资源描述:
《Aligning context-based statistical models of language with brain activity during reading 》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、Aligningcontext-basedstatisticalmodelsoflanguagewithbrainactivityduringreadingLeilaWehbe1,2,AshishVaswani3,KevinKnight3andTomMitchell1,21MachineLearningDepartment,CarnegieMellonUniversity,Pittsburgh,PA2CenterfortheNeuralBasisofComputation,CarnegieMellonUniversity,Pit
2、tsburgh,PA3InformationSciencesInstitute,UniversityofSouthernCalifornia,LosAngeles,CAlwehbe@cs.cmu.edu,vaswani@usc.edu,knight@isi.edu,tom.mitchell@cs.cmu.eduAbstractimpressivegoals.Modelslikedeepneuralnet-worksandvectorspacemodelshavebecomepop-Manystatisticalmodelsfor
3、naturallanguagepro-ulartosolvediversetaskslikesentimentanaly-cessingexist,includingcontext-basedneuralnet-worksthat(1)modelthepreviouslyseencontextsisandmachinetranslation.Becauseofthecom-asalatentfeaturevector,(2)integratesuccessiveplexityofthesemodels,itisnotalways
4、clearhowwordsintothecontextusingsomelearnedrepresen-toassessandcomparetheirperformancesastheytation(embedding),and(3)computeoutputproba-bilitiesforincomingwordsgiventhecontext.Onmightbeusefulforonetaskandnottheother.theotherhand,brainimagingstudieshavesug-Itisalsonot
5、easytointerprettheirveryhigh-gestedthatduringreading,thebrain(a)continu-ouslybuildsacontextfromthesuccessivewordsdimensionalandmostlyunsupervisedrepresenta-andeverytimeitencountersawordit(b)fetchesitstions.Thebrainisanothercomputationalsystempropertiesfrommemoryand(c
6、)integratesitwiththatprocesseslanguage.Sincewecanrecordbrainthepreviouscontextwithadegreeofeffortthatisinverselyproportionaltohowprobablethewordis.activityusingneuroimaging,weproposeanewdi-Thishintstoaparallelismbetweentheneuralnet-rectionthatpromisestoimproveourunde
7、rstand-worksandthebraininmodelingcontext(1anda),ingofbothhowthebrainisprocessinglanguagerepresentingtheincomingwords(2andb)andin-tegratingit(3andc).Weexplorethisparallelismtoandofwhattheneuralnetworksaremodelingbybetterunderstandthebrainprocessesandtheneu-aligningthe
8、braindatawiththeneuralnetworksralnetworksrepresentations.Westudythealign-representations.mentbetweenthelatentvectorsusedbyneuralnet