machine translation introduction

《machine translation introduction》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、SpellCorrection&MachineTranslationContentSpellCorrectionMachineTranslationIntroductionStatisticalMachineTranslation:IBMModelsPhrase-BasedTranslationModelsSpellCorrectionIntroductionGivenaword,wearetryingtochoosethemostlikelyspellingcorrectionfor

2、thatword(the"correction"maybetheoriginalworditself).Thereisnowaytoknowforsure(forexample,should"lates"becorrectedto"late"or"latest"?),whichsuggestsweuseprobabilities.Wewillsaythatwearetryingtofindthecorrectionc,outofallpossiblecorrections,thatmaximiz

3、estheprobabilityofcgiventheoriginalwordw:argmaxcP(c

4、w)ByBayes'Theoremthisisequivalentto:argmaxcP(w

5、c)P(c)/P(w)SinceP(w)isthesameforeverypossiblec,wecanignoreit,giving:argmaxcP(w

6、c)P(c)SpellCorrectionIntroductionTherearethreepartsofthisexpressio

7、n.P(c),theprobabilitythataproposedcorrectioncstandsonitsown.ThisiscalledthelanguagemodelSoP("the")wouldhavearelativelyhighprobability,whileP("zxzxzxzyyy")wouldbenearzero.P(w

8、c),theprobabilitythatwwouldbetypedinatextwhentheauthormeantc.Thisistheerr

9、ormodelargmaxc,thecontrolmechanism,whichsaystoenumerateallfeasiblevaluesofc,andthenchoosetheonethatgivesthebestcombinedprobabilityscore.HowtoworkP(c)1Wewillreadabigtextfile,big.txt,whichconsistsofaboutamillionwords2extracttheindividualwordsfromth

10、efile3trainaprobabilitymodel,whichisafancywayofsayingwecounthowmanytimeseachwordoccursenumeratingthepossiblecorrectionscofagivenwordweditdistance:thenumberofeditsitwouldtaketoturnoneintotheotherTheliteratureonspellingcorrectionclaimsthat80to95%of

11、spellingerrorsareaneditdistanceof1Forawordoflengthn,therewillbendeletions,n-1transpositions,26nalterations,and26(n+1)insertions,foratotalof54n+25(ofwhichafewaretypicallyduplicates).HowtoworkP(w

12、c)mistakingonevowelforanotherismoreprobablethanmistak

13、ingtwoconsonants;makinganerroronthefirstletterofawordislessprobable,etc.definedatrivialmodelthatsaysallknownwordsofeditdistance1areinfinitelymoreprobablethanknownwordsofeditdistance2,andinfinitelylessprobablethanaknownwordofeditdistance0.Thefunctio