T55.4
.W2 no.CSWEY
A
Compararive Analysis
of
Design
Rationale
Representations
JintaeLee
Kum-Yew
Lai
March
1992
WP
#
84-92
INTERNATIONAL
CENTER
FOR
RESEARCH ON
MANAGEMENT
OF
TECHNOLOGY
Massachusetts Institute ofTechnology Sloan Schoolof
Management
The
International
Center
for
Research
on
the
Management
of
Technology
A
Compararive Analysis
of
Design
Rationale
Representations
JintaeLee
Kum-Yew
Lai
March
1992
WP
#
84-92
Sloan
WP#
3295
&
3405
CCS TR#
121
A
revised
and
condensed
version
of thisreport appears
inthe
special issue ofHuman-Computer
Interaction
on
design
rationale, v.6(3-4),pp.251-280.
©
1992 Massachusetts
Institute ofTechnology
Sloan School
ofManagement
Massachusetts
Institute ofTechnology
38
Memorial
Drive, E56-390
d/l.l.T.
A
revisedand
condensedversion ofthisreportappearsinthe special issue ofHuman-Computer
Interactionon
designrationale,v.6(3-4),pp.251-280A
Comparative Analysis
of
Design
Rationale
Representations
JintaeLee
and
Kum-Yew
LaiCenterfor Coordination Science
and
MIT
Artificial Intelligence LaboratoryABSTRACT
A
fewrepresentationshave been usedforcapturingdesignrationale. Itisimportanttoknow
inwhat ways
theyareadequateor limitedso thatwe know how
toimprove
them. Inthispaper,we
developaframework
forevaluating designrationalerepresentations basedon
a set of genericdesigntasks.We
build theframework
byprogressively differentiating theelementsofdesignrationalethat,
when made
explicit,support anincreasingnumber
ofthedesigntasks.With
thisframework,we
evaluatetheexpressivenessof the existing representations.We
alsopresentalanguage,
DRL,
thatwe
believeisthemost expressiveof the existing representationswithout beingtoo
complex
forhuman
users.We
alsodiscuss the limitations ofDRL
asopen
problemsfor further research.1.
INTRODUCTION
As
the articles inthisissuepointout,anexplicit representation ofdesignrationalecan bringmany
benefits. Sucha representationcanlead to a betterunderstandingofthe issuesinvolved[McCall1987;
Yakemovic
&
Conklin1990], of thedesign space[MacLean
etal.1989],and
of the principlesunderlyinghuman-computer
interaction [Carroll&
Campbell
1989]. Itcanalsoprovidea basis for learning,justification,
and
computational support[Fischeretal.1989;Lee1990a].
The
extent towhich
we
canactuallyreapthese benefits,however,would depend
largely
on
thelanguagewe
useforrepresenting designrationale. If,forexample, designrationale
were
representedinfreetext,the benefitswe
obtain fromitwould
notbedifferentfrom
what
we
alreadygetfromthenoteson
paperthatwe
takeindesign meetings. Also, theAuthors' present address: JintaeLee
and
Kum-Yew
Lai,CenterforCoordinationScience,MassachusettsInstitute ofTechnology, E40-140, 1
Amherst
Street,Cambridge,
MA
02139.CONTENTS
1.
INTRODUCTION
12.
WHAT
DO WE
WANT
TO
DO
WITH
DESIGN RATIONALE?
33.
WHAT'S
IN
DESIGN
RATIONALE?
54.
EXISTING
REPRESENTATIONS
154.1. IBIS (Issue Based Information System) 16
4.2. Toulmin's
Model
ofArgumentation
214.3.
QOC
(Question,Option,and
Criteria) 244.4. Other Representations 26
5.
DRL
(DECISION
REPRESENTATION
LANGUAGE)
295.1. Introduction 29
5.2. Description of
DRL
305.3. Evaluationof
DRL
as aDesign RationaleLanguage
355.4. Relation to
Other
Studies 406.
CONCLUSIONS
42kindsofcomputational supportthat
we
can providedepends on
what
a representationmakes
explicit
and
how
formaltherepresentationis.A
few systemshave beenbuiltand
actuallyusedtocapturedesignrationale orarguments
[Kunz
&
Rlttel1970;McCall1987;Conklin&
Begeman
1988; Fischer etal.1989;Lee1990a, 1990b;McCall1990],
and most
ofthem
usedrepresentationsbased
on
the earlierstudiesofdesignactivity[Kunz
&
Rittel1970]orofargumentation[Toulmin1958].
However,
thereisno
systematicattempttojustifythechoiceoftheserepresentationsordiscuss the rationale forusingthem.
This paper is motivated by thefollowing questions:
How
adequate
are the existing representations?Do
theyallow ustoeasilyrepresentwhat
we
want
torepresent? Ingeneral,how
do
we
evaluatealanguageforrepresentingdesignrationale? Thispaperisan attempttoanswer
thesequestionsby identifying theelementsofdesignrationale thatcould bemade
explicit
and by
exploringtheconsequencesofmaking them
explicit. Laying outtheseelements providesaframework
forplacing the existingmeanings
ofdesignrationaleinperspective, aswell asprovidinga
framework
forevaluatingarepresentationlanguagefordesignrationale,asComparative
AnalysisofDesignRationaleRepresentations 3We
proceedinthefollowingway. First,we
identify the tasks thatwe
mightwant
tosupport usingadesignrationale representation.Throughout
thepaper,we
willusethese tasks as a referenceagainstwhich
we
would
evaluatetherepresentations thatwe
discuss. InSection3,we
characterizedesignrationaleby presentingprogressively richermodels.We
startwithasimple
model
ofdesignrationale,where
anartifactisassociatedwithabody
ofreasonsforthechoiceof thatartifact.
We
thenelaboratethissimplemodel by
incrementallydifferentiatingand
making
explicitwhat
isimplicitinthebody
of reasons.As
we
do
so,we
discusswhat
eachresulting
model
allows ustodo. Thesemodels
ofdesignrationaleprovideaframework
inwhich
todefinethescopeof arepresentationand
itsadequacy
withinitsscope.^ Usingthisframework,
we
evaluatethe existingrepresentationsinSection4. In Section5,we
presentalanguage, called
DRL, which
we
believe ismore
expressive thanmost
of theexisting representationsand
overcomesmany
oftheirlimitations.As
we
describeDRL,
we
alsodiscussitscurrent limitations,
which
we
present asopen
problemsforfuture research.2.
WHAT DO WE WANT
TO
DO
WITH DESIGN RATIONALE?
To
evaluatea representation,we
needtoknow
what
tasksitisdesignedtosupport.The
tasksthatadesignrationalerepresentationcanorshould support can be describedin
many
ways
at different levels of abstractions.^ For example,[Mostow
1985] lists thefollowing tasks:documentation, understanding,debugging,verification,analysis,explanation,modification,
and
automation. [Fischeret al.1991] pointsoutthatdocumenting
designrationalecan support maintenanceand
redesignofanartifact,reuseofthedesignknowledge,and
criticalreflectionduring design process.
[MacLean
etal. 1991]listtwo
majorbenefitsfrom
designrationalerepresentation: aidtoreasoning
and
aid tocommunication.The
tasksofachievingthesebenefitsareelaboratedfurther intermsofsubtasks suchasenablingthedesignerstoenvisage
theavailabledesignalternatives in a
more
structuredway, withtheargumentsforand
againstthem.
Another
way
ofcharacterizing the tasksistolistthequestionsthatwe
oftenneedtoanswer
toaccomplishthegeneraltasksmentioned aboveinthedesignprocess.
To
theextent thatwe
want
our designrationalerepresentationtohelpanswer
these questions,answeringtheseInthispaper,
we
use the terms,model andrepresentation,in the followingway. Modelisan algebraicconcept,andrepresentationalinguisticone. Sincethesamestructurecan be describedinmany
ways,amodelcanhaveseveralrepresentations. Therefore,when we
wanttodiscussastructureindependentof a particular
way
of describingit,we
usetheterm model.On
theotherhand,we
usetheterm representationto refer toaparticularnotationfordescribing thestructure.There are
many
gooddiscussionsofthe benefitsortasks forwhichadesignratioanlecan be used.4
questions
become
thetasks that the representationshould support.The
followingisasetofrepresentative questionsthat
we
gatheredfrom our experiments with designrationale[Lee1991a],from walking through examples [Lewisetal.1991],
and
fromcreating scenarios [Carroll&
Rosson1990].•
What
isthe status of the currentdesign?•
What
didwe
discusslastweek and what do
we
needtodo
today?•
What
are the alternativedesignsand
what
are theirprosand
cons?•
What
are thetwo
mostfavorable alternatives so far?•
Sun
Microsystemsjust released theirX/NeWs
server.How
would
the releasechange our evaluations?•
What
ifwe
do
not considerportability?•
Why
isportabilityimportantanyway?
•
What
are the issues thatdepend on
thisissue?•
What
are theunresolvedissues?What
arewe
currentlydoing aboutthem?
• What'stheconsequencesofdoing
away
withthispart?•
How
did other people deal withthisproblem?What
canwe
learnfrom
thepastdesigncases?
The
abovelistof thequestionsisby no
means
completeastherearemany
possiblepathsthatwe
did notwalk
throughand
many
scenarios thatwe
did notconstruct.We
alsoleftoutthosequestionswhich,though important,
do
notseem
tobethe jobofdesignrationaletoanswer
(e.g.How
canwe
compute
thetotalcostofthisdesign?) Nevertheless,thequestionsinthelistprovideauseful
framework
forassessing theexpressivenessofthe differentrepresentations.When we
discuss the limited or increasedexpressivenessof agivenrepresentation,we
refertothose questionsthatcanorcannot be
answered
asaresult. Ifourtaskincludes answeringaquestion
which
isnot representedinthelist,thenwe
canalways evaluatethe representationsby
askingwhether
theywould
supportansweringthequestionand
ifnot,what
additionalobjects, attributes,or relations
would
havetobemade
explicit.We
want
toemphasize
further thatwe
areassessingonlytheexpressivenessofthe existing representations ofdesignrationale. Thatitisdesirabletoanswer
thequestionsinthelistdoes notmean
thatany
representation fordesignrationaleshouldsupporttheansweringofthese questions. Eachrepresentationmust weigh
the costsand
benefitsinvolvedin tradeoffsamong
threegeneral dimensions: expressiveness,
human
usability,and
computational tractability (Figure1). Thesetradeoffsshouldinturn be motivatedby
the tasks thatareintendedtobeComparative
AnalysisofDesign Rationale Representations 5accomplished usingthe representation. In short,
we
arenotdictatingwhat any
existingrepresentationshouldorshould nothave.
However,
we
do hope
thattheanalysispresentedin thispaperwould make
the architects of a representationfordesignrationalemore
consciousofwhat
theirlanguage canorcannotexpressand
why.
Domain
Figure1. Elementsincomputer supportedactivities.
To
besure,we
cannotseparateour concern withexpressiveness entirelyfromotherconcerns suchas
human
usability orcomputationaltractability. For example,ifalanguageismeant
tobeusedby peopletocapture designrationales,butifitistoo
complex
forpeopletomanage, thenthereisnot
much
pointinevaluatingitsexpressiveness.Whether any
ofthe representations thatwe
discussfallsinto thatcategoryisan empiricalquestion. Allthelanguages discussed herehave beenactuallyusedby
people,butthatisno
guaranteethattheywillallsucceedattheir "industrial strength"use [Conklin
&
Yakemovic
1991]. Nevertheless,we
believeitwould
bedifficulttoevaluatetradeoffs
among
thethreedimensions withoutcalibratingindividual dimensions suchasexpressiveness(cf.[Brachman
&
Levesque
1984]on
tradeoffsbetween
expressivenessand
computationaltractabilityforgeneralknowledge
representation).3.
WHAT'S
IN
DESIGN
RATIONALE?
What
is design rationale? Design rationalefor an artifacthas at least threedifferentmeanings
initscurrentusage: a historicalrecordofthereasonsfor thechoiceofanartifact[Yakemovic
&
Conklin1990], asetofpsychologicalclaimsembodied
by anartifact[Carroll&
Rosson1990],and
a description of thedesign space[MacLean
etal.1989].'^Therepresentationusedin[Yakemovic
&
Conklinl990] describeslogical aswellas historicalaspects ofdesignrationale,aswe
discussfurtherlater. For example,thesupportrelationhas both thelogicalaspectandthehistoricalaspect because,
We
associate[Yakemovic&
Conklin 1990],aswellas[Lee1990a;Potts
&
Bruns1988] withthe historicalrecord becauseoneof theirgoalistocaptureand documenttheactualprocessofdesign.6
Designrationale often
means
the historicalrecordofthe analysis that ledtothechoiceoftheparticular artifactor the featurein question.
To
illustrate, letus takeas anexample
a particularfeature of theMacintosh operating system,namely
theplacementofallthewindow
commands
intheglobalmenubar
atthe top of thescreen."*By
awindow
command,
we
mean
acommand
specifictoawindow,
e.g.SAVE
isawindow
command
thatsavesthecontents of thewindow.
A
designrationalefor thisfeature inthesenseof historicalrecordwould
be somethinglike: "Theissue of
where
toputthewindow commands
was
raisedby Mark on
January20.Kevin proposedtheidea of incorporating
them
intothe globalmenu
baratthetop of the screenand
pointedoutthatitsaves screenspace(e.g.,asopposed
toputtingthecommands
on
eachwindow,
as in the Starenvironment). Julieobjectedbecauseitrequires alongmouse
travelfromthecurrently active
window
inexecutingacommand.
But,we
decidedtohavethe globalmenu
bar
anyway
because people generally agreedthat thattheadvantage, together with otherssuchas
more
efficientimplementation,outweighthe objection."We
canprovidemore
structuretothishistoricalrecord, as
we
discussinthe rest of thepaper.Such
structureisusually designedtomake
explicitthe logical structure(e.g.,anargument
supportsaproposal)and/or
the historical structure(e.g.,aproposalreplacesanotherproposal).
Another
meaning
ofdesignrationaleisthesetofpsychological claimsembodied by
anartifactinthesenseof [Carroll
&
Kellogg1989], i.e."theclaimsthatwould have
tobetrueifthe artifactistobesuccessful"or"theclaimsaboutthepsychological consequencesforthe user" [Carroll&
Rosson
1990]. These claimsare differentfromthe historical record; theclaimsneed notbe presentinthe historical record;eveniftheywere, theywould
havetobeextractedfromtherecord
and
formulatedina testableform. Forexample,thedesignrationale inthissensewould
besomethinglike:"Theglobalmenu
makes
theenvironment
easier tousebecauseitreducesscreenclutter."or
"Dimming
the irrelevantitemsintheglobalmenu
makes
iteasier tolearnaboutthe
commands."
The
thirdmeaning
ofdesignrationaleisone
usedby [MacLean
etal.1989],namely
how
agivenartifactislocated inthespaceofpossibledesignalternatives:
what
are theotherpossible alternatives,how
are these alternatives related,what
are the tradeoffsamong
them?
InourMacintosh
example, thedesignrationale in thissensewould
besome
descriptionof thelogicallypossible alternatives for placing
window
commands,
how
theyare related,and what
the tradeoffsare. Itisoftendifficult toprovidesuchadescriptioninasystematic way, but an
* Since theunit ofanartifactisoftenambiguous(e.g.,Macintosh,MacintoshOS,Macintosh
window
system, Macintoshmenubar,or the positionofthemenubar),
we
willusethetermartifactin theComparative
AnalysisofDesign Rationale Representations 7example
isfoundin[Cardetai.1990;McKinlay
etal.1990],which
providesavocabularyof theprimitives
and
a set ofcomposition operatorsfordescribingthedesign spaceofpossibleinputdevices. This
meaning
ofdesignrationaleseems
differentfrom
the firstmeaning
in itsemphasis
on
designrationalenotbeingarecordbuta construction,and
fromthesecondinitsemphasis not
on
a particularartifactbuton
the relationamong
possible alternatives.Models
of
Design
Rationale
We
now
developa series ofprogressivelyrichermodelsofdesignrationale,which
provideaframework
inwhich
we
canplace the three differentmeanings
ofdesignrationale.The
firsttwo meanings
arediscussed immediately below.The
thirdmeaning
ofdesignrationale,asapossibilityspace,isdiscussedshortlyafter.
Designrationaleinthe
most
generalsenseisan explanationofwhy
anartifactisdesignedtheway
itis. Soinourfirstmodel
ofdesignrationale,anartifactisassociatedwithabody
ofreasonsas
shown
in Fig. 2a.There are differentkindsofreasons that
we
can giveforan
artifact.The
reasonscanbehistorical orlogical,roughlycorrespondingtothe
meanings
ofdesignrationale, respectively, asa historicalrecordand
as asetofclaimsembodied by
anartifact.^The
recordoftheprocessthatledtothechoiceofanartifacttellsus
one
kindofreasonwhy
that artifactwas
chosen.The
free textexample
aboutthewindow commands
aboveisan example. Ifwe
wanted
a logicaljustification,
we
would
haveto extractitfromthe record;butatleastsucharecordtellsusthehistoricalcircumstances
and
sequencethat ledtothe design,and
providesa basisfromwhich
to inferthe logical reasons.On
theother hand,we
can representthe logicalreasonsdirectly,i.e.thereasonsjustifyingthechoiceof
an
artifactno
matterhow
or inwhat
order theywere
articulated.
The
setofclaimsembodied by
anartifactisanexample
becausetheseclaimswould
justifythedesignoftheartifact. These claimsare logical also in thesensethat thecontextin
which
theyaretruehavetobemade
explicit. For example,aclaim should not say "Theglobalmenu
isbetterbecauseitleads to smallerimplementation" ifitreallymeans
"TheWe
believethatthe distinctionbetweenhistoricalandlogicalreasons breaksifwe
pushittoofarbecauseapurelyhistoricalrecord persedoes notreallygiveusareason. Itwouldgive usareason onlytotheextentthat
we
canextractsomelogicalstructure outofit. Nevertheless,we
believethatthe distinctionis useful for the purposeofevaluating representationsbecausefor a given representation
we
wouldliketoknow
whatitmakesexplicitand whatwe
haveto inferfromit.(a)
MODEL
I:Anartifactisassoaated withabody ofalltheargumentsrelevant to thedesignof the artifact(b)
MODEL
2:Alternativesandtheirrelationsaremadeexplicitandthearguments aboutindividualahematives can bedifferentiated.Ct^'^
''^'k^lr^'2.^*%a
CTi)
"^^3
Figure 2(a)-(c). Progressively
more
differentiatedmodels
ofdesignrationaleComparative
AnalysisofDesign Rationale Representations^'ne,^
(d)
MODEL
4:Criteriausedforevaluationsandtheirrelationsaremadeexplicitandthearguments aboutthem can befurther differentiated in theargumentspace.
(e)
MODEL
5;Individualissuesaremadeexplicit,eachofwhichcontains the alternatives, evaluations,andcriteriausedindiscussing the issue.Apart of theargument spaceincludes themeta-arguments aboutthe issuesandtheir relations.
Figure 2(d)-(e).Progressively
more
differentiatedmodels
ofdesignrationale(continuedfrom previouspage).
global
menu
is better in the context of theMacintosh
because it leads to smaller implementation. This is important in asystem
like theMacintosh
which
has a smallmemory."
Extracting these logicalreasonsisnotan easytask;onceidentified,however, they10
The
internalstructure of thesereasons can bemade
explicittodifferentdegrees.At one extreme, theycan be completelyundifferentiated.An
example
isthe naturallanguagedescription thatwe
gaveearlier asan
example
of a historical record. Ifwe
were
tomake
the historicalrelations
more
explicit,we
candifferentiatefurtherbymaking
explicitsuchrolesand
relationssuchas:Initiator,Second-Motioner,Initiates,
and
Replaces.An
example
thatisnothistoricalistherepresentationused by[Carroll
&
Rosson1991] fordescribing theclaims"embodied"inanartifact. Inthisrepresentation, theclaimsthemselvesarerepresentedinnaturallanguage, but
theclaimsare
grouped
bythequestions theyanswer
to:what
canIdo,how
doesthatwork,and
how
do
Ido
this?We
canalsoimaginearepresentationwhere
the logicalsupportrelationscan bemade
more
explicit by providing such constructsasLogically-Implies,Supports,Denies,Qualifies,
and
Presupposes.We
willusetheterm.Argument
Space,torefertowhat
we
havecalled a
body
of reasons,becausethereasonsarecapturedeither as a historicalrecordof thevarious
arguments
relevant for thedesignoftheartifactorlogicalarguments
underlyingthe design.Thereis
much
we
cando
with ourfirstmodel
ofdesignrationale.A
representationbasedon
thismodel
can help usanswer
thequestion,"What
didwe
discusslastweek and what do
we
needtodo
today?" Sucha representationcanalsohelp usanswer
the questions:How
didotherpeopledealwiththisproblem?
Can
we
learnfromthe pastdesigncases? Carrolland
Rosson[Carroll&
Rosson
1991]provideagood
example.They
report in detailhow
theirrepresentationofdesignrationale,mentioned above, suggested
many
issues forredesigninganartifact(theView
MatcherinSmalltalk).They
discusshow
these issuescan be couchedas adesignhypothesis,which
can betestedand
compiledtoform,inthelongrun,"acontextualizedscienceoutofpractice"of
human
computer
interaction.Our
firstmodel,however, doesnothelpverymuch
withtheother questions,thoughwe
shouldqualifythisstatementimmediately. Sayingthatitdoes not help
much
isnottosaythatwe
cannot
answer
these questions.Of
course,iftheuserworks
hard enough,and
aslongasthe representationbasedon
themodel
hasenough
informationcaptured,evenintheformof naturallanguagefreetext,
we
cananswer
these questions. Sothe realissueishow much
themodel
itselfhelp us
answer
thesequestionseitherby helpingus see the structure better orby enablingustodefinecomputationalservices thathelp us
answer
thequestions.We
willseehow
more
differentiated
models
allow ustoanswer
thesequestionsmore
easily,although theyincrease the cost insome
otherways
[Conklin&
Yakemovic
1991].Comparative
AnalysisofDesign Rationale Representations 11Our
secondmodel
(Fig.2b) differsfromthefirstbymaking
multiplealternativesand
theirrelations explicit. Design involves formulatingseveral alternatives,
comparing
them,and
merging
them,asmany
of thequestionsinourlistindicate.Inourfirstmodel, onlya singlealternativeis
made
explicit at agiventime,and
themultiplealternativesarepresentonlyimplicitlyinthe
argument
space.Our
secondmodel
make
these alternativesexplicit,includingtheonesthathave beenrejected.
Once
the alternativesbecome
explicit,we
cantalkabouttheir attributes(e.g.currentstatussuchas "rejected" or'waiting for
more
information'),make
the relations
among
thealternatives explicit(e.g. specialization, historical precedence), define computational operationson
them
(e.g.comparing
alternatives, displaying thealternatives that specializethisalternative),oreven argue about
whether
analternativeisworth
considering.The
alternatives,otherthantheonefinallychosen,are interestingbecausemany
of the issuesand
theknowledge
usedinevaluatingthem
are useful in other contexts, forexample,
when
situational constraintschange.We
usetheterm. AlternativeSpace,torefer tothissetofmultiplealternatives
and
theirrelations.Theserelations
among
the alternativescanalsobehistorical orlogical. Historical relationsmay
be not onlythe linearsequencethatwe
usually describe as versions,butalsomore complex
relationssuchas layers
and
contexts[Bobrow
&
Goldstein1980].The
logical relationsmay
include Specializes,Generalizes,Elaborates, or Simplifies.
Or
alternatives can be relatedthroughadesignspace [McKinlayet al.1990].
To
the extent thatwe
want
a representationtorepresentthese different alternatives
and
their relations,we
saythat the alternativespaceiswithinthescope ofthe representation. glBlS, forexample, seemstoincludethe alternative
space withinitsscope because oneofitsgoalsis"tocapturealternative resolutions (including
those
which
arelaterrejected),[and] trade-off analysisamong
these alternatives"[Conklin&
Begeman
1988].The
constructsinglBlSforrepresentingthe alternative spaceconsistof:Position,with
which
we
can describe multiplealternatives,and
the specialization relationamong
the Positions.By now,
we
have analternativespace connectedtotheargument
space, asshovwiin Fig.2b. Foreachofthe alternatives, there arearguments describingthereasonsforitscurrentevaluation,
justas inourfirst
model
there arearguments describingtheevaluationstatus of that single alternative,i.e.,thatitwas
chosen.Some
ofthearguments
can be shared;forexample, anargument
can support analternativewhiledenying
another; soitisbetter tothinkofthearguments aboutthe different alternativesforminga single large
argument
space, asshown
inOnce
the alternativespaceisrepresented,we
canimaginehow we
canmake
asystemthathelpus
answer
some
ofthequestionsinourlist.To answer "What
are the alternativedesignsand
what
are their prosand
cons?,"we
canassociate anargument
space with each ofthealternativesthroughthe linkssuchasSupportsor Objects-To, asingIBIS.
To answer
"Why
do
we
even considerthis alternative,and
how
isitrelated totheone
thatwe
discussedlastweek?"
we
needtosome
historical relations,suchas Rqjlaces,among
the alternatives.With
therepresentation of the alternative space,
we
canimaginehow
we
canmake
asystem help usanswer
some
ofthequestionsposedinSection2.To answer "What
are the alternativedesignsand
what
are theirprosand
cons?,"we
canassociateanargument
spacewith eachofthealternativesthroughthe linkssuchasSupports orObjects-To, asingIBIS.
To answer
"Why
do
we
even considerthisalternative,and
how
isitrelatedto theone
thatwe
discussedlastweek?"
we
needtousesome
historical relations(e.g.Replaces)or structural relations(e.g.Is-A-Part-Of),
among
the alternatives.However,
oncewe make
explicitmultiplealternatives,we
needtoarticulatemore
carefullywhat
theargument
spaceisabout(Fig. 2b). Inourfirstmodel
when
we
had
a singleartifact,namely
thechosenone, theargument
space contained reasonsforthechoiceof thatartifact.Similarly,the
arguments
fortheotheralternativesareaboutwhy
theywere
not chosen,or,to generalize,why
theyhave
theirparticularevaluationstatus, e.g."Still inConsideration", "Waiting forMore
Information","Rejected".These
evaluation status could benominal
categories(suchas theabove examples),ordinal categories(suchas"Very
Good,"
"Good,"and
"Poor") oracontinuousmeasure
(suchastheprobability that the alternative willachieveagivensetof goals).
Therefore,
we
introducetheevaluationspace(Fig. 2c),where
theevaluationstatus aremade
explicit
and
related. Usually,we
do
notand
need notspecifyany
elaborate relationamong
theevaluationmeasures
we
use. Often,the implicitordinalrelationamong
thesevalues(e.g."Very Good," "Good," "Poor," "Very Poor")issufficient
when
we
leaveitforthehuman
usertoassignthesevaluestothe alternatives.
However,
ifwe
want
todefineany
computationalservice that
manages
these values, forexample
thatautomatically propagatesand merges
them
toproduceahigherlevelsummary,
thenwe
needtobe verycarefulaboutwhat
thesevaluesmean.
We
needtospecifythe units ofmeasurement,acalculus forcombining them,and
amodel
specifyingwhat
theymean.Even
inthe casewhere
these actions arelefttohuman,
forexample,ifthe
human
userisexpectedtocombine
thesevaluestoproduce
ahigherlevelsummary
measure, thenwe
need
to setdown
what
these valuesmean
so that theirComparative
AnalysisofDesign Rationale Representations 13Making
theevaluation spaceexplicitallows ustodifferentiatetwo components
of theargument
space:(1)
arguments
aboutwhy
an
alternativehas its currentevaluationstatus,and
(2)arguments aboutthe alternativesthemselves,e.g.,
why we
shouldorshould noteven consideran
object asan
alternativeorwhether
this alternativeis reallyaspecialcaseofanotheralternative. That is,as
shown
in Fig. 2c, these different kinds ofarguments
can bedifferentiated in the
argument
space.With
therepresentation of theevaluationspace,we
cannow
answer
questionssuchas:"What
are thetwo most
favorable alternatives sofar?"and
"SunMicrosystems just released their
X/NeWs
server.How
would
the releasechange
ourevaluations?"
We
canalso explainhow
an evaluationwas
made
by pointingtotheargumentsinthe
argument
space behind the artifact in question,and
explaininghow
this particularevaluationmeasureisderivedor
computed
fromthem
or relatedtoothermeasures.Our
modelssofardo
notmake
explicit thecriteriausedinproducinganevaluation.However,
thecriteriausedfor theevaluation
and
theirrelations are usually quiteimportanttorepresentexplicitly. For example,itisimportant to
know
that theargument
"We
do
notneed
toduplicate
menu
items"isapro-argumentforthe alternative"GlobalMenu
atthetopoftheScreen"becauseo/thegoal ofreducingscreenclutter,
which
isusedas a criterion for evaluation.By making
this criterionexplicit,we
cangroup
alltheargumentsthatappealto thiscriterionand weigh them
againstone
another. Ifthe criterionchangesorbecomes
lessimportant, thenwe
cando
appropriatethingsto allthearguments
thatpresupposethisgoal(forexample,making
theseargumentslessimportant).Knowing
how
thiscriterionisrelatedtoothers(e.g."Reducing ScreenClutter"isa
way
ofachieving "EasytoUse"), alsoallows ustoassignproper importancetothiscriterionorchangeitsimportance,when
the relatedcriteriachanges.We
usethetermcriteriaspacetorefertothesecriteria
and
their relations.As
Fig.2d shows, oncewe
have
the criteriaspace explicit,we
can further differentiate theargument
spaceby
groupingthosearguments
which
areaboutthecriteriaand
their relations.Hence,itisimportantthatalanguage
whose
scope includesthecriteriaspace representthedifferent attributes of thecriteria
and
the relationshipamong
them. For example,itshould allow ustorepresenttheimportanceofthesecriteriaand
the synergistic or tradeoff relationsamong
them.A
setofcriteriacan besub-criteria ofanotherinthesensethat satisfyingthem
facilitatesthe satisfaction of thelatter. Thesesub-criteriacan berelated
among
themselvesinvarious ways.
They
can be mutually exclusive inthe sensethat satisfyingone makes
itimpossibletosatisfyothers.
They
can be independentofeachotherinthesensethat satisfyingtheirparentcriterion invarious
ways
as well.They
can be exhaustiveinthesensethat satisfyingallofthem
isequivalenttosatisfyingthe parent.With
thecriteriaspacerepresented,we
cannow
seehow
thesystemmight beabletohelp usanswer
thequestionssuchas:what
ifwe
do
not considerportability? orwhy
isportabilityimportant
anyway?
The answer might
be"Ifwe
giveup
thegoalof portability,thentheevaluationof thealternative
X
changesto"High" becausealltheseclaimsthatargueagainstX were
basedon
theimportanceof portability."or"Portabilityisimportant becauseitisasubgoalofanother importantgoal,
Have
awide
distribution." Theseanswers can be derivedfrom
arepresentationiftherepresentationmakes
explicitthe relationbetween
evaluations,criteria,
and
arguments.Of
course, representation of thecriteriaisnotsufficient foransweringthese questions. It isnotobvious
how
thesequestionscan be answered evenifsome
parts ofthecriteriaspacearerepresentedexplicitly.
However,
the explicitrepresentationof thecriteriaspaceseemsanecessaryconditionif
we
aretoanswerthese questions. Atleast,we
would
havetheinformation necessarytodefinean operationthat willgiveorsuggesttheanswersto these questions.
We
willgivesome
examplesofsuchoperationslaterinSection5.Sofar,
we
have
identifiedand
discussedthe structure of a singledecision underlyingan
artifact,
namely which
ofthe alternativedesignsshouldwe
choose?However,
withtherepresentation ofsuchlocalstructures alone,
namely
itsargument
space, alternativespace,and
criteriaspace,
we
stillcannot asksome
ofthequestionsin thelistsuchas:What
are theunresolved issues
and
what
arewe
currentlydoing aboutthem?
What
are the issues thatdepend
on
thisissue?To answer
suchquestions,we
needamore
global picture ofhow
individual issues are related.
A
decision often requiresand
/orinfluencesmany
other decisions.Forexample,adecisioncan beasub-decisionofanotherifthelatterrequires
making
thefirstdecision.
A
decisioncan bea specialization ofanotherifthefirstdecisionisamore
detailed case of thesecond. Itisimportanttocapturehow
these decisions are related,and
we
usethetermIssueSpacetorefer tothem.
A
unitin thisissuespaceis,therefore,a single decision thathasasitsinternal structure theother spaces, as
shown
in Fig. 2e.Once
we
have anissue asanexplicitelement,
we
canassociate the attributessuchas "Status"and
"ActionsTaken"
withissues
and answer
questionssuchas"What
are theunresolvedissuesand what
arewe
currentlydoing about them?" Representingthe
dependency
relationamong
the issues willallow ustoanswer
thequestion"What
arethe issues thatdepend
thisissue?"Therearestill
some
questionsthatwe
have notyetcovered suchas:How
didotherpeopledealComparative
AnalysisofDesign Rationale Representations 15spacessofaridentified,thespacesofarguments,alternatives,evaluations,criteria,
and
issue,cancontain
enough
informationtoanswerthese questions.InSection5,we
discusscomputational operationsthathelp usanswer
theseadditionalquestionsby
exploiting the structure of thefivespaces.
4.
EXISTING
REPRESENTATIONS
Thereareonlyahandful ofrepresentationsthathave been used forrepresenting design
rationale[Toulmin1958;
Kunz
&
Rittel1970;Marshall1987;McCall1987;Conklin&
Begeman
1988; Potts
&
Bruns 1988;Lee 1990a;MacLean
etal.1991]. Most,though notall,ofthese representations areheavily influencedbythe IBISstructure forrepresentingissues[Kunz
&
Rittel1970]or
by
Toulmin'smodel
ofargumentation[Toulmin1958].[MacLean
etal.89;MacLean
91] alsoproposea representationbasedon
theirexperienceswith constructing designrationales. Inthenextthreesubsections(4.1-4.3),
we
discuss these three representationsindetailby definingtheirscope
and
evaluatingthem
withintheframework
outlined above. In thenext subsection (4.4),we
discuss a few other studies thatbearon
designrationale representation. InSection5,we
presentalanguage,calledDRL, which
we
believeovercomes
many
ofthe limitations of the existinglanguagesthatwe
discussin thissection.Some
qualificationsareinorderbeforewe
proceed.Our
intentionisnotto crihcize the existingrepresentations,but onlytoevaluate
them
asadesignrationalerepresentationlanguage.Even
then,our evaluation is
mainly
with respect to theirexpressiveness.We
would
liketoemphasize
againthattheadequacyof a representationcan be evaluated only withrespecttoaset of tasks.
One
representationmay
bemuch
more
expressivethananother; as a resultwe may
beableto
do more
withit.However,
ifwhat
itenablesisnotinthesetof desirable tasks,orifitenables thoseonly
by
sacrificingothermore
importantconstraints,thentheadditional expressivenessisnotworthwhat
itgivesus.What
we
discussbelow
isonlywhat
the existing representationsallowordo
not allow ustodo;We
do
notintendtomake
value judgments about whethertherepresentationsshouldorshould notdo.Some
oftheselanguageswere
designed withdifferent goals,which
inturndeterminethe tradeoffsadoptedinthe designs. Therefore,thefollowingdiscussionshould not be construedas a criticism of these representations,but only
asan articulation oftheir scope
and
theiradequacy
asdesign rationalerepresentation languageswithrespecttoexpressiveness.We
alsowant
tomake
clearthatwhenever
we
saythata representationcannot expresssome
information, it does notmean
that people cannot infer that informationfrom
theComparative
AnalysisofDesign Rationale Representations 16representation.
Take
a naturallanguagerepresentation ofdesignrationale thatwe
gaveasanexample
inthebeginningofthepaper. Ifwe
keepa detailedenough
recordofwhat
happened,orevenavideo recordingof the
whole
designprocess,we
canalwaysretrievetheinformationthathasbeen ever expressedif
we
work
hardenough.When we
saythat arepresentation cannot expresssome
information,we
mean
that therepresentationdoes not provideconstructsthat
make
theinformationexplicit insuchaway
thathelp peoplesee thestructure betterorthat
makes
itamenable
tocomputational manipulation.4.1.
IBIS
(IssueBased
Information
System)
The
IBISstructurewas
originally developedin[Kunz
&
Rittel 1970] forthepurpose
ofrepresentingdesigners'argumentationactivities. Sincethen,
some
variationsofithave been usedby
afew systemsforrepresentingdesignrationale.One
variationistherepresentationusedby gIBIS [Conklin
&
Begeman
1988],"ahypertexttool forexploratorypolicy discussion. SincegIBISismost well-known
[Conklin&
Yakemovic
1991]and
can be regardedas amodem
incarnation of theoriginal IBIS ,we
use gIBIS as the contextfor discussing theIBISrepresentation.^ Othervariationsinclude
PHI
(ProceduralHierarchyof Issues)[McCall1987]and
theone
usedby[Potts&
Bruns1988] for the rationalemodule
in theirrepresentation.We
discuss
them
briefly inSection4.4.The
goal ofgIBISistocapture"thedesignproblems,alternative resolutions, tradeoff analysisamong
these alternatives,and
the tentativeand
firmcommitments
thatwere
made
in theprocessof thedecision
making
process". Figure3shows
the objectsand
relations thatformthelanguageofgIBIS
and
Figure4shows
anexample
representation. IngIBIS,one
raisesanIssuesuchas
where
toputthewindow
commands.
Positionsare created to Responds-to the issue(e.g.,"intheglobal
menubar
atthetop of thescreen.",and
"Commands
atthetopof eachwindow.")Arguments can becreatedtoSupportor Object-to a Position. For example,the
argument
"Don't needtoduplicatecommands
foreachwindow
"supportsthefirstposition,and
theargument
"Requireslong
mouse
travel."objects toit.Also,anIssuecanberelated to other objects asshown
in the figure.
The
gIBISmodel
extends the original IBISmodel by
introducing: the generalize/specializerelationamong
Arguments
aswellasamong
Positions,an
additionalWe
believethatthe distinctionbetweenhistoricalandlogicalreasonsbreaksifwe
pushittoofarbecauseapurelyhistoricalrecordpersedoes notreallygive usareason. Itwouldgiveusareason onlytotheextentthat
we
canextractsomelogicalstructureoutofit. Nevertheless,we
believethatthedistinction is useful for thepurposeof evaluating representationsbecausefor a given representation
we
wouldliketoknow
whatitmakesexplicitand whatwe
havetoinferfromit.Comparative
AnalysisofDesign Rationale Representations 17C_J_^>
Figure3.
The
glBIS vocabulary.f Commandsatth^lCommandsatthd
Itopofthewindow.IlsideotthewindowI
C ^"-/"•" responds-to
^ ^ 'I commanasrea
(^N°J—
responds-toV
jistjacting?_J
Dimmedcommandsstilldistracting
bytheirmerepresence.
Other type
which
serveasan"escape"mechanism
forrepresentingthingsnot representable withthe available constructs,and
anExternaltypeforexternal objectssuchasdocuments
orsketches. Inthefollowing,ourdiscussion referstothestructure
common
tobothIBISand
gIBIS exceptwhen
we
discuss the generalize/specialize relation for Positionand
Argument.The
scopeof thegIBISrepresentationdepends
on what
anissueis. Ifwe
takeanissueinavery general sense tomean
any
question that takes a set of Positions, then theIssue/Position/Argument structurecan representa fairlylargepart of thedesignrationale spaces.
The
internalstructure ofthese spaces,however,isnot welldifferentiated in gIBIS, aswe
discussbelow.The
alternativespaceisrepresentedingIBISby
Positionsand
the relationamong
them. SincemultiplePositionscan becreated for agivenissue,gIBIS allowstherepresentation ofmultiple
alternatives,thusoffering at least the richness ofour
model
2.The
onlyrelation,however,among
thePositions thatwe
can representingIBISistheSpecializes/Generalizes relation,althoughthereareotherrelations thatcan connectaPositiontoobjectsofother types(e.g.
Questions or ls-Suggested-By)7 Thus, gIBIScan
answer
thequestion"What
are theotheralternativesbeing considered?" but notthequestion
"How
arethese alternatives related?"unlessthey
happened
toberelated viathe Specializes relation.The
unit oftheargumentspaceingIBISisArgument.An
Argument
(e.g."Thecommands
notapplicabletothecurrent
window
would
beshown and
distracting.")can supportor objecttoaPosition(e.g."Intheglobal
menubar
at thetopof the screen"),indicating areasonwhy
thepositionisorisnota
good
one.Thereisno
construct,however,forqualifyingan argument. For example,we
cannotindicate thatthe existingArgument,"theglobalmenu
reducesthescreenclutter",isvalidonlyifthe
Argument
"theglobalmenu
containingallpossible actionitemsdoes not
become
so bigitself"isvalid.The
relationamong
thearguments
isagainlimitedtothe Specializes relation. Thislimitationimplies that
an Argumentcannot
Supportor Object-toanotherArgument
directly. For example,the
Argument
"Irrelevantmenus
can bedimmed"
isanargument
objectingtothe originalArgument,
"Thecommands
not applicable to the currentwindow
would
beshown
and
distracting."
To
expressthisrelationshipamong
ArgumentsingIBIS,we
havetocreateanIssueabout
whether
the originalArgument
isrightornot,create Positions "yes"and
"no",and
thenComparative
AnalysisofDesignRationaleRepresentations 19argue aboutthese Positions.^
A
possible criticism ofthisrepresentationisthatitleadstothe proliferation ofIssues. For example,inordertojustpointoutthe factualinaccuracyof a claim,we
would have
tocreate threenew
objects: anew
Issue,a"No"Position,and
tinArgument
supportingthe
"No"
Position.However,
we
might beabletoreducethecomplexitywithaninterface thathidestheintermediatedetails.
A
more
serious limitation ofthisrepresentationisthatit does not help us
answer
questionssuchas"Show
me
alltheclaimsthatrespondto thisclaim."We
might
trytoanswer
thesequestionsby
followingtheIs-Suggested-Bylink thatconnectsthe originalclaimand
the issues thatcontaintherespondingclaims.However,
theIs-Suggested-By link istoogeneral for that
purpose
because itdoes not allow usto distinguish the issues that containrespondingarguments fromthose thatdo
not.IngIBIS
we
cannotargue aboutrelationalclaims. For example,thereisnoway
ofsayingthatwe
agreewithA
and
B,but notthatA
SupportsB
becausea linkingIBISisnotanobject thatyou
canargueabout.To
illustratewith another example,supposewe
want
toargue aboutwhether
portabilityshould bea criterion forevaluatingtheposiHonsforagivenissue.The
onlyway
we
cando
soinIBISisbycreating anew
Issue,say"Shouldwe
considerportabilityasacriteria?",
and
suggestpossibleanswersintheformof positions,say"Yes,we
should" or"No,we
shouldnot". Then,we
canargueaboutthese positions.We
canalso relatethisnew
issuetothe originalissue
by
anIs-Suggested-By relation.However,
thisrepresentationagaindoes notmake
explicitthe relationbetweenthenew
issueand
thecriteriabeing questioned. Ifwe
make
a relationanexplicit object,then
we
canarguedirectlyaboutthe relation, Is-a-Criteria-for,between
the object portabilityand
the original issue.Beingabletoargue aboutrelationalclaimsisimportantforotherreasonsas well. People often
want
tosaythattheymay
agreewithA
and
B but not withtheclaimthatA
supportsB. For example,one
may
agreethat the globalmenubar
isabad
ideaand
thatseeingirrelevantcommands
isdistractingbut notthatthesecond claim supportsthefirst. Thatis,one
may
believe thatthe global
menubar
isabad
ideabut not becauseitdistracts theuserby showing
irrelevant
commands;
hemay
point outthatthe globalmenubar
does nothave
toshow
theirrelevant
commands.
The
usermay
want
tomake
the distinction clearbecause theydo
notwant
tobe construedasdenyingeither of theclaimsalone.The
distinctionisevenmore
crucialwhen
therepresentationistobeusedtoprovide computationalservices. Mixingdenial ofnode
Anotheralternativeistoshowthefirstargumentassupportingthepositionthatthe secondobjects to,instead of objectingtothesecondargumentdirectlybecausean argumentcannotrespondto
anotherargumentingIBIS. This representation does notmakeexplicitthe relationbetweenthetwo arguments. Probably forthatreason, gIBISdoesnotencouragethisrepresentation (Conklin, personalcommunication)
20
withdenial of linkjustisnotrightsemantically. If
you
define acomputationalservicebasedonthistypeof carelesssemantics,suchasevaluation
management
thatpropagatesand
mergesevaluations,
what you
getwould
notmake
much
sense.The
evaluationspaceusedby
gIBISconsistsofsome
nominalcategoriessuchas"Rejected"and
"Chosen"
assignedtothe Positions.We
could usefinercategories,suchas"WaitingforMore
Information," butbeyond
thatitisnotclearwhat more
sophisticatedevaluationmeasures canbe defined
on
thegIBISstructure.We
couldassignuncertaintymeasurestothearguments,and
try toderive
some
ordinalrankingamong
thepositionsbasedon
supportingor objectingarguments.
However,
itisdifficult toimaginewhat
thecalculuswould
belike.The
criteriaspaceisbeyond
thescopeof gIBIS. gIBISprovidesno
vocabularyfordescribingthecriteria
we
useinevaluatingthe alternatives. Thislack of explicitrepresentation ofcriteriaingIBISisa serious limitation for adesignrationalerepresentationlanguage. Sincecriteriaare
notexplicit,
we
cannotargue about them;we
cannot representthereasonsforhavingthesecriteria; nor can
we
indicateany
relationship,suchasmutuallyexclusiveness,among
thecriteria. Again,
we
couldindirectlyrepresent these relationsby
creating additionalIssues. Forexample.Fig.4
shows an
Issue("Should smallimplementation bea goal?") relatedtoouroriginal Issue viaan Is-Suggested-by link. But, again, this representationdoes not
make
explicitthe relation
between
thisissueand
the criterion in question. Itwillbedifficultforpeopletoseethisrelationship;itwillbe even
more
difficultfor amachine.Lackof explicitrepresentation of thecriteriaentailsother limitations. Forexample,
when
goalschange,thereis
no
easyway
toaccommodate
the changes. It ismore
difficult to isolate the realdisagreements
among
people becausethecriteriathey useintheirarguments remainimplicit.As
we
will see in thenextsection,explicit representationofgoals can providemodular
representation ofarguments,multipleviewpoints,
and
a basis forrelevance matching.InglBlS,theunit of the issuespaceisanIssue,
and
gIBIS providesseveralconstructs fordescribingthe relationsaboutissues.
An
IssuecanGeneralize, Specialize,Replace,Question,and
Suggestanother Issue. Figure4shows
that the original Issueaboutthe placementofwindow commands
suggests other Issues:whether
thecommand
pop-up
menu
should beconsideredasanalternative
and
whetherwe
shouldhavesmallimplementationas a goal.An
IssuecanalsoQuestiona PositionoranArgument; anIssuecan be Suggested-bya Position orby an
Argument.
Theserelationsarequiteimportantindescribingamore
global relationshipComparative
AnalysisofDesign Rationale Representations 21differentiated, the relationssuchasReplace
and
Specializeseem
essential. Itwould
benice,however,if
we
cansomehow
show
whetheragivensetof relationsiscompleteoradequate.4.2.
Toulmin's
Model
of
Argumentation
Stephen Toulmin,a philosopher,proposeda
model
ofargument
in1958. Since then, themodel
hasbeen adoptedfor
many
purposes, includingthecomputationalrepresentation ofarguments[Bimbaum
et al.1980;Lowe
1986;Marshall1987; Streitzet al.1989]. Figure 6shows
anexample
Toulmin
representation ofan argument.^CLAIM
commands neednot be duplicated for eachwindowX
usually Placingwindow commandsintheglobalmenureduces
screenclutter.
No
duplicationsavesspace.
unless
5
on-account-oftlieglobalmenu
containingallpossible
commandsbecomes
so big
Common
Sense KnowledgeFigure5.
An
example
ofanargument
inToulmin'srepresentation.A
Claimisthemain
assertionbeingmade
(e.g.,"Placingthewindow commands
intheglobalmenu
reducesthescreenclutter.")We
supporttheClaimby
producingaDatum,
"Commands
need
notbe duplicatedforeach window."^^We
justify the leapfrom
theDatum
to theClaimby
producingaWarrant,e.g. "Ifyou
need notduplicate things, thatsavesspace."A
Warrantisusuallyaninference ruleor aprinciple that licenses the inferencefrom
Datum
toClaim.A
Warrantisnotoften absolute; soifwe
want
toexpresshow
stronglythisWarrantsupportstheinference
from
theDatum
totheClaim,we
usea Qualifier. Inour example,the QualifierisWe
will usetheterm Toulmin's representation rather than Toulmin's modelbecausewe
areinterested inthemodeltotheextentithasbeenorcan be usedas arepresentation.
Aspointedoutin
[Newman
&
Marshall90],Toulmin wantsthedatumtobeasingularfact,suchassomethingthatyoucan pointto,because
we
thinkhestillwantstoretainthesyllogisticform,where onegetstheconclusionbyhavingastheminorpremise,asingularfact.Sohe seemsshllundertheboundof syllogism,atleast atthetimeof his writing,despitehis rebellionagainstit.Thedatumofour
exampleis amoregeneral statement. But
we
usethisexamplefirstbecause thedistinctionitselfbetweenasingularfactandauniversalstatementhasbeencalled intoquestioninphilosophyand
22
"usually",
which
qualifies the strength of the inference.We
indicate the excepting circumstanceswhere
theWarrantfailsthroughRebuttal(e.g.,"Theobject tobesharedmight havetobelargerthanthesum
ofthese objects locally existing.') IfaWarrantiscalled intoquestion,then
we
appealtoaBacking,i.e.,abody
ofknowledge
thattelluswhy
ingeneralwe
shouldaccept thewarrant. Inour example,
we
appealtoourcommonsense
knowledge. Othertypesofbackingmight bephysical laws, historicalfacts,orlegalprecedents.
The
scopeof theToulmin
structure as adesignrationalerepresentationislimitedtowhat
we
called the
argument
space.The
goal ofToulmin
structureistomake
explicithow
aclaimissupported. Thereis
no
attempttorepresentalternatives,goals,any
evaluationmeasure,orhow
theyare related.As
such,we
evaluateToulmin'srepresentationonlyasa representationmodule
fortheargument
space,which might
interfacewith representationsfor theotherelementsofdesignrationale.
Even
then,we
should keepinmind
thatthe originalgoalofToulmin
was
to delineate the logicalsupportstructureofan argument, notnecessarilytoprovidea representation tobeused dynamicallyforcapturingdesignrationale.
As
a representation for theargument
spaceofdesignrationale,Toulmin's hasmany
limitations.Forexample,
we
can onlydeny
aClaimby supportingitsnegation.Suppose
you want
todeny
thattheglobal
menu
reducesthescreenclutterbecausetheglobalmenu
hastocontainallpossible
command
items(letusassume
thatthisisbeforepull-downmenus were
invented).The
onlyway
todo
soisbycreating aClaim"theglobalmenu
does not reducethescreen clutter"and
presents "the global
menu
hastocontainallthemenu
items" as theDatum
forthisnew
Claim.Whether denying "A
reducesB"means
thesame
assupporting"A
does not reduceB",itiscertainlyan
awkward way
ofdenyingaClaim;and
inany
case,withoutsome
construct that indicates that thetwo
Claimsare thenegationofeachother,thedenial relationshipwould
not beexplicit atall.Even
withsuchconstruct,suchawkward
representationwould
slowdown
human
comprehensionormachine
computation.Some
people[Bimbaum
etal.1980]have solvedthis
problem by
explicitlynaming
the original linkbetween
Datum and
ClaimasSupportsand
introducinganotherlinkcalledAttacks. So Toulmin's representation should be extendedatleast this
much
toaccommodate
denyingclaims.Itisalsonoteasyto qualifyaclaimusing Toulmin'srepresentation.
We
canqualify aWarrant using Qualifierand
Rebuttal,butifwe
want
to qualify aClaim or aDatum,
thereisno
correspondingconstruct.We
couldofcoursebuild into theClaimany
qualificationso that theClaimitselfisoftheform, "Global
menu
reducesscreen clutter provided that the globalmenu
isComparative
AnalysisofDesignRationaleRepresentations 23purposeof
making
thesecomponents
distinctistoallowustosee the relationsmore
clearly, toformallymanipulatethese
components and
relations,and
toincrementallyconstruct pieces of theargument.Toulmin'srepresentation also suffersfrom buildingintoobjecttypescontext-dependentroles,
suchaswarrantor backing.
The
differencebetween warrantand
backingisnotinthenatureofthe objectitselfbutonlyinthefactthatbackingissomethingthatsupports warrant.
The same
objectcan beaWarrantor aBackingoreven Claimor
Datum, depending on
the context. Soeither
we
assign atypetoanobjectbasedon
the inherent properties of the objectand
indicateitscontext-dependentrolein
some
otherway,aswe
show
inSection5,oranobjecttype should dynamicallychange
depending
onthe context.Most
oftheToulmin-basedrepresentations[Marshall1987; Streitzetal.1989;
Newman &
Marshall1990] take thelatterapproach. For example,ifwe
want
tosupportaDatum
object,we
firstchangeitsrolefromDatum
toClaim,theninstantiatethefive-component schema,
and
fillitin.However,
thenitisnotclearwhy
we
need
aseparatetypecalledBacking. Ifwe
want
tosupport ordeny
aWarrant,why
notchangeitsroletoClaim,instantiatethe
now
four-components(thatis,theToulmin
structurewithout Backing) schema,
and
useaDatum
tobacktheClaim? Itisnotclearwhat
isspecialabout Warrantthat
we
needaseparatetypeforitssupportwhen
we
do
notneedaseparatetypetosupportother thingssuchas aDatum.
We
understandthatToulmin's reasonforintroducingBacking as a separatetypeistoprovideacategoryfor thekindof
knowledge
thatisrarely called into question,suchas the written law,common
senseknowledge,or physical laws.We
agreethatforanargument
toberesolved, theremust
be somethingthattheinvolvedpartiescan agreeand
appealto.However,
we
believethatitismisleadingto representsuch
knowledge
as a distinctcomponent
ofa rigid structuresuchasToulmin's. For
one
thing,thefive-componentschema
leadspeopletobelieve thatanargument
hastomake
explicitallthesecomponents
tobe complete. Thisbelief often forcespeopletounnecessarily representtheobvious;for
example
many
warrants,when
explicitly represented, are often asobviousas theoneinour example, "Noduplicationsaves space". Alsooften in the process,peopleare ledtoargue about whether somethingisawarrantorbacking,
ratherthan aboutthesubstanceof theargument. Conversely,theclosednatureof the
schema
leadspeopletobelieve thatifan
argument
hasthese fivecomponents,itiscomplete. But,havingfive
components
isarbitrary,becausetheremighthavetobe multipleData,multipleWarrants, orchainofWarrants before
we
get tosomethingthatwe
agreeon.We
believe thatitisbettertodefinetherepresentation inan open-ended
way
thatallowsasmany
objectstobe createdasneeded
withoutnecessarilyspecifyingtheboundary
ofanargument and
withoutrequiring theobvioustobe representedunlessitneedstobe,e.g.,unlessitiscalled intoquestion
orarguedabout.
Thereare
numerous
otherlimitationswithToulmin'srepresentation, as[Newman
&
Marshall1990] pointout fromtheirexperiencesofusing Toulmin'sforrepresentinglegalarguments. For example, they
had
tomake
Rebuttalhaveatleastfouradditionalmeanings
torepresentthedifferenttypes of objections,eachcorrespondingto
what
isbeingobjectedto.We
believe that these differenttypesofobjectionsshouldexplicitlyindicatewhat
theyobjectto,and
thatwe
can
do
soina gracefulway,aswe
hope
toshow
inSection5.4.3.
QOC
(Question,
Option,
and
Criteria)QOC
isarepresentationproposed by[MacLean
et al.1989,1991] for "constructing"designrationale.Designrationalein
QOC
issaidtobe notarecordofthedesignprocess,butir\steadisaco-productofdesignthathastobeconstructedalongsidethe artifactitself. Thisemphasis
makes
QOC
differentfrom,saygIBIS,whose
goalat leastincludescapturingthe rationale asitunfolds. I DATA I In globalmenubar allopofscreen Oneach window Positiveevaluation Negativeevaluation DATA: commandsnot applicable
tocurrentwindowwould
Comparative
AnalysisofDesign Rationale Representations 25The
majorconstructs ofQOC
are straightforwardand
map
clearlytotheframework
proposedin thispaper. Figure6shows
anexample
representedinQOC.^^ The
unit of the issuespaceinQOC
isaQuestion.The
unit of the alternativespaceisan Option. Questionsand
Options roughly correspondtoIssuesand
PositionsingIBIS.However,
unlike gIBIS,QOC
canrepresent thecriteria space with Criteria.A
Criteria (e.g.Reduce
ScreenClutter)issaid to bea"bridgingcriteria"if itisa
more
specificone
thatderivesitsjustificationfromamore
generalone
(e.g.Easytouse).The
units ofevaluation spaceare links labelledwith "+"and
"-,"correspondingtowhether an option doesordoes not achieveagivencriterion.
The
constructsforrepresentingthe
argument
spaceareData, Theory,and
Mini-Theory.One
supportstheevaluation ("+" or"-") ofan option withrespecttoa criterion
by
appealingtoempiricalData(e.g.
"The
mouse
isaFitt'sLaw
Device") ortoan accepted Theory(e.g."Fitt'sLaw").When
thereis
no
relevantdataathand
orexistingtheorytodraw
on,thedesignersmay
havetoconstructa Mini-Theory,
which
isan approximate
explanation of partof thedomain.
[MacLean
etal.1991]provideanilluminatingdiscussion of the otherformsof justifications fordesign,suchas variousformsofdependencies
and
metaphors, nospecificconstructs arediscussedforrepresentingthem.
QOC
aswe
understandithasanumber
of limitations as a representationlanguage. First,in theargumentspace,constructs like Data, Theory,orMini-Theory
do
notseem
tocapturemany
aspectsofarguments. For example,it isnotclear
how
an argument
suchas "Irrelevantconnmands
canbedimmed"
should betreatedgiventhatitisneither a piece of empiricaldatanoratheory.
Nor
is itclearwhetherand
how we
can argueabouttheories, or individualclaimsintheories. Inthe alternative space, thereisa referencetocross-option dependency, but
no
specificconstructs arediscussedforrepresentingit. Inthecriteriaspace. Bridging Criterion
seems
to treatedas a fixedtype, ratherthanarolethata criterionplays withrespect toanother criterion.
However,
building a role into a fixed type results in unnecessaryir\flexibility. Thatis,beingaBridgingCriterionisnotaproperty inherentinthe objectitself
butisarelationthatthe objecthas toanothercriterion.
As
such,one
should nothave
toclassify agivenobject asaBridgingCriterionbutinstead indicateitsrolethroughthe relation
ithaswith anothercriterion. Otherwise,
we
have
to unnecessarilychange
object types dependingon
whichobjectwe
focus on.The
QOC
constructs for the evaluation space are"+"and
"-"links. These evaluationsare said tobe supportedby
appealingtoempiricaldataand
acceptedtheories. Itisnotclear,however.1^