form‐meaning systemacity and language evoluon · form‐meaning systemacity and language evoluon...

Post on 07-May-2018

225 Views

Category:

Documents

4 Downloads

Preview:

Click to see full reader

TRANSCRIPT

Form‐meaningsystema/cityandlanguageevolu/on

MónicaTamarizmonica@ling.ed.ac.uk

Tübingen,1September2009

Language as an adaptive system

Language structure is influenced by many pressures

LANGUAGE

Conveyvariedmeanings!

ExpressmyidenFty!

Beeasytolearn!

Beeasytoproduce! Beeasytounderstand!

Language as an adaptive system

LANGUAGE

Expressivity

Learnability

Conveyvariedmeanings!

ExpressmyidenFty!

Beeasytolearn!

Beeasytoproduce! Beeasytounderstand!

Language structure is influenced by many pressures

Overview

1.  Form-meaning systematicity

2.  Exploring systematicity in Natural Language: A corpus study

3.  Systematicity and compositionality in Artificial Languages: An experimental approach

4. Implications for language evolution

1. Systematicity between forms and meanings in language

Systematicity

•  TherelaFonshipbetweentwospacesissystemaFcifthestructureofonespacereflectsthestructureoftheother

•  Thereforeknowingaboutonestructureprovidesuswithsomeknowledgeabouttheother

Systematicity in language PhonologicalformssystemaFcallyreflectmorphosyntax

DoesitreflectsemanFcs??

present past 3rd person

talk talked talks

play played plays

study studied studies

A systematic lexicon

Wordsthatsoundsimilartendtohavesimilarmeanings

Meaningspace

SystemaFcityreflectsapreferenceforisomorphic/topographic/systemaFcrepresentaFonsinthenervoussystem

Systematic lexicon – Why?

FuncFonsofsystemaFcity:

• Helplearn/understandnewitems

• AllowcreaFvityandgeneralisaFon

Systematic lexicon – What for?

Meaningspace

2. Exploring systematicity in Natural Language: A corpus study

(Tamariz 2005, 2008)

Hypothesis:ThestructureofthementallexiconshowssystemaFcitybetweenformsandmeanings

Method:QuanFtaFvecorpusstudy

Data:Threesubsetsofa1MwordSpanishtranscribedspeechcorpus(MarcosMarin,1992)

252CVCV

146CVCCV

148CVCVCV

Twostudies:

SyntaxandsemanFcs

SemanFcsonly

Looking for systematicity in the lexicon

wordsoffrequency>=20

Measuring systematicity

DistribuFonalsimilarity(Landauer&Dumais,1997;MacDonald,2000)

Phonologicalsimilarity(Tamariz2005,2008)

kasa

lima

nene

kita

kasa

kita lima

nene

Measuring systematicity

DistribuFonalsimilarity(Landauer&Dumais,1997;MacDonald,2000)

Phonologicalsimilarity(Tamariz2005,2008)

k-a-s-a k-i-t-a

v1, v2, v12 c1, c2, c12 stress1, stress2

strv1, strv2 (syntax)

Patterns of coocurrence with:

394 words (Stx+Sem) 320 content words (Sem)

of freq >=200

Measuring systematicity

Phonologicalsimilarity

DistribuFonalsimilarity

FisherdivergenceFD=1.4

(significanceestablishedwithMonteCarloanalysis)kasa

lima

nene

kita

kasa

kita lima

nene

Result

Result

Wordsthatsoundsimilardotendtohavesimilarmeanings

ThereissystemaFcitybetweenformandmeaninginthelexicon

Result

Wordsthatsoundsimilardotendtohavesimilarmeanings

However,toomuchsystemaFcitymayposeaproblemforcomprehension!

ThereissystemaFcitybetweenformandmeaninginthelexicon

Systematicity vs. disambiguation

kiri

kili kini

kisi

•  Thelexiconstructureisunderpressure

1.  TobesystemaFc(whichhelpsprocessingandlearning)

2.  ToavoidambiguiFesderivedfromsystemaFcity

Hypothesis:

Thestructureofthementallexiconisshapedbythesetwoopposedpressures

•  SomeaspectsofformssystemaFcallyreflectthestructureofwordmeanings

•  Otheraspectsofformsshouldhave“negaFvesystemaFcity”withrespecttowordmeanings

The phonological correlates of systematicity

Method

MeasuresystemaFcitybetween

•  DistribuFonalsimilarityand

•  Phonologicalsimilarityintermsof

‐  Consonants

‐  Vowels

‐  Stresspagerns

E.g.Dowordsthatsharethefirstconsonanttendtohavesimilarmeanings?

Results TheimpactofdifferentphonologicalparametersonsystemaFcity

consonantsvowelsstress

Allvaluesp<0.01exceptwherestated

CVCVwords

Results TheimpactofdifferentphonologicalparametersonsystemaFcity

consonantsvowelsstress

Allvaluesp<0.01exceptwherestated

CVCVwords

• ConsonantshaveposiFveimpact

• Vowels,negaFveimpact

• StressedvowelinthepenulFmatesyllable,negaFveimpact

• Otherstress,posiFveimpact

Results TheimpactofdifferentphonologicalparametersonsystemaFcity

consonantsvowelsstress

Allvaluesp<0.01exceptwherestated

CVCVCVwords

• ConsonantshaveposiFveimpact

• Vowels,mostlynegaFveimpact

• StressedvowelinthepenulFmatesyllable,negaFveimpact

• Otherstress,posiFveimpact

Results TheimpactofdifferentphonologicalparametersonsystemaFcity

consonantsvowelsstress

Allvaluesp<0.01exceptwherestated

CVCCVwords

• Consonantshavemixedimpact

• Vowels,mostlynegaFveimpact

• StressedvowelinthepenulFmatesyllable,negaFveimpact

• Otherstress,posiFveimpact

Discussion IntheSpanishlexicon…

•  Thestructureofconsonants,stresspagernandthestressedfinalvowelhasadaptedtothepressureforsystemaFcity

•  Thestructureofvowels,especiallythestressedvowelinthepenulFmatesyllablehasadaptedtothepressureforeasydisambiguaFon

•  TheseresultssupportthehypothesisthatthelexiconisanadapFvesystem,evolvingundermulFplepressures

(OtherlanguagesmayhavefounddifferentsoluUonstothisconflict)

3. Systematicity and compositionality in Artificial Languages:

An experimental approach

Systematicity and compositionality

HolisFc

lop=

mer=

sus=

pina=

e=

koF=

•  ComposiFonalityisaspecialtypeofsystemaFcitybetweenformsandmeanings

“InacomposiFonalsystem,themeaningofanexpressionisafuncFonofthemeaningofitscomponentparts,plusthewaythepartsarecombinedtogether”

+

+

+

lopi loto

napi nato

mepi meto

+

+

+

lop mer

sus pina

e koF

ComposiFonal

lo‐=

na‐=

me‐=

‐pi=

‐to=+

+

+

+

Compositionality and Expressivity

ComposiFonalityfacilitatesemergentcategoriesandgeneralisaUon

(Ifthemeaningspaceisstructured)

+

+

+

lopi loto

napi nato

mepi meto

+

+

+

lop mer

sus pina

e koF

ComposiFonal

lo‐=

na‐=

me‐=

‐pi=

‐to=+

• Extrameanings:

1=outsideshape

2=insert

• GeneralisaFonpossible

HolisFc

lop=

mer=

sus=

pina=

e=

koF=

• Noextrameanings

• GeneralisaFonnotpossible

+

+

+

Why is language compositional?

•  ThestructureoflanguageevolvedoverthousandsofhumangeneraFonswholearnitandthenuseit

•  WeassumethatiniFallylanguagewasnotcomposiFonal(e.g.Wray,2002)

•  Howdiditcometohavethisstructure?

–  Theroleofindividuallearning

–  Theroleofculturaltransmission

The role of individual learning (Tamariz & Smith 2008)

•  Hypothesis–  ThereisalearningbiasfavouringcomposiFonalmappings

betweensignalsandmeanings

•  PredicFon–  MorecomposiFonallanguagesareeasiertolearn

•  Test–  ArFficialLanguageLearningexperiment–  ConstructarFficiallanguageswithdifferentlevelsof

composiFonalityandseewhicharebegerlearned

Artificial languages

MEANINGS

27differentmeanings

+

FORMS

3‐syllablepseudowords

27differentforms

Procedure

Training

hune

lene

letupi

miko

lekono

TesFng

1.Signals

Task:Typename

2.Meanings

Task:Selectpicture

lekono

_

27screensx3Fmes

Variables

•  Independentvariable

4levelsofinputlanguagecomposiFonality

LANG1 LANG2 LANG3 LANG4

RegMap0.154 0.456 0.754 1.00

•  Dependentvariable

Learnability–  MeasuringhowwellthecomposiFonallanguagestructurewas

learned

Measuring Compositionality (RegMap)

RegMap measures the confidence that a signal and a meaning are unambiguously and consistently associated

lene

letu

hune

leko

Co-occur Freq

le - 3 2 2

- ne 2 - ko 2

miko

RegMap(X | Y ) = 1− H (X | Y )log(nx )

Results

Learnability of the compositional language structure

Similarity of compositional language structure of input and output languages

Results and discussion

MorecomposiFonallanguagestructuresarelearntbeger

ParFcipantsseemstohavealearningbiasforcomposiFonality

IndividuallearningmayhaveplayedaroleinlanguagebecomingcomposiFonal

Learnability of the compositional language structure

Results and discussion

MorecomposiFonallanguagestructuresarelearntbeger

ParFcipantsseemstohavealearningbiasforcomposiFonality

IndividuallearningmayhaveplayedaroleinlanguagebecomingcomposiFonal

Learnability of the compositional language structure

Butwheredoesthisbiascomefrom?

The role of enculturation (Brown & Tamariz, 2008)

•  Hypothesis–  ThelearningpreferenceforcomposiFonalstructureisgivenby

culturalknowledgesuchasliteracy

•  PredicFon–  LiteratesarebegeratlearningcomposiFonallanguagesthan

illiterates

•  Test–  MusicalArFficialLanguageLearningexperiment

–  ConstructarFficialmusicallanguageswithdifferent levelsofcomposiFonalityandseewhicharebeger learned

Procedure

Training

hune

miko

TesFng

Signals

Task:Singtune

27screensx3Fmes

lene

lekono

27signalspresented

letupi

Variables

•  Independentvariables–  3levelsofinputlanguagecomposiFonality

LANG1 LANG2 LANG3

RegMap 0.004 0.66 1.00

–  2levelsofmusicalliteracy(literate/illiterate)

•  Dependentvariable–  LearnabilityofthecomposiFonalstructureoflanguages

Results

ItseemsthatliteracyenhancesthepreferenceforlearningcomposiFonalsystems

Learnability of the compositional language structure

Discussion: The role of individual learning

•  TheseresultsindicatethatindividuallearningmayincreaselanguagecomposiFonality

•  Thiseffectmaybeenhancedbyculturalknowledgesuchasliteracy

•  ButlanguageevolvedasitwastransmigedovermanygeneraFons…

The role of cultural transmission (Kirby, Cornish & Smith, 2008)

•  Hypothesis–  Culturaltransmissioncontributestotheemergenceof

systemaFcstructureinlanguage

•  PredicFon–  Theeffectsofindividuallearningareenhancediniterated

learning

•  Test–  IteratedArFficialLanguageLearningexperiment

–  MeasuresystemaFcityandlearnability

kimako

koni

kanige

kuni

winige

komako

Genera/on0:randomsignals

MEANINGS

27differentmeanings

Languages and procedure

Genera/on1

•  OneoftheiniFalrandomlanguages

Results

huhunigu kemuniwa

kihupo

wakiki pokiheku waguhuki

nihu niguki koni

muwapo powa

hukinimu

wako hukeko pohumu

muko kokeguke

kimu

kekewa konihuke

kopo

huwa hukike

ponikiko

kowagu kokihuko kiwanike

•  ThesamelanguagesomegeneraFonslater

Results

•  LanguagesareincreasinglysystemaFcandlearnable

Results

Learning error Systematicity

•  LanguagesareincreasinglysystemaFcandlearnable•  Composi/onalityisapparent,butitisnotquan/fied

Results

Learning error Systematicity

•  ForeachgeneraFon,calculateRegMapofallpairsofmeaningandformdimensions

E.g.forGeneraFon2:

Quantifying compositionality (Cornish, Tamariz & Kirby, in press)

Segment 1

Results

Emergence of compositionality over cultural transmission

Segment 2

Emergence of compositionality over cultural transmission

Results

Segment 3

Emergence of compositionality over cultural transmission

Results

Discussion

•  WehaveobservedapreferenceforindividualstolearnsystemaFclanguages

•  ThatpreferencemaybeenhancedbyculturalinsFtuFonssuchasliteracyandteaching

•  CulturaltransmissioncanmediatetheaccumulaFonoflanguagesystemaFcityoverFme

Conclusion

•  Manypressures,someFmesconflicFng,havelestheirmarkonlanguagestructureoveritslongevoluFonaryhistory.

•  Thebalanceofsystema/cityinlanguageisanadaptaFontomanyfactors

–  Neuralconstraints(preferenceforisomorphic,systemaFcmappings)

–  Learnedbiases(livinginastructured,culturalworld,welearntoexpectstructure)

–  LanguagefuncFon(theneedtoproduceandunderstandflexibly)

–  CulturaltransmissionovergeneraFons

References •  Tamariz(2005).ExploringtheadapFvestructureofthementallexicon(unpublished

PhDthesis).

•  Tamariz(2008).ExploringsystemaFcitybetweenphonologicalandcontext‐cooccurrencerepresentaFonsofthementallexicon.TheMentalLexicon,3(2):259‐278.

•  Kirby,Cornish&Smith(2008).CumulaFveCulturalEvoluFonintheLaboratory:anexperimentalapproachtotheoriginsofstructureinhumanlanguage.PNAS,105(31):10681‐10686.

•  Tamariz&Smith(2008).QuanFfyingtheregularityofthemappingsbetweensignalsandmeanings.In:A.D.M.Smith,K.Smith,R.FerreriCancho,Eds.,TheevoluUonoflanguage,pp.315‐322.

•  Tamariz(inpress).CouldarbitraryimitaFonandpagerncompleFonhavebootstrappedhumanlinguisFccommunicaFon?InteracUonStudies.

•  Cornish,Tamariz&Kirby(2010).ComplexadapFvesystemsandtheoriginsofadapFvestructure:whatexperimentscantellus.SpecialissueonLanguageasacomplexAdapFveSystem.LanguageLearning:59:4S1.

•  Brown,J.E.andTamariz,M.(2010).TheroleofliteracyandpracFceintheevoluFonoflinguisFcstructure.In:A.D.M.Smith,B.deBoer,M.Schouwstra,Eds.,TheevoluUonoflanguage.

Form‐meaningsystema/cityandlanguageevolu/on

MónicaTamarizmonica@ling.ed.ac.uk

Tübingen,1Sept2009

top related