How much does Geography Influence LanguageVariation?

John Nerbonne

Center for Language and Cognition, University of Groningen

Language, Space and Geography III:26-28 Nov. 2009, Freiburg

John Nerbonne 1/32

Overview

Introduction to dialectometryLevenshtein distance

Examining the influence of geographyUnderstanding Seguy’s curve (distance)Incorporating areas

John Nerbonne 2/32

Why dialectometry?

Solve problems of earlier dialectologyNon-overlapping distributionsSelection of features too arbitrairy“Atomism” (Coseriu), idiosyncratic words (Bloomfield)

Introducing replicable proceduresFollowing Seguy, Goebl, Schiltz, Kretzschmar, Shackleton, ...

Aggregating over categorical variables (same vs. different)

Seeking law-like relations in linguistic variation

John Nerbonne 3/32

Beyond categorical differences: Levenshtein distance

Levenshtein distance enables analysis of phonetic transcriptionswithout manual alignment

—move from categorical to numerical analysis of data.One of the most successful methods to determine sequencedistance (Levenshtein, 1964)

biological molecules, software engineering, ...

Levenshtein distance: the minimum number of insertions,deletions and substitutions to transform one string into the other

John Nerbonne 4/32

Example of the Levenshtein distance

mO@lk@ delete @ 1mOlk@ subst. O/E 1mElk@ delete @ 1mElk insert @ 1mEl@k

4

m O @ l k @m E l @ k

1 1 1 1

John Nerbonne 5/32

Calculating dialect distances

To determine the aggregate distance between dialects:We determine the distance between each dialect pair for everysingle word (in sample, e.g. dialect atlas)We sum these distances for every word (hundreds of them) andcompare them

Besides dialect distances, this also yields interesting soundcorrespondences contained in the alignments (more on thatelsewhere)

Note that a 100-word comparison already yields about 500 soundcorrespondences

John Nerbonne 6/32

Example

Based on Dutch pronunciation data from theGoeman-Taeldeman-Van Reenen-Project data (GTRP; Goemanand Taeldeman, 1996)

We use 562 words for 424 varieties in the Netherlands

Wieling, Heeringa & Nerbonne (2007) An Aggregate Analysis ofPronunciation in the Goeman-Taeldeman-van Reenen-ProjectData. In: Taal en Tongval 59(1), 84-116

John Nerbonne 7/32

Distribution of sites

John Nerbonne 8/32

Analytical steps

Obtain the distances between each of the ≈ 90, 000 pairs ofvarieties

n.b. this involves 500× 52 segment comparisons≈ 1.1× 109 segment comparisons in total

Organize these in a 400× 400 tableSeek groups (dialect areas) or continuum-like relations, e.g. byapplying clustering or multi-dimensional scaling, respectively

John Nerbonne 9/32

MDS map

John Nerbonne 10/32

Large body of work using Levenshteindistance—positive aspects

Dutch, German, American English, Norwegian, Swedish,Afrikaans, Sardinian, Tuscan, Sino-Tibetan, Chinese, Bulgarian,Bantu, Central Asian (Turkic & Indo-Iranian), ...Development of consistency measure (Cronbach’s α) indictingwhether data set is sufficiently largeNovel reflection, work on validation aimed at assessing degree ofdetection of SIGNALS OF PROVENANCE

Gooskens & Heeringa (2004) Perceptive Evaluation of LevenshteinDialect Distance Measurements using Norwegian Dialect Data.Language Variation and Change 16(3), 189-207.

John Nerbonne 11/32

The Influence of Geography

Regression designDependent variable: varietal distance, as measured by aggregatecategorical distance or Levenshtein distanceIndependent variable: geographical distance, regarded as anoperationalization of the chance of social contactStatistical caution: correlations involving averages are muchhigher than correlations involving individual items

True, but we’re interested in the properties of entire varieties(dialects), not just in their individual features

John Nerbonne 12/32

Inspiration: Jean Seguy

Seguy (1971) La relation entre la distance spatiale et la distancelexicale. Revue de Linguistique Romane 35(138), 335-357:Aggregate variation increases sublinearly with respect togeography

COURSE MOYENNE

Y = 36Vlog(x + 11

so

.0

J

10

1

~ 1. 6 . I) IS 10 1~ 30 3~ .0 .~ 50 55 60 ~ 10 1S 10 is 90 95 100 IDS 110

John Nerbonne 13/32

Sublinear spread is general

0 100 300 500

0.00

0.10

0.20

Bantu

0 100 200 300 400 500

0.00

00.

002

0.00

4

Bulgaria

0 200 400 600 800

0.04

0.08

0.12

Germany

0 200 600 1000

0.0

0.2

0.4

LAMSAS / Lowman

0 50 100 200 300

0.01

0.03

0.05

0.07

The Netherlands

0 100 200 300 400 500

1.0

2.0

3.0

4.0

Norway

John Nerbonne 14/32

Refuting Trudgill’s “Gravity hypothesis”

Moon

DeimosPhobos

Venus

Earth

Mars

Sun

According to Trudgill (1972) diffusion follows an inverse square

law, with the consequence that linguistic distance should likewise

increase with the square of the distance. Population size plays

the role of mass.

John Nerbonne 15/32

Trudgill’s “Gravity hypothesis”

Sublinear aggregate relation incompatible with a quadraticinfluence

J.Nerbonne (to appear) Measuring the Diffusion of Linguistic Change.Phil. Transactions of the Royal Society B

John Nerbonne 16/32

How much does geography influence languagevariation?

Area Corr.(l,geo) r2

Gabon Bantu 0.47 0.22Bulgaria 0.49 0.24Germany 0.57 0.32Eastern U.S. 0.51 0.26Netherlands 0.62 0.38Norway 0.41 0.16

But note that linguistic distance in Norway correlates more highly with19th century travel times (r = 0.54, r2 = 0.29)Gooskens (2005) Traveling time as a predictor of linguistic distance.Dialectologia et Geolinguistica 13.

John Nerbonne 17/32

Geography and language variation?

Geography accounts for 22− 38% of aggregate linguistic variation.General — sublinear — characterization of relation betweengeographical distance and linguistic differencesLike population geneticists’ “isolation by distance” (Wright, 1943;Malecot, 1955)

John Nerbonne 18/32

Wrede’s (1926-56) German Dialect Areas

Aachen

Adorf

Ahrbergen

Albersloh

AllnaAltenberg

Altentrüdingen

Altlandsberg

Altwarp

Astfeld Atzendorf

BardenflethBarßel

Bempflingen

Bennin

Benzingerode

Billingsbach

Bockelwitz

BonnBorstendorf

Breddin

Brelingen

Bremscheid

Brockhausen

Bruch

Burgrieden

Büdesheim

Clausthal

Cursdorf

Dexbach

Diekhusen

Dreeke

Ebertshausen

Ebstorf

Eichenhofen

Einöllen

Eisenbach

Endenburg−Lehnacker

Engelsbach

Engter

Ensheim

Eschelbronn

Eversen

Exdorf

Falkenthal

Freden

Frohnhausen

Gelting

Gerbstedt

Gersdorf

Gornsdorf

Gottsbüren

Gresenhorst

Groß Ballhausen

Groß Jamno

Groß Mohrdorf

Großwechsungen

Großenwiehe

Grünlichtenberg

Grünow

Götz

Haddorf

Hagen

Hammah

Hartenfels

Hartenstein

HelmersHerbornseelbach

Herdecke

Hermannsreuth

Hermeskeil

Herrentrup

Herßum

Herzfeld

Hesselteich

Hohenkörben

Hohwacht

Holmkjer

Horheim

Huddestorf

Hundisburg

Höchstädt

Hönebach

Iggingen

Igling

Iversheim

Jacobsdorf

Jeetzel

Jesteburg

Jever

Jonsdorf

Jürgenshagen

Kemmern

Kirch Kogel

Klafferstraß

Klein Rossau

Kleinbottwar

Kruft

Kuhstedt

Köln

LampertswaldeLandgrafroda

Langenbruck

LeuthLinz

Lippen

Lippramsdorf

Lohra

Lohrhaupten

Maibrunn

Mastershausen

Maxweiler

Mirow

Mockern

Mörshausen

Neu Schadow

Niederasphe

Niederauerbach

Niedernhausen

Nürnberg

Oberau

Oberhomberg

Obermaiselstein

Oberviehbach

Oberwürzbach

OdenspielOffdilln

Ohrdorf

Oiste

Orferode

Osterbruch

Osterfeld

Osterhagen

Ottowind

Peterskirchen

Pfaffenrot

Pielenhofen

Pretzsch

Pöttmes

Rammenau

Ramsau

Reelkirchen

Reetz

Ringgau−Röhrda

Rosenthal

Roßwein

Ruhlsdorf

Schachach

Schauenstein

Schellroda

Schmannewitz

Schnepfenbach

Schraden

Schwabstedt

Schömberg

Schönbeck

Seelbach

Siebenbach

Stöcken

Tann

Theuma

Treffelstein

Ulbering

Ungstein

Unterellen

Valdorf

Verchen

Vielbrunn

Vietmannsdorf

Volkershausen

Wangerooge

Warpe

Wasbek

Wasbüttel

Wehrsdorf

Weidenbach

Wemb

Wentorf

Weyhers

Wieda

Wildpoldsried

Winterspelt

Wittelsberg

Wolgast

Wüllen

Zellingen

John Nerbonne 19/32

Influence of Dialect Areas?

We add to the regression design variables indicating whether twovarieties belong to the same or different dialect areas.For example var56-dff has the value 1 if the two sitescompared differ in that one is in area 5 (Bavaria) and the other inarea 6 (Southwest)Do dialect areas contribute to the explanation of linguisticdifferences?

(Ongoing work.)

John Nerbonne 20/32

Influence of Dialect Areas?

Preliminary results: Dialect areas contribute substantially to theexplanation of aggregate linguistic distance. r2 increases fromabout 32% (based only on geographic distance) to about 45%(based on geographic distance and areal differencs).

(Ongoing work.)

John Nerbonne 20/32

How much does geography influence languagevariation?

Pure distance models explain 22% - 38% of aggregate linguisticvariation.Areal distinctions are somewhat collinear, but nonethless addsubstantially to simple models, perhaps as much as 50% (moving30% to 45%, for example).Naturally, there is also subdialectal variation (social, sexual,individual), but few systematic data collections.Emerging questions:

What is the linguistic structure of the dialect differences we find?Do typological constraints play a (confounding) role?Can we tease apart geographical and historical explanations, andhow?

John Nerbonne 21/32

Questions?

Thank You!

John Nerbonne 22/32

Criticisms of Levenshtein-based work

Binary segment distances too roughFrequent concern in Groningen (Heeringa Diss., 2004)

Segment distances based on phonetic features, phonologicalfeatures, canonical spectrogramsHigh correlation with rough measures when compared at aggregate(varietal) levelBut no substantial improvement in aggregate distance measures(validation wrt dialect speakers’ judgments)compare height measurements in in., cm, mm, µm

Difficult problem in general — due inter alia to fine detail inatlases, e.g., 1,300 different vowels in LAMSASNew procedure (Jelena Prokic) induces segment weights fromdata

John Nerbonne 23/32

Inducing segment distances

Sound correspondences were obtained using the Levenshteinalgorithm using a Pointwise Mutual Information procedure(Wieling et al., 2009; included in RuG/L04)

Levenshtein algorithm:

l E I k @ nl i k h 8 n

1 1 1 1

Segment distances based on Pointwise Mutual Information:

PMI(x , y) = log2

(p(x , y)

p(x) p(y)

)

John Nerbonne 24/32

Evaluating segment weight induction

Evaluation with respect to alignment correctness—more sensitive than aggregate correlations with judgments

50% less error using alignments with induced weightsCompetitive with sophisticated bio-informatic techniques from,(pair Hidden Markov Models)Future project: evaluate the segment weights against linguisticcriteria, compare weights induced from different data sets

Martijn Wieling, Jelena Prokic, & John Nerbonne “Evaluating thepairwise string alignments of pronunciations” LaTeCH-SHELT&R,2009.

John Nerbonne 25/32

Gabon Bantu

0 100 300 500

0.00

0.10

0.20

Bantu

Area: BantuData: 53 sites, 160 wordsSource: Van der Veen, LyonNote: Late settlement

John Nerbonne 26/32

Bulgaria

0 100 200 300 400 500

0.00

00.

002

0.00

4

Bulgaria

Area: BulgariaData: 482 sites, 54 wordsSource: Stoykov’s atlasNote: Long Turkish domination

John Nerbonne 27/32

Germany

0 200 400 600 800

0.04

0.08

0.12

Germany

Area: GermanyData: 186 sites, 201 wordsSource: Kleiner DeutscherLautatlas (Goschel)

John Nerbonne 28/32

LAMSAS / Lowman

0 200 600 1000

0.0

0.2

0.4

LAMSAS / Lowman

Area: Eastern Seaboard, USData: 357 sites, 145 wordsSource: Mid & South Atlantic,LAMSAS (Lowman)Note: Settlement in last 400 yr.

John Nerbonne 29/32

The Netherlands

0 50 100 200 300

0.01

0.03

0.05

0.07

The Netherlands

Area: The NetherlandsData: 424 sites, 562 wordsSource: Goeman-Taeldeman-van Reenen Atlas

John Nerbonne 30/32

Norway

0 100 200 300 400 500

1.0

2.0

3.0

4.0

Norway

Area: NorwayData: North Wind & Sun15 sites, 58 wordsSource:www.ling.hf.ntnu.no/nos

John Nerbonne 31/32

References

Inderjit Dhillon. 2001. Co-clustering documents and words using bipartite spectral graph partitioning. Proceedings of theseventh ACM SIGKDD international conference on Knowledge discovery and data mining, pp. 269–274. ACM New York.

Ton Goeman, and Johan Taeldeman. 1996. Fonologie en morfologie van de Nederlandse dialecten. Een nieuwemateriaalverzameling en twee nieuwe atlasprojecten. Taal en Tongval, 48:38–59.

Vladimir Levenshtein. 1965. Binary codes capable of correcting deletions, insertions and reversals. Doklady AdademiiNauk SSSR, 164:845–848.

Wilbert Heeringa. 2004. Measuring Dialect Pronunciation Differences using Levenshtein Distance. Ph.D. thesis,Rijksuniversiteit Groningen.

Martijn Wieling and John Nerbonne. 2009. Bipartite spectral graph partitioning to co-cluster varieties and soundcorrespondences in dialectology. In: Monojit Choudhury (ed.) Proceedings of the TextGraphs-4 Workshop at the 47thMeeting of the Association for Computational Linguistics, August 2009, Singapore. Available viahttp://www.martijnwieling.nl.

Martijn Wieling, Jelena Prokic, and John Nerbonne. 2009. Evaluating the pairwise string alignment of pronunciations. In:Lars Borin and Piroska Lendvai (eds.) Language Technology and Resources for Cultural Heritage, Social Sciences,Humanities, and Education (LaTeCH - SHELT&R 2009) Workshop at the 12th Meeting of the European Chapter of theAssociation for Computational Linguistics. Athens, 30 March 2009, pp. 26-34. Available via http://www.martijnwieling.nl.

John Nerbonne 32/32