Adler
NELS 35 poster, October 2004
Faithfulness and Perception in Loanword Adaptation:
A Case Study from Hawaiian
Allison Adler, MIT
anadler@mit.edu
1. Why study loans from English into Hawaiian?
o We can examine the treatment of sounds borrowed into a language with a uniquely
limited consonant inventory:
 No coronal stops: *t, d
 No oral fricatives: *f, v, s, z, , 
Table 1: Consonant inventory of Hawaiian (and English)
labial
coronal
dorsal
stop
p (b)
(t, d)
k (g)
fricative
(f, v)
(s, z, , )
nasal
m
n
glide
w
liquid
laryngeal

h
()
l (r)
o Hawaiian also has restricted syllable structure: (C)V(V) only
o So… how are non-native segments adapted? How are consonant clusters and codas
adapted?
2. Methods & Data
o Preliminary corpus study: approx. 250 lexicalized loans from English into Hawaiian,
taken from dictionary (Pukui & Elbert, 1979)
o Elicited adaptations: non-lexicalized borrowings elicited from two native EnglishHawaiian bilinguals (n=200 and n=100)
3. Selected findings of this study
o Positional variation: Consonants adjacent to a vowel or sonorant are overwhelmingly
preserved – i.e., there is frequent epenthesis to save these segments; other consonants
variably deleted
o Unattested featural adaptations:
 changes in sonority, nasality: */b/m, */t/n, l
 change in major articulator: */t,d/p, *//m
o Attested/frequent featural adaptations: devoicing, change in place of articulation
 blow
[polóu]
 colby
[kolopí:]
 tinker [kiníkə]
 delve
[ke:lewə]
 green
[kəlínə]
Adler
NELS 35 poster, October 2004
o Fricatives variably map to [h] or nearest stop (modulo place of articulation):

variation in singleton onset and coda fricatives
C  stop
Ch
crease
kəlíki
crease
kəlíki
shark
ká:like
pressure
pəlé:hu
fork
póka
proof
pulúhu
half
hápə
floor
holólə
o Strident fricatives follow this pattern but can also exhibit deletion in non-salient
positions:


variation in /sC/ clusters: s  k, h, Ø
speech
kəpíke ~ həpíke
despot
kepáe
story
kokóli ~ kóli
dispatch
kipá:ki
stymie
həkáimi
beast
píkə
sloop
kəlúpe ~ həlúpe
grasp
kələhépe
non-prevocalic stridents (n=89):
s  k: 37%
s  h: 29%
s  Ø: 34%
4. Existing approaches to loanword adaptation
4.1 Phonological: Theory of Constraints and Repair Strategies: Paradis (1988, et seq.)
 Relies on standard feature geometry to motivate phonological closeness as the basis
for segmental adaptations
 Principles of Minimality, Preservation, Threshold in loan adaptation
4.2 Perceptual: P-map: Steriade (2001); also Fleischhacker (2000), Kang (2002),
Kenstowicz (2001, 2003)
 Hypothesis: adaptations in loanwords are generally perceptually-based; highly
perceptible changes or deletions are avoided. Can capture positional asymmetries due
to differential cue distribution by context
 cf. perceptual similarity/confusability of Steriade (2001) and perceptual cue studies
(e.g., Steriade (1999), Côté (2000), Fleischhacker (2000)) and early work on speech
perception.
4.3 Articulatory: Perceptual Assimilation Model: Best (1994, 1995, et seq.)
 Articulatory phonology motivates a model of non-native speech perception: assumes
that listeners access gestural information in speech stream and categorize non-native
sounds based on articulatory similarity to native phonological categories
Adler
NELS 35 poster, October 2004
5. How do existing approaches fare on this data?
5.1 Adaptation of coronal stops:
5.1.1 Phonological approach (TCRS):
Problem: Counting feature changes yields the wrong result – not all
changes are perceptually equal: predict t  k or p, d  n
Table 2: Predicted adaptation of coronal stops under TCRS
Constraint: *[cor], [-son] Predicted repair: t  k, p; d  n
Adaptation
voiceless /t/
Cp
1. delink coronal
2. insert labial
Ck
1. delink coronal
2. insert dorsal
Cn
1. insert [+son]
2. insert [+nasal]
3. delink [-voice];
insert [+voice]
voiced /d/
1. delink coronal
2. insert labial
3. delink [+voice];
insert [-voice]
1. delink coronal
2. insert dorsal
3. delink [+voice];
insert [-voice]
1. insert [+son]
2. insert [+nasal]
5.1.2 Perceptual approach (P-map):
Predictions based on cue distribution by context are upheld, but
perceptual information alone is not definitive: t  k or p



Correctly excludes mapping to sonorants [n, l]: change in sonority is highly
perceptible and therefore disfavored
CAN account for positionally-based deletion based on reduced perceptibility
of place of articulation of word-final, unreleased stops.
Problem: acoustic/perceptual info does not conclusively choose between [p, k]
as possible outputs: perception studies of stop recognition and confusability
are not able to single out any cue from release burst or formant transitions as
constant across contexts
5.1.3 Perceptual Assimilation Model:
Correctly selects as optimal the mapping which changes place of
articulation but not major articulator: t k



Based on “tube geometry” structure (see Diagram 1 below), the most minimal
change is from coronal to dorsal, as observed in the data
Problem: tube geometry not conclusive on minimal change for all segments
Beyond the oral node, minimal featural change unclear: what determines that
changing the position of the glottis (voiced vs. unvoiced) is more minimal
than changing the position of the velum (oral vs. nasal)?
Adler
NELS 35 poster, October 2004
Diagram 1: Tube geometric representation of the oral cavity
Oral Cavity
Lips
Tongue
(labial)
Central
Tip
(coronal)
Lateral
Body
(dorsal)
5.2 Adaptation of fricatives:
5.2.1 TCRS:
Variation predicted for /s/ – but not the observed mappings
Table 3: Predicted adaptation of /s/ under TCRS
Constraint: *[+strident]
Predicted repair: s  l? otherwise, variation
sp
Delink:
1. +strident
2. +continuant
3. Coronal
sk
Delink:
1. +strident
2. +continuant
3. Coronal
Insert
4. Labial
Insert
4. Dorsal



sn
sl
Delink:
Delink:
1. +strident
1. +strident
2. +continuant
Insert
3. +sonorant
4. +nasal
(5. +voice)
Insert
2. +sonorant
3. +lateral
(4. +voice)
sh
Delink:
1. Oral node
2. +strident
3. +consonantal
Insert
4. +sonorant
Preservation Principle favors adaptation with the least featural deletion has
taken place – but these mappings are unattested (s  n, l)
Mapping to [k] and [p] equal: Coronal, Dorsal, Labial are all equidistant in the
feature geometry
Another problem: subjective counting of feature changes – i.e., what counts as
a “step” in the repair process?

“Default insertion” of features not always counted: depending on
what counts as a step, prediction changes

How to characterize “debuccalization” to [h]?
5.2.2 Perceptual approach:
Incorrectly predicts retention of all stridents because of their high
inherent salience

Otherwise, positional variation is expected due to weakness of cues in wordfinal or non-prevocalic position.
Adler
NELS 35 poster, October 2004
5.2.3 PAM:
Predictions unclear for unfamiliar stricture type; also, system not designed
to account for segments in context



PAM predicts that fricatives will be treated as speech-like even if they cannot
be assimilated to a single native category
BUT it is not clear which features will be targeted in attempting to map an
unfamiliar stricture type to a native sound.
Problem: system describes context-free categorization of sounds, thus is not
designed to handle positionally-based variation of strident fricatives.
6. Combining articulatory and perceptual similarity: A comprehensive account
Perceptual and articulatory information are necessary and
complementary factors in forming the judgments of similarity which
allow borrowers to select a mapping for non-native segments
6.1 Mapping of coronals /t,d/  k
 Articulatory information determines the closest mapping for the coronal stops even
though perceptual information is not conclusive. Given the options available in
Hawaiian, the optimal mapping preserves the major articulator by realizing /t/ as a
lingual [k] rather than a labial [p]
 Cue-based, perceptual information informs judgments of similarity, predicting
positional asymmetries in deletion vs. epenthesis
6.2 Mapping and deletion of stridents: /s/  k, h, Ø
 These findings do not actually call into question the high salience of stridents. A
trade-off exists between the pressure to retain salient input segments (like stridents)
and the need to change their features:
 Deletion is a viable option only because no native segment adequately captures
exactly those features which make these segments highly salient. Thus, the feature
loss/change is as dispreferred as segment loss
 Deletion is not viable for non-strident fricatives because, in the absence of the
[+strident] feature, the available native options are closer (albeit still imperfect)
mappings
 The cross-linguistics dispreference for deletion is a real phenomenon but it derives its
power from the fact that languages are frequently able to map non-native segments
without changing their most salient features. Deletion is a viable option only because
no native segment adequately captures the salient features of stridents.
 The chosen mappings for input fricatives reflects a trade-off between retaining the
presence of a stricture in the oral tract and retaining their salient [+continuant]
feature. These two features are incompatible in the native Hawaiian inventory (recall
that [h] is the only fricative in Hawaiian)
 An OT-based approach of violable constraints is well-suited to capturing the
variation observed
Adler
NELS 35 poster, October 2004
6.2 The constraint system and selected tableaux
6.2.1 Adaptation of coronal stops
 Ident[±sonorant]: The output correspondent of input [sonorant] is [sonorant]
 Ident[articulator]: The output correspondent of input [articulator] is
[articulator]
 Result: segments are mapped to a sound of equal sonority and same major
articulator
Tableau 1: Adaptation of coronal stops
a. Ident[±son] >> Ident[place]: map to [-sonorant]
/t/
Ident[±son] *t
a. t
b. n
c. l
d. ? p
e. ? k
Ident[place]
*!
*!
*!
*
*
b. Ident[artic], Ident[place]: map to [k], not [p]
/t/ *t
a. t
b. h
c. 
d. p
e.  k
Ident[artic]
Ident[place]
*!
*!
*!
*
*
*
*
*!
6.2.2 Adaptation of [+strident] fricatives
 Max-C[+strident]: A strident consonant in the input has a correspondent in the
output
 Ident[±strident]: The output correspondent of input [strident] is [strident]
 Ranking: Ident[±stri], MaxC[+stri]
 Result: Deletion and strident feature change are equally (dis)favored
Tableau 2: Adaptation of [+strident] fricatives
*s
Ident[±stri] MaxC[+stri]
/s/
a. s
b. p
c.  h
d.  k
e.  Ø
*!
*
*
*
*
Adler
NELS 35 poster, October 2004
6.2.3 Adaptation of fricatives generally
 Max-fricative: A fricative consonant in the input has a correspondent in the
output
 Ident[±continuant]: The output correspondent of input [continuant] is
[continuant]
 Ranking: Ident[articulator], Ident[±continuant], Max-fricative
 Result: there is variation in preserving [+continuant] and the presence/location
of a stricture in the oral tract
Tableau 3: Adaptation of all fricatives
a. deletion of stridents is not ruled out
*s
Ident[artic]
Ident[±cont] Max-fric
*
*
*!
/s/
a. s
b. p
c.  h
d.  k
e.  Ø
*!
*
*
b. deletion of nonstridents not optimal: Max/Ident[strident] irrelevant
*f
Max
C[-stri]
Ident[artic] Ident[±cont] Maxfric
/f/
a. f
b. Ø
c. k
d.  h
e.  p
*!
*!
*
*
*
*!
*
6.2.4 Positional/Contextual asymmetry
 Max-C/V: A consonant followed by a vowel in the input has a correspondent in
the output
 Max-C//R, N: A consonant adjacent to a sonorant in the input has a
correspondent in the output
 Ranking: Max-C/V, Max-C//R, N >> Ident
 Result: salient, well-cued consonants are retained despite the need for feature
change