Upload
vuonghanh
View
216
Download
0
Embed Size (px)
Citation preview
How Much Homophony Is Normal?!
Abby KaplanUniversity of California, Santa Cruz
CUNY Conference on the Word in PhonologyJanuary 14, 2010
1 Introduction: Contrast Maintenance in Phonology
Phonological patterns sensitive to the need to maintain contrast
– Dispersion Theory (Flemming 2002): constraints that require contrast amongphonemic categories oppose constraints against expending articulatory e!ort
– Phonological patterns sometimes serve to maximize contrasts among potentiallexical items (Padgett 2003; Padgett and Tabain 2005)
– Some patterns seem especially prone to avoiding neutralization (e.g., lenition,Gurevich (2004))
This sensitivity usually modeled at level of potential lexical items, not actual ones
– Don’t want to predict unattested phonological patterns:
Dispersed inventories only for lexical items in minimal pairs
Rules that apply only where they don’t create homophones
“These questions arise when we take the domain of explanation to bethe set of actual lexical items in a language. But this is in fact notthe practice in generative phonology. Instead, theories model the set ofpossible words of a language....” (Padgett 2003, 78-79)
– Could there be other types of homophony avoidance?
!Many thanks to Aaron Kaplan, Yongeun Lee, Jaye Padgett, Dan Silverman, Paul Willis, and membersof Phlunch at UC Santa Cruz. All shortcomings are my own. This research was funded by an NSF GraduateResearch Fellowship.
1
Silverman (to appear) argues that phonological patterns avoid creating homophonesamong actual words
– Korean has many neutralizing phonological alternations
– But these alternations result in only a relative handful of homophones
– Homophony avoidance in lexical statistics rather than formalization of phonolog-ical pattern
Remaining question: how can we be sure number of homophones in Korean is reallyunexpectedly low?
" Perhaps any set of neutralizations would create very few homophones
To establish a ‘baseline’ level of expected homophony, count homophones produced byalternative phonological rules and compare to actual rule; at least two ways to do this:
1. Hand-select a few alternative rules for comparison
– Silverman’s method; alternatives produce more homophony than actual rules
– Pro: can select phonologically plausible rules that are similar to actual rules
– Con: results highly dependent on which alternatives we happen to pick
2. Compute homophony for large number of alternative rules (‘brute-force’ method)
– Method used for this talk
– Pro: cover a lot more ground than hand-selection approach
– Con: hard to filter out implausible rules
2 Method
2.1 Korean
Why Korean?
– Large number of neutralizing phonological alternations
(1) a. (i) /nAÙ-i/ # [nAÙi] ‘day.nom’(ii) /nAÙ-k’wA/ # [nAtk’wA] ‘day and’
b. (i) /nAÙh-i/ # [nAÙhi] ‘face.nom’(ii) /nAÙh-k’wA/ # [nAtk’wA] ‘face and’
– Writing system roughly morphophonemic: orthography can be used to approxi-mate underlying forms
2
Tab
le1:
Surf
ace
real
izat
ions
ofunder
lyin
gco
da-
onse
tse
quen
ces
inK
orea
n
ph
pp’
tht
t’s
s’Ù
hÙ
Ù’
kh
kk’
mn
lh
$p
ph
p’
pth
pt’
ps’
pÙh
pÙ’
pkh
pk’
pm
pn
pl
ph
pp
hp
hh
ph
ps
ps’
hps’
lplp
hlp
lph
lphh
lph
h
ths’
Ùh
kh
tntl
hh
$t
tht
thth
hth
st’
Ù’sh
ss’
s’h
s’Ù
k’Ùh
ÙÙ
hÙh
hÙh
k
kph
kp’
kth
kt’
ks’
kÙh
kÙ’
kmkn
kl
khk
k’
k’k
hkh
hkh
ks
ksh
ks’
lklk
hlk
mm
ph
mp
mp’
mth
mt
mt’
ms
ms’
mÙh
mÙ
mÙ’
mkh
mk
mk’
mm
mn
ml
mn
np
hnp
np’
nth
ntnt
’ns
ns’
nÙh
nÙ
nÙ’
nkh
nk
nk’
nm
nn
nl
nnh
nhh
nÙ
nÙh
nÙ
NNp
hNp
Np’
Nth
NtNt
’Ns
Ns’
NÙh
NÙNÙ
’Nk
hNk
Nk’
NmNn
NlN
lm
lph
lp
lp’
lth
lt
lt’
ls
ls’
lÙh
lÙ
lÙ’
lkh
lk
lk’
lmln
ll
lml
llh
lhh
lth
lthh
lth
lsls
hls
$p
hp
p’
tht
t’s
s’Ùh
ÙÙ’
khk
k’m
nl
$
3
Table 1 shows surface realizations of underlying coda-onset sequences
– Reflect nine ordered rules from descriptions in Sohn (1994): resyllabification, [h]-aspiration, coda neutralization, sibilation, tensification, consonant cluster simpli-fication, decoronization, [h]-weakening, reduction
– Some other rules suppressed:
Non-neutralizing alternations (e.g., intervocalic voicing)
Morphologically conditioned alternations (e.g., lateralization)
Data on Korean lexicon
– Data from Korean National Database (Lee 2006) (uninflected stems in Koreanorthography; no morpheme boundaries)
– Phonological rules applied via Java scripts (collaboration w/ Paul Willis)
– Rules applied only stem-internally: application to codas may be bled by vowel-initial su"xes, of which there are many (Albright and Kang 2009)
2.2 Procedure
Same procedure for each rule; illustrated here with data for overall neutralizationpattern
2.2.1 Step 1: Count Homophones Created by Rule
Measures of homophony
Homophones Number of words in lexicon w/ at least one homophone
Weighted Homophones Sum of frequencies of words w/ at least one homophone
Homophone Pairs Number of pairs of homophones in lexicon
Homophone Sets Number of (maximal) sets of words in lexicon that all neutralizeto the same thing
Table 2 shows levels of homophony underlyingly and after all rules have applied
" ‘New’: number of homophones added by rules
For results of simulations, only ‘homophones’ and ‘weighted homophones’ shown
" ‘Homophone pairs’ and ‘homophone sets’ very similar to ‘homophones’
4
Table 2: Homophony in Korean lexicon
Measure Underlying Surface NewHomophones 6201 6646 445Weighted Homophones 291681 319835 28154Homophone Pairs 4308 4692 384Homophone Sets 2794 2975 181
2.2.2 Step 2: Simulate Alternative Rules
Three simulations; each creates 1000 random patterns w/ same number of neutraliza-tions as actual rule
‘a’-series simulation: neutralize random sets of coda-onset sequences
– Table 3 illustrates this for overall neutralization pattern
– Randomly distribute onset-coda sequences in table; superimpose grid from actualpattern
– All sequences in same box neutralize with each other
– (Some cells neutralize w/ nothing – left blank; sequences here don’t matter)
– Figure 1 shows results for overall pattern
Histograms show level of homophony across simulated patterns
Vertical lines show actual number of homophones
Title gives percentile rank of actual number of homophones among simulatedpatterns: smaller percentile # actual level of homophony surprisingly low
" For both measures, every simulated pattern resulted in more homophony thanactual pattern
5
Tab
le3:
Exa
mple
ofneu
tral
izat
ion
pat
tern
sin
an‘a
’-se
ries
sim
ula
tion
ÙÙ’
hÙ
np’
lkh
lspp
tp’
Ùllh
khth
Ù’s’
s’th
s’n
Ùph
lphth
lsth
k’p
tht’
lpht
mÙh
Ùhl
ls’
nhh
psn
nÙs
ksp
kth
lkp
hÙh
msl
tht’
k’kh
mn
Ùkh
lmp’
ksp’
Ùhk’
mm
pk’
kss’
nkh
lmÙ
k’s’
llks
khpsÙ
’p
h$
sth
lhp
hÙt
’k’
Ù’Ùn
tÙh
$Ùh
pk
pkh
nÙl
NÙh
$tkp
hn
Ùp’
hk’
mh
nl
lkn
s’m
khp
hNh
lkm
Nthp’
php
hÙp
kÙh
nt’
kspss
’hs’
ls$
psk
hlp
Ùhkh
s’pl
hs
lk$
s’th
lthk’
NÙ’
lkÙ
ms’
Nkh
lÙh
lp$
lphs
Np’
lmth
pp
hs’
Ùm
k’s’
p’
lhs’
kst’
lthl
ÙhÙh
lth$
lkÙh
k’t
ps$
lpp
$$k’
p’
lkh
lsn
phm
$s’
lhÙh
lkkh
s’k
lÙm
ts’
llp
nm$
lss’
kst
s’t’
h$
ssnm
kÙ’
khp’
lsÙh
pÙ’
k’$
lskh
pst
nhp
lpp’
mt’
Ù$lm
ktn
spsm
lphÙh
khÙ
lphl
hts’$
lt’
k’p’
lmt’
$sÙh
Ùpsk
phÙ’
ksÙ
php
kkkp
phk
khÙh
Ùht’
ss’
t$psk
’ls
Ùlt
hk
lht’
ks$
nht
h$Ù
lmÙh
sph
Nnlm
nhs
tpp
hth
$k’
nn
hh
ths
$Ù’
nÙk
’lp
t’psp
lhlm
hlt
hp’
nÙ’
np
k’m
mk
mp
hm
psh
lms
$nkh
nÙh
lpl
kk’
kh$
Ùhk
lpÙ
k’k
ksk’
lphkh
nh
Ùls
lp
hn
nÙs
’ht
hth$
lhs
hÙh
kht’
lkth
nÙÙ
’lp
hÙ’
k’th
phh
sknh
Ùhnhk’
Ùspt’
nhp
hÙh$
lthÙ’
thl
tklp
sNt
’th
nk
$kh
lsk
thth
nk’
lth
ksth
lthÙ
s’t
ÙÙh
lks
tÙ’
nÙk
hlh
ltÙ
ÙÙ$l
nÙp
Ùknh
k’k’
lpt
lsp
kÙk’
sÙh
sls
k’Ùk
’s’
ph
sÙpsm
thh
ktsp
’n
Ùklp
hn
ps
np
hth
khn
Ùtst
’nt
hth
pÙt
Nllp
hk’
ksÙ’
Ùp’
lthkh
ÙhÙ’
lmp
lsÙ
s’kh
$kÙh
p’
lpm
ksm
nh$
ksn
s’n
$th
lhm
khs
Ns’
lss
k’Ùh
lhÙ’
nhl
lphk
$p’
nhk
hp
ksh
nhn
lkÙ’
ph
lsm
psp
hlm$
lthh
thk’
$hls
p’
lpth
$mlm
tnp
Ùhp
s’Ù’
nhkh
sk’
n$
nÙm
Nph
ms
lhp’
s’Ùh
phÙh
hk
lpn
ÙÙh
lthm
nthn
ml
thm
$ph
s’p
lths’
pÙ
pht’
Ùmm
Ù’pn
ltht’
lp’
hÙ’
lh$
$plk
p’
nÙ$
hkh
khk
skh
ps’
tt’
lst’
psl
stk’$
tmlp
s’kh
th
nht
lpÙ’
thp
hlp
hs’
phÙ
lmp
hkh
tlp
ph
k’l
s’k’
tkh
tk’
Nth
mp
hpsp
’kh
khls
ph
ltts
mth
khl
nÙt
’lm
s’Nk
phs’
ns’
$t’
lhn
khÙ’
Ùhkh
lphÙ
lph
snÙh
s’nhm
k’p
hlk
llk
t’Ùh
hnh
Ù’s’
sk’
hnhs’
mÙ
khk’
ltht
thÙ
pst
hp
hkh
php’
lphm
lml
pp’
nÙ
hp
hkk
hks
klm
Ù’tl
Ùs’
pt
p$
k’Ù
pst
’kh
pn
ÙÙp
hk’
lphp’
nÙh
NÙs$
Nk’
lhh
Ùhp
hts
’kl
thÙh
lk’
pss
Ùht
ns
nÙt
hht
tph
lph$
psÙ
hs’
hkh
hkm
ksÙh
lnth
nkn
lths
thp’
lmm
lhk
Ùhn
lph
pm
kss
lphp
sÙ’
lkt
lthÙh
ksl
knlh
thpsÙ
lthth
lphh
lthn
6
Figure 1: Amount of homophony in simulation a
Percentile: 0
# of Homophones
# of
Pat
tern
s
500 1000 1500 2000
015
030
0Percentile: 0
# of Weighted Homophones
# of
Pat
tern
s
30000 50000 70000 90000
010
025
0‘b’-series simulation: randomly shu#e onsets, codas separately
– Table 4 illustrates for overall pattern
– Neutralized sequences more similar than in ‘a’-series: more phonologically plau-sible
– Figure 2 shows results
" Almost every simulated pattern results in more homophony than actual pat-tern
Figure 2: Amount of homophony in simulation b
Percentile: 0
# of Homophones
# of
Pat
tern
s
1000 3000 5000
010
025
0
Percentile: 0.4
# of Weighted Homophones
# of
Pat
tern
s
40000 80000 120000
010
020
0
7
Tab
le4:
Exa
mple
ofneu
tral
izat
ion
pat
tern
sin
a‘b
’-se
ries
sim
ula
tion
ps’
l$
Ùh
ph
nk
Ù’
t’k
hs
k’
ht
p’
Ùth
m
ppp
ps’
pl
p$
pÙh
pp
hpn
pk
pÙ’
pt’
pkh
ps
pk’
ph
pt
pp’
pÙ
pth
pm
nnp
ns’
nl
n$
nÙh
np
hnn
nk
nÙ’
nt’
nkh
ns
nk’
nh
ntnp’
nÙ
nth
nm
lklk
plk
s’lk
llk$
lkÙh
lkp
hlk
nlk
klk
Ù’lk
t’lk
khlk
slk
k’lk
hlk
tlk
p’
lkÙ
lkth
lkm
ph
php
phs’
phl
ph$
phÙh
php
hp
hn
phk
phÙ’
pht’
phkh
phs
phk’
phh
pht
php’
phÙ
phth
phm
lmlm
plm
s’lm
llm$
lmÙh
lmp
hlm
nlm
klm
Ù’lm
t’lm
khlm
slm
k’lm
hlm
tlm
p’
lmÙ
lmth
lmm
mm
pm
s’m
lm$
mÙh
mp
hm
nm
km
Ù’m
t’m
khm
sm
k’m
hm
tm
p’
mÙ
mth
mm
NNp
Ns’
NlN$
NÙh
Nph
NnNk
NÙ’
Nt’
Nkh
NsNk
’Nh
NtNp
’NÙ
Nth
Nmk
kpks
’kl
k$kÙ
hkp
hkn
kkkÙ
’kt
’kk
hks
kk’
khkt
kp’
kÙkt
hkm
hhp
hs’
hl
h$
hÙh
hp
hhn
hk
hÙ’
ht’
hkh
hs
hk’
hh
hthp’
hÙ
hth
hm
ks
ksp
kss’
ksl
ks$
ksÙh
ksp
hks
nks
kks
Ù’ks
t’ks
khks
sks
k’ks
hks
tks
p’
ksÙ
ksth
ksm
lhlh
plh
s’lh
llh$
lhÙh
lhp
hlh
nlh
klh
Ù’lh
t’lh
khlh
slh
k’lh
hlh
tlh
p’
lhÙ
lhth
lhm
kh
khp
khs’
khl
kh$
khÙh
khp
hkh
nkh
kkh
Ù’kh
t’kh
khkh
skh
k’kh
hkh
tkh
p’
khÙ
khth
khm
Ùh
Ùhp
Ùhs’
Ùhl
Ùh$
ÙhÙh
Ùhp
hÙh
nÙh
kÙh
Ù’Ùh
t’Ùh
khÙh
sÙh
k’Ùh
hÙh
tÙh
p’
ÙhÙ
Ùhth
Ùhm
ssp
ss’
sls$
sÙh
sph
snsk
sÙ’
st’
skh
sssk
’sh
stsp
’sÙ
sth
smt
tpts
’tl
t$tÙ
htp
htn
tktÙ
’tt
’tk
hts
tk’
thtt
tp’
tÙtt
htm
nÙ
nÙp
nÙs
’n
Ùln
Ù$n
ÙÙh
nÙp
hn
Ùnn
Ùkn
ÙÙ’
nÙt
’n
Ùkh
nÙs
nÙk
’n
Ùhn
Ùtln
Ùp’
nÙÙ
nÙt
hn
Ùmlp
hlp
hp
lphs’
lphl
lph$
lphÙh
lphp
hlp
hn
lphk
lphÙ’
lpht’
lphkh
lphs
lphk’
lphh
lpht
lphp’
lphÙ
lphth
lphm
lth
lthp
lths’
lthl
lth$
lthÙh
lthp
hlt
hn
lthk
lthÙ’
ltht’
lthkh
lths
lthk’
lthh
ltht
lthp’
lthÙ
lthth
lthm
llp
ls’
lll$
lÙh
lph
lnlk
lÙ’
lt’
lkh
lslk
’lh
ltlp
’lÙ
lth
lm$
$p$s
’$l
$$$Ù
h$p
h$n
$k$Ù
’$t
’$k
h$s
$k’
$h$t
$p’
$Ù$t
h$m
ps
psp
pss
’psl
ps$
psÙ
hpsp
hpsn
psk
psÙ
’pst
’psk
hpss
psk
’psh
pst
psp
’psÙ
pst
hpsm
lsls
pls
s’ls
lls$
lsÙh
lsp
hls
nls
kls
Ù’ls
t’ls
khls
sls
k’ls
hls
tls
p’
lsÙ
lsth
lsm
s’s’
ps’
s’s’
ls’$
s’Ùh
s’p
hs’
ns’
ks’
Ù’s’
t’s’
khs’
ss’
k’s’
hs’
ts’
p’
spÙ
s’th
s’m
k’
k’p
k’s’
k’l
k’$
k’Ùh
k’p
hk’
nk’
kk’
Ù’k’
t’k’
khk’
sk’
k’k’
hk’
tk’
p’
k’Ù
k’th
k’m
thth
pth
s’th
lth$
thÙh
thp
hth
nth
kth
Ù’th
t’th
khth
sth
k’th
hth
tth
p’
thÙ
thth
thm
ÙÙp
Ùs’
ÙlÙ$
ÙÙh
Ùph
ÙnÙk
ÙÙ’
Ùt’
Ùth
ÙsÙk
’Ùh
ÙtÙp
’ÙÙ
Ùth
Ùmlp
lpp
lps’
lpl
lp$
lpÙh
lpp
hlp
nlp
kkp
Ù’lp
t’lp
thlp
slp
k’lp
hlp
tlp
p’
lpÙ
lpth
lpm
nh
nhp
nhs’
nhl
nh$
nh
Ùhnhp
hnhn
nhk
nh
Ù’nht
’nht
hnhs
nhk’
nhh
nht
nhp’
nh
Ùnht
hnhm
8
‘c’-series simulation: randomly shu#e segments; combine according to neutralization‘template’ for rule
– Template: specification of which coda-onset sequences neutralize with each other
– Example from tensification rule: for each ordered pair in set B, first memberpreceded by anything from set A neutralizes with second member preceded bysame thing
" So, neutralizing pairs include {k+k, k+k’}, {t+k, t+k’}, {k+t, k+t’}, etc.
Figure 3: Template for tensification
{/A + B1/, /A + B2/}
A = {k, t, p, s, kt, nt, lk, lp, lt, pt, ks,ns, ls, ps}
B = {%k, k’&, %t, t’&, %p, p’&, %s, s’&,%Ù, Ù’&, %sh, s’h&}
– In ‘c’-series simulation: randomly assign new segments to sets A and B (and Cand so on, if template has more sets); restrictions:
Identical segments across sets stay that way (i.e., every [k] in figure 3 becomes[s] in figure 4)
Segments that appear as codas in template must be possible codas, those asonsets must be possible onsets (i.e., every member of set A must be a possiblecoda)
– Example given in figure 4
" New neutralizing pairs include {s+s, s+kh}, {n+s, n+kh}, {s+n, s+s’h}
Figure 4: Example of new sets for tensification template in a ‘c’-series simulation
A = {s, n, t, k, ns, $, N, p, m, ps, ls, lt, pt, lk}
B = {%s, kh&, %n, s’h&, %t, khh&, %k, th&, %t, Nh&, %mh, phh&}
In theory, ‘b’-series simulations phonologically more plausible than ‘a’-series, ‘c’-seriesmore plausible than ‘b’-series
9
3 Results
For each rule, I give:
– Description of rule
– Template for rule
– Homophony produced by rule
– Results of three simulations
Display of simulation results
– Two graphs: homophones measure on left, weighted homophones on right
– Density curves (' smoothed histograms) show distribution of homophony in 1000patterns run in each simulation
– Three curves, one for each simulation (‘a’, ‘b’, ‘c’)
– Vertical bar: actual level of homophony
– Legend notes percentile rank of actual level in each simulation
Note:
– Rules apply in ordered fashion, each rule to output of previous rule
– Possible onsets and codas at each stage di!er accordingly
– ‘Total’ homophones listed for each rule are total homophones in lexicon after ruleapplies
10
3.1 Rule 1: Resyllabification
Resyllabify single coda consonant into following onsetless syllable or syllable with initial[h] (Sohn 1994, 164)
Almost all simulations produce far less homophony than actual rule; most producenone at all!
Small additional peak in histograms for ‘c’-series near actual pattern
" Suggests a single pair of neutralized coda-onset sequences responsible for most ofhomophony produced by resyllabification
Figure 5: Homophony afterresyllabification
Measure Total NewHomophones 6330 129Weighted 295832 4151Pairs 4417 109Sets 2847 53
Figure 6: Template for resyllabification
{/A + B/, /B + A/}
{/C1 + B/, /C2 + C3/}
A = {k, n, t, l, m, p, s, Ù, Ùh, kh,th, ph, h, k’, s’}
B = {$}
C = {%ks, k, s&, %nÙ, n, Ù&, %nh, n,h&, %lk, l, k&, %lm, l, m&, %lp, l, p&,%ls, l, s&, %lth, l, th&, %lph, l, ph&,%lh, l, h&, %ps, p, s&}
Figure 7: Amount of homophony in simulation series 1
0 50 100 150 200
0.00
00.
010
0.02
00.
030
# of Homophones
Dens
ity
a (99)b (99.4)c (99.2)
0 2000 6000 10000
0e+0
04e−0
48e−0
4
# of Weighted Homophones
Dens
ity
a (97.2)b (99.1)c (98.6)
11
3.2 Rule 2: [h]-Aspiration
Fuse plain noncontinuant obstruent with adjacent [h] into homorganic aspirated ob-struent (Sohn 1994, 166)
In ‘a’-series, most simulations yield less homophony than actual pattern: over half of‘a’-series simulations yield none
In ‘b’- and ‘c’-series, actual level of homophony in lower half of simulations
Note that the more phonologically plausible the simulation, the more homophony ittends to produce (c > b > a)
Figure 8: Homophony after[h]-aspiration
Measure Total NewHomophones 6340 10Weighted 296460 628Pairs 4425 8Sets 2851 4
Figure 9: Template for [h]-aspiration
{/A1 + B/, /B + A1/,/C + A2/, /C + A3/}
{/D + A2/, /D + A3/}
A = {%k, kh, kh&, %t, th, th&, %p,ph, ph&, %Ù, Ùh, Ùh&}
B = {h}
C = {$}
D = {k, n, t, l, m, p, s, N, Ù, Ùh,kh, th, ph, k’, s’}
Figure 10: Amount of homophony in simulation series 2
0 200 400 600 800
0.00
00.
004
0.00
8
# of Homophones
Dens
ity
a (86.5)b (36.2)c (22.3)
0 10000 20000 30000
0.00
000
0.00
015
# of Weighted Homophones
Dens
ity
a (90.4)b (51.4)c (36.6)
12
3.3 Rule 3: Coda Neutralization
Syllable-final obstruents become plain stops; /h/, /s/ # [t] (Sohn 1994, 165)
Simulated patterns yield more homophony than actual pattern; less concentrated atlow end of scale
Exception: weighted homophones measure for ‘a’-series
Note that percentile rank of actual level of homophony is generally greater for weightedhomophones measure
Figure 11: Homophony aftercoda neutralization
Measure Total NewHomophones 6378 38Weighted 305133 8673Pairs 4475 50Sets 2863 12
Figure 12: Template for coda neutralization
{/A1 + B/, /A2 + B/,/A3 + B}
{/C1 + B/, /C2 + B/,/C3 + B/, /C4 + B/,/C5 + B/, /C6 + B/,
/C7 + B/}
{/D1 + B/, /D2 + B/}
A = {%k, k’, kh&, %ls, lth, lh&}
B = {k, n, t, l, m, p, s, N, $, Ù,Ùh, kh, th, ph, k’, t’, p’, s’, Ù’,nh, lh, mh, sh, Ùhh, khh, thh,phh, hh, k’h, s’h}
C = {%t, s, Ù, Ùh, th, h, s’&}
D = {%p, ph&, %nÙ, nh&, %lp,lph&}
Figure 13: Amount of homophony in simulation series 3
0 1000 2000 3000 4000
0e+0
04e−0
48e−0
4
# of Homophones
Dens
ity
a (4.5)b (11)c (1.4)
−20000 20000 60000 100000
0.00
000
0.00
006
0.00
012
# of Weighted Homophones
Dens
ity
a (83.3)b (27)c (7.4)
13
3.4 Rule 4: Sibilation
/t/ becomes [s] before [s] or [s’] (Sohn 1994, 165)
Since all /s/s became [t]s in Coda Neutralization, this rule is non-neutralizing
3.5 Rule 5: Tensification
Plain obstruents become tense after obstruents (Sohn 1994, 173)
Most simulated patterns produce more homophony than actual rule
Again, more plausible simulations tend to yield more homophony (c > b > a)
Figure 14: Homophony af-ter tensification
Measure Total NewHomophones 6382 4Weighted 305140 7Pairs 4477 2Sets 2865 2
Figure 15: Template for tensification
{/A + B1/, /A + B2/}
A = {k, t, p, s, kt, nt, lk, lp, lt,pt, ks, ns, ls, ps}
B = {%k, k’&, %t, t’&, %p, p’&, %s, s’&,%Ù, Ù’&, %sh, s’h&}
Figure 16: Amount of homophony in simulation series 5
0 500 1000 1500
0.00
000.
0015
0.00
30
# of Homophones
Dens
ity
a (21.8)b (7.4)c (3.2)
−10000 10000 30000 50000
0e+0
04e−0
58e−0
5
# of Weighted Homophones
Dens
ity
a (15.1)b (5.2)c (2.4)
14
3.6 Rule 6: Consonant Cluster Simplification
Delete one consonant from all coda clusters (Sohn 1994, 170)
– If first consonant in cluster is /l/ and second is non-coronal stop, delete /l/ (thissimplifies the facts somewhat)
– Otherwise, delete second consonant
Most simulated patterns produce far more homophony than actual rule
Again, note that percentile rank for actual rule is greater for weighted homophonesmeasure
Figure 17: Homophony af-ter consonant cluster sim-plification
Measure Total NewHomophones 6399 17Weighted 308910 3770Pairs 4499 22Sets 2871 6
Figure 18: Template for simplification
{/A1 + B/, /A2 + B/,/A3 + B/, /A4 + B/}
{/C1 + B/, /C2 + B/,/C3 + B/}
A = {%k, kt, lk, ks&, %l, lm, lt, ls&,%p, lp, pt, ps&}
B = {k, n, t, l, m, p, s, N, $, Ù, Ùh,kh, th, ph, h, k’, t’, p’, s’, Ù’, nh,lh, mh, Nh, sh, Ùhh, khh, thh, phh,hh, k’h, s’h}
C = {%n, nt, ns&}
Figure 19: Amount of homophony in simulation series 6
0 1000 2000 3000
0e+0
04e−0
48e−0
4
# of Homophones
Dens
ity
a (0)b (3.9)c (1.1)
0 20000 60000
0e+0
04e−0
58e−0
5
# of Weighted Homophones
Dens
ity
a (22)b (17.1)c (14.2)
15
3.7 Rule 7: Decoronization
/t/ assimilates in place to a following stop (Sohn 1994, 175)
Decoronization creates practically no homophony (just one pair of words)
Seems to be a lower limit: no simulation produced no homophones
Figure 20: Homophony af-ter decoronization
Measure Total NewHomophones 6401 2Weighted 308923 13Pairs 4500 1Sets 2872 1
Figure 21: Template for decoronization
{/A1 + B/, /A2 + B/}
{/A1 + C/, /A3 + C/}
A = {%t, k, p&}
B = {k, N, kh, k’, Nh, khh, k’h}
C = {m, p, ph, p’, mh, phh}
Figure 22: Amount of homophony in simulation series 7
0 100 300 500
0.00
00.
010
0.02
0
# of Homophones
Dens
ity
a (0)b (0)c (0)
0 5000 10000 15000
0e+0
04e−0
48e−0
4
# of Weighted Homophones
Dens
ity
a (0)b (0)c (0)
16
3.8 Rule 8: [h]-Weakening
Delete /h/ between sonorants
This rule creates more homophones than any other
Actual level of homophony always in top half of simulations
Recurrence of two patterns
– Phonologically more plausible simulations yield more homophony
– Percentile rank of actual homophony greater for weighted homophones measure
Figure 23: Homophony af-ter [h]-weakening
Measure Total NewHomophones 6628 227Weighted 319453 10530Pairs 4677 177Sets 2967 95
Figure 24: Template for [h]-weakening
{/A + C1/, /A + C2/}
{/B + C1/, /B + C2/}
{/B + D1/, /B + D2/}
A = {k, t, p, s, n, l, m, N}
B = {$}
C = {%n, nh&, %N, Nh&, %l, lh&, %m,mh&}
D = {%$, h&}
Figure 25: Amount of homophony in simulation series 8
0 500 1000
0.00
00.
004
0.00
8
# of Homophones
Dens
ity
a (98.4)b (77.5)c (53.7)
−10000 0 10000 30000
0.00
000
0.00
015
0.00
030
# of Weighted Homophones
Dens
ity
a (99.1)b (88.2)c (72.2)
17
3.9 Rule 9: Pre-Tense/Aspirate Reduction
Delete plain obstruents before homorganic tense or aspirated obstruents (Sohn 1994,175)
Actual level of homophony seems to fall in about the middle of simulated patterns
Many simulations create no homophony
Again, note greater percentile rank for actual level for weighted homophones measure
Figure 26: Homophony af-ter pre-tense/aspirate re-duction
Measure Total NewHomophones 6646 18Weighted 319835 400Pairs 4692 15Sets 2975 8
Figure 27: Template for pre-tense/aspirate reduction
{/A + B/, /C + B/}
{/D + E/, /C + E/}
{/F + G/, /C + G/}
A = {k}
B = {kh, k’, khh, k’h}
C = {$}
D = {t, s}
E = {Ùh, th, t’, s’, Ù’, Ùhh, thh,s’h}
F = {p}
G = {ph, p’, phh}
Figure 28: Amount of homophony in simulation series 9
0 200 400 600 800
0.00
00.
004
0.00
80.
012
# of Homophones
Dens
ity
a (40.4)b (46)c (43.9)
−5000 0 5000 15000
0e+0
02e−0
44e−0
4
# of Weighted Homophones
Dens
ity
a (76.2)b (67.2)c (67.8)
18
4 Discussion
4.1 Does Korean Have Low Homophony?
Rules 2, 3, 5, 6, and 7 produce less homophony than simulated patterns, as does overallpattern
But rules 1, 8, and 9 produce more
To assess overall results across simulations: linear mixed-e!ects model predicting per-centile rank of actual level of homophony for each rule/measure
– Predictors: measure of homophony, simulation series
– Random e!ect of rule
Two models: one for raw percentiles, one for arcsin-transformed percentiles
" Similar results; only raw percentiles presented here
Results
– Intercept: 40
" Below 50: suggests rules yield less homophony than the median, but notsignificantly di!erent from 50
– Weighted homophones measure associated with significantly greater percentilesthan other measures (p = .011)
– Percentile rank for ‘b’-series simulations significantly less than for ‘a’-series (p =.0014), for ‘c’-series significantly less than for ‘b’-series (p = .0000014)
Tentative conclusion: actual level of homophony is indeed low
– Low intercept in model
– More rules yield less homophony than their simulations than yield more
– Overall pattern has less homophony than simulations
4.2 Possible Mechanisms of Homophony Avoidance
As discussed above, we don’t want to build homophony avoidance (as opposed toneutralization avoidance) into formal phonological patterns
These results suggest only a gradient avoidance of homophony
How might this situation come about? Possibilities:
19
1. Given phonetic precursor to a rule, the less homophony the rule would create, themore likely the precursor is to be phonologized
2. Rules tend to neutralize contrasts that are already perceptually suboptimal; lexi-con is already optimized to avoid homophones based on hard-to-perceive contrasts
3. Words in dense neighborhoods tend to resist alternation (Ussishkin and Wedel2009)
4.3 Other Patterns in the Data
The more phonologically plausible the rule, the more homophony it creates
– If true, this trend might argue against explanation 2 above
– Caveat: most simulated rules still not very plausible
– Future research: how to better filter rules for phonological plausibility?
Level of actual homophony looks less surprisingly low when homophones are weightedby frequency
– In other words, few words are homophonized, but they tend to be especiallyfrequent
– Possible explanation: short words more likely both to be homophonized and tobe frequent
" Unlikely: should be just as true of simulated patterns as of actual patterns
– Looks like an anti-functional tendency
5 Conclusion
Neutralizing alternations in Korean appear to produce less homophony than expected
Thus, phonological rules may be sensitive to contrast among actual words, not justpotential ones
References
Adam Albright and Yoonjung Kang. Predicting innovative alternations in Korean verbparadigms. In Current Issues in Unity and Diversity of Languages: Collection of thePapers Selected from the CIL 18, Held at Korea University in Seoul, pages 1–20, Seoul,Korea, 2009. Linguistic Society of Korea.
Edward S. Flemming. Auditory Representations in Phonology. Outstanding Dissertations inLinguistics. Routledge, New York, NY, 2002.
20
Naomi Gurevich. Lenition and Contrast: The Functional Consequences of Certain Phono-logically Conditioned Sound Changes. Outstanding Dissertations in Linguistics. GarlandPublishing, New York, NY, 2004.
Yongeun Lee. Sub-Syllabic Constituency in Korean and English. PhD thesis, NorthwesternUniversity, 2006.
Jaye Padgett. Contrast and post-velar fronting in Russian. Natural Language and LinguisticTheory, 21(1):39–87, 2003.
Jaye Padgett and Marija Tabain. Adaptive dispersion theory and phonological vowel reduc-tion in Russian. Phonetica, 62:14–54, 2005.
Daniel Silverman. Neutralzation and anti-homophony in Korean. Journal of Linguistics, toappear.
Ho-Min Sohn. Korean. Descriptive Grammars. Routledge, New York, NY, 1994.
Adam Ussishkin and Andrew Wedel. Lexical access, e!ective contrast, and patterns in thelexicon. Ms., University of Arizona, 2009.
21