
• The infants attend longer to sentences with
  unfamiliar structures.
• Conclusions
• Results suggest that infants can represent,
  extract, and generalize abstract algebraic rules.

Mean time spent looking in the direction of the
consistent and inconsistent stimuli in each
condition for experiments 1, 2, and 3.
9
Conclusions about generalisation extraction

• Ample evidence that humans extract
  generalisations from patterns of data in the real
  world
• These are directly captured in rules
• These are not captured insightfully (or sometimes
  at all) by switch-network models (surface
  constraints, connectionist networks)

10
Evidence for rules
11
Internal evidence

• A typical line of argumentation
• When does glottalization occur?
• sat
• Atlantic, atmosphere, coat-tails
• tap, atrocious
• Since glottalization/unrelease is predictable, we
  dont want this to be part of the underlying
  representation, under the assumption that
  speakers dont store redundant information.
• If this is the case, we need a rule to glottalize
  stops in the appropriate environments.
• What form should this rule take?

12
External evidence

• Productivity
• Child and adult Wug tests, e.g. Pinker and
  Ullmann on novel plurals
• Speech therapy
• Click girl undoing her problem with lightning
  quickness
• Syllable deletion in speech errors
• unanímity jun??nIm?Ri ? unámity ju?nQm?Ri
• treméndously tHr??mEnd?sli ? trémenly
  ?tHrEm?nli
• specifícity spEs??fIs?Ri ? specífity
  spE?sIf?Ri
• What is the error in each case?
• We need a rule to assign a new stress in these
  words if there were no rules, we should expect
  the forms to be stressless
• What sort of rules do we need to account for the
  outcome of these errors?
• First-language acquisition phenomena
• Over-regularization (goed for went, etc.)
• Transfer in second-language acquisition
• Speakers have trouble suppressing L1 rules
• Japanese/Korean palatalization, epenthesis
• English aspiration
• Hard to explain this phenomenon (L1
  non-suppression) without rules

13
What about extraction of generalisations from
less clear patterns?

• Morphophonemic rules
• Static patterns

14
What about generalisations that have exceptions?

• English Vowel Shift is productive for some
  speakers for some vowels
• Cena 1978, Jaeger 1980, McCawley 1986
• Pierrehumbert 2002
• English /k/ ? s / _ i in Latinate contexts
• electric-ity vs cheek-y ? cheesy
• Is the rule active, or just a historical remnant?
• Method
• ADJ?N (back formation)
• In Pierres entire career as a curator, he had
  never before seen such a perfect example of
  hovacity. It was an electrifyingly ______
  sculpture.
• N?ADJ (forward formation)
• Before Pierre stood an electrifyingly hovac
  sculpture. In his entire career as curator, he
  had never before seen such a perfect example of
  ______.
• Results
• The alternation was productive, but only for
  Latinate and semi-Latinate targets.

15
What about generalisations that show no
alternations?

• Esper 1925
• Test subjects break up nonce words into morphemes
  based on phonotactics of their L1
• Moreton 1999
• Speakers have active knowledge of constraint on
  monosyllables ending in lax vowel, which they use
  in speech perception
• Pater and Tessier 2003
• toy grammars easier to acquire when their
  alternations conform to phonotactic
  generalizations in their L1
• Dell et al 2000
• speech errors conform to phonotactics of data in
  toy language
• Cebrian 2002
• native English speakers, and Catalan learners of
  English, use this restriction in interpreting the
  morphological composition of nonce words.
• Vaux 2003
• Productivity of MSCs
• Kaun and Harrison 1999 on Tuvan reduplication

16
Tuvan overwriting reduplication

• Common assumption among phonologists
• Non-alternating structure is stored as such in
  underlying forms.
• Alternating structure is not stored in URs.
• Alternation Condition (Kiparsky 68), Lexicon
  Optimization (PS 93)
• Kaun and Harrison 1999
• Observation Tuvan VH all vowels in a root agree
  wrt ?back
• Question does vowel harmony apply to
  non-alternating forms?
• Method teach subjects Jocular Reduplication see
  if new V triggers root harmony
• Replace first vowel of root with a nom book ?
  nom-nam
• If root vowel is a, replace it with u at
  name ? at-ut
• Results harmonic forms reharmonize, disharmonic
  forms dont
• Harmonic words idik boot ? idik-adik (not
  adik)
• Disharmonic words maina car ? maina-muina
  (mui/una)

17
Tuvan overwriting reduplication

• Conclusions
• Disharmonic forms are fully specified
  underlyingly
• Harmonic forms are not (Free Ride, McCarthy
  2004)
• Theoretical implication
• Generalisations can be formed over
  non-alternating phonological material

i d i k m a i n a
-bk b -b b
18
Technical aspects of rule formalism
19
The formal statement of rules

• Rules take the general form A ? B /X_Y
• A target of rule, an element in UR
• ? becomes
• B what the segment containing A becomes
• / in the environment of
• _ position of the target A
• X element left-adjacent to A (can be absent)
• Y element right-adjacent to A (can be absent)
• word boundary
• Ø zero/nothing
• /X/ underlying form
• X surface form
• stray segment
• (X) optional segment
• a,ß,? variables

20
Key rule types

• Insertion
• Ø ? A / B _ C
• Insert A between any BC sequence
• Ø ? A / _
• Insert A word-finally
• Deletion
• A ? Ø / B _ C
• Delete A between B and C
• A ? Ø / _
• Delete A word-initially
• Alpha Rule
• aX ? -aX / B _ s
• Invert the feature specification for X when it
  occurs after B at the end of a syllable

21
Desiderata in rules

• Keys
• Elsewhere Case UR
• Use as few rules as possible
• This includes trying to collapse rules dealing
  with (seemingly) separate phenomena, such as the
  English plural and other voice assimilation
  processes
• Be as general as possible
• E.g. try stops rather than p t
• Be as predictive as possible
• a rule that merely describes the facts is
  essentially useless
• The last two points normally boil down to the
  same thing (use as few features as possible,
  etc.)
• Linguists generally temper their rule
  formulations with consideration of what is
  (typologically) plausible

22
Choosing a UR

• Relevance of Elsewhere Case
• English aspiration
• Insertion of material is less common than
  deletion
• Generalisation avoid insertion/creation of
  arbitrary elements
• Consideration of rule typology
• Final devoicing, palatalization, etc.
• An example
• Sound X occurs only at the ends of words, while
  sound Y occurs anywhere but at the ends of words.
  Which of the following rules is most likely to be
  involved?
• X ? Y / ___
• Y ? X / ___
• X ? Y everywhere but / ___

23
Use as few rules as possible

• English aspiration
• p ? ph, t ? th, k ? kh 3 rules
• -voice, -cont ? spread glottis 1 rule

24
Use as few features as possible

• Voicing neutralization
• Russian voiced obstruents become voiceless
  word-finally
• Voiced obstruents voice, -sonorant,
  consonantal
• Relatively specific formulation
• voice, -son, cons ? -voice, -son, cons / _
  
• More general/predictive formulation, using fewer
  features
• -son ? -voice / _
• voiceless obstruents vacuously undergo the rule

25
Spanish spirantization

• Noun definite gloss
• banca baNka la banca la BaNka bank
• demora demoRa la demora la DemoRa delay
• gana gana la gana la ?ana desire
• What are the segments targeted by the rule?
• In what environment(s) do they undergo the rule?
• The set of sounds that undergoes this change is
  the voiced stops, i.e. the natural class of
  consonantal, -sonorant, -continuant, voice
  segments.
• The set of sounds produced by the rule is the
  voiced fricatives, i.e. the natural class of
  consonantal, -sonorant, continuant, voice
  segments.
• The set of sounds that triggers the change is
  vowels and r, i.e. the natural class of
  continuant segments.
• We could therefore say
• cons, -son, -cont, voice ? cons, -son,
  cont, voice / cont_
• However, we want to be as general and efficient
  as possible. Therefore
• voice ? continuant / continuant _

26
English plural formation

• Formation of regular plurals of nouns in English
• cat cats
• dog dogz
• ash ash?z
• Possible analyses
• 1. Memorize each word and its plural form.
• 2. Memorize 3 plural endings assign each word
  to class 1, 2, or 3.
• 3. plural ?
• s after p t k ...
• z after b d g ...
• ?z after
• 4. several general rules (holding over domains
  broader than the plural)
• Plural selection plural ? /-z/
• Epenthesis Ø ? ? / _ cf. knish
• Voicing Assimilation -son ? avoice / _
  avoice s cf. fif-th
• Predictions?
• Analyses 1 and 2 predict that speakers will be
  unable to deal with foreign and made-up words.

27
What does each model predict?

• Some theories of English plural formation
• rule-based
• pl ? -z / aeioubdgmn?ð _
• -s / ptk?f _
• -?z / szcj _
• pl ? -z / voice, -strident _
• -s / -voice, -strident
  _
• -?z / strident _
• pl ? -?z / strident _
• -s / -voice _
• -z / elsewhere
• rule 1 pl ? /-z/
• rule 2 Ø ? ? / _
• rule 3 cons ? -voice / -voice _
• probabilistic (? analogical, connectionist)
• wug PL ? 70 of g-final words take -z ? 70
  wugz
• wug PL ? 70 of g-final words take -z ? 100
  wugz
• memory-based

unordered
ordered
28
Conclusions
29
Avoiding insertion/creation

• Proto-Polynesian C ? Ø / _
• Synchronic analysis
• Passive /-ia/, gerundive /-a?a/
• V ? Ø / V _
• Better to have C-deletion rule than to have many
  allomorphs for the passive and the gerundive
• The allomorphy analysis also incorrectly predicts
  the existence of roots selecting -tia but -ma?a
• NB Maori actually did later choose the allomorphy
  analysis, and then made -tia its default form

Hale, Ken. 1973. Deep-surface canonical
disparities in relation to analysis and change
An Australian example. Current Trends in
Linguistics 11401-458.
30
References

• Armbruster, Thomas. 1978. The Psychological
  Reality of the Vowel Shift and Laxing Rules
  Dissertation Abstracts International.
  391516A-17A.
• Aske, Jon. 1990. Disembodied Rules versus
  Patterns in the Lexicon Testing the
  Psychological Reality of Spanish Stress Rules
  Berkeley Ling. Soc.. Berkeley 30-45. Proceedings
  of the Sixteenth Annual Meeting of the Berkeley
  Linguistics Society, February 16-19, 1990
  General Session and Parasession on the Legacy of
  Grice. Hall, Kira (ed.) Koenig, Jean-Pierre
  (ed.) Meacham, Michael (ed.) Reinman, Sondra
  (ed.) Sutton, Laurel A. (ed.).
• Berent, Iris, Steven Pinker, G. Ghavami, and S.
  Murphy. 2006. The Dislike of Regular Plurals in
  Compounds Phonological Familiarity or
  Morphological Constraint? Manuscript, Harvard
  University.
• Bernstein Ratner, N. 1984 Phonological rule usage
  in mother-child speech. Journal of Phonetics
  12245-254.
• Cena, R. 1978. When is a phonological
  generalization psychologically real? Bloomington
  Indiana University Linguistics Club.
• Dell, Gary, Reed, K.D., Adams, D.R., Meyer, A.
  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 61355-1367.
• Gallistel, C. Randy. 2003. Conditioning from an
  information processing perspective. Behavioural
  Processes 61.31234 1-13.
• Gallistel, C.Randy. 2006. The nature of learning
  and the functional architecture of the brain. In
  Q. Jing, et al (Eds) Psychological Science Around
  the World, vol 1. Proceedings of the 28th
  International Congress of Psychology. Sussex
  Psychology Press.
• Hale, Ken. 1973. Deep-surface canonical
  disparities in relation to analysis and change
  An Australian example. Current Trends in
  Linguistics 11401-458.
• Hauser, Marc, Daniel Weiss, and Gary Marcus.
  2002. Rule learning by cotton-top tamarins.
  Cognition 86B15B22.
• Hetzron, Robert. 1972. The Shape of a Rule and
  Diachrony. Bulletin of the School of Oriental and
  African Studies 35.3451-475.
• Iverson, Greg. 1994. The Reality of Linguistic
  Rules (Studies in Language Companion Series, no.
  26), ed. with S. Lima R. Corrigan. Amsterdam
  John Benjamins.
• Janda, Richard, Brian Joseph, and Neil Jacobs.
  1992. Systematic hyperforeignisms as maximally
  external evidence for linguistic rules. In
  Iverson et al, the reality of linguistic rules.
• Marcus, Gary, S. Vijayan, S. Bandi Rao, and P.
  Vishton. 1999. Rule learning by seven-month-old
  infants. Science 283.5398.
• Minsky, Marvin and Seymour Papert. 1969.
  Perceptrons. Cambridge MIT Press.
• Moreton, Elliott. 1999. Evidence for phonological
  grammar in speech perception. In J. J. Ohala, Y.
  Hasegawa, M. Ohala, D. Granville, and A. C.
  Bailey (eds.), Proceedings of the 14th
  International Congress of Phonetic Sciences, San
  Francisco, pp. 2215-2217.
• Pater, J. and A.-M. Tessier. 2003. Phonotactic
  Knowledge and the Acquisition of Alternations. In
  M.J. Solé, D. Recasens, and J. Romero (eds.)
  Proceedings of the 15th International Congress on
  Phonetic Sciences, Barcelona. 1777-1180.
• Pierrehumbert 2002, an unnatural process. LabPhon
  8.
• Pinker and Prince. 1994. Regular and irregular
  morphology and the psychological status of rules
  of grammar. In S. D. Lima, R. L. Corrigan, G. K.
  Iverson (eds.), The reality of linguistic rules,
  321-51. Amsterdam Benjamins.

31
Truly random control

• Shows that
• statistical correlations of the sort if CS then
  US do not drive generalisation formation
• Categorical generalisations can be extracted from
  gradient distributions
• vs 33 response as one might expect for Group 1

32
Blocking

• Shows that something beyond statistical
  association is taking place

33
Esper 1925

• Method
• Ss learn names of 16 objects, each having one of
  four different shapes and one of four different
  colors
• Ss trained on 14 object-name associations but
  tested on 16 to see if they generalize what they
  learned
• 3 experimental conditions
• names presented to Group 1
• naslig, sownlig, nasdeg, sowndeg, where nas- and
  sown- coded color and -lig and -deg coded shape
• Since these names consisted of two phonologically
  legal morphemes, this group could simplify their
  task by learning not 16 names but 8 morphemes (if
  they could discover them) plus the simple rule
  that the color morpheme preceded the shape
  morpheme in each name.
• Names presented to Group 2
• bi-morphemic names, as with Group 1
• unlike group 1, the morphemes were not
  phonologically legal for English, e.g., nulgen,
  nuzgub, pelgen, pezgub (where nu- and pe- were
  color morphemes and -lgen and -zgub were shape
  morphemes, the latter two violating English
  morpheme structure constraints)
• Names presented to Group 3 (a control group)
• names with no morphemic structure
• no recourse but to learn 16 idiosyncratic names
• Results
• As expected, group 1 learned their names much
  faster and more accurately than group 3.
• Performance of Group 2 was similar to (and
  marginally worse than) that of group 3
• Analysis of the errors of group 2, including how
  they generalized what theyd learned to the two
  object-name associations excluded from the
  training session, revealed that they tried to
  make phonologically legal morphemes from the
  ill-formed ones. 
• Demonstrates (i) psychological reality of MSCs
  (ii) ability to conduct morphological analysis

34
Korean borrowing of Coda t

• Korean word-final t ? /t, th, t, c, ch, c,
  s, s/
• Surface word-final postvocalic t in loans and
  nonce words invariably assigned to /s/ (Martin
  1992, Kang 1998, Hayes 1998, Iverson Lee 2004)
• supermarket ? nom. sup?makhet, dat.
  sup?makhese
• What appears to be involved in the Korean case is
  that speakers know that surface word-final ts
  most often come from underlying /s/ in their
  native lexicon, and they therefore assign all new
  words to the same pattern.