资源描述：
《Neural Machine Translation》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、NeuralMachineTranslationbyJointlyLearningtoAlignandTranslateDzmitryBahdanauKyungHyunChoYoshuaBengioJacobsUniversity,GermanyUniversitedeMontr´eal´UniversitedeMontr´eal´CIFARSeniorFellowAbstractNeuralmachinetranslationisarecentlyproposedapproachtomachin
2、etransla-tion.Unlikethetraditionalstatisticalmachinetranslation,theneuralmachinetranslationaimsatbuildingasingleneuralnetworkthatcanbejointlytunedtomaximizethetranslationperformance.Themodelsproposedrecentlyforneuralmachinetranslationoftenbelongtoafam
3、ilyofencoder–decodersandencodesasourcesentenceintoafixed-lengthvectorfromwhichadecodergeneratesatranslation.Inthispaper,weconjecturethattheuseofafixed-lengthvectorisabottleneckinimprovingtheperformanceofthisbasicencoder–decoderarchitec-ture,andproposeto
4、extendthisbyallowingamodeltoautomatically(soft-)searchforpartsofasourcesentencethatarerelevanttopredictingatargetword,withouthavingtoformthesepartsasahardsegmentexplicitly.Withthisnewapproach,weachieveatranslationperformancecomparabletotheexistingstat
5、e-of-the-artphrase-basedsystemonthetaskofEnglish-to-Frenchtranslation.Furthermore,qualitativeanalysisrevealsthatthe(soft-)alignmentsfoundbythemodelagreewellwithourintuition.1IntroductionNeuralmachinetranslationisanewlyemergingapproachtomachinetranslat
6、ion,recentlyproposedbyKalchbrennerandBlunsom(2013),Sutskeveretal.(2014)andChoetal.(2014b).Unlikethetraditionalphrase-basedtranslationsystem(see,e.g.,Koehnetal.,2003)whichconsistsofmanysmallsub-componentsthataretunedseparately,neuralmachinetranslationa
7、ttemptstobuildandarXiv:1409.0473v2[cs.CL]4Sep2014trainasingle,largeneuralnetworkthatreadsasentenceandoutputsacorrecttranslation.Mostoftheproposedneuralmachinetranslationmodelsbelongtoafamilyofencoder–decoders(Sutskeveretal.,2014;Choetal.,2014a),withan
8、encoderandadecoderforeachlan-guage,orinvolvealanguage-specificencoderappliedtoeachsentencewhoseoutputsarethencom-pared(HermannandBlunsom,2014).Anencoderneuralnetworkreadsandencodesasourcesen-tenceintoafixed-lengthvector.Adecoderthenoutputsatrans
