Should I use machine translation.docx

Should I use machine translation.docx

《Should I use machine translation.docx》由会员分享，可在线阅读，更多相关《Should I use machine translation.docx（14页珍藏版）》请在冰点文库上搜索。

ShouldIusemachinetranslation

ShouldIusemachinetranslation?

AlanK.Melby

Machinetranslationissomewhatlikethechildwho,whengood,wasverynicetohavearoundbutwho,whenbad,wasjustawful.Forsomeguidanceindecidingwhethermachinetranslationislikelytosucceedinyourparticularenvironment,pleasereadon.

Contents

∙Whatismachinetranslation?

∙Howismachinetranslationtypicallyused?

∙TheTranslationTripod-Figure1

oSourceText

oSpecifications

oTerminology

∙Quality

∙TheBlackBoxMyth

∙TerminologyManagement

∙AnotherMajorFactor

∙MakingtheDecision

∙Whataboutthefuture?

∙Ambiguity

oTheMeaningof"Cut"-Figure2

oCutOffaPiece-Figure3

oTheMeaningof"Bank"-Figure4

oWhichpengoesinthebox?

-Figure5

oWhichpen?

-Figure6

∙AttacksonAmbiguity

oConsidertheWord"Bus"-Figure7

∙Objectivism

oDifferentCutsofMeet-Figure8

oLightShiningIn-Figure9

∙Summary

∙FortheUnconvincedortheCurious

Whatismachinetranslation?

Peoplewhoneeddocumentstranslatedoftenaskthemselveswhethertheycoulduseacomputertodothejob.

Whenacomputertranslatesanentiredocumentautomaticallyandthenpresentsittoahuman,theprocessiscalledmachinetranslation.Whenahumancomposesatranslation,perhapscallingonacomputerforassistanceinspecifictaskssuchaslookingupspecializedwordsandexpressionsinadictionary,theprocessiscalledhumantranslation.

Thereisagrayareabetweenhumanandmachinetranslation,inwhichthecomputermayretrievewholesentencesofpreviouslytranslatedtextandmakeminoradjustmentsasneeded.However,eveninthisgrayarea,eachsentencewasoriginallytheresultofeitherhumantranslationormachinetranslation.Wewillreservethelabel"machinetranslation"forthecasewhenboththeinitialtranslationofthesentencesandsubsequentmanipulationsareperformedbyacomputer.Allelsewewillcall"translatortools".

Howismachinetranslationtypicallyused?

Machinetranslationishighlyappealingwhenitsqualityisacceptableforsomepurposewithlittleornohumanrevision.Butslowdown!

Ninechancesoutoften,machinetranslationwillnotworkforyou!

Ibasethatestimateonthefactthat,currently,machinetranslationisusedforlessthantenpercentofthepublicationqualitytranslationproducedworldwideeachyear.Thisfigureishardtomeasure,butfewwouldclaimthatitisovertenpercent.Whyisthisso?

Wewillbegintoanswerthisquestionbysuggestionthatformachinetranslationtobeappropriate,itmustbesittingonastable"tripod".

TheTranslationTripod

Atranslationprojectcanbethoughtofassittingonatripodwhosethreelegsarethesourcetext,thespecifications,andtheterminology.Ifanyofthethreelegsisremoved,theprojectfallsdown.[Figure1:

16kGIF]

1.Sourcetext

Obviously,notranslationcanbedonewithoutasourcetext（i.e.,thedocumenttobetranslated）.Butformachinetranslation,anadditionalbasicrequirementisthatthesourcetextbeavailableinmachine-readableform.Thatis,itmustcomeondisketteorcartridgeortapeorbymodemandendupasatextfileonyourdisk.Afaxofthesourcetextisnotconsideredtobeinmachine-readableform,evenifitisinacomputerfile.Afaxinacomputerfileisonlyagraphicalimageofthetext,andthecomputerdoesnotknowwhichdotscomposetheletteraortheletterb.Conversionofasourcetextonpaperorinagraphicalimagefiletomachine-readableformusingimagedcharacterrecognition（ICR）isnotusuallyaccurateenoughtobeusedwithouthumanediting,andhumaneditingisexpensive,addinganunacceptablecostcomponenttothetotalcostofmachinetranslation.Thus,formachinetranslationtobeappropriate,itisusuallynecessarytoobtainthewordprocessingordesktoppublishingfilefromtheorganizationthatcreatedthesourcetext.Butthisisonlyoneofmanyrequirements.

2.Specifications

Alltranslationsprojectshavespecifications.Theproblemisthattheyareseldomwrittendown.Specificationstellhowthesourcetextistobetranslated.Onespecificationthatisalwaysgiveniswhatlanguagetotranslateinto.Butthatisinsufficient.Shouldtheformatofthetargettext（i.e.,thetranslation）bethesameasthatofthesourcetextordifferent?

Whoistheintendedaudienceforthetargettext?

Doestheleveloflanguageneedtobeadjusted?

Intechnicaltranslation,perhapsthemostimportantspecificationiswhatequivalentstousefortechnicalterms.Arethereothertargettextswithwhichthistranslationshouldbeconsistent?

Whatisthepurposeofthetranslation?

Ifthepurposeisjusttogetageneralideaofthecontentofthesourcetext,thenthespecificationswouldinclude"indicativetranslationonly."Anindicativetranslationisusuallyforthebenefitofonepersonratherthanforpublicationandneednotbeahigh-qualitytranslation.Thus,publication-qualitytranslationsarehigh-qualitytranslations（andareusuallytheresultofhumantranslation）,whileindicativetranslationsarelow-qualitytranslations（andareusuallytheresultofmachinetranslation）.Thesetwotypesoftranslationarenotnormallyincompetitionwitheachother,sincearequesteroftranslationwilltypicallywantonetypeortheotherforagivendocumentandagivensetofspecifications.Sometimes,thetwotypesarecomplementary,suchaswhenanindicativetranslationisusedtodecidewhetherornottorequestahigh-qualitytranslationofaparticulardocument.Inthisenvironment,anindicativetranslationmayberequestedforanumberofdocuments,and,usingtheindicativetranslations,therequestermayselectoneortwodocumentsforpublicationqualitytranslation.

Aspreviouslymentioned,indicativetranslationsareusuallydoneusingmachinetranslationandhigh-qualitytranslationsareusuallydoneusinghumantranslation.Thisfactrevealsabasicdifferencebetweenhumansandcomputers.Humans,withproperstudyandpractice,aregoodatproducinghigh-qualitytranslationsbuttypicallycanonlytranslateafewhundredwordsanhourtoapproximatelyathousandwordsanhour,dependingonsuchfactorsasthedifficultyofthesourcetext.Evenwithveryfamiliarmaterial,humantranslatorsarelimitedbyhowfasttheycantypeordictatetheirtranslations.Computersaregoodatproducinglow-qualitytranslationsveryquickly.Somemachinetranslationsystemscantranslatetensofthousandsofwordsanhour.Butastheyare"trained"byaddingtotheirdictionariesandgrammars,theyreachaplateauwherethequalityoftheoutputdoesnotimprove.Byupgradingtoamorepowerfulcomputer,thespeedoftranslationimprovesbutnotthequality.Byupgradingtoa"morepowerful"humantranslator,thequalityoftranslationimprovesbutnotnecessarilythespeed.Herewehaveaclassiccaseofatrade-off.Youcanhavehighspeedorhighqualitybutnotboth.

Indicativetranslation（highspeed,lowcost,butlowquality）representsanewandgrowingmarketbutdoesnotsubstantiallyoverlapwiththeexistingmarketforpublicationqualitytranslation.Theexistingmarket,variouslyestimatedat10,000,000,000to20,000,000,000USdollarsworld-wideperyear,isprimarilyforhigh-qualitytechnicaltranslation.If,ontheonehand,yourspecificationsincludelowquality（barelyunderstandable）translation,thenmachinetranslationisforyou,andyoucanstopreadingrighthere.If,onthemorelikelyhand,yourspecificationsincludehigh-qualitytranslation,thenitisnotobviousthatmachinetranslationisappropriateforyourcurrenttranslationjob.Herequalitywouldbemeasuredbywhetherthetargettextisgrammatical,accurate,understandable,readable,andusable.Usabilitycanbemeasuredbyselectingtasks,suchasmaintenanceoperations,whichcanbeaccomplishedbyasource-languagereaderwiththehelpofthesourcetextandseeingwhetherthosesametaskscanbeperformedbyatarget-languagereaderwiththehelpofthetargettext.Suchmeasurementsarenotoriouslyexpensive,butaskilledreviewercanaccuratelypredictusabilitysimplybystudyingthesourceandtargettexts.Grammaticality,andunderstandability,andreadability,whichareprogressivelymorestringentrequirements,canbemeasuredbyatarget-languagemonolingualperson.Butaccuracyrequirestheassistanceofaskilledbilingualpersonwhoexaminesboththesourceandtargettexts.

3.Terminology

Thetreatmentofterminologycouldhavebeenincludedsoleyunderspecifications.Butterminologyissoimportantthattheactualterminologicaldatabase（alsocalleda"termbase"）suppliedwithasourcetexthasbeenlistedasathirdessentialcomponentofatranslationjob.Theaspectofterminologythatdoesfitunderspecificationsistherequirementthatthetranslationjobuseacertaintermbaseinordertoachievedesiredconsistency.LetmeexplainwhatImeanbyconsistency.Translationrequesterstypicallywanttheterminologyintheirtranslateddocumentstomeshcloselywithterminologyinrelateddocuments.Forexample,asoftwarecompanywillwantallrevisionsofasoftwaremanualtousethesametermsastheoriginal,toavoidconfusingreaders.Translationrequestersshouldtrackallterminologyrelevanttoagivendocumentanddeliverthatterminologytothetranslationprovideralongwithspecificationsandsourcetext.Thespecificationcomponentofthejobtellswhatappropriatetermbasetouseand,asisalltoocommon,tellswhattodoifasource-texttermismissingfromthetermbase.Theterminologycomponentofthejobcontainsthetermbaseitself.

Nowwecandefineanappropriatetranslationjob（forahumanor