[PDF] Automatic Identification of Cognates and False Friends in French

View - Download Automatic Identification of Cognates and False Friends in French

Previous PDF

Next PDF

inFrenchandEnglish

DianaInkpenandOanaFrunza

SchoolofInformationTechnologyandEng.

UniversityofOttawa

Ottawa,ON,K1N6N5,Canada

DepartmentofComputingScience

UniversityofAlberta

Edmonton,AB,T6G2E8,Canada

kondrak@cs.ualberta.ca

Abstract

Cognatesarewordsindierentlanguagesthat

havesimilarspellingandmeaning.Theycan helpasecond-languagelearneronthetasksof vocabularyexpansionandreadingcomprehen- sion.Thelearneralsoneedstopayattention topairsofwordsthatappearsimilarbutarein factfalsefriends:theyhavedierentmeaning weproposeamethodtoautomaticallyclassify apairofwordsascognatesorfalsefriends.We focusonFrenchandEnglish,butthemethods areapplicabletootherlanguagepairs.Weuse featuresforclassication.Westudytheimpact andcombiningthemthroughmachinelearning techniques.

Keywords:similaritymeasures,machine

learning,cognatesandfalsefriends,second- languagelearning,machinetranslation.

1Introduction

Whenlearningasecondlanguage,astudentcan

benetfromknowledgeinhis/herrstlanguage (Gass87)(Ringbom87).Cognates{wordsthat arealsopairsofwordsthatappearsimilar,but havedierentmeaninginsomeorallcontexts: 96).

Cognateshavealsobeenemployedinnatural

sentencealignment(Simardetal.92;Melamed

99),inducingtranslationlexicons(Mann&

drak&Dorr04).Allthoseapplicationsdepend onaneectivemethodofidentifyingcognatesby computinganumericalscorethatre ectsthelike- tures.Thenweexplorevariouswaystocombine ingtechniquesfromtheWekapackage(Witten& asCognates;otherwise,theyareassumedtobe

False-Friends.

AlthoughFrenchandEnglishbelongtodier-

entbranchesoftheIndoeuropeanfamilyoflan- guages,theyshareanextraordinaryhighnum- berofcognates.Thecognatesderivefromsev-

LatinandGreekoriginthatpermeatethevocab-

veryold,\genetic"cognatesgobackalltheway toProto-Indoeuropean,e.g.,mere-motherand pied-foot.Othercognatescanbetracedtothe lapseoftheRomanEmpire,andbytheperiodof

FrenchdominationofEnglandaftertheNorman

conquest.

WhileourfocusisonFrenchandEnglish,the

methodsthatwedescribearealsoapplicableto otherlanguagepairs.Nowadays,newtermsre- latedtomoderntechnologyareoftenadopted hasthesamemeaning.

2RelatedWork

Previousworkonautomaticcognateidentica-

tionismostlyrelatedtobilingualcorporaand translationlexicons.Simardetal.(Simardet al.92)usecognatestoalignsentencesinbi- texts.Theyemployaverysimpletest:French-

Englishwordpairsareassumedtobecognatesif

McKelvie(Brew&McKelvie96)extractFrench-

sures.MannandYarowsky(Mann&Yarowsky thebasisofcognatepairs.Theyfoundthatedit hiddenMarkovmodelsandstochastictransduc- guagesbycombiningthephoneticsimilarityof

Kondrak&Dorr04)reportthatasimpleaverage

performsallindividualmeasuresonthetaskof theidenticationofdrugnames.

ForFrenchandEnglish,substantialwork

oncognatedetectionwasdonemanually.

LeBlancandSeguin(LeBlanc&Seguin96)

(Dubois81).6,447ofthecognateshadidenti- cognatesappeartomakeupover30%ofthevo- cabulary.

Theuseofcognatesinsecondlanguageteach-

versionfromEnglishtoFrenchwerealsoproved rules.Anexampleis:cal!queinpairssuchas

3.1Denitions

Weadoptthefollowingdenitions.Thedeni-

arepairsofFrenchandEnglishwords,respec- tively.

Cognates,orTrueFriends(VraisAmis),are

pairsofwordsthatareperceivedassimilarand reconnaissance.

FalseFriends(FauxAmis)arepairsofwords

intwolanguagesthatareperceivedassimilar buthavedierentmeanings,e.g.,main\hand"- main,blesser\toinjure"-bless.

PartialCognatesarepairsofwordsthathave

thesamemeaninginbothlanguagesinsomebut notallcontexts.Theybehaveascognatesoras ineachcontext.Forexample,inFrench,fac- whileetiquettecanalsomean\label".

GeneticCognatesarewordpairsinrelated

intheancestor(proto-)language.Becauseof periodsoftime,geneticcognatesoftendierin formand/ormeaning,e.g.,pere-father,chef- otheratsomepointoftime,suchasconcierge.

Unrelatedpairsarewordsthatexhibitnoor-

e.g.,glace-chair.

3.2OrthographicSimilarityMeasures

subjective.Inthissection,webrie ydescribethe measuresthatweuseasfeaturesforthecognate classicationtask.

IDENTisabaselinemeasurethatreturns1

ifthewordsareidentical,and0otherwise.

PREFIXisasimplemeasurethatreturnsthe

lengthofthecommonprexdividedbythe lengthofthelongerstring.1E.g.,thecom- ofSimardetal.(Simardetal.92)approach. monprexforfactoryandfabriquehaslength

2(thersttwoletters)which,dividedbythe

lengthof8,yields0:25.

DICE(Adamson&Boreham74)iscalcu-

latedbydividingtwicethenumberofshared letterbigramsbythetotalnumberofbigrams inbothwords:

DICE(x;y)=2jbigrams(x)\bigrams(y)j

jbigrams(x)j+jbigrams(y)j wherebigrams(x)isamulti-setofcharac- terbigramsinwordx.E.g.,DICE(colour, couleur)=6/11=0.55(thesharedbigrams areco,ou,ur).

TRIGRAMisdenedinthesamewayas

DICE,butemploystrigramsinsteadofbi-

grams.

XDICE(Brew&McKelvie96)isalsodened

inthesamewayasDICE,butemploys\ex- tendedbigrams",whicharetrigramswithout themiddleletter.

XXDICE(Brew&McKelvie96)isanexten-

sionoftheXDICEmeasurethattakesinto accountthepositionsofbigrams.Eachpair ofsharedbigramsisweightedbythefactor: 1

1+(pos(a)pos(b))2

wherepos(a)isthestringpositionofthebi- grama.2

LCSR(Melamed99)standsfortheLongest

CommonSubsequenceRatio,andiscom-

putedbydividingthelengthofthelongest commonsubsequencebythelengthofthe longerstring.E.g.,LCSR(colour,couleur) =5/7=0.71

NEDisanormalizededitdistance.Theedit

distance(Wagner&Fischer74)iscalculated bycountsuptheminimumnumberofedit operationsnecessarytotransformoneword intoanother.Inthestandarddenition,the anddeletions,allwiththecostof1.Anor- malizededitdistanceisobtainedbydividing thetotaleditcostbythelengthofthelonger string. proximationtophoneticnamematching.

SOUNDEXtransformsallbuttherstletter

tonumericcodesandafterremovingzeroes

Forthepurposesofcomparison,ourimple-

mentationofSOUNDEXreturnstheeditdis- tancebetweenthecorrespondingcodes.

BI-SIM,TRI-SIM,BI-DIST,andTRI-DIST

belongtoafamilyofn-grammeasures(Kon- drak&Dorr04)thatgeneralizeLCSRand

NEDmeasures.Thedierenceliesincon-

sideringletterbigramsortrigramsinstead ofsingleletter(i.e.,unigrams).Forexam- ple,BI-SIMndsthelongestcommonsub- sequenceofbigrams,whileTRI-DISTcalcu- latestheeditdistancebetweensequencesof trigrams.n-gramsimilarityiscalculatedby theformula: s(x1:::xn;y1:::yn)=1 nP n i=1id(xi;yi) whereid(a;b)returns1ifaandbareidenti- cal,and0otherwise.

4TheData

pairsofFrenchandEnglishwords(seeTable1). andexpressions.(Afterexcludingmulti- wordexpressions,wemanuallyclassied203 pairsasCognatesand527pairsasUnre- lated.)

2.Amanuallyword-alignedbitext(Melamed

98).(Wemanuallyidentied258Cognate

pairsamongthealignedwordpairs.)

3.AsetofexercisesforAnglophonelearnersof

French(Treville90)(152Cognatepairs).

nates"(314False-Friends).

Aseparatetestsetiscomposedof1040pairs

(seeTable1),extractedfromthefollowing sources:

1.Arandomsampleof1000wordpairsfrom

anautomaticallygeneratedtranslationlex- icon.(Wemanuallyclassied603pairsas

Cognatesand343pairsasUnrelated.)

TrainingsetTestset

Cognates613(73)603(178)

False-Friends314(135)94(46)

Unrelated527(0)343(0)

Total14541040

Table1:Thecompositionofdatasets.Thenum-

identical(ignoringaccents).

EnglishFalseCognates"(94additionalFalse-

Friends).

Inordertoavoidanyoverlapbetweenthetwo

sets,weremovedfromthetestsetallpairsthat happenedtobealreadyincludedinthetraining set.Thedatasethasa2:1imbalanceinfavour intheexperimentspresentedinSection5).All tionpairs.Itwouldhavebeeneasytoaddmore sampletranslationlexicons.

5Evaluation

Wepresentevaluationexperimentsusingthe

twodatasetsdescribedinSection4:atrain- returnedbyallthe13measures.Then,inorder tocombinethemeasures,werunseveralmachine learningclassiersfromtheWekapackage.

5.1ResultsontheTrainingDataSet

measure,weneedtochooseaspecicsimilarity fromtheUnrelatedpairs.FortheIDENTmea-

OrthographicThresholdAccuracy

similaritymeasure

IDENT143.90%

PREFIX0.0384592.70%

DICE0.2966989.40%

LCSR0.4580092.91%

NED0.3484593.39%

SOUNDEX0.6250085.28%

TRI0.047688.30%

XDICE0.2182592.84%

XXDICE0.1291591.74%

BI-SIM0.3798094.84%

BI-DIST0.3416594.84%

TRI-SIM0.3484595.66%

TRI-DIST0.3484595.11%

Averagemeasure0.1477093.83%

Table2:Resultsofeachorthographicsimilar-

Thelastlinepresentsanewmeasurewhichisthe

averageofallmeasuresforeachpairofwords. thebestsplit.Thevaluesofthethresholdsob- tainedinthiswayarealsoincludedinTable2.

Thetrainingdatasetformachinelearningex-

ingclassiersfromtheWekapackage:OneRule(a sionofSupportVectorMachine.

TheDecisionTreeclassierhastheadvantage

aninstanceasUnrelatediftheBI-SIMvalueis greaterthan0.3.Sinceallmeasuresattemptto todissimilarpairs,thepresenceofsuchanode eringthecondencelevelthresholdfromthede- faultCF=0:25untilweobtainedatreewithout

ClassierAccuracyonAccuracy

trainingsetcross-val

Baseline63.75%63.75%

OneRule95.94%95.66%

NaiveBayes94.91%94.84%

DecisionTrees97.45%95.66%

DecTree(pruned)96.28%95.66%

IBK99.10%93.81%

AdaBoost95.66%95.66%

Perceptron95.73%95.11%

SVM(SMO)95.66%95.46%

TRI-SIM<=0.3333

|TRI-SIM<=0.2083:UNREL(447.0/17.0) |TRI-SIM>0.2083 ||XDICE<=0.2:UNREL(97.0/20.0) ||XDICE>0.2 |||BI-SIM<=0.3:UNREL(3.0) |||BI-SIM>0.3:CG_FF(9.0)

TRI-SIM>0.3333:CG_FF(898.0/17.0)

Figure1:ExampleofDecisionTreeclassier,

CF=16%).

ure1).Ourhypothesiswasthatthelattertree wouldperformbetteronatestset.

Theresultspresentedintherightmostcolumn

ontrainingdataset(thedataisrandomlysplit in10parts,aclassieristrainedon9partsand ingset:theyarearticiallyhigh,duetoover- training.ThebaselinealgorithmintheTable

3alwayschoosesthemostfrequentclassinthe

dataset,whichhappenedtobeCognates/False-

Friends.Thebestclassicationaccuracy(for

OneRule,andAdaBoost(95.66%).Theperfor-

manceequalstheoneachievedbytheTRI-SIM measurealoneinTable2.

Erroranalysis:Weexaminedthemisclassi-

edpairsfortheclassiersbuiltonthetraining data.Thereweremanysharedpairsamongthe

60{70pairsmisclassiedbyseveralofthebest

formduetochangesoflanguageovertime.False theword,whichisastrongclueofcognation.

Also,thepresenceofanidenticalprexmade

unlessthewordrootsarerelated.

5.2ResultsontheTestSet

obtainedonthetestsetdescribedinSection4. asfeatures.Theclassiersaretheonesbuilton thetrainingset.

Therankingofmeasuresonthetestsetdif-

fersfromtherankingobtainedonthetraining set,whichmaybecausedbytheabsenceofge- averageoforthographicmeasures.Thepruned

DecisionTreeshowninFigure1achieveshigher

stillbelowthesimpleaverage.Amongthein- dividualorthographicmeasures,XXDICEper-

Englishcognatesreportedin(Brew&McKelvie

set.Weconcludethatourclassiersaregeneric enough:theyperformverywellonthetestset.

5.3ResultsontheGeneticCognates

Dataset

Greenberg(Greenberg87)givesalistof\most

ofthecognatesfromFrenchandEnglish".The todemonstratethatFrenchandEnglisharege- cognatesbetweenthosetwolanguages.Wetran- scribedthelistof82cognatepairsfromIPAto standardorthography.Weaugmentedthelist

DataCorpus5and17pairsthatweidentiedour-

selves.Thenallistcontains113truegenetic cognatesthatgobacktoProto-Indoeuropean6.

ClassierAccuracyAccuracy

(measureorongeneticontest combination)cognatessetset

IDENT1.76%55.00%

PREFIX36.28%90.97%

DICE13.27%93.37%

LCSR24.77%94.24%

NED23.89%93.57%

SOUNDEX39.82%84.54%

TRI4.42%92.13%

XDICE15.92%94.52%

XXDICE13.27%95.39%

BI-SIM29.20%93.95%

BI-DIST29.20%94.04%

TRI-SIM35.39%93.28%

TRI-DIST34.51%93.85%

Averagemeasure36.28%94.14%

Baseline|66.98%

OneRule35.39%92.89%

NaiveBayes29.20%94.62%

DecisionTrees35.39%92.08%

DecTree(pruned)38.05%93.18%

IBK43.36%92.80%

AdaBoost35.39%93.47%

Perceptron42.47%91.55%

quotesdbs_dbs1.pdfusesText_1

[PDF] false friends english french list pdf

[PDF] false friends english french pdf

[PDF] false friends list

[PDF] famille d'hier et d'aujourd'hui

[PDF] famille de maupassant

[PDF] famille negriere nantes

[PDF] famille traditionnelle et famille moderne

[PDF] family visitor visa uk

[PDF] famous english short stories pdf

[PDF] famous english story books pdf free download

[PDF] famous novels in english pdf

[PDF] famous short stories in english literature pdf

[PDF] fantasia air economie antwoorden

[PDF] fao statistique algerie

[PDF] fardeau excessif immigration canada