Ngram-based Statistical Machine Translation relies on a standard Ngram language model of tuples to estimate the translation process. In training, this translation model requires a segmentation of each parallel sentence, which involves taking a hard decisio
Linguistictuplesegmentationinngram-basedstatisticalmachinetranslation
Adri`adeGispert,Jos´eB.Mari no
TALPResearchCenter
UniversitatPolit`ecnicadeCatalunya(UPC),Barcelona
{agispert|canton}@gps.tsc.upc.edu
Abstract
Ngram-basedStatisticalMachineTranslationreliesonastandard
Ngramlanguagemodeloftuplestoestimatethetranslationpro-cess.Intraining,thistranslationmodelrequiresasegmentationofeachparallelsentence,whichinvolvestakingaharddecisionontuplesegmentationwhenawordisnotlinkedduringwordalign-ment.Thisisespeciallycriticalwhenthiswordappearsinthetargetlanguage,asthisharddecisioniscompulsory.
Inthispaperwepresentathoroughstudyofthissituation,comparingforthe rsttimeeachoftheproposedtechniquesintwoindependenttasks,namelyEnglish–SpanishEuropeanParlia-mentProceedingslarge-vocabularytaskandArabic–EnglishBasicTravelExpressionssmall-datatask.Inthefaceofthiscomparison,wepresentanovelsegmentationtechniquewhichincorporateslin-guisticinformation.Resultsobtainedinbothtasksoutperformallprevioustechniques.
IndexTerms:statisticalmachinetranslation,tuplesegmentation,n-gram-basedSMT,linguisticinformation
cannotbeallowed,andacertainharddecisionmustbetakenre-gardingtuplesegmentation,asin gure1.
Literatureoffersexamplesofcriteriatodecidethissegmenta-tion,rangingfromsimplylinkingthetargetwordsoftheNULL-sourcetupletothepreviousornexttuple(ifthereisany)deter-ministically,tocomparingtheIBMmodel1scoreassignedtotheresultingtuplesoftwocompetingsegmentations

[6].
Figure1:AhardsegmentationdecisionmustbetakenHowever,toourknowledgetheimpactofthissegmentationdecisionontranslationqualityisnotstudied,northealternativesegmentationstrategiesarecompared.Here,wecarryoutthiscomparison,exploringtheimpactofthisharddecisioninNgramtranslationmodelestimationandintranslationquality,andwepro-poseanovelstrategywhichfollowsalinguisticcriteriatoindi-rectlyreducethemodelentropy.WeperformthiscomparisoninanEnglish-to-Spanish(andviceversa)large-vocabularytask,andadditionallyinanArabic-to-Englishsmall-vocabularytask.Fur-thermore,westudytheeffectontranslationofNULLtokenswhentheseoccurinthetargetsideofthetuple.
1.Introduction
Ngram-based(orTuple-based)StatisticalMachineTranslation(SMT)hasprovedtobeastate-of-the-artalternativeapproachtophrase-basedmodelsinperformancecomparisons[1,2].ItsmaindistinctivetreatistheestimationofthecoretranslationmodelbymeansofastandardNgramlanguagemodel,de nedonthespecialbilinguallanguageexpressedbytuples[3].
Accordingtoliterature,tuplesareunitscontainingoneormoresource-languagewordsandoneormoretarget-languagewords(in-cludingtheNULLtoken,whichisinfactnoword).ThismodelhasitsoriginsinmachinetranslationusingFinite-StateTransduc-ers[4],whosetheoreticalfoundationexpressesthatweseekthetargetsentence’e’maximising:
NYn=1
2.TupleSegmentationstrategies
AccordingtotheaforementionedconceptualframeworkofNgram
translationmodel,itseemsclearthattheidealtuplesegmentationstrategyshouldtakeaglobaldecisionbasedonthesegmentationforallotherNULL-sourcecases,attemptingtoobtainthatsetoftuplesandNgramswhichbetterrepresentedtheunseenuniverseofevents,meaningtheonewithlessentropy.However,nofeasiblealgorithmcanperformthatcalculationinareasonabletimegivencurrentcomputationalpower,asthiswouldinvolveawholemodelreestimationforeachparticularsegmentationalternative.
SofartwosegmentationstrategiestosolvethesourceNULLproblemhavebeenproposed,whicharepresentedinthenextsec-tions,togetherwithanovelstrategybasedonalinguisticcriterion.2.1.DeterministicalwaysNEXT
Averypragmaticandsimpleapproachistoalwaysjointhetar-getwordsinvolvedinNULLlinks(NULL-linkedwords)tothenexttuple,ifthereisany(otherwisetothepreviousone),as rstdonein[7].Besidessimplicityandextremeef ciency,wedonotobserveanyotheradvantageofthisapproach,whichontheotherhanddoesnotfollowanylinguisticorstatisticalcriterion.
p((f,e)n|(f,e)n x+1,...,(f,e)n 1)
(1)
wherethen-thtupleofasentencepairisreferredas(f,e)n.Inordertoestimatethislanguagemodelparameters,givenaparallelcorpusandacertainwordalignment,andincontrasttophrase-basedapproaches,thetrainingrequiresauniquesegmentationofeachbilingualpairofsentencesintoasequenceoftuplessothatthenaturalorderofbothlanguagesisnotviolated[5].
Thestandardtupleextractionalgorithm[5]de nesauniquesetoftuplesexceptwhenevertheresultingtuplecontainsnosourceword(NULL-sourcetuple).Inordertore-usetheseunitsinde-codingnewsentences,thesearchshouldallowfornoinputwordtogenerateunits,andthisisnotthecase.Therefore,theseunits
热门文档