probmach

MachTranslat(2017)31:1–2DOI10.
1007/s10590-017-9198-yEditorialQunLiu1·XiaodongHe2·HermannNey3Publishedonline:20June2017SpringerScience+BusinessMediaB.
V.
2017Inthepastfewyears,wehavewitnessedanexplosioninthenumberofapplicationsofneuralnetworksinmachinetranslation(MT).
OneofthefundamentaldifferencesbetweentheneuralnetworkapproachesandtraditionalapproachestoMTisthewayinwhichtheydenethetranslationproblem:inneuralnetworkapproaches,thetranslationproblemisdenedinacontinuousspace,whereasintraditionalapproachesitisdenedinasymbolicspace.
Thisdifferencehasbroughtaboutgreatchangesinthisarea.
Themostobviouschangeisthenewstate-of-the-artcreatedbyneuralMT(NMT).
Butbeyondthis,manyotherchangesmayalsohaveprofoundandlong-termimpactsinthisareaandshouldnotbeovershadowed.
Inthevepapersinthisspecialissue,weareveryhappytoseethevariousaspectsofresearchbyusingneuralnetworkapproachestoMT.
Wordembeddingistherststeptoconvertthedenitionofthetranslationprob-lemfromasymbolicspacetoacontinuousspace.
Thepaper"Therepresentationalgeometryofwordmeaningsacquiredbyneuralmachinetranslationmodels"analysesthedifferencesbetweenthewordembeddingsderivedfromamonolingualcorpusandthosefromaparallelcorpusbyNMTtraining.
TheauthorsfoundthatwordembeddingsBQunLiuqun.
liu@dcu.
ieXiaodongHexiaohe@microsoft.
comHermannNeyney@informatik.
rwth-aachen.
de1ADAPTCentre,SchoolofComputing,DublinCityUniversity,Dublin,Ireland2DeepLearningTechnologyCenter,MicrosoftResearch,Redmond,WA,USA3LehrstuhlfürInformatik6,RWTHAachen,Ahornstr.
55,52056Aachen,Germany1232Q.
Liuetal.
learnedonaparallelcorpusbytheNMTsystemcapturemoresemanticinformationthanwordembeddingsderivedfromamonolingualcorpus.
Expressionsoflanguagesinacontinuousspacebreakdownthebarriersbetweenlanguagesandothermodalitieslikespeech,imageandvideo,andmakeitpossibletomakeuseofinformationfromothermodalitiestoimproveMT.
Inthisspecialissue,weseetwopapersextendingMTresearchinthisway.
Thepaper"Zero-resourceMachineTranslationbyMultimodalEncoder-decoderNetworkwithMultimediaPivot"demonstratesthepossibilitytouseimagesratherthanathirdlanguageasapivotforMTbetweenlanguageswithoutaparallelcorpus.
IntermsofthetypeofimageinformationusedinMT,thepaper"Chinese-SpanishNeuralMachineTranslationEnhancedwithCharacterandWordBitmapFonts"goesevenfurtherbyusingimagesofbitmapfontsofChinesecharactersratherthanpicturesofrealworldobjects.
IncontrasttobuildingMTsystemsunderapureneuralframework,neuralnetworkapproachescanalsobeintegratedintoatraditionalstatisticalMT(SMT)framework.
Thepaper"AComparisonofDiscriminativeTrainingCriteriaforContinuousSpaceTranslationModels"treatstranslationmodels,whichareoneofthekeycomponentsinSMT,inacontinuousspaceandinvestigateshowtoimprovethesetranslationmodelsbyexploringanewdiscriminativetrainingprocedure.
Inthepaper"Anovelandrobustapproachforpro-droplanguagetranslation",theauthorsproposeaneuralnetworkapproachtorestoredroppedpronounsinthesourcelanguageandsignicantlyimprovetranslationqualityintheresultantSMTsystem.
Wehopeyouenjoyalltheseexcellentpapersinthisspecialissue.
123