资源描述:
《candide-机器翻译系统》由会员上传分享,免费在线阅读,更多相关内容在教育资源-天天文库。
1、TheCandideSystemforMachineTranslationAdamL.Berger,PeterF.Brown,*StephenA.DellaPietra,VincentJ.DellaPietra,JohnR.GiUett,JohnD.Lafferty,RobertL.Mercer,*HarryPrintz,LuboiUreiIBMThomasJ.WatsonResearchCenterP.O.Box704YorktownHeights,NY10598ABSTRACT~English-to-FrenchIf_[French-to-EnglishWepresenta
2、noverviewofCandide,asystemforautomaticeChannel"-]Decoder6translationofFrenchtexttoEnglishtext.CandideusesmethodsofinformationtheoryandstatisticstodevelopaFigure1:TheSource-ChannelFormalismofTranslation.probabilitymodelofthetranslationprocess.Thismodel,HerefistheFrenchtexttobetransla
3、ted,eistheputativewhichismadetoaccordascloselyaspossiblewithalargeoriginalEnglishrendering,and6istheEnglishtranslation.bodyofFrenchandEnglishsentencepairs,isthenusedtogenerateEnglishtranslationsofpreviouslyunseenFrenchsentences.Thispaperprovidesatutorialinthesemethods,Thisformalismcanbeexplo
4、itedtoyieldFrench-to-Englishdiscussionsofthetrainingandoperationofthesystem,andtranslationsasfollows.LetuswritePr(eIf)fortheprobabil-asummaryoftestresults.itythatewastheoriginalEnglishrenderingoftheFrenchf.GivenaFrenchsentencef,theproblemofautomatictransla-1.Introductiontionreducestofindingt
5、heEnglishsentencethatmaximizesCandideisanexperimentalcomputerprogram,nowinitsP.r(eIf).Thatis,weseek6=argmsxePr(eIf).fifthyearofdevelopmentatIBM,fortranslationofFrenchByvirtueofBayes'Theorem,wehavetexttoEnghshtext.OurgoalistoperformfuRy-automatic,high-qualitytext-to-texttranslation.H
6、owever,becausewe=argmaxPr(eIf)=argmaxPr(fIe)Pr(e)(1)arestillfarfromachievingthisgoal,theprogramcanbeusedeeinbothfully-automaticandtranslator's-assistantmodes.ThetermPr(fle)modelstheprobabilitythatfemergesOurapproachisfoundeduponthestatisticalanalysisoflan-fromthechannelwheneisitsinp
7、ut.Wecallthisfunctionthetranslationmodel;itsdomainisallpairs(f,e)ofFrenchguage.Ourchieftoolsaxethesource-channelmodelofcom-munication,parametricprobabilitymodelsoflanguageandandEnglishword-strings.ThetermPr(e)modelstheaprioritranslation,andanassortmentofnumericalalgorithmsfo