jelena proki´c michael cysouwcysouw.de/home/presentations_files/cysouw_prokicichl.pdf · jelena...

String comparisonPatterns of sound changes

Conclusions

Automatic detection of patterns of sound changes

Jelena Prokic Michael Cysouw

Ludwig-Maximilians-Universitat Munchen

20th International Conference on Historical LinguisticsOsaka, July 25-30, 2011

Jelena Prokic, Michael Cysouw Automatic detection of patterns of sound changes


Conclusions

Overview

1 String comparison

2 Patterns of sound changes

3 Conclusions



Conclusions

Levenshtein distance

One of the most successful methods to determine sequence distance

(Levenshtein, 1964)

biological molecules, software engineering, ...

Levenshtein distance: minimum number of insertions, deletions and

substitutions to transform one string into the other

Syllabicity constraint add: vowels never substitute for consonants

m O @ l k @

m E l @ k

1 1 1 1

One of the most commonly used methods in quantitative language

comparison, including automatic approaches to historical linguistics



Conclusions

Positive aspects

Allows automatic alignment of the strings

Very fast and easy to implement

Gives good general picture of the relatedness between languagevarieties



Conclusions

Negative aspects

Usually based on 0/1 segment differences

Yields too little insight into the linguistic basis of differences

There is no model of linguistic change



Conclusions

Multisequence Alignment

We first need to align not only pairs of strings, but large sets.Gusfield “holy grail of string algorithms”

—no perfect solution

Softwares for automatic multisequence alignment in linguistics:Prokic et al. (2009)Johann-Mattis (2010)Steiner et al. (2011)



Conclusions

Example of multi-aligned strings

Aldomirovtsi: v "e>tS e r d - n "o l "7 s n o

Asparuhovo-Lom: v "e>tS e r d - n "o l "e s n o

Asparuhovo-Prov: vj "e>tS @ r d "7 n u lj "e s n u

Babyak: v "e>tS e r d - n "o ? ? ? ? ?

Bachkovo: v "e >ts e r d "A n u lj "e s n u


23 24 25 26 27

41.5

42.0

42.5

43.0

43.5

44.0

X

Y

0

5

10

15

20

25

Frequency of @

23 24 25 26 27

41.5

42.0

42.5

43.0

43.5

44.0

X

Y0

5

10

15

20

25

30

Frequency of A

23 24 25 26 27

41.5

42.0

42.5

43.0

43.5

44.0

X

Y

0

5

10

15

20

25

Frequency of e

23 24 25 26 27

41.5

42.0

42.5

43.0

43.5

44.0

X

Y0

5

10

15

20

25

Frequency of i

23 24 25 26 27

41.5

42.0

42.5

43.0

43.5

44.0

X

0

1

2

3

4

5

6

Frequency of m_j

23 24 25 26 2741.5

42.0

42.5

43.0

43.5

44.0

X

14

15

16

17

18

19

20

Frequency of m


Conclusions

Example of multi-aligned strings

Aldomirovtsi: v "e>tS e r d - n "o l "7 s n o

Asparuhovo-Lom: v "e>tS e r d - n "o l "e s n o

Asparuhovo-Prov: vj "e>tS @ r d "7 n u lj "e s n u

Babyak: v "e>tS e r d - n "o ? ? ? ? ?

Bachkovo: v "e >ts e r d "A n u lj "e s n u



Conclusions

Do various word positions vary at different rate?

Yes (not surprising)

Can we measure that?

Yes: Shannon index



Conclusions

Shannon index

Information entropy of the distribution

H = −i=1�

S

pi ln pi

S - number of phones, pi - relative abundance of phone i .



Conclusions

Variability under the ”Levenshtein” model

0 200 400 600 800

0.6

0.8

1.0

1.2

1.4

word positions

variability



Conclusions

Actual variability

word positions

variability

0.0

0.5

1.0

1.5



Conclusions

Variability of vowels and consonants

c mt v vs

0.0

0.5

1.0

1.5

2.0

sw in

dex



Conclusions

Variability: onset vs. nucleus vs. coda

c n o

0.0

0.5

1.0

1.5

2.0

sw in

dex



Conclusions

Computing highly correlated positions

How regular are sound changes?

Completely identical positions were not found in the data set

Word positions that show high correlation can be identified using

Mutual Information (MI)



Conclusions

Mutual Information

The mutual information of two random variables is a quantity thatmeasures the mutual dependence of the two variables.

I =�

y∈Y

�

x∈X

p(x , y) logp(x , y)

p(x)p(y)

p(x,y) is the joint probability distribution function of X and Y, and p(x)and p(y) are the marginal probability distribution functions of X and Yrespectively.


02

4

02

4


Conclusions

Computing highly correlated positions

word positions

variability

0.0

0.5

1.0

1.5


04

8

i- b_j“e bi

“7-

-tSi ni “7r

d_Zi

ib_jb_j“AZd d_j@@w

v“i j-

u- z--@ -u “7-

@j b_j“e be

“7-

-ts\@ n“e “Ar

z\@

@b_jb_j“ez\g deex

v“i j@-f -v

v@ v“o @vej b“e b“e

-A

AtSe n_je -“r_=

Ze

ebb“e:Zd d“o:“o:-

v“i -e

u- z-

-i v“o Av

04

8

mi r“7 r“ASt

“7l “7ln`kk@@v

v----“7 r_j@

@v_jv_j“A

di uv_jr_j@

“7- di u--@m_j@ r- r“A -t

“Al_j “Aln`kk@@v

v----“A r_j@

@v_jv_j“e

d_j@ uv_jr_j@

“At de @xx@me “r_=- r“ASt

“u- “u-n`kkAA-

-uunn“e re

evv“e

de ovre

“o- J\e o--A

04

8

“7- s_j@d_j“e“etid_Zd_Z“op-

k@

r_j“A “7rr-i- je

-i umm“i“ic

cZZiil_jl_j“A

@- s_j@d“e“etez\z\“ep-

k@

r_j“e “Arr-@d -i -u umm“i

“ik

kz\z\“e“elle

em se deet“ed_Zd_Z“ep-

kA

r“e -“r_=“r_=-ed -e -e -mm“i

“ik

kZZeell“e

04

8

“uwZ“e“ell-

veimm--“e

-imi @k l_j“u -“e r“7“7- @d_jd_j“A ul_j St jkk@ si

“ufz\“e“ell-

v_j“eemml_jl_j“A

him_j@ en_j l_j“u j“e r--- @d_jd_j“e ul_j s\t jcc@ se

“if Z----“7

ve“emmn_jn_jA

vume ik l_j“u j“e ---“r_= Add“e ol St -ccA se


Conclusions

Results

Each position was compared to all others

Highly correlated positions were linked into groups

Four groups of highly correlated positions were detected



Conclusions

Cluster 1: back vowels



Conclusions

Cluster 2: front vowels



Conclusions

Cluster 3: the ‘yat’ vowel



Conclusions

Cluster 4: presence of velar fricative



Conclusions

Conclusions

Occurrence of geographical patterns from the data

MI is not content dependent, but highly correlated columns have

similar phonetic content

This approach enables us to automatically identify layers of sound

changes


jelena proki´c michael cysouwcysouw.de/home/presentations_files/cysouw_prokicichl.pdf · jelena...

Documents