资源描述:
《Convolutional Neural Networks for Sentence Classification 》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、ConvolutionalNeuralNetworksforSentenceClassificationYoonKimNewYorkUniversityyhk255@nyu.eduAbstractlocalfeatures(LeCunetal.,1998).Originallyinventedforcomputervision,CNNmodelshaveWereportonaseriesofexperimentswithsubsequentlybeenshowntobeeffectiveforNLPconvolutionalneuralnetwo
2、rks(CNN)andhaveachievedexcellentresultsinsemantictrainedontopofpre-trainedwordvec-parsing(Yihetal.,2014),searchqueryretrievaltorsforsentence-levelclassificationtasks.(Shenetal.,2014),sentencemodeling(Kalch-WeshowthatasimpleCNNwithlit-brenneretal.,2014),andothertraditionalNLPt
3、lehyperparametertuningandstaticvec-tasks(Collobertetal.,2011).torsachievesexcellentresultsonmulti-Inthepresentwork,wetrainasimpleCNNwithplebenchmarks.Learningtask-specificonelayerofconvolutionontopofwordvectorsvectorsthroughfine-tuningoffersfurtherobtainedfromanunsupervisedneu
4、rallanguagegainsinperformance.Weadditionallymodel.ThesevectorsweretrainedbyMikolovetproposeasimplemodificationtothear-al.(2013)on100billionwordsofGoogleNews,chitecturetoallowfortheuseofbothandarepubliclyavailable.1Weinitiallykeepthetask-specificandstaticvectors.TheCNNwordvecto
5、rsstaticandlearnonlytheotherparam-modelsdiscussedhereinimproveupontheetersofthemodel.Despitelittletuningofhyper-stateofthearton4outof7tasks,whichparameters,thissimplemodelachievesexcellentincludesentimentanalysisandquestionresultsonmultiplebenchmarks,suggestingthatclassificat
6、ion.thepre-trainedvectorsare‘universal’featureex-1Introductiontractorsthatcanbeutilizedforvariousclassifica-tiontasks.Learningtask-specificvectorsthroughDeeplearningmodelshaveachievedremarkablefine-tuningresultsinfurtherimprovements.Weresultsincomputervision(Krizhevskyetal.,fina
7、llydescribeasimplemodificationtothearchi-2012)andspeechrecognition(Gravesetal.,2013)tecturetoallowfortheuseofbothpre-trainedandinrecentyears.Withinnaturallanguageprocess-task-specificvectorsbyhavingmultiplechannels.ing,muchoftheworkwithdeeplearningmeth-odshasinvolvedlearningwo
8、rdvectorrepresenta-OurworkisphilosophicallysimilartoRazaviantionsthroughneu