欢迎来到天天文库
浏览记录
ID:39713775
大小:3.54 MB
页数:7页
时间:2019-07-09
《10_14_recursive_network_training》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、NeuralnetworksNaturallanguageprocessing-recursivenetworktrainingtoformawhole.Weintroduceamax-margin2structurepredictionarchitecturebasedonre-cursiveneuralnetworksthatcansuccessfullyRECURSIVENEURALNETWORKrecoversuchstructurebothincomplexsceneimagesaswellassentences.Thesamealgo-
2、Topics:recursiveneuralnetwork(RNN)rithmcanbeusedbothtoprovideacompeti-•Idea:recursivelymergepairsofword/phraserepresentationstivesyntacticparserfornaturallanguagesen-tencesfromthePennTreebankandtoout-performalternativeapproachesforsemanticscenesegmentation,annotationandclassifi
3、-cation.Forsegmentationandannotationouralgorithmobtainsanewlevelofstate-of-the-wordrepresentations{artperformanceontheStanfordbackgrounddataset(78.1%).Thefeaturesfromtheim-ageparsetreeoutperformGistdescriptorsfor•Weneed2thingsFigure1.Illustrationofourrecursiveneuralnetworkar-S
4、ocher,Lin,NgandManning,2011sceneclassificationby4%.chitecturewhichparsesimagesandnaturallanguagesen-‣amodelthatmergespairsofrepresentationstences.Segmentfeaturesandwordindices(orange)are‣amodelthatdeterminesthetreestructurefirstmappedintosemanticfeaturespace(blue)andthen1.Introd
5、uctionrecursivelymergedbythesameneuralnetworkuntiltheyrepresenttheentireimageorsentence.BothmappingsandRecursivestructureiscommonlyfoundindifferentmergingsarelearned.modalities,asshowninFig.1.Thesyntacticrulescanberecursivelysplitintosmallercarregionsdepict-ofnaturallanguageare
6、knowntoberecursive,withingpartssuchastiresandwindowsandthesepartscannounphrasescontainingrelativeclausesthatthem-occurinothercontextssuchasbeneathairplanesorinselvescontainnounphrases,e.g.,...thechurchwhichhouses.Weshowthatrecoveringthisstructurehelpshasnicewindows....Similarl
7、y,onefindsnestedhier-inunderstandingandclassifyingsceneimages.Inthisarchicalstructuringinsceneimagesthatcapturebothpaper,weintroducerecursiveneuralnetworks(RNNs)part-ofandproximityrelationships.Forinstance,carsforpredictingrecursivestructureinmultiplemodali-areoftenontopofstree
8、tregions.Alargecarregionties.Weprimarilyfocusonsceneunderstan
此文档下载收益归作者所有
点击更多查看相关文章~~