Narrative prose generation

2024年5月7日发(作者：)

NarrativeProseGeneration

ay

DepartmentofComputerScience

NorthCarolinaStateUniversity

Raleigh,NC27695-8206USA

cbcallaw@

Abstract

Storygenerationisexperiencingarevival,despite

disappointingpreliminaryresultsfromthepreced-

heprinciplereasonsfor

previousinadequacieswasthelowlevelofwriting

quality,whichresultedfromtheexcessivefocusof

ghthesesys-

temsleveragednarrativetheoryviacorporaanaly-

ses,theyfailedtothoroughlyextendthoseanaly-

result

wasnarrativesthatwererecognizableasstories,but

whoseprosequalitywasunsatisfactory.

However,theblameforpoorwritingqualitycannot

belaidsquarelyatthefeetofstorygrammars,as

naturallanguagegenerationhasto-datenotfielded

systemscapableoffaithfullyreproducingeitherthe

varietyorcomplexityofnaturallyoccurringsto-

perpresentstheA

UTHOR

architecture

foraccomplishingpreciselythattask,theS

TORY

-

B

OOK

implementationofanarrativeprosegenera-

tor,andabriefdescriptionofaformalevaluationof

thestoriesitproduces.

1Introduction

Despiteextensiveresearchinthefieldsofstorygeneration

andnaturallanguagegeneration,collaborativeresearchbe-

reason

forthisisthedifficultnatureoftheproblemsencounteredre-

spectivelyinthesefienerators

[

Meehan,1977;

Yazdani,1982;Lebowitz,1985;Turner,1994;Lang,1997

]

,

typicallyaddressthemacro-scaledevelopmentofcharacters

andplot,slowlyrefiningfromthetopmostnarrativegoallevel

downtoindividualdescriptionsandcharacteractionsbypro-

ile,workin

naturallanguagegeneration(NLG)focusesonlinguisticphe-

nomenaattheindividualsentencelevel,andonlyrecently

haveNLGsystemsachievedtheabilitytoproducemulti-

mainsisasubstantialgapbetween

thenarrativeplansproducedbystorygeneratorsandthere-

quirementsofNLGsystems.

Thisisexplainedbythehistoricresearchprogramsofthese

twodistinctfienerationoriginallydescendsfrom

DepartmentofComputerScience

NorthCarolinaStateUniversity

Raleigh,NC27695-8206USA

lester@

theapplicationofplanningformalismstotheworkofsoci-

olinguistssuchasVladimirPropp

[

Propp,1968

]

,whocre-

atedstorygrammarstocapturethehigh-levelplotelements

found

[

ork(Figure1)inthisarea

Meehan,1977;Yazdani,1982;Lebowitz,1985

]

focuseson

thecreationofcharactersandtheirinteractionswithplotele-

atestofthese,Lebowitzstates,“Eventually,we

expectU

NIVERSE

tobeabletogenerateconnectedstoriesin

mo-

ment,weareconcentratingongeneratingplotoutlines,and

leavingproblemsofdialogueandotherlow-leveltextgener-

ationforlater.”Moreover,eventhemostrecentstorygenera-

tionsystems,suchasM

INSTREL

andJ

OSEPH

[

Turner,1994;

Lang,1997

]

,focusoncharactersandplotwhengenerating

text,withoutconsideringtheactuallinguisticstructuresfound

inthetextstheyareattemptingtomimic(Figure2).

However,thelackofprogressinachievingcomputer-

producedstoriescharacterizedbyhigh-qualityproseisfar

thannarrativegeneration,most

full-scaleNLGsystems

[

Hovy,1993;Young,1996;Ho-

racek,1997;LesterandPorter,1997;Mittaletal.,1998;

Callawayetal.,1999

]

insteadfocusonexplanationgener-

ation,creatingscientificorinstructionaltextwhichsignif-

icantlydiffersinthedistributionandfrequencyofsyntac-

tic,semantic,andorthographicfeaturesfromthatfoundin

narrativeprose(althoughafewprojectsdoaddresssomeof

,

[

KantrowitzandBates,1992;Robin,1994;

Doran,1998;Casselletal.,2000

]

).Inaddition,themost

advancedofthesesystemsarestillnotcapableofproducing

morethantwoparagraphsoftext,whilethevastmajorityof

naturallyoccurringnarrativesareatleastseveralpageslong.

Finally,noneofthesesystemsareintendedtoacceptnarrative

plansfromatypicalstorygenerator.

TobridgethegapbetweenstorygeneratorsandNLGsys-

tems,wehavedeveloped

[

theA

UTHOR

narrativeprosegen-

erationarchitectureCallaway,2000

]

tocreatehigh-quality

narrativeprosecomparableto,andinsomecasesidenticalto,

chitecture

hasbeenimplementedinS

TORY

B

OOK

,anend-to-endnarra-

tiveprosegenerationsystemthatutilizesnarrativeplanning,

sentenceplanning,adiscoursehistory,lexicalchoice,revi-

sion,afull-scalelexicon,andthewell-knownF

UF

/S

URGE

[

Elhadad,1992

]

surfacerealizertoproducemulti-pagesto-

riesintheLittleRedRidingHoodfairytaledomain.

ONCEUPONATIMEGEORGEANTLIVED

ASANEST

IRDLIVEDINTHE

ASSOMEWATERINARIVER.

WILMAKNEWTHATTHEWATERWASINTHE

KNEWTHATTHEWATERWAS

WILMAWASVERY

ANTEDTOGETNEARSOME

LEWFROMHERNESTACROSS

AMEADOWTHROUGHAVALLEYTOTHERIVER.

ASN’T

VERYTHIRSTYANYMORE.

Figure1:ProsegeneratedbyT

ALE

-S

PIN

,1977

Narrativeprosedifferslinguisticallyfromtextfoundinex-

planatoryandinstructionalpassagesinanumberofways:

The

nying

existence

difficulties

of

of

character

orthographic

dialogue

markers

with

[

the

Doran,

accompa-

1998;

Callaway,2001

]

,speaker-hearerrelationships,locu-

tionalrelationsandmannerclauses,interjections,and

tance,the

followingwouldneverbefoundinexplanatorytext:

“Bewarethewolves,”hermothersaidinahushedvoice.

Since

islittle

explanatory

needtoinclude

textlacks

personal

dramatic

pronouns,

characters,

highly

there

id-

iomatictextaboutpersonalneeds,orintentionaldesires

suchaswanting,needing,orknowing.

Without

abletoget

character

byusing

dialogue,

onlypresent

explanatory

verbtenses

text

with

isusually

anoc-

casionalreferencetoeventsinthepastwhendiscussing

r,dialogueandthecom-

plexinteractionsbetweencharactersopensuptheneed

toperformatleastsimplistictemporalreasoningandre-

alizationsincomplexpresent,futureandpasttenses.

Because

innarrative

human

(e.g.

authors

,Hemingway

employ

vs.

widely

Joyce)

differing

asopposed

styles

to

explanatoryorinstructionaltextwhichtriestoadhereto

stricterconventions,anarrativeprosegeneratorshould

becapableofmimickingthosedifferenttypesofstyles.

Finally,

common

a

prose

narrative

formatting

prose

conventions,

generatormust

such

conform

asknowing

to

whentoforceparagraphbreaksandbeingabletogener-

atewrittenstylisticeffectslikeonomatopoeia,regional

dialects,,“Ewwww!”“B-b-but,it’s

s-s-scary!”“Mom,youCAN’Tdothat!”)

S

TORY

B

OOK

iscapableofreproducingthesephenomena,

anddoingsoinbothgrammaticallycorrectEnglishandpass-

ableSpanish

[

Callawayetal.,1999;Callaway,2001

]

.

Uponreceivingahigh-levelstoryspecificationfromanar-

rativeplanner,S

TORY

B

OOK

(1)structuresitintoparagraph

andsentence-sizedchunks,(2)conductsadiscoursehistory

analysistodetermineindefinitereferencesandpronominal-

izations,(3)performsalexicalchoiceanalysistoincrease

varietyamongconceptsandeventrelations,(4)mapsac-

tors,propsandeventstosemantic/syntacticrolesinfulllin-

guisticdeepstructures,(5)revisesparagraph-sizedgroupsof

onedayithappenedthatpeasant

is

happened,

response,peasanttookawalkinthe

tfoundapitwhenhe

is

happened,peasantdesiredtopunish

onse,peasantmadeit

hisgoalthatwifewouldbeinthe

s

tlivedalone.

Figure2:ProsegeneratedbyJ

OSEPH

,1997

deepstructuresviaaggregationandreorderingtoeliminate

theshort,choppysentencescharacteristicoftextproducedby

discourseplanningsystems,and(6)performssurfacereal-

izationwithintegratedformattingtoproducenarrativeprose

similartothatfoundinstorieswrittenbyhumanauthors.

ToevaluatethequalityofthenarrativesthatS

TORY

B

OOK

produces,wecreatedasimplifiednarrativeplannercapable

ofgeneratingtwoLittleRedRidingHoodstoriesexpressed

intherequiredhigh-levelstoryspecificreated

fiveversionsofS

TORY

B

OOK

variouslyablatingthediscourse

history,lexicalchoice,andrevisioncomponentstoproducea

totalof10storyversionswhichwerethenformallyevaluated

ultsshowedsignificantdiffer-

encesbetweentheinclusionorablationofindividualarchi-

tecturalcomponents.

2NarrativeRepresentation

Whilemostresearchersinstorygenerationutilizeplanning

mechanisms

[

orstorygrammars,agrowingliteratureonnar-

ratologyPropp,1968;Segre,1988;Bal,1997

]

positsthat

narrativeconsistsofthefabula,orsumtotalofknowledgeand

factsaboutanarrativeworld,andthesuzjet,ortheordering

andspecificsaboutwhattheauthorpresentsandatwhichpo-

sition(s)

UTHOR

archi-

tectureadoptsthisviewandcomputationalizesittodescribe

therequirementsofanarrativeplannerandanarrativeprose

generator:thenarrativeplannerisresponsibleforcreating

boththefabulaandsuzjet,whilethenarrativeprosegenerator

isresponsibleforconvertingthemintotextuallyrecognizable

narratives.

Anarrativeworldisalsopopulatedwithalargenumber

ofscenes,characters,props,locations,events,anddescrip-

TORY

B

OOK

implementationexplicitlyrepre-

sentsthisknowledge,whichformsthebasisofthefabula.

Initially,thefabulacontainsonlyontologicalinformation,in-

cludingtheexistenceofbroadconceptssuchasforest,cot-

tage,andperson,andconceptrelationslikenext-to,mother-

of,andmoves-toward.S

TORY

B

OOK

assumesthatanarrative

plannerisresponsibleforconstructingthespecificconceptin-

,LittleRedRiding

HoodlivesinCottage001,whichisherhouse,whileher

grandmother(Grandmother001)livesinadifferenthouse,

Cottage002.