26
Narrative Story Generation A Skeleton-Based Model for Promoting Coherence 4-11-2018 Jingjing Xu, Xuancheng Ren, Yi Zhang, Qi Zeng, Xiaoyan Cai, Xu Sun MOE Key Lab of Computational Linguistics, School of EECS, Peking University School of Automation, Northwestern Polytechnical University [email protected] A Skeleton-Based Model for Promoting Coherence Among Sentences in Narrative Story Generation

A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018

JingjingXu,XuanchengRen,Yi Zhang,QiZeng, Xiaoyan Cai,XuSun

MOEKeyLabofComputationalLinguistics,SchoolofEECS,PekingUniversitySchoolofAutomation,NorthwesternPolytechnical University

[email protected]

ASkeleton-BasedModelforPromotingCoherenceAmongSentencesinNarrativeStoryGeneration

Page 2: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 2 of28

p Introduction

Ø Task

Ø Challenge

p Skeleton-BasedModel

Ø Overview

Ø Skeleton-BasedGenerativeModule

Ø SkeletonExtractionModule

Outline

p Experiment

Ø Dataset

Ø Baselines

Ø Results

p Analysis

Ø IncrementalAnalysis

Ø ErrorAnalysis

p Conclusion

Page 3: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

Introduction

Page 4: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 4 of28

NarrativeStoryGeneration

TaskDefinitionInput:Ashort description ofasceneoranevent.Output:Arelevant narrativestoryfollowingtheinput.

Input:Fans cametogethertocelebrate theopening ofanewstudio foranartist.

Output:Theartist providedchampagne influtesforeveryone.Friendstoasted andcheered theartistassheopened hernewstudio.

Input: LastweekIattendedawedding forthefirsttime.

Output: Therewerealotoffamilies there.Theywerealltakingpictures together.Everyonewasveryhappy.Thebride andgroom gottorideinalimothattheyrented.

Page 5: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 5 of28

q Highrequirementsontightsemanticconnectionamongsentences

q Ifthegiveninputisabout“artistandstudioopening”

Theartistprovidedchampagneforeveryone

Fanstoastedandcheeredtheartist

Thefanswereveryhappytoseethisevent

Thefootballgameissoexciting

Theweatherissocoldandalltreesarecoveredwithheavysnow

Difficulty

Page 6: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 6 of28

Motivation

q Theconnectionamongsentencesismainlyreflectedthroughkeyphrases (orskeleton)

q Weexplicitlymodeltherelationamongskeletonsinthiswork.

Page 7: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 7 of28

q Motivatedbythisfact:

1.Wedonotgenerateasentencefromlefttoright

2.Instead,wefirstgenerateaskeleton

3.Expandtheskeletontoacompleteandfluentsentence

Skeleton-basedModel

Page 8: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

Approach

Page 9: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 9 of28

q Ourmodelcontainstwoparts

Ø Skeleton-basedgenerativemodule

1. Input-to-skeletoncomponent

2. Skeleton-to-targetcomponent

Ø Skeletonextractionmodule

Overview

Page 10: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 10 of28

q GenerationProcess

Skeleton-BasedGenerativeModule

Fanscametogethertocelebratetheopeningofanewstudioforanartist.

The artist provided champagne in flutes for everyone.

artist champagne

Friends toasted cheered artist.

Friends toasted and cheered the artist as she opened her new studio.

Page 11: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 11 of28

q Input-to-SkeletonComponent

Ø Learnthedependencybetweensourceinputsandtargetskeletons

Ø Model:ASeq2Seq

Ø TrainingLoss:Cross-entropyloss

q Skeleton-to-TargetComponent

Ø Expandaskeletontoacompletesentence

Ø Model:Seq2Seq

Ø Encoder:LSTM

Ø Decoder:LSTM

Ø TrainingLoss:Cross-EntropyLoss

TheStructureofSkeleton-BasedGenerativeModule

Page 12: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 12 of28

q SkeletonsassupervisorysignalØ Inreal-worlddatasets,thehuman-annotatedskeletonisusuallyunavailable

q Rule-basedapproaches

Ø Advantage:Simple

Ø Drawback:Difficulttodefinetheunifiedrulesofextractingskeletons

q Toaddressthisproblem,webuildaskeletonextractionmoduletoautomaticallyexploresentenceskeletons

Skeleton-ExtractiveModule

Page 13: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 13 of28

q Reformulateskeletonextractionasasentencecompressionproblem

Ø Advantage:sentencecompressionrequiressystemstokeepcore

information,whichisconsistentwithourtarget.

Ø Drawback:sentencecompressionresultsusuallyaregrammatical,which

limitsthequalityofskeletons.

Skeleton-ExtractionModule

Thelady wearingthepinkshirtdecidedtostopplayingthevideoandchattedwithotherguests.

Lady decidedtostopplayingthevideo

compress

Page 14: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 14 of28

q Sequencetosequencemodel

Ø Encoder:LSTM

Ø Decoder:LSTM

q Input:originaltext

q Output:theskeletonoftheoriginaltext

q Lossfunction:cross-entropyloss

ThestructureofSkeletonExtractionModule

Page 15: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 15 of28

qMotivation

q Usethefeedbackofgenerativemoduletoimproveit

q Challenge

q Lossisnolongerdifferentiable

q Approach

q ReinforcementLearningMethod

q Treatthefeedbackofgenerativemoduleasreward

q Optimizetheextractionmodelby

Howtoimprovethequalityofskeleton-extractionmodule

Page 16: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 16 of28

qMultiplicationofthecross-entropyloss

• K:Upperboundofthereward,

• R1:Cross-entropylossesintheinput-to-skeletoncomponent

• R2:Cross-entropylossesintheskeleton-to-targetcomponent,respectively.

Reward

Page 17: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

Experiment

Page 18: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 18 of28

q Source:

ü VisualStorytelling

q Dataset:

ü Input:thefirstsentence

ü Output:thefollowingsentences

Dataset

Having a good time bonding and talking.

[M] got exhausted by the heat.

Sky illuminated with a brilliance of gold and orange hues.

Page 19: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 19 of28

q Entity-EnhancedSeq2SeqModel(EESeq2Seq)

ü Combineentitycontextandtextcontexttogether

q Dependency-TreeEnhancedSeq2SeqModel(DE-Seq2Seq)

ü Basedondependencyparsinglabels

q Generalized-TemplateEnhancedSeq2SeqModel(GE-Seq2Seq)

ü Useexistingknowledgebases(WordNet,VerbNet)togetageneralized

sentencerepresentation

Baselines

Page 20: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 20 of28

q AutomaticEvaluation

Results

Page 21: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 21 of28

q HumanEvaluation

ü 100items

ü Fluencyandcoherence

Results

Page 22: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 22 of28

q Input1:[Female]andherfriendhadanightoutonthetown.

Ø EE-Seq2Seq:Theywereveryhappytoseeus.[Male]and[female]were

gettingmarried today.

Ø DE-Seq2Seq:Wewalkedthroughthestreetsofthecityandfoundaplaceto

sit.Thewholefamilycameouttosupporttheirfamilies.

Ø GE-Seq2Seq:MyfriendsandIwenttoabarlastnight.[Female]wasso

happytobethere.

Ø ProposedModel:Theywenttothebar.Theyhadagreattime.

Generatedexamples

Page 23: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

Analysis

Page 24: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 24 of28

q 4typesoferrors

ü Irrelevantscene

ü Chaoticsyntax

ü Chaotictimeline

ü Repeatedscene

ErrorAnalysis

Page 25: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 25 of28

A. Weproposeanewskeleton-basedmodelforgeneratingcoherentnarrativestories

B. Experimentalresultsshowthatourmodelsignificantlyimprovesthequalityofgeneratedstories

C. Erroranalysisshowsthattherearestillmanychallengesinnarrativestorygeneration,whichwe

wouldliketoexploreinthefuture

Conclusion

Page 26: A Skeleton-Based Model for Promoting Coherence Among ... · A Skeleton-Based Model for Promoting Coherence Narrative Story Generation 4-11-2018 Jingjing Xu, Xuancheng Ren, YiZhang,

NarrativeStoryGenerationASkeleton-BasedModelforPromotingCoherence 4-11-2018 26 of28

Thank You!

Ifyouhaveanyquestion,[email protected]