A beginner’s guide to Optimality Theory
Moira Yip UCL
1
Part I
The Basics
2
The philosophy
Q: Why does phonology exist? A: To improve the form of phonological outputs Phonologists have always said things like ‘The glottal stop is inserted because the syllable needs an onset” OT builds this directly into the grammar.
3
The pioneers…
4
Alan Prince, n.d.
Paul Smolensky
Alan Prince
5
Structure of an OT Grammar
Output-based theory, for a given input all possible outputs are compared against a set of constraints One-step, non-derivational Universal set of constraints Two types: Markedness constraints, that penalize disfavored outputs, e.g. NoCoda Faithfulness constraints, that penalize changes to input: e.g. NoDeletion
6
How do languages differ?
Every language uses the same set, but prioritizes them differently Constraints may compete. For a CVCCV input, compare two syllabifications
Conflicts resolved by ranking constraints. English ranks NoCoda >> *Complex Onset (re.ply), but Cantonese reverses this (yap.lej)
7
CV.CCV violates *ComplexOnset but CVC.CV violates NoCoda Neither output is perfect
Why aren’t all syllables [ta]?
Faithfulness constraints stop wholesale return to an unmarked state. Main trio:
NoDeletion (Max) NoInsertion (Dep) NoChange (Ident)
8
Example
Suppose input is /kæt/ Output [kæt] violates NoCoda [kæ] violates NoDeletion [kæ.ti] violates NoInsertion If NoDeletion, NoInsertion >> NoCoda [kæt] will win.
9
Extensions of Faithfulness
Faithfulness relates input to output, but extended to: Base-reduplicant relations Paradigms and sets of related words Loanword to source language
10
How it works
Maori:
Active
hopu aru
Passive -ia
hopukia arumia
Gerund -aŋa
hopukaŋa ‘to catch’ arumaŋa ‘to follow’
NoCoda >>NoDeletion
11
A tableau
/hopuk/ a. ho.pu
b. ho.puk *!
NoCoda
NoDeletion *
Higher-ranked constraints to left Shows the winner Shading shows cells that are irrelevant because higher constraint has decided *= violation, *! = fatal violation
12
German final devoicing
lan[t] ta[k] landen tage ‘to land’ ‘days’ NoDeletion NoChange *
*!
/tag/ *VoicedCoda a. tak
b. Ta
c. Tag
*!
13
Factorial typology I
NoCoda, Onset, *Del, *Ins
*Del, *Insert >> NoCoda, Onset
Onsets opt., and codas permitted. NoCoda, Onset >> *Del, *Insert No codas, and onset obl. NoCoda >> *Del, *Insert >> Onset No codas, onsets opt. Onset >> *Del, *Insert >> NoCoda Onsets obl, codas permitted
English Hua, Siona
Hawaiian Cantonese
14
Factorial typology cont
*Del >> Onset/NoCoda >> *Insert
Repairs by epenthesis
*Insert >> Onset/NoCoda >> *Del
Arabic
Repairs by deletion
Maori
15
Factorial typology cont..
For n constraints, n! grammars, where n! = n(n-1)(n-2)…. But some constraints may not interact, so their relative ranking may have no effects, and there will not be n! languages. This is true for Onset and NoCoda above
16
Submerged constraints: The Loch Ness monster phenomenon
English violates NoCoda and Onset freely (‘eat’) Yet they govern syllabification of intervocalic C: ’guitar’ [g.ta:], not [gt.a:] Conclusion: Constraints may be ‘below the surface’, but are not ‘turned off’ Distinguishes OT from a parameter approach
17
Thanks to Mary Pearce
18
The Emergence of The Unmarked (TETU)
A language may violate constraints in most of the grammar, but observe them in certain circumstances Reduplication: only CV syllables may surface, even if language allows CCV or CVCC elsewhere: Tagalog ta-trabaho Ponapean ke-kens Faith-IO >> Markedness >> Faith-BR
19
Part II: Issues and results
(phonology and morphology only!)
20
Are all rankings possible grammars?
Factorial typology says ‘Yes’. However, some are thought to be fixed in UG, usually because they are rooted in phonetics. E.g. All languages prefer vocalic nuclei, but some allow sonorants (English ‘little’ [l.tl]), a very few allow stops (Berber [tf.tkt]) One-way entailment: If stops then sonorants, but not vice-versa *Nucleus/stop >> *Nucleus/Sonorant >> 21 *Nucleus/vowel
How can we limit the constraint set?
All should be plausibly universal The resulting factorial typology should not predict unlikely grammars Introduction of a new constraint should be a last resort, only when an effect cannot be shown to result from the interaction of existing constraints.
22
Are all constraints universal and innate?
Maybe. Some proposals instantiate morphemes as constraints, and these must be language-specific and learnt. E.g. English: Plural=s, Possessive=s. For possessive plurals, one [s] satisfies both constraints, so there is no need to add two: cats’ , * cats’s There may be schemas, from which the child ‘builds’ a constraint. E.g. Align L/R (X,Y) where X, Y are morphological or prosodic categories.
23
The opacity problem:
In phonology, opacity refers to cases where the context for a process is not surface apparent (or not surface-true). Tiberian Hebrew: /deš/ becomes [deše], as a result of epenthesis into a final cluster (cf /melk/ > [melex]) and deletion of final glottal stop (cf. /qara/ > [qara]). Ordered rules handle this easily. /deš/ deše deše BUT: In non-derivational OT, the easiest way to satisfy *CC and * ] is to choose [deš].
24
Why [deš] can never win
/deš/ *CC deš deše
*]
NoDel NoIns * * *!
25
Solutions? Not yet, but watch this space…
Deny the phenomenon (Mielke and Hume, Green) Sympathy theory (McCarthy) Constraint conjunction (Kirchner, Ito and Mester) Output-output constraints (Steriade, Burzio) Stratal OT (Kiparsky) Comparative markedness (McCarthy) Candidate chains (McCarthy)
26
John McCarthy:
Somewhere in between these pictures he did some OT……………………..
2001
1971
27
What is an ‘output’?
How much phonetic detail do outputs contain? Maybe a great deal. Work of Steriade and her students shows faithfulness to fine phonetic details of duration, etc, and role of perceptual cues in the grammar.
28
Language variation
Some models allow free ranking, so that two candidates tie, and suggest that this accounts for 50/50 free variation.(Antilla) Others propose that constraints have a range of ranking positions, and that one can read off this the actual frequencies of different variants (Stochastic OT: Hayes, Boersma)
29
Constraint A
Constraint B
Constraint A >> Constraint B most of the time, but occasionally B >> A, for example if we select the rankings at the arrows
30
First-language acquisition
Many data-based studies, including Gnanadesikan 1995, Hayes 1999, Boersma 2001, Dinnsen and Gierut 2001, Goad 2001, Stemberger and Bernhardt 2002, Pater and Wehrle 2001, Also much work on computational issues and learning algorithms, especially Tesar and Smolensky, Prince, Samek-Lodovici, Pulleyblank and Turkel, and others
31
Initial state
Language acquisition must involve re-ranking, as the grammar evolves Likely starting point: Markedness >> Faithfulness. Early utterances unmarked. Adult ‘dog’ is child [d], with NoCoda>>NoDeletion. Re-ranking happens when child ‘notices’ NoCoda violations in adult, and demotes NoCoda below NoDeletion. Gradual Learning Algorithm (GLA) (Boersma and Hayes)
32
Stages in acquisition
Each stage should involve a demotion of a markedness constraint on the basis of positive evidence But stages are not abrupt and discrete: stochastic OT. GLA makes only tiny adjustments, so constraints ‘slide’ past each other, allowing for a period of variation.
33
What can go wrong?
Deviant constraint set? Deviant initial ranking? Deficit in learning algorithm? A few studies, such as Specific Language Impairment (SLI), phonological disorders, Prader-Willi Syndrome (Gamble, MFM 2005)
34
SLI: Marshall 2005
Some SLI children (selected for morphosyntactic deficits) also have phonological disorder. Past tense affixation may be omitted if it would create a voiced cluster: rob, not robbed. English observes this constraint in monomorphemic words. *VoicedCluster>> Realize-affix Constraint set seems to be normal. They seem to have defective re-ranking algorithms; unmarked utterances persist after usual age. 35
Common vs rare phonological disorders (Kinney 2004)
Common disorders: Markedness >> Faithfulness: *Fric >> Nochange: Stopping Less common : Unusual rankings of faithfulness constraints. NoDel >> NoCoalescence, so /sp/ > [f] Rarest: Reversal of usual markedness hierarchies: *s, *h, *t >> *fŋ, so /s/ > fŋ
36
Types of disorder
M >> F: normal, but delayed. Learning algorithm at fault F 2>>F1: possibly skewed initial state M2 >>M1: Oddly structured constraint set
37
Stuttering
Diment, Howell, and Harris 2005, and many other studies from Pete Howell’s lab. Not OT, but nonetheless show a clear role of phonological markedness in increased stuttering rates More likely in forms which violate:
NoCoda *Fricative *ComplexOnset Both : sCC clusters
Onset
38
BUT: Less likely in syllables which violate:
The assessment and treatment of phonological disorders (Barlow 2001)
Generally, markedness >> faithfulness, producing stopping (‘shoe’ > [tu]), cluster reduction, coda deletion (‘sled’ > [l]), gliding (‘rain’ > [wen]) Specific treatment goals target constraints that need to be demoted. Most marked aspects targeted first, producing greatest change. For example, if final f,l,r are targeted, all of *Fric, *liquid, NoCoda will be demoted at once. Effects thus widespread: If ‘tell’ and ’Tess’ can be correctly produced, ‘sled’ will be too.39
Summary
OT developed for adult language competence, Implies a theory of acquisition, including possible pathways Suggests ways of thinking about the locus of disorders May predict that certain interventions may be most effective
40
Today’s programme
Alan Prince, Rutgers Hubert Truckenbrodt, Tübingen Vieri Samek-Lodovici, UCL
41