Convolutional Neural Networks for Sentence Classification

《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,CNNmodelshaveWereportonaseriesofexperimentswithsubsequentlybeenshowntobeeffectiveforNLPconvolutionalneuralnetworks(CNN)andhave

2、achievedexcellentresultsinsemantictrainedontopofpre-trainedwordvec-parsing(Yihetal.,2014),searchqueryretrievaltorsforsentence-levelclassificationtasks.(Shenetal.,2014),sentencemodeling(Kalch-WeshowthatasimpleCNNwithlit-brenneretal.,2014),andothertraditionalNLPtlehyperparametertuningandstati

3、cvec-tasks(Collobertetal.,2011).torsachievesexcellentresultsonmulti-Inthepresentwork,wetrainasimpleCNNwithplebenchmarks.Learningtask-specificonelayerofconvolutionontopofwordvectorsvectorsthroughfine-tuningoffersfurtherobtainedfromanunsupervisedneurallanguagegainsinperformance.Weadditionallym

4、odel.ThesevectorsweretrainedbyMikolovetproposeasimplemodificationtothear-al.(2013)on100billionwordsofGoogleNews,chitecturetoallowfortheuseofbothandarepubliclyavailable.1Weinitiallykeepthetask-specificandstaticvectors.TheCNNwordvectorsstaticandlearnonlytheotherparam-modelsdiscussedhereinimpro

5、veupontheetersofthemodel.Despitelittletuningofhyper-stateofthearton4outof7tasks,whichparameters,thissimplemodelachievesexcellentincludesentimentanalysisandquestionresultsonmultiplebenchmarks,suggestingthatclassification.thepre-trainedvectorsare‘universal’featureex-1Introductiontractorsthatc

6、anbeutilizedforvariousclassifica-tiontasks.Learningtask-specificvectorsthroughDeeplearningmodelshaveachievedremarkablefine-tuningresultsinfurtherimprovements.Weresultsincomputervision(Krizhevskyetal.,finallydescribeasimplemodificationtothearchi-2012)andspeechrecognition(Gravesetal.,2013)tecture

7、toallowfortheuseofbothpre-trainedandinrecentyears.Withinnaturallanguageprocess-task-specificvectorsbyhavingmultiplechannels.ing,muchoftheworkwithdeeplearningmeth-odshasinvolvedlearningwordvectorrepresenta-OurworkisphilosophicallysimilartoRazaviantionsthroughneu