Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Inferen&alorganiza&oningrammarsystems:Wordstructure,paradigmorganiza&onandlearningFarrellAckermanUniversityofCaliforniaatSanDiego
(RobMaloufSanDiegoStateUniversity)
1
1
Minélkülönösebbvalami,annálkevésberejtélyes- SherlockHolmeskalandjai
Thebigques+on:Howdowedemonstratethatthemorecuriousaphenomenonseems,thelessmysteriousitactuallyis?
Q1:Whatensuresthelearnabilityofcomplexinflec+onalsystems? H1:(Partly)theinferen+alorganiza+onthatemergesandinheresincomplexinflec+onalsystems:I(ntegra+ve)complexityasconstrainedbytheLowEntropyConjecture.
Contributoryques+ons:
Q2:Whatisthenatureofwordinternalstructure?Q3:Whatisthenatureofparadigmorganiza+on? Q4:Whatisthenatureoflearningasitrelatestoparadigmorganiza+on?
2
Outlineofthepresenta&on
AndtheendofallourexploringWillbetoarrivewherewestarted Andknowtheplaceforthefirst&me.T.EliotLiOleGiddingV
1.Theempiricalchallengeandtheproblemsitraises:Finnish
2.WordsandParadigmsasPart/Wholerela+ons:Uralicand WesternNilo+c
3.Quan+fyingrela+onsbetweenwords:PiteSaami
4.BacktoFinnish
5.Conclusions
3
1.Theempiricalchallengeandtheproblemsitraises
4
Afundamentallearnabilityproblem
5
5
Afundamentallearnabilityproblem
6
6
Afundamentallearnabilityproblem
7
7
Thebasicchallenge
Q1:HowcanaFinnishchildreliablypredicte.g.,thesingulartransla+ve1sgpersonpossessorkoirakseni`turnintomydog’givenknowledgeofe.g.,thenomina+vesingularkoirafortheLEXEMEcat,aswellasallcorrectlyproduceallotherformsofthislexeme?
Q2:Howdoessheextrapolatetothefullset(s)offormsforallotherLEXEMEs?
Alterna+vely,howdospeakersreducetheuncertaintyofpredic+ngonewordformfromknowledgeofanother(pa\ernof)wordforms?
8
8
Expressionsofdespair(seeBaermanToAppearonSeriforanotherbigchallenge)
AndersenonDinka(WesternNilo+c):(but,seeBaermanonNuer2012onrelatedlanguageNuer) Fromtheverybeginningoflinguis&cresearchonDinka,ithasbeennotedthatnumberinflec&onofnounsinthislanguageisirregular.Mi\erutzner(1866:15)andBeltrame(1880:22–24)statedthatthereisnogeneralruleforformingthepluralfromthesingular,andbothauthorsmadeobserva+onsaboutthetypesofphone+cdifferencesexis+ngbetweenthesingularformandthepluralformofanoun…Thatnumberinflec+onofsimplena+venouns,suchasthose…above,isindeedirregularandunpredictable,hasrecentlybeenestablishedbyLaddetal.(2009).Thepluralformcannotbepredictedfromthesingularform,norcanthesingularformbepredictedfromthepluralform,andthenumberinflec&onmayappeartobetotallyirregular.2014:226
Carstairs-McCarthyonPolish:
IfthePluralweretakenintoaccount,too,theamountofblurringcouldnotdecreaseandmightevenincrease,asissuggestedbydeBray'sgloomycomment(1980,p.273):"thestudenthastolearnforeach[noun],apartfromtheNom.sing.,atleasttheGen.sing.andNom.andGen.p1.aswell,andpreferablyalsotheDat.andLoc.sing.2000:818
VajdaonKet(EurasianIsolate)
ThelexicalentryofeachKetfiniteverbthereforecontains,inaddi&ontoitspurelyderiva&onalmorphemeshapes,aformulathatpredeterminestheconfigura&onofitsactantagreementmarkers.Thisformulacannotbepredictedinthegrammarbasedonanyoverallsetofsyntac&cfunc&ons,seman&croles,orotherformalstemfeatures....Thisstrategynecessitatestheunprecedentedclaimthatmostinflec+on-bearingposi+onsintheKetverbarespecifiedidiosyncra+callybythelexicalentryofeachindividualstem.Thoughfiniteverbmorphemesareagglu&na&veonaphonologicallevel,withseparatorelementso^enappearingbetweenthem,seman&callytheyexhibitanetworkofextendedandmul&pleexponencesthatrivalsthemostfusionaloflanguages.
9
9
Asimple(minded)answer
A:IfeachverbLEXEMEhasasinglestemandeachdis+nctmorphosyntac+ccategorye.g.,casehasauniquemarkerorexponentforeachofitsvalues,i.e.,NOM,TRANS...,thesolu+oniseasy:
Morphotac+cschema:[lexicalstemN-CASE-POSS]
1.koira-kse-nidog-TRANS-1SG.POSS `intomydog’
ReflectsCorbe\’sopera+veno+onofcanonicalinflec+onalencoding,i.e.,thelogicalextremeinthreedimensionsofwordencoding(Brownet.al.2013,Corbe\2015)
a)uniquemorphotac+cs, b)uniquestem,andc)uniquemorphologicalmarkerthatmodelse.g.agglu+na+vesystemslikeTurkish
Canonicalinflec+onalencodingfunc+onsasalocusfromthesimplestencodingpa\ernbetweenmorphosyntac+ccategoriesandformsfortrackingcross-linguis+cdevia+onsfromit:ithasnonorma&ve(oruniversal)theore&calstatus,simplyaheuris&ctaxonomiconeand,thereby,differsfrommanymorpheme-basedproposals(seeSiddiqi&Harleyeds.MorphologicalMetatheory(toappear)forcompe+ngviews)
10
10
Problemswiththesimpleanswer
StraighqorwardalignmentsofmorphologicalfeatureCONTENTandFORMaremuchrarerthanorensupposedandtheore+callymodeled(Ma\hews1991,Brownet.al.2013,Stump2001,2013,2016,Blevins2016,Harris(ToAppear)onmul+pleexponence),Corbe\2015,AckermanandBonami(ToAppear),among
others)and,
requisitesegmenta&onsofinternalstructureintomorphemicunitsandstructuralcons&tuentsareorenmorear+factualandindeterminatethanrecognized.(seeHocke\1987,Morpurgo-Davies1978,Ma\hews1991,Bochner1993,Anderson1990,Corbe\2009,2015,Stump2016,Blevins2016,BonamiandBenjamin2016,mongothers).
BothreflectLounsbury’s(1953)fic+veagglu+na+veanalogueasamodelforstructuralistmorphemicmorphology.
Relatedlyandworse,languagescommonly
(1)havemul&pleinflec&onal/declensionalclasses,and
(2)displayZipfiandistribu&onsofwordformswithinandacrosssthoseclasses.
11
11
Mul&pledeclensionclassesinFinnish(Paunonen1976,Thymé1993,Thyméet.al.1994,Ackermanet.al.2009)
Pa\ernsofformvaria+ondefinedifferentdeclensionalclasses:Compare(8),(32),and(10)withrespecttoNomSg,GenSg,andPartSg.
Q:Isn’tallthisvariabilityworsethanuseless,andarguablyanimpedimenttolearnability?
1.ThenumbersintheglosscolumnrefertodeclensionclassesaspresentedintheSoome-ees+sõnaraamat(Finnish-EstonianDic+onary)KaljuPihel&ArnoPikamäe(eds.)1999:758-771Tallin:Valgus.Anearlyexplora+onofPCFPasrelatestoFinnishisfoundinPauonen1976.
12
12
Itgetsworse:Zipfiandistribu&onofforms(Zipf1935,1949,Kornai1992,Yang2004,Chan2008,EllisandO’Donnell2011,Kurumadaet.al.2013,BonamiandBeniamine2015,Blevinset.al.2015,Finlay2015,LignosandYang(toAppear),Yang(toappear)amongothers)
Oneofthemostpuzzlingfactsabouthumanlanguageisalsooneofthemostbasic:wordsoccuraccordingtoafamouslysystema+cfrequencydistribu+onsuchthattherearefewveryhighfrequencywordsthataccountformostofthetokensintext(e.g.“a",“the",“I",etc.),andmanylowfrequencywords(e.g.“accordion",“catamaran",`ravioli".Piantadosi2015:1
Nosurprisethatinflec+onalmorphologyfollowsthesamedistribu+on:
Afewlexemesoccurfrequently,withskeweddistribu+onsoftheirinflectedforms,andmostlexemesandinflectedformsoccurrarely,ifatall.
13
Itgetsworse:Zipfiandistribu&onofforms(Yang2004,Chan2008,EllisandO’Donnell2011,LingnosandYang(toAppear),Kurumadaet.al.2013,BonamiandBeniamine2015,Blevinset.al.2015,Finlay2015,amongothers)
Ascorporaincreaseinsize,theydonotconvergeonuniformlypopulatedparadigms.Instead,theyreinforcepreviouslya\estedformsandclasseswhileintroducingprogressivelyfewernewitems.Thisdistribu+onreflectsthefactthatinflectedvariantsofopen-classitemsobeyZipf’slawatallobservedsamplesizes.Blevinset.al.ToAppear.
CzechNa+onalCorpusSYN2010(OlivierBonamip.c.)
100millionmorphologicallytaggedwords64,302dis+nctnounlexemes561,668dis+nctnounwordforms900,228possiblewordforms(7cases,2numbers)Only66lexemesoccurwithfullparadigmsNosingleformisobservedforeverylexemeOnly110lexemesoccurinthevoc.pl(butmorefrequentinspokenlanguage,sameformasnom.pl)
Largerstudiesrevealsimilardistribu+onsinFrenchandPortuguese.
14
Oldrecogni&onandsolu&onformul&pleclassesandpredic&onfromskeweddistribu&ons(Paul 1891, Anttila 1989, Wurzel 1989, Fertig 2013, among others)
Principalparts=defTheprincipalpartsofalexemeLareasetofcellsinL’srealizedparadigmfromwhichonecancategoricallydeducetheremainingcellsinL’srealizedparadigm.
(StumpandFinkel2013:11-13)
15
All4La+ndeclensionclassesarepar++onedinto4PrincipalParts
PrincipalPartsare(categorically)predic+veofallinflectedformsofagivenlexeme.
15
Ahypothesis:PrincipalPartsinthelightofInforma&onTheory
Morphologicalsystemsconsistofwordsthatareinforma+veaboutotherwords(Wurzel1987onimplica+onalrela+onsand“paradigmstructureconstraints”)
Informa+venessisquan+fiableintermsofcondi+onalentropyinInforma+on-Theory:wordsareinforma+veaboutotherwordstoaspecifiabledegree.
Thecategoricalityofpredic+venessdefini+onalofPrincipalPartsissimplyareflexofthepossibilitythatinsomesystemssomewordsareassociatedwithnounpredictability,i.e.arefullypredic+ve,withrespecttotheirrela+onstoother(formsof)words.(seeAckermanandMalouf2013,2015,BonamiandBenjamin(toappear)andBlevin(toappearfordiscussion)
16
16
Generalformula&onoftheimplica&onproblem:TheParadigmCellFillingProblem(PCFP)
Speakersoflanguageswithcomplexmorphologyandmul+pleinflec+onclassesconfrontalargelearningtaskwhosesolu+onraisesfundamentalques+onsaboutthestructureofwords,andtheorganiza+onofmorphologicalsystems.Thistaskreceivesageneralformula+onastheParadigmCellFillingProblem(PCFP)inAckermanetal.(2009):
PARADIGMCELLFILLINGPROBLEM:Givenexposuretoaninflectedwordformofanovellexeme,whatlicensesreliableinferencesabouttheotherwordformsinitsinflec+onalandderiva+onal)family?
NotmuchaproblemforlanguageslikeEnglish,withsimplemorphology,orTurkishwithagglu+na+vemorphologyandstraighqorwardmappingbetweenmorphosyntac+cproper+esandforms.
Acentralproblemofmorphologyandlearnability,thoughmuchoftheliteratureinmorphologyandlearnabilityfocusesonword(andwordinternalmorpheme)segmenta+on,leavingsomewhatmysteriouswhattodowithrespecttoproduc+vityoncemanyinflectedwordshavebeenlearned.(Saffranet,al.1996,Fedzechkinaet.al.2011,FinleyandWiemers2013,2015,Gerken2005,GerkenandKnight2015,Gerkenet.al.2015,GagliardiandLidz2014,LidzandGagliardi2015,andmanyothers.)
17
17
Uncertaintyreduc&on
Howdospeakersreliablyresolve/diminishuncertain+esintheselec+onoftheappropriateformforapreviouslyunencounteredword?
Theproblemseemsincreasinglydifficult:(i)thelargerthenumberofmorphosyntac+cproper+esalanguagecontains, (ii)thegreaterthenumberofallomorphicvariantsitusestoencodethem,and(iii)themoreextensivetheconjuga+onclassesandsubclasses,i.e.,dis+nc+vepa\erns,overwhichwordscanbedistributed.
Morphologicalcomplexityiscommonlycalculatedbyconsideringthesefactors
Ackerman&Malouf(2013)refertothisperspec+veonmorphologicaltypologyasEnumera&veComplexityorE-complexity:thisistheclassifica+onandquan+fica+onofmorphologicalphenomenabyreferencetofactors(i)–(iii)above,whichallfigureintheformalshapesofwords.(seeCarstairs-Mcarthy’sParadigmEconomyPrincipleandtheNoBlurPrinciplefora
highlyar+culatedformofthisview.)
18
18
Gesngtoasolu&onforPCFP(forsolu&onswithinadifferenttheore&calperspec&veseeAllenet.al.,Allen
Hypothesis1:Thecross-linguis+callyprofuseshapevaria+onintheinternalstructureofwordformsandtheexternalpa\ernsofrelatednessbetweenwordformsareconstrained(inpart)byimplica+veorganiza+on(=I(ntegra&ve)Complexity)amongwordsasquan+fiedinlowcondi+onalentropyvalues(=LowCondi&onalEntropyConjecture(LCEC).(AckermanandMalouf2013,2015)
Hypothesis2:TheLCECisonestrategyofaddressingthelearnabilityof(complex)morphologicalsystemsgiventhesparsenessofthedatareflectedinZipfiandistribu+onsofinflec+onalwordforms.
Hypothesis3:GiventhedemonstrableSparsenessofS+muliandtheRichnessoftheS+muli(=thecurseofdimensionality(AslinandNewport2009:17))morphologicalsystemsmustbeorganizedinwaysthatfacilitatelearnability.
Hypothesis4:Tounderstandmorphologicallearning,onemustunderstandinflec+onalorganiza+onandtheLCECisanemergentconstraintshaping(many)morphologicalsystems.(forsimilardirec+onsseeRaczet.al.2015,PaterandMoreton2012,AllenandHanson2016,Allen2016,AllenandBeckerms,PulleyblankandArchangeli(ToAppear),Wedel2007,2009).
19
19
Ageneralsolu&onharkingbacktotradi&onalWP
Q:Whatshouldmorphologistsmodel?
Hypothesis:Wordsandtheirpa\ernsoforganizedrela+ons(paradigmstructure)aretheprimaryobjectsofmorphologicalanalysis. Hocke\(1967)asks:IsacompactlyelegantIAanalysisofcomplexmorphologicalcomposi+onsinYawalmaniusefullyandinstruc+velyreplacedbythelesscompactWP?
Tocoverthecomplexalterna&onsofYawelmanibyprincipal-parts-and-paradigmswouldtakemuchmorespacethanisoccupiedinthefirstsec&onsofthispaperbythemorphophoneme-and-rewrite-rulepresenta&on.Buttherewouldbeanetgaininrealism,forthestudentofthelanguagewouldnowberequiredtoproducenewformsinexactlythewaythena&veuserofthelanguageproducesorrecognizesthem-byanalogy.Therewouldremainthisdifference:thesitua+onforthestudentisar+ficiallysimplified.Heisenabledtooperate,inhisanalogizing,intermsofaneatminimalsetofreferenceparadigmsandafixedpoint-of-departuresetofprincipalparts.Thena&veuserofthelanguage,ofcourse,doesnotdothis.Heoperatesintermsofallsortsofinternallystoredparadigms,manyofthemdoubtlessonlypar&al;andhemayfirstencounteranewbasicverbinanyofitsinflectedforms.Forthena&veuser,theformsthatwehaveforconvenienceselectedtobeour'principalparts'havenosuchfavoredposi&on.Theyareaslikelytobecreatedanalogically,asneeded,asareanyoftheotherforms.HockeO1967:221
20
Ageneralperspec&veharkingbacktotradi&onalWP
ThisconcurswithChomsky’sfavorableviewofWPoverIAandhisargueddispreferenceforthe“ill-advised”constructsrequiredinmorpheme-basedproposals:
“I know of no compensa+ng advantage for the modern descrip+ve reanalysis of tradi+onal paradigma+cformula+ons in termofmorphemesequences. This [=morphemicanalysis - FA] seems, therefore tobean ill-advised theore&cal innova&on... It seems that in inflec+onal systems, the paradigma+c analysis has manyadvantages and is to bepreferred… It is difficult to say anythingmoredefinite, since therehavebeen so fewaOemptstogivepreciseandprincipleddescrip&onofinflec&onalsystemsinawaythatwouldhavesomebearingonthetheore&calissuesinvolvedhere.”(Chomsky1965:174)
But,
a)51yearshavepassed.
b)Manyparadigma+capproacheshavebeendevelopedandappliedwithrigorandprecisionoverwidevarie+esofphenomenaandbroadarraysoflanguages,so
c)itisnolonger“difficulttosayanythingmoredefinite.”
21
…wordsarenotmerelywholesmadeupof parts,butarethemselvesconstruableas partswithrespecttosystemsofformsin whichtheypar&cipate. MaOhews(1991:204)
2.WordsandParadigmsasPart/Wholerela&ons:UralicandWesternNilo&c
22
Part/wholerela&ons:Theinternalstructureofwords
Internalcomposi&onofwords:wordsarecommonlythesmallestunitsofmeanings(seeRobins1959,Ma\hews1991,Blevins2016).
Cross-linguis+cimportanceofinternalstructureformorphologyisnotintheiden+fica+onofexponentsformeaningfulbits,i.e.,morphemesandorennotincons+tuentstructure,butinwaysthattheorganiza+onofexponentsfacilitatepa\ernsofdiscriminabilitythathelptodis+nguishandrelate(classesof)words.
Some+mes,thisinvolves(semi-)classicmorphemiccomposi+on,butmorecommonlyitrequiresconsideringwordsasrecombinantgestalts,i.e.,wholesconsis+ngofconfigura+onsofredeployedelements(segmental,suprasegmental)thateachalonedonotcontributeinvariantmeaningsindependentofthewordcontextsinwhichtheyoccur.
ExternalpaOernsofworddistribu&on:rela+onsbetweenwordsasinstruc+veasinstruc+veaboutmorphologicalorganiza+on.
23
23
Part/wholerela&ons1:Internalcomposi&onofwords
Hungarian(Uralic)nominalmorphology:
Stemadj-NOMLZ-POSS-CASEMorphotac+cs
1.bátor-ság-om-ról‘aboutmybravery‘
Hungarian:(classic)agglu+na+onofmorphemicsuffixestolexicalstems
24
24
Part/wholerela&ons1:Internalcomposi&onofwords
Mari(Uralic)verbalmorphology:singular2ndpastparadigmsforkol‘die’.
1stpastaffirma+ve1stpastnega+ve
1kolə̂-s-̆ə̂m`Idieds-̆ə̂mkolə̂`Ididn’tdie’
2kolə̂-s-̆ə̂c`youdied‘s-̆ə̂ckolə̂`youdidn’tdie’
3kolə̂-s̆`s/hedied‘ə̂s̆kolə̂`s/hedidn’tdie’
Synthe&cexpressionPeriphras&cexpression
Thesamepiecesdeployedindifferentmorphologicalconfigura+onsconveydifferentpolarityvaluesforverbs:theytakeondifferentfunc+onsinthewordcontextinwhichtheyoccur.
Theorganiza+on(gestalt)oftheelementsisasimportantastheelementsthemselves.
MariismorphologicallymoreE-complexthanEnglish,sinceithasmorphologicalmarkersforseveralpasttensesandperson/number,aswellasdifferentparadigmsforaffirma+veandnega+vepolarity.
25
25
Part/wholerela&ons1:Internalcomposi&onofwords
AgardialectofDinka(EasternNilo+c)nominalmorphology:Andersen(2014)(seealsoRemijsen2005,Laddet.al.2009,Storch2005,amongothers)
Caseandnumberdis+nguishedbywordinternalinterac+onsamongfourparameters:(i)vowellength,(ii)tone,(iii)voicequalityofthevowel,and(iv)vowelqualityalterna&ongrade.
Contrastsbetween(pairsof)wordsdisclosethepa\ernednatureofmorphologicalorganiza+on.
DinkaismorphologicallymoreE-complexthanEnglish,sinceithasmanymorestrategiesforexpressingnumbercontrasts. 26
Mirrorimagepa\erning:
Singularfor‘elbow’hastriplelengthforitsvowelsandlowtonewhilethisisthewordinternalpa\ernforthepluralof‘thigh’.Thedoublelengthvowelsandlowtoneforthesingularof‘thigh’parallelsthesamepa\ernforthepluralof‘elbow’.
26
Shilluk(WesternNilo&c)(Remijsen&Ayoker2014)
Shillukpresentsarichsystemofmorphologicalmarkingwithasmallsegmentalfootprint.Morphologicalexponenceischaracterizednotbytheconcatena+onofdiscretemorphemes,butratherbyastackingofmorphologicalopera+onswithinaconfineddomain,consis+ngofthestemandalimitedwindowofaffixes.Insuchasystem,morphologicalmarkingisrestrictedbythefactthatthestemsyllablecanhaveonlyonespecifica+oneachfortone,length,vowelquality,ATR,andthestem-finalconsonant,andthesespecifica+onsneedtoconveybothlexicalandmorphologicalinforma+on.Onewayinwhichtheuseoftheseresourcesismaximizedisthroughdistributedexponence,wherebyamorphophonologicalpa\ernrealizestwomorphosyntac+cvalues.Remijenet.al.ToAppear:17
27
27
SomeShillukverbclasses(Remijsen&Ayoker2014)
Remijsenet.al.ToAppear:17
28
28
Part/wholerela&ons1:Internalcomposi&onofwords
Firstdeclensionpar++vesinEstonian(Blevins2016:83) Alloftheformsarediscriminablydis+nct,exhibi+ngthesamepa\ernsofdis+nc+veness,includingpa\ernsofsyncre+smforPartSgandIllSg.
Firstdeclensionstemsandthemevowels
Therearenopiecesofformthatareuniquelyassociatedwithmorphosyntac+ccaseandnumberproper+es,butthereisasetofphonologicalresourcesthatgetreusedtocreatedis+nc+vewordshapes.
29
Dis+nc+veforms
29
Part/wholerela&ons1:Summary
Internalstructure:discrimina+vefunc+on,nota(necessarily)composi+onalfunc+on:dis+nguishingwordsfromoneanotherorviasyncre+smsiden+fyingpa\ernsofiden+ty.
Thewordisfrequentlythesmallestmeaningfulunitandthepa\ernsthatcharacterizeanddis+nguish(typesof)wordsmakeswordsdiscriminablefromoneanother:
Wordsaremadeupofreusedpiecesreconfiguredtoservedifferentpurposesandtofacilitatediscriminabilityamongwords:Gurevich(2006:44)onGeorgian
Themeaningofthewholewordlicensestheexponentstobeused,butthereisnoprecondi+onthatthemeaningsoftheexponentshavetocombinetocomprisethemeaningofthewhole.Composi&onalitymay,indeed,emerge,butasasideproductratherthanacentralprinciple,orperhapsasaneffec&velearningstrategy.Thewholeitselfmaycontributemeaningtothemeaningsoftheparts,ormayoverridethemeaningsoftheparts.
Theadop+onofwordsasindependentandnecessaryunitsofanalysisalsopermitswords,inturn,tobepartsoflargersystems.
Consequenceofpermi�ngwordstobecontrastedwithwords:thepossibilityofdiscoveringmorphologicalorganiza+on(paradigma+csystemsorinforma+onniches)inthesystemsofrela+onsbetweenwords.
30
30
Part/Wholerela&ons2:ExternalpaOernsofworddistribu&on
Blevins2006:549
Blevins2006:551
31
Pervasivepa\ernsofimplica+veinterdependencies
31
Cooldata,butsowhat?
Pesetsky(2009:464)expressesappropriateskep+cismaboutsimilar“cabinetsofcuriosi+es”iden+fiedinEvansandLevinson’s(2009)againstlanguageuniversals,sincetheimportanceofunusualvariantsoffamiliarconstruc+onsw/oanalysisishardtoevaluate.
So,whatsortofanalysiscanbedeveloped?
32
32
3.Quan&fyingrela&onsbetweenwords:PiteSaami
33
33
PiteSaami
AllnaturallanguagesshowacertaindegreeofwhatBaermanetal.(2010:2)call“gratuitous”morphologicalcomplexityandWurzel(1986:76)describesas“ballast”inthelinguis+csystem.
PiteSaami(Wilbur2014:109)
34
PiteSaamihaseightnominaldeclensionsshowingdis+nctgradeandsuffixpa\erns.
Sincetheassignmentoflexicalitemstopar+culardeclensionsislargelyarbitrary(thoughinfluencedbyphonologicalfactors),theseclassesaddcomplexitytotheinflec+onalsysteminawaythatservesno(evident)communica+vepurpose.
34
ExtractedpaOernofformvaria&ondistribu&on
PiteSaami(Wilbur2014:102)
Pa\ernofstemandaffixvaria+on(stemspar++ontheinflec+onalfeaturespace(BonamiandBoyé
2006,2007))
35
35
Informa&on-Theore&cinsights
Q:Whatpoten+atesthelearnabilityofthesystem?
A:Systemicrela+onsamongwordsorganizedintermsoflowcondi+onalentropies,i.e.,thepredictabilityofanunknownform,givenknowledgeofanotherformofthatword.
H:Quan+fy“predic+on”orasareduc+oninuncertainty,orinforma+onentropy(Shannon1948)
Complexinflec+onalsystemsareorganizedinwaysthatfacilitatethereliableguessingfromknownformstounknownformsandthisorganiza+oncanbemadevisiblebyusingInforma+on-Theore+cmeasuresofEntropyandCondi&onalEntropy.
Whatweneedtoknow:
1)TheParadigmCellEntropyforeachcelli.e.,howthechoicesofexponentsforeachcellcanbecalculatedintermsofthedegreetowhichachoicebetweenthemisuncertain.
36
36
Informa&on-Theore&cinsights3
(2)theCondi&onalEntropybetweenpairsofcells,sothatwecandeterminehowmuchtheknowledgeoftheforminonecellreducestheuncertaintyassociatedwithselec+ngapairedcell,H(Y|X)
Howmuchinforma+ondoesthepresenceofXhaveforpredic+ngY,orhowsurprisedarewethatgivenXwegetY?
What’simportantissurprisal:measureoftheamountofinforma+onexpressedbyapar+cularoutcomemeasuredinbits,where1bitisachoicebetween2equiprobableoutcomes.3
Theintui+on:Outcomeswhicharelessprobable(hardertopredictandmoreuncertain)havehighersurprisal,
Surprisalis0bitsforoutcomeswhichalwaysoccur(p(x)=1)andapproaches∞forveryunlikelyevents(asp(x)approaches0.
3)TheAverageCondi&onalEntropyastheaveragedsumofallofthepairwisecondi+onalentropiesprovidesuswithageneral“interpredictablitybetweenforms”measure.
3.Forpresentpurposeswewillassumeallofthetasksbelowreflectequiprobabledistribu+ons.Thisactuallyprovidestheupperboundofentropyvalueswithfrequencyinforma+onand/orseman+candphonologicalcondi+onslikelyreducingentropy(seereferencesforexplora+onsalongtheselines)
37
37
Calcula&ngParadigmCellEntropy
Compara+veDeclensionClassesinPiteSaami(Wilbur2014:109)
38
Cellscanberealizedbydifferentnumbersofallomorphs:Illa+vesingularhas5allomorphs.
Givenequiprobabilityassump+ons,themorethenumberofchoices,thehighertheentropy.(formorerealis+cassump+onsseeAckermanet.al.2009,Bonami2015,Sims2016,amongothers)Eachmemberofarealiza+onsetisassociatedwithanentropyvaluereflec+ngthedegreeofsurpriseofitsselec+on.
38
An(un)realis&ctaskanditsentropies:ParadigmCellEntropy
Aspeaker’stask:Guessthewordformforaspecificcell.
Thiswilldependonthenumberofpossiblerealiza+onsforeachcell:
Illa+vesingularonlyhasfivepossiblerealiza+onsandanentropyof2.250bits,whilethemostdiversecellshaveanentropyat3.00bits(=8choices)
Averageentropyacrossallcellsis2.66;thisaverageisameasureofhowdifficultitisforaspeakertoguesstherealiza+onofanyonewordformofanypar+cularlexemeintheabsenceofanyinforma+onaboutthatlexeme’sdeclension=ParadigmcellentropyAnentropyof2.66bitsisequivalenttoselec+ngamongonly22.26=6.31equallylikelyalterna+ves:
PiteSaamihaseightdeclensions,butselec+ngtherealiza+onforapar+cularwordformofalexemeisasdifficultasachoiceamongali\lemorethansixequallylikelyalterna+ves.
39
39
Arealis&cspeaker’stask:Condi&onalEntropy
Quan+fyingthepredictabilityofoneformgiventheother:measurethesizeofthesurpriseassociatedwiththeseformsusingcondi+onalentropyH(Y|X),theuncertaintyinthevalueofYgiventhatwealreadyknowthevalueofX:
Itmeasureshowmuchentropyremainsforagivenforminagivencellifaforminanothercellisalreadyknown.
Thepoint:ThesmallerH(Y|X)is,themorepredictableYisonthebasisofX,i.e.,thelesssurprisedoneisthatYisselectedgivenknowledgeofX.
WhereXcompletelydeterminesY,thecondi+onalentropyX(Y|X)is0bits:giventhevalueofX,thereisnoques+onremainingastowhatthevalueofYis.(ThinkofPrincipalParts)
But,ifXgivesusnoinforma+onaboutYatall,thecondi+onalentropyX(Y|X)isequaltoH(Y):giventhevalueofX,wearejustasuncertainaboutthevalueofYaswewouldbewithoutknowingXatall.
40
40
AverageCondi&onalEntropiesinPiteSaami
41
ForPiteSaamiexample,H(P)is0.116bits,equivalenttoachoiceamongonly20.116=1.08equallylikelydeclensions.
WhilePiteSaamihaseightnominaldeclensionsfromthepointofviewofalexicographertryingtodescribethelanguage,foraspeakertryingtousethesystemithasonaverageonlyslightlymorethanone:thisistheI(ntegra+ve)-complexityofthisparadigm.
Providesaquan+ta+vemeasureofthedescrip+onsandinsightsaboutparadigmstructure(Wurzel1989,Ma\hews1991)andawaytocalculateI-complexityofsystem.
41
Iden&fyingdegreesofinforma&vity
42
ColumnsaveragesE[col]areameasureofpredictedness,orhowdifficultitistoguesstherealiza+onofacell(onaverage)givenknowledgeofsomeothercell
Rowaveragesindicateacell’spredic&veness:theaverageuncertaintyinanotherparadigmcellgivenknowledgeofthatcell.
Thenomina+vesingularisverypredic+vebuthardertopredict:onitsbasisallotherformsarecompletelypredictable,makingitaprincipalpartintheclassicalsense.
42
I-complexitygeneraliza&onsacrosslanguages?2
TheLowEntropyConjectureisthepredic+onthatlanguageswillshowlowaveragecondi+onalentropyalongthelinesdiscoveredfortheselanguages.
Ackerman&Malouf(2013)refertothismeasureascalcula+ngtheI(-ntegra+ve)-complexityofamorphologicalsystem:itismeasureoftransparencyoftherela+onsofpa\ernsofwordsinparadigm.
2.Addi+onalconfirmingempiricalresults:PiteSaami(AckermanandMaloufabove),Estonian(Blevins,Baerman),TundraNenets(Ackermanetal.),Tlingit(Cable),Murrinhpatha(MansfieldandNordlinger),Portuguese(BonamiandLuis),French(BonamiandBeniamine),Nuer(Baerman),Voro(Baerman),PalantlaChinantecandKadiweu(SimsandParker).
43
43
Observa&ons
TheLCECreflectsastrategyforhowcomplexmorphologicalsystemsareorganizedinawaytoinsurelearnability,despiteluxuriantE-complexity(manydifferentforms,manydifferentclasses)andskewedZipfiandistribu+onofsparses+muli.
Representsthediscoveryofarobustprinciple,asta+s+callanguageuniversal,ofcross-linguis+cmorphologicalorganiza+onthatonlybecomesclearwhentheinternalstructureofwordsandtheirexternaldistribu+onsarerecognizedasprimarytheore+calobjectsofanalysis-wordsarenotepiphenomena.
Morphologicalsystemsmustbesimpleinwaysthatallowthemtobelearnedandusedbyna+vespeakers,irrespec+veofhowcomplexwordsandparadigmsmayappearaccordingtoexternalmeasures.
Speakersmustgeneralizebeyondtheirdirectexperience:
Morphologicalsystemsmustpermitspeakerstomakeaccurateguessesaboutunknownformsoflexemesbasedononlyasmallsampleofknownforms.
ThisistheIntegra&veComplexityofasystem:therela+veinforma+vityassociatedwitheachformandhowthisdefineslanguagepar+cularpa\ernsofinterpredictabilitybetweenforms.
44
Animportantobserva&onconcerningwordinternalstructure
Resultsdonotdependon(universal)assump+onsaboutpar+cularformalproper+esofwords,onlyonwhetheraspeakercandiscriminatebetweentwoforms:
specificformtype,eithersynthe+corperiphras+c,andthemannerofdiscrimina+on(whetherbyaffixes,tones,stress,ablaut,orensemblesofthese)isirrelevant,aslongasfortheformsarediscriminablefromoneanother.
45
Animportantobserva&onconcerningwordinternalstructure
A(hypothe+cal)straighqorwardlyagglu+na+velanguagehasanaveragecondi+onalentropyof0bits,asexpected:
asdoesahypothe+calfusionallanguage.
46
Organizedoraccidentalmappings?MonteCarloSimula&ons
Language Declensions Cells Realiza&ons Paradigmentropy
BootstapAvg
Bootstrapp
Fur 19 12 80 0.517 1.316 0.001Pite-Saami 8 14 70 0.116 0.322 0.001Russian 4 12 26 0.538 0.541 0.383
Fur
Resultsbasedonuniformtypefrequencies(MaloufandAckerman2010,2013,2015)
Pite-Saami
Simula+onscomparetheentropiesassociatedwithactual
47
4.BacktoFinnish
48
Finnishimplica&veorganiza&on(Ackermanet.al.2009)
Givens+mulustuohta‘birchbark(part.sg)’,thereiscorrectassignmenttoclass32basedontheanalogicalpropor+onkuusta:tuohta::kuusi:TUOHI.
Ifthes+mulusisanon-diagnos+cwordform,correctclassassignmentisunderdetermined.S+mulusnuken‘puppet(gen.sg)’couldbeassignedeithertoclass9orclass8,basedonthecompe+nganalogicalpropor+onsnallen:nuken::nalle:NUKKEversusoven:nuken::ovi:NUKKI.
Ifthes+mulicomprisethepairnuken‘puppet(gen.sg)’andnukkeja‘puppet(part.pl)’,thentherecorrectassignmentofthiswordtoclass9.(=jointentropy(BonamiandBeniamine2015,
1.ThenumbersintheClasscolumnrefertodeclensionclassesaspresentedintheSoome-ees+sõnaraamat(Finnish-EstonianDic+onary)KaljuPihel&ArnoPikamäe(eds.)1999.Tallin:Valgus.
49
49
Finnishimplica&veorganiza&on(Ackermanet.al.2009)
50
Most Predicted
Most Predictive
Most Predicted
Most Predictive
50
5.Conclusions
51
Currentdirec&ons
(1)iden+fylargerdatasetssupplementedwithfrequencyinforma+ontoserveasobjectsofmeasurementtoreplacethemeasurementofformsderivedfromdescrip+vegrammars,aswellasdevelopingappropriatetoolsfortheirmeasurement(Bonami2014,Boye
ms.,Sims2015,Bonami&Beniamine2015,BaermanonSeri(toappear),amongothers)
(2)morecarefullyexplorethenatureofthephonological/phone+cs+mulicons+tu+veofwordinternalstructureand,moregenerally,inquireabouttheappropriateformsthatwordsasobjectsofanalysisshouldtake=morphophone+cs(Lehiste1972,Kempsetal.2005,Blazej&Cohen-Goldberg2015,Seyfarthetal.2015,Plaget.al.2015,amongothers)
(3)explorehowanalogicalinferencemayrelyuponimplica+veorganiza+oninthelearningofcomplexmorphologicalsystems(Baayen&Ramscar2015,Ramscaretal.2015),and
(4)iden+fycross-linguis+cgeneraliza+onsconcerningpossibleconstraintsontheorganiza+onofmorphologicalsystems(Stump&Finkel2013,Baermanetal.2015,Ackerman&Malouf2015,Sims&Parkertoappear).
52
52
AndtheendofallourexploringWillbetoarrivewherewestarted Andknowtheplaceforthefirst&me.T.EliotLiOleGiddingV
Thebigques+on:
Q1:Whatensuresthelearnabilityofcomplexinflec+onalsystems? A1:(Partly)theinferen+alorganiza+onthatemergesandinheresincomplexinflec+onalsystems:I-complexityasconstrainedbytheLowEntropyConjecture.
Contributoryques+ons:
Q2:Whatisthenatureofwordinternalstructure?A1:Discrimina+veQ3:Whatisthenatureofparadigmorganiza+on? A3:Implica+ve Q4:Whatisthenatureoflearningasitrelatestoparadigmorganiza+on? A4:Analogical(implicitintoday’spresenta+on,butseepa\erngeneraliza+onliteratureonsta+s+cal
learning)
53
Asurprisinglysimplecase
InChiquihuitlánMazatec,verbsaremarkedforpersonandaspectbyacombina+onoftones,finalvowel,andstemforma+ve(Jamieson1982,Capen1996,Baerman&Corbe\2010)
Posi+veparadigmforba3se2‘remember’
NEUTRAL INCOMPLETIVE
SG PL SG PL
1INCL ča2sẽ2 ča2sẽ42
1 ba3sæ1 ča2sĩ24 kua3sæ1 ča4sĩ24
2 ča2se2 ča2sũ2 ča4se2 ča4sũ2
3 ba3se2 kua4se2
54
NEUTRAL INCOMPLETIVE
SG PL SG PL
1INCL 2-2 4-421 3-1 2-24 3-1 4-242 2-2 2-2 4-2 4-2
3 3-2 4-2
NEUTRAL INCOMPLETIVE
SG PL SG PL
1INCL -ẽ -ẽ
1 -æ -ĩ -æ -ĩ2 -e -ũ -e -ũ3 -e -e
NEUTRAL INCOMPLETIVE
SG PL SG PL
1INCL ča- ča-1 ba- ča- kua- ča-2 ča- ča- ča- ča-3 ba- kua-
-s-‘remember’
ToneclassB31
Finalvowel-e
Stem-forma+ve11
55
Implica&onalrela&ons
Eachoftheseseparateinflec+onalsystemsshowconsiderablecomplexity
Eachlexicalitemisamemberofsomeconjunc+onineachofthesethreesystems
Therearepoten+ally6×10×18=1,080meta-conjuga+ons
Baerman&Corbe\reportthat109area\estedinCapen(1996)
56
Implica&onalrela&onsTheseappeartobeindependentsystemsofinflec+onclasses
ByBaerman&Corbe\’scount,mostjointconjuga+onshaveonlyoneortwomembers;themostfrequenthas22
Knowingwhichclassalexemebelongstoinonedimensionprovidesrela+velyli\leinforma+onaboutanotherdimension:
Expectedentropyforchoosingaclassinadimensionis2.469bits
Expectedcondi+onalentropyforchoosingaclassinadimensionknowingtheclassinanotherdimensionis2.154bits
Jamiesonoffersdiachronicexplana+onsforthedevelopmentofthiscomplexity,buthowisitmaintained?
57
Implica&onalrela&onsConsideralessabstractproblem:giventhestemforma+ve,finalvowel,andtonepa\ernofawordform,guessthestemforma+ve,finalvowel,andtonepa\ernforsomeotherwordform
Thisturnsouttobemucheasier:fortheposi+veneutralforms,theexpectedentropyis4.920bitsbuttheparadigmentropyisonly0.709bits
Everywordformprovidesinforma+onaboutallthreedimensions
Jamieson’sinflec+onclassesshowahighdegreeofinter-paradigmsyncre+sm,solis+nglexemesbyclassgreatlyoverstatesthevaria+on
ComparedtoModernGreek,wri+ngadic+onaryofChiquihuitlánMazatecissignificantlyharder(E-complexity),butspeakingitisn’t(I-simplicity)
58
Somecaveats
Entropycalcula+onsdependonmany,manyassump+ons
Iden+fica+onandenumera+onofforms
Frequenciesoflexemesandwordforms
Choiceof(sub-)paradigms
Generalizingfromasingle,randomlyselectedform
Thenumbersshouldbeinterpretedwiththisinmind
Whatisclear,however,isthatparadigmentropiesaremuchlowerthantheycouldbe
59
Tes&ngentropyclaims
Theimplica+onalstructureoftheparadigmsiscrucialtoreducingparadigmentropy
Howcanwetestthis?
Nullhypothesis:ParadigmentropyoflanguageLisindependentofparadigmorganiza+on
Ifthisistrue,thenL0,aversionLwiththesameformsandthesameclassesbutadifferentorganiza+on,shouldhavemoreorlessthesameparadigmentropy
Bootstraptest:samplewithreplacementfromthespaceofpossibleL0’s,andcomparetotheobservedL
60
Chiquihuitlán Mazatec
Paradigm entropy (bits)
Per
cent
of T
otal
0
5
10
15
20
0.7 0.8 0.9 1.0 1.1 1.2
61
Language Cells Realiza&ons Declensions Declensionentropy
Expectedentropy
Paradigmentropy
BootstapAvg Bootstrapp
Amele 3 31 24 4.585 2.882138071081.105330679521.32749714946 0.001Arapesh 2 41 26 4.700 4.070538035410.6299016827340.629901682734 1Burmeso 12 24 2 1.000 1 0 0 1Fur 12 80 19 4.248 2.394762454520.5168808443561.31619569963 0.001Greek 8 12 8 3.000 1.621 0.644 0.891 0.001Kwerba 12 26 4 2.000 0.8639723955630.428030303030.522673331721 0.001Mazatec 6 356 109 6.768 4.92 0.709 1.1 0.001Ngi+ 16 68 10 3.322 1.936641648330.4839928968691.01926185459 0.001Nuer 6 12 16 4.000 0.8636758201390.7926130922220.810940247458 0.16Russian 12 26 4 2.000 0.9111528644480.5375256392580.541181039402 0.383
62
Externalfactors
Amele(Roberts1987)isdescribedinWALSashaving31differentclassesofpossessivesuffixesplusapostposi+on
HeinandMüller(2009)arguethatfactoringoutphonologicallypredictablealterna+onsreducesthisto23suffixedclasses
H&M’sparadigmshaveanentropyof1.105bits!
But,somefacts:
Possessivesuffixesonlyapplytoaclosedclassof109inalienablypossessednouns
Acombina+onofalmost(butnotquite)categoricalseman+candphonologicalpa\ernsgeneratemostoftheclasses
Manyclasseshaveonlyasinglemember
63
Prospects
ParadigmentropymeasuresthecomplexityofaparadigmwithrespecttotheParadigmCellFillingProblem
TherearemanywaysthatmorphologicalsystemscanbeE-complex,but(perhaps)onlyonebasicprincipleofI-Simplicity,thoughmanywaystogetthere.
Ques+ons:
Whatistherangeofparadigmentropiesinrealtypologicallydiverselanguages?
Whatarethewaysthatparadigmscanbeorganizedtomanagecomplexity(andkeepparadigmentropylow)?
Arethereotheraspectsofmorphologicalsimplicitythatcanbequan+fied?
64