![cs224d深度學習到自然語言處理_第1頁](http://file4.renrendoc.com/view/9625e5229b44affe291b7d1f9d42d48c/9625e5229b44affe291b7d1f9d42d48c1.gif)
![cs224d深度學習到自然語言處理_第2頁](http://file4.renrendoc.com/view/9625e5229b44affe291b7d1f9d42d48c/9625e5229b44affe291b7d1f9d42d48c2.gif)
![cs224d深度學習到自然語言處理_第3頁](http://file4.renrendoc.com/view/9625e5229b44affe291b7d1f9d42d48c/9625e5229b44affe291b7d1f9d42d48c3.gif)
![cs224d深度學習到自然語言處理_第4頁](http://file4.renrendoc.com/view/9625e5229b44affe291b7d1f9d42d48c/9625e5229b44affe291b7d1f9d42d48c4.gif)
![cs224d深度學習到自然語言處理_第5頁](http://file4.renrendoc.com/view/9625e5229b44affe291b7d1f9d42d48c/9625e5229b44affe291b7d1f9d42d48c5.gif)
版權(quán)說明:本文檔由用戶提供并上傳,收益歸屬內(nèi)容提供方,若內(nèi)容存在侵權(quán),請進行舉報或認領(lǐng)
文檔簡介
CS224d:DeepNLP
Lecture15:
Applica8onsofDLtoNLP
RichardSocher
richard@metamind.io
Overview
?
Modeloverview:Howtoproperlycomparetoothermodelsand
chooseyourownmodel
?
?
?
?
Wordrepresenta8on
Phrasecomposi8onObjec8vefunc8on
Op8miza8on
?
?
?
?
?
CharacterRNNsontextandcode
MorphologyLogic
Ques8onAnswering
Image–Sentencemapping
Lecture1,Slide2
RichardSocher
5/27/15
Modeloverview:WordVectors
?
Random
?
Word2Vec
?
Glove
?
Dimension–oRendefinesthenumberofmodelparameters
?
Orworkdirectlyoncharactersormorphemes
Lecture1,Slide3
RichardSocher
5/27/15
Modeloverview:PhraseVectorComposi8on
?
Composi8onFunc8ongovernshowexactlywordandphrase
vectorsinteracttocomposemeaning
WVscore
W
s
p
?
Averaging:p=a+b
Lotsofsimplealterna8ves
c1
c2
?
Recursiveneuralnetworks
?
Convolu8onalneuralnetworks
?
Recurrentneuralnetwork
Lecture1,Slide4
RichardSocher
5/27/15
Composi8on:BigramandRecursivefunc8ons
?
ManyrelatedmodelsarespecialcasesofMV---RNN
?
MitchellandLapata,2010;Zanzo^oetal.,2010:
?
BaroniandZamparelli(2010):Aisanadjec8vematrixandbis
anounvector
?
RNNsofSocheretal.2011(ICML,EMNLP,NIPS)arealso
specialcases
?
Recursiveneuraltensornetworksbringquadra8cand
mul8plica8veinterac8onsbetweenvectors
Addi8onalchoiceforrecursiveneuralnets
?
Dependencytreesfocusmoreonseman8cstructure
1.
Cons8tuencyTree
2.
DependencyTree
3.
BalancedTree
Composi8on:CNNs
?
Severalvariantsalso:
?
?
?
Nopoolinglayers
Poolinglayers:simplemax---poolingordynamicpoolingPoolingacrossdi?erentdimensions
?
SomewhatlessexploredinNLPthanRNNs2
?
Notlinguis8callynorcogni8velyplausible
Lecture1,Slide7
RichardSocher
5/27/15
Composi8on:RecurrentNeuralNets
?
Vanilla
?
GRU
?
LSTM
?
ManyvariantsofLSTMs
“LSTM:ASearchSpaceOdyssey”byGre?etal.2015
Lecture1,Slide8
RichardSocher
5/27/15
Modeloverview:Objec8vefunc8on
?
Max---margin
?
Cross---entropy
Supervisedtopredictaclass
Unsupervised:predictsurroundingwords
?
Auto---encoder
Myopinion:UnclearbenefitsforNLP
Unlessencodinganothermodality
Lecture1,Slide9
RichardSocher
5/27/15
Op8miza8on
?
Ini8aliza8on(wordvectorandcomposi8onparameters)!!
?
Op8miza8onalgorithm
?
?
?
?
?
SGD
SGD+momentumL---BFGS
AdaGrad
Adelta
?
Op8miza8ontricks
Regulariza8on(somedefineaspartofmodel)
Dropout
Lecture1,Slide10
RichardSocher
5/27/15
CharacterRNNsontextandcode
h^p://karpathy.github.io/2015/05/21/rnn---e?ec8veness/
Lecture1,Slide11
RichardSocher
5/27/15
CharacterRNNsontextandcode
?
Haven’tyetproducedusefulresultsonrealdatasets
?
ShowsthatRNNscanmemorizesequencesandkeepmemory
(mostlyLSTMs)
?
Mostinteres8ngresultssimplytrainondatasetandsample
fromitaRerwards(firstshownbySutskeveretal.2011:Genera8ngTextwithRecurrentNeuralNetworks)
?
ResultsfromanLSTM(karpathy.github.io)
Lecture1,Slide12
RichardSocher
5/27/15
Shakespeare
PANDARUS:
Alas,IthinkheshallbecomeapproachedandthedayWhenli^lesrainwouldbea^ain'dintobeingneverfed,Andwhoisbutachainandsubjectsofhisdeath,
Ishouldnotsleep.
SecondSenator:
Theyareawaythismiseries,produceduponmysoul,Breakingandstronglyshouldbeburied,whenIperish
Theearthandthoughtsofmanystates.
DUKEVINCENTIO:
Well,yourwitisinthecareofsideandthat.
Lecture1,Slide13
RichardSocher
5/27/15
Wikipedia
NaturalismanddecisionforthemajorityofArabcountries'
capitalidewasgroundedbytheIrishlanguageby[[JohnClair]],[[AnImperialJapaneseRevolt]],associatedwithGuangzham'ssovereignty.HisgeneralswerethepowerfulrulerofthePortugalinthe[[ProtestantImmineners]],whichcouldbesaidtobedirectlyinCantoneseCommunica8on,whichfollowedaceremonyandsetinspiredprison,training.Theemperortravelledbackto[[An8och,Perth,October25|21]]tonote,theKingdomofCostaRica,unsuccessfulfashionedthe[[Thrales]],[[Cynth'sDajoard]],knowninwestern[[Scotland]],nearItalytotheconquestofIndiawiththeconflict.
Lecture1,Slide14
RichardSocher
5/27/15
Latex(hadtobefixedmanually)
Lecture1,Slide15
RichardSocher
5/27/15
Code!(Linuxsourcecode)
Lecture1,Slide16
RichardSocher
5/27/15
Morphology
?
Be^erWordRepresenta8onswithRecursiveNeuralNetworks
forMorphology–Luongetal.(slidesfromLuong)
?
Problemwithwordvectors:
poorlyes8materareandcomplexwords.
dis8nct
dis8nctnessa?ect
una?ected
di?erentdis8nc8vebroadernarrower
morphologiespesawatcleRspathologiesexacerbateimpactscharacterize
unno8ceddwarfedmi8gated
uniquebroaddis8nc8veseparate
companionroskamhitoshienjoyedallowpreventinvolveenable
mon8sheathskrystal
Lecture1,Slide17
RichardSocher
5/27/15
Limita8onsofexis8ngwork
?
?
Treatrelatedwordsasindependenten88es.
Representunknownwordswithafewvectors.
35323
23828
22687
352634791285
569
406
175 141
108
WordfrequenciesinWikipediadocs(986mtokens)
Luong’sapproach–networkstructure
vT
f(W[x;x;...;x]+b)
1 2
n
v
W,b
unfortunately
Wm,bm
the
bank
was
closed
Wm,bm
unfortunate
ly
close
d
Wm,bm
p=f(Wm[xstem;xaffix]+bm)
un
fortunate
?
NeuralLanguageModel:simplefeed---forwardnetwork(Huang,etal.,2012)
withranking---typecost(Collobertetal.,2011).
?
MorphologyModel:recursiveneuralnetwork(Socheretal.,2011).
MorphologyModel
NeuralLanguageModel
Analysis
?
Blendswordstructureandsyntac8c---seman8cinforma8on.
commen8ng
una?ecteddis8nctdis8nctnessheartlessness
saudi---owned
insis8nginsistedfocusing
unno8ceddwarfedmi8gateddi?erentdis8nc8vebroadermorphologiespesawatcleRs
?
avatarmohajirkripalani
commentedcommentscri8cizing
undesiredunhinderedunrestricteddivergentdiversedis8nc8vedis8nc8venesssmallnesslargenesscorrup8veinhumanityine?ectual
saudi---basedsyrian---controlled
Solu8onstotheproblemofpolysemouswords
?
ImprovingWordRepresenta8onsViaGlobalContextAnd
Mul8pleWordPrototypesbyHuangetal.2012
Lecture1,Slide21
RichardSocher
5/27/15
Naturallanguageinference
Claim:Simpletasktodefine,butengagesthefullcomplexityof
composi8onalseman8cs:
?
?
?
?
?
?
?
?
Lexicalentailment
Quan8fica8onCoreferenceLexical/scopeambiguity
Commonsenseknowledge
Proposi8onalawtudesModality
Fac8vityandimplica8vity
Lecture1,Slide22
RichardSocher
5/27/15
Firsttrainingdata
?
?
?
?
?
?
?
Trainingdata:
dance
waltztangosleep
waltz
entails
neutralentailscontradicts
entails
move
tangodance
dance
dance
Memoriza8on(trainingset): Generaliza8on(testset):
?
?
dance
waltz
???move
???tango
sleep???waltz
tango???move
Lecture1,Slide23
RichardSocher
5/27/15
Naturallanguageinference:defini8ons!
Lecture1,Slide24
RichardSocher
5/27/15
AminimalNNforlexicalrela8ons
?
Wordsarelearnedembeddingvectors.
?
OneplainRNNorRNTNlayer
?
SoRmaxemitsrela8onlabels
?
LearneverythingwithSGD.
Lecture1,Slide25
RichardSocher
5/27/15
Recursioninproposi8onallogic
Experimentalapproach:Trainonrela8onalstatementsgenerated
fromsomeformalsystem,testonothersuchrela8onalstatements.
Themodelneedsto:
?
?
Learntherela8onsbetweenindividualwords.(lexicalrela8ons)
Learnhowlexicalrela8onsimpactphrasalrela8ons.
Thisneedstoberecursivelyapplicable!
?
a≡a,
a^(nota),
a≡(not(nota)),
...
Lecture1,Slide26
RichardSocher
5/27/15
NaturallanguageinferencewithRNNs
?
Twotrees+learnedcomparisonlayer,thenaclassifier:
Lecture1,Slide27
RichardSocher
5/27/15
NaturallanguageinferencewithRNNs
Lecture1,Slide28
RichardSocher
5/27/15
Ques8onAnswering:QuizBowlCompe88on
?
QUESTION:
HeleRunfinishedanovelwhose8tlecharacterforgeshisfather'ssignaturetogetoutofschoolandavoidsthedraRbyfeigningdesiretojoin.AmorefamousworkbythisauthortellsoftheriseandfallofthecomposerAdrianLeverkühn.AnotherofhisnovelsfeaturesthejesuitNapthaandhisopponentSe^embrini,whilehismostfamousworkdepictstheagingwriterGustavvonAschenbach.NamethisGermanauthorofTheMagicMountainandDeathinVenice.
?
Iyyeretal.2014:ANeuralNetworkforFactoidQues8on
AnsweringoverParagraphs
Ques8onAnswering:QuizBowlCompe88on
?
QUESTION:
HeleRunfinishedanovelwhose8tlecharacterforgeshisfather'ssignaturetogetoutofschoolandavoidsthedraRbyfeigningdesiretojoin.AmorefamousworkbythisauthortellsoftheriseandfallofthecomposerAdrianLeverkühn.AnotherofhisnovelsfeaturesthejesuitNapthaandhisopponentSe^embrini,whilehismostfamousworkdepictstheagingwriterGustavvonAschenbach.NamethisGermanauthorofTheMagicMountainandDeathinVenice.
ANSWER:ThomasMann
?
RecursiveNeuralNetworks
?
Followdependencystructure
PushingFactsintoEn8tyVectors
QantaModelCanDefeatHumanPlayers
LiteratureQues8onsareHard!
VisualGrounding
?
Idea:Mapsentencesandimagesintoajointspace
?
Socheretal.2013:
GroundedComposi8onalSeman8csforFindingandDescribingImageswithSentences
Discussion:Composi8onalStructure
?
RecursiveNeuralNetworkssofar
usedcons8tuencytrees
whichresultsinmoresyntac8callyinfluencedrepresenta8ons
?
Instead:Usedependencytreeswhichcapturemore
seman8cstructure
Convolu8onalNeuralNetworkforImages
?
?
CNNtrainedonImageNe
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權(quán)益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網(wǎng)頁內(nèi)容里面會有圖紙預(yù)覽,若沒有圖紙預(yù)覽就沒有圖紙。
- 4. 未經(jīng)權(quán)益所有人同意不得將文件中的內(nèi)容挪作商業(yè)或盈利用途。
- 5. 人人文庫網(wǎng)僅提供信息存儲空間,僅對用戶上傳內(nèi)容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內(nèi)容本身不做任何修改或編輯,并不能對任何下載內(nèi)容負責。
- 6. 下載文件中如有侵權(quán)或不適當內(nèi)容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 2025年度幼兒園兒童安全教育教材及課程開發(fā)協(xié)議
- 2025年度城市綜合體物業(yè)管理服務(wù)合同協(xié)議范本
- 2025年產(chǎn)品質(zhì)量保證協(xié)議文本
- 2025年倉儲場地續(xù)租合同樣本
- 2025年市場擴張戰(zhàn)略策劃咨詢協(xié)議
- 市場調(diào)研與分析服務(wù)框架協(xié)議
- 2025年飲料酒項目規(guī)劃申請報告模范
- 2025年中藥材市場分析與采購服務(wù)合同
- 2025年滌綸短纖項目規(guī)劃申請報告模范
- 2025年鼻毛修剪器項目規(guī)劃申請報告模稿
- 勞動感悟800字作文30篇
- 尚書全文及譯文
- 華師大版初中數(shù)學中考總復習全套課件
- 動物外科與產(chǎn)科
- 上下樓梯安全我知道安全教育課件
- 手術(shù)風險及醫(yī)療意外險告知流程
- 綜合實踐活動六年級下冊 飲料與健康課件 (共16張PPT)
- 《醫(yī)院重點專科建設(shè)專項資金管理辦法》
- 最新短視頻運營績效考核表KPI(優(yōu)選.)
- 設(shè)備基礎(chǔ)隔振設(shè)計探討
評論
0/150
提交評論