资源描述：
《Word Representation Models for Morphologically Rich Languages in Neural Machine Translation》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、WordRepresentationModelsforMorphologicallyRichLanguagesinNeuralMachineTranslationEkaterinaVylomova,1TrevorCohn,1andXuanliHe1andGholamrezaHaffari21DepartmentofComputingandInformationSystems,UniversityofMelbourne2FacultyofInformationTechnology,MonashUniversityevylo
2、mova@gmail.comtcohn@unimelb.edu.auxuanlih@student.unimelb.edu.augholamreza.haffari@monash.eduAbstractwithaheavytaildistribution.ForexampleinRus-sian,thereareatleast70wordsfordog,encodingDealingwiththecomplexwordformsinmor-case,gender,age,number,sentimentandothers
3、e-phologicallyrichlanguagesisanopenprob-leminlanguageprocessing,andisparticularlymanticconnotations.Manyofthesewordsshareaimportantintranslation.Incontrasttomostcommonlemma,andcontainregularmorphologicalmodernneuralsystemsoftranslation,whichaffixation;consequently
4、muchoftheinformationre-discardtheidentityforrarewords,inthispa-quiredfortranslationispresent,butnotinanacces-perweproposeseveralarchitecturesforlearn-sibleformformodelsofneuralMT.ingwordrepresentationsfromcharacterandInthispaper,weproposeasolutiontothisprob-morph
5、emelevelworddecompositions.Wein-lembyconstructingwordrepresentationscompo-corporatetheserepresentationsinanovelma-chinetranslationmodelwhichjointlylearnssitionallyfromsmallersub-wordunits,whichoc-wordalignmentsandtranslationsviaahardcurmorefrequentlythanthewordst
6、hemselves.Weattentionmechanism.Evaluatingontrans-showthattheserepresentationsareeffectiveinhan-latingfromseveralmorphologicallyrichlan-dlingrarewords,andincreasethegeneralisationca-guagesintoEnglish,weshowconsistentim-pabilitiesofneuralMTbeyondthevocabularyob-pro
7、vementsoverstrongbaselinemethods,ofservedinthetrainingset.Weproposeseveralneu-between1and1.5BLEUpoints.ralarchitecturesforcompositionalwordrepresenta-tions,andsystematicallycomparethesemethodsin-1IntroductiontegratedintoanovelneuralMTmodel.Modelsofend-to-endmachi
8、netranslationbasedonMorespecifically,wemakeuseofcharacterse-neuralnetworkshavebeenshowntoproduceexcel-quencesormorphemesequencesinbuildingwordlenttranslations,r
