Page163
anintermediary,abstractrepresentationofthemeaningoftheoriginal,and
generatingfromitthetargetlanguagetext.Theparsingprocessinvolves
successiveprogramsforidentifyingwordstructure(morphology)andsentence
structure(syntax)andforresolvingproblemsofambiguity(semantics).
Accordingtothenatureoftheintermediaryrepresentation,twospecificindirect
approachescanbedistinguished:thetransferbasedapproachandthe
interlinguaapproach.
TransferbasedMTconsistsofthreebasicstages:(i)parsinganinputsentence
intoaformalmeaningrepresentationwhichstillretainsthedeepstructure
characteristicsofthesourcetext;(ii)‘transferring’,i.e.converting,theSTformal
representationintoonewhichcarriesthedeepstructurecharacteristicsofthe
targetlanguage,and(iii)generatingatargetsentencefromthetransferred
meaningrepresentation.Mostoftoday’smajorcommercialmainframesystems,
includingMETAL,SYSTRAN,andLogos,adoptthisapproach.Twowidely
knownresearchprojects,Eurotra(fundedbytheCommissionoftheEuropean
Communities)andAriane(atGETAinGrenoble),alsousedthisapproach
(Hutchins1999).
IninterlinguaMT,theabstractrepresentationofthemeaningoftheoriginalis
createdusingan‘interlingua’orpivotlanguage,i.e.an(ideally)source/target
languageindependentrepresentation,fromwhichtargettextsinseveraldifferent
languagescanpotentiallybeproduced.Translationthusconsistsoftwobasic
stages:ananalyser‘transforms’thesourcetextintotheinterlinguaanda
generator‘transforms’theinterlinguarepresentationintothetargetlanguage.
Themostobviousadvantageofthisapproachisthat,fortranslationsinvolving
morethanonelanguagepair,notransfercomponenthastobecreatedforeach
languagepair.Theinterlinguaisusedtoprovideasemanticrepresentationfor
thesourcelanguagewhichhasbeenabstractedfromthesyntaxofthelanguage.
However,findinglanguageindependentwaysofrepresentingsemanticmeaning
isanextremelydifficulttaskwhichgenerallyinvolveseithermakingarbitrary
decisionsastowhatspecificlanguage(natural,artificial,orlogical)
conceptualizationsshouldbetakenasthebasis,ormultiplyingthedistinctions
foundinanyofthelanguagesconcerned,withtheresultthatavastamountof
informationisrequired.Inthelattercase,onewillobtain,forexample,several
primitiveinterlingualitemsrepresenting‘wear’asaconceptbecausethe
Japanesetranslationofthisverbdependsonwheretheobjectisworn,sothata
differentverbwillberequireddependingonwhethertheobjectwornisahator
gloves,forexample(Dorretal.2006).Thetremendousdifficultiesinvolvedin
findinglanguageneutralwaysofrepresentingsemanticmeaningledsome
researcherstoarguethatinterlinguaMTmaynotbeaviableoptionwithinthe
rulebasedMTparadigm;butsuccessfulinterlingualsystemsdoexist,thebest
knownbeingtheFujitsusysteminJapan.
AvariantofinterlingualMTisknowledgebasedMT(KBMT),which
producessemanticallyaccuratetranslationsbuttypicallyneeds,forthepurpose
ofdisambiguation,massiveacquisitionofvariouskindsofknowledge,especially
nonlinguisticinformationrelatedtothedomainsofthetextstobetranslatedand
generalknowledgeabouttherealworld.Thisknowledgeisusuallyencoded
usingpainstakingmanualmethods.ExamplesofKBMTsystemsinclude
Caterpillar(CarnegieMellonUniversity)andULTRA(NewMexicoState
University).
Inthe1990s,researchersbegantoexplorethepossibilityofexploiting
CORPORAofalreadytranslatedtextsforautomatictranslation.Corpusbased
MTcanbeclassifiedintotwocategories:statisticalMTandexamplebased
MT.Instatisticalmachinetranslation(SMT),wordsandphrases(word
sequences)inabilingualparallelcorpusarealignedasthebasisfora‘translation
model’ofword–wordandphrase–phrasefrequencies.Translationinvolvesthe
selection,foreachinputword,ofthemostprobablewordsinthetarget
language,andthedeterminationofthemostprobablesequenceoftheselected
wordsonthebasisofamonolingual‘languagemodel’(Hutchins2006).Since
thetranslationengineworksonthebasisofcorpora,buildingqualitybilingual
textcorporaisessentialtothesuccessofSMT.Wheresuchcorporaare
available,impressiveresultscanbeachievedwhentranslatingtextsofasimilar
kindtothoseinthetrainingcorpus.
ExamplebasedMT(EBMT)systemsalsousebilingualparallelcorporaas
theirmainknowledgebase,atruntime.Inthiscase,translationisproducedby
comparingtheinput