Computational Narrative - and what we can learn from ...ouyangj/JessicaOuyang_candidacy.pdf ·...

Computational Narrativeand what we can learn from narratology

Jessica Ouyang

Columbia University

September 2015

1/1

Introduction

storytelling is a fundamental and universal form of communication

I how are narratives different from non-narratives?I what do narratives tell us about authors and audiences?I can we generate narratives automatically?

2/1

Outline

where did we come from?I literary and linguistic studies of narrative

where are we now?I the current state of computational narrative research

where are we going?I how emotion detection can and should help us improve

computational narrative

3/1

NarratologyStructuralism

Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Propp (1928)

I there is a limited set offunctions that are thebuilding blocks of a fairy tale

I the sequence of functions isalways the same in everyfairy tale

some functions are more specificthan others

not all functions are equallyimportant

Vladimir Propp (1968). Morphology of the Folktale.4/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Todorov (1969)

I the structural approach tonarrative seeks an abstracttheory of the structure ofliterary discourse such thatall existing works areparticular realized instancesof that structure

I the minimal complete plot isa shift from one equilibirumstate to another

T. Todorov (1969). Structural analysis of narrative.5/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Barthes (1975)

I there must exist a commonnarrative structure,otherwise narratives are justrandom sequences of events

I we can reliably produce andrecognize narratives, so theycannot be random

R. Barthes (1975). An introduction to the structural analysis of narrative.6/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Barthes (1975)three levels of analysis

I functions

I charactersI narration

functions↙ ↘

events states↙ ↘

causal temporal

what are the decision points of anarrative?



Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae


I functionsI characters

I narration

characters are defined by actionsI have perspectives on

sequences of actionsI participate in relationshipsI form subject/object,

giver/recipient, andassistant/opposer pairs



Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae


I functionsI charactersI narration

the discourse of the narrativeI who is the narrator?

I the authorI an omniscient observerI a character

I who is the reader?



Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

the story so far –I narrative has two parts

I surface form(narration/discourse)

I deep structure(functions/characters)

I every existing narrative is arealization of a deepstructure

I eg. Hamlet and The LionKing are two realizationsof the same deepstructure

this is the structuralist position

8/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

but wait –I even if the deep structure is

the same, the surfacerealization is very different

I even if the surfacerealization is the same, thereader’s interpretation canbe very different

I L. Bohannan (1966).Shakespeare in the Bush.

this is the contextualist position

9/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Smith (1980)

I narratives are not juststructures, but also acts

I a narrative has a narratorand reader, both of whommust be interested in thenarrative

I what does the narratorwant to convey?

I what does the reader takeaway from the narrative?

B. H. Smith (1980). Narrative versions, narrative theories.10/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Polanyi (1981)

I a narrative must be tellableI the durative-descriptive

structure shows thesociocultural context

I evaluation devices indicateimportant material thatforms an adequateparaphrase of the story

the surface realizationindicates the teller’s intendeddeep structure

L. Polanyi (1981). What stories can tell us about their teller’s world.11/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

the lay of the land –

StructuralismI literatureI functions and actions

ContextualismI personal narrativeI discourse and context

12/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Chatman (1990)

I contextualists assume a realauthor/real audiencerelationship, which does notapply to literature

the theory might not apply toliterature, but this does not makeit less true in its intended domain

S. Chatman(1990). What can we learn from contextualist narratology?13/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Chatman (1990)

I there is less variety andinnovation of the form byamateur storytellers than byliterary authors

this makes amateur stories moresuitable for defining a generalstructure for narrative



Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Chatman (1990)

I personal narratives do notdistinguish between thesurface realization and thedeep structure

this makes them perfect forstudying deep structure



Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Labov (2013)abstract

contextualistorientation

structuralistcomplicating action

structuralistmost reportable event

contextualistevaluation

contextualistresolution

structuralistcoda

contextualistW. Labov (2013). The Language of Life and Death.

14/1

OutlineStructuralism

Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

NarratologyI structuralism: literature,

functions, character actionsI contextualism: personal

narrative, discourse, context

Computational NarrativeI coming up next

Emotion DetectionI later

15/1

Computational NarrativeStructuralism

Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Lehnert (1981)plot units

I vertices: affect statesI edges: causal links

–

M

m

⇒ Proppian functions

W. Lehnert (1981). Plot units and narrative summarization.16/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Elson (2012)Story Intention Graph (SIG)

I three layers of annotationI textual ⇐ narrationI timeline ⇐ functionI interpretive⇐ character

I annotators felt there weremultiple interpretations forsome narratives

⇒ Barthes’s layers of analysis⇒ Contextualism

D. Elson (2012). DramaBank: annotating agency in narrative discourse.17/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Halpin, Moore, and Robertson(2004)

I 103 stories rewritten bychildren

I does the pupil understandthe point of the story andemphasize important linksand details?

⇐

I extract and compare theevents of the rewritten storywith exemplar

⇐

⇒ contextualist question⇒ structuralist answer

H. Halpin, J. D. Moore, and J. Robertson (2004). Automatic analysis ofplot for story rewriting.

18/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae



I does the pupil understandthe point of the story andemphasize important linksand details? ⇐

I extract and compare theevents of the rewritten storywith exemplar

⇐

⇒ contextualist question

⇒ structuralist answer


18/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae



I does the pupil understandthe point of the story andemphasize important linksand details? ⇐

I extract and compare theevents of the rewritten storywith exemplar ⇐

⇒ contextualist question⇒ structuralist answer


18/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Chambers and Jurafsky (2008)narrative chain

I partially ordered set ofevents with a commonprotagonist

⇒ Barthes: characters aredefined by their actions

system average rank

verb co-occurence 1826protagonist 1160

but is this good performance?how to interpret these numbers?

N. Chambers and D. Jurafsky (2008). Unsupervised learning of narrativeevent chains.19/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

McIntyre and Lapata (2009)

I construct DAG from eventchains (Chambers andJurafsky)

I generate narratives bywalking the DAG

The giant guards the child. Thechild rescues the son from thepower. The child begs the sonfor a pardon. The giant criesthat the son laughs the happinessout of death. The child hears ifthe happiness tells a story.

N. McIntyre and M. Lapata (2009). Learning to tell tales: a data-drivenapproach to story generation.

20/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Reidl and Young (2010)

I Intent-Driven PartialOrder Causal Linkplanning problem

I for a plan to be complete, itmust not contain anycharacter actions that arenot part of a frame ofcommitment

The genie has a frighteningappearance. The genie appearsthreatening to Aladdin. Aladdinwants the genie to die.

M. Reidl and R. M. Young (2010). Narrative planning: balancing plot andcharacter.

21/1

Computational Narrative

the story so far –

work data task approach

Lehnert – summarization plot unitsElson fables annotation SIGHaplin rewritten stories grading event similarityChambers Gigaword script extraction event chainMcIntyre fairy tales generation event chainReidl – generation IPOCL

structuralist ↗

22/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Montfort (2011)I interactive fiction system

that supports several modelsof narration

I reverse chronologicalorder

I flashback and flashforwardI tense generated

automatically based on aninternal state time ofnarration and the time ofthe event

N. Montfort (2011). Curveship’s automatic narrative style.23/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Gordon and Swanson (2009)

I ICWSM 2009 Spinn3r datsetI annotated 5002 posts

I 240 stories (4.8%)I simple unigrams normalized

by frequencyI POS tags and title text

not statistically significantI created dataset of 960,098

stories (precision = 0.75)

⇒ narratives of personalexperience

A. S. Gordon and R. Swanson (2009). Identifying personal stories inmillions of weblog entries.

24/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Swanson et al (2014)

I identify orientation,complicating action, andevaluation

I several rounds of annotationto achieve agreement

I problematic clause typesI clauses containing of

multiple elementsI implied actionsI stative descriptions

resulting from local actionI subjective language

R. Swanson et al (2014). Identifying narrative clause types in personalstories.

25/1


Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

Sagae et al (2013)

I diegetic versus extradiegeticI subjective versus objective

task accuracy

six-way 58%binary subjectivity 78%

binary diegetic 81%

K. Sagae et al (2013). A data-driven approach for classification ofsubjectivity in personal narrative.

26/1

OutlineStructuralism

Propp

Todorov

Barthes

Chatman

Smith Polanyi

LabovContextualism

Lehnert

Elson

Chambers Halpin

Generation

McIntyre Reidl

MontfortGordon

Swanson Sagae

NarratologyI structuralism: literature,

functions, character actionsI contextualism: personal

narrative, discourse, context

Computational NarrativeI previous work has focused

on structuralist approachesI contextualist approaches are

needed

Emotion DetectionI coming up next

27/1

Emotion DetectionRubin Mishne

Tokuhisa

basic emotionsLiu

Chaumartin

Strapprava

Alm

Calix

Ulinski

Liu, Lieberman, and Selker (2003)I six basic emotions: happy, sad,

anger, fear, disgust, surpriseI Ekman (1993)

I evaluationI email client that illustrates

sentences with facesI faces generated by emotion

system rated more interactiveand intelligent than random faces

what was the rating scale?

H. Liu, H. Lieberman, and T. Selker (2003). A model of textual affectsensing using real-world knowledge.

28/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Rubin, Stanton, and Liddy (2004)I Circumplex Theory of Affect

(Watson and Tellegen 1985)I positive affect axis (active, elated

v. drowsy, dull)I negative affect axis (distressed,

fearful v. calm, placid)

unintuitive axes

too much overlap among labels in thesame octant

V. Rubin, J. Stanton, and E. Liddy (2004). Discerning emotions in texts.29/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Mishne (2005)

I top 40 LiveJournal moods (eg.amused, excited, contemplative,sick, anxious, ecstatic)

too much subjectivity in self-reportedmoods

G. Mishne (2005). Experiments with mood classification in blog posts.30/1


Tokuhisa

basic emotionsLiu

Chuamartin

STrapparava

Alm

Calix

Ulinski

Tokuhisa, Inui, and Matsumoto(2008)

I happiness, pleasantness, relief, fear,sadness, disappointment,unpleasantness, loneliness, anxiety,anger

I fatal v. non-fatal errors

too much overlap among labels

R. Tokuhisa, K. Inui, and Y. Matsumoto (2008). Emotion classificationusing massive examples extracted from the web.

31/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Chaumartin (2007)

I fine- v. coarse-grain evaluationI WordNet Affect and SentiWordNet

scores weighted by syntactic role ofword

I high accuracy, moderate precision,low recall

fine-grained control of emotion ispossible

F.-R. Chaumartin (2007). UPAR7: a knowledge-based system for headlinesentiment tagging.

32/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Strapparava and Mihalcea (2008)I LSA representation of WordNet

AffectI outperformed by Chaumartin on

fine-grained evaluationI performs best on coarse-grained

evaluationI strict WN keyword has best

precision (38.28)I synset augmented WN has best

recall (90.22)

C. Strapparava and R. Mihalcea (2008). Learning to identify emotions intext.

33/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Alm, Roth, and Sproat (2005)

I 185 children’s storiesI emotions in adjacent sentences can

affect the current sentenceI strongest feature group included

thematic story type

⇐

different types of stories should be toldin different ways

C. Alm, D. Roth, and R. Sproat (2005). Emotions from text: machinelearning for text-based emotion prediction.

34/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Alm, Roth, and Sproat (2005)

I 185 children’s storiesI emotions in adjacent sentences can

affect the current sentenceI strongest feature group included

thematic story type ⇐

different types of stories should be toldin different ways

C. Alm, D. Roth, and R. Sproat (2005). Emotions from text: machinelearning for text-based emotion prediction.

34/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Calix et al (2010)

I mutual information between wordsand emotions

I learned list of words outperformedhand-crafted word lists

I performance varied by author

authors express emotions in differentways

R. Calix, S. Mallepudi, B. Chen, and G. Knapp (2010). Emotionrecognition in text for 3-d facial expression rendering.

35/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Calix et al (2010)

I mutual information between wordsand emotions

I learned list of words outperformedhand-crafted word lists

I performance varied by author ⇐

authors express emotions in differentways

R. Calix, S. Mallepudi, B. Chen, and G. Knapp (2010). Emotionrecognition in text for 3-d facial expression rendering.

35/1


Tokuhisa

basic emotionsLiu

Chaumartin

Strapparava

Alm

Calix

Ulinski

Ulinski, Soto, and Hirschberg (2012)

I 4443 LiveJournal entriesI 660 annotated WordsEye images

I literal descriptionI basic emotionI explanation for why that emotion

was chosenI LIWC classes, tf-idf of (word

n-gram, POS n-gram) pairsI cross-domain classification

M. Ulinski, V. Soto, and J. Hirschberg (2012). Finding emotion in imagedescriptions.

36/1

Conclusion

structuralist narratology examines functions and character actionsin literature; contextualist narratology examines discourse andcontext in personal narrative

most work in computational narrative has focused on structuralistapproaches, but contextualist approaches should be explored,especially with the amount of online data now available

in particular, work on emotion in text can help in computationalnarrative tasks by uncovering authors’ intents and opinions

questions?

37/1

Computational Narrative - and what we can learn from ...ouyangj/JessicaOuyang_candidacy.pdf ·...

Documents

Transcript of Computational Narrative - and what we can learn from ...ouyangj/JessicaOuyang_candidacy.pdf ·...