Predicting phonotactic difficulty in second language

acquisition

Katarzyna Dziubalska-Kołaczyk

Adam Mickiewicz University, Poznańdkasia@ifa.amu.edu.pl

Predicting phonotactic difficulty in second language

acquisition

Katarzyna Dziubalska-KołaczykGrzegorz Krynicki

Adam Mickiewicz University, Poznańdkasia@ifa.amu.edu.plkrynicki@ifa.amu.edu.pl

3

Aim of the paper

to demonstrate thatuniversal phonotactic preferences

guide the acquisition of consonant clusters in a second language

4

Empirical evidence

young learners of English (L2 English) with the following L1’s:– independent: Japanese, Korean, Vietnamese– Sino-Tibetan: Chinese– Austronesian: Kosraean, Marshallese,

Palauan, Ponapean, Samoan, Tagalog, Trukese, Visayan

– Dravidian: Tamil– Polish

5

Outline of the talk

1. Hypothesis

2. Description of the experiment

3. Introduction to B&B phonotactics

4. Phonotactic calculator

5. Analysis of the selected data

6. Preliminary conclusions

6

1. Hypothesis1. a degree of difficulty in pronouncing L2 clusters

would correlate with the universal characteristics of a given consonantal cluster

2. the more preferred a cluster, the easier and less susceptible to modifications it is expected to be

3. NAD is expected to be a universal criterion, underlying the performance of all subjects, and surpassing other relevant factors, such as the structure of the subjects’ mother tongue, their experience with English or their other capacities and motivations

degree of preference is measured by the NAD Principle

7

2. Description of the experiment

53 subjects; 15 subjects analysed here aged 11-13 native speakers of 15 various languages; 10

here recorded reading 83 times an English carrier

sentence I haven’t seen a xxx before! each time containing a different bi-syllabic nonce word

each word contained just one double or triple consonant cluster

all positions (initial, medial and final) and representative combinations were covered

8

Text for subjects

Read the following sentences aloud:

– I haven’t seen a kyati before!– I haven’t seen a shwepy before!– I haven’t seen a chluppy before!– I haven’t seen a katewt before!– I haven’t seen a petewm before!– …

9

a sound file demo

a Ponapean speaker (Micronesia)

10

3. B&B phonotactics

a universal model of phonotactics within Beats & Binding Phonology (Dziubalska-Kołaczyk 2002) – a syllable-less theory of phonology embedded in Natural Phonology

intersegmental cohesion determines syllable structure, rather than being determined by it (if one insists on the notion of the ”syllable”)

1111

B&B phonotactics the phonotactic preferences specify the universally

required distances between segments within clusters which guarantee, if respected, preservation of clusters (cf. intersegmental cohesion)

clusters, in order to survive, must be sustained by some force counteracting the overwhelming tendency to reduce towards CV's (CV preference)

this force is a perceptual contrast defined as NAD Principle (cf. Dziubalska-Kołaczyk 2002, 2003, Dressler & Dziubalska-Kołaczyk 2007, in press, Dziubalska-Kołaczyk & Krynicki 2007, Bertinetto et al. 2007)

1212

B&B phonotactics the universal preferences specify the optimal

shape of a particular cluster in a given position by referring to the

Net Auditory Distance Principle (NAD Principle)NAD = |MOA| + |POA| + |Lx|

whereby MOA, POA and LX are the absolute values of differences in the Manner of Articulation, Place of Articulation and Voicing of the neighbouring sounds respectively

13

B&B phonotactics

Example:NAD (C1,C2) ≥ NAD (C2,V)

In word-initial double clusters, the net auditory distance (NAD) between the two consonants should be greater than or equal to the net auditory distance between a vowel and a consonant neighbouring on it.

1414

Table of consonants

5laryngeal(glottal)

4radical

3dorsal

2coronal

1labial

semiVaffricate

Vapproximantsonorant stopfricativestop

sonorantobstruent01234

1515

B&B phonotactics consider the preference for initial double clusters

NAD (C1,C2) ≥ NAD (C2,V) let us now define two Net Auditory Distances

between the sounds (C1, C2) and (C2, V) whereC1 (MOA1, POA1, Lx1) C2 (MOA2, POA2, Lx2)V (MOA3, Lx3)

in terms of the following metric for (C1, C2) cluster |MOA1 - MOA2| + |POA1 - POA2| + |Lx1 - Lx2|

& |MOA2 – MOA3| + |Lx2 – Lx3|

for (C2, V) cluster

1616

B&B phonotactics

Example:in CCV in E. try

t = (4, 2, 0), r = (1, 2, 1), V = (0, 0, 1)NAD (C1, C2) = |4-1| + |2-2| + |0-1| = 3+0+1=4NAD (C2, V) = |1-0| + |1-1| = 1+0=1

thus, the preference NAD (C1,C2) ≥ NAD (C2,V)

is observed because 4 > 1

NAD Principle makes finer predictions than the ones based exclusively on sonority:

prV > trV, krV > trV, trV > drV, etc.

17

B&B phonotactics

the universal NAD Principle leads to predictions about language-specific phonotactics, its acquisition and change

specifically, it also allows to predict and explain the order of difficulty in the acquisition of second language phonotactics which appears to be universally valid and as such calls for similar remedies across languages

18

English frequent initial doubles according to NAD Principle

1919

Selected Polish clusters according to NAD Principle

Cluster types in Polish acc. to NAD

5 4 3 3 3 3 2 2 2

11 3 3 4 4

4 4 4

43 0 0

-1 -1 -2 -2 -2

-4

-2

0

2

4

6

8

10

12

pr fr lv mʂ rd fk mb ʂk sk

MOA+POA+Lx C2V NAD

20

4. Phonotactic calculator

for the purposes of B&B phonotactics, Krynicki developed the phonotactic calculator

its purpose is to enable fine-tuning and developing the theory by statistical analysis of phonetic dictionaries and phonetically annotated corpora from various languages

21

Phonotactic Calculator - requirements

various cluster lengths at all word positions formulating phonotactic hypotheses feedback on predictability of a phonotactic hypothesis choice or customization of

available phone sets, features of each phone and scores for each feature

available phonetic dictionaries and languages (PolSynt, Festvox, Festival)

metrics used for calculating distances between phones (taxicab, euclidean)

accepted phonetic alphabets (IPA, SAMPA)

22

23

24

25

5. Analysis of the selected data

a total of 1245 utterancesproduced by 15 childreneach reading 83 sentences containing a

nonsense word with a 2- or 3-consonant cluster

in 767 of these utterances (61,6%) the speakers modified or avoided the cluster that was assumed to be the correct pronunciation of the nonsense word

26

Error types

error description

number of errors in the corpus symbol

vowel insertion between the elements of a consonant cluster or at the end of a cluster that was expected to be pronunced word-finally 234 @

reducing the number of consonants in the cluster (from 3 to 2 or 1 and from 2 to 1) 218 $

unintelligible pronunciation 154 ?

substitution of consonant in a cluster by consonants not present in the expected cluster 152 #

substantial mispronunciations 119 %

pause insertion between the elements of a consonant cluster or at the end of a cluster that was expected to be pronunced word-finally 24 .

deletion of the cluster 4 ∅

omission of the word 2 omitted

total 907  

27

Summary statistics for six preferences

preference number

number of cluster that apply to a given preference

number of clusters that follow the preference percentage

1 17 17 100%

2 13 13 100%

3 38 27 71%

4 5 3 60%

5 5 3 60%

6 5 2 40%

28

Part 1 of the hypothesisA degree of difficulty in pronouncing L2 clusters correlates with the

universal characteristics of a given consonantal cluster.

To a certain degree the amount of correlation between the number of errors students make when producing a cluster and the NAD parameters between the components of that cluster can be illustrated by means of cluster ranking in terms of their NAD differences and their difficulty.

Ranking of clusters can be performed first with respect to the NAD criterion and then with respect to linearly scaled percentage of clusters in which speakers made errors.

Although statistically not significant, the trend line indicates the expected direction of change and degree of slope between difficulty and NAD measure for finals.

correlation for final double clusters

y = -0,0302x - 2,5577R 2 = 0,007

y = -0,978x + 17,385R 2 = 0,92

-10

-5

0

5

10

15

20

a w

m

a w

θ

a w

l

a j n

a w

t

a w

t͡�

ʃ

a l k

a m

t

a r t͡

�ʃ

a m

p

a l s

a r p

a r f

NA D(V C ) - NA D(C C )Inc orrec tL iniowy (NA D(V C ) - NA D(C C ))L iniowy (Inc orrec t)

31

Part 2 of the hypothesis:Linear Regression

The more preferred a cluster, the easier and less susceptible to modifications it is.

The error of complex mispronunciation annotated in the corpus involved combination of other various errors, epenthesis, substitution, metathesis and other.

There is a significant correlation between the NAD differences in a word-medial cluster and the frequency of the complex mispronunciation errors made in it by the speakers (P-value in the ANOVA = 0,0282; R-squared = 11,2148%).

mispronunciation & medial clusters

32

Plot of Fitted Model

-9 -6 -3 0 3 6

NAD_VC minus NAD_CC

0

1

2

3

4

5

6

mis

pron

Part 2 of the hypothesis:Analysis of variance and median

• NAD(VC) - NAD(CC) turns out to have statistically significant influence on the number of reduction errors students made in word-final clusters

0 1

Means and 95,0 Percent LSD Intervals

VC minus CC le mdn[VC minus CC]

0

1

2

3

4

5

Red

uctio

n

34

reduction & word-final clusters

ANOVA F=11.86, p=0.006Kruskal-Wallis T=7,46, p=0,006

Difference↗

Preference↘

35

Part 3 of the hypothesis

NAD is expected to be a universal criterion, underlying the performance of all subjects.

If a child produces a consonant cluster different from the expected one, this new cluster will usually follow phonotactic preferences (grand mean of 79.7% compared to 78.3% for expected clusters).

all initials medials finals

The number of all expected consonant clusters 83

78,3%

2290,9

%

4367,4

%

1888,9

%The number of expected consonant clusters that followed phonotactic preferences 65 20 29 16

Total number of elicited consonant clusters 158

79,7%

2253,6

%

4360,1

%

1858,3

%Total number of elicited consonant clusters that followed phonotactic preferences 126 12 26 11

The number of cases when there was a 0 in the expected cluster but more than 0 in the elicited cluster 10 0 0 0

The number of cases when both the expected and the elicited cluster were a 1 37 9 20 8

The number of cases when both the expected and the elicited cluster were a 0 7 0 7 0

The number of clusters for which no speaker produced an qualifiable utterance. 22 10 14 8

This suggests that phonotactic preferences underlie the performance of the subjects of various linguistic backgrounds and may be universal.

More research is necessary to show whether the speakers of different languages displayed significant differences in their following of the preferences.

38

6. Preliminary conclusions

universal phonotactic preferences guide speakers in producing SL clusters

the scale of preference in the acquisition of a given type of cluster allows for fine-tuning of SL learning/teaching materials

many aspects of the analysis remain to be continued– comparison with the L1’s of the subjects– data from further subjects– detailed analysis of the errors: which types of

improvements are preferred

• f j a h l a tT ʃ j a tT ʃ w a k j a k r a p l a p w a ʃ w a k m a m j a t l a tT ʃ l a s r a l j a m w a t n a

• f j a h l a tT ʃ w a p w a ʃ w a t n a k j a k m a tT ʃ j a tT ʃ l a m w a t l a s r a l j a m j a p l a k r a

correlation for initial double and triple clusters

41

correlation for initial double clusters

y = -0,2586x + 5,5919R 2 = 0,9229

y = -0,1152x + 10,743R 2 = 0,0399

-6

-1

4

9

14

19

f j a

h l a

t͡�

ʃ j a

t͡�

ʃ w

a

k j a

k r a

p l a

p w

a

ʃ w

a

k m

a

m j

a

t l a

t͡�

ʃ l a

s r a

l j a

m w

a

t n a

NA D(C C ) - NAD(C V )Inc orrec tL iniowy (NAD(C C ) - NA D(C V ))L iniowy (Inc orrec t)

correlation for medial double clusters

y = -0,1754x + 4,2098R 2 = 0,917

y = -0,0928x + 9,2304R 2 = 0,0864

-6

-1

4

9

14

19

v t t

vv

t tS

vv

tS tS

vv

p f v

v t S

vv

s s

vv

tS t

vv

p tS

vv

S tS

vv

p s

vv

m m

vv

p k

vv

S f

vv

f p v

v s

t vv

f tS

vv

n l v

v s

n v

v l l

vv

m n

vv

r l v

v s

k v

v s

p v

v w

w v

v tS

m v

v m

r v

v tS

h v

v m

f v

v r n

vv

j l v

v l m

vv

n t v

v j w

vv

m tS

vv

l s v

v n

h v

v m

k v

v r h

v

NA D(V C ) - NAD(C C ) Inc orrec t

L iniowy (NAD(V C ) - NA D(C C )) L iniowy (Inc orrec t)