a memory-sensitive classification model of errors in early...
TRANSCRIPT
![Page 1: A Memory-Sensitive Classification Model of Errors in Early ...tetreaul/Presentations-and-Posters/0527.pdf• Test : the final 10% of exercises for each user Linguistic Memory Categorical](https://reader035.vdocuments.mx/reader035/viewer/2022062913/5e442c4c03dda3383766ed92/html5/thumbnails/1.jpg)
Acquiringasecondlanguage(L2)asanadultisnotoriouslydifficult.Byunderstandingwhereindividuallearnersmakemistakes,wecanimproveefficiencyanddurabilityofL2learning• Linguisticfactors:
• E.g.Cognates,concretewordsareeasier(deGroot&Keizer,2000)whileinterlingualhomographsareharder(Dijkstra,Timmermans&Schriefers,2000)
• Memoryfactors:• Sincelanguageislearned,itmustbestoredinmemory.
• Whatimprovesmemoryingeneralshouldalsoimprovememoryforlanguage
• Spacedrepetition:words(andotheritems)arerememberedbetterwhentheyareencounteredrepeatedly,withtemporalgapsinbetween(vs.repeatedallatonce).
• Longergapsarebetter(e.g.Cepeda etal.2006)• Robustoverseconds,minutes,days,weeks,years(e.g.Cepeda etal.
2008)• Appliestoawidevarietyofmaterials(e.g.Donovan&Radosevich,
1999)• Includinglanguage(e.g.Ullman&Lovelett,2018)
• RetrievalPractice:Recallinginformationfrommemorymakesthatinformationeasiertorecallinthefuture
• Duolingo frequentlypromptsuserstoretrievefrommemory• Retrievalpracticeenhancestheefficacyofspacedrepetition
BybetterunderstandingthefactorsthatinfluencelearningandretentionofL2,systemslikeDuolingo can:• DevotemoreresourcestothemostdifficultaspectsoftheL2(foreachlearner)• Schedulereviewoflearnedmaterialwhenitisofmostbenefittothelearner• Leveragetheirownusers’datatoimproveunderstandingofthelearningprocess,andin
turnimprovelearningoutcomes
A Memory-Sensitive Classification Model of Errors in Early Second Language Learning
INTRODUCTION
Brendan Tomoschuk, Jarrett T. Lovelett1University of California, San Diego
Contact: [email protected], [email protected]
ENGINEERED FEATURES
MODEL
RESULTS & DISCUSSION
DATA
REFERENCES
AUROC F1 Log-lossSLAMEnglish .7730 .1899 .3580
English .8286 .4242 .3191
SLAMFrench .7707 .2814 .3952
French .8228 .4416 .3561SLAMSpanish .7456 .1753 .3862
Spanish .8027 .4353 .3571
EnglishFrench
Spanish
userMeanError
userVarError
countries
timePerToken
days
userTrial
tokenMeanError tim
e
tokenVarError
format:prevFormat
lagTr1Tr2
nthOccurance
stemLag1
tokenLag1
stemLag1:stemLag2
lagTr1Tr2:morphoComplexity
stemLag1:stemLag2:lagTr1Tr2
stemLag2
tokenLag2
sentLength
0.000
0.025
0.050
0.075
0.100
0.000
0.025
0.050
0.075
0.100
0.000
0.025
0.050
0.075
0.100
Feature
Importance
EngineeredNewOld
Thefirst30daysofeachuserslearningbrokenarebrokendowninto:• Training:eachuser’sfirst80%ofsessions• Development:thenext10%ofeachuser’s
data• Test:thefinal10%ofexercisesforeachuser
Linguistic Memory Categorical InteractionsorthoLength Wordlength incharacters nthOccurance Number oftimesatokenhasbeenseen pos Partofspeech stemLag1xstemLag2
phonLength Wordlengthinphonemes userTrial Number oftrialsauserhasseen format Trialformat(seeFigure 1) stemLag1xstemLag2xlagTr1Tr2
orthoNei Numberoforthographicwordneighbors tokenLag1 Amountoftimesincetokenlastseen prevFormat Previoustrialformat lagTr1Tr2xmorphoComplexity
phonNei Numberofphonological wordneighbors tokenLag2 Amountoftimebetweenlasttime awordhasbeenseenandthetimebeforethat
client User’sclient(collapsed tomobileorweb) lagTr1Tr2xmorphoLag1
logWordFreq log-transformed wordfrequency stemLag1 Amountoftimesincestemmed tokenhasbeenseen
userMeanError Averageof a user’saccuracyacrosstrials FormatxprevFormat
logOrthoNeiFreq Average log-transformedwordfrequencyoforthographicneighbors
stemLag2 Amountoftimebetweenlasttime astemmedtokenhasbeenseenandthetimebeforethat
userVarError Variance inauser’saccuracyacrosstrials orthoNei xformat
logPhonNeiFreq Average log-transformedwordfrequencyofphonologicalneighbors
morphoLag1 Amountoftimesincemorphological featureshavelastbeenseen
phonNei xformat
Edit Distance Levenshtein distancebetweentranslationsofword
lagTr1Tr2 Amount oftimebetweenfirstandsecondtrialscontainingthattoken
formatxclient
Interlingual homograph Whether agiventranslationwasidenticaltoadifferentwordinthesourcelanguage
morphoComplexity xpos
morphoCompexity Numberofmorphological features
Concreteness Subject ratingsofhowperceptibleanentityis
Threegroupswereanalyzedseparately:• English-speakinglearnersofSpanish• English-speakinglearnersofFrench• Spanish-speakinglearnersofEnglish
POPULATIONS
THREESETS
ReverseTranslate ReverseTap Listen
Randomforestclassifier• Eachdecisiontreebranchedanumberoftimesequaltothesquarerootofthetotalnumberof
features• Anensembleof1000treeswascreatedforeachofthethreelanguagedatasets• Eachtreebrancheduntilleaveswerepure(containedonlyasinglelabel:“error”or“noerror”)• Out-of-bagerrorwasusedtoestimatepredictionerroroftheclassifier• TheclassifierwastrainedinPython3,usingsklearn.ensemble.RandomForestClassifier()
• Themoretimeusersspendpertoken(onaverage)withinanexercise(timePerToken)themorelikelytheyaretomakeerrorsinthatexercise
• Usersmakemoreerrorsonaveragethelongerthey’vespentusingtheapp(Days,userTrial).Perhapsbecauseitemdifficultyalsoincreaseswithexperience.
• Wordsthatrepeatmoreoften(nthOccurrence)arerememberedbetter.• Themoretimethatpassedsincethepreviousoccurrenceofaword,thehighertheerrorrate
(tokenLag1, tokenLag2)• Contraspacingeffect:perhapsmoreconsiderationoffullitemhistoryis
needed(orgapstoolong;seeCepeda etal.2008)• Thereseemstobeacosttoswitchingformats:errorratesarehigherwhenthecurrenttask
typeisdifferentfromtheprevious(format:prevFormat)• Futuremodelswillincludeablationexperimentsandwordembeddings
• Mostimportantfeatures: userMeanError; userVarError:• meanandvarianceofeachuser’serrorrate(undereachcombinationof
levelsofasmallsetoffeatures)• Computationalsavingsoverfittingamorecomprehensiverandomeffect
structure(i.e.randomeffectsforallusers,alltokens,andalluser-tokencombinations,atminimum)
Figure1.ExamplesofDuolingo exercisesanderrormarkingspresentinthedata
Figure2.Importancemeasuresforeachofthetop20features.
Table2.ModeloutcomescomparedtoSLAMbaselines.
Table1.Namesanddescriptionsoftheengineeredfeatures
BenAmbridge,AnnaL.Theakston,ElenaV.m.Lieven,andMichaelTomasello.2006.Thedistributedlearningeffectforchildren’sacquisitionofanabstractsyntacticconstruction.CognitiveDevelopment,21(2),174–193.HarryP.Bahrick andElizabethPhelphs.1987.RetentionofSpanishvocabularyover8years.JournalofExperimentalPsychology:Learning,Memory,andCognition,13(2),344–349.DavidA.Balota,JanetM.Duchek,andRondaPaullin.1989.Age-relateddifferencesintheimpactofspacing,lag,andretentioninterval.PsychologyandAging,4(1),3–9.MarcBrysbaert,AmyBethWarriner,andVictorKuperman.2014.Concretenessratingsfor40thousandgenerallyknownEnglishwordlemmas. Behaviorresearchmethods, 46(3),904-911.ShanaK.Carpenter.2009.Cuestrengthasamoderatorofthetestingeffect:Thebenefitsofelaborativeretrieval.JournalofExperimentalPsychology:Learning,Memory,andCognition,35(6),1563–1569.ShanaK.CarpenterandEdwardL.DeLosh.2006.Impoverishedcuesupportenhancessubsequentretention:Supportfortheelaborativeretrievalexplanationofthetestingeffect.MemoryandCognition,268-276. NicholasJ.Cepeda,HaroldPashler,EdwardVul,JohnTWixted,andDougRohrer.2006.Distributedpracticeinverbalrecalltasks:Areviewandquantitativesynthesis.PsychologicalBulletin,132(3),354-380.Nichlaos J.Cepeda,EdwardVul,DougRohrer,JohnT.Wixted,andHaroldPashler.2008.Spacingeffectsinlearning:Atemporalridgelineofoptimalretention.PsychologicalScience,19(11),1095-1102.WilliamL.Cull.2000.Untanglingthebenefitsofmultiplestudyopportunitiesandrepeatedtestingforcuedrecall.AppliedCognitivePsychology,14(3),215-235.AnnetteDeGroot andRineke Keijzer.2000.Whatishardtolearniseasytoforget:Therolesofwordconcreteness,cognatestatus,andwordfrequencyinforeign-languagevocabularylearningandforgetting. LanguageLearning, 50(1),1-56.TonDijkstra,MarkTimmermans,andHerbertSchriefers.2000.Onbeingblindedbyyourotherlanguage:Effectsoftaskdemandsoninterlingual homographrecognition. JournalofMemoryandLanguage, 42(4),445-464.JohnJ.Donovan andDavidJ.Radosevich.1999.Ameta-analyticreviewofthedistributionofpracticeeffect:Nowyouseeit,nowyoudon't.JournalofAppliedPsychology,84(5),795-805.HermannEbbinghaus.1964.Memory:Acontributiontoexperimentalpsychology(H.A.Ruger,C.E.Bussenius,&E.R.Hilgard,Trans.).NewYork,NY:Dover.(Originalworkpublishedin1885). JeffreyD.Karpicke andHenryL.Roediger.2007.Expandingretrievalpracticepromotesshort-termretention,butequallyspacedretrievalenhanceslong-termretention.JournalofExperimentalPsychology-LearningMemoryandCognition,33(4),704-719.ThomasK.Landauer andRobertA.Bjork.1978.OptimumrehearsalpatternsandnamelearningInM.M.Gruneberg,P.E.Morris,&R.N.Sykes(Eds.),Practicalaspectsofmemory(pp.625-632).London:AcademicPress.Viorica Marian,JamesBartolotti,SarahChabal,andAnthonyShook.2012.CLEARPOND:Cross-linguisticeasy-accessresourceforphonologicalandorthographicneighborhooddensities. PloS one, 7(8),e43230.CorneliusP.ReaandVitoModigliani.1985.Theeffectofexpandedversusmassedpracticeontheretentionofmultiplicationfactsandspellinglists.HumanLearning:JournalofPracticalResearchandApplications,4(1),11-18.B.Settles,C.Brust,E.Gustafson,M.Hagiwara,andN.Madnani.2018.SecondLanguageAcquisitionModeling.In ProceedingsoftheNAACL-HLTWorkshoponInnovativeUseofNLPforBuildingEducationalApplications(BEA).ACL.MichaelT.Ullman andJarrettT.Lovelett.2016.Implicationsofthedeclarative/proceduralmodelforimprovingsecondlanguagelearning:Theroleofmemoryenhancementtechniques.SecondLanguageResearch,39(1),39-65.EleanorVanderLinde,BarbaraA.Morrongiello,andCarolynRovee-Collier.1985.Determinantsofretentionin8-week-oldinfants.DevelopmentalPsychology,21(4),601–61.
Settlesetal.2018