WissenschaftskollegBerlin - uni-tuebingen.de

The phylogeny of word meaningsInferring the directionality of semantic change from word lists

Gerhard Jägerjoint work with Alla Münch and Johannes Dellert

Seminar für Sprachwissenschaft, Tübingen

Wissenschaftskolleg Berlin

January 26, 2016

Gerhard Jäger (Tübingen) Phylogeny of word meanings 1/26/2016 1 / 65

Evolution and language change

“The formation of different languages and of distinctspecies, and the proofs that both have been developedthrough a gradual process, are curiously parallel. [...] Wefind in distinct languages striking homologies due tocommunity of descent, and analogies due to a similarprocess of formation. The manner in which certainletters or sounds change when others change is very likecorrelated growth. [...] The frequent presence ofrudiments, both in languages and in species, is still moreremarkable. [...]Languages, like organic beings, can be classed in groupsunder groups; and they can be classed either naturallyaccording to descent, or artificially by other characters.Dominant languages and dialects spread widely, and leadto the gradual extinction of other tongues.”

(Darwin, The Descent of Man)



Vater Unser im Himmel, geheiligtwerde Dein Name

Onze Vader in de Hemel, laat UwNaam geheiligd worden

Our Father in heaven, hallowed beyour name

Fader Vor, du som er i himlene!Helliget vorde dit navn





Middle High German:Got vater unser, dâ du bist in demhimelrîche gewaltic alles des dir ist,geheiliget sô werde dîn nam

Old High German:Fater unser thû thâr bist in himile, sigiheilagôt thîn namo

Gothic:Atta unsar þu in himinam, weihnainamo þein


The comparative method in historical linguistics

Genetic language relationships

● In most cases, we do not have written records of earlier stages

● Regular sound correspondences provide evidence for genetic relationship though● Correspondences indicate common ancestor +

different sound shifts● The more cognates two languages share and the

fewer sound shifts separate them, the closer they are related


Example: High German vs. West Germanic


How historical linguists do it

Example: Polynesian languages

● Taken from Crowley & Bowern (2010)





Guidelines for reconstruction

● Only establish sound correspondences if you are reasonably sure the words are cognate

● Assume sound shifts that are plausible (are known to occur frequently)

● Assume as few sound changes as possible for reconstructing a proto-language

● The reconstructed proto-language should have a typologically plausible sound system



Polynesian example

● Vowels in Proto-Polynesian are unchanged in daughter languages (otherwise we would stipulate unnecessary sound shift)

● Likewise, p, m and n are unchanged● Majority rule:

● pp. *t, *N, *v → hw. k, n, w

● lenition is more likely than fortition● also, Proto-Polynesian has p and t, so it should also have a

k, hence:● pp. *k → sm., hw. 7 (rather than *7 → tg./rg. k)



Polynesian example

● majority rule:● pp. *f → rg. 7, hw. h

● not enough data to reconstruct the l and r● majority rule:

● pp. *h, *7 → sm., rg., hw. 0

● change s → h is known to be more common than h → s, hence (against majority rule):● pp. *s → tg./hw. h, rg. 7



Polynesian example

● constructing a tree

Proto-Polynesian

Tongan Samoan Rarotongan Hawaian

t->kN->nv->wk->7f->hh->07->0s->h

k->7h->07->0

f->7h->07->0s->7

s->h



Polynesian example

● constructing a tree

Proto-Polynesian

TonganSamoan

Rarotongan Hawaian

t->kN->nv->wf->hs->h

k->7h->07->0

f->7h->07->0s->7

s->h



Polynesian example

Proto-Polynesian

TonganSamoan

Rarotongan

Hawaian

t->kN->nv->wf->hs->h

k->7

f->7s->7

s->h

7->0h->0



Polynesian example

● reconstruction seems reasonable because● only one shift is assumed twice (s->7), and this type

is known to occur frequently● reconstruction assumes (pull-) chain shifts

– Rarotongan and Proto-Samoan/Hawaian restore the lost 7– Hawaiian additionally restores the lost k and h

● this procedure started from a reconstructed proto-language; usually tree construction and reconstructon of ancestral forms go hand in hand


How computational biologists do it



Transition probabilities in a two-state model

In a symmetric two-state model with rate of change r per unit time, where

0 1r

r we have Prob (1 | 0, t, r) = 12(1− e−2 r t)

0.0 0.5 1.0 1.5 2.00.0

0.2

0.4

0.6

0.8

1.0

2 (1 1 − e

r t

r t

Pro

bab

ilit

y o

f ch

an

ge

)− 2rt

When rt is small then to very good approximation the probability of theevents in the branch is rt or 1 − rt . The latter is nearly 1.

Week 3: Parsimony variants, compatibility, statistics and parsimony – p.26/46



Rooted, bifurcating, labelled treesspecies number of trees

1 12 13 34 155 1056 9457 10,3958 135,1359 2,027,025

10 34,459,42511 654,729,07512 13,749,310,57513 316,234,143,22514 7,905,853,580,62515 213,458,046,676,87516 6,190,283,353,629,37517 191,898,783,962,510,62518 6,332,659,870,762,850,62519 221,643,095,476,699,771,87520 8,200,794,532,637,891,559,37530 4.9518 ×1038

40 1.00985 ×1057

50 2.75292 ×1076

Genome 570, Phylogenetic Inference – p.33/45



computer doggedlylooks at one damned tree after anothercomputes the likelihood of the data given that tree, andmoves on to the next tree

exhaustive search is impossible, given the astronomical number ofdifferent treesno guarantee to find the single best tree, but there are good heuristicshighly informative results:

estimates of branch lengthsconfidence values for branches of a treeprobabilistic reconstructions of ancestral states...



English Dutch Frenchtwo twee deuxthree drie troismountain berg montdog hond chientree boom arbretooth tand dentcold koud froid

⇒

English Dutch FrenchA A AA A AA B AA B BA B CA A AA A B

Different cognate classes for a given meaning are treated just likedifferent alleles of a gene.Once the cognacy data are in this format, the full gamut ofcomputational phylogenetics can be unleashed.


Spectacular results


Skeptically received by traditional historicallinguists


A closer look at cognate classes

based on expertjudgments

usually cover ca.200 concepts(Swadesh list)

publicly availablefor very fewlanguage families(mainlyIndo-European andAustronesian) http://ielex.mpi.nl/


http://ielex.mpi.nl/

Phylogenetic inference with cognate classes

defines aclassification oflanguages intodiscrete states foreach Swadeshconceptexample:Indo-European,concept foot

AshkunSogdianOsseticDigor OsseticIron OsseticWakhi

BaluchiKurdishZazakiTadzikPersianPashtoWaziri

Old PersianAvestan

Vedic Sanskrit Kashmiri

HindiLahndaPanjabi StUrduSindhiMarwariGujaratiMarathiOriyaBengaliBihariNepaliKhaskuraGypsy GkSinghalese

Old Prussian LatvianLithuanian OLithuanian St

Bulgarian PBulgarianMacedonianMacedonian PSerbocroatianSerbianSerbocroatian PSlovenian

Slovenian PRussianRussian PUkrainian P

PolishUkrainianByelorussianByelorussian P

SlovakCzech ECzechSlovak PCzech PPolish PUpper SorbianLower Sorbian

Old Church Slavonic

Old BretonOld CornishOld Welsh

CornishBreton SeBreton ListBreton StWelsh CWelsh N

GaulishOld Irish

Irish AIrish BGaelic ScotsManx

OscanUmbrian

VlachRumanian List

Dolomite LadinoRomanshLadinFriulianItalian

WalloonFrenchProvencalCatalanBrazilianPortuguese StSpanish

Sardinian LSardinian CSardinian N

Latin

Gothic

AfrikaansFlemishDutch ListFrisian

GermanStandard German MunichSchwyzerduetschLetzebuergeschPennsylvania Dutch

Old High GermanOld English English

Old Gutnish

Old Norse Icelandic StFaroeseOld Swedish

StavangerskNorwegianDanishDanish FjoldeGutnish LauOevdalianSwedishSwedish UpSwedish Vl

Tocharian ATocharian B

Ancient Greek

Greek MlGreek DGreek MdTsakonianGreek ModGreek K

Classical ArmenianArmenian ModArmenian List

LycianLuvianHittite

●

●

●●●

●●●●●●●

●●

●

●

●●●

●●●●●●●●●●

● ●●●

●●●●●●

●●●●

●●●●

●●●●●

●●●

●●●

●

●●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●●●●●

●●

●

●●●●●●

●●●

●●●



cognate classes are mathematically treated like alleles of a gene, ordifferent DNA lettersevolution as continuous time Markov chain:

evolution along continuous time line (no discrete generations)a language spontaneously, at each point in time

split into two daughter languages, orchange it’s state for any concept

state transition probability density described by rate matrix

for a given tree structure (with branch lengths) and rate matrix, eachdistribution of observed states at the leaves has a well-definedlikelihood



Inference methods:Maximum Likelihood:

find the tree structure/rate matrix combination maximizing thelikelihood of observed distribution of cognacy classes over languages

Bayesian inference:assume prior distributions over tree structures and rate matricesmay incorporate historical, archaeological, linguistic etc. informationgenerate posterior distribution over trees


Practical issues

for n cognate classes, n2 − n transition rates have to be fitted ⇒often insufficient datatherefore binary recoding of cognate classification:

each cognate class is a separate featurevalues are 1 (present in a language) and 0 (absent)

only two rates need to be fitted per concept


Practical issues


Ashkun

SogdianOsseticDigor OsseticIron OsseticWakhi


Old PersianAvestan


Hindi

LahndaPanjabi StUrduSindhiMarwariGujaratiMarathiOriyaBengaliBihariNepaliKhaskuraGypsy GkSinghalese





SlovakCzech ECzechSlovak PCzech P

Polish PUpper SorbianLower Sorbian

Old Church Slavonic


Cornish

Breton SeBreton ListBreton StWelsh CWelsh N

GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino

RomanshLadinFriulianItalian

WalloonFrenchProvencal

Catalan

BrazilianPortuguese StSpanish


Latin

Gothic


GermanStandard German Munich

SchwyzerduetschLetzebuergeschPennsylvania Dutch


Old Gutnish


StavangerskNorwegian

DanishDanish Fjolde

Gutnish LauOevdalianSwedishSwedish UpSwedish Vl


Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:A


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:B


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:C


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:D


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:E


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:F


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:G


Ashkun



Old PersianAvestan


Hindi








Old Church Slavonic


Cornish


GaulishOld Irish


OscanUmbrian

VlachRumanian List

Dolomite Ladino



Catalan



Latin

Gothic





Old Gutnish



DanishDanish Fjolde



Ancient Greek



LycianLuvianHittite

●

●●●●●

●●●●●●●

●●

● ●

●

●●●●●●●●●●●●●●

● ●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●●●

●

●●●●●

●●

●●●●

●●

●●

●

●●●●

●●●

●

●●●

●●●

●

●

●●●●

●●

●●●

●● ●

●

● ●●●

●●

●●

●●●●●

●●

●

●●●●●●

●●●

●●●

foot:J


Parallel evolution

by definition, each cognate class can emerge only oncetransition rate 0→ 1 should thus be exceedingly smallno parallel evolution 0→ 1 to be expectedonly expected exception: undetected borrowingsHowever (see next slide...)


Parallel evolution

SogdianOsseticDigor OsseticIron OsseticWakhiShughniSariqoli

BaluchiZazakiTadzikPersianPashtoWaziri

Avestan

Vedic SanskritKashmiri

Hindi

LahndaPanjabi StUrduSindhiMarwariGujaratiMarathiOriyaBengaliBihariNepaliKhaskuraGypsy Gk

Old PrussianLatvianLithuanian OLithuanian St






Old Church Slavonic

Cornish


Old Irish

Irish AIrish BGaelic Scots

VlachRumanian List

Dolomite Ladino



Catalan


Sardinian LSardinian C

Latin




Old High GermanOld English

Old Gutnish

Old NorseIcelandic StFaroese

Old Swedish


DanishDanish Fjolde


Albanian T

Ancient Greek

Greek MlGreek DGreek MdGreek ModGreek K


●●●●●●●

●●●●●●

●

●●

●

●●●●●●●●●●●●●

●●●●

●●●●●●●●

●●●●

●●●●

●●●●●

●●●

●

●

●●●●●

●

●●●

●●

●

●●●●

●●●

●

●●●

●●

●

●●●●

●●

●●●

●●

●

●●●

●

●●

●●

●●●●●

●

●

●●●●●

●●●

leg:Q


Parallel evolution

apparently independent emergence of leg:Q in Greek, Icelandic,Romanian and Proto-Indo-Iraniantraditional scholarship:

members of class leq:Q are cognate to members of foot:Bboth cognate classes descend from Proto-Indo-European *ped-parallel semantic change foot → leg in several branches ofIndo-European


Testing for correlated evolution

Following Pagel and Meade (2006) (see also Dunn et al. 2011), usingsoftware BayesTraits(http://www.evolution.rdg.ac.uk/BayesTraits.html)

model comparison

leg:B = 0

leg:B = 1

foot:Q = 0

foot:Q = 1

foot:B=0leg:Q=0

foot:B=1leg:Q=0

foot:B=0leg:Q=1

foot:B=1leg:Q=1

log Bayes Factor ≈ 243 in favor of second model ⇒ strong evidencedownsides:

strong evidence in favor of dependence, but weak evidence in favor ofdirectionalityinformation about Swadesh entries of a single family does not affordscaling up


http://www.evolution.rdg.ac.uk/BayesTraits.html

Polysemy networks

A polysemy network (cf. Dellert and Münch 2015, see also List et al.2013) is

a graph over concepts (expressed by glosses)where each link represents the fact that at least one language has alexical item which can denote both concepts (or rather: can betranslated by both glosses)if we count the number of instances in different language families, linkstrength measures cross-linguistic colexification between concepts

Polysemy networks are a potential source of evidence for plausibility ofsemantic shifts:

intuition: every instance of semantic shift needs to pass anintermediate stage where the word in question is polysemousa massively cross-linguistic polysemy network provides us with asnapshot of semantic evolution in action

⇒ Polysemy networks might be a good model for modeling semanticchange in computational historical linguistics!


Polysemy networks

Question: How well does a polysemy network model attestedsemantic shifts?Data:

a massively cross-linguistic polysemy network (TUE) developed inTübingen within EVOLAEMP (Dellert, 2014)the Catalogue of Semantic Shifts (ZAL) maintained by the Russianacademy of sciences (Zalizniak, 2008; Zalizniak et al., 2012)

Method: for each attested change, determine the length of theshortest paths between the relevant concepts in the network


Tübingen Polysemy Network (TUE)

is derived from a large German-based dictionary database containingmore than 750,000 entries in 114 languageshas near-complete coverage of a 1,000-item list of basic concepts in 88languagesis built on more than 10,000 translations each for 31 languages from10 primary language families of Eurasiais available to other researchers upon request


Tübingen Polysemy Network (TUE): Some data

32,653 concepts and 47,647 linksa central connected componentof 13,073 nodes (!)1,640 smaller components and14,391 unconnected islandsan average of 2.4 neighbors foreach nodea typical shortest path lengthof 8


Small world characteristics


The Catalogue of Semantic Shifts

crosslinguistically recurring shifts collected manually by expertsno restrictions or preferences in the range of meanings or languages(de facto: focus on Eurasian)contained 3,650 semantic shifts at the time of publicationclassified into six types, each supported by up to 40 realizationspublicly available via a web interface:http://semshifts.iling-ran.ru/


http://semshifts.iling-ran.ru/

Procedure

English version of the catalogue was extracted from the websiteconsiderable semi-automated cleanup workthe total number of shift pairs in the result: 6,174

Preperatory steps

metalanguages of thecatalogue: English and Russian

primary language of theTübingen Polysemy Network:German

electronic English-Germandictionary needed to be useedas an intermediary forcomparing the links in bothresources

Experiment

evaluate pairs of Englishglosses in the catalogue againstthe network by determining theshortest paths in thenetwork between any pair ofGerman translationequivalents

do this separately for eachtype, and compare recall


Example

Bein::N

Bein[Tisch]::N

Fuß::N

Fuß[Maß]::NFüßchen::N

Keule::N

Keule[Gastronomie]::N

Knochen::N

Lammkeule::N Oberschenkel::NPfote::N

Schenkel::NSohle::N

Stamm::N

Stütze::N

Tatze::N

Unterschenkel::N


Results

Types of Semantic Shifts in the CataloguePolysemyDiachronic Semantic EvolutionSyncretismCognatesBorrowingsMorphological Derivation

Only Diachronic Semantic Evolution, Cognates, and Borrowings containinstances of semantic change in the stricter sense.


Results

Polysemysynchronic polysemy: A and Bare meanings of a polysemouswordZAL: tree — forest / TUE:Baum ‘tree’ — Holz ‘wood’ —Wald ‘forest’Example: Tuvan yjash ‘tree;forest’

Baum::N

Holz::N

Stab::N

Stange::N

Stock::N

hölzern::A

Brennholz::N

Dschungel::N

Wald::N

Forst::N


Results

Diachronic semantic evolutionDiachronic semantic evolution ofa word in one language or froman ancestor language to adescendant languageZAL: board, plank — table, desk/ TUE: Brett ‘board’ — Tisch‘table’Example: Latin tabula ‘plate,tablet’ → Italian tavola ‘table’

Bord::N

Brett::N

Platte::NScheibe::NSchild::N

Tafel::N

Tisch::N Schreibtisch::N


Results

Diachronic semantic evolutionDiachronic semantic evolution ofa word in one language or froman ancestor language to adescendant languageZAL: catch — hunt / TUE:jagen ‘hunt’ — fangen ‘catch’Example: Latin capto ‘variousmeanings related to catching’ →Italian cacciare ‘to hunt’

anfangen::V

beginnen::V

fangen::V

fassen::V

greifen::V

angeln::V

fischen::V

bekommen::V

empfangen::V ergreifen::V

erhalten::V

kriegen::V

nehmen::V

eilen::V

fahren::V

jagen::V

laufen::V

rennen::V

halten::V treiben::V

verfolgen::V


Results

Diachronic semantic evolutionDiachronic semantic evolution ofa word in one language or froman ancestor language to adescendant languageZAL: mountain — nose / TUE:Berg ‘mountain’ — Hügel ‘hill’— Gipfel ‘summit’ — Spitze‘tip’ — Nase ‘nose’Example: Hungarian orr ‘nose’← ‘top, mountain’

Abhang::N

Hang::NHügel::N

Anfang::NEnde::N

Kopf::NSpitze::N

Anhöhe::N

Gipfel::N

Haufen::N

Höhe::N

Berg::NFels::NFelsen::N

Gebirge::N

Dach::N

Deckel::N

Schluss::N

Scheitel::N

oberer::A

Nase::N

Schnabel::N

Schnauze::N

Stachel::N


Results

Cognatesmeanings A and B belong towords of two sister languageswhich have developed from thesame root in their commonancestorZAL: hear — understand /TUE: hören ‘hear’ - verstehen‘understand’Example: French entendre ‘tohear’ — Spanish entender ‘tounderstand, to beknowledgeable’

Verständnis::N

begreifen::Vverstehen::V

erfassen::V erkennen::V

fassen::V

wahrnehmen::V

empfinden::Verfahren::V

fühlen::Vhören::V

spüren::V

vernehmen::Vwissen::V

umfassen::V

vermuten::Verreichen::V

können::V

jm_gehören::V

zuhören::V


Results

BorrowingsB is the meaning of a borrowedword, while A is the meaning ofits source in the donor languageZAL: bell — watch/clock /TUE: Glocke ‘bell’ - Uhr‘watch/clock’Example: Medieval Latin clocca‘bell’ → English clock ‘clock,watch’

Armbanduhr::N

Uhr::N

Glocke::N

Klingel::N Stunde::N

Uhrzeit::N

Zeit::N


Results

Table: Shifts in the catalogue covered by the polysemy network.

Semantic Number of No Path length Path lengthshift type instances path 1 or 2 3 and morePolysemy 2315 20.6 % 35.0 % 44.4 %Semantic 107 26.2 % 33.6 % 40.2 %EvolutionMorphological 597 28.5 % 29.0 % 42.5 %DerivationSyncretism 43 25.6 % 55.8 % 18.6 %Borrowing 58 31.0 % 41.4 % 27.6 %Cognates 393 23.7 % 37.9 % 38.4 %


Towards detecting directional biases in meaningchange

Assumptionsmeaning change proceeds along the edges of the polysemy graphfor each possible transition, there is a fixed transition rate, acrosslanguage families and times → ultimately founded in cognition

Research questionsCan we estimate those rates from data?Can we detect directional biases?

RequirementMany data — across language families and beyond the Swadesh list


The NorthEuraLex database

Massive data collection effort ofthe Tübingen EVOLAEMPproject(currently) translations of 1,017concepts into 103 (mostly)Northern Eurasian languages (cf.Dellert, 2015)everything transcriped in IPA(so far) no manual cognatecoding


Automatic cognate classification

clustering of the wordsfor a given conceptheuristic technique basedon phonetic similarityso far only roughapproximation to expertclassification

Language Word Cognate class

Dra.Southern Dravidian.Kannada paada Fuß:12IE.Baltic.Lithuanian peeda Fuß:12IE.Greek.Greek po8i Fuß:12IE.Indic.Hindi pEEr Fuß:12IE.Indic.Hindi paaw Fuß:12IE.Iranian.Ossetic fad Fuß:12IE.Iranian.Pashto psa Fuß:12IE.Iranian.Persian paa Fuß:12IE.Romance.French pye Fuß:12IE.Romance.Italian piede Fuß:12IE.Romance.Portuguese pE Fuß:12IE.Romance.Spanish pye Fuß:12Kor.Korean.Korean pal Fuß:12Ura.Permic.Udmurt p3d Fuß:12Alt.Tungusic.Manchu pothxo Fuß:6Brs.Burushaski.Burushaski ut Fuß:6Brs.Burushaski.Burushaski utis Fuß:6IE.Armenian.Armenian votkh Fuß:6IE.Germanic.Danish foo78 Fuß:6IE.Germanic.Dutch vut Fuß:6IE.Germanic.English fut Fuß:6IE.Germanic.German fuus Fuß:6IE.Germanic.Icelandic foutir Fuß:6IE.Germanic.Norwegian fuut Fuß:6IE.Germanic.Swedish fuut Fuß:6


Getting a handle on meaning change

two cognate classes Aand B, for concepts cAand cB, are merged if

cA and cB areneighbors in thepolysemy graphin at least onelanguage, A and Bare instantiated by thesame phonetic string

Language Word Cognate class

Dra.Southern Dravidian.Kannada paada footIE.Baltic.Lithuanian peeda footIE.Greek.Greek po8i footIE.Indic.Hindi pEEr footIE.Indic.Hindi paaw footIE.Iranian.Ossetic fad footIE.Iranian.Pashto psa footIE.Iranian.Persian paa footIE.Romance.French pye footIE.Romance.Italian piede footIE.Romance.Portuguese pE footIE.Romance.Spanish pye footKor.Korean.Korean pal footUra.Permic.Udmurt p3d footAlt.Tungusic.Nanai b3gdi Bein:9IE.Greek.Greek po8i Bein:9IE.Indic.Bengali pa Bein:9IE.Iranian.Pashto psa Bein:9IE.Iranian.Persian paa Bein:9Krt.Kartvelian.Georgian phExi Bein:9Ura.Permic.Udmurt p3d Bein:9

0see also (Dellert, 2016) for a different approachGerhard Jäger (Tübingen) Phylogeny of word meanings 1/26/2016 61 / 65

Getting a handle on meaning change

treat each merged cognate class as feature and meaning as valuefit a model with two parameters:

transition rate foot → legtransition rate leg → foot

Language Word meaning

Alt.Tungusic.Nanai b3gdi legDra.Southern Dravidian.Kannada paada footIE.Baltic.Lithuanian peeda footIE.Greek.Greek po8i foot legIE.Indic.Bengali pa legIE.Indic.Hindi pEEr/paaw footIE.Iranian.Ossetic fad footIE.Iranian.Pashto psa foot legIE.Iranian.Persian paa foot legIE.Romance.French pye footIE.Romance.Italian piede footIE.Romance.Portuguese pE footIE.Romance.Spanish pye footKor.Korean.Korean pal footKrt.Kartvelian.Georgian phExi legUra.Permic.Udmurt p3d foot leg


Results

no evidence of directional bias forfoot ↔ legmoon ↔ months

foot/leg

bias

−1.0 −0.5 0.0 0.5 1.0 1.5

mean = −0.021

95% HDI−0.674 0.808

moon/month

bias

−1.5 −1.0 −0.5 0.0 0.5 1.0 1.5

mean = 0.0238

95% HDI−0.964 0.871

0model fitted using BayesTraits (Pagel and Meade, 2014); tree sample generatedwith BayesPhylogenies (Pagel and Meade, 2015)


Results

clear evidence of directional bias forsun ↔ daytongue ↔ language

sun/day

bias

1 2 3 4 5 6

mean = 2.13

95% HDI0.852 3.37

tongue/language

bias

−2 0 2 4 6

mean = 2.28

95% HDI0.593 4.11


Conclusion

method suffers from poor quality of cognate classificationresults therefore to be taken with a pound of saltstill, proof of concept that crosslinguistic lexical data provide a handleon studying laws of meaning evolution

Great potential for combining cognitive with phylogeneticmethods!


References

Atkinson, Q. D. and R. Gray (2005). Curious parallels and curious connections —phylogenetic thinking in biology and historical linguistics. Systematic Biology,54(4):513–526.

Bouckaert, R., P. Lemey, M. Dunn, S. J. Greenhill, A. V. Alekseyenko, A. J.Drummond, R. D. Gray, M. A. Suchard, and Q. D. Atkinson (2012). Mappingthe origins and expansion of the Indo-European language family. Science,337(6097):957–960.

Crowley, T. and C. Bowern (2010). An introduction to historical linguistics.Oxford University Press, Oxford.

Dellert, J. (2014). Evaluating cross-linguistic polysemies as a model of semanticchange for cognate finding. Proceedings of the Workshop on semantictechnologies for research in the humanities and social sciences (STRiX).November 24-25, Gothenburg, Sweden.

Dellert, J. (2015). Compiling the Uralic dataset for NorthEuraLex, alexicostatistical database of Northern Eurasia. Proceedings of the FirstInternational Workshop on Computational Linguistics for Uralic Languages.January 16, Tromsø, Norway.

Dellert, J. (2016). Using causal inference to detect directional tendencies insemantic evolution. In 11th International Conference on the Evolution ofLanguage (EvoLang 2016). New Orleans.Gerhard Jäger (Tübingen) Phylogeny of word meanings 1/26/2016 65 / 65

References

Dellert, J. and A. Münch (2015). Evaluating the potential of a large-scalepolysemy network as a model of plausible semantic shifts. In H. Baayen,G. Jäger, M. Köllner, J. Wahle, and A. Baayen-Oudshoorn, eds., Proceedingsof QITL-6. University of Tübingen, Tübingen.

Dunn, M., S. J. Greenhill, S. Levinson, and R. D. Gray (2011). Evolved structureof language shows lineage-specific trends in word-order universals. Nature,473(7345):79–82.

Felsenstein, J. (2004). Inferring Phylogenies. Sinauer Inc. Publishers, Sunderland.List, J.-M., A. Terhalle, and M. Urban (2013). Using network approaches to

enhance the analysis of cross-linguistic polysemies. In Proceedings of the 10thInternational Conference on Computational Semantics (IWCS 2013), pp.347–353. Association for Computational Linguistics, Potsdam, Germany.

Pagel, M., Q. D. Atkinson, A. S. Calude, and A. Meade (2013). Ultraconservedwords point to deep language ancestry across Eurasia. Proceedings of theNational Academy of Sciences, 110(21):8471–8476.

Pagel, M. and A. Meade (2006). Bayesian analysis of correlated evolution ofdiscrete characters by reversible-jump Markov chain Monte Carlo. TheAmerican Naturalist, 167(6):808–825.

Pagel, M. and A. Meade (2014). BayesTraits 2.0. software distributed by theauthors.Gerhard Jäger (Tübingen) Phylogeny of word meanings 1/26/2016 65 / 65

Pagel, M. and A. Meade (2015). BayesPhylogenies 2.0. software distributed bythe authors.

Pereltsvaig, A. and M. W. Lewis (2015). The Indo-European Controversy.Cambridge University Press.

Zalizniak, A. A. (2008). A catalogue of semantic shifts: Towards a typology ofsemantic derivation. In M. Vanhove, ed., From polysemy to semantic change.Towards a typology of lexical semantic associations, pp. 217–232. JohnBenjamins, Amsterdam and Philadelphia.

Zalizniak, A. A., M. Bulakh, D. Ganenkov, I. Gruntov, T. Maisak, and M. Russo(2012). The catalogue of semantic shifts as a database for lexical semantictypology. Linguistics, 50(3):633–669.


Date post:	26-Oct-2021
Category:	Documents
Upload:	others
View:	4 times
Download:	0 times

WissenschaftskollegBerlin - uni-tuebingen.de

Documents