64
Using Computational Modeling to Understand Language Acquisition another one KI $y Noun Every ki.y didn’t Who does… is pre.y? UC Computational Social Science

Using Computational Modeling to Understand Language

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

UsingComputationalModelingtoUnderstandLanguageAcquisition

anotherone

KI$y

Noun

Everyki.ydidn’t…

Whodoes…ispre.y?

UCComputationalSocialScience

Languageacquisition:Howhumanslearnlanguageknowledge

Languageacquisition:Howhumanslearnlanguageknowledge

Firstlanguageacquisition=Learningnativelanguage(s)

Happensasayoungchild

Languageacquisition:Howhumanslearnlanguageknowledge

Secondlanguageacquisition=Learningnon-native/foreignlanguage(s)

Happensasanolderchildoradult

Firstlanguageacquisition

Firstlanguageacquisition

Howdochildrenacquiretheknowledgeaboutlanguagethattheydofromthelanguagedatatheyhave?

Babiesareamazingatlearninglanguage

Whyfirstlanguageacquisition?

Andtheylearnalot!

Babiesareamazingatlearninglanguage

Likewhat?

Andtheylearnalot!

Likewhat?

Everythingyouknowaboutyournativelanguage(s).

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Youknowhowtoidentifywordsinfluentspeech(speechsegmentation)

Youknowhowtopronouncewords(metricalphonology)

KI$ykiTTY

whatapre$yki$y!

speechsegmentaDon

Youknowthatcertainwordsbehavelikeotherwords(syntacticcategorization)

ki$y

owl

penguin

whatapre$y___!Noun

whatapre$yki$y!

speechsegmentaDon

KI$y

kiTTY

metricalphonology

Youknowhowtointerpretwordsincontext(syntax,semantics)

“Ohlook—apre$yki$y!”“Look—there’sanotherone!”

whatapre$yki$y!

speechsegmentaDon

KI$y

kiTTY

metricalphonologyki$y

owlpenguin

Noun

syntacDccategorizaDon

Thiski.ywasboughtasapresentforsomeone.

Lilythinksthiski.yispre.y.

WhodoesLilythinktheki.yforispre.y?

Youknowhowtoputwordstogethertoaskquestions(syntax)

“Ohlook—apre$yki$y!”“Look—there’sanotherone!”

syntax,semanDcs

whatapre$yki$y!

speechsegmentaDon

KI$y

kiTTY

metricalphonologyki$y

owlpenguin

Noun

syntacDccategorizaDon

WhodoesLilythinktheki.yforispre.y?

syntax

Youknowhowtoidentifytherightinterpretationincontext(pragmatics)

“Everyki$ydidn’tsitonthestairs”

NotallkiVessatonthestairs.

NokiVessatonthestairs.x

whatapre$yki$y!

speechsegmentaDon

KI$y

kiTTY

metricalphonologyki$y

owlpenguin

Noun

syntacDccategorizaDon

“Ohlook—apre$yki$y!”“Look—there’sanotherone!”

syntax,semanDcs

pragmaDcs

“Everyki$ydidn’tsitonthestairs”

KI$y

kiTTY

metricalphonologyki$y

owlpenguin

Noun

syntacDccategorizaDon

NotallkiVessatonthestairs.

whatapre$yki$y!

speechsegmentaDon

“Ohlook—apre$yki$y!”“Look—there’sanotherone!”

syntax,semanDcs

WhodoesLilythinktheki.yforispre.y?

syntax

syntax

speechsegmentaDon

metricalphonology

syntacDccategorizaDon

syntax,semanDcs

pragmaDcs

Sohowexactlydochildrenlearnallthis?

Weknowtheydoitrelativelyquickly.

syntax

speechsegmentaDon

metricalphonology

syntacDccategorizaDon

syntax,semanDcs

pragmaDcs

Sohowexactlydochildrenlearnallthis?

Muchofthelinguisticsystemisalreadyknownbyage4.

Theyalsodon’tseemtogetalotofexplicitinstruction.Andwhentheydo,theydon’treallypayattentiontothingsthatdon’timpactmeaning.

Sohowexactlydochildrenlearnallthis?

Child:Wantotheronespoon,Daddy.Father:Youmean,youwanttheotherspoon.Child:Yes,Iwantotheronespoon,pleaseDaddy.Father:Canyousay“theotherspoon”?Child:Other…one…spoon.Father:Say“other”.Child:Other.Father:“Spoon.”Child:Spoon.Father:“Otherspoon.”Child:Other…spoon.Nowgivemeotheronespoon?

(FromMartinBraine)

Theyalsodon’tseemtogetalotofexplicitinstruction.Andwhentheydo,theydon’treallypayattentiontothingsthatdon’timpactmeaning.

Sohowexactlydochildrenlearnallthis?

Whatthey’redoing:Extractingpatternsandmakinggeneralizationsfromthesurroundingdatamostlyjustbyhearingexamplesofwhat’sallowedinthelanguage.

Wecanalsothinkaboutthisasaninformationprocessingtask.

Wecanalsothinkaboutthisasaninformationprocessingtask.

Giventheavailableinput,

Input

Lookatthatkitty!There’sanotherone.

Wheredidhehide?Whathappened?

processing&generalization

Giventheavailableinput, informationprocessingdonebyhumanminds

Input

Lookatthatkitty!There’sanotherone.

Wheredidhehide?Whathappened?

Wecanalsothinkaboutthisasaninformationprocessingtask.

Input

Lookatthatkitty!There’sanotherone.

Wheredidhehide?Whathappened?

Giventheavailableinput, informationprocessingdonebyhumanmindstobuildasystemoflinguisticknowledge

syntax

words&morphemes

metricalphonology

syntacDccategories

semanDcs

pragmaDcs

processing&generalization

Wecanalsothinkaboutthisasaninformationprocessingtask.

Giventheavailableinput,tobuildasystemoflinguisticknowledge whoseoutputweobserve

Where’sthekitty?

Thatone’sreallycute.

syntax

words&morphemes

metricalphonology

syntacDccategories

semanDcs

pragmaDcs

processing&generalization

Input

Lookatthatkitty!There’sanotherone.

Wheredidhehide?Whathappened?

informationprocessingdonebyhumanminds

Wecanalsothinkaboutthisasaninformationprocessingtask.

Tounderstandhowchildrensolvethisacquisitiontask,weneedtothinkmoreaboutallthecomponentsinvolved.

Where’sthekitty?

Thatone’sreallycute.

syntax

words&morphemes

metricalphonology

syntacDccategories

semanDcs

pragmaDcs

processing&generalization

Input

Lookatthatkitty!There’sanotherone.

Wheredidhehide?Whathappened?

AframeworkthatmakescomponentsoftheacquisiDontaskmoreexplicit

AdaptedfromLidz&Gagliardi2015

Distinguishesbetweenthingsexternaltothechildthatwecanobserve(inputsignal,child’sbehavior)vs.thingsinternaltothechild(everythingelse).

Perceptualencoding:

AdaptedfromLidz&Gagliardi2015

Turningtheinputsignalintoaninternallinguisticrepresentation=perceptualintake.

Perceptualencoding:

AdaptedfromLidz&Gagliardi2015

Involvesusingcurrentknowledgeofthelanguage(thedevelopinggrammar)…

Perceptualencoding:

AdaptedfromLidz&Gagliardi2015

Involvesusingcurrentknowledgeofthelanguage(thedevelopinggrammar)deployedinrealtimetoparsetheinput…

Perceptualencoding:

AdaptedfromLidz&Gagliardi2015

Involvesusingcurrentknowledgeofthelanguage(thedevelopinggrammar)deployedinrealtimetoparsetheinput,oftendrawingonextralinguisticsystems(likeworkingmemory,auditoryprocessing,etc.)

AdaptedfromLidz&Gagliardi2015

syllableswithstress

Perceptualencoding

Highvs.Midvs.LowrelaRvepitch

speakeridenRty

Mainvs.secondarystress

wˈʌ ɾə pɹˈɪ ɾi kˈɪ ɾiMs s

L L H H M M

(Mom)

AdaptedfromLidz&Gagliardi2015

GeneratingobservablebehaviorInvolvesthecurrentlinguisticrepresentationsandthedevelopinggrammarbeingusedbytheproductionsystem.

AdaptedfromLidz&Gagliardi2015

GeneratingobservablebehaviorTheseareusedinrealtimetogeneratelinguisticbehavior(utterances)andnon-linguisticbehavior(pointing,looking,etc.).Thesebehaviorsrequirelinguisticsystems(utterancegeneration)andextralinguisticsystems(motorcontrol,attention,decision-making,etc.)

AdaptedfromLidz&Gagliardi2015

Inference=learning

Thisishowchildrenlearnfromthecurrentdatainordertoupdatethedevelopinggrammar.

AdaptedfromLidz&Gagliardi2015

Inference=learning

Constraintsonchildren’shypothesesandfiltersontheirattentioncausethemtoheedasubsetoftheperceptualintake—thisistheacquisitionalintake.

AdaptedfromLidz&Gagliardi2015

syllableswithstress

Highvs.Midvs.LowrelaRvepitch

speakeridenRty

Mainvs.secondarystress

wˈʌ ɾə pɹˈɪ ɾi kˈɪ ɾiMs s

L L H H M M

(Mom)wˈʌ ɾə pɹˈɪ ɾi kˈɪ ɾi

syllableswithstress

perceptualintake

acquisiDonalintake

AdaptedfromLidz&Gagliardi2015

Inference=learning

Inferencehappensovertheacquisitionalintake,usingextralinguisticabilities(statisticallearning,probabilisticinference,hypothesistesting,etc.)…

AdaptedfromLidz&Gagliardi2015

Inference=learningInferencehappensovertheacquisitionalintake,usingextralinguisticabilities(statisticallearning,probabilisticinference,hypothesistesting,etc.)togeneratethemostup-to-dateideasaboutthelanguage’sgrammar.

AdaptedfromLidz&Gagliardi2015

Thiswholeprocesshappensoverandoveragainthroughoutthelearningperiod

AdaptedfromLidz&Gagliardi2015

Aninformativecomputationalmodeloflanguageacquisitioncapturestheseimportantpiecesinanempirically-groundedway.

Theoretical

Corpus

Experimental

AdaptedfromLidz&Gagliardi2015

Whenwehaveaninformativecomputationalmodel,itwillconnectthechild’sinputtothechild’soutputinjustthisway.

External ✔

AdaptedfromLidz&Gagliardi2015

Wecanthenlook“underthehood”toseewhatinternalpiecesmadethatpossible—thispartishardtodoinrealchildren’sminds!

Internal

AdaptedfromLidz&Gagliardi2015

Upshot:Withcomputationalmodeling,wecanunderstandmorepreciselyhowthelearningstrategiesthatchildrenusework.

✔✔

speechsegmentaDon

Somethingswe’velearnedbymodel-buildingthisway

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

speechsegmentaDonSomethingswe’velearned

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

InvesDgaDngaBayesianinferencestrategyfortheveryearlystagesofspeechsegmentaDonoccurringaroundsixmonths

speechsegmentaDonSomethingswe’velearned

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

TheintuiDonofBayesianinference(appliedtospeechsegmentaDon)

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Thebestanswer(basedontheu$eranceyoujustheard)…

speechsegmentaDonSomethingswe’velearned

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Thebestanswer(basedontheu$eranceyoujustheard)dependsonyourpriorbeliefsaboutwhatgoodanswerslooklike…

TheintuiDonofBayesianinference(appliedtospeechsegmentaDon)

speechsegmentaDonSomethingswe’velearned

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Thebestanswer(basedontheu$eranceyoujustheard)dependsonyourpriorbeliefsaboutwhatgoodanswerslooklikeandhoweasilyananswerexplainsthedataobservedintheu$erance.

TheintuiDonofBayesianinference(appliedtospeechsegmentaDon)

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Mathemahcallyencodedpreferences: wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾi

Bayesianinference

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾiMathemahcallyencodedpreferences:

(1)Prefershorterwords

Bayesianinference

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾi

Mathemahcallyencodedpreferences:

(1)Prefershorterwords

(2)Preferlexiconswithfewerwords

Bayesianinference

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾi

Mathemahcallyencodedpreferences:

(1)Prefershorterwords

(2)Preferlexiconswithfewerwords

FindthebestsegmentaDon

Bayesianinference

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾi

Mathemahcallyencodedpreferences:

(1)Prefershorterwords

(2)Preferlexiconswithfewerwords

Bayesianinference

FindthebestsegmentaDonthatbalancesthesepreferences

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

wʌɾə pɹɪɾi kɪɾi

wʌ ɾə pɹɪɾikɪɾi

wʌɾə pɹɪɾikɪɾi

Mathemahcallyencodedpreferences:

(1)Prefershorterwords

(2)Preferlexiconswithfewerwords

FindthebestsegmentaDonthatbalancesthesepreferences

andcangeneratetheobservablefluentspeechuXerances

Bayesianinference

Somethingswe’velearned

Strategy:IdenDfyalistofwordforms(=lexicon)thatbestgeneratestheobservablefluentspeechuXerances

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Isitusefulforchildren?✓ModeledlearnerswithoutcognitivelimitationsontheirinferenceandmemorycanusethisstrategytosegmentfairlywellwhengivenrealisticEnglishchild-directedspeechdatatolearnfrom.

Bayesianinference

Theinferredlexicons,whilenotperfect,areveryusefulforsubsequentstagesoflanguageacquisition.

Somethingswe’velearned

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Isituseful?✓

ModeledlearnerswithcognitivelimitationsontheirinferenceandmemorycanstillusethisstrategyandsegmentEnglishquitewell.

Isituseablebychildren?

Bayesianinference

Somethingswe’velearned

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Isituseful?✓ Isituseable?✓

Doesitworkfordifferentlanguages?

Itsegmentswellforlanguageswithdifferentmorphologyandsyllableproperties:Spanish,Italian,German,Hungarian,Japanese,Farsi

Bayesianinference

Somethingswe’velearned

speechsegmentaDon

whatapre$yki$y!

= wʌɾəpɹɪɾikɪɾi wʌɾ ə pɹɪɾi kɪɾi

Phillips&Pearl2012,2014a,2014b,2015a,2015b,Pearl&Phillipsinpress

Isituseful?✓ Isituseable?✓ Doesitworkfordifferentlanguages?✓

Bayesianinference

BayesianinferenceseemstobeagoodproposalforaveryearlyspeechsegmentaDonstrategy.

Somethingswe’velearned

Recap

Languageacquisihonisanintereshngareaofresearchinhumancognihonbecauseit’sreallyhardandliXlehumansarereallygoodatit.

syntax

speechsegmentaDon

metricalphonology

syntacDccategorizaDon

syntax,semanDcs

pragmaDcs

RecapLanguageacquisihon

isintereshng

Tounderstandhowitworks,wecanbuildcognihvecomputahonalmodelsthatcapturetheimportantcomponentsoftheprocessandthenlookinsidetoseeexactlyhowtheywork

AdaptedfromLidz&

RecapLanguageacquisihon

isintereshng

Adaptedfrom

Modelscancaptureimportantcomponentsand

wecanlookinside

SomerecentfindingswiththisapproachsuggestBayesianinferenceisaplausibleearlyspeechsegmentahonstrategythat’suseful,useable,

andworksformanylanguages

Thankyou!

anotherone

KI$y

Noun

Everyki.ydidn’t…

Whodoes…ispre.y?

UCComputationalSocialScience