Transactions of the Association for Computational Linguistics, 1 (2013) 353–366. Action Editor: Patrick Pantel.

In angehen An

Transactions of the Association for Computational Linguistics, 1 (2013) 353–366. Action Editor: Patrick Pantel.
Submitted 5/2013; Überarbeitet 7/2013; Published 10/2013. C(cid:13)2013 Verein für Computerlinguistik.

DistributionalSemanticsBeyondWords:SupervisedLearningofAnalogyandParaphrasePeterD.TurneyNationalResearchCouncilCanadaInformationandCommunicationsTechnologiesOttawa,Ontario,Kanada,K1A0R6peter.turney@nrc-cnrc.gc.caAbstractTherehavebeenseveraleffortstoextenddistributionalsemanticsbeyondindividualwords,tomeasurethesimilarityofwordpairs,phrases,andsentences(kurz,tuples;orderedsetsofwords,contiguousornoncontiguous).Onewaytoextendbeyondwordsistocom-paretwotuplesusingafunctionthatcom-binespairwisesimilaritiesbetweenthecom-ponentwordsinthetuples.Astrengthofthisapproachisthatitworkswithbothrela-tionalsimilarity(analogy)andcompositionalsimilarity(paraphrase).Jedoch,pastworkrequiredhand-codingthecombinationfunc-tionfordifferenttasks.Themaincontributionofthispaperisthatcombinationfunctionsaregeneratedbysupervisedlearning.Weachievestate-of-the-artresultsinmeasuringrelationalsimilaritybetweenwordpairs(SATanalo-giesandSemEval2012Task2)andmeasur-ingcompositionalsimilaritybetweennoun-modifierphrasesandunigrams(multiple-choiceparaphrasequestions).1IntroductionHarris(1954)andFirth(1957)hypothesizedthatwordsthatappearinsimilarcontextstendtohavesimilarmeanings.Thishypothesisisthefounda-tionfordistributionalsemantics,inwhichwordsarerepresentedbycontextvectors.Thesimilarityoftwowordsiscalculatedbycomparingthetwocor-respondingcontextvectors(Lundetal.,1995;Lan-dauerandDumais,1997;TurneyandPantel,2010).Distributionalsemanticsishighlyeffectiveformeasuringthesemanticsimilaritybetweenindivid-ualwords.Onasetofeightymultiple-choicesyn-onymquestionsfromthetestofEnglishasafor-eignlanguage(TOEFL),adistributionalapproachrecentlyachieved100%accuracy(BullinariaandLevy,2012).Jedoch,ithasbeendifficulttoextenddistributionalsemanticsbeyondindividualwords,towordpairs,phrases,andsentences.Movingbeyondindividualwords,therearevari-oustypesofsemanticsimilaritytoconsider.Herewefocusonparaphraseandanalogy.Paraphraseissimilarityinthemeaningoftwopiecesoftext(AndroutsopoulosandMalakasiotis,2010).Anal-ogyissimilarityinthesemanticrelationsoftwosetsofwords(Turney,2008A).Itiscommontostudyparaphraseatthesentencelevel(AndroutsopoulosandMalakasiotis,2010),butweprefertoconcentrateonthesimplesttypeofparaphrase,whereabigramparaphrasesaunigram.Forexample,doghouseisaparaphraseofkennel.Inourexperiments,weconcentrateonnoun-modifierbigramsandnoununigrams.Analogiesmaptermsinonedomaintotermsinanotherdomain(Gentner,1983).Thefamiliaranal-ogybetweenthesolarsystemandtheRutherford-Bohratomicmodelinvolvesseveraltermsfromthedomainofthesolarsystemandthedomainoftheatomicmodel(Turney,2008A).Thesimplesttypeofanalogyisproportionalanal-ogy,whichinvolvestwopairsofwords(Turney,2006B).Forexample,thepairhcook,rawiisanal-ogoustothepairhdecorate,plaini.Ifwecookathing,itisnolongerraw;ifwedecorateathing,Es

l

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

P

:
/
/

D
ich
R
e
C
T
.

M

ich
T
.

e
D
u

/
T

A
C
l
/

l

A
R
T
ich
C
e
–
P
D

F
/

D
Ö

ich
/

.

1
0
1
1
6
2

/
T

l

A
C
_
A
_
0
0
2
3
3
1
5
6
6
6
7
5

/

/
T

l

A
C
_
A
_
0
0
2
3
3
P
D

.

F

B
j
G
u
e
S
T

T

Ö
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

354

isnolongerplain.Thesemanticrelationsbetweencookandrawaresimilartothesemanticrelationsbetweendecorateandplain.Inthefollowingexper-iments,wefocusonproportionalanalogies.Erk(2013)distinguishedfourapproachestoextenddistributionalsemanticsbeyondwords:Inthefirst,asinglevectorspacerepresentationforaphraseorsentenceiscomputedfromtherepresen-tationsoftheindividualwords(MitchellandLap-ata,2010;BaroniandZamparelli,2010).Inthesec-ond,twophrasesorsentencesarecomparedbycom-biningmultiplepairwisesimilarityvalues(Socheretal.,2011;Turney,2012).Dritte,weightedinferencerulesintegratedistributionalsimilarityandformallogic(Garretteetal.,2011).Vierte,asinglespaceintegratesformallogicandvectors(Clarke,2012).Takingthesecondapproach,Turney(2012)intro-ducedadual-spacemodel,withonespaceformea-suringdomainsimilarity(similarityoftopicorfield)andanotherforfunctionsimilarity(similarityofroleorusage).Similaritiesbeyondindividualwordsarecalculatedbyfunctionsthatcombinedomainandfunctionsimilaritiesofcomponentwords.Thedual-spacemodelhasbeenappliedtomea-suringcompositionalsimilarity(paraphraserecogni-tion)andrelationalsimilarity(analogyrecognition).Inexperimentsthattestedforsensitivitytowordorder,thedual-spacemodelperformedsignificantlybetterthancompetingapproaches(Turney,2012).Alimitationofpastworkwiththedual-spacemodelisthatthecombinationfunctionswerehand-coded.Ourmaincontributionistoshowhowhand-codingcanbeeliminatedwithsupervisedlearning.Foreaseofreference,wewillcallourapproachSuperSim(supervisedsimilarity).Withnomodifi-cationofSuperSimforthespecifictask(relationalsimilarityorcompositionalsimilarity),weachievebetterresultsthanprevioushand-codedmodels.Compositionalsimilarity(paraphrase)comparestwocontiguousphrasesorsentences(n-grams),whereasrelationalsimilarity(analogy)doesnotrequirecontiguity.Weusetupletorefertobothcon-tiguousandnoncontiguouswordsequences.Weapproachanalogyasaproblemofsupervisedtupleclassification.Tomeasuretherelationalsim-ilaritybetweentwowordpairs,wetrainSuperSimwithquadruplesthatarelabeledaspositiveandneg-ativeexamplesofanalogies.Forexample,thepro-portionalanalogyhcook,raw,decorate,plainiislabeledasapositiveexample.Aquadrupleisrepresentedbyafeaturevector,composedofdomainandfunctionsimilaritiesfromthedual-spacemodelandotherfeaturesbasedoncorpusfrequencies.SuperSimusesasupportvectormachine(Platt,1998)tolearntheprobabilitythataquadrupleha,B,C,diconsistsofawordpairha,biandananalogouswordpairhc,di.Theprobabilitycanbeinterpretedasthedegreeofrelationalsimilar-itybetweenthetwogivenwordpairs.Wealsoapproachparaphraseassupervisedtupleclassification.Tomeasurethecompositionalsimi-laritybeweenanm-gramandann-gram,wetrainthelearningalgorithmwith(m+n)-tuplesthatarepositiveandnegativeexamplesofparaphrases.SuperSimlearnstoestimatetheprobabilitythatatripleha,B,ciconsistsofacompositionalbigramabandasynonymousunigramc.Forinstance,thephrasefishtankissynonymouswithaquarium;thatis,fishtankandaquariumhavehighcompositionalsimilarity.Thetriplehfish,tank,aquariumiisrepre-sentedusingthesamefeaturesthatweusedforanal-ogy.Theprobabilityofthetriplecanbeinterpretedasthedegreeofcompositionalsimilaritybetweenthegivenbigramandunigram.WereviewrelatedworkinSection2.Thegen-eralfeaturespaceforlearningrelationsandcompo-sitionsispresentedinSection3.TheexperimentswithrelationalsimilarityaredescribedinSection4,andSection5reportstheresultswithcompositionalsimilarity.Section6discussestheimplicationsoftheresults.WeconsiderfutureworkinSection7andconcludeinSection8.2RelatedWorkInSemEval2012,Task2wasconcernedwithmea-suringthedegreeofrelationalsimilaritybetweentwowordpairs(Jurgensetal.,2012)andTask6(Agirreetal.,2012)examinedthedegreeofseman-ticequivalencebetweentwosentences.Thesetwoareasofresearchhavebeenmostlyindependent,althoughSocheretal.(2012)andTurney(2012)presentunifiedperspectivesonthetwotasks.Wefirstdiscusssomeworkonrelationalsimilarity,thensomeworkoncompositionalsimilarity,andlastlyworkthatunifiesthetwotypesofsimilarity.

l

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

P

:
/
/

D
ich
R
e
C
T
.

M

ich
T
.

e
D
u

/
T

A
C
l
/

l

A
R
T
ich
C
e
–
P
D

F
/

D
Ö

ich
/

.

1
0
1
1
6
2

/
T

l

A
C
_
A
_
0
0
2
3
3
1
5
6
6
6
7
5

/

/
T

l

A
C
_
A
_
0
0
2
3
3
P
D

.

F

B
j
G
u
e
S
T

T

Ö
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

355

2.1RelationalSimilarityLRA(latentrelationalanalysis)measuresrela-tionalsimilaritywithapair–patternmatrix(Turney,2006B).Rowsinthematrixcorrespondtowordpairs(A,B)andcolumnscorrespondtopatternsthatconnectthepairs(“afortheb”)inalargecor-pus.Thisisaholistic(noncompositional)approachtodistributionalsimilarity,sincethewordpairsareopaquewholes;thecomponentwordshavenosep-araterepresentations.Acompositionalapproachtoanalogyhasarepresentationforeachword,andawordpairisrepresentedbycomposingtherepresen-tationsforeachmemberofthepair.Givenavocabu-laryofNwords,acompositionalapproachrequiresNrepresentationstohandleallpossiblewordpairs,butaholisticapproachrequiresN2representations.Holisticapproachesdonotscaleup(Turney,2012).LRArequiredninedaystorun.Bollegalaetal.(2008)answeredtheSATanal-ogyquestionswithasupportvectormachinetrainedonquadruples(proportionalanalogies),aswedohere.However,theirfeaturevectorsareholistic,andhencetherearescalingproblems.Herda˘gdelenandBaroni(2009)usedasupportvectormachinetolearnrelationalsimilarity.Theirfeaturevectorscontainedacombinationofholisticandcompositionalfeatures.Measuringrelationalsimilarityiscloselycon-nectedtoclassifyingwordpairsaccordingtotheirsemanticrelations(TurneyandLittman,2005).SemanticrelationclassificationwasthefocusofSemEval2007Task4(Girjuetal.,2007)andSemEval2010Task8(Hendrickxetal.,2010).2.2CompositionalSimilarityToextenddistributionalsemanticsbeyondwords,manyresearcherstakethefirstapproachdescribedbyErk(2013),inwhichasinglevectorspaceisusedforindividualwords,phrases,andsentences(Lan-dauerandDumais,1997;MitchellandLapata,2008;MitchellandLapata,2010).Inthisapproach,giventhewordsaandbwithcontextvectorsaandb,weconstructavectorforthebigramabbyapplyingvec-toroperationstoaandb.MitchellandLapata(2010)experimentwithmanydifferentvectoroperationsandfindthatelement-wisemultiplicationperformswell.Thebigramabisrepresentedbyc=a(cid:12)B,whereci=ai·bi.However,element-wisemultiplica-tioniscommutative,sothebigramsabandbamaptothesamevectorc.Inexperimentsthattestforordersensitivity,element-wisemultiplicationper-formspoorly(Turney,2012).Wecantreatthebigramabasaunit,asifitwereasingleword,andconstructacontextvectorforabfromoccurrencesofabinalargecorpus.Thisholisticapproachtorepresentingbigramsperformswellwhenalimitedsetofbigramsisspecifiedinadvance(beforebuildingtheword–contextmatrix),butitdoesnotscaleup,becausetherearetoomanypossiblebigrams(Turney,2012).Althoughtheholisticapproachdoesnotscaleup,wecangenerateafewholisticbigramvectorsandusethemtotrainasupervisedregressionmodel(Guevara,2010;BaroniandZamparelli,2010).Givenanewbigramcd,notobservedinthecorpus,theregressionmodelcanpredictaholisticvectorforcd,ifcanddhavebeenobservedseparately.WeshowinSection5thatthisideacanbeadaptedtotrainSuperSimwithoutmanuallylabeleddata.Socheretal.(2011)takethesecondapproachdescribedbyErk(2013),inwhichtwosentencesarecomparedbycombiningmultiplepairwisesimilar-ityvalues.Theyconstructavariable-sizedsimilar-itymatrixX,inwhichtheelementxijisthesim-ilaritybetweenthei-thphraseofonesentenceandthej-thphraseoftheother.Sincesupervisedlearn-ingissimplerwithfixed-sizedfeaturevectors,thevariable-sizedsimilaritymatrixisthenreducedtoasmallerfixed-sizedmatrix,toallowcomparisonofpairsofsentencesofvaryinglengths.2.3UnifiedPerspectivesonSimilaritySocheretal.(2012)representwordsandphraseswithapair,consistingofavectorandamatrix.Thevectorcapturesthemeaningofthewordorphraseandthematrixcaptureshowawordorphrasemod-ifiesthemeaningofanotherwordorphrasewhentheyarecombined.Theyapplythismatrix–vectorrepresentationtobothcompositionsandrelations.Turney(2012)representswordswithtwovectors,avectorfromdomainspaceandavectorfromfunc-tionspace.Thedomainvectorcapturesthetopicorfieldofthewordandthefunctionvectorcapturesthe

l

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

P

:
/
/

D
ich
R
e
C
T
.

M

ich
T
.

e
D
u

/
T

A
C
l
/

l

A
R
T
ich
C
e
–
P
D

F
/

D
Ö

ich
/

.

1
0
1
1
6
2

/
T

l

A
C
_
A
_
0
0
2
3
3
1
5
6
6
6
7
5

/

/
T

l

A
C
_
A
_
0
0
2
3
3
P
D

.

F

B
j
G
u
e
S
T

T

Ö
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

356

functionalroleoftheword.Thisdual-spacemodelisappliedtobothcompositionsandrelations.Hereweextendthedual-spacemodelofTur-ney(2012)intwoways:Hand-codingisreplacedwithsupervisedlearningandtwonewsetsoffea-turesaugmentdomainandfunctionspace.Movingtosupervisedlearninginsteadofhand-codingmakesiteasiertointroducenewfeatures.Inthedual-spacemodel,parameterizedsimilar-itymeasuresprovidedtheinputvaluesforhand-craftedfunctions.Eachtaskrequiredadifferentsetofhand-craftedfunctions.Theparametersofthesimilaritymeasuresweretunedusingacustomizedgridsearchalgorithm.Thegridsearchalgorithmwasnotsuitableforintegrationwithasupervisedlearningalgorithm.TheinsightbehindSuperSimisthat,givenappropriatefeatures,asupervisedlearn-ingalgorithmcanreplacethegridsearchalgorithmandthehand-craftedfunctions.3FeaturesforTupleClassificationWerepresentatuplewithfourtypesoffeatures,allbasedonfrequenciesinalargecorpus.Thefirsttypeoffeatureisthelogarithmofthefrequencyofaword.Thesecondtypeisthepositivepoint-wisemutualinformation(PPMI)betweentwowords(ChurchandHanks,1989;BullinariaandLevy,2007).Thirdandfourtharethesimilaritiesoftwowordsindomainandfunctionspace(Turney,2012).Inthefollowingexperiments,weusethePPMImatrixfromTurneyetal.(2011)andthedomainandfunctionmatricesfromTurney(2012).1Thethreematricesandthewordfrequencydataarebasedonthesamecorpus,acollectionofwebpagesgath-eredfromuniversitywebsites,containing5×1010words.2Allthreematricesareword–contextmatri-ces,inwhichtherowscorrespondtoterms(wordsandphrases)inWordNet.3Thecolumnscorrespondtothecontextsinwhichthetermsappear;eachmatrixinvolvesadifferentkindofcontext.1Thethreematricesandthewordfrequencydataareavail-ableonrequestfromtheauthor.Thematrixfilesrangefromtwotofivegigabyteswhenpackagedandcompressedfordistri-bution.2ThecorpuswascollectedbyCharlesClarkeattheUniver-sityofWaterloo.Itisabout280gigabytesofplaintext.3Seehttp://wordnet.princeton.edu/forinfor-mationaboutWordNet.Lethx1,x2,…,xnibeann-tupleofwords.Thenumberoffeaturesweusetorepresentthistupleincreasesasafunctionofn.Thefirstsetoffeaturesconsistsoflogfrequencyvaluesforeachwordxiinthen-tuple.Letfreq(xi)bethefrequencyofxiinthecorpus.WedefineLF(xi)aslog(freq(xi)+1).Ifxiisnotinthecorpus,freq(xi)iszero,andthusLF(xi)isalsozero.Therearenlogfrequencyfeatures,oneLF(xi)featureforeachwordinthen-tuple.Thesecondsetoffeaturesconsistsofpositivepointwisemutualinformationvaluesforeachpairofwordsinthen-tuple.WeusetherawPPMImatrixfromTurneyetal.(2011).Althoughtheycomputedthesingularvaluedecomposition(SVD)toprojecttherowvectorsintoalower-dimensionalspace,weneedtheoriginalhigh-dimensionalcolumnsforourfeatures.TherawPPMImatrixhas114,501rowsand139,246columnswithadensityof1.2%.ForeachterminWordNet,thereisacorrespondingrowintherawPPMImatrix.ForeachunigraminWord-Net,therearetwocorrespondingcolumnsintherawPPMImatrix,onemarkedleftandtheotherright.Supposexicorrespondstothei-throwofthePPMImatrixandxjcorrespondsthej-thcolumn,markedleft.Thevalueinthei-throwandj-thcol-umnofthePPMImatrix,PPMI(xi,xj,links),isthepositivepointwisemutualinformationofxiandxjco-occurringinthecorpus,wherexjisthefirstwordtotheleftofxi,ignoringanyinterveningstopwords(thatis,ignoringanywordsthatarenotinWordNet).Ifxi(orxj)hasnocorrespondingrow(orcolumn)inthematrix,thenthePPMIvalueissettozero.Turneyetal.(2011)estimatedPPMI(xi,xj,links)bysamplingthecorpusforphrasescontainingxiandthenlookingforxjtotheleftofxiinthesampledphrases(andlikewiseforright).Duetothissam-plingprocess,PPMI(xi,xj,links)doesnotnecessar-ilyequalPPMI(xj,xi,Rechts).Forexample,supposexiisararewordandxjisacommonword.WithPPMI(xi,xj,links),whenwesamplephrasescontain-ingxi,wearerelativelylikelytofindxjinsomeofthesephrases.WithPPMI(xj,xi,Rechts),whenwesamplephrasescontainingxj,wearelesslikelytofindanyphrasescontainingxi.Although,intheory,PPMI(xi,xj,links)shouldequalPPMI(xj,xi,Rechts),theyarelikelytobeunequalgivenalimitedsample.

l

D
Ö
w
N
Ö
A
D
e
D

F
R
Ö
M
H

T
T

P

:
/
/

D
ich
R
e
C
T
.

M

ich
T
.

e
D
u

/
T

A
C
l
/

l

A
R
T
ich
C
e
–
P
D

F
/

D
Ö

ich
/

.

1
0
1
1
6
2

/
T

l

A
C
_
A
_
0
0
2
3
3
1
5
6
6
6
7
5

/

/
T

l

A
C
_
A
_
0
0
2
3
3
P
D

.

F

B
j
G
u
e
S
T

T

Ö
N
0
8
S
e
P
e
M
B
e
R
2
0
2
3

357

Fromthen-tuple,weselectallofthen(n−1)pairs,hxi,xji,suchthati6=j.Wethengener-atetwofeaturesforeachpair,PPMI(xi,xj,links)andPPMI(xi,xj,Rechts).Thusthereare2n(n−1)PPMIvaluesinthesecondsetoffeatures.Thethirdsetoffeaturesconsistsofdomainspacesimilarityvaluesforeachpairofwordsinthen-tuple.Domainspacewasdesignedtocapturethetopicofaword.Turney(2012)firstconstructedafrequencymatrix,inwhichtherowscorrespondtotermsinWordNetandthecolumnscorrespondtonearbynouns.Givenatermxi,thecorpuswassam-pledforphrasescontainingxiandthephraseswereprocessedwithapart-of-speechtagger,toidentifynouns.Ifthenounxjwastheclosestnountotheleftorrightofxi,thenthefrequencycountforthei-throwandj-thcolumnwasincremented.Thehypoth-esiswasthatthenounsnearatermcharacterizethetopicsassociatedwiththeterm.Theword–contextfrequencymatrixfordomainspacehas114,297rows(Bedingungen)and50,000columns(nouncontexts,topics),withadensityof2.6%.ThefrequencymatrixwasconvertedtoaPPMImatrixandthensmoothedwithSVD.TheSVDyieldsthreematrices,U,Σ,andV.AtermindomainspaceisrepresentedbyarowvectorinUkΣpk.Theparameterkspecifiesthenum-berofsingularvaluesinthetruncatedsingularvaluedecomposition;thatis,kisthenumberoflatentfactorsinthelow-dimensionalrepresentationoftheterm(LandauerandDumais,1997).WegenerateUkandΣkbydeletingthecolumnsinUandΣcorrespondingtothesmallestsingularvalues.TheparameterpraisesthesingularvaluesinΣktothepowerp(Caron,2001).Aspgoesfromonetozero,factorswithsmallersingularvaluesaregivenmoreweight.Thishastheeffectofmakingthesimilaritymeasuremorediscriminating(Turney,2012).Thesimilarityoftwowordsindomainspace,Dom(xi,xj,k,P),iscomputedbyextractingtherowvectorsinUkΣpkthatcorrespondtothewordsxiandxj,andthencalculatingtheircosine.Optimalper-formancerequirestuningtheparameterskandpforthetask(BullinariaandLevy,2012;Turney,2012).Inthefollowingexperiments,weavoiddirectlytun-ingkandpbygeneratingfeatureswithavarietyofvaluesforkandp,allowingthesupervisedlearningalgorithmtodecidewhichfeaturestouse.FeaturesetSizeofsetLF(xi)nPPMI(xi,xj,handedness)2N(n−1)Dom(xi,xj,k,P)12N(n−1)nknpFun(xi,xj,k,P)12N(n−1)nknpTable1:Thefoursetsoffeaturesandtheirsizes.Fromthen-tuple,weselectall12n(n−1)pairs,hxi,xji,suchthatie l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 358 n-tupleLFPPMIDomFunTotal11000122411011022633123303306754424660660134855401100110022456660165016503366Table2:Numberoffeaturesforvarioustuplesizes.corpus.ThePPMIfeaturesarebasedondirectco-occurrencesoftwowords;thatis,PPMIisonlygreaterthanzeroifthetwowordsactuallyoccurtogetherinthecorpus.Domainandfunctionspacecaptureindirectorhigher-orderco-occurrence,duetothetruncatedSVD(LemaireandDenhi`ere,2006);thatis,thevaluesofDom(xi,xj,k,P)andFun(xi,xj,k,P)canbehighevenwhenxiandxjdonotactuallyco-occurinthecorpus.Weconjec-turethatthereareyethigherordersinthishierarchythatwouldprovideimprovedsimilaritymeasures.SuperSimlearnstoclassifytuplesbyrepresentingthemwiththesefeatures.SuperSimusesthesequen-tialminimaloptimization(SMO)supportvectormachine(SVM)asimplementedinWeka(Platt,1998;Wittenetal.,2011).4Thekernelisanormal-izedthird-orderpolynomial.Wekaprovidesproba-bilityestimatesfortheclassesbyfittingtheoutputsoftheSVMwithlogisticregressionmodels.4RelationalSimilarityThissectionpresentsexperimentswithlearningrela-tionalsimilarityusingSuperSim.Thetrainingdatasetsconsistofquadruplesthatarelabeledaspositiveandnegativeexamplesofanalogies.Table2showsthatthefeaturevectorshave1,348elements.Weexperimentwiththreedatasets,acollectionof374five-choicequestionsfromtheSATcol-legeentranceexam(Turneyetal.,2003),amodi-fiedten-choicevariationoftheSATquestions(Tur-ney,2012),andtherelationalsimilaritydatasetfromSemEval2012Task2(Jurgensetal.,2012).54Wekaisavailableathttp://www.cs.waikato.ac.nz/ml/weka/.5TheSATquestionsareavailableonrequestfromtheauthor.TheSemEval2012Task2datasetisavailableathttps://sites.google.com/site/semeval2012task2/.Stem:word:languageChoices:(1)paint:portrait(2)poetry:rhythm(3)Notiz:Musik(4)tale:Geschichte(5)week:yearSolution:(3)Notiz:musicTable3:Afive-choiceSATanalogyquestion.4.1Five-choiceSATQuestionsTable3isanexampleofaquestionfromthe374five-choiceSATquestions.Eachfive-choiceques-tionyieldsfivelabeledquadruples,bycombiningthestemwitheachchoice.Thequadruplehword,lan-guage,Notiz,musiciislabeledpositiveandtheotherfourquadruplesarelabelednegative.Sincelearningworksbetterwithbalancedtrain-ingdata(JapkowiczandStephen,2002),weusethesymmetriesofproportionalanalogiestoaddmorepositiveexamples(LepageandShin-ichi,1996).Foreachpositivequadruple,ha,B,C,di,weaddthreemorepositivequadruples,hb,A,D,ci,hc,D,A,bi,andhd,C,B,ai.Thuseachfive-choicequestionpro-videsfourpositiveandfournegativequadruples.Weuseten-foldcross-validationtoapplySuper-SimtotheSATquestions.ThefoldsareconstructedsothattheeightquadruplesfromeachSATquestionarekepttogetherinthesamefold.Toansweraques-tioninthetestingfold,thelearnedmodelassignsaprobabilitytoeachofthefivechoicesandguessesthechoicewiththehighestprobability.SuperSimachievesascoreof54.8%correct(205outof374).Table4givestherankofSuperSiminthelistofthetoptenresultswiththeSATanalogyquestions.6Thescoresrangingfrom51.1%to57.0%arenotsig-nificantlydifferentfromSuperSim’sscoreof54.8%,accordingtoFisher’sexacttestatthe95%confi-dencelevel.However,SuperSimanswerstheSATquestionsinafewminutes,whereasLRArequiresninedays,andSuperSimlearnsitsmodelsautomat-ically,unlikethehand-codingofTurney(2012).6SeetheStateoftheArtpageontheACLWikiathttp://aclweb.org/aclwiki. l D O w N O A D e D F R O M H T T P : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 359 AlgorithmReferenceCorrectKnow-BestVeale(2004)43.0k-meansBic¸ici&Yuret(2006)44.0BagPackHerda˘gdelen&Baroni(2009)44.1VSMTurney&Littman(2005)47.1Dual-SpaceTurney(2012)51.1BMIBollegalaetal.(2009)51.1PairClassTurney(2008B)52.1PERTTurney(2006A)53.5SuperSim—54.8LRATurney(2006B)56.1HumanAveragecollegeapplicant57.0Table4:Thetoptenresultsonfive-choiceSATquestions.4.2Ten-choiceSATQuestionsInadditiontosymmetries,proportionalanalogieshaveasymmetries.Ingeneral,ifthequadrupleha,B,C,diispositive,ha,D,C,biisnegative.Forexample,hword,Sprache,Notiz,musiciisagoodanalogy,buthword,Musik,Notiz,languageiisnot.Wordsarethebasicunitsoflanguageandnotesarethebasicunitsofmusic,butwordsarenotnecessaryformusicandnotesarenotnecessaryforlanguage.Turney(2012)usedthisasymmetrytoconvertthe374five-choiceSATquestionsinto374ten-choiceSATquestions.Eachchoicehc,diwasexpandedwiththestemha,bi,resultinginthequadrupleha,B,C,di,andthentheorderwasshuf-fledtoha,D,C,bi,sothateachchoicepairinafive-choicequestiongeneratedtwochoicequadruplesinaten-choicequestion.Nineofthequadruplesarenegativeexamplesandthequadrupleconsistingofthestempairfollowedbythesolutionpairistheonlypositiveexample.Thepurposeoftheten-choicequestionsistotesttheabilityofmeasuresofrela-tionalsimilaritytoavoidtheasymmetricdistractors.Weusetheten-choicequestionstocomparethehand-codeddual-spaceapproach(Turney,2012)withSuperSim.Wealsousethesequestionstoper-formanablationstudyofthefoursetsoffeaturesinSuperSim.Aswiththefive-choicequestions,weusethesymmetriesofproportionalanalogiestoaddthreemorepositiveexamples,sothetrainingdatasethasninenegativeexamplesandfourposi-tiveexamplesperquestion.Weapplyten-foldcross-validationtothe374ten-choicequestions.Ontheten-choicequestions,SuperSim’sscoreFeaturesAlgorithmLFPPMIDomFunCorrectDual-Space001147.9SuperSim111152.7SuperSim011152.7SuperSim101152.7SuperSim110145.7SuperSim111041.7SuperSim10005.6SuperSim010032.4SuperSim001039.6SuperSim000139.3Table5:Featureablationwithten-choiceSATquestions.is52.7%(Table5),comparedto54.8%onthefive-choicequestions(Table4),adropof2.1%.Thehand-codeddual-spacemodelscores47.9%(Table5),comparedto51.1%onthefive-choicequestions(Table4),adropof3.2%.Thedif-ferencebetweenSuperSim(52.7%)andthehand-codeddual-spacemodel(47.9%)isnotsignificantaccordingtoFisher’sexacttestatthe95%confi-dencelevel.TheadvantageofSuperSimisthatitdoesnotneedhand-coding.TheresultsshowthatSuperSimcanavoidtheasymmetricdistractors.Table5showstheimpactofdifferentsubsetsoffeaturesonthepercentageofcorrectanswerstotheten-choiceSATquestions.Includedfeaturesaremarked1andablatedfeaturesaremarked0.Theresultsshowthatthelogfrequency(LF)andPPMIfeaturesarenothelpful(butalsonotharmful)forrelationalsimilarity.Wealsoseethatdomainspaceandfunctionspacearebothneededforgoodresults.4.3SemEval2012Task2TheSemEval2012Task2datasetisbasedonthesemanticrelationclassificationschemeofBejaretal.(1991),consistingoftenhigh-levelcategoriesofrelationsandseventy-ninesubcategories,withparadigmaticexamplesofeachsubcategory.Forinstance,thesubcategorytaxonomicinthecate-goryclassinclusionhasthreeparadigmaticexam-ples,flower:tulip,emotion:rage,andpoem:sonnet.Jurgensetal.(2012)usedAmazon’sMechanicalTurktocreatetheSemEval2012Task2datasetintwophases.Inthefirstphase,Turkersexpandedtheparadigmaticexamplesforeachsubcategorytoan l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 360 AlgorithmReferenceSpearmanBUAPTovaretal.(2012)0.014Duluth-V2Pedersen(2012)0.038Duluth-V1Pedersen(2012)0.039Duluth-V0Pedersen(2012)0.050UTD-SVMRink&Harabagiu(2012)0.116UTD-NBRink&Harabagiu(2012)0.229RNN-1600Mikolovetal.(2013)0.275UTD-LDARink&Harabagiu(2013)0.334ComZhilaetal.(2013)0.353SuperSim—0.408Table6:SpearmancorrelationsforSemEval2012Task2.averageofforty-onewordpairspersubcategory,atotalof3,218pairs.Inthesecondphase,eachwordpairfromthefirstphasewasassignedaprototypical-ityscore,indicatingitssimilaritytotheparadigmaticexamples.ThechallengeofSemEval2012Task2wastoguesstheprototypicalityscores.SuperSimwastrainedonthefive-choiceSATquestionsandevaluatedontheSemEval2012Task2testdataset.Foragivenawordpair,wecreatedquadruples,combiningthewordpairwitheachoftheparadigmaticexamplesforitssubcategory.WethenusedSuperSimtocomputetheprobabilitiesforeachquadruple.Ourguessfortheprototypicalityscoreofthegivenwordpairwastheaverageoftheprobabilities.Spearman’srankcorrelationcoef-ficientbetweentheTurkers’prototypicalityscoresandSuperSim’sscoreswas0.408,averagedoverthesixty-ninesubcategoriesinthetestingset.Super-SimhasthehighestSpearmancorrelationachievedtodateonSemEval2012Task2(seeTable6).5CompositionalSimilarityThissectionpresentsexperimentsusingSuperSimtolearncompositionalsimilarity.Thedatasetscon-sistoftriples,ha,B,ci,suchthatabisanoun-modifierbigramandcisanoununigram.Thetriplesarelabeledaspositiveandnegativeexam-plesofparaphrases.Table2showsthatthefea-turevectorshave675elements.Weexperimentwithtwodatasets,seven-choiceandfourteen-choicenoun-modifierquestions(Turney,2012).77Theseven-choicedatasetisavailableathttp://jair.org/papers/paper3640.html.Thefourteen-choicedatasetcanbegeneratedfromtheseven-choicedataset.Stem:fantasyworldChoices:(1)fairyland(2)fantasy(3)Welt(4)phantasy(5)universe(6)ranter(7)souringSolution:(1)fairylandTable7:Anoun-modifierquestionbasedonWordNet.5.1Noun-ModifierQuestionsThefirstdatasetisaseven-choicenoun-modifierquestiondataset,constructedfromWordNet(Tur-ney,2012).Thedatasetcontains680questionsfortrainingand1,500fortesting,atotalof2,180ques-tions.Table7showsoneofthequestions.Thestemisabigramandthechoicesareuni-grams.Thebigramiscomposedofaheadnoun(Welt),modifiedbyanadjectiveornoun(fantasy).Thesolutionistheunigram(fairyland)thatbelongstothesameWordNetsynsetasthestem.Thedistractorsaredesignedtobedifficultforcur-rentapproachestocomposition.Forexample,iffan-tasyworldisrepresentedbyelement-wisemultipli-cationofthecontextvectorsforfantasyandworld(MitchellandLapata,2010),themostlikelyguessisfantasyorworld,notfairyland(Turney,2012).Eachseven-choicequestionyieldssevenlabeledtriples,bycombiningthestemwitheachchoice.Thetriplehfantasy,Welt,fairylandiislabeledpos-itiveandtheothersixtriplesarelabelednegative.Ingeneral,ifha,B,ciisapositiveexample,thenhb,A,ciisnegative.Forexample,worldfantasyisnotaparaphraseoffairyland.Theseconddatasetisconstructedbyapplyingthisshufflingtransfor-mationtoconvertthe2,180seven-choicequestionsinto2,180fourteen-choicequestions(Turney,2012).Theseconddatasetisdesignedtobedifficultforapproachesthatarenotsensitivetowordorder.Table8showsthepercentageofthetestingquestionsthatareansweredcorrectlyforthetwodatasets.Becausevectoradditionandelement-wisemultiplicationarenotsensitivetowordorder,theyperformpoorlyonthefourteen-choicequestions.Forbothdatasets,SuperSimperformssignificantly l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 361 CorrectAlgorithm7-choices14-choicesVectoraddition50.122.5Element-wisemultiplication57.527.4Dual-Spacemodel58.341.5SuperSim75.968.0Holisticmodel81.6—Table8:Resultsforthetwonoun-modifierdatasets.betterthanallotherapproaches,exceptfortheholis-ticapproach,accordingtoFisher’sexacttestatthe95%confidencelevel.8Theholisticapproachisnoncompositional.Thestembigramisrepresentedbyasinglecontextvec-tor,generatedbytreatingthebigramasifitwereaunigram.Anoncompositionalapproachcannotscaleuptorealisticapplications(Turney,2012).Theholisticapproachcannotbeappliedtothefourteen-choicequestions,becausethebigramsintheseques-tionsdonotcorrespondtotermsinWordNet,andhencetheydonotcorrespondtorowvectorsinthematricesweuse(seeSection3).Turney(2012)founditnecessarytohand-codeasoundnesscheckintoallofthealgorithms(vectoraddition,element-wisemultiplication,dual-space,andholistic).Givenastemabandachoicec,thehand-codedcheckassignsaminimalscoretothechoiceifc=aorc=b.Wedonotneedtohand-codeanycheckingintoSuperSim.Itlearnsautomat-icallyfromthetrainingdatatoavoidsuchchoices.5.2AblationExperimentsTable9showstheeffectsofablatingsetsoffea-turesontheperformanceofSuperSimwiththefourteen-choicequestions.PPMIfeaturesarethemostimportant;bythemselves,theyachieve59.7%correct,althoughtheotherfeaturesareneededtoreach68.0%.Domainspacefeaturesreachthesec-ondhighestperformancewhenusedalone(34.6%),buttheyreduceperformance(from69.3%to68.0%)whencombinedwithotherfeatures;Jedoch,thedropisnotsignificantaccordingtoFisher’sexacttestatthe95%significancelevel.SincethePPMIfeaturesplayanimportantroleinansweringthenoun-modifierquestions,letustake8TheresultsforSuperSimarenewbuttheotherresultsinTable8arefromTurney(2012).FeaturesAlgorithmLFPPMIDomFunCorrectDual-Space001141.5SuperSim111168.0SuperSim011166.6SuperSim101152.3SuperSim110169.3SuperSim111065.9SuperSim100014.1SuperSim010059.7SuperSim001034.6SuperSim000132.9Table9:Ablationwithfourteen-choicequestions.PPMIfeaturesubsetsha,biha,cihb,ciCorrect11168.001159.910165.411067.510062.601058.100155.600052.3Table10:PPMIsubsetablationwithfourteen-choices.acloserlookatthem.FromTable2,weseethattherearetwelvePPMIfeaturesforthetripleha,B,ci,whereabisanoun-modifierbigramandcisanoununigram.Wecansplitthetwelvefeaturesintothreesubsets,onesubsetforeachpairofwords,ha,bi,ha,ci,andhb,ci.Forexample,thesubsetforha,biisthefourfeaturesPPMI(A,B,links),PPMI(B,A,links),PPMI(A,B,Rechts),andPPMI(B,A,Rechts).Table10showstheeffectsofablatingthesesubsets.TheresultsinTable10indicatethatallthreePPMIsubsetscontributetotheperformanceofSuperSim,buttheha,bisubsetcontributesmorethantheothertwosubsets.Theha,bifeatureshelptoincreasethesensitivityofSuperSimtotheorderofthewordsinthenoun-modifierbigram;forexam-ple,theymakeiteasiertodistinguishfantasyworldfromworldfantasy.5.3HolisticTrainingSuperSimuses680trainingquestionstolearntorec-ognizewhenabigramisaparaphraseofaunigram;itlearnsfromexpertknowledgeimplicitinWordNet l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 362 Stem:searchengineChoices:(1)searchengine(2)suchen(3)engine(4)searchlanguage(5)searchwarrant(6)dieselengine(7)steamengineSolution:(1)searchengineTable11:Aquestionbasedonholisticvectors.synsets.ItwouldbeadvantageoustobeabletotrainSuperSimwithlessrelianceonexpertknowledge.Pastworkwithadjective-nounbigramshasshownthatwecanuseholisticbigramvectorstotrainasupervisedregressionmodel(Guevara,2010;BaroniandZamparelli,2010).Theoutputoftheregressionmodelisavectorrepresentationforabigramthatapproximatestheholisticvectorforthebigram;thatis,itapproximatesthevectorwewouldgetbytreat-ingthebigramasifitwereaunigram.SuperSimdoesnotgeneratevectorsasoutput,butwecanstilluseholisticbigramvectorsfortraining.Table11showsaseven-choicetrainingquestionthatwasgeneratedwithoutusingWordNetsynsets.Thechoicesoftheformabarebigrams,butwerepre-sentthemwithholisticbigramvectors;wepretendtheyareunigrams.Wecallabbigramspseudo-unigrams.AsfarasSuperSimisconcerned,thereisnodifferencebetweenthesepseudo-unigramsandtrueunigrams.ThequestioninTable11istreatedthesameasthequestioninTable7.Wegenerate680holistictrainingquestionsbyrandomlyselecting680noun-modifierbigramsfromWordNetasstemsforthequestions(searchengine),avoidinganybigramsthatappearasstemsinthetestingquestions.Thesolution(searchengine)isthepseudo-unigramthatcorrespondstothestembigram.InthematricesinSection3,eachterminWordNetcorrespondstoarowvector.ThesecorrespondingrowvectorsenableustotreatbigramsfromWordNetasiftheywereunigrams.Thedistractorsarethecomponentunigramsinthestembigram(searchandengine)andpseudo-unigramsthatshareacomponentwordwiththestem(searchwarrant,dieselengine).Toconstructtheholistictrainingquestions,weusedWordNetasaCorrectTraining7-choices14-choicesHolistic61.854.4Standard75.968.0Table12:ResultsforSuperSimwithholistictraining.sourceofbigrams,butweignoredtherichinfor-mationthatWordNetprovidesaboutthesebigrams,suchastheirsynonyms,hypernyms,hyponyms,meronyms,andglosses.Table12comparesholistictrainingtostandardtraining(thatis,trainingwithquestionslikeTable11versustrainingwithquestionslikeTable7).Thetestingsetisthestandardtestingsetinbothcases.Thereisasignificantdropinperformancewithholistictraining,buttheperformancestillsurpassesvectoraddition,element-wisemultiplication,andthehand-codeddual-spacemodel(seeTable8).Sinceholisticquestionscanbegeneratedauto-maticallywithouthumanexpertise,weexperi-mentedwithincreasingthesizeoftheholistictrain-ingdataset,growingitfrom1,000to10,000ques-tionsinincrementsof1,000.Theperformanceonthefourteen-choicequestionswithholistictrain-ingandstandardtestingvariedbetween53.3%and55.1%correct,withnocleartrendupordown.Thisisnotsignificantlydifferentfromtheperformancewith680holistictrainingquestions(54.4%).Itseemslikelythatthedropinperformancewithholistictraininginsteadofstandardtrainingisduetoadifferenceinthenatureofthestandardquestions(Table7)andtheholisticquestions(Table11).Wearecurrentlyinvestigatingthisissue.Weexpecttobeabletoclosetheperformancegapinfuturework,byimprovingtheholisticquestions.However,itispossiblethattherearefundamentallimitstoholistictraining.6DiscussionSuperSimperformsslightlybetter(notstatisticallysignificant)thanthehand-codeddual-spacemodelonrelationalsimilarityproblems(Section4),butitperformsmuchbetteroncompositionalsimilarityproblems(Section5).TheablationstudiessuggestthisisduetothePPMIfeatures,whichhavenoeffectonten-choiceSATperformance(Table5),buthavea l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 363 largeeffectonfourteen-choicenoun-modifierpara-phraseperformance(Table9).Oneadvantageofsupervisedlearningoverhand-codingisthatitfacilitatesaddingnewfeatures.Itisnotclearhowtomodifythehand-codedequationsforthedual-spacemodelofnoun-modifiercomposi-tion(Turney,2012)toincludePPMIinformation.SuperSimisoneofthefewapproachestodistri-butionalsemanticsbeyondwordsthathasattemptedtoaddressbothrelationalandcompositionalsimilar-ity(seeSection2.3).Itisastrengthofthisapproachthatitworkswellwithbothkindsofsimilarity.7FutureWorkandLimitationsGiventhepromisingresultswithholistictrainingfornoun-modifierparaphrases,weplantoexperimentwithholistictrainingforanalogies.Considertheproportionalanalogyhardistohardtimeasgoodistogoodtime,wherehardtimeandgoodtimearepseudo-unigrams.Toahuman,thisanalogyistriv-ial,butSuperSimhasnoaccesstothesurfaceformofaterm.AsfarasSuperSimisconcerned,thisanalogyismuchthesameastheanalogyhardistodifficultyasgoodistofun.Thisstrategyautomat-icallyconvertssimple,easilygeneratedanalogiesintomorecomplex,challenginganalogies,whichmaybesuitedtotrainingSuperSim.Thisalsosuggeststhatnoun-modifierparaphrasesmaybeusedtosolveanalogies.Perhapswecanevaluatethequalityofacandidateanalogyha,B,C,dibysearchingforatermesuchthathb,e,aiandhd,e,ciaregoodparaphrases.Forexample,considertheanalogymasonistostoneascarpenteristowood.Wecanparaphrasemasonasstoneworkerandcarpenteraswoodworker.Thistransformstheanalogytostoneworkeristostoneaswoodworkeristowood,whichmakesiteasiertorecognizetherelationalsimilarity.AnotherareaforfutureworkisextendingSuper-Simbeyondnoun-modifierparaphrasestomeasur-ingthesimilarityofsentencepairs.WeplantoadaptideasfromSocheretal.(2011)forthistask.Theyusedynamicpoolingtorepresentsentencesofvary-ingsizewithfixed-sizefeaturevectors.Usingfixed-sizefeaturevectorsavoidstheproblemofquadraticgrowthanditenablesthesupervisedlearnertogen-eralizeoversentencesofvaryinglength.SomeofthecompetingapproachesdiscussedbyErk(2013)incorporateformallogic.TheworkofBaronietal.(2012)suggestswaysthatSuperSimcouldbedevelopedtodealwithlogic.WebelievethatSuperSimcouldbenefitfrommorefeatures,withgreaterdiversity.Oneplacetolookforthesefeaturesishigherlevelsinthehierar-chythatwesketchinSection3.Ourablationexperimentssuggestthatdomainandfunctionspacesprovidethemostimportantfeaturesforrelationalsimilarity,butPPMIvaluesprovidethemostimportantfeaturesfornoun-modifiercomposi-tionalsimilarity.Explainingthisisanothertopicforfutureresearch.8ConclusionInthispaper,wehavepresentedSuperSim,aunifiedapproachtoanalogy(relationalsimilarity)andpara-phrase(compositionalsimilarity).SuperSimtreatsthembothasproblemsofsupervisedtupleclassifi-cation.Thesupervisedlearningalgorithmisastan-dardsupportvectormachine.ThemaincontributionofSuperSimisasetoffourtypesoffeaturesforrep-resentingtuples.Thefeaturesworkwellwithbothanalogyandparaphrase,withnotask-specificmod-ifications.SuperSimmatchesthestateoftheartonSATanalogyquestionsandsubstantiallyadvancesthestateoftheartontheSemEval2012Task2chal-lengeandthenoun-modifierparaphrasequestions.SuperSimrunsmuchfasterthanLRA(Turney,2006B),answeringtheSATquestionsinminutesinsteadofdays.Unlikethedual-spacemodel(Tur-ney,2012),SuperSimrequiresnohand-codedsimi-laritycompositionfunctions.Sincethereisnohand-coding,itiseasytoaddnewfeaturestoSuperSim.Muchworkremainstobedone,suchasincorporat-inglogicandscalinguptosentenceparaphrases,butpastworksuggeststhattheseproblemsaretractable.InthefourapproachesdescribedbyErk(2013),SuperSimisaninstanceofthesecondapproachtoextendingdistributionalsemanticsbeyondwords,comparingwordpairs,phrases,orsentences(ingen-eral,tuples)bycombiningmultiplepairwisesimi-larityvalues.Perhapsthemainsignificanceofthispaperisthatitprovidessomeevidenceinsupportofthisgeneralapproach. l D O w N O A D e D F R O M H T T P : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 364 ReferencesEnekoAgirre,DanielCer,MonaDiab,andAitorGonzalez-Agirre.2012.Semeval-2012Task6:Apilotonsemantictextualsimilarity.InProceedingsoftheFirstJointConferenceonLexicalandCompu-tationalSemantics(*SEM),pages385–393,Montr´eal,Canada.IonAndroutsopoulosandProdromosMalakasiotis.2010.Asurveyofparaphrasingandtextualentailmentmethods.JournalofArtificialIntelligenceResearch,38:135–187.MarcoBaroniandRobertoZamparelli.2010.Nounsarevectors,adjectivesarematrices:Representingadjective-nounconstructionsinsemanticspace.InProceedingsofthe2010ConferenceonEmpiricalMethodsinNaturalLanguageProcessing(EMNLP2010),pages1183–1193.MarcoBaroni,RaffaellaBernardi,Ngoc-QuynhDo,andChung-chiehShan.2012.Entailmentabovethewordlevelindistributionalsemantics.InProceedingsofthe13thConferenceoftheEuropeanChapteroftheAsso-ciationforComputationalLinguistics(EACL2012),pages23–32.IsaacI.Bejar,RogerChaffin,andSusanE.Embretson.1991.CognitiveandPsychometricAnalysisofAna-logicalProblemSolving.Springer-Verlag.ErgunBic¸iciandDenizYuret.2006.Clusteringwordpairstoansweranalogyquestions.InProceedingsoftheFifteenthTurkishSymposiumonArtificialIntelli-genceandNeuralNetworks(TAINN2006),Akyaka,Mugla,Turkey.DanushkaBollegala,YutakaMatsuo,andMitsuruIshizuka.2008.WWWsitstheSAT:Measuringrela-tionalsimilarityontheWeb.InProceedingsofthe18thEuropeanConferenceonArtificialIntelligence(ECAI2008),pages333–337,Patras,Greece.DanushkaBollegala,YutakaMatsuo,andMitsuruIshizuka.2009.MeasuringthesimilaritybetweenimplicitsemanticrelationsfromtheWeb.InProceed-ingsofthe18thInternationalConferenceonWorldWideWeb(WWW2009),pages651–660.JohnBullinariaandJosephLevy.2007.Extract-ingsemanticrepresentationsfromwordco-occurrencestatistics:Acomputationalstudy.BehaviorResearchMethods,39(3):510–526.JohnBullinariaandJosephLevy.2012.Extract-ingsemanticrepresentationsfromwordco-occurrencestatistics:Stop-lists,stemming,andSVD.BehaviorResearchMethods,44(3):890–907.JohnCaron.2001.ExperimentswithLSAscor-ing:Optimalrankandbasis.InProceedingsoftheSIAMComputationalInformationRetrievalWork-shop,pages157–169,Raleigh,NC.KennethChurchandPatrickHanks.1989.Wordasso-ciationnorms,mutualinformation,andlexicography.InProceedingsofthe27thAnnualConferenceoftheAssociationofComputationalLinguistics,pages76–83,Vancouver,BritishColumbia.DaoudClarke.2012.Acontext-theoreticframeworkforcompositionalityindistributionalsemantics.Compu-tationalLinguistics,38(1):41–71.KatrinErk.2013.Towardsasemanticsfordistributionalrepresentations.InProceedingsofthe10thInterna-tionalConferenceonComputationalSemantics(IWCS2013),Potsdam,Germany.JohnRupertFirth.1957.Asynopsisoflinguistictheory1930–1955.InStudiesinLinguisticAnalysis,pages1–32.Blackwell,Oxford.DanGarrette,KatrinErk,andRayMooney.2011.Inte-gratinglogicalrepresentationswithprobabilisticinfor-mationusingmarkovlogic.InProceedingsofthe9thInternationalConferenceonComputationalSemantics(IWCS2011),pages105–114.DedreGentner.1983.Structure-mapping:Atheoreticalframeworkforanalogy.CognitiveScience,7(2):155–170.RoxanaGirju,PreslavNakov,ViviNastase,StanSzpakowicz,PeterTurney,andDenizYuret.2007.Semeval-2007task04:Classificationofsemanticrelationsbetweennominals.InProceedingsoftheFourthInternationalWorkshoponSemanticEvalua-tions(SemEval2007),pages13–18,Prague,CzechRepublic.EmilianoGuevara.2010.Aregressionmodelofadjective-nouncompositionalityindistributionalsemantics.InProceedingsofthe2010WorkshoponGEometricalModelsofNaturalLanguageSemantics(GEMS2010),pages33–37.ZelligHarris.1954.Distributionalstructure.Word,10(23):146–162.IrisHendrickx,SuNamKim,ZornitsaKozareva,PreslavNakov,Diarmuid´OS´eaghdha,SebastianPad´o,MarcoPennacchiotti,LorenzaRomano,andStanSzpakow-icz.2010.Semeval-2010task8:Multi-wayclassifica-tionofsemanticrelationsbetweenpairsofnominals.InProceedingsofthe5thInternationalWorkshoponSemanticEvaluation,pages33–38,Uppsala,Sweden.Amac¸Herda˘gdelenandMarcoBaroni.2009.Bagpack:Ageneralframeworktorepresentsemanticrelations.InProceedingsoftheEACL2009GeometricalModelsforNaturalLanguageSemantics(GEMS)Workshop,pages33–40.NathalieJapkowiczandShajuStephen.2002.Theclassimbalanceproblem:Asystematicstudy.IntelligentDataAnalysis,6(5):429–449.DavidA.Jurgens,SaifM.Mohammad,PeterD.Tur-ney,andKeithJ.Holyoak.2012.SemEval-2012 l D o w n o a d e d f r o m h t t p : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 365 Task2:Measuringdegreesofrelationalsimilarity.InProceedingsoftheFirstJointConferenceonLexi-calandComputationalSemantics(*SEM),pages356–364,Montr´eal,Canada.ThomasK.LandauerandSusanT.Dumais.1997.AsolutiontoPlato’sproblem:Thelatentseman-ticanalysistheoryoftheacquisition,induction,andrepresentationofknowledge.PsychologicalReview,104(2):211–240.BenoˆıtLemaireandGuyDenhi`ere.2006.Effectsofhigh-orderco-occurrencesonwordsemanticsimilar-ity.CurrentPsychologyLetters:Behaviour,Gehirn&Cognition,18(1).YvesLepageandAndoShin-ichi.1996.Saussuriananalogy:Atheoreticalaccountanditsapplication.InProceedingsofthe16thInternationalConferenceonComputationalLinguistics(COLING1996),pages717–722.KevinLund,CurtBurgess,andRuthAnnAtchley.1995.Semanticandassociativepriminginhigh-dimensionalsemanticspace.InProceedingsofthe17thAnnualConferenceoftheCognitiveScienceSociety,pages660–665.TomasMikolov,Wen-tauYih,andGeoffreyZweig.2013.Linguisticregularitiesincontinuousspacewordrepresentations.InProceedingsofthe2013Confer-enceoftheNorthAmericanChapteroftheAssocia-tionforComputationalLinguistics:HumanLanguageTechnologies(NAACL2013),Atlanta,Georgia.JeffMitchellandMirellaLapata.2008.Vector-basedmodelsofsemanticcomposition.InProceedingsofACL-08:HLT,pages236–244,Columbus,Ohio.AssociationforComputationalLinguistics.JeffMitchellandMirellaLapata.2010.Compositionindistributionalmodelsofsemantics.CognitiveScience,34(8):1388–1429.TedPedersen.2012.Duluth:Measuringdegreesofrelationalsimilaritywiththeglossvectormeasureofsemanticrelatedness.InFirstJointConferenceonLexicalandComputationalSemantics(*SEM),pages497–501,Montreal,Canada.JohnC.Platt.1998.Fasttrainingofsupportvectormachinesusingsequentialminimaloptimization.InAdvancesinKernelMethods:SupportVectorLearn-ing,pages185–208,Cambridge,MA.MITPress.BryanRinkandSandaHarabagiu.2012.UTD:Deter-miningrelationalsimilarityusinglexicalpatterns.InFirstJointConferenceonLexicalandComputationalSemantics(*SEM),pages413–418,Montreal,Canada.BryanRinkandSandaHarabagiu.2013.Theimpactofselectionalpreferenceagreementonsemanticrela-tionalsimilarity.InProceedingsofthe10thInterna-tionalConferenceonComputationalSemantics(IWCS2013),Potsdam,Germany.RichardSocher,EricH.Huang,JeffreyPennington,AndrewY.Ng,andChristopherD.Manning.2011.Dynamicpoolingandunfoldingrecursiveautoen-codersforparaphrasedetection.InAdvancesinNeuralInformationProcessingSystems(NIPS2011),pages801–809.RichardSocher,BrodyHuval,ChristopherManning,andAndrewNg.2012.Semanticcompositionalitythroughrecursivematrix-vectorspaces.InProceed-ingsofthe2012JointConferenceonEmpiricalMeth-odsinNaturalLanguageProcessingandComputa-tionalNaturalLanguageLearning(EMNLP-CoNLL2012),pages1201–1211.MireyaTovar,J.AlejandroReyes,AzucenaMontes,DarnesVilari˜no,DavidPinto,andSaulLe´on.2012.BUAP:Afirstapproximationtorelationalsimilar-itymeasuring.InFirstJointConferenceonLexi-calandComputationalSemantics(*SEM),pages502–505,Montreal,Canada.PeterD.TurneyandMichaelL.Littman.2005.Corpus-basedlearningofanalogiesandsemanticrelations.MachineLearning,60(1-3):251–278.PeterD.TurneyandPatrickPantel.2010.Fromfre-quencytomeaning:Vectorspacemodelsofsemantics.JournalofArtificialIntelligenceResearch,37:141–188.PeterD.Turney,MichaelL.Littman,JeffreyBigham,andVictorShnayder.2003.Combiningindependentmod-ulestosolvemultiple-choicesynonymandanalogyproblems.InProceedingsoftheInternationalCon-ferenceonRecentAdvancesinNaturalLanguagePro-cessing(RANLP-03),pages482–489,Borovets,Bul-garia.PeterD.Turney,YairNeuman,DanAssaf,andYohaiCohen.2011.Literalandmetaphoricalsenseidentifi-cationthroughconcreteandabstractcontext.InPro-ceedingsofthe2011ConferenceonEmpiricalMeth-odsinNaturalLanguageProcessing,pages680–690.PeterD.Turney.2006a.Expressingimplicitsemanticrelationswithoutsupervision.InProceedingsofthe21stInternationalConferenceonComputationalLin-guisticsand44thAnnualMeetingoftheAssociationforComputationalLinguistics(Coling/ACL-06),pages313–320,Sydney,Australia.PeterD.Turney.2006b.Similarityofsemanticrelations.ComputationalLinguistics,32(3):379–416.PeterD.Turney.2008a.Thelatentrelationmappingengine:Algorithmandexperiments.JournalofArtifi-cialIntelligenceResearch,33:615–655.PeterD.Turney.2008b.Auniformapproachtoanalo-gies,synonyms,antonyms,andassociations.InPro-ceedingsofthe22ndInternationalConferenceonComputationalLinguistics(Coling2008),pages905–912,Manchester,Vereinigtes Königreich. l D O w N O A D e D F R O M H T T P : / / D ich R e C T . M ich T . e d u / t a c l / l A R T ich C e - P D F / d o i / . 1 0 1 1 6 2 / t l a c _ a _ 0 0 2 3 3 1 5 6 6 6 7 5 / / t l a c _ a _ 0 0 2 3 3 P D . F B j G u e S T T O N 0 8 S e P e M B e R 2 0 2 3 366 PeterD.Turney.2012.Domainandfunction:Adual-spacemodelofsemanticrelationsandcompositions.JournalofArtificialIntelligenceResearch,44:533–585.TonyVeale.2004.WordNetsitstheSAT:Aknowledge-basedapproachtolexicalanalogy.InProceedingsofthe16thEuropeanConferenceonArtificialIntel-ligence(ECAI2004),pages606–612,Valencia,Spain.IanH.Witten,EibeFrank,andMarkA.Hall.2011.DataMining:PracticalMachineLearningToolsandTech-niques,ThirdEdition.MorganKaufmann,SanFran-cisco.AlisaZhila,Wen-tauYih,ChristopherMeek,GeoffreyZweig,andTomasMikolov.2013.Combininghet-erogeneousmodelsformeasuringrelationalsimilar-ity.InProceedingsofthe2013ConferenceoftheNorthAmericanChapteroftheAssociationforCom-putationalLinguistics:HumanLanguageTechnolo-gies(NAACL2013),Atlanta,Georgia.
PDF Herunterladen