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
2.2.IBMmodel1weight
Beingindependentofwordposition,IBMmodel1probabilitiesprovideaprobabilisticlexiconbetweenpairsofwordofeachlan-guage(see[8]fordetailsinthesemodels).Thisinformationcanbeusedtoweightandcomparetheresultingtuplesfromtwocom-petingsegmentations,isin[6].
Whilethisapproachisappealinginthatitusesbilingualinfor-mation,observationofthesesituationsleadstoadifferentconclu-sion;manyNULL-linkedwordsrepresentarticles,prepositions,conjunctionsandotherparticleswhosemainfunctionistoensurethegrammaticalcorrectnessofasentence,complementingothermoreinformativewords.Therefore,theirprobabilitiestotranslatetoanotherwordarenotverymeaningful.2.3.EntropyofthePOSdistribution
Alternatively,fromalinguisticpointofview,onecanthinkofthistuplesegmentationproblemaroundsourceNULLsasamonolin-gualdecisionrelatedtowhetherthegiventargetwordismorecon-nectedtotheprecedingortothefollowingword.Intuitively,wecanexpectthatagoodcriteriontoperformtuplesegmentationlaysinpreservinggrammatically-connectedphrases(suchas,forin-stance,articlestogetherwiththenountheyprecede)inthesametuple,asthismayprobablyleadtoasimpli cationofthetransla-tiontask.Onthecontrary,splittinglinguisticunitsintoseparatetu-pleswillprobablyleadtoatuplevocabularyincreaseandahighersparseness,producingapoorerNgrammodel.
Inthisdirection,weproposetotakethesegmentationdecisionaccordingtotheentropyoftheforwardandbackwardPart-Of-Speech(POS)distributions,whichwede neconditionedtocon-text.Indetail,giventhetuplesequencedescribedasfollows:
<...sj>
|
<...ti 1>
NULL|ti
<sj+1...>
|
<ti+1...>
Then,wecantakeatuplesegmentationdecisionfavouringthemostPOS-entropiccase.Therationalebehindthisisthat,iffbHPOS>HPOS,wehaveobservedthe rstsequenceofwordscomprisedof(ti 1,ti)inmoregrammaticallydifferentsituationsthanthelattersequencecomprisedof(ti,ti+1).Therefore,wecaninducethatti 1andtitendtobemoreoftenconnectedthantiandti+1,andshouldbelongtothesametranslationtuple.Anal-fb
ogously,onecanconcludethecontraryifHPOS<HPOS.
Whilethisisamonolingualdecisiononthetargetlanguagemorphology,beinglinguistically-guided,thePOSentropyap-proachismuchmorecorrelatedwithhumanintuition.2.4.CriteriaforNULLsintarget
Whereasthesegmentationdecisionisrequiredwhenatargetwordisunlinked(orlinkedtoNULL),thisisnotsowhentheunlinkedwordisinthesourcetarget,inwhichcasetheseunitsareallowedinthetuplevocabularyforNgramestimation.
However,onecanthinkofapplyingthesamecriteriontore-moveNULLsinthetargetsideoftuples,possiblyaddressingomis-sionerrorsintranslation.Aimingatevaluatingtheimpactofthisdecision,wehavealsoappliedthePOSentropystrategytoseg-menttupleswithunlinkedsourcewords.
3.Experimentalframework
Inordertocompareeachsegmentationstrategyandevaluateitsimpactontranslationquality,experimentshavebeencarriedoutusingtwoparallelcorpora,differinginlanguagepairandcor-pussize.Ontheonehand,weusedaSpanish–Englishlarge-vocabularycorpus,containingtheEuropeanParliamentProceed-ingsfrom1996toSeptember2004,andontheother,anArabic-to-Englishsmall-vocabularycorpus,whichcontainsasmallpartoftheBasicTravelExpressionsCorpus.EnglishhasbeentaggedusingTnTtagger1,andSpanishusingFreeLinganalysistool2.
sent
TrainsetEnglishSpanishDevsetEnglishSpanishTestsetEnglishSpanish
1.22M5048401094
words33.37M34.96M15.3k15.4k22.75k26.88k
vocab104.8k151.4k2.30k2.74k4.1k4.0k
avglen27.328.6OOVs2022OOVs44113
refs13322
wheresjmeanswordinpositionjinsourcesentence,andequiva-lentlytimeanswordinpositioniintargetsentence,wecande nea’forward’entropyofthePOSdistributioninpositioni+1given(ti 1,ti)asinequation2:
Xf
f
HPOS= pPOSlogpf(2)POS
POS
where
pfPOS
N(ti 1,ti,POSi+1)=P
POS N(ti 1,ti,POSi+1)
(3)
istheprobabilityofobservingacertainPart-Of-Speechfollowingthesequenceofwordsde nedbytiandti+1.
Equivalently,wecande nea’backward’entropyofthePOSdistributioninpositioni 1given(ti,ti+1)asinequation4:
Xb
bHP= pPOSlogpb(4)OSPOS
POS
Table1:EuropeanParliamentEnglish-Spanishcorpusstatistics.StatisticsofeachcorpuscanbefoundinTables1and2,in-cludingnumberofsentences,runningwords,vocabularysizes,av-eragesentencelengthandout-of-vocabulary(OOVs)wordsand
numberofreferencetranslationsfordevelopmentandtestsets.Notethebigsizedifferencebetweentasks.3.1.Translationmodelresults
where
pbPOS=P
N(POSi 1,ti,ti+1)
POS N(POSi 1,ti,ti+1)
(5)
AcomparisonofNgramtranslationmodelperformanceforeachtaskisshowninrowsnamed’alwaysNEXT’,’IBM1weight’and
1Available2Available
istheprobabilityofobservingacertainPart-Of-Speechprecedingthesequenceofwordsde nedbyti 1andti.
atdoc.guandang.net~thorsten/tntatdoc.guandang.net
热门文档