[PDF] Lexical Query Paraphrasing for Document Retrieval

View - Download Lexical Query Paraphrasing for Document Retrieval

Previous PDF

Next PDF

IngridZukerman

SchoolofComputerScienceandSoftwareEng.

MonashUniversity

Clayton,VICTORIA3800

AUSTRALIABhavaniRaskutti

TelstraResearchLaboratories

770BlackburnRoad

Clayton,VICTORIA3168

AUSTRALIA

Abstract

Wedescribeamechanismforthegenerationof

thenusedtoranktheparaphrases.Weevaluated ourmechanismusing404querieswhoseanswers resideintheLATimessubsetoftheTREC-9cor- pus.Therewasa14%improvementinperfor- mancewhenparaphraseswereusedfordocument retrieval.

1Introduction

paraphraseswillmatcharelevantdocument. onyms.WeuseWordNet(Milleretal.,1990)and asaresultofqueryparaphrasing. ourevaluationandconcludingremarks.

CouncilgrantDP0209565.2RelatedResearch

Thevocabularymis-matchbetweenuserqueries

queryexpansion.Twocommontechniquesfor queryexpansionareblindrelevancefeedback (Buckleyetal.,1995;Mitraetal.,1998)and wordsensedisambiguation(WSD)(Mihalceaand mentsusingaquerygivenbyauser,andthencon- documents.WSDoftenprecedesqueryexpansion usedautomaticallyconstructedthesauri.

Theimprovementsinretrievalperformancere-

portedin(Mitraetal.,1998)arecomparableto disambiguationerrors. theaboveapproachesinthattheexpansionofa phrases.LikeHarabagiuetal.(2001),weuse

WordNettoproposesynonymsforthewordsin

whichwordstoparaphrase.Incontrast,weuse trievalprocess.

3Resources

onymsforthewordsineachquerywasobtained fromWordNet(Milleretal.,1990).

TimesportionoftheNISTTextResearchCollec-

tion(//trec.nist.gov).Thiscorpuswassmall formedforthedocumentsintheLATimescollec- obtainedfromWordNet(Milleretal.,1990). timesitappearsinthecorpus.Thelemma-pair ???thenumberoftimes ???appearsbefore ???in afive-wordwindowinthecorpus(notcounting maintainsadifferententryforthelemmapair naryowingtodiskspacelimitations.

4ParaphrasingandRetrievalProcedure

thefollowingsteps:

1.Tokenize,tagandlemmatizethequery.

2.Generatesynonymsforeachcontentlemmain

thequery(stopwordsareignored). entsynonymcombinations,computeascorefor eachparaphrase,andranktheparaphrasesac- cordingtotheirscore.Thelemmatizedquery plusthe19topparaphrasesareretained.

Documentsarethenretrievedforthequeryand

strainthenumberofsynonymsgeneratedforit.

Inordertodeterminetheeffectoftagginger-

allythewrongtags,andranoursystemwithboth

4.2Proposingsynonymsforeachword

ThefollowingtypesofWordNetsynonymswere

generatedforeachcontentlemmainaquery: synonyms,attributes,pertainymsand seealsos(Milleretal.,1990).1Forexample, accordingtoWordNet,asynonymfor"high"is anddonotgeneratesynonymsforpropernounsor stopwords.

4.3Paraphrasingqueries

inturn,andproposesasynonymfromthosecol-

Thesearequerieswhereallthewordsexceptone

arestopwordsorclosed-classwords.

4.4Computingparaphrasescores

Thescoreofaparaphraseisbasedonhowcommon

wouldberepresentedbyPr ditionalprobabilitiesareoftenused,e.g., Pr theinteractionbetweenalemma ???andonlyone oftheresultsinmostcases. abilities).Forinstance,if ??appears10timesinthe corpusand ???appears4times, (where ?isanormalizingconstant).Incontrast, if ???appears200timesinthecorpusand ???ap- pears30times, contributeahigherscoretotheparaphrase.

Toaddresstheseproblems,weproposeusingthe

yields posedoflemmas Pr Pr frequencies,yielding Pr ??freq? where ?isanormalizingconstant.2Sincethiscon- wedropitfromconsideration,anduseonlythe

Thus,ourparaphrasescoringfunctionis

4.4.1Experimentalparameters

Whencalculatingthescoreofaparaphraseus-

freq theorderof ??and ??(asitappearsintheparaphrase) shouldbeenforced;and(2)howtohandle ??pairs asexperimentalparametersofthesystem. 2???? forcetheorderof ???whencalculatingfreq? isdeterminedbytheweight?orderasfollows: freq ?freq? (3) wherefreq ??????isthefrequencyofthelemma- pair ??????when ??isfollowedby ??.?order ?order ???countsequallytheorderinthepara- weightsof0,1and0.5for ?order(Section6).

Absentlemma-pairs.Whenalemma-pairisnot

66millionpairs-hadafrequencyof1butwere

thedictionary.

4.5Retrievingdocumentsforeachquery

inthatforeachquery ?,weadjustthescoresofthe paraphrasesof ?(obtainedfromEquation2).Our

1.Foreachparaphrase

??of?(? #para?),where ???isthelemmatizedquery: (a)Extractthecontentlemmasfrom contentlemmasinparaphrase (b)Foreachlemma,computeascoreforthere- trieveddocumentsusingastandardIRmea- sure,e.g.,TermFrequencyInverseDocument

Frequency(TFIDF)(SaltonandMcGill,

1983).Lettfidf

????bethescoreof document???retrievedforlemma ).Whenadocument???isretrieved bymorethanonelemmainaparaphrase ?,itsTFIDFscoresareadded,yieldingthe score ???tfidf????? ????.Thisscoreindi- cateshowwell ???matchesthelemmasin paraphrase ??.Inordertotakeintoaccount theplausibilityof ??,thisscoreismultiplied by ???-thescoreof ??obtainedfrom

Equation2.Thisyields

??,thescoreof document ???forparaphrase ?????tfidf????? ????(4)

2.Foreachdocument

???,addthescoresfromeach paraphrase(Equation4),yielding para?????? tfidf???? ????(5)

Anoutcomeofthismethodisthatlemmas

performance(Section6).

5SampleResults

obtainedwith ?order ??,AbsAdjDiv=10,and werenotfoundinthedictionary. spitetheirlowoverallscoreandabsentlemma- forthelemmasinthisquery.Thesecondquery onymfor"invent"and"video"asasynonymfor rectparaphrases.Thethirdqueryisanextreme exampleofthisbehaviour,whereWordNetsyn-

Score#AbsParaphrase

WhoistheGreekGodoftheSea?

9.20E+020whobethegreekgodofthesea?

6.90E+001whobethegreekgodoftheocean?

5.00E-011whobethegreecegodofthesea?

1.00E-022whobethegreecedeityofthesea?

1.00E-022whobethegreecedivinityofthesea?

1.00E-022whobethegreeceimmortalofthesea?

1.00E-022whobethegreeceidolofthesea?

8.00E-032whobethegreekdeityofthesea?

8.00E-032whobethegreekdivinityofthesea?

8.00E-032whobethegreekimmortalofthesea?

8.00E-032whobethegreekidolofthesea?

Whoinventedtelevision?

7.00E+000whoinventtelevision?

1.60E+010whomanufacturetelevision?

1.60E+010whomanufacturevideo?

1.10E+010whomanufacturetv?

9.00E+000whoinventtv?

2.00E+000whodevisetelevision?

2.00E+000whoforgetv?

1.00E-021whoinventvideo?

1.00E-021whoinventtelly?

1.00E-021whocontrivetelevision?

1.00E-021whocontrivetv?

WhenwasBabeRuthborn?

6.06E+030whenbebaberuthbear?

3.39E+040whenbebaberuthpay?

1.97E+040whenbebaberuthstand?

1.09E+040whenbebaberuthhold?

2.42E+030whenbebaberuthcarry?

1.21E+030whenbebaberuthhave?

4.24E+021whenbebaberuthsupport?

9.09E+011whenbebaberuthexpect?

6.06E+001whenbebaberuthbrook?

6.06E+001whenbebaberuthwear?

3.03E-012whenbebaberuthdeliver?

Howtallisthegiraffe?

4.00E+000howtallbethegiraffe?

2.00E+000howlargebethegiraffe?

2.00E+000howbigbethegiraffe?

2.00E+000howhighbethegiraffe?

1.00E-011howgrandiloquentbethegiraffe?

1.00E-011howmagniloquentbethegiraffe?

1.00E-011howimprobablebethegiraffe?

1.00E-011howmarvelousbethegiraffe?

6Evaluation

ontheTRECLATimescollection,usingTREC task,sinceourultimategoalistoanswerques- where1105documentswerejudgedrelevantto48 ofthe50TREC-6keyword-basedqueries.

Ourresultsshowthatqueryparaphrasingim-

Forthequestionansweringtask,underthesame

wereretrievedimprovedfrom169to182. factorsonretrievalperformance. paraphrasegeneration). versusautomaticPoStagging(Brill,1992), whichtaggedcorrectly84%ofthequeries. ?Out-of-orderweight(?order)-howmuchwe shouldtakeintoaccountthewordorderina intermediate). muchweshouldpenalizelemma-pairsthatare adjacentinthequerybutabsentfromthecorpus (samepenaltyasnon-adjacentabsentlemma- pairs,alittlehigher,alothigher). ?Querylength-howthenumberofwordsinthe queryaffectsretrievalperformance.

1paraphrase(Set1),thenthequeryplus2para-

phrases(Set2),andsoon,uptoamaximumof trievalenginefrom1to20documents.

6.1WordNetCo-locations

05101520290

295
300
305
310
315
320
325
330
335

Number of paraphrases

Total number of correct documents

Correct Documents Vs Number of Paraphrases

Col

ColScore

NoCol paraphrases(20retrieveddocuments)

NoColandColScore.UndertheColsetting,our

mechanismcheckedwhetheralemma-pairinthe onymssuchas"folate"and"vitaminm"forthe

ColScoreisahybridsetting,whereWordNetwas

phrases,butnotforgeneratingthem. asafunctionofthenumberofparaphrasesina set(from0to19).Thevaluesfortheotherfac- torswere: ?order=1,AbsAdjDiv=2,andmanually- trievedwhenonlythelemmatizedquerywassub- bersofparaphrases,whenthemaximumnumberof

051015200

50
100
150
200
250
300
350

Number of retrieved documents

Total number of correct documents

Correct Documents Vs Number of Retrieved Documents

NoPara

Col

ColScore

NoCol retrieveddocuments(maximumparaphrases) trievedperquery(from1to20).AsforFigure1, moredocumentsareretrieved. as"vitaminm"and"vitaminbc",whichappeared "bigleague"inonly3ofthe19paraphrases)en- restofourevaluation.6.2Taggingaccuracy

ThePoS-taggerincorrectlytagged64ofthe404

mis-taggingwhichhadthelargestimpactonthe asotherPoSandviceversain24cases,andthe hadanounmis-taggedasanotherPoS.

6.3Out-of-orderweight

weight, ?order(Equation3):1,0and0.5.The thequery"howmanydogspullasledintheIdi- theweightoftheorderedpairs. theirorderinthecorpus.Thus,whenanordered

6.4Penaltyforabsentadjacentlemma-pairs

foreachabsentadjacentlemma-pair. tainedforAbsAdjDiv=10.quotesdbs_dbs30.pdfusesText_36

[PDF] comment utiliser op-cit et ibid

[PDF] quand mettre un alinéa mémoire

[PDF] citer un cours

[PDF] citer ses sources

[PDF] svt 3eme l homme face aux micro organismes controle

[PDF] différence entre bactérie et virus svt

[PDF] la diversité des microorganismes

[PDF] comment la constitution protège les droits fondamentaux

[PDF] en france comment les droits de l'homme sont ils protégés

[PDF] garantie des droits définition

[PDF] garantie des droits fondamentaux

[PDF] quels sont les droits fondamentaux

[PDF] les acteurs transnationaux dans les relations internationales

[PDF] isonomie

[PDF] l'organisation de la défense nationale schéma