Class 6: Underlying forms Adam Albright (albright@mit.edu) LSA 2017 - - PowerPoint PPT Presentation

Class 6: Underlying forms

Adam Albright (albright@mit.edu)

LSA 2017 Phonology University of Kentucky

Announcements

▶ For those taking this class for credit

▶ Option 1: assignment 2 due today, assignment 3 will be posted

today (due next Monday 7/31)

▶ Option 2: short paper/squib due next Monday 7/31

▶ Game plan for the last few sessions

▶ T

• day: underlying forms

▶ Thursday: learning constraint rankings ▶ Next Monday: phonological typology

References 1/46

A pedagogically conventional phonology problem

Sg. Pl. Gloss diːp diːbə ‘thief’ liːt liːdɐ ‘song’ kriːk kriːɡə ‘war’ braːf braːvə ‘good’ (adj.) kraɪs kraɪzə ‘circle’ Sg. Pl. Gloss pʁɪntsiːp pʁɪntsiːpjən ‘principle’ gəbiːt ɡəbiːtə ‘area’ blɪk blɪkə ‘glance’ ʃaːf ʃaːfə ‘sheep’ ɡlaɪs ɡlaɪsə ‘track’ ▶ Step 1: morpheme segmentation

Step 2: notice that root-final voicing is unpredictable Step 3: notice that root-final voicing is contextually neutralized Step 4: hypothesize underlying representations (UR’s) for morphemes

/diːb/, /liːd/, /kʁiːɡ/, /pʁɪntsiːp/, /ɡəbiːt/

Step 5: devise rules/constraint rankings that change UR’s into

• bserved surface forms in the right contexts

Final devoicing

References 2/46

A pedagogically conventional phonology problem

Sg. Pl. Gloss diːp diːb–ə ‘thief’ liːt liːd–ɐ ‘song’ kriːk kriːɡ–ə ‘war’ braːf braːv–ə ‘good’ (adj.) kraɪs kraɪz–ə ‘circle’ Sg. Pl. Gloss pʁɪntsiːp pʁɪntsiːp–jən ‘principle’ gəbiːt ɡəbiːt–ə ‘area’ blɪk blɪk–ə ‘glance’ ʃaːf ʃaːf–ə ‘sheep’ ɡlaɪs ɡlaɪs–ə ‘track’ ▶ Step 1: morpheme segmentation ▶ Step 2: notice that root-final voicing is unpredictable

Step 3: notice that root-final voicing is contextually neutralized Step 4: hypothesize underlying representations (UR’s) for morphemes

/diːb/, /liːd/, /kʁiːɡ/, /pʁɪntsiːp/, /ɡəbiːt/

Step 5: devise rules/constraint rankings that change UR’s into

• bserved surface forms in the right contexts

Final devoicing

References 2/46

A pedagogically conventional phonology problem

Sg. Pl. Gloss diːp diːb–ə ‘thief’ liːt liːd–ɐ ‘song’ kriːk kriːɡ–ə ‘war’ braːf braːv–ə ‘good’ (adj.) kraɪs kraɪz–ə ‘circle’ Sg. Pl. Gloss pʁɪntsiːp pʁɪntsiːp–jən ‘principle’ gəbiːt ɡəbiːt–ə ‘area’ blɪk blɪk–ə ‘glance’ ʃaːf ʃaːf–ə ‘sheep’ ɡlaɪs ɡlaɪs–ə ‘track’ ▶ Step 1: morpheme segmentation ▶ Step 2: notice that root-final voicing is unpredictable ▶ Step 3: notice that root-final voicing is contextually neutralized

Step 4: hypothesize underlying representations (UR’s) for morphemes

/diːb/, /liːd/, /kʁiːɡ/, /pʁɪntsiːp/, /ɡəbiːt/

Step 5: devise rules/constraint rankings that change UR’s into

• bserved surface forms in the right contexts

Final devoicing

References 2/46

A pedagogically conventional phonology problem

Sg. Pl. Gloss diːp diːb–ə ‘thief’ liːt liːd–ɐ ‘song’ kriːk kriːɡ–ə ‘war’ braːf braːv–ə ‘good’ (adj.) kraɪs kraɪz–ə ‘circle’ Sg. Pl. Gloss pʁɪntsiːp pʁɪntsiːp–jən ‘principle’ gəbiːt ɡəbiːt–ə ‘area’ blɪk blɪk–ə ‘glance’ ʃaːf ʃaːf–ə ‘sheep’ ɡlaɪs ɡlaɪs–ə ‘track’ ▶ Step 1: morpheme segmentation ▶ Step 2: notice that root-final voicing is unpredictable ▶ Step 3: notice that root-final voicing is contextually neutralized ▶ Step 4: hypothesize underlying representations (UR’s) for

morphemes

▶ /diːb/, /liːd/, /kʁiːɡ/, /pʁɪntsiːp/, /ɡəbiːt/

Step 5: devise rules/constraint rankings that change UR’s into

• bserved surface forms in the right contexts

Final devoicing

References 2/46

A pedagogically conventional phonology problem

Sg. Pl. Gloss diːp diːb–ə ‘thief’ liːt liːd–ɐ ‘song’ kriːk kriːɡ–ə ‘war’ braːf braːv–ə ‘good’ (adj.) kraɪs kraɪz–ə ‘circle’ Sg. Pl. Gloss pʁɪntsiːp pʁɪntsiːp–jən ‘principle’ gəbiːt ɡəbiːt–ə ‘area’ blɪk blɪk–ə ‘glance’ ʃaːf ʃaːf–ə ‘sheep’ ɡlaɪs ɡlaɪs–ə ‘track’ ▶ Step 1: morpheme segmentation ▶ Step 2: notice that root-final voicing is unpredictable ▶ Step 3: notice that root-final voicing is contextually neutralized ▶ Step 4: hypothesize underlying representations (UR’s) for

morphemes

▶ /diːb/, /liːd/, /kʁiːɡ/, /pʁɪntsiːp/, /ɡəbiːt/

▶ Step 5: devise rules/constraint rankings that change UR’s into

• bserved surface forms in the right contexts

▶ Final devoicing

References 2/46

Reverse engineering phonological systems

▶ We observe the set of forms that speakers produce ▶ We wish to infer

▶ A grammar: determines set of potential forms ▶ A lexicon: determines what speakers actually produce

▶ Division of labor

▶ We have argued that grammar should be restrictive: concentrate

probability mass on the type of outputs that are seen

▶ However, it’s likely that not all unattested strings/words are

ungrammatical: English [gæt], [kɪf], [dæk], …

▶ Evidence: acceptability judgments (the blick test), readily adopted

as names, truncations, loan words, and slang terms ▶ Some differences in attestedness are accidents of the lexicon

References 3/46

Reverse engineering grammars and lexicons

The data that we observe imposes certain conditions on the analysis

▶ The set of surface forms

liːt, braːf, ʃaːfə, kraɪzə, liːdɐ, kriːk, braːvə, ɡlaɪsə, ɡəbiːtə, blɪk, pʁɪntsiːp, diːbə, kriːgə, ʃaːf, diːp, kraɪs, blɪkə, gəbiːt, ɡlaɪs, pʁɪntsiːpjən

▶ Condition: they must be in the range of the function ▶ I.e., grammar must assign non-zero probability to these outputs,

given some input(s)

▶ We don’t know what input(s), and it doesn’t matter for purposes of

satisfying this condition ▶ Paradigmatic relations

liːt ∼ liːd-, gəbiːt ∼ gəbiːt

▶ Desideratum: analyze morphologically related forms as derived

from the same input

▶ Creates an additional condition: there must be input(s) for which

the grammar favors these pairs of surface forms, depending on context

▶ Again, we don’t know what input(s)

References 4/46

Reverse engineering grammars and lexicons

▶ Other information about mappings: loan adaptation, etc.

English [klʌb] ⇒ German [klʊp]

▶ Condition: for some input like the source form, grammar must

favor surface form ▶ These sources of data impose conditions on the solution (the

range of the function, and, to a lesser extent, mappings), but do not directly tell us how speakers represent the actual morphemes

• f their language.

References 5/46

Underdetermination

▶ German: voiced pulmonic stops (b,d,g) do not contrast with

voiced implosive stops (ɓ,ɗ,ɠ)

▶ Conjecture: implosives map to their pulmonic counterparts

/ɗa/ *[+implos] Ident([±voi]) *[voi]/ [+son] Ident([±implos])

• a. ɗa

*! W * L

☞ b. da

* *

• c. ta

*! W L *

▶ German has voiced obstruents before sonorants, but not before

• bstruents or word-finally

/bad/ *[voi]/ ¬[+son] Ident([±voi]) *[voi]/ [+son] a. bad *! W L

☞

b. bat * * c. pat **! W L

References 6/46

Underdetermination

Putting this together /ɓaɗ/ *[+implos] *[voi] Ident *[voi] Ident / ¬[+son] (±voi) / [+son] (±implos)

• a. ɓaɗ

*!* W * W L * L

• b. ɓad

*! W * W L * * L

• c. ɓat

*! W * * * L

• d. baɗ

*! W * W L * * L

• e. bad

*! W L * **

☞ f. bat

* * **

• g. pat

**! W L ** /ɓaɗ+ə/ *[+implos] *[voi] Ident *[voi] Ident / ¬[+son] (±voi) / [+son] (±implos)

• a. ɓaɗə

*!* W ** L

• b. baɗə

*! W ** * L

☞ c. badə

** **

• d. batə

*! W * L **

References 7/46

Underdetermination

Putting this together /bad/ *[+implos] *[voi] Ident *[voi] Ident / ¬[+son] (±voi) / [+son] (±implos)

• a. ɓaɗ

*!* W * W L * ** W

• b. ɓad

*! W * W L * * W

• c. ɓat

*! W * * * W

• d. baɗ

*! W * W L * * W

• e. bad

*! W L *

☞ f. bat

* *

• g. pat

**! W L /bad+ə/ *[+implos] *[voi] Ident *[voi] Ident / ¬[+son] (±voi) / [+son] (±implos)

• a. ɓaɗə

*!* W ** ** W

• b. baɗə

*! W ** * W

☞ c. badə

**

• d. batə

*! W * L

References 8/46

Underdetermination

▶ Intuitively: since Ident([±implos]) is ranked below all Markedness

constraints on [±implos], output does not depend on whether input is [+implos] or [−implos]

▶ Therefore, this grammar is consistent with multiple assumptions

about the lexicon

▶ Different considerations favor different solutions

▶ Economy: since [±implosive] is predictable in German, omit

specification from the lexicon altogether

▶ Transparency/concreteness: since segments are always

[−implosive] in German, represent them as such in the lexicon

References 9/46

Output-Driven Maps

T esar (2008, et seq)

▶ T

ableaux above illustrate a nice property of the “core” OT system that we’ve been using here

▶ If /A/ maps unfaithfully to [B], then /B/ must also map to [B] ▶ T

esar: output-driven map

▶ Intuition about why: harmonic bounding

▶ This reduces the search space for underlying forms, and can also

helps learners break into the system to start making inferences about rankings

References 10/46

Underscoring an important point

▶ Richness of the Base (ROTB) does not say that any logically

possible input string could be the UR for any existing form

▶ A much more limited condition

▶ Every logically possible input string must map onto some legal

surface string (the grammar is robust)

▶ Some legal surface strings may be the winning output for more

than one input string ▶ This raises a question: which UR’s do speakers actually use to

encode existing morphemes, from among the possibly infinitely many inputs that could map onto them?

References 11/46

Economy of the lexicon

One sensible stance

▶ If feature value [±F] is predictable, don’t list it in the lexicon

▶ List only what’s unpredictable: economical encoding ▶ Globally (implosives are never contrastive in German) ▶ Perhaps even contextually (morpheme-final stops are never labial

• r dorsal in Navajo)

▶ Requires a separate statement defining possible lexical

specifications—i.e., an inventory

▶ Phoneme-based approaches ▶ Morpheme structure rules in feature-based approaches Halle

(1959); Chomsky and Halle (1968)

▶ In OT: Rasin and Katzir (2016)

References 12/46

The duplication problem

▶ A simple model

▶ MSR excluding implosives from the German lexicon ▶ Grammar need not say anything else about implosives

▶ Prince and Smolensky (2002) and others: lexical economy alone

is not sufficient

▶ A common situation: a class of segments is not used for lexical

contrasts, and is avoided/repaired whenever it is (expected to be) created through morpheme concatenation or phonological processes

▶ The duplication problem: exclude from inventory and repair in

grammar ▶ It is difficult to illustrate this problem with implosives in German,

since there is no process that is plausibly expected to create implosives in German

▶ Instead, let’s turn to the German vowel inventory

References 13/46

German *[æ]

The “Standard” German monophthong inventory, roughly Front Back iː, yː uː ɪ, ʏ ʊ eː, øː ə

• ː

ɛ, œ ɔ a, aː

▶ No back unrounded vowels: *ɨ, *ɯ, *ɤ ▶ No front low vowel: *æ

References 14/46

German *[æ]

Umlaut

▶ In certain morphological contexts, such as before -çən ‘dimin’,

back vowels become front bluːmə blyːmçən ‘flower’ hʊnt hʏntçən ‘dog’ bʁoːt bʁøːtçən ‘bread’ ʁɔk ʁœkçən ‘skirt’ bakə bɛkçən ‘cheek’ *bækçən

▶ If the stem has /a/, the result of umlaut is [ɛ] (with raising), not

*[æ] (fronting) bakə bɛkçən ‘cheek’ *bækçən

▶ Constraint ruling out morphemes with *[æ] also applies to the

• utput of (morpho)phonological processes such as umlaut

▶ McCarthy (2002, pp. 71–76): surface-oriented markedness avoids

need to state *æ constraint separately for lexical and derived [æ]

References 15/46

Avoiding duplication

Same ranking rules out lexical and ‘derived’ [æ] /mæt/ *

[ −back +low ]

Umlaut Ident([±back]) Ident([±low]) a. mæt *! W n.a. L

☞

b. mɛt n.a. * c. mat n.a. *! W L /bak+çən/ *

[ −back +low ]

Umlaut Ident([±back]) Ident([±low]) a. bækçən *! W * L

☞

b. bɛkçən * * c. bakçən *! W L L

▶ Grammar does the work of ruling out surface contrast between

[æ] and [ɛ]: both /æ/ and /ɛ/ surface as [ɛ]

▶ No independent constraint on UR’s to rule out listing /æ/ in UR’s. ▶ What (if anything) precludes listing some or all [ɛ] morphemes

with /æ/?

References 16/46

Economy of derivation

Comparing converging paths to same output [bat] /ɓaɗ/ *[+implos] *[voi] Ident *[voi] Ident /

¬[+son]

(±voi) / [+son] (±implos)

☞ bat

* * ** /bad/ *[+implos] *[voi] Ident *[voi] Ident /

¬[+son]

(±voi) / [+son] (±implos)

☞ bat

* * /bat/ *[+implos] *[voi] Ident *[voi] Ident /

¬[+son]

(±voi) / [+son] (±implos)

☞ bat

*

▶ Same output, so derivations differ only in faithfulness violations ▶ Fewest violations: input most similar to output ▶ I.e., /bat/ → [bat] is the most harmonic derivation of [bat] ▶ Lexicon optimization: choose UR that yields the most harmonic

derivation

References 17/46

Lexicon optimization and allophony

▶ English aspiration: two values of [±spread glottis] in

complementary distribution

▶ Unaspirated after fricatives: explain [ɛksˈpleɪn] ▶ Aspirated initially and before stressed sonorants1 ▶ Unaspirated elsewhere: rapid [ˈɹæpəd]

▶ Exercise: prove to yourself that the most harmonic lexical entry

for cat has an aspirated initial stop and unaspirated final stop ([kʰæt])

1“Stressed sonorant” here is meant to include the liquid or glide in clusters such as

apply [əˈpʰlaɪ], where the stop is not immediately before the stressed vowel, but it is before a stressed sonorant+V sequence.

References 18/46

Lexical vs. derivational complexity

▶ Economy and transparency are both sensible principles, that yield

very different solutions regarding the ‘optimal’ encoding of existing morphemes

▶ Challenge: testable predictions that differentiate these two

approaches?

▶ On-going debate

▶ See McCarthy (2002, pp. 71–76) ▶ Rasin and Katzir (2016): conceptual and learnability arguments in

favor of limiting how morphemes are encoded (constraints on URs)

References 19/46

Abstractness, more generally

▶ Recall the condition imposed by morphologically related pairs of

forms: e.g., [liːt], [liːd-ɐ]

▶ There must be input(s) for which the grammar favors these pairs

• f surface forms, depending on context

▶ For German grammar above, there are many: /liːd/, /liːɗ/, etc.

▶ What if different pairs appear to yield ranking paradoxes? E.g.,

T urkish (Inkelas et al., 1997) Sg. Pl. Acc. 1sg.Poss Gloss sanat sanatlar sanatɨ sanatɨm ‘art’ kanat kanatlar kanadɨ kanadɨm ‘wing’ sop soplar sopu sopum ‘clan’ jop joplar jobu jobum ‘club’ gurup guruplar gurubu gurubum ‘group’ etyd etydler etydy etydym ‘etude’

References 20/46

Abstract UR’s to resolve inconsistencies

T urkish (Inkelas et al., 1997) Sg. Pl. Acc. 1sg.Poss Gloss sanat sanatlar sanatɨ sanatɨm ‘art’ kanat kanatlar kanadɨ kanadɨm ‘wing’ sop soplar sopu sopum ‘clan’ jop joplar jobu jobum ‘club’ gurup guruplar gurubu gurubum ‘group’ etyd etydler etydy etydym ‘etude’

▶ Three patterns: [−voi], [+voi], alternation ▶ Lexicon optimization

▶ Underlying [−voi] vs. [+voi], with high-ranking Ident([±voi]) ▶ However, this won’t capture alternating forms

▶ A popular approach: posit some other representation

References 21/46

Abstract UR’s to resolve inconsistencies

▶ Infinite number of possible underlying distinctions that we could

contemplate

▶ Set that could conceivably resolve inconsistencies depends on

theory of faithfulness constraints

▶ E.g., suppose that we can distinguish Ident([±voi])/[+implosive]

≫ Ident([±voi])/[-implosive]

▶ Ident([±voi])/[+implosive] violated by /ɗ/ → [ɗ

̥ ] (=[ƭ]) (perceptually more salient change?)

▶ Ident([±voi])/[+implosive] violated by /d/ → [t] (perceptually less

salient change?) ▶ *[+implosive], Ident([±voi])/[+implosive] ≫ *[voi]/

¬[+son] ≫ Ident([±voi]), Ident([±implosive])

Invariant [t] /t/ Variable [t] ∼ [d] /d/ Invariant [d] /ɗ/

References 22/46

Abstract UR’s to resolve inconsistencies

▶ Suppose that Ident([±voi]) is violated equally by changing an

underlying [±voi] specification (/+voi/ → [−voi]) and by filling in an value where there was none (/∅ voi/ → [+voi])

▶ Equally ‘costly’ to realize underspecified /∅ voi/ (‘D’) as [+voi] or

[−voi]

/maD/ Ident([±voi]) a. mat * b. mad *

▶ Ident([±voi]) ≫ *[voi]/

¬[+son]

Invariant [t] /t/ [−voi] Variable [t] ∼ [d] /D/ [∅ voi] Invariant [d] /d/ [+voi]

▶ This is the actual proposal of Inkelas et al. (1997)

References 23/46

Limits on abstraction

▶ The possibility of arbitrarily many underlying contrasts to capture

distantly related surface distinctions raised conceptual qualms and learnability concerns in early generative phonology

▶ Most egregious: positing a feature value that never surfaces

▶ E.g., /[+implosive]/ for voiced stops ▶ This value never surfaces, and segments that bear it always

surface as [−implosive,+voi] (Kiparsky, 1968: absolute neutralization)

▶ The sole reason for positing it was to create a distinction between

two classes of voiced stops, and exempt one of them from final devoicing ▶ Kiparsky (1968): proposed ban on absolute neutralization ▶ Alternative: simply treat stems like /etyd/ as exceptions to final

devoicing

References 24/46

The allure of abstractness

▶ In spite of vigorous debate about the alternation condition and

abstractness more generally (Kiparsky, 1968; Hyman, 1970; Schane, 1974), and widespread skepticism regarding abstract UR’s, the literature is full of proposals that violate the alternation condition

▶ Kiparsky (1968) ‘How abstract is phonology?’ ▶ Hyman (1970) ‘How concrete is phonology?’ ▶ Harms (1973) ‘How abstract is Nupe?’ ▶ Schane (1974) ‘How abstract is abstract?’

▶ One compelling reason: segments with multiple exceptional

properties

▶ Positing one underlying difference allows us to write a grammar

that distinguishes the relevant segments in multiple ways

References 25/46

Hyman (1970): Nupe (Kwa, Central Nigeria)

▶ Inventory

p, b t, d k,ɡ f, v s, z ʃ, ʒ h kp, ɡb ts, dz tʃ, dʒ m n l, r w j i ī ū u e

• ā

a aː

▶ Consonants are coarticulatorily rounded before o,u and

fronted/palatalized before i,e

▶ kʷu, pʷu, rʷu, mʷu, … ▶ kʲi, pʲi, rʲi, mʲi, …

(Clearest with dorsals, as in English)

References 26/46

Hyman (1970): Nupe (Kwa, Central Nigeria)

▶ A very limited contrast: Ca vs. Cʷa vs. Cʲa

ēɡʷā ‘hand’ tʷá ‘to trim’ ēɡʲà ‘blood’ tʲá ‘to be mild’ ēɡā ‘stranger’ tá ‘to tell’

▶ Underlying Cʷ,Cʲ?

▶ Would need to neutralize before front/round vowels (easy to do

with Markedness)

▶ Hyman (1970) worries that now there are both underlying and

derived Cʷ, Cʲ

▶ Since it’s predictable sometimes, he asks: could it be predictable

here, too?

References 27/46

Hyman (1970): Nupe (Kwa, Central Nigeria)

▶ Unifying all labialized, palatalized C’s

▶ We know palatalization is triggered by a front V, and labialization

by a round V

▶ Could the three a’s really be three different vowels?

▶ Regular /a/, front /a/, rounded /a/

High Back Low Round /a1/

− + + − = a

/a2/

− − + − = æ (ɛ)

/a3/

− + + + = ɒ (ɔ)

▶ Absolute neutralization: /æ/, /ɒ/ → [a]

References 28/46

Hyman (1970): Nupe (Kwa, Central Nigeria)

Corroborating evidence: nominalizing reduplication ɡʲí to eat ɡʲīɡʲí eating ɡʲē to be good ɡʲīɡʲē goodness ɡʷú to puncture ɡʷūɡʷú puncturing ɡʷò to receive ɡʷūɡʷò receiving tʷá to trim tʷūtʷá trimming tʲá to be mild tʲītʲá mildness tá to tell tʲītá telling

▶ Process: copy CV, and raise the vowel to high

▶ CV → C

[ +high αround ]

CV ▶ In top forms, CʲV vs. CʷV is predictable, and carried over to

reduplication

▶ As before: Ca vs. Cʲa vs. Cʷa is not predictable from frontness of

[a], but can be attributed to /Ca/ vs. /Cæ/ vs. /Cɒ/

▶ This corresponds to reduplicant vowel [i] (non-back vowel) vs. [u]

(back vowel)

References 29/46

Hyman (1970): Nupe (Kwa, Central Nigeria)

Corroborating evidence: loanword adaptation from Yoruba Yoruba Nupe Gloss kɛ̀kɛ́ kʲàkʲá ‘bicycle’ ɛ̀ɡbɛ̀ èɡbʲà (a Yoruba town) tɔ̄rɛ̄ tʷārʲā ‘to give a gift’ kɔ́bɔ̀ kʷábʷà ‘penny’

▶ Corroborates assumption that /ɛ/, /ɔ/ map to [ʲa], [ʷa]

▶ Does not prove that they are ever encoded as such in the Nupe

lexicon! But at least it’s consistent with that hypothesis ▶ This is not just sensitivity to coarticulatory labialization and

palatalization in the source language: Hausa has processes, but no reason to posit /ɛ/, /ɔ/; borrowed as e,o

References 30/46

Nupe without absolute neutralization?

▶ Alternative: CV vs. CʲV vs. CʷV are contrastive, but contextually

neutralized before non-low vowels

▶ Neutralize to CʲV before front vowels ▶ Neutralize to CʷV before back vowels

▶ Loanword adaptation of [ɛ], [ɔ]: simply the closest grammatical

match

▶ CʲV, CʷV preserve coarticulatory palatalization/rounding triggered

by source vowel, and height of source vowel ▶ Reduplication: copy the consonant, and choose the only legal

high vowel

▶ *Cʲu, *Cʷi favor Cʲi, Cʷu

▶ See also Harms (1973) for another possible approach

References 31/46

Greenlandic (Underhill, 1976)

sg pl 3sg.poss.sg 3sg.poss.pl pl.refl Gloss illu illut illua illui illutik ‘house’ iki ikit ikia ikii ikisik ‘wound’ ini init inaa inai initik ‘place’ tusarti tusartit tusartaa tusartai tusartitik ‘listener’ siut siutit siutaa siutai siutitik ‘ear’ qimmiq qimmit qimmia qimmii qimmitik ‘dog’ irniq irnirit irnira irniri irnitik ‘son’ Some stem-final [i]’s are consistently [i], and trigger assibilation

• f intervocalic /t/ (ikisik)

Some stem-final [i]’s alternate with [a] in poss., fail to trigger assibilation (initik), and fail to trigger dorsal deletion (irnira) “Funny” [i]’s are also consistently found after clusters (tusarti), but delete after legal singleton codas (siut ) Hypothesis: abstract vowel “ï” Also useful in accounting for other lexical distinctions: nouns ending in k/q, certain verbal forms, etc.

References 32/46

Greenlandic (Underhill, 1976)

sg pl 3sg.poss.sg 3sg.poss.pl pl.refl Gloss illu illut illua illui illutik ‘house’ iki ikit ikia ikii ikisik ‘wound’ ini init inaa inai initik ‘place’ tusarti tusartit tusartaa tusartai tusartitik ‘listener’ siut siutit siutaa siutai siutitik ‘ear’ qimmiq qimmit qimmia qimmii qimmitik ‘dog’ irniq irnirit irnira irniri irnitik ‘son’

▶ Some stem-final [i]’s are consistently [i], and trigger assibilation

• f intervocalic /t/ (ikisik)

▶ Some stem-final [i]’s alternate with [a] in poss., fail to trigger

assibilation (initik), and fail to trigger dorsal deletion (irnira)

▶ “Funny” [i]’s are also consistently found after clusters (tusarti),

but delete after legal singleton codas (siut∅)

▶ Hypothesis: abstract vowel “ï” ▶ Also useful in accounting for other lexical distinctions: nouns

ending in k/q, certain verbal forms, etc.

References 32/46

Reanalysis

▶ Given the right grammar, speakers have two rational options

when learning a new noun ending in [i]

▶ Posit underlying /i/ ▶ Posit underlying /ï/

▶ Grammar doesn’t care which one a give noun has; designed to

yield an attested pattern of alternation for all possible inputs

▶ Other principles for deciding

▶ Lexicon optimization ▶ Statistically predominant value (Harrison and Kaun, 2000)

▶ Underhill (1976): evidence that speakers sometimes choose

“wrong” (i.e., differently from their parents)

• p. 355 “…we find that in many of the cases where we would want

to set up underlying /ï/, there is a tendency for speakers to reshape the forms, so that they are treated as though they contained /i/.…

References 33/46

Alternative without absolute neutralization?

▶ Just one high non-back vowel: /i/ ▶ Exceptions

▶ Some final /i/’s lower in the possessive (ina-a) ▶ Some final /i/’s fail to trigger assibilation (initik) ▶ Some final /i/’s fail to trigger dorsal deletion (irnira)

▶ Change: learners no longer treat a particular lexical item as

exceptional

▶ Overregularization: grammatically expected form replaces

exceptional form

References 34/46

Lexical reanalysis, or grammatical regularization?

▶ The beauty of the /ï/ analysis is that it captures a number of

exceptional properties with a single abstract segment

▶ Just decide /i/ or /ï/, and grammar determines how it will behave

w.r.t. final lowering, assibilation, dorsal deletion, etc. ▶ According to Rischel and Underhill, not all exceptional properties

are equally susceptible to reanalysis/change

▶ Exceptional final i/a is reliably retained ▶ Exceptional dorsals often become subject to deletion: /tikiq+a/

[tikira] ⇒ [tikia] ▶ Predictions

▶ Behavior of a morpheme should follow from whatever underlying

value has been assigned to it

▶ When one property changes, all other properties should

automatically too (consistency)

References 35/46

Variability

▶ Underhill points out that some morphemes show variability

Noun Possessive Gloss ilik iliŋa ∼ ilia ‘helper’ tikiq tikira ∼ tikia ‘index finger’ tupiq tuqqa ∼ tupia ‘tent’

▶ UR’s approach

▶ Variable underlying form: /ilik/, /ilïk/ ▶ Grammar applies predictably, based on the UR that’s chosen

▶ Exceptions approach

▶ One underlying form: /ilik/ ▶ Grammar applies variably, depending on whether morpheme is

treated as exceptional or not

Gouskova and Becker (2013): two different probabilities speakers could learn

Probability of one UR vs. another: ‘morpheme-bound knowledge’ (doesn’t automatically generalize to new morphemes) Probability that a process applies or fails to apply: ‘grammatical knowledge’ (easy to generalize)

References 36/46

Variability

▶ Underhill points out that some morphemes show variability

Noun Possessive Gloss ilik iliŋa ∼ ilia ‘helper’ tikiq tikira ∼ tikia ‘index finger’ tupiq tuqqa ∼ tupia ‘tent’

▶ UR’s approach

▶ Variable underlying form: /ilik/, /ilïk/ ▶ Grammar applies predictably, based on the UR that’s chosen

▶ Exceptions approach

▶ One underlying form: /ilik/ ▶ Grammar applies variably, depending on whether morpheme is

treated as exceptional or not ▶ Gouskova and Becker (2013): two different probabilities speakers

could learn

▶ Probability of one UR vs. another: ‘morpheme-bound knowledge’

(doesn’t automatically generalize to new morphemes)

▶ Probability that a process applies or fails to apply: ‘grammatical

knowledge’ (easy to generalize)

References 36/46

Gouskova and Becker (2013)

Russian mid vowel alternations ʃatʲór ʃatrú ‘tabernacle (nom/dat sg) kavʲór kavrá ‘carpet (nom/gen sg)’ kalʲéts kalʲtsó ‘ring (gen pl/nom sg)’ matór matóru ‘motor (nom/gen sg)’ ɡravʲór ɡravʲórə ‘engraver (nom/gen sg)’ mʲétr mʲétru ‘meter (nom/gen sg)’

▶ Mid vowels unpredictably delete when vowel-initial suffixes are

added

▶ A common analysis: different underlying vowels

▶ Mirrors historical source ▶ Deleting vowels were once high lax vowels (‘yers’)

▶ Grammar-based alternative: same underlying vowel

▶ Forces learners to seek markedness constraints that can help

predict when deletion occurs

▶ T

• the extent that it’s not fully predictable, must also posit

exceptionality (rules approach <100% of the time)

References 37/46

T esting generalization

Berko (1958): the “Wug” test This is a wug. Now there are two of them There are two Adult native speakers: automatically generate wugs, even if they have never heard it before Similarly for verbs: John likes to wug. Yesterday, he

Speakers automatically generate wugged

Conclusion: knowledge of past tense -ed affix generalizes beyond the collection of words that you know (walked, helped, jumped, asked, …)

References 38/46

T esting generalization

Berko (1958): the “Wug” test This is a wug. Now there are two of them There are two

▶ Adult native speakers: automatically generate wugs, even if they

have never heard it before

▶ Similarly for verbs:

John likes to wug. Yesterday, he

Speakers automatically generate wugged

Conclusion: knowledge of past tense -ed affix generalizes beyond the collection of words that you know (walked, helped, jumped, asked, …)

References 38/46

T esting generalization

Berko (1958): the “Wug” test This is a wug. Now there are two of them There are two

▶ Adult native speakers: automatically generate wugs, even if they

have never heard it before

▶ Similarly for verbs:

John likes to wug. Yesterday, he

▶ Speakers automatically generate wugged

▶ Conclusion: knowledge of past tense -ed affix generalizes beyond

the collection of words that you know (walked, helped, jumped, asked, …)

References 38/46

Gouskova and Becker (2013)

A wug test for Russian yers

▶ Examined lexicon to find phonological properties that help to

predict whether a given noun undergoes vowel alternations

▶ Of particular interest: cases where vowel deletion would create or

help repair violations of a markedness constraint

▶ E.g., [matrós] ∼ [matrósəf] (*matrs-, with CrC cluster) ▶ Quality of vowel, clusters, etc.

▶ A wug test

▶ Made up nonce words varying these properties ▶ Presented in carrier sentences, which varied inflection so as to

(potentially) trigger vowel alternations

▶ Participants rated forms with/without vowel alternations

References 39/46

Gouskova and Becker (2013)

Results

▶ Participants do generalize mid vowel alternations according to

probability that such alternations are triggered or blocked in the lexicon, by factors such as vowel quality, cluster identity, etc.

▶ Participants do not generalize mid vowel alternations according

to other features of morphemes that also correlate in various ways: gender, inflectional pattern of noun, etc.

▶ Interpretation: speakers have exactly the knowledge that they

would need if they were forced to try to distinguish alternating vs. non-alternating nouns using their grammar

References 40/46

Innovative inconsistency

▶ The fact that multiple properties line up in Greenlandic is

probably not a coincidence!

▶ Absolute neutralization is controversial as a feature of synchronic

grammars, but it’s a common sound change

▶ Comparative evidence suggests that ‘ï’ actually was a distinct

vowel in the past, but merged with [i] in Greenlandic

▶ Language may retain multiple reflexes of that earlier distinction

▶ If the synchronic analysis is in terms of exceptions to individual

processes, then it’s conceivable that a morpheme could regularize w.r.t. one process, but retain other irregularities2

▶ Result: not consistent with either reconstruction

2Though co-phonologies could help keep morphemes associated with ‘clusters’ of

properties…

References 41/46

Inconsistent innovation

Korean final obstruents: manner and laryngeal neutralization Unaffixed Suffixed (adult) Gloss

• t̚
• s’-i

‘clothes(-nom)’ pit̚ pis’-i ‘comb(-nom)’ nat̚ nacʰ-i ‘day(-nom)’ k’ot̚ k’ocʰ-ɨl ‘flower(-acc)’ pat̚ patʰ-e ‘field(-loc)’

▶ Historically, nouns ended in a variety of obstruents

▶ Including, for coronals, (/t/), /tʰ/, /c/, /cʰ/, /s/

▶ Final neutralization: [t̚] ▶ Faced with unsuffixed form, speaker must determine how the final

segment behaves when a suffix is added

▶ In terms of UR’s: is it an /s/? is it a /cʰ/? is it a /tʰ/? (etc.) ▶ In terms of grammatical processes: does it assibilate? Does it

aspirate? does it become [tʃʰ]? ▶ If learners solve in terms of UR’s, then their productions should be

consistent with their decision

References 42/46

Inconsistent innovation

In actuality…

▶ For many speakers, the morpheme that was historically /patʰ/

‘field’ preferentially uses different final consonants in different inflected forms. (Similar facts hold of other nouns, as well)3

▶ Acc. pas-ɨl (innovative s) ▶ Dat. patʰ-e (retained tʰ)

▶ It is not possible to say that these speakers have simply

reanalyzed the morpheme as /pas/, but they also don’t consistently retain /patʰ/

▶ Innovation is more consistent with grammatical regularization

▶ This morpheme now assibilates before high vowels: paʃ-i ‘nom’,

pas-ɨl ‘acc’ ▶ Of course, it’s always possible to say that the UR is reanalyzed

/pas/, and the irregular locative [patʰe] is retained

▶ This analysis no longer derives related forms from the same UR ▶ Makes no predictions about possible combinations of realizations

3Caveat: significant variability between speakers, and (especially) between dialects References 43/46

The upshot

▶ Overt data imposes conditions on how attested morphemes are

encoded, but can’t tell us directly what solution humans prefer

▶ Theoretical considerations: economy, transparency

▶ Sometimes the conditions imposed by the data can’t actually be

met

▶ No possible assumption about underlying forms that would let us

derive two related forms from the same underlying form

▶ In such cases, we’re forced to say that there’s an exception

▶ The ‘breaking point’ for when we’re forced to call something an

exception depends on how powerful/abstract we allow URs to be

▶ Hopefully, testable predictions: is something treated by speakers

as an exception?

References 44/46

References

Berko, J. (1958). The child’s acquisition of English morphology. Word 14, 150–177. Chomsky, N. and M. Halle (1968). The sound pattern of English. New York: Harper and Row. Gouskova, M. and M. Becker (2013). Nonce words show that Russian yer alternations are governed by the grammar. Natural Language and Linguistic Theory 31, 735–765. Halle, M. (1959). The Sound Pattern of Russian. The Hague: Mouton. Harms, R. T. (1973). How abstract is Nupe? Language 49, 439–446. Harrison, K. D. and A. Kaun (2000). Pattern-responsive lexicon optimization. In

• M. Hirotani, A. Coetzee, N. Hall, and J. Kim (Eds.), Proceedings of the Northeast

Linguistics Society (NELS) 30. Amherst, MA: University of Massachusetts, Amherst, Graduate Linguistic Student Association. ROA 392. Hyman, L. (1970). How concrete is phonology? Language 46, 58–76. Inkelas, S., O. Orgun, and C. Zoll (1997). The implications of lexical exceptions for the nature of grammar. In I. Roca (Ed.), Derivations and Constraints in Phonology, pp. 393–418. Oxford University Press. Kiparsky, P . (1968). How abstract is phonology? T echnical report, Bloomington.

References 45/46

References

McCarthy, J. J. (2002). A Thematic Guide to Optimality Theory. Cambridge University Press. Prince, A. and P . Smolensky (1993/2002). Optimality Theory: Constraint Interaction in Generative Grammar. T echnical report, Rutgers RuCCS-TR-2/University of Colorado, Boulder CU-CS-696-93. ROA 537, 8/2002 version. Rasin, E. and R. Katzir (2016). On evaluation metrics in Optimality Theory. Linguistic Inquiry 47, 235–282. Schane, S. A. (1974). How abstract is abstract? In A. Bruck, R. A. Fox, and M. W. La Galy (Eds.), CLS 10: Papers from the T enth Regional Meeting of the Chicago Linguistic Society (Vol. 2: Papers from the Parasession on Natural Phonology), pp. 297–317. Chicago Linguistic Society. Underhill, R. (1976). The case for an abstract segment in Greenlandic. International Journal of American Linguistics 42, 349–358.

References 46/46