![Page 1: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/1.jpg)
Using language-wide phonotactics to learn affix-specific phonology
Brian W. Smith UC Santa Cruz
1
![Page 2: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/2.jpg)
Today
• A case of phonologically-conditioned suffixation, in which…
• learners have very little data
• the distribution of affixes can be learned through a combination of…(1) extending language-wide phonotactics(2) learning small affixal differences
• No sublexicons are necessary for this data
2
![Page 3: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/3.jpg)
Case study
• Two suffixes, each with two allomorphs
-licious -alicious[lɪʃəs] [əlɪʃəs]
-thon -athon [θɑn] [əθɑn]
3
![Page 4: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/4.jpg)
Phonological conditioning
• Both suffixes conditioned by phonology
• Schwaful variant is more likely after stressed syllables, consonants
• Schwaless variant is more likely after unstressed syllables, vowels
• Shown in the following slides using data from GlOWbE (Davies et al. 2013)
• GloWbe: data from 2012-2013, 60% blogs
4
![Page 5: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/5.jpg)
Effect of segment: -(a)licious
• -alicious ∕ C_
appolicious good-a-liciouscraftilicious bookalicious
• -licious ∕ V_
roylicious bow-liciousskalicious rawlicious
5
![Page 6: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/6.jpg)
Effect of stress: -(a)licious
• -alicious ∕ σ́_
spookalicious swoon-a-liciousnomalicious meadilicious
• -licious ∕ σ̆_
dietlicious summerliciousTwilightlicious Jerseylicious
6
![Page 7: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/7.jpg)
7
Prop
ortio
n sc
hwa
V-final σ̆
C-final σ̆
C-final σ́
0.220.790.90
0.78
0.210.10
-alicious-licious
V vs. C
Rate of schwa in -(a)licious
![Page 8: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/8.jpg)
8
Prop
ortio
n sc
hwa
V-final σ̆
C-final σ̆
C-final σ́
0.220.790.90
0.78
0.210.10
-alicious-licious
σ̆ vs. σ́
Rate of schwa in -(a)licious
![Page 9: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/9.jpg)
Rate of schwa in -(a)thon
9
V vs. C
Prop
ortio
n sc
hwa
V-final σ̆
C-final σ̆
C-final σ́
0.290.79 1.00
0.71
0.21
-athon-thon
![Page 10: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/10.jpg)
Rate of schwa in -(a)thon
10
Prop
ortio
n sc
hwa
V-final σ̆
C-final σ̆
C-final σ́
0.290.79 1.00
0.71
0.21
-athon-thon
σ̆ vs. σ́
![Page 11: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/11.jpg)
Idiosyncratic differences
• Despite the fact that phonological conditioning is similar across the suffixes, the suffixes differ in their overall rate of schwa
• This difference holds across all phonological contexts
11
![Page 12: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/12.jpg)
-(a)licious and -(a)thon
12
0.00
0.25
0.50
0.75
1.00
hero hero cactus cactus police police
-alicious -licious -athon -thon
![Page 13: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/13.jpg)
Summary
• -(a)licious and -(a)thon are conditioned by phonological context in the same ways-athon C _ -thon V _ -alicious σ́ _ -licious σ̆ _
• But! -athon is used more often than -alicious across all phonological contexts
13
} }
![Page 14: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/14.jpg)
Whence phonological conditioning?
• Affixes prefer some roots over others
• Two explanations:
• Language-wide grammar
• Subcategorization
14
![Page 15: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/15.jpg)
Language-wide grammar
• Phonological conditioning comes from the phonological grammar (e.g., Mester 1994, Kager 1996, Mascaró 1996)
• In OT, markedness constraints: one set of constraints for allomorphy, alternations, and phonotactics
• E.g., choice of suffix avoids hiatus and stress clash, driven by *HIATUS and *CLASH
15
![Page 16: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/16.jpg)
Subcategorization
16
• Lexical listing / subcategorization frames (e.g. Paster 2006, Embick 2010)
• -alicious ↔ C __ -licious ↔ V __
• Sublexicons: every suffix can have its own GateKeeper grammar (Becker, earlier)
![Page 17: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/17.jpg)
Three arguments for language-wide grammar approach
• 1. Cross-suffix similarityMany suffixes are subject to the same
phonological conditions
• 2. Poverty of the stimulusBoth suffixes are very rare, with uneven distributions in a corpus
• 3. The same constraints that condition the suffixes also play a role in alternations and the distribution of words in the lexicon
17
![Page 18: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/18.jpg)
A solution
• Problem: the phonological conditioning of the suffixes persists despite a lack of learning data
• Using the pre-existing phonotactic grammar to choose between suffixal forms solves this problem — learners don’t need many -(a)licious words data to learn -(a)licious!
18
![Page 19: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/19.jpg)
Cross-suffix similarity
Experiment:-(a)licious and -(a)thon
19
![Page 20: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/20.jpg)
Experiment
• Goals
• Test for effects of segment and stress beyond words in corpus
• Estimate probabilities of -(a)licious and -(a)thon across contexts
20
![Page 21: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/21.jpg)
Experiment
• Web-based forced choice presented through Ibex Farm
21
![Page 22: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/22.jpg)
Item design
• Four stress and segmental contexts
• 10 roots of each type (plus 40 fillers)
22
EXAMPLE FINAL SEGMENT STRESS
cactus C 10police C 01hero V 10
![Page 23: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/23.jpg)
Item design
• Four stress and segmental contexts
• 10 roots of each type (plus 40 fillers)
23
Effect of final stressEXAMPLE FINAL SEGMENT STRESS
cactus C 10police C 01hero V 10
![Page 24: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/24.jpg)
• Four stress and segmental contexts
• 10 roots of each type (plus 40 fillers)
EXAMPLE FINAL SEGMENT STRESS
cactus C 10police C 01hero V 10
Item design
24
Effect of segment
![Page 25: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/25.jpg)
Experiment
• 109 participants after exclusions
• All self-identified as native speakers of English
• Only included data for American participants
25
![Page 26: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/26.jpg)
26
Prop
ortio
n sc
hwa
V-final σ̆
C-final σ̆
C-final σ́
0.070.550.930.93
0.45
0.07
-alicious-licious
V vs. C σ̆ vs. σ́
Summary for -(a)licious
![Page 27: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/27.jpg)
Summary for -(a)thon
27
V vs. C σ̆ vs. σ́Pr
opor
tion
schw
a
V-final σ̆
C-final σ̆
C-final σ́
0.010.210.71 0.99
0.79
0.29
-athon-thon
![Page 28: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/28.jpg)
-(a)licious and -(a)thon
• Comparing -(a)licious and -(a)thon we find roughly the same phonological conditioning
• We also see that schwa is used more often in -(a)thon
28
![Page 29: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/29.jpg)
-(a)licious and -(a)thon
29
0.00
0.25
0.50
0.75
1.00
hero hero cactus cactus police police
-alicious -licious -athon -thon
![Page 30: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/30.jpg)
-(a)licious and -(a)thon
• The greater preference for schwa in -(a)thon holds across items and participants
• Including fillers, which have other stress patterns, e.g. 102, 201, 010
• True for 78/80 words in the experiment
• True for 87% of participants (95/109)
30
![Page 31: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/31.jpg)
By item
31
0.00
0.25
0.50
0.75
1.00
0.00 0.25 0.50 0.75 1.00Proportion -alicious
Prop
ortio
n -a
thon
![Page 32: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/32.jpg)
32
-(a)licious -(a)thon
0.00
0.25
0.50
0.75
1.00
0.00
0.25
0.50
0.75
1.00
corpusexperim
ent
hero
cactus
police
hero
cactus
police
Stem type
Prop
ortio
n sc
hwa
Figure 9: Comparison of experimental and corpus results for -(a)liciousand -(a)thon
For both suffixes, schwa is more likely in the experiment than the corpus, but otherwise, theresults are similar.
5 Experiment part 2: the Rhythm RuleIn this section, I present the rest of the experimental results, focusing on the interaction between-(a)licious, -(a)thon, and the Rhythm Rule, a phonological alternation that resolves stress clash.Both suffix selection and the Rhythm Rule conspire to avoid stress clash, providing an additionalargument for the clash-driven nature of -(a)licious allomorphy.
These results also support a model in which UR selection and the phonological grammaroccur in parallel. They provide an example of the chicken-egg effect (McCarthy 2002), a casewhere two processes must both apply first, creating an ordering paradox.
(33) The chicken-egg effect, a consequence of parallelism (McCarthy 2002)The application of process A depends on knowing the output of process B, and theapplication of process B depends on knowing the output of process A.
The Rhythm Rule must paradoxically apply both before and after the form of the suffix ischosen.
(34) Chicken-egg effect in -(a)licious and the Rhythm Rule
27
![Page 33: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/33.jpg)
Conspiracies
33
• Experiment shows that speakers use roughly the same phonological criteria for both -(a)thon and -(a)licious
• Many suffixes in English seem subject to the same constraints
• -(a)holic, -(e)teria, -(o)rama, etc.
• And well-established derivational suffixes
• -(e)ry, -ese, -al, -eer, -ee, -ette, -ize, -ify(Raffelsiefen 2005 and earlier work)
![Page 34: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/34.jpg)
Subcat?
34
• Under subcategorization, similarity across suffixes and alternations is a coincidence -alicious ↔ C __ -licious ↔ V __ -athon ↔ C __ -thon ↔ V __ -ery ↔ C __ -ry ↔ V _
![Page 35: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/35.jpg)
Poverty of the stimulus
35
![Page 36: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/36.jpg)
Poverty of the stimulus
• Speakers agree on the phonological conditioning of the suffixes
• But if the corpus data is representative: data is scarce
• -(a)licious and -(a)thon are not very common to begin with
• Especially uncommon with roots of certain shapes
36
![Page 37: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/37.jpg)
Uncommon…
• In GlOWbE, licious- and thon-words are uncommon. Out of 500 million words for American and Canadian English (combined):
• 933 tokens for -(a)licious
• 1866 tokens for -(a)thon
• Assume a speaker hears 30,000 words/day…
• 30 licious/year, 60 thon/year
37
![Page 38: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/38.jpg)
Variety in root shape…
• For both suffixes, more than half of the types are hapaxes (182/310 for -(a)licious)
• The top 10 most-frequent words account for >50% of the tokens
38
![Page 39: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/39.jpg)
And most of the frequent -(a)thon words have schwa
39
only one with -thon
![Page 40: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/40.jpg)
For -(a)licious, the most frequent words don’t have schwa
40
BOOTY-licious
CANDY-licious
DIVA-liciousJERSEY-licious
JEW-licious SAM-licious
SLIM-a-licious
SUMMER-liciousTACO-licious
YUMMY-licious
0
25
50
75
CANDY-licious JEW-licious TACO-licious SAM-licious DIVA-licious JERSEY-liciousSUMMER-liciousYUMMY-licious SLIM-a-licious BOOTY-liciousreorder(Var1, Freq)
Freq
only one with -alicious
![Page 41: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/41.jpg)
Poverty of the stimulus
• If a speaker gets 30–60 of these tokens per year and doesn’t get a variety of phonological contexts, learning the “correct” subcategorization frames will be difficult at best
41
![Page 42: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/42.jpg)
The same constraints are active in suffixation, alternations,
phonotactics
42
![Page 43: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/43.jpg)
MaxEnt model
• A model of -(a)licious and -(a)thon with handpicked constraints
• Fit on experimental probabilities
• Note: probably not how learners acquire the distribution in the real world
• Using MaxEnt Grammar Tool (Wilson & George 2009)
43
![Page 44: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/44.jpg)
Markedness constraints
• *CLASH: Assign a violation for every σ́σ́ sequence
• *LAPSE:Assign a violation for every σ̆σ̆ sequence
• *HIATUS: Assign a violation for every V.V sequence
44
![Page 45: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/45.jpg)
Constraints to capture preference for schwa
• Analyze schwa alternation as listed allomorphs
• Constraints encode which listed allomorph is default (UR constraints, Pater at al. 2012)
• UR = /əlɪʃəs/ (-alicious)
• UR = /lɪʃəs/ (-licious)
• UR = /əθɑn/ (-athon)
• UR = /θɑn/ (-thon)
45
![Page 46: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/46.jpg)
Learned weights
46
Constraint Weight Constraint Weight
*CLASH 2.61 -athon 1.97
*HIATUS 2.31 -alicious 0.37
*LAPSE 0.43 -licious 0.18
-thon 0.13
![Page 47: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/47.jpg)
*Clash > *Hiatus > *Lapse
• Order of *CLASH, *HIATUS, and *LAPSE is mirrored in English lexicon
• On the next slide: counts from CMU dictionary, number of 3+ syllable words that violate the constraint
• Important: constraint weights were determined using only the experimental probabilities for -(a)licious and -(a)thon
47
![Page 48: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/48.jpg)
CMU violators
48
Constraint Weight Number of violators
*CLASH 2.61 1,597 8%
*HIATUS 2.31 2, 792 13%
*LAPSE 0.43 8, 702 41%
![Page 49: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/49.jpg)
Defaults
• For both -athon and -alicious, the default is the schwaful form
• The preference for -athon over -thon is greater than the preference for -alicious over -licious
49
Constraint Weight
-athon 1.97
-alicious 0.37
-licious 0.18
-thon 0.13
![Page 50: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/50.jpg)
Defaults
• Phonotactics alone can’t explain the distribution of -(a)licious or -(a)thon
• No phonotactic motivation for baseline difference in schwa rates
• Speakers then must learn from observation
• -athon is more common than -thon
• -licious is more common than -alicious
50
![Page 51: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/51.jpg)
*CLASH in English
• A small sample
• Rhythm Rule (Liberman and Prince 1977)thìrtéen → thìrteen mén
• optional that (Lee and Gibbons 2007)I know (that) Lucy went vs. I know (that) Louise went
• genitive alternation (Shih et al. 2015)the car’s wheel vs. wheel of the car
• historical change (Schlüter 2005)
51
![Page 52: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/52.jpg)
*HIATUS in English
• English avoids hiatus, especially when the left vowel is lax
• Observable in phonotactics, repairs, and allomorphy
52
![Page 53: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/53.jpg)
*HIATUS in English
• Radio Rule: no hiatus where V1 is lax √ radio, boa *ɹedɪ.o, bɔ.ə (Chomsky & Halle 1968)
• Glottal stop epenthesismora-[ʔ]ist (Plag 1999), sea [ʔ] otter (Davidson & Erker 2014)
• Intrusive R and intrusive Ldraw[r]ing [dɹɑɹɨŋ] (McCarthy 1993); draw[l]ing [dɹɔlɨŋ] (Gick 2002)
• a/an allomorphy and function word reduction an apple, a pear; [ði] apple, [ðə] pear (overview in Smith 2015)
53
![Page 54: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/54.jpg)
*HIATUS in -(a)licious
• As in the rest of English, hiatus is especially bad when the left vowel is lax
54
0.00
0.25
0.50
0.75
1.00
schwa full vowel allFinal segment
Prop
ortio
n sc
hwa
![Page 55: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/55.jpg)
*HIATUS in -(a)thon
• As in the rest of English, hiatus is especially bad when the left vowel is lax
55
0.00
0.25
0.50
0.75
1.00
schwa full vowel allFinal segment
Prop
ortio
n sc
hwa
0.00
0.25
0.50
0.75
1.00
schwa full vowel allFinal segment
Prop
ortio
n sc
hwa
![Page 56: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/56.jpg)
Learning with sparse data
56
![Page 57: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/57.jpg)
How to learn -(a)licious and -(a)thon
• The proposal:1. Take the pre-existing phonotactic grammar2. For each suffix, learn the rate of allomorphs from available data
57
![Page 58: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/58.jpg)
BLICK grammar
• Weights of phonological constraints come from BLICK (Hayes 2012)
• a MaxEnt phonotactic grammar of English based on CMU pronouncing dictionary (Weide 1994)
• constraints are a mix of hand-picked and machine-generated
58
![Page 59: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/59.jpg)
Learning the rate of schwa
• For each suffix, set the weight of morpheme-specific constraints to match the overall probability of schwa
• Here, fitted on token frequency not type
• Token frequency provides significantly better fit than type frequency for -(a)licious (difference is largely due to booty-licious)
59
![Page 60: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/60.jpg)
Model’s performance on -(a)thon
• Model captures the relative likelihood of schwa across contexts
• Overpredicts schwa in CACTUS and HERO roots
60
%schwa Target Model
POLICE-thon 99 99
CACTUS-thon 79 94
HERO-thon 48 87
SODA-thon 10 1
![Page 61: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/61.jpg)
Model’s performance on -(a)thon
61
Grammar containsa constraint against
obstruent-θ sequence
%schwa Target Model
POLICE-thon 99 99
CACTUS-thon 79 94
HERO-thon 48 87
SODA-thon 10 1
![Page 62: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/62.jpg)
Model’s performance on -(a)thon
62
%schwa Target Model
POLICE-thon 99 99
CACTUS-thon 79 94
HERO-thon 48 87
SODA-thon 10 1
Grammar doesn’tcontain a general constraint against
hiatus
![Page 63: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/63.jpg)
Model’s performance on -(a)licious
63
%schwa Target Model
POLICE-licious 93 94
CACTUS-licious 45 46
HERO-licious 9 46
SODA-licious 4 0
• Model doesn’t capture a difference between HERO and CACTUS
![Page 64: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/64.jpg)
Model’s performance on -(a)licious
64
%schwa Target Model
POLICE-licious 93 94
CACTUS-licious 45 46
HERO-licious 9 46
SODA-licious 4 0
Grammar doesn’tcontain a general constraint against
hiatus
![Page 65: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/65.jpg)
Improving the BLICK grammar
• Assume HERO+ə violates the stress-sensitive constraint against hiatus in the grammar (*V́V)
• Assume constraint against obstruent followed by θ doesn’t operate across morpheme boundaries
65
![Page 66: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/66.jpg)
With improved BLICK grammar
66
Target Model
POLICE-thon 99 99
CACTUS-thon 79 87
HERO-thon 48 51
SODA-thon 10 1
Target Model
POLICE-licious 93 94
CACTUS-licious 45 46
HERO-licious 9 12
SODA-licious 4 0
![Page 67: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/67.jpg)
Conclusion
67
![Page 68: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/68.jpg)
Take-away
• Sometimes, affix-specific phonology doesn’t require learning much affix-specific information
• In the account here, the only affix-specific information is the rate of schwa
• Using the phonotactic grammar solves sparse data problems, and also accounts for similarities between suffixes and phonotactics
68
![Page 69: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/69.jpg)
Predictions
• Under the strongest form of the language-wide grammar hypothesis, we should find a single grammar for all licious-like suffixes
• Same constraints for every one
• Same relationship between constraints
• Only differences are in default forms
69
![Page 70: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/70.jpg)
Predictions
• Any constraint that’s active in English should have an effect on -(a)licious
• Liquid OCP (which has effects in derivational morphology)
• Syllable contact
70
![Page 71: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/71.jpg)
A problem
• Phonotactics aren’t going to work for every case of affixation, e.g. English comparative -er
• How does the learner decide which approach to employ?
71
![Page 72: Using language-wide phonotactics to learn affix-specific](https://reader030.vdocuments.mx/reader030/viewer/2022012207/61dec2cbb0ada254f443ec6c/html5/thumbnails/72.jpg)
Thank you
72