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
alwaysNEXTIBM1weightPOSentropytrgNULLtrgNULLpos
tupvcb2.11M2.04M2.08M2.35M2.18M
Eng→Spa
%1–2–3grams17.6–44.4–38.018.0–44.7–37.317.8–44.3–37.923.2–45.1–31.719.0–44.5–36.5
tuplen1.16-1.101.16-1.091.16-1.111.25-1.191.18-1.14null31192466228201625tupvcb2.15M2.08M2.11M2.42M2.16M
Spa→Eng
%1–2–3grams14.1–41.5–44.414.2–41.4–44.414.2–41.5–44.319.9–44.1–36.014.7–41.6–43.7
tuplen1.14-1.061.13-1.051.13-1.061.26-1.221.14-1.08null27612318219401977
Table4:TuplevocabularyandNgramtranslationstatisticsforeachsegmentationstrategy.
NULLsintargetdidnotyieldimprovements,itappearsthatthebestperformingstrategyistheonethat’uses’lessunitstoNULLintranslation.Afurtherstudyofthereasonsbehindthisistobecarriedout,tryingtounderstandtherealcontributionofthisunitsandhoworwhetherfurtherimprovementscouldbeachievedbyremovingsomeofthem.
Finally,otherissuesrelatedtothetuplengramtranslationmodelarepruningstrategiesandsmoothingtechniques.Thoughoutofthescopeofthispaper,theyarealsocrucialtounderstandthecoremodelofngram-basedSMT.Therefore,athoroughstudyoftheirrespectiveimpactintranslationqualityistobecarriedout.
E→SS→EA→E
IBM1weightPOSentropyIBM1weightPOSentropyalwaysNEXTPOSentropy
BLEU0.47140.47440.54700.54660.39740.4024mWER40.2240.5634.4134.4440.1640.05NIST9.839.8510.7410.727.237.39
Table5:Resultswithadditionalfeatures.
Inthelarge-vocabularySpa-Engtasksthetargetlanguageandlexiconmodelsaddrobustnesstothewholesystembypenalisingtupleswith’wrong’segmentations,oratleasttheircatenationtobuildupthetranslationoutput.YettheproposedsegmentationachievesslightlybetterresultsintheEng→Spadirection.
However,small-vocabularytasksseemmoresensitivetoseg-mentationevenwhencombiningthecoretranslationmodelwithadditionalfeatures,asimprovementsaremoresigni cantandhigherthanusingonlythetranslationmodel.
5.Acknowledgements
ThisworkwaspartlyfundedbyEuropeanUnion(TC-STARproject,FP6-506738),GeneralitatdeCatalunyaandEuropeanSo-cialFund.
6.References
[1]P.KoehnandC.Monz,“Sharedtask:StatisticalMachine
TranslationbetweenEuropeanLanguages,”Proc.oftheACLWorkshoponBuildingandUsingParallelTexts(ACL’05),pp.119–124,2005.[2]M.EckandC.Hori,“OverviewoftheIWSLT2005Evalua-tionCampaign,”Proc.oftheInt.WorkshoponSpokenLan-guageTranslation,IWSLT’05,pp.11–32,2005.[3]J.Mari no,R.Banchs,J.Crego,A.deGispert,P.Lambert,
M.R.Costa-juss`a,andJ.Fonollosa,“BilingualN–gramsta-tisticalmachinetranslation,”Proc.oftheMTSummitX,pp.275–282,2005.[4]E.Vidal,“Finite-statespeech-to-speechtranslation,”Proc.of
1997IEEEInt.Conf.onAcoustics,SpeechandSignalPro-cessing,pp.111–114,1997.[5]J.M.Crego,J.Mari no,andA.deGispert,“Finite-state-based
andphrase-basedstatisticalmachinetranslation,”Proc.ofthe8thInt.Conf.onSpokenLanguageProcessing,ICSLP’04,pp.37–40,2004.[6]J.M.Crego,J.Mari no,andA.Gispert,“TALP:TheUPC
tuple-basedSMTsystem,”Proc.oftheInt.WorkshoponSpo-kenLanguageTranslation,IWSLT’05,pp.191–198,2005.[7]A.deGispertandJ.Mari no,“TALP:Xgram-basedSpoken
LanguageTranslationSystem,”Proc.oftheInt.WorkshoponSpokenLanguageTranslation,IWSLT’04,pp.85–90,2004.[8]P.Brown,S.DellaPietra,V.DellaPietra,andR.Mercer,
“Themathematicsofstatisticalmachinetranslation,”Compu-tationalLinguistics,vol.19,no.2,pp.263–311,1993.
4.Conclusion
Thispaperdelvesintothedetailsoftuplesegmentation,aneces-sarystepwhentrainingNgram-basedSMTsystems.Apartfromcomparingallpreviouslypresentedsegmentationcriteria,anovelstrategybasedontheentropyofthePOSdistributionisproposed.
A rstconclusionofourstudyisthattranslationmodelper-formanceissigni cantlyaffectedbytuplesegmentation,andtheimpactofsegmentationstrategiesdependsonlanguagepairandcorpusused,growingaccordingtothepercentageoftupleswithNULLinsourceside.Secondly,theproposedlinguisticcrite-rionperformssigni cantlybetterthansegmentationstrategiesal-readypresentedinliterature,behavingconsistentlyacrossdifferenttranslationtasks.
RegardingNULLsintargetside,weconcludethattheypro-videcontexttothemodelandremovingthemthroughresegmen-tationisnotbene cial.Finally,whentranslationmodelislog-linearlycombinedwithotherfeatures,directimpactofsegmenta-tionissmallerforlarge-vocabularytasks,asfeaturescanpartiallycompensateabadtuplesegmentation.
Possibly,themaindrawbackoftheproposedsegmentationstrategyisthatitrequiresaPOStaggingtool.Inthisdirection,onecanthinkofmakingitindependentfromthisbyusinginducedclassi cationstrategies.Oppositetothat,ifchunkingtoolsareavailable(forexampleinEnglish),wewouldliketoinvestigatewaysofusingchunkinginformationtobettersegmenttuples.
AnotherinterestingissueisrelatedtoNULLsintargetside.Whereasthepresentedexperimentspartiallyorfullyremoving
热门文档