northwestern university capturing qualitative science ... · learning (burstein, 1985), and problem...

NORTHWESTERNUNIVERSITY

CapturingQualitativeScienceKnowledgewithMultimodalInstructionalAnalogies

ADISSERTATION

SUBMITTEDTOTHEGRADUATESCHOOL

INPARTIALFULFILLMENTOFTHEREQUIREMENTS

forthedegree

DOCTOROFPHILOSOPHY

FieldofComputerScience

By

MariadelosAngelesChang

EVANSTON,ILLINOIS

September2016

2

©Copyrightby(MariaChang)2016

AllRightsReserved

3ABSTRACT

Thisthesisexploresacommunicationmethodthatisrelevanttolearningbyreadingsystems

andeducationalsoftwaresystems:instructionalanalogy.Itiswidelyrecognizedthatanalogical

reasoningplaysavitalroleinourabilitytodetectsimilaritiesanddifferencesandtotransferknowledge

betweentopics.Buildingsoftwarethatcanunderstandtheseanalogiescanadvancethestateoftheart

inknowledgecaptureandisanimportantsteptowardhuman-levellearningbyreading.However,this

goalinvolvesseveralchallenges.Instructionalanalogiesareexpressedinnaturallanguageandareoften

accompaniedbyrichspatialrepresentations.Successfulinterpretationoftheseanalogiesrequires

substantialbackgroundknowledgeabouteverydayexperiencesandthephysicalworld.Lastly,the

targetknowledgeforinstructionalanalogiesforintroductoryscienceisoftenconceptualandqualitative.

Theclaimsofthisthesisarethatacombinationofsimplifiednaturallanguageunderstanding,sketch

understanding,andstructure-mappingcanbeusedtobuildqualitative,conceptualknowledgefrom

multimodalinstructionalanalogies,andthatsuchknowledgecanbeusedtoanswerquestionsaboutthe

newdomain.Amodelofinstructionalanalogyinterpretationwasdevelopedandtestedonasetof11

multimodalanalogies.Themodelachieves83%accuracyonqueriesoftargetknowledge.Ablation

experimentsindicatethattheuseofbasedomainknowledge,implicitbackgroundknowledge,and

sketchknowledgemakevaryingcontributionstowardaccuracy.Theknowledgeacquiredfromthese

analogieswasusedtoanswer11outof14relatedquestionsonmiddleschoolscienceexams.

Continuedresearchinthisareacanimprovereadingsystemsandopensthedoortonewkindsof

intelligenttutoringsystems.

4ACKNOWLEDGEMENTS

IthanktheNationalScienceFoundationandtheSpatialIntelligenceandLearningCenterfortheir

generousfinancialsupportofmyresearch.

Ithankmythesiscommittee,DedreGentner,ChrisRiesbeck,andMatthewEasterdayfortheir

constructivefeedback.IespeciallythankDedreGentnerwhosupportedmeasifIwasoneofherown

students.

IthankKenForbus,forbeinganoutstandingadvisorandforremindingmethateverythingisharduntil

it’seasy.Thatbattleisalwaysworthit.

IthankNadineGaab,HectorMuñoz-Avila,andYolandaGilfortheirmentorshipandconstructive

feedback.

Ithankmyfellowstudents,researchers,andstaffattheQualitativeReasoningGroupandtheCognitive

SystemsDivision,fortheirfriendship,intellectualinspiration,andsupport:SubuKandaswamy,Jon

Wetzel,DavidBarbella,JasonTaylor,AndrewLovett,MaddyUsher,TomHinrichs,JennStedillie,and

CarrieOst.IthankthenextgenerationofQRGstudents,JoeBlass,ChenLiang,CJMcFate,IrinaRabkina,

andMaxCrouse,forkeepingitfresh.IespeciallythankScottFriedman,MattMcLure,andMark

Cartwrightforbeinggreatfriendstomeand,invaryingways,knowingjustwhattodotomakemydays

better.

Ithankmyfamilyandfriendsfortheirgenuineinterestinmyworkandwell-being.Myparents,Peppina

LianoandJorgeArturoChang,haveshownmeunconditionalsupport.Tomysiblings,Blanca,Todd,

Aurora,Andres,George,andJose,thankyouforremindingmehowcrazyourfamilyisandhowthat’sa

goodthing.Tomysecondfamily,RickandSuePorteus,Karissa,Mike,Jacob,andAshleyReardon,I’mso

5honoredtohaveallofyouinmycorner.MarcoandElliotHidalgo-Greenberger,youknowyoutwoare

whatI’llmissmostaboutChicago.Thankyou,MattO’Connor,forallthepoliticalnonsensethatyou

knowIlove.Ithankmyfriendsbackhome,LeahRubin,JeremiahRomm,andDianeScheiner,for

providingthefriendshipthatIneedevenwhenwe’refaraway.Ithankmydogs,JulepandMarmalade,

foralwayslisteningtomypracticetalks.

Mostofall,Ithankmywife,ShaunaPorteus,forbeingthemostincrediblylovingandsupportivepartner

Icouldhavehopedfor.

6

Formyparents,

PeppinaLiano&JorgeArturoChang

whotrulyembodystrength,courage,andsacrifice.

Formywife,

ShaunaPorteus

wholovesandsupportsmethrougheverything,overandoveragain.

7Contents

ABSTRACT.....................................................................................................................................................3

ACKNOWLEDGEMENTS................................................................................................................................4

ListofTables...............................................................................................................................................10

ListofFigures.............................................................................................................................................11

Chapter1: Introduction........................................................................................................................14

1.1 Motivation..................................................................................................................................15

1.1.1. StructureMappingTheory.................................................................................................18

1.1.2. ImpactinAI.........................................................................................................................23

1.2 ClaimsandContributions...........................................................................................................24

1.3 Preview.......................................................................................................................................27

Chapter2: Background.........................................................................................................................27

2.1 Structure-MappingEngine.........................................................................................................27

2.2 Cyc..............................................................................................................................................28

2.3 QualitativeReasoning................................................................................................................32

2.4 EANLU.........................................................................................................................................34

2.5 CogSketch...................................................................................................................................34

2.6 Companions................................................................................................................................37

Chapter3: CharacterizationofInstructionalAnalogies.......................................................................37

3.1 TheFARGuide............................................................................................................................38

3.2 FocusedAlignmentandInference..............................................................................................38

3.3 SpatialRepresentations..............................................................................................................43

3.4 BackgroundKnowledge..............................................................................................................44

3.5 TargetKnowledge.......................................................................................................................46

Chapter4: InterpretationofMultimodalInstructionalAnalogies.......................................................48

8

4.1 Problem......................................................................................................................................48

4.2 Approach....................................................................................................................................49

4.2.1. BackgroundKnowledge......................................................................................................50

4.2.2. SketchInterpretation.........................................................................................................52

4.2.3. TextInterpretation.............................................................................................................60

4.2.4. MultimodalIntegration......................................................................................................61

4.2.5. CaseExtraction...................................................................................................................65

4.2.6. Type-levelSummarization..................................................................................................66

4.2.7. MappingandInferenceEvaluation....................................................................................67

4.3 Evaluation...................................................................................................................................69

4.3.1. ModelInput........................................................................................................................69

4.3.2. ModelPerformance............................................................................................................72

4.4 Discussion...................................................................................................................................78

Chapter5: InstructionalAnalogiesforQuestionAnswering................................................................79

5.1 Problem......................................................................................................................................80

5.2 Approach....................................................................................................................................80

5.2.1. DetectingQuestionTypes..................................................................................................82

5.2.2. InterchangeabilityofStatements.......................................................................................83

5.2.3. ObjectPropertiesviaSubParts..........................................................................................84

5.2.4. ObjectPropertiesviaBaseDomainSpeculation................................................................87

5.2.5. ExampleSituationAnalysis.................................................................................................88

5.2.6. AbductiveHypotheses........................................................................................................93

5.2.7. SolvingSequence................................................................................................................96

5.3 Evaluation...................................................................................................................................98

5.4 Discussion.................................................................................................................................101

Chapter6: RelatedWork....................................................................................................................104

6.1 ProblemSolvingandQuestionAnswering...............................................................................104

6.2 KnowledgeCapture..................................................................................................................106

6.3 TutoringSystems......................................................................................................................108

9

6.4 Summary..................................................................................................................................110

Chapter7: Conclusion........................................................................................................................111

7.1 ClaimsRevisited........................................................................................................................113

7.2 OpenQuestionsandFutureWork............................................................................................115

REFERENCES.............................................................................................................................................120

APPENDIXA:InstructionalAnalogies.......................................................................................................127

APPENDIXB:ScienceQuestions...............................................................................................................155

10ListofTables

Table1:ExamplesofstatementsthatuserulemacropredicatesinCycLandtheirmeaningsinEnglish.29

Table2:NumberofanalogiesintheFARguidethatinvolvereasoningaboutfunctionalsimilarities,

physicalsimilarities,andphysicalmodels.Notethatthethreecategoriesarenotmutuallyexclusive...40

Table3:NumberofanalogiesintheFARguidethatusespatialrepresentations.....................................44

Table4:AnalogybetweenATP/ADPandarechargeablebattery,Harrison&Coll(2007),p.94...............45

Table5:Analogyusedtoexplaintherelativesizeofcellsandsomeoftheirparts,Harrison&Coll(2007),

p.118..........................................................................................................................................................45

Table6:Modelfragmentsusedforinterpretinganalogies........................................................................52

Table7:Newvisualconceptualpartrelations...........................................................................................53

Table8:Neweventrolerelationsbasedonarrowlabelandarrowdirection.Foreachrow,ifthereexists

anobjectnamedarrow-objectthatbelongstothespecifiedeventtypeandpointsfromarrow-

srcandtoarrow-dest,thenthespecifiedfactscanbeinferredon-demand....................................55

Table9:GoldstandardqueriesinnaturallanguageandintheCycrepresentationallanguageforthe

enzyme/keyanalogy..................................................................................................................................72

Table10:Summaryoftextandcasesizesfor9biologyanalogiesand2electricityanalogies..................73

Table11:Numberoffactsclassifiedasbelongingtothebasedomain,belongingtothetargetdomain,

andbeingambiguous.................................................................................................................................74

Table12:Numberoffactsinbaseandtargetdomainsaftertype-levelsummarizationandanalogical

mapping.....................................................................................................................................................74

Table13:Performanceofindividualanalogiesoneachofthefourexperimentalconditions...................76

Table14:Questionsusedtoevaluatequalitativescienceknowledge.......................................................99

11ListofFigures

Figure1:AnanalogyusedtoexplainhowenzymesandsubstratesworkfromHarrison&Coll(2007),p.

96...............................................................................................................................................................16

Figure2:Ananalogybetweenabookatrestonatableandahandpushingdownonaspring,

ConceptualPhysics,Hewitt(p.28).............................................................................................................18

Figure3:AnexamplesketchwithpartialrepresentationsfromCogSketch.Colorsindicateseparate

glyphsdrawnbytheuser.User-providedlabelsfromtheCycknowledgebasegiveexplicitconceptual

informationabouttheglyphs.CogSketchdoesnotperformsketchrecognitionoftheink.Instead,it

reliesonthelabelsprovidedbytheuser.Inthisexample,thedifferentcoloredglyphsareallgrouped

togetherandlabeledagain,allowingCogSketchtorepresenteachindividualpartwithitsownlabel(e.g.

CellNucleus),aswellasthegroupofglyphswithanotherlabel(i.e.Cell).................................................36

Figure4:AnexampleofananalogyfromtheFARguide(Harrison&Coll,2007,p.234).Thisanalogy

expressessimilaritiesinfunctions/behaviorsandspatialrelationsthroughaphysicalmodel.................42

Figure5:Overviewofthesystem.Theinputtothemodelisasketch(i.e.CogSketchsketchfile)anda

textfilewithasimplifiedEnglishdescriptionoftheanalogy.Lightblueitemsrepresentsystemsor

knowledgesourcesthatpre-datethisthesis.Darkblueitemsindicatenewtechniquesorknowledge

sourcesdevelopedforthisthesis...............................................................................................................50

Figure6:Anillustrationshowinghownewvisualconceptualrelationsalloweventstobedrawnas

arrows.Thetoparrowrepresentsusingabattery.Thebottomarrowrepresentschargingabattery...56

Figure7:Thesketchassociatedwithananalogyusedtoexplainhomeostasis.Thepersoniswalkingup

anescalatorthatismovingdown.Themodelfragmentthatdescribesthisscenariohasthethree

consequencesshownbelowthesketch.Thefirststatementmeansthatthefasterthepersonwalks,the

greatertheirheight.Thesecondstatementmeansthatthefastertheescalatormoves,thelesserthe

12person’sheight.Thethirdstatementrepresentsaquantitycorrespondence:whenthepersonand

escalatormoveatthesamespeed,theheightofthepersonisnotchanging..........................................57

Figure8:TheCogSketchsketchfortheenzyme-keyanalogy.Eachpurpleboxisanindividualsubsketch,

whichisusedtorepresentanindividualstate.Thearrowsbetweenstatesareenablementrelationsto

indicatetheorderofthesequence............................................................................................................58

Figure9:Anillustrationofhowgenericeventinterpretationworks.Eventinformationfromthetext

modalityisassumedtoapplytoelementsinthevisualmodality.Providingelectricityisprojectedonto

windmillsbecausetheyarespecializationsofpowerstations...................................................................63

Figure10:Resultsfromablationexperimentsbyanalogy.Theboldlinerepresentsaverageaccuracy

acrossallanalogies.....................................................................................................................................76

Figure11:AnexamplequestioninnaturallanguageandinCycL..............................................................82

Figure12:Asubsetofthequeriesthatareexecutedforquestion1.Groundqueriesarefirstcreatedfor

allansweroptions.Sincenonesucceedforthisquestion,theQAsystemattemptsquerieswithsubpart

typesinplaceoftheoriginalcollection.Sincethosealsofail,theQAsystemattemptsqueryvariants,

withvariablesintroducedtoincreasethechancesofreturningaresult.Oneoftheresultsthatis

returnedisthatNucleihavemanygenes.Thisissimilartothegroundqueryforansweroption1,andis

thereforeusedasevidenceforthatanswer..............................................................................................86

Figure13:Anexamplequestionaboutecosystems.TherighthandcolumnshowstheCycL

representationofthescenarioandtheansweroptions............................................................................90

Figure15:Abductivehypothesisstatementsusedtoassumefunctionsandcollectionmembership.

Theseareimplementedinhornclauses,withthefirstabductivehypothesisstatementbrokenintothree

separatehornclausestohandlethedisjunction.InEnglish,thefirststatesthatifsomethinghas

subparts,thenareasonablehypothesisisthatitcanmakethesubpartsavailable,changethesubparts,

13ortakecareofthem.Forexample,knowinghavebatterieshaveenergysuggeststhatbatteriesmake

energyavailable.Thesecondstatesthatifthereissomethingthathassubpartsanditisknowntohave

subpartsofaparticulartype,thenwecanassumethatthesubpartsareofthatsubparttype.For

example,ifthereisacircuitwithahiddenpart,andweknowthatcircuitstendtohavewires,thena

reasonablehypothesisisthatthehiddenpartmightbeawire.................................................................94

Figure14:Anobjectpropertyquestionthataboutthefunctionofthecellnucleus.................................95

Figure16:Anoverviewofhowmultiplechoicequestionsareanswered.Thefirststepexecutesthe

groundqueriesdirectly.Ifananswerisfound(followingtheboldarrow),thentheprocessisdone.If

not,itcontinueswithabasicanalysisofthequestiontodetermineifitisaboutobjectpropertiesoran

examplesituation.Foreach,differentstrategiesareemployed.Theobjectpropertystrategyuses

subpartsandbasedomainentitiestocomeupwithvariantsoftheoriginalquerybyplacingvariablesin

differentargumentpositions.Ifananswerisnotfounddirectlyfromsubpartsorbasedomainentities,

thesystemexaminesthequeryvariantsandusedthemtocomputeevidenceforeachoftheanswer

options.......................................................................................................................................................96

14Chapter1: Introduction

Thisthesisexploresacommunicationmethodthatisrelevanttolearningbyreadingsystemsand

educationalsoftwaresystems:instructionalanalogy.Itiswidelyrecognizedthatanalogicalreasoning

playsavitalroleinourabilitytodetectsimilaritiesanddifferencesandtotransferknowledgebetween

topics.Instructionalanalogiesareveryfrequentlyusedintextbooksandbyeducatorstocommunicate

newideastopeople,especiallyinscience,technology,engineering,andmathematics(STEM)domains.

Whatisthenatureoftheseanalogies?Whatarethereasoningrequirementsforinterpretingthem?Do

theyprovideknowledgethatisusefultoanintelligentsystem?Thesearethemainquestionsthatthis

thesisaddresses.

Thisthesisshowsthat(atleast)thefollowingcapabilitiesareneededtointerpretinstructionalanalogies:

1) Multimodalreasoning,includingnaturallanguageunderstanding,spatialreasoning,andthe

abilitytocombineinformationfrombothmodalities

2) Useofbackgroundknowledge,includingpartialknowledgeabouteverydayconceptsor

experiences

3) Analogicalreasoning,includingtransferringknowledgefromonetopictoanother

Withoutlimits,theserequirementscanexpandtoAI-completeproblems.However,thetaskconstraints

imposedbyinstructionalanalogiescanbeusedtocontextualizetheserequirements,andtodevelop

heuristicsthatmakeinterpretationfeasible.Thisthesisdemonstrateshowtheserequirements,

constrainedbythereasoninggoalsofinstructionalanalogies,canbefulfilledwithacombinationof

simplifiednaturallanguageunderstanding,sketchunderstanding,andanalogicalreasoninggovernedby

thestructure-mappingtheoryofanalogy(Gentner,2010).

151.1 Motivation

Analogicalreasoningisahallmarkofinterpersonalcommunicationandlearning.Analogiesarepowerful

toolsforcommunicatingabstractideas,highlightingimportantsimilaritiesanddifferences,andfocusing

inference.Instructionalanalogiesarecomparisonsthatareoftenusedtoteachintroductoryscience

topics(Glynn,2008;Harrison&Coll,2007;Zeitoun,1984).Thepurposeofinstructionalanalogiesisto

describenew,unfamiliarconceptsintermsoffamiliarones.Theytendtousebasic,everyday

experiencesorobjectsasasourceofknowledge(oftenreferredtoasabasedomain)thatcanbe

projectedontoanewtopic(oftenreferredtoasatargetdomain).Forexample,inbiology,enzyme

activationisdescribedasaprocessthatissimilartounlockingalock.Instructionalanalogiesareoften

presentedexplicitly,withindividualsimilaritiesidentified,e.g.Anenzymemoleculeislikeakey/A

substratemoleculeislikealock,Figure1.Analogiesmayalsobeimplicit,wherethecomparisonis

deeplyembeddedinthelanguageusedtodescribethetopic,e.g.Theelectricalcurrentflowsthrough

thewire.Insomecases,itcanbeverydifficulttodescribethingswithoutimplicitlyinvokingananalogy

orconceptualmetaphor(Lakoff&Johnson,2003).

Instructionalanalogiesmayalsotaketheformofcomparisonsbetweenhighlysimilarthings.

Figure2showsaclassicbridginganalogy(Clement,1993).Thespatialarrangementofthetwoimagesis

intendedtoconveyananalogybetweenahandpressingdownonaspringandabookatrestonatable.

Itisalsointendedtosuggestthat,justasthespringpushesuponthehand,thetablepushesuponthe

book.Thisanalogycanbeusedtoteachstudentsaboutforcesonobjectsinstaticequilibrium.Unlike

analogiesforenzymeactionandelectriccurrent,thisanalogyinvolvestwoscenariosofthesame

domain.Forthisreason,itisreferredtoasawithin-domainanalogy,whereasanalogiesbetweentwo

completelydifferenttopicsarereferredtoascross-domainanalogies.

16

Spatialrepresentationsprovideadditionalbenefitstoinstructionalanalogies.Asillustratedby

Figure2,thespatialarrangementofelementsinadrawingordiagramcaninvitecomparisonbetween

individualitemsthatarelaidoutinasimilarway.ForSTEMdomainsinparticular,studentsoftenhave

tolearnaboutconceptsthattheycannotobservedirectly.Ingeoscience,forexample,studentsmust

understandprocessesthatcannotbefullyobservedvisuallybecausetheyaremostlyinvisible(e.g.the

carboncycle),theyoccuroverthousandsofyears(e.g.subduction),ortheyexistonverylargescales

Lockandkey EnzymesandsubstratesKey EnzymeMoleculeLock(padlock) SubstrateMoleculeNotchedpartofkey(hasuniqueshape) Activesite(hasuniquechemicalmakeup)Keysunlockonlyspecificlocks EnzymesreactonlywithspecificsubstratesKeybreaksapart(unlocks)apadlock Enzymeactionbreaksapartsubstrate

molecules.Keycomesoutofthelockunchangedandcanbereused.

Theenzymecomesoutofthereactionunchangedandcanbereused.

Figure1:AnanalogyusedtoexplainhowenzymesandsubstratesworkfromHarrison&Coll(2007),p.96.

17(e.g.layersoftheearth).Similarissuesareencounteredinbiology(e.g.microbiology)andphysics(e.g.

electricity).Spatialrepresentationsprovideawayforstudentstovisuallyobservethingsthatthey

wouldnotbeabletoseeotherwise.

Instructionalanalogiesalsotendtofocusonqualitativeknowledge,whichisimportantfor

solvingproblemswithpartialinformationandforrepresentingcontinuousphenomenawithdiscrete

labels.Qualitativeknowledgeallowsustounderstandscenarioswithlittleornodetailedquantitative

information.Theconnectionbetweenanalogicalreasoningandthedevelopmentofabstractknowledge

isimportantbecauseabstract,qualitativeknowledgeisahallmarkofexpertknowledge(Chietal.,

1981).Thishasimpactedteachingstrategiesandassessments.Forexample,theforceconcept

inventoryisatestofqualitativephysicsknowledgethatwasdevelopedafterdiscoveringthat

quantitativeassessmentswerepoormeasuresofphysicsexpertise(Hestenes&Halloun,1995;Hestenes

etal.,1992).Analogiesareusefulforbuildingqualitativeknowledgebecausetheyenablepeopletouse

afamiliarvocabularytopartiallydescribenewtopicsand,asdescribednext,theyhelppeoplefocuson

abstractsimilaritiesanddifferences.

18

1.1.1. StructureMappingTheory

Structure-mappingtheory(SMT)(Gentner,1983)providesanaccountofhowanalogicalreasoning

worksandwhyitisuseful.AccordingtoSMT,thereisanimportantdistinctionbetweensurface

similarityandstructuralsimilarity.Surfacesimilarityreferstosharedattributesbetweentwothings,e.g.

color,overallshape.Structuralsimilarity(alsocalledrelationalsimilarity)referstosharedrelational

structure,e.g.thattwoobjectscausesomethingelsetohappen.Thedistinctionbetweenrelational

similarityandsurfacesimilarityisimportantbecauseSMTclaimsthatrelationalsimilarityisthetypeof

similaritythathasthegreatestinferentialpower.Thepresenceofsharedsystemsofrelationsis

referredtoassystematicity,andtheimportanceofsharedsystemsofrelationsfordetectingstructural

alignmentisreferredtoasthesystematicityprinciple.Onecanseehowthedistinctionbetweensurface

andstructuralsimilaritymapstoinstructionalanalogiesdescribedearlier:within-domainanalogieshave

highsurfacesimilarity(andpossibly,butnotnecessarily,highstructuralsimilarity),andcross-domain

analogies(iftheyareeffectiveinstructionally)havehighstructuralsimilarity(andpossibly,butnot

necessarily,somesurfacesimilarity).

Figure2:Ananalogybetweenabookatrestonatableandahandpushingdownonaspring,ConceptualPhysics,Hewitt(p.28).

19

Psychologicalevidencesupportstheclaimthatstructuralalignmentdriveswhatanalogical

inferencespeoplemake.Inaseminalstudyonproblemsolving,GickandHolyoak(1980)demonstrated

thatcross-domainanalogiesbetweenstoriescouldbeusedforproblemsolving.Theydemonstratednot

onlythatpeoplecanusestoriesfromdistantdomainstosolvenewproblems,butthatthelevelof

structuralsimilarityimpactsthelikelihoodofarrivingatananalogoussolution.Inasimilarstudy,

BlanchetteandDunbar(2002)showedthatpeoplehaveaselectivebiasforinferencesthathavehigh

systematicity,i.e.aresupportedbyhighstructuralsimilarity.Thestructuralsimilaritythatsupports

analogicalinferencecanbespatialaswell(Gentneretal.,2015).InastudyattheChildren’sMuseumin

Chicago,childrenwereshownpairsoftowersthatvariedintermsoftheirsurfaceandstructural

similaritytoeachother.Inthiscase,structuralsimilaritydidn’tinvolvesemanticrelationsinastory,but

spatialrelationsbetweenpartsofthetowers.Childrenweretoldtocomparethetwotowerstoeach

otherandwereasked,“whichoneisstronger?”(i.e.morestable?).Ineachpair,onlyonetoweruseda

diagonalbraceinitsconstruction,whichmadeitmorestablethanthetowerthatlackedadiagonal

brace.Thepairswereeitherhighlyalignable,meaningthattheirpartssharedmanyspatial

relationships,orlessalignable,meaningthattheirpartssharedfewerspatialrelationships.Children

betweentheages6and8whowereshownthehighlyalignablepairsweremorelikelytoapplythe

diagonalbraceprincipletoatowerrepairtaskthanchildrenwhowereinthelowalignmentcondition.

Thisindicatesthattheimportantdifferencebetweenhighlyalignabletowerswasmademoresalientby

thepresenceofmanysharedspatialrelations.OtherstudiesonSMTillustratetheimportanceof

systematicity(Clement&Gentner,1991),theimpactofrelationallanguageoncategorylearning

(Gentner&Namy,1999),andthecareerofsimilaritywherebytheabilitytoattendtoabstractrelational

similaritiesincreaseswithageandexpertise(Gentner&Ratterman,1991).

20

Structure-mappingtheoryalsosuggeststhatthealignmentofrelationalstructuresenables

importantsimilaritiestoemerge,whileabstractingawayirrelevantdetails(Kotovsky&Gentner,1996;

Namy&Gentner,2002).Analogicalreasoningcommonlyreferstotransferringknowledgefroma

knowntopictoanunknowntopic,butresearchershaveshownthatevenwhenpresentedwithtwo

scenariosthatarebothonlypartiallyunderstood,mutualalignmentcanpromotedeepunderstanding

(Kurtzetal.,2001;Mason,2004).Thisleadstotheformationofabstractknowledge,orschemas,which

canreadilybetransferredtonovelsituations.GickandHolyoak(1983)foundthatabstractschemasfor

solvingparticularstoryproblemscanbeinducedbycomparingtwoanalogousstories,andthattheact

ofcomparinganalogousstoriesismoreeffectivethanstudyingandsummarizingoneanalogousstory

beforebeingremindedofitattransfertime.Lowensteinetal.(1999)conductedatransferlearning

experimentwithbusinessschoolstudentsandfoundthatcomparingnegotiationscenarioswasmore

effectivefortransferthanreadingidenticalscenariosindependently.Gadgil,Nokes-MalachandChi

(2012)foundthatcomparisonofdifferentmentalmodelsofthecirculatorysystemwasmoreeffective

thanself-explanationatundoingmisconceptions.Inanotherinvestigation,analogicalcomparisonand

self-explanationwereusedtopromotefartransferinphysicsproblemsolving(Nokes-Malachetal.,

2013),andbothwereindependentlymoreeffectivethansimplyreadingfromworkedexamples.

Additionally,thequalityoftheschemasproducedbyparticipantspredictshowoftentheyareableto

transferthatschematonewexamples(Gentneretal.,2003;Gick&Holyoak,1983;Loewensteinetal.,

1999).Thesestudiesindicatethatanalogicalcomparisonsfacilitatetheacquisitionofabstract

knowledge.Giventheimportanceofqualitativeknowledgeinearlyscienceeducation,itmakessense

thatanalogiesareappealinginstructionaltools.

Despiteevidenceoftheirefficacy,therearecircumstancesthatcancauseanalogiestobe

misleading.Spiroetal.(1989)developedataxonomyofeighttypesofanalogypitfalls,wherethereisa

21tendencytooversimplifycomplexideasbasedonindividualanalogies.Thesepitfallsincludevarious

waysofover-projectinginformationfromthebasedomain(i.e.theknowntopic)tothetargetdomain

(i.e.thenewtopic)ormissingimportantpartsofthetargetdomainbecausetherearenocorresponding

partsinthebasedomain.Thesepitfallscanbethoughtofasproblemswiththealignmentprocessoras

inadequateevaluationofcandidateinferences.Therearealsotwopitfallsthatrelatetothespecific

terminologythatisusedinanalogies.Sometermshavecommon-knowledgeconnotationsthatconflict

withtheirtechnicalmeanings.Spiroetal.provideanexamplefrombiology,wheretheword

“compliance”referstotheflexibilityofbloodvessels,andstudentsdevelopmisconceptionsaboutblood

vessels“givingwayto”or“surrenderingto”blood.Thistypeofpitfallisduetotherecruitmentof

erroneousorirrelevantbackgroundknowledgethatmakesitswayintotheanalogicalmapping.While

thetaxonomyidentifiedbySpiroandcolleaguesdescribesthevariousmisconceptionsthatcanarise

fromanalogies,structure-mappingcanprovideadditionalinsightintowhythesemisconceptionstake

place.Thesystematicityprincipleindicatesthatcommonrelationalstructuresarecriticaltomaking

structuralalignments.Itisthereforecriticaltomakecommonrelationsknownandblockpotential

mismatches.Ourtendencyfornoticingsurfacesimilarityalsoplaysaroleinthesepitfalls,sincewe

processlocalsurfacematchesbefore(orfasterthan)relationalones(Goldstone&Medin,1994).Iftwo

itemsinamappingareintendedtocorrespond,butacompetingmatchsharesgreatersurfacesimilarity,

itismoredifficulttoarriveattheintendedmapping.Suchanalogiesarecalledcross-mapped

comparisonsbecausecommonsurfaceattributesinviteanunintendedcorrespondence(Gentner&

Toupin,1986).Cross-mappedcomparisonsaremoredifficultfornovices(Gentner&Ratterman,1991),

indicatingthat,especiallywhenteachinganewtopic,surfacesimilarityshouldbeasconsistentas

possiblewithrelationalmatches.Otherfactorsthatimpedetheabilitytodetectrelationalmatches

includeexcessivecognitiveload(Markman&Gentner,1993)andunfamiliaritywiththebasedomain.

22Thesepitfallsandrecommendedsolutionsareconsistentwiththeteachingstrategiesidentifiedbythe

educationliterature,whicharedescribednext.

Therearethreeprominentguidesforteachingwithanalogies:theGeneralModelofAnalogy

Teaching(GMAT)(Zeitoun,1984),theTeachingwithAnalogies(TWA)guide(Glynn,2008),andtheFocus

ActionReflection(FAR)guide(Harrison&Coll,2007).Thegeneralprinciplesforallthreeguidesarevery

similarandconsistent.Theyallstresstheimportanceofassessinglearners’knowledgepriorto

instructionaswellasexplicitlyidentifyingsimilaritiesanddifferencesbetweenthetopicsintheanalogy.

TouseGlynn’sterms,simpleanalogiesonlygivethestudentthebaseandthetarget,e.g.“Ananimalcell

islikeacity,”whereaselaborateanalogiessystematicallymappropertiesofonedescriptionto

propertiesoftheother.Elaborateanalogiesarepreferred.Itisimportanttoidentifydifferencesaswell

becauselearnersneedtobemadeawareofhowtheanalogybreaksdown.Withoutknowingthe

differences,theanalogymaybetakentoofar,leadingtoincorrectinferencesormisconceptions.The

useofmultipleanalogiesisalsorecommendedtorefineknowledgeandfillingapsthatotheranalogies

leaveopen.Theseareimportantpartsoftheteachingprocessbecausetheyaddresssomeofthepitfalls

identifiedbySpiroetal.(1989).HolyoakandRichland(2014)takethesepracticesevenfurther,

recommending(1)multiplebases/sourcesforanalogiespresentedinorderofdifficultytofacilitate

progressivealignment(Gentner&Medina,1998),(2)techniquesforreducingworkingmemoryload(e.g.

usingvisualrepresentations),and(3)theusevisualrepresentationsandgesturesthathighlight

relationalcommonalities,makingthemhighlyalignable.Theserecommendationsareconsistentwith

theliteraturesupportingtheuseofprogressivealignment,relationallanguage,andalignabilityto

promotelearning(Gentneretal.,2015;Gentner&Ratterman,1991;Kotovsky&Gentner,1996).In

general,theseseparaterecommendationsarecomplementary.TheFARGuideisaparticularlyuseful

23resourcebecauseitcontainsacollectionofinstructionalanalogiesformanydifferentsciencedomains.

Forthisreason,itisusedasasourceofmaterialforthisthesis.

Theeducationandcognitivescienceliteratureindicatesthatmultimodalinstructionalanalogies

playimportantrolesforpositivelearningoutcomesinpeople(Alfierietal.,2013).InteractiveAIsystems

shouldthereforeknowhowtointerprettheseanalogiessothattheycancommunicatemorenaturally

withpeopleandsothattheycancaptureknowledgeinthesamewaysthatpeopledo.However,

currentAIsystemslacktheabilitytounderstandormakeuseofmultimodalinstructionalanalogies.

1.1.2. ImpactinAI

Usinganalogytotransferknowledgebetweendomains,aspeopledo,isalongstandingproblem

inAIwithmanyapplications.Researchersbelievedearlyonthattheabilitytoevaluateanewscenario

withrespecttosomepriorexperiencecouldbeveryusefulinplanning(Veloso&Carbonell,1993),

learning(Burstein,1985),andproblemsolving(Falkenhainer,1990;Klenk&Forbus,2013).Inaddition

tobeingabletouseanalogiestotransferknowledge,intelligentsystemsneedtoknowhowtointerpret

analogiesforhuman-levellearningbyreading(Barbella&Forbus,2011).Aswemovetowardthegoalof

buildinginteractiveintelligentsystems,analogicalreasoningbecomesanimportantrequirementfor

learningfromhumancollaborators,whomayusenovelanalogiestoexplaincomplexphenomena.

Similarly,interactiveintelligentsystemsneedtheabilitytocommunicateviaanalogytoengageinrich

interactionswiththeirhumancollaboratorsortoexplainnewconceptstothem.Giventhatanalogies

aresuchpowerfultoolsforlearning,buildinganalogy-basedtutoringsystemshasbeenproposed(Lulis

etal.,2004;Murrayetal.,1990),butofthethreethathavebeenimplemented,allareintendedto

operateinasingledomainandonarestrictedtopic(i.e.supportforces(Murrayetal.,1990),howthe

circulatorysystemworks(Lulisetal.,2004),andtopicsindatastructures(Harsleyetal.,2016)).

24

Aswithmanyformsofknowledgecapture,reasoningaboutanalogieswithinonemodalityisnot

enough.Multimodalinstructionalanalogiesarepervasiveintextbooksandclassrooms.Analogiesin

textbooksareoftenaccompaniedbypicturesordiagrams,andtheuseofspatialrepresentationsispart

ofbestpracticesforteachingwithanalogies,asmentionedabove(Holyoak&Richland,2014).The

importanceofspatialrepresentationsisnotsurprisingbecausetheylightenworkingmemoryload

(Larkin&Simon,1987)andsupportstudents’spatialreasoning,whichhasbeenshowntobeastrong

predictorforsuccessinmanySTEMdomains(Waietal.,2009).Foranalogiesinparticular,spatial

representationscanfacilitatealignment,makinganalogiesliketheoneshowninFigure2veryeasyfor

peopletoalign.Itisthereforeimportantforanintelligentsystemtobeabletotakeadvantageofthe

spatialinformationthatisdepictedwithinstructionalanalogies.

Intelligentsystemsthatcancombineanalogicalreasoning,qualitativereasoning,andspatial

reasoning,canadvancethestateoftheartinknowledgecapture.Understandingtherequirementsfor

reasoningaboutmultimodalinstructionalanalogiesisanimportantsteptowardthisgoal.Theclaimsof

thisthesisarethatacombinationofsimplifiednaturallanguageunderstanding,spatialreasoning,and

structure-mappingcanbeusedtobuildqualitative,conceptualknowledgefrommultimodal

instructionalanalogies,andthatsuchknowledgecanbeusedtoanswerquestionsaboutthenew

domain.Weexaminethoseclaimsindetailnext.

1.2 ClaimsandContributions

Claim1:Structure-mappingcanbeusedtobuildqualitativeknowledgefrommultimodalinstructional

analogies.

ThisclaimissupportedbyexperimentsinChapter4.Itcanbedecomposedintothreesubclaims:

1) Multimodalinstructionalanalogiesusevisualrepresentationstofacilitateinterpretation.

25

2) Instructionalanalogiesusebackgroundknowledgetofacilitateinterpretation.

3) Multimodalintegrationandanalogyinterpretationcanbeachievedusingstructure-mapping.

Aqualitativeanalysisofcommonanalogiesusedtoteachintroductoryscienceprovidesevidencethat

visualrepresentationsarepervasiveandthataccurateinterpretationreliesonimplicitbackground

knowledge.Basedonthisanalysis,Iimplementedananalogyinterpretationsystemthattakesasinput

simplifiednaturallanguagepassagesanddigitalsketches.Thissystemusesspatialandconceptual

informationdepictedineachsketchandexpressedinnaturallanguagetobuildandinterpret

instructionalanalogies.Theresultingknowledgeisqualitativebecauseitincludescausalrelationships

betweencontinuousquantitiesandtype-levelrulesthatmaynotbeuniversallytrue,butare

nonethelessusefulforreasoning.Structure-mappingservestwomainpurposesinthissystem.Itisthe

processbywhichvisualandtext-basedrepresentationsarealignedandintegrated.Itisalsotheprocess

bywhichtheinstructionalanalogyisinterpreted.Theinterpretationallowsthesystemtoconstruct

factsaboutthenewtopicthatuseontologiesfromalargescaleknowledgebase.Variationsinthe

reasoningcomponentsofthesystemdemonstratetherelativecontributionof:

1) SpatialInformation/Multimodalintegration

2) Backgroundknowledge

3) Explicitbasedomainknowledge

Contribution1:Interpretationrequirementsandstrategiesformultimodalinstructionalanalogies.

Asaresultoftheexperimentsusedtosupportclaim1,asetofinterpretationstrategiesareidentified

formultimodalinstructionalanalogyinterpretation.Thesestrategiesare:

1) Theuseofvisualrepresentationsasanaturallanguagedisambiguationheuristic

26

2) Theuseofvisual-conceptualelaborationforinterpretingsketches

3) Theuseofstructure-mappingforcombiningvisualandsketch-basedrepresentations

4) Theuseofdialogueactsandeventinterpretationsfordeterminingcross-domainmatch

constraints

5) Type-levelsummarizationoffactspriortoanalogicalmapping

Claim2:Qualitativeknowledgecapturedviamultimodalinstructionalanalogiescanbeusedtoanswer

questions.

ThisclaimissupportedbyexperimentinChapter5,whichdemonstratesthattheknowledgecaptured

fromasetofmultimodalinstructionalanalogies,althoughqualitativeandincomplete,canbeusedto

answerquestionsfrommiddleschoolscienceexams.Itisnotthecasethattheknowledgecaptured

frommultimodalinstructionalanalogiesissufficientforansweringall(ormost)oftheexamquestions.

Theexperimentsdosuggest,however,thatthequalitativeknowledgeplussomesimplequestion

answeringheuristicsareusefulforansweringquestionsaboutbasicfunctionsandpropertiesof

conceptsinthetesteddomain.

Contribution2:Questionansweringstrategiesusingknowledgecapturedviaanalogy.

Questionansweringiscarriedoutwiththefollowingbasicheuristics:

1) Detectingobjectpropertiesviapart-wholereasoningandbasedomainreasoning

2) Reasoningaboutexamplesituationswithqualitativebackgroundknowledge

3) Usingconceptandrelationhierarchiestoestimatesemanticsimilarityofstatements

271.3 Preview

Chapter2coversthetheoreticalbackgroundforthisworkandexplainsthesystemsusedinthisthesis.

Chapter3providesaqualitativeanalysisoftheanalogiesintheFARguide.Chapter4describesmy

approachforinterpretingmultimodalinstructionalanalogiesandChapter5describesmyapproachfor

answeringmiddleschoolsciencequestionsusingtheknowledgecapturedinChatper4.Chapter6

coversrelatedworkinanalogicalreasoning,multimodalreasoning,andquestionanswering.Iclosewith

ageneraldiscussionandsuggestionsforfuturework.

Chapter2: Background

2.1 Structure-MappingEngine

Thestructuremappingengineisacomputationalmodelofanalogyandsimilaritybasedonthe

structure-mappingtheoryofanalogy.Thealgorithmhasbeendescribedindetailelsewhere

(Falkenhaineretal.,1989;Forbusetal.,1994),buthereIdescribethegeneralinputsandoutputsaswell

ashowconstraintsandanalogycontrolpredicatesareusedinthistask.

Thestructure-mappingengine(SME)takesasinputtwostructureddescriptions:abaseanda

target.Thestructureddescriptionsconsistofexpressionsthatdescribethedomain.Givenabaseanda

target,SMEproducesoneormoremappingsbetweenthem.Eachmappinghasasetof

correspondences,whichsayhowthingsinthebasecorrespondtothingsinthetarget,astructural

evaluationscore,whichisanestimateofmatchquality(i.e.systematicity),and(optionally)candidate

inferences,whicharethingsthataretrueinthebaseandhypothesizedtobetrueinthetarget.There

canalsobereversecandidateinferences,whicharethingsthataretrueinthetargetandhypothesized

tobetrueinthebase.Essentially,candidateinferences(regularorreverse)representdifferences

28betweenthebaseandthetarget.Theyhavethepotentialtointroducenewtermsviaanalogyskolems.

Analogyskolemsaretermsthatexistinonedescription,butnottheother.

SMEprovidesoptionalmatchconstraintsthatoperateasadvicetoencodetaskdemands.

Partitionconstraintsrequirethatitemsofthesametypecorrespondtoeachother(e.g.applestoapples

andorangestooranges).Theseareusefulforliteralsimilaritymatcheswheretypesorsurface

attributesmatter.Requiredcorrespondencesareconstraintsonindividualitems(e.g.apple1must

correspondtoapple2).Theseareusefulinsituationswherecorrespondencesaregivenexplicitly,asis

oftenthecaseininstructionalanalogies.Partitionconstraintsandrequiredcorrespondencesareused

theinterpretationmodeldescribedinChapter4.

SMEalsousesavocabularyofanalogycontrolpredicates,whichallowforcertainpredicatesto

beignoredorforcertainpredicatestobedeclaredasnon-atomicfunctions.Anon-atomicfunctionisa

predicatethatresultsinanon-atomicterm,whichisstatementthatshouldbeconsideredasingleterm.

Forexample,ifthereisapredicate,LiquidFn,thattakesoneargumentanddenotestheliquidformof

thatargument,wemightwanttotreatfactslike,(LiquidFn Nitrogen),asbeinguniqueterms.

DeclaringLiquidFnanon-atomicfunctionmakesthathappen.

2.2 Cyc

ResearchCyc(Cyc)isalargescaleknowledgebase(KB)witharepresentationallanguagecalledCycL.1

TheCycKBhasanontologythatdescribesconceptandrelationalhierarchiesthatcanbeusefulfor

reasoning.InCyc,conceptsarerepresentedusingCollections,whicharelikesets.Entitiesareinstances

ormembersofcollections.Forexample,ifIhaveapieceoffruitinmyhand,thatpieceoffruitisan

1www.cyc.com

29entitythatbelongstothecollectionFruit.Microtheoriesarelogicalenvironmentsthatcanbeusedto

controlreasoning.Forexample,toquerytheknowledgeaboutafictionalworldinCyc,onemustrestrict

thequerytoonlyexaminefactsinthemicrotheorythatrepresentsthatfictionalworld.Predicatesare

usedtoformstatementsandtheycanbeinstance-levelortype-levelpredicates.Instance-level

predicatesoperateoninstances,whiletype-levelpredicatescanoperateoncollectionsandpredicates.

Rulemacropredicateexamples Meaning(relationAllExists cityMayor City

Person) Allcitieshaveamayor.

(relationAllExistsRange citizens City Person Many-Quant)

Allcitieshavemanycitizens.

(relationAllExistsCount physicalParts Person Eye 2)

Allpeoplehave2eyes.

(relationAllInstance citizens Person PlanetEarth)

AllpeoplearecitizensofplanetEarth.

(relationExistsAll doneBy (ControllingFn City) CityGovernment)

Allcitygovernmentshaveacitythattheycontrol.

(relationExistsRangeAll anatomicalParts Mammal Foot-AnimalBodyPart AFew-Quant)

Allmammalshaveafewfeet.

(relationExistsCountAll birthChild BirthEvent Person 1)

Everypersonisbornonce.

(relationExistsInstance temporallyCoexist Martian PlanetMars)

Loosely:ThereareMartiansonMars.Precisely:ThereexistsaMartianthattemporallycoexistswiththeplanetMars.

(relationInstanceAll temporallyCoexist PlanetMars Martian)

AllMartiansareonplanetMars.

(relationInstanceExists bordersOn NileRiver Country)

TherearecountriesthatshareaborderwiththeNileriver.

(relationInstanceExistsRange citizens Tokyo-PrefectureJapan Person Millions-Quant)

TherearemillionsofTokyocitizens.

(relationInstanceExistsCount presidentOfCountry UnitedStatesOfAmerica Person 1)

ThereisonepresidentoftheUS.

(relationAll repeatedEvent Sunrise) Sunrisesrepeat.(relationExists inProgressEvent

Writing) Thereiswritinghappeningrightnow.

Table1:ExamplesofstatementsthatuserulemacropredicatesinCycLandtheirmeaningsinEnglish.

30 CycLhasaclassoftype-levelpredicatesthatexpandintorules.Thesepredicatesarecalledrule

macropredicates(Table1).Theyareusefulforexpressinggeneralrulesthatapplytoanentire

collection.Sinceintroductoryscienceknowledgeoftencontainsinformationaboutclassesofthingsor

processes,theyareanimportantpartoftherepresentationalsystemusedforthisthesis.

Introductoryscienceknowledgealsoincludesinformationaboutfunctionsandbehaviors.These

aretypicallyrepresentedinCycaseventsorsituations(i.e.instancesofthecollectionEventand/or

instancesofthecollectionSituation).WhenanactionoreventisdescribedinCyc,itisdescribedusing

aneo-Davidsonianrepresentation,whichmeansthattheeventitselfisanentityandpropertiesofthat

eventarerelatedtotheeventseparately.Forexample,thesentence“Janekickedtheball”couldbe

describedusingathree-argumentpredicatefortheverb“kick.”

(kicked Jane the-ball)

However,theneo-Davidsonianrepresentationwouldlooklikethis:

(isa kick123 KickingAnObject)

(doneBy kick123 Jane)

(objectActedOn kick123 the-ball)

Thiswayofrepresentingtheeventismoreflexible.Ifmoreinformationwereadded,allthatisneeded

isonemorestatement,ratherthananewkickpredicatewithanotherargumentslot.Forexample,

“JanekickedtheballtoBob”couldberepresentedwiththepreviousthreestatementsandthe

following:

(toAgent kick123 Bob)

31Neo-Davidsonianrepresentationsareusedinthemodelforinterpretinginstructionalanalogies.

Functionalpredicatesarepredicatesthatdenoteotherthingsdynamically.Collectiondenoting

functionsarefunctionalpredicatesthatdenotecollectionsdonotalreadyexistintheknowledgebaseor

thatneedtobedefinedcompositionally.Forexample,thefunctionalpredicate

SubcollectionOfWithRelationToTypeFncanbeusedtodenotesubcollectionswithparticularrelation

toothercollections.Todenotethecollectionofeventswhereaballiskicked,wecanusethis:

(SubcollectionOfWithRelationToTypeFn KickingAnObject objectActedOn Ball)

ThistermdenotesthecollectionofeventsthatareinstancesofKickingAnObjectandarerelatedtoan

instanceofBallviaobjectActedOn.Tosaythataparticularevent,kick123,belongstothiscollection,

wecouldusethisstatement:

(isa kick123

(SubcollectionOfWithRelationToTypeFn KickingAnObject objectActedOn Ball))

Whileneo-Davidsonianrepresentationsareusefulforflexiblyrepresentingindividualevents,collection

denotingfunctionsareusefulforpreciselydescribingcollectionsofevents.Pairedwithrulemacro

predicates,collectiondenotingfunctionscanbeusefulforbuildingrules.Forexample,tosaythatall

childrenkickballs,wecouldusethisstatement:

(relationExistsAll performedBy

(SubcollectionOfWithRelationToTypeFn KickingAnObject objectActedOn Ball)

HumanChild)

32Collectiondenotingfunctionsandrulemacropredicatesarebothusedinthissystemtorepresenttype

levelknowledge.

2.3 QualitativeReasoning

QualitativeProcessTheory(QPT)(Forbus,1984)isarepresentationalsystemforcarvingupcontinuous

quantitiesandprocessesintoqualitativerepresentations.Continuousquantitiesincludethingslikeage,

height,weight,accountbalance,calories,etc.Eachofthesequantitiescanhaveanumericalvalue

associatedwiththem.However,weoftenreasonaboutquantitieswithoutknowingtheexactnumerical

valuesbyusingtermslike“short”or“low-calorie.”InCyc,suchquantitiesareinstancesofthecollection

ScalarInterval,whichisthecollectionofthingsthatcanberankedonsomescale.Thisincludes

quantitiesliketemperature,whichcanhavenumericalvalues,andquantitieslikehappiness,whichdon’t

havenumericalvalues.

InQPT,processescausechangesinquantities.Processescanberepresentedusingmodel

fragments(Friedman&Forbus,2011;Klenk&Forbus,2009),whichrepresentpartialinformationabout

aprocesstype.Eachmodelfragmenthasparticipants(i.e.theentitiesinvolved),constraintsonwhat

canbeinvolved,conditionsfortheprocesstoactuallyhappen,andconsequencesoftheprocess.Model

fragmentscanalsorepresentconceptualorphysicalviewsthatprovideaperspectiveonasituation.

Suchmodelfragmentsarenotprocesses,sowhiletheymayinvolvecausalrelationships,theydonot

directlycausequantitiestoincreaseordecrease.Theconsequencesofmodelfragmentstypically

includecausalrelationships,ordinalrelationships,orcorrespondencesbetweenquantities.

Causalityisembeddedinrelationshipscalledinfluences.Directinfluencesdescribechangesto

extensiveparameters,e.g.“Agrowthprocessesdirectlyinfluencesweight.”UsingQPTsyntax,this

statementwouldbe(I+ (WeightFn dog) (RateFn dog-growing)),wherethetokendog-growing

33wouldrefertotheprocessofthedoggrowing.Therateofthegrowthprocessisacausalantecedentof

theweightofthedog.Alooserformofcausalitycanberepresentedwithindirectinfluences,whichcan

describechangestoextensiveorintensiveparametersthatarenotnecessarilyadditive.Indirect

influencesarerepresentedbypredicatesforpositiveandnegativequalitativeproportionalities,denoted

bythepredicatesqpropandqprop-.Forexample,inthewinter,thehigherthetemperature,the

happierIam:

(qprop (HappinessFn me) (TemperatureFn Chicago))

Thefactthatthissentenceisqualifiedwith“inthewinter”indicatesthatthisqualitativeproportionality

couldbeaconsequencetoamodelfragmentdescribingwinter,meaningthatwhenthatmodel

fragmentisactive,therelationshipholds.Inthesummer,therelationshipmightchangeto:

(qprop- (HappinessFn me) (TemperatureFn Chicago))

ThisnegativequalitativeproportionalityindicatesthatIamhappierwhenitiscooler.Asisthecasewith

directinfluences,thesecondargumentofanindirectinfluenceisthecauseandthefirstargumentisthe

effect.Correspondencesrepresentpointsofequalitybetweenquantities.Considerthefollowing

expression:

(qpCorrespondence (PopulationFn GrantPark-Chicago) Zero

(TemperatureFn Chicago) Zero))

ThiscorrespondencemeansthatwhenthetemperatureinChicagoiszero,thenumberofpeopleat

GrantPark(whichisoutside)iszero.Causalinfluencesenableaqualitativeanalysisofasituationin

termsofwhatquantitiesarechanging.Givenascenariowithatleastoneactiveprocess,thescenario

canbecharacterizedintermsofwhatquantitiesareincreasing,decreasing,andstayingthesame.This

34characterizationiscalledinfluenceresolution.Qualitativesimulationalsoenableslimitanalysis,which

involvesthepredictionoffuturestates,althoughthisaspectofQPTisnotusedforthisthesis.QPTis

usefulforsupportinginferencewithpartialinformation,whichisexactlythetypeofinformationthat

existswhenatopicisfirstbeinglearned.

2.4 EANLU

TheExplanationAgentNaturalLanguageUnderstandingsystem(EANLU)(Tomai&Forbus,2009)

producessemanticrepresentationsforsimplifiednaturallanguagetext.EANLUtakesapragmatic

approachtonaturallanguageunderstanding.Semanticinterpretationoftextismadetractablebyusing

sentenceswithsimplifiedsyntax.Inotherwords,thegoalisnottohavecompletecoverageofnatural

languageinputs,butrathertohaveverybroadcoverageoftheknowledgethatcanbeexpressedtothe

systemusingsimplesentences.ForeachsentenceEANLUprocesses,itgenerateschoicesetsto

representthepossibleinterpretationchoicesthatcanbemade.Severalheuristicsformakingsemantic

interpretationchoicesarebuiltintoEANLU.Forthisthesis,semanticchoicesareselectedusingtwo

simpleheuristics:informationgainandfavoredcontext.TheseheuristicsandthewayEANLUisusedto

interpretmultimodalinstructionalanalogiesaredescribedinsection4.2.3.

2.5 CogSketch

CogSketchisadomain-independentsketchunderstandingsystem(Forbusetal.,2011).Sketch

understandingreferstotheprocessofusingspatialreasoningtointerpretsketches,independentof

sketchrecognition.Sketchunderstandingisintendedtocoverthetypesofsketchesthatarequalitative,

abstract,andmaynotnecessarilyreflectrealisticimages.CogSketchachievesthisbyusingqualitative

spatialreasoningoverinkthatismanuallysegmentedandlabeledbytheuser.InCogSketch,thebasic

buildingblocksofasketcharecalledglyphs.Todrawaglyph,theuserdrawsinkusingamouseorstylus

35andtellsthesoftwarewhentheglyphisdonebyclickingaFinishGlyphbutton.Thismeansthattheuser

manuallysegmentsinkintomeaningfulobjects.Inkeditingtoolsallowtheusertomergeandre-

segmentinkastheydraw.Groupingtoolsalsoallowtheusertogroupconceptualitemstogether(e.g.

twowheelsandaframecanbegroupedtogethertorepresentabicycle).Theusercanalsoprovide

conceptuallabelsforglyphsusingcollectionsandrelationsfromtheCycknowledgebasesothatthe

softwarehasamodeloftheuser’sintentthatistiedtotheCycontology.Bylabelinganindividualglyph

withacollectionfromCyc,theuserisindicatingthattheyhavedrawnaninstanceofthatcollection.For

example,inFigure3,theinnermostcirclerepresentsacell’snucleus.ToconveythistoCogSketch,the

userwoulddrawthenucleus,clickFinishGlyphtoindicatethattheglyphisdone,andlabelitwiththe

collectionCellNucleusfromtheCycknowledgebase.Additionally,thepartsofthecellinFigure3can

begroupedtogetherbytheuser.BylabelingthegroupedglyphswiththeCyccollectionCell,theuser

indicatesthat,whengroupedtogether,theglyphsrepresentanindividualcell.Althoughthisapproach

requiresextradrawingandlabelingeffort,ithastwoadvantagesoverrecognitionofrawink.Thefirstis

thatitavoidssegmentationandrecognitionerrorsbecausetheuserexplicitlytellsthesoftwarehowto

groupinkandwhattheywanttheinktorepresent.Thesecondisthatthisdraw-and-labelinterfaceis

amenabletoeducationalsettingsbecausestudentsarerequiredtolabelsketchesandexplicitlyprovide

their(possiblyincorrect)interpretationofwhattheyhavedrawn.Also,inkrecognitionwithoutlabeling

wouldnotworkacrossmultipledomains,sincethemappingfromshapestoconceptsismanytomany.

Itisespeciallyproblematicwhenanewdomainisbeingintroduced,sincetrainingrecognizersrequires

manyexamples.

CogSketchautomaticallygeneratesqualitativespatialrepresentationsforwhatisdrawnina

sketch.Topologicalrelations(e.g.intersection,containment)andpositionalrelations(e.g.above,right

of)areautomaticallycomputedbetweenadjacentglyphs.Spatialrelationsbetweennonadjacentglyphs

36canbecomputedon-demand.TheconceptuallabelsprovidedbytheuserarealsousedbyCogSketchso

thatspatialandconceptualinformationexistinthesamereasoningenvironment(i.e.inthesame

microtheory).CogSketchsketchescan(optionally)consistofmultiplesubsketches,whichareusefulfor

representingindividualstatesinasketch.Eachsubsketchhasitsownmicrotheorysothatthingsabout

thesameobjectcanbetrueinonesubsketch/stateandnottrueinanother.

Forinterpretingthevisualportionofinstructionalanalogies,IcreatedsketchesusingCogSketch.

Eachsketchwasmanuallyorganizedintoindividualglyphs,andlabeledwithconceptsfromtheCyc

knowledgebase.Whenobjectsweredepictedasindividualpartsofanotherobject,glyphgroupingwas

Figure3:AnexamplesketchwithpartialrepresentationsfromCogSketch.Colorsindicateseparateglyphsdrawnbytheuser.User-providedlabelsfromtheCycknowledgebasegiveexplicitconceptualinformationabouttheglyphs.CogSketchdoesnotperformsketchrecognitionoftheink.Instead,itreliesonthelabelsprovidedbytheuser.Inthisexample,thedifferentcoloredglyphsareallgroupedtogetherandlabeledagain,allowingCogSketchtorepresenteachindividualpartwithitsownlabel(e.g.CellNucleus),aswellasthegroupofglyphswithanotherlabel(i.e.Cell).

37usedtorepresentpart-wholerelationships.Inotherwords,CogSketchdidnotperformanysketch

recognition.Rather,itgeneratedqualitativespatialrepresentationsforconceptuallylabeledglyphs.

2.6 Companions

TheCompanionCognitiveArchitecture(Forbusetal.,2009)isbasedontheideathatintelligentsystems

aresocialorganismsthatcollaboratewithothersandlearnoverextendedperiodsoftime(e.g.from

experience).InaCompanion,analogyisacentralreasoningmechanism.EachCompanioniscapableof

usingmultiplemodesofinteraction.IthasanaturallanguageinterfacethatisbuiltuponEANLUanda

sketchinginterfacethatusesCogSketch.IusetheCompanioncognitivearchitecturetomodelthe

interpretationofmultimodalinstructionalanalogiesandtoanswermultiplechoicequestions.

InstructionalanalogyinterpretationiscoordinatedbyHTNplans.Questionansweringstrategiesare

representedbyacost-basedAND/ORtree,whereANDnodesrepresentsubgoalsforsolvingaquestion

(i.e.theymustallsucceed)andORnodesrepresentdifferentstrategiesforsolvingasinglequestion(i.e.

onlyonemustsucceed).TherearecostsassociatedwithORnodes,sothatsomestrategiescanbe

prioritizedoverothers.

Chapter3: CharacterizationofInstructionalAnalogies

Ingeneral,thegoalofinstructionalanalogiesistobuildknowledgeaboutanewtopic.However,thisis

achievedinmanydifferentways.Someanalogieshighlightfunctionalsimilaritiestohelpnovices

understandcomplexphenomena.Otheranalogiessimplyprovideaphysicalmodelforsomethingthatis

notdirectlyobservable.Inthischapter,Iprovideaqualitativeanalysisofthereasoningrequirementsof

instructionalanalogiesfoundinaguideforin-serviceteachers.

383.1 TheFARGuide

TheFARguide(Harrison&Coll,2007)containsanalogiesusedforteachingmiddleschoolandsecondary

schoolscience,collectedfromtheeducationalliterature,withrecommendationsforhowtousethem

effectivelyintheclassroom.Theanalogiescoverseveraltopicsoverfourdomains:Biology(12),

Chemistry(13),EarthandSpaceScience(10),andPhysics(14).

Eachanalogyusesacommonsensedomain,topic,orphysicalmodeltodescribeasciencetopic.The

basicgoalistogetthelearnertotransfertheirknowledgeabouteverydaysituationstotheirknowledge

aboutscience.However,theanalogiesvaryintermsofwhatsimilaritiesshouldbeattendedtoand

whattypesofinferencesshouldbemade.Theyalsovaryinwhatkindofspatialrepresentationsare

associatedwiththem.Althoughsimilaritiesanddifferencesareoftenexplicitlystatedineachanalogy,

thereisoftenbackgroundknowledgethatthelearnerisexpectedtouse.Thefollowingsectionsprovide

aqualitativeanalysisoftheseproperties,whichprovidesinsightintowhatthereasoningrequirements

areforinterpretation.

3.2 FocusedAlignmentandInference

Asothershavenoted(Spiroetal.,1989),analogiescansometimesleadtoinaccurateconclusions.The

authorsoftheFARguide(andotherguidesintheeducationliterature)agreewiththiscaveat,and

thereforerecommendthatteachersexplicitlyidentifyindividualsimilaritiesanddifferences.Thisis

especiallyimportantforcross-domainanalogies,wheretherearemanyirrelevantsimilaritiesand

differencesthatcandistractstudentsorleadthemtoincorrectconclusions.Asaresult,thereasoning

goalsandtargetknowledgeofananalogydetermineswhatsimilaritiesanddifferencesaredescribed.

Table2showsthenumberofanalogiesintheFARthatbelongtodifferentcategoriesofreasoninggoals.

Outofthe49analogiesintheFARguide,eachgenerallyfitintoone,ormoreofthesecategories:

39

1) Analogiesaboutfunctions,e.g.providingenergy,buildingsomething

Theseanalogiesdescribefunctionsorbehaviorsintermsofknownphenomena.Forexample,inthe

analogyacellislikeacity,partsofthecityarecomparedtopartsofthecellbecausetheyhave

similarfunctions.Powerstationsarelikemitochondriabecausetheybothprovideenergy.

Constructioncompaniesarelikeribosomesbecausetheybothbuildthings:homesandproteins,

respectively.

2) Analogiesaboutphysicalproperties,e.g.containment,relativesize

Theseanalogiesdescriberelativesize,shape,orphysicalconfigurations.Forexample,intheanalogy

acellislikeplanetEarth,partsoftheplanetarecomparedtopartsofthecelltoillustraterelative

sizeandproperpartrelations.Acellisalreadysmall,butrelativetochromosomes,thecellisvery

large.Similarly,theplanetismuchlargerthananindividualtownorcity.

3) Physicalmodelanalogies

Thesetypesofanalogiesaremorecommonindomainsthataimtoexplainprocessesorthingsata

levelofdetailthatcannotbedirectlyobserved.Forexample,analogiesaboutelectricityuse

physicalmodelstoexpressthefunctionsofthepartsofasimpleseriescircuitandthecausal

relationshipsbetweenquantities(e.g.voltageandcurrentflow).Itisnotthecasethatelectricity

cannotbeobservedatall,sinceonecouldbuildacircuitandobserveitsbehavior.However,the

explanationforhowthatartifactworksisinitiallydependentonthephysicalpropertiesofa

differentprocesswhereindividualbehaviorsaremoredirectlyobservable.Studentsdonotsee

electricityflowing,buttheyhaveseenwaterflow.

40

MostoftheinstructionalanalogiesintheFARguideinvolvesimilaritiesinfunctionsor

behaviors.Thisisnotsurprisingbecausecross-domainanalogiesusuallylacksurfacesimilarity.In

biology,forexample,proteinsynthesisisdescribedasfollowingamasterplantobuildahouse.The

focusofthisanalogyisnotonsurfacecharacteristics.Instead,theanalogyfocusesontherolesthat

DNA,thecellnucleus,andaminoacidsplayinbuildingproteins.Studentsfamiliarwiththenotionof

blueprintsandmasterplanscanusethatknowledgetointerpretDNAasinstructionsforbuilding

proteins.Anotherexampleisinphysics,whereconservationofelectriccurrentinasimpleseriescircuit

isexplainedusingananalogytoapassengertrain.Inasimpleseriescircuit,withabatteryandalight

bulb,theconversionofenergybythebatteryandlightbulbarerepresentedbypassengersenteringand

exitingthetrain.Thetraincontinuestomovesolongasitisinaclosedtrack,whichmimicscurrentina

closedcircuit.Inthisanalogy,it’snotthephysicalconfigurationthatmatters,butratherthe

requirementsformovement(i.e.flow)inacircuit.

Otheranalogiesfocusonsimilaritiesinrelativesize,structure,and/orshape.Theseanalogies

areusedtodealwiththedifficultiesinobservingquantitiesatverylargeorverysmallscales.For

example,tohelpstudentsunderstandAvogadro’snumber,theFARguidedescribesananalogytograins

ofricethatfilltheUnitedStatesororangesthatfilltheplanet.Thisisintendedtohelpstudents

Topic FunctionalSimilarities

PhysicalSimilarities

PhysicalModel

Total

Biology 8 3 2 12Chemistry 7 8 4 13Earth/SpaceScience 8 8 6 10Physics 14 2 9 14Total 37 21 21 49Table2:NumberofanalogiesintheFARguidethatinvolvereasoningaboutfunctionalsimilarities,physicalsimilarities,andphysicalmodels.Notethatthethreecategoriesarenotmutuallyexclusive.

41understandhowlargeAvogadro’snumberis,andtobeabletoconceptualizeamoleasasingleunitof

something.Similarly,tounderstandthelengthofthehumangenome,instructorsmayalsousean

analogytoaverylongroadtrip.Inthisanalogy,regionsalongthelongtrip(e.g.states)represent

chromosomes.Monotonouspartsofthetrip(e.g.longstretchesofemptyland)representrepetitive

DNAthatdoesnotcodeforproteins.Interestingparts,likebusycities,representgenesthatplayan

importantroleinproteinsynthesis.Thisanalogyisintendedtohelpstudentsunderstandthatthereare

varyingpartsofthegenome,andthatmostofitconsistsofDNAthatdoesnotcodeforproteins.

Switchingbetweentwoverydifferentscalesisintendedtohelpstudentsunderstandtherelative

magnitudeofthingstheycannotdirectlyobserve.

Anothercharacteristicofsomeanalogiesisthattheyuserealphysicalmodelsordemonstrations

toillustrateanewtopic.Sometimestheseanalogiesareenrichedversionsofanalogiesthatconvey

similaritiesinsize,shape,orstructure.Althoughothertimes,thereislittlephysicalsimilarityandthe

focusisonfunctionsandbehaviors.Forexample,ananalogycanbeusedtoteachstudentsaboutthe

interdependenceofecosystemsandfoodwebs.Byhavingstudentsphysicallyconnecttoeachother

withstring,theycanunderstandhowdisturbancesinoneareacanaffectmanyothers.Theconnections

betweenstudentsbearnophysicalresemblancetorealecosystems,butthegoalistohavethem

transfertheirobservationsaboutmovementandstabilitytothestabilityofecosystems.Ontheother

hand,amorephysicallyfocusedanalogywithaphysicaldemonstrationisusedtoteachstudentsabout

thegeologicrecordoftheearth.Thisanalogyworksbyhavingstudentscreateatimelineinrealspace

toillustratethelargedistances(i.e.times)between,forexample,theextinctionofdinosaursandthe

presentera.Likethegenomeanalogy,thegeologicrecordanalogyillustratestherelativemagnitudeof

differentspansoftime,butitachievesthiswitharealphysicalmodel.

42

Notethatthesecategoriesarenotmutuallyexclusive.Someanalogiesinvolvesimilarphysical

featuresaswellassimilarfunctions.Forexample,inearthscience,layersofspongesareusedtomodel

salinityinsoil(Figure4).Usingwater,salt,andaheatsource,waterevaporationandsaltcrystal

formationcanbeobserved.Thisallowsstudentstounderstanddrylandsalinityintermsofanobserved

physicalmodel.Thephysicalmodelisaimedatillustratingthattheheatcauseswaterevaporation,

whichleavesbehindsaltcrystals.However,thesecausalrelationshipsarealsotiedtospatial

information.Thewaterrisestodrylayersthatareabovewetones.Thewatertableisatthelowest

levelofsponges.Thephysicalconfigurationofthespongesalsoreflectstherelativephysicalstructure

ofthelayersofsoil.Forthesereasons,thisanalogyinvolvesbothfunctionalandphysicalsimilarities.

Thisanalogyalsohappenstoachievethisthroughtheuseofarealphysicaldemonstration.

Figure4:AnexampleofananalogyfromtheFARguide(Harrison&Coll,2007,p.234).Thisanalogyexpressessimilaritiesinfunctions/behaviorsandspatialrelationsthroughaphysicalmodel.

43

ThewiderangeofanalogiesfoundintheFARguideillustratethattheycanbecharacterizedby

multipleoverlappingcategories.Thereasoninggoalsoftheseanalogiesindicatethatamodelof

interpretationneedstorepresentfunctionalandphysicalpropertiesatalevelofabstractionthat

transcendsdomain.Inotherwords,eventhoughthecross-domainfunctionalandphysicalproperties

arenotliterallythesame,theyneedtoberepresentedasiftheyare.Thisissimilartothewaynovices

representthisinformationbeforetheylearnthedomain’stechnicalvocabulary.Forexample,

mitochondriaarethesiteswherephosphatemoleculesbondwithadenosinediphosphate(ADP)to

createadenosinetriphosphate(ATP),whichisthemostcommonsourceofcellularenergy.Ratherthan

representingthisprocessatthislevelofdetail,itcansimplyberepresentedasmitochondriaproviding

energy.Similarly,onecouldgointogreaterdetailabouthowpowerstationsgenerateelectricity,but

instead,sayingthatpowerstationsprovideenergyismoreamenabletoamatchwithourdescriptionof

mitochondria.Languageplaysacriticalroleindeterminingwhatlevelofabstractionisusedanda

successfulinterpretationmustmakethisdeterminationcorrectly.

3.3 SpatialRepresentations

MostoftheanalogiesintheFARguideareaccompaniedbypicturesorspatialrepresentationsofsome

kind(Table3).Forthepurposesofthisinvestigation,Idefinespatialrepresentationstobeanynon-

linguisticinputthatprovidesspatialinformation.Thisincludesphotos,sketches,diagrams,andphysical

demonstrationswherethelearnervisuallyobservessomething.Ofthe49analogiesintheFARguide,45

hadaspatialrepresentationofsomekindassociatedwithit.Inmostcases,thespatialrepresentation

depictedthebasedomain(44).Infewercases(13),therewasaspatialrepresentationforthetarget.In

allbutoneoftheanalogies,anytimetherewasaspatialrepresentationforthetargettherewasalsoone

forthebase.

44

Althoughthereweresomeanalogiesthatwereaccompaniedwithpictures,thevastmajority(39outof

45)werepresentedwitheitherdiagramsoraphysicaldemonstrationofsomekind.Thepresenceof

diagramsindicatestheneedtoderiveconceptualinformationfromabstractsymbolslikearrowsand

topologicalrelationslikecontainment.Forexample,thediagramfortheanalogybetweenATPanda

batteryusesarrowstoindicatethetransformationofATPtoADPandviceversa.Thepresenceof

physicaldemonstrationsindicatestheneedtounderstandphysicalprocessesorsituationsinawaythat

supportstheacquisitionofqualitativeknowledge.Forexample,intheanalogyfordescribing

interdependenceintheecosystem,studentsneedtounderstandthatconnectionsbetweenstudents

indicateinterdependenceandthatjustasinthephysicalmodel,achangeintheecosystemripples

throughtoimpactconnectedspecies.Thepresenceanduseofspatialrepresentationstoconvey

conceptualinformationindicatesthatspatialreasoningisanimportantaidinunderstandingthese

analogies.TheexperimentsinChapter4investigatehowoftenthespatialrepresentationsareactually

requiredforinterpretation.

3.4 BackgroundKnowledge

Instructionalanalogiesvaryinhowtheyarepresentedandinhowmuchimplicitbackgroundknowledge

isusedbythelearner.Implicitbackgroundknowledgeisknowledgeaboutthebasedomainthatisn’t

explicitlystatedorshowninthepresentationoftheanalogy.AllanalogiesintheFARguideare

Topic BaseSpatialRepresentation

TargetSpatialRepresentation

NoSpatialRepresentation

Biology 10 6 2Chemistry 11 4 1Earth/SpaceScience 10 0 0Physics 13 3 1Total 44 13 4

Table3:NumberofanalogiesintheFARguidethatusespatialrepresentations.

45presentedintableformat,sothemanyoftheimportantcorrespondencesareusuallyexplicit.However,

eventhistypeofexplicitrepresentationcanleaveoutimportantinformation.Takeforexamplethe

analogybetweenATPandbatteriesinTable4andtheanalogybetweenthecellandEarthinTable5.

TheanalogyusedtoexplainATPismuchmoredetailedandcontainsmoretechnicallanguage.

TheprimarygoalistounderstandhowATPprovidesenergyforvariouscellactivitiesandthatitcanbe

usedrepeatedly.Understandingthisprocessintermsofbatteriescanbeausefulfoundation,butthe

sameinformationisalsogivenexplicitlyinthetextoftheanalogy.Thisisnottosaythatthebase

domainofbatteriesservesnopurpose,butitdoesdemonstratethatsometimesexplanationsofthe

RechargeableBattery ATPandADPAchargedbatteryisenergysufficient. ATPisenergysufficient.Batteriesmoveenergytowhereit’sneededtopoweranelectronicdevice.

ATPismovedtowhereenergyisneededinthecell.

Achargedbatteryconvertsintoaflatbatterywhenenergyisused.

ATPconvertstoADPwhenenergyisused.

Arechargeablebatterycanbeusedoverandoveragain.

ATPcanbeusedoverandoveragain:ADP+PàATP

Abatteryrechargeristhesitewhereenergyisreintroducedtotheflatbatteries

MitochondriaarethesiteswhereenergyisusedtochangeADPtoATP.

Rechargedbatteriescanbeusedforavarietyofmachinesforavarietyoftasks.

ATPcanbeusedatmanysitesinthecell(ribosomesorcellmembrane)fortaskssuchasproteinproductionoractivetransport.

Table4:AnalogybetweenATP/ADPandarechargeablebattery,Harrison&Coll(2007),p.94

TheEarthtoastreetaddress ThecelltoacodonTheEarth AcellAcontinent ThecellnucleusAstate AchromosomeAcity ChromosomalDNAfragmentStreetAddress CodonTable5:Analogyusedtoexplaintherelativesizeofcellsandsomeoftheirparts,Harrison&Coll(2007),p.118

46basedomainareincludedexplicitlywiththeanalogy.Thesestatementsserveasindicatorsforhowthe

mappingbetweenthetwodomainsshouldwork.Incontrast,thecell/earthanalogyinTable5onlylists

correspondencesbetweenentities.Itismuchlessdetailed.Thepurposeofthisanalogyisto

understandtherelativesizeofthecellanditsparts,eventhoughsizeandpart-wholerelationshipsare

notexplicitinthelanguage.Thisindicatesthatthelearnerisexpectedtofillinthegapswith

backgroundknowledge(andvisualinformationifthereisany).Aswithspatialinformation,experiments

inChapter4provideevidencethatbackgroundknowledgedoesimpactinterpretation.

3.5 TargetKnowledge

Thetargetknowledgeforinstructionalanalogiesatthisagelevel(i.e.middleschool)isqualitative.

Specifictechnical,quantitativeinformationisrarelyaddressedintheseanalogies.Analogiesthatfocus

onfunctionalorbehavioralsimilaritiesrelyonqualitativerepresentationsandabstract(sometimes

metaphorical)language.Analogiesthatfocusonphysicalsimilaritieshavetodowithrelative,not

absolute,sizeandstructure.Usingthemitochondriaandpowerstationexamplefromearlier,the

productionofenergyisdescribedasanactofmakingsomethingavailabletothecell,ratherthana

chemicalreactionthatcreatesATPmolecules.Inanotheranalogy,homeostasisiscomparedtowalking

upadownescalator.Thespeedoftheescalatorgoingdownandthespeedofthepersongoingupmust

beequalinordertomaintainhomeostasis.Thequantitativevaluesofthosespeedsandtheresulting

bodytemperaturearenotcommunicatedintheanalogy.Instead,itiscommunicatedthatthespeedof

theescalatorhasanegativeinfluenceonthebody’stemperatureandthespeedofthepersonhasa

positiveinfluenceonthebody’stemperature.Thesequalitativerelationshipsconveythemeaningof

47thisanalogy.Theonlyquantitativevaluethatisgivenisthenormalbodytemperatureofthehuman

body,becauseitrepresentsboundaryconditionbetweenqualitativelydifferentstates(i.e.hyperthermia

andhypothermia).Inthecell/earthanalogy(Table5),thecellanditspartsarecomparedtotheEarth

anditspartstoconveytheverylargesizedifferencesbetweensmallpartsandbigparts.However,

quantitiesareonlyusedtoindicatethattherearemanysubpartsinbothdomains.Theactual

cardinalityvaluesarenotused,andforgoodreason,sinceusingthosevaluesmightbringastudentto

concludethateverycellhassixorsevennuclei.Inallcases,thingsthatareexplicitlymentionedinthe

analogyprovidecluesastowhatisimportant.However,becauseimplicitbackgroundknowledgeis

sometimesrequired,justbecausesomethingisn’texplicitlymentioned,doesnotmeanitisunimportant.

Thetargetknowledgeforinstructionalanalogiesisintendedtobegeneralsothatitcanbeapplied

tonewsituationsordomaininstances.ThispresentsachallengeforAIsystemsattemptingtolearn

fromtheseanalogiesbecausetheanalogiesdonotexplicitlyidentifywhichinformationisgeneraland

whichinformationisincidental.Thisisaknownprobleminnaturallanguageunderstanding,referredto

astheproblemofgenerics(Leslie,2008),wherestatementsaboutparticularthingsmaybeintendedto

applytoall(ormost)otherthingsofthesamekind.Forexample,givenadiagramthatdepictsan

enzyme,theanalogymayexpress,inlanguage,thattheenzymehasanactivesite.Despitetheuseof

thedefinitearticle,the,thefactthattheenzymehasanactivesiteisintendedtobeunderstoodasa

48propertyofallenzymes.Successfullyinterpretingtheseanalogiesdependsonknowinghowto

generalizecertainpartsoftheanalogy,andknowingwhichpartstogeneralize.

Chapter4: InterpretationofMultimodalInstructionalAnalogies

Buildingsoftwarethatcanunderstandinstructionalanalogieshastwopotentialbenefits:knowledge

captureandeducation.InthischapterIdescribemyapproachforbuildingasystemthatcan

understandmultimodalinstructionalanalogies,usingexamplesfromtheFARguide.

4.1 Problem

Thereareseveralchallengesinbuildinganalgorithmforinterpretingmultimodalinstructionalanalogies.

ThesechallengesoverlapwiththereasoningrequirementsidentifiedinChapter3:focusedalignment,

spatialreasoning,backgroundknowledge,andqualitative/type-levelrepresentations.Asshownin

Chapter3,instructionalanalogiesoftenconsistofnaturallanguageandvisualrepresentations.

Interpretingthesetwosourcesofinformationindependently(includingnaturallanguageprocessingand

spatialreasoning)isachallenge,asiscombiningtheminameaningfulway.Informationaboutthebase

andtargetdomainsisofteninterspersedanditisuptothereadertofigureoutwhichinformation

describesthebasedomainandwhichinformationdescribesthetarget.Peoplealsobringbackground

informationintotheforegroundsothatitcanbeusedintheanalogy.Thisallowsthemtogobeyond

whatisexplicitlystated.Usingqualitativeand/ortype-levelknowledgeenablesindividualstogeneralize

theknowledgebeingdescribedinanindividualanalogytoanentireconceptortopic.Buildingthis

abilityintoaninstructionalanalogylearningsystemwouldenableittomakeinferencesabouttypesof

thingsratherthanindividualexamples.Thereisalsothechallengeofbuildingtheanalogyitself.Inthis

thesis,Iusethestructure-mappingenginetobuildmappingswithconstraintsthatareguidedbythe

49typesofcross-domainanalogiesfoundintheFARguide.Lastly,onceamappingiscreatedviastructure-

mapping,inferencesneedtobeevaluatedtodeterminewhichonescanbeacceptedascorrect.

Thereisnosinglewaytointerpretaninstructionalanalogy.Differentapproachescanvarywith

respecttomodality,levelofhuman-computerinteraction,andlevelofincrementalreasoning.Ichose

tobuildasystemthatanalyzedsimplifiedEnglishtextandaCogSketchsketchfiletobuildqualitative

knowledgewithouthumanintervention.Ichosethisapproachfortworeasons.Thefirstreasonisan

assumptionthatlearningindependentlyrepresentsthemostbasicwaytoillustratetheeffectivenessof

alearningsystem.Thatis,ifasystemcanlearnatopicindependently,thatprovidesthestrongest

evidencethatitslearningapproachissufficient.Thesecondreasonisthatexcludinghuman

interventionduringthelearningprocesscouldrevealtheobstaclestobuildingindependentlearningby

readingsystemsthatcanmakeuseofinstructionalanalogiesfoundintextbooksandonlinesources.

Pointsoffailureprovideinsightintohowthelearningapproachcouldbeimproved,eitherthrough

improvedsketchand/ornaturallanguageunderstandingorthroughhumanintervention.

4.2 Approach

ThegeneralapproachforinterpretationissummarizedinFigure5.Asatestbedforthisproblem,I

gatheredanalogiesfromtheFARguideusedtodescribetopicsinbiology(9)andelectricity(2).Each

analogyconsistsofnaturallanguagetextdescribingsimilaritiesanddifferences.Mostanalogiesalso

consistofsomekindofspatialrepresentation(eitherapictureordiagram).Eachanalogywasmanually

adaptedfromitsoriginalformtosimplifiedEnglishandaCogSketchsketch.Theinterpretationmodel

takesasinputatextfileandasketchfile,andproducessymbolicknowledgetorepresenttheanalogy

describedinthosefilesandinformationaboutthenewlylearnedtargetdomain.Theinterpretation

processoccursinfivesteps:sketchinterpretation,textinterpretation,multimodalintegration,case

50extraction,andmapping.Eachofthesestepsisoutlinedbelow,afteradescriptionofthebackground

knowledgethatisgiventothesystemindependentoftheindividualanalogies.

4.2.1. BackgroundKnowledge

AsillustratedinChapter3,instructionalanalogiesdependonbackgroundknowledgeabouteveryday

conceptsorexperiences.Thisgeneralknowledgecanberepresentedusingtype-levelrulemacro

predicatesinCycandmodelfragments.Cychasmanyrulemacropredicatesbuiltin(examplesshownin

Table1).However,outofthebox,theydonotprovidesufficientcoverageforthetypesofbasic

Figure5:Overviewofthesystem.Theinputtothemodelisasketch(i.e.CogSketchsketchfile)andatextfilewithasimplifiedEnglishdescriptionoftheanalogy.Lightblueitemsrepresentsystemsorknowledgesourcesthatpre-datethisthesis.Darkblueitemsindicatenewtechniquesorknowledgesourcesdevelopedforthisthesis.

51knowledgethatwererelieduponfortheinstructionalanalogiesintheFARguide.Tosupplement

existingknowledgeinCyc,rulemacropredicatesliketheonesshowninTable1weremanuallywritten

andaddedtotheknowledgebase.Alloftherulemacropredicatestatementscapturedknowledge

aboutthebasedomainonly.Forexample,inananalogythatcomparesacelltoacity,background

knowledgeaboutcities(e.g.thatallcitieshavemayors)isimportedintothesystem’sinterpretationof

theanalogy,eventhoughsuchknowledgemaynotbeexplicitlystated.Thebackgroundknowledge

statementsdonotrepresentrulesthatareuniversallytrue.Thatis,itispossiblethattherearecities

withoutmayors,butforthepurposesofinterpretingscienceanalogies,thesestatementsareaccepted

aspartofthebasicknowledgethatisassumedtobeknownbythelearner.

Thesecondsourceofbackgroundknowledgewasrepresentedwithmodelfragments,whichare

usefulforrepresentingcommonsenseknowledgeaboutprocessesandscenarios.Aswasthecasefor

rulemacropredicates,modelfragmentswereonlyusedtorepresentknowledgeinthebasedomain.

Modelformulationhasbeenusedpreviouslywithsketches(Changetal.,2014;Klenketal.,2011),andI

extendedthe2-dimensionalmechanicsdomaintheoriesinChangetal.(2014)withthosedescribedin

Table6.TheanalogiesintheFARguidedidnotalwaysinvolvequalitativeprocessesandquantities,but

whentheydid,themodelfragmentdefinitionsinTable6wereneededtocapturethephenomenabeing

described.Forexample,theanalogydepictedinFigure7requiredaveryspecifictypeofsituation:

walkingupanescalatorthatismovingdown.Thisspecificscenarioisintendedtoteachstudentsabout

homeostasis.Althoughthisscenarioisveryspecific,themodelfragmentusedtodescribeitismadeup

oftwomoregeneralpurposemodelfragmentsformovement.Theescalatormovingdownis

representedwithaninstanceofamotionprocess,whilethepersonmovingupisrepresentedwith

anotherinstanceofamotionprocess.Whenthesetwoprocessesareactiveundertherightconditions

(i.e.thepersoniswalkingup,theescalatorismovingdown,andthepersonisontheescalator),thenthe

52morespecificmodelfragmentisalsoactive.Tointerprettheoneoftheanalogiesaboutelectricity,the

systemusedamodelfragmentthatdescribedliquidflowingoutofacontainer.Tointerpretananalogy

aboutecosystems,thesystemusedamodelfragmentthatdescribedaphysicalroleplayscenario,

wherestudentsformawebmadeofstring.

Backgroundknowledge,whetherfromrulemacropredicatesormodelfragments,provide

implicitknowledgeaboutthebasedomain.Wheninterpretinganalogies,thisimplicitbackground

knowledgeisusefulbecauseitallowsbasedomainknowledgetobeprojectedontothetargetdomain

evenifthatbasedomainknowledgeisnotexplicitlystated.

4.2.2. SketchInterpretation

SketchinterpretationinvolvedanalyzingtheCogSketchsketchfileassociatedwithananalogyand

performingspatialandconceptualreasoningtoextractrelevantfactsfromthesketch.Thefacts

includedacombinationofinformationautomaticallycomputedbyCogSketch(i.e.topologicaland

positionalrelations)alongwithinferencesthatweregeneratedfromvisualconceptualreasoning,

Analogy ModelFragmentType QuantitiesEscalator/Homeostasis Walkingupanescalator

thatismovingdown(compositional;moregeneralmotionprocessesareparticipantstothisfragment)

Height/Elevation,speedofwalker,speedofelevator

Voltage/Pressure Waterflowingoutofacontainer

Pressure,depth,flowrate

Ecosystem/Web Astudentwebroleplayingscenario;movementwithinthewebincreasesinstability.

Stability,speedofmovement

Table6:Modelfragmentsusedforinterpretinganalogies.

53qualitativemodeling,andmulti-statereasoning.Theextensionsthatwereusedtoenablethesethree

typesofreasoningaredetailednext.

4.2.2.1 Extensionstovisualconceptualrelations

ThefirstextensionrequiredwritingrulestobridgespatialrelationsthatarecomputedbyCogSketch

withontologiesinCyc.BecauseCogSketchusesaconceptuallabelinginterface,asignificantamountof

conceptualknowledgecomesdirectlyfromtheuser.Thespatialrelationsthatarecomputedby

CogSketch(automaticallyandon-demand)arecombinedwithconceptualinformationtomakenew

inferences,whicharereferredtoasvisual-conceptualrelationsbecausetheirmeaningdependson

spatialandconceptualinformation(Forbusetal.,2005).Therearetwocategoriesofvisual-conceptual

relationsthatIimplemented:partrelations,andrelationsresultingfromdrawnarrows.

Partrelationsprovideageneralwayforrepresentingthingsandtheirparts.Partscanbephysicalorconceptual.Table7showsthenewvisualconceptualrelationsthatweredevelopedtosupportpart-whole

reasoning.Theinferencechainforpartrelationsstemsfromtopologicalrelations,whichare

automaticallycomputedbyCogSketch,andbyglyphgrouping,whichisinitiatedbytheuser.Proper

partrelationsinthercc8vocabulary(Cohnetal.,1997)indicatewhenoneobjectcontainsanother.If

Relation SpatialAntecedents ConceptualAntecedentsvisualSubsetOf Hollowcontainment SetsorCollections,

Containers,EventsvisualSubsetOf Glyphgrouping None

possessiveRelation Visualsubset None

physicalParts Visualsubset Partiallytangibleentities

subsetOf Visualsubset SetsorCollections

subEvents Visualsubset Events

Table7:Newvisualconceptualpartrelations.

54thecontainingglyphrepresentsaset,container,orevent,thenthecontainedglyphisa

visualSubsetOfthecontainerglyph.Forexample,ifaglyphofabottlecontainsaglyphofwater,then

thewaterglyphisavisualSubsetOfthebottleglyph.Glyphgroupingcanalsoindicatevisualsubsets,

althoughunliketopologicalrelations,glyphgroupingisanexplicituseraction,soitcanbeusedtoinfer

visualSubsetOfnomatterwhattheconceptuallabelsare.Forexample,thecelldrawninFigure3uses

glyphgroupingtoindicatethatwhenthoseglyphsaregroupedtogether,theyrepresentacell.Inthat

case,eachofthegroupedglyphsisavisualSubsetOftheglyphthatrepresentsthecell.Thevisual

subsetrelationisaspatialantecedentfortherestofthepartrelations,whichcanbeinferreddepending

ontheconceptuallabelsoftheglyphs.IftheglyphsareinstancesofthePartiallyTangiblecollection

inCyc,thenthephysicalPartsrelationshipholds.Iftheglyphsaresetsorcollections(i.e.instancesof

SetOrCollectioninCyc),thenthesubsetOfrelationholds.Iftheglyphsareevents(i.e.instancesof

EventinCyc),thenthesubEventsrelationholds.Possession(possessiveRelation)isthemost

generalofthepartrelations,holdingforsets,collections,containers,andanythingthatisinvolvedina

glyphgroupingaction.Forexample,becausethecellnucleusinFigure3isavisualSubsetOfthecell,

therelation(possessiveRelation the-cell the-nucleus)holds.TheuseofpossessiveRelation,

ratherthanorinadditiontomorespecificrelationslikephysicalParts,matchesthewaysuch

relationshipsareoftenexpressedinnaturallanguage,e.g.“Thecellhasanucleus.”Thisstatement,

althoughimprecise,isalsoatalevelofabstractionthattranscendsdomain,whichisusefulfor

interpretinginstructionalanalogies.

Arrowsareusedinavarietyofwaysinsketches.Sometimestheyareusedtoindicatemotion,

e.g.apersonwalkingupanescalator.Sometimestheyareusedtoindicatethetransferofenergyora

changeinstate,e.g.abatterychangingfromchargedtodead.Sometimestheyareusedtoindicate

causality,e.g.onestatecausesorleadstoanother.InCogSketch,eacharrowhasaconceptuallabel,but

55usually,additionalinformationcanbeinferredbasedonwhatthearrowrepresentsandhowitis

oriented.CogSketchalreadyinterpretsarrowsthatarelabeledasdirectionofmovementannotations

andarrowsthatarelabeledasexplicitrelations,butpriortothisthesisitdidnotreasonaboutarrowsas

events.WithintheCycontology,neo-Davidsonianrepresentationsareusedtoindicatethematicrole

fillers,andwecanthusmakehypothesesabouttherolesfilledintheseeventsbasedonhowthearrow

isdrawn(e.g.Figure6).Myextensiontothesecapabilitieswastowriterulesthatcapturethemeaning

oftherolerelationsinTable8.

Thesetofvisualconceptualrelationsdescribedherearebynomeanscomplete,buttheyserve

asavocabularyforarrowsindiagrammaticrepresentationsthataresufficientforachievingtheresults

describedbelow.

EventType InferredOnDemandCreationevent (doneBy arrow-object arrow-src)

(outputsCreated arrow-object arrow-dest) (products arrow-obj arrow-dest)

Destructionevent (doneBy arrow-obj arrow-src) (inputsDestroyed arrow-obj arrow-dest)

MakingSomethingAvailable

(doneBy arrow-object arrow-src) (objectActedOn arrow-object arrow-dest) (transferredObject arrow-object arrow-dest)

Usinganobject (doneBy arrow-object arrow-src) (instrument-generic arrow-object arrow-dest)

Intrinsicstatechangeevent

(objectOfStateChange arrow-obj arrow-src) (toState arrow-obj arrow-src)

Table8:Neweventrolerelationsbasedonarrowlabelandarrowdirection.Foreachrow,ifthereexistsanobjectnamedarrow-objectthatbelongstothespecifiedeventtypeandpointsfromarrow-srcandtoarrow-dest,thenthespecifiedfactscanbeinferredon-demand.

56

4.2.2.2 QualitativeModeling

Usingthedomaintheoriesfor2Dmechanics(Changetal.,2014)withextensionsshowninTable6,

modelformulation(Friedman&Forbus,2011)wasusedtomodelanalogiesthatinvolvedqualitative

causalrelationships.Threenewmodelfragmenttypesweredefinedforthissystem(Table6)andthey

arepartofthegeneralbackgroundknowledgeaboutthebasedomain.Todetermineifqualitative

modelingwasnecessaryforaparticularanalogy,thesystemcheckedforthepresenceofquantitiesin

theanalogy.Iftherewereanyquantities,thesystemsearchedformodelfragmentsthatmightapplyto

thesketchedscenario.Therelevanceofaparticularmodelfragmenttypewasdeterminedbyits

constraintsandconditions.Foranyactivemodelfragments,thesystemassertsconsequencesofthe

modelfragment,whichtypicallyindicatecausalrelationshipsbetweenquantities.Forexample,inthe

homeostasis/escalatoranalogy,oneoftheinferencesisthatthereisapositivequalitative

proportionalitybetweentheheight(i.e.level)ofthepersonwalkingandtheirspeed,aswellasa

negativequalitativeproportionalitybetweentheheight(i.e.level)ofthepersonwalkingandthespeed

Figure6:Anillustrationshowinghownewvisualconceptualrelationsalloweventstobedrawnasarrows.Thetoparrowrepresentsusingabattery.Thebottomarrowrepresentschargingabattery.

Given:(isa top-arrow

IntrinsicStateChangeEvent) (isa bottom-arrow

IntrinsicStateChangeEvent) Availableondemand:(objectOfStateChange top-arrow

charged-battery) (toState top-arrow

drained-battery) (objectOfStateChange bottom-arrow

drained-battery) (toState bottom-arrow charged-

battery)

57oftheescalator(Figure7).Thereisalsoaqualitativecorrespondencethatindicatesthatwhenthe

speedsofthepersonandtheescalatorareequal,therateofchangeoftheperson’sheightiszero.

Theserelationshipsarepartofthetargetknowledgeforthisanalogy.Qualitativemodelingwasusedto

interpretthreeoutofelevenanalogiesinthecurrentdataset.

4.2.2.3 InterpretingMulti-stateProcesses

Manyoftheanalogiesdescribedprocesses.Inoneoftheanalogiesexamined,thediagramdepictedthe

processusingmultiplestates(Figure1,Figure8).Theenzyme/keyanalogydepictsthreestatesinthe

targetdomain:(1)theenzymeandsubstrateareindependent,(2)theenzymebindswiththesubstrate,

(3)theenzymebreaksthesubstrateapartandunbinds.

(qprop ((QPQuantityFn Height) ?person) (RateFn ?person-motion)) (qprop- ((QPQuantityFn Height) ?person) (RateFn ?esc-motion))) (qpCorrespondence

(RateFn ?person-motion) (RateFn ?esc-motion) (QPDerivativeFn ((QPQuantityFn Height) ?person)) Zero))

Figure7:Thesketchassociatedwithananalogyusedtoexplainhomeostasis.Thepersoniswalkingupanescalatorthatismovingdown.Themodelfragmentthatdescribesthisscenariohasthethreeconsequencesshownbelowthesketch.Thefirststatementmeansthatthefasterthepersonwalks,thegreatertheirheight.Thesecondstatementmeansthatthefastertheescalatormoves,thelessertheperson’sheight.Thethirdstatementrepresentsaquantitycorrespondence:whenthepersonandescalatormoveatthesamespeed,theheightofthepersonisnotchanging.

58

Thefirstchallengeofrepresentingmulti-stateprocesseshastodowithcontextualizing

knowledgecorrectly.CogSketchalreadysupportspropercontextualizationthroughsubsketches.In

Figure8,eachpurpleboxisanindividualsubsketchwithitsownlogicalreasoningenvironment.This

allowsthesketchtorepresentthesameobjectunderdifferentconditions.

Thesecondchallengeariseswhentherearerelationshipsbetweensubsketchesthatneedtobe

generalizedtoanentireconcept.Becauseinstructionalanalogiesatthisageleveltendtofocuson

qualitative,type-levelknowledge,itisimportantthatrelationshipscanbegeneralizedtoanentire

concept,ratherthanjustanindividualinstance.Inthiscase,thearrowsrepresentenablement

relationshipstoconveythatonestateleadstotheotherfromlefttoright.Thepurposeofthisanalogy,

however,isnotjusttolearnonthissinglesituation,butonothersituationslikeit,i.e.enzymeactivation

events.Conceptuallabelsprovidesomeadditionalinformation:thefinalstateislabeledasabreaking

Figure8:TheCogSketchsketchfortheenzyme-keyanalogy.Eachpurpleboxisanindividualsubsketch,whichisusedtorepresentanindividualstate.Thearrowsbetweenstatesareenablementrelationstoindicatetheorderofthesequence.

59event(becausethesubstratemoleculebreaksapart).Thefirsttwostatesaresimplysub-eventsofan

enzymeactivationevent.Iftheenablementrelationshipsweregeneralizedusingonlythisinformation,

wewouldgetunderspecifiedtype-levelrelationships,e.g.“subeventsofenzymeactivationsenablesub

eventsofenzymeactivation.”Abetterapproachwouldbetoanalyzethesubeventstounderstand

whatmakesthemdifferentfromeachother.

Thesystemdoesthisviastructure-mapping.Eachsubsketchiscomparedtotheotherstolook

forpropertiesthatareuniquetoeachstate.Thisallowsthemodeltodescribeeachstatemore

precisely.Morespecifically,foreachstate,themodelfindspairwisecandidateinferencesbycomparing

ittoeachoftheotherstatesusingthestructure-mappingengine.Eachcandidateinferenceisafactthat

istrueinthebasestate,butnotthetargetstate.Eachreversecandidateinferenceisafactthatistrue

inthetargetstate,butnotthebasestate.Thesefactscanbegeneralizedintoexistentialrulemacro

predicates(liketheonesshowninTable1)thatqualifythecollectionsofeachstate.Forexample,when

comparingthefirststatetothesecond,wherethefirstisthebaseandthesecondisthetarget,thereis

areversecandidateinferencethattheenzymeintersectswiththesubstratemolecule.Thisisanother

wayofsayingthatanimportantdifferencebetweenthefirststateandthesecondstateisthatthe

enzymeintersectswiththesubstrateinthesecondstate.Themodeltranslatesthisintoapropertyof

thefirststate:theenzymeandthesubstratedonotintersect.Thisfactcanthenbegeneralizedintoa

type-levelrulemacropredicate,whichstatesthattheredoesnotexistanenzymethatintersectswitha

substratemolecule.Thisrulemacropredicatebecomesageneralpropertyofthefirststate.Now,

insteadofknowingonlythatthefirststateisasub-eventofenzymeactivation,themodelnowknows

thatthefirststateisasub-eventofenzymeactivationwheretheenzymeisnotincontactwiththe

substratemolecule.Byqualifyingthecollectionsofallthreestates,theenablementrelationsbecome

moremeaningful.Wheretheyusedtoexpressthatoneeventenabledanother,itcannowexpressthat

60astatewherethemoleculesarenotincontactenablesastatewheretheyareincontact(whichinturn

enablesthefinalstatewherethesubstratehasbrokenapart).

4.2.3. TextInterpretation

TextinterpretationwascompletedusingtheexistingEANLUpipeline,withonenewapproachfor

disambiguation.EANLUhasseveraloptionaldisambiguationheuristicsthatguidetheselectionof

interpretationschoices.Thetwoheuristicsthatareusedtomakedisambiguationchoicesfor

instructionalanalogiesarefavoredcontextandinformationgainheuristics.Thefavoredcontext

heuristictakesasinputamicrotheoryname,whichcanbeusedasaframeofreferenceforinterpreting

thecurrentreading.Forinstructionalanalogies,theonlyinputtothefavoredcontextheuristicisthe

nameofthemicrotheorythatcontainstheinformationstoredinthesketch.SinceglyphsinCogSketch

haveconceptuallabelsthataretakendirectlyfromtheCycknowledgebase,theyprovidedirect

evidenceforparticularinterpretationchoices.Givenaninterpretationchoice,thefavoredcontext

heuristiccountsthenumberofcollectionsandpredicatesinthechoicethatarepresentinthesketch.

Thenumberofcommoncollectionsandpredicatesisthepreferenceweightassignedtothatchoice.

Thismeansthatchoicesthatinvolvemorethingsthatarealsomentionedinthesketcharegiven

preference.Alloftheseheuristicspredatethisthesis,exceptfortheuseofpredicates(inadditionto

collections)inthefavoredcontextheuristic.Lastly,IusedBarbella’s(Barbella&Forbus,2011)

analogicaldialogueactinterpretationtodetectwhentheanalogyisbeingintroduced,when

correspondencesarebeingintroduced,andtodetectwhichkeyelementsarepartsofthebaseand

targetdomainrespectively.

614.2.4. MultimodalIntegration

Oncetheinitialsketchandtextinterpretationiscomplete,thetworepresentationsneedtobemerged

totakeadvantageofinformationfrombothsources.Thisproblemcanbecharacterizedasanalignment

problem(Lockwood&Forbus,2009).IextendedtheapproachusedinLockwood’sworkbyusing

languagethathasbeenautomaticallydisambiguatedandbyusingeventinterpretationstoprovide

supportforaccuratealignments.

Structure-mappingisusedtoaligntheknowledgegatheredfromtheanalogy’stextandsketch

sources.Aligningthesetworepresentationscanbecharacterizedasaverynear,within-domain

analogy.Thetwodescriptionsareliterallydescribingthesamethings.Usingpartitionconstraintsforall

collectionsisawaytoensurethattermscanonlymaptoeachotheriftheybelongtothesame

collection(s),forexample,keysmatchwithkeys,lockswithlocksandsoon.Thisworksverywellwhen

theconceptuallabelsinthesketchareidenticaltotheinterpretationsgeneratedbyEANLU.However,

evenwhenthereisconsiderableoverlapinconceptualinformation,eachmodalitytypicallyincludes

uniqueinformation.Forexample,theenzyme/keyanalogyshowninFigure1(withCogSketchsketch

showninFigure8)hassubstantialoverlapinconceptualinformationbecauseeachglyphinthesketch

hasalabel(EnzymeMolecule,SubstrateMolecule,EnzymaticBindingSite,etc.).Theselabels

guaranteethattheenzymemoleculementionedinthetextmodalitywillalignwiththeenzymeshown

inthesketchbecausepartitionconstraintsareused.However,thereareseveralthingsthatare

mentionedinthetextthatarenotpresentinthesketch,e.g.thebindingsite’suniquechemicalmakeup,

thekey’suniqueshape,thefactthatenzymescanbereused,etc.Similarly,thereisinformationinthe

sketchthatisnotpresentinthetext,e.g.thespatialintersectionoftheenzymeandsubstrate,the

sequentialnatureofenzymeactivation,etc.Thedifferencesbetweenthetwomodalitiesmeanthat

thereareopportunitiesformismatches.Thisisespeciallytrueforanalogiesthatinvolveconceptsthat

62arenotexplicitlydrawn,e.g.energy.Thesystemreducesthechancesformismatchesbyreasoning

abouteventsinthetextinterpretationandprojectingthatinformationtothesketchmodality.

Genericeventinterpretationtakestheneo-Davidsonianeventinterpretationsgeneratedby

EANLUandprojectsthemtothesketchmodality(Figure9).Itisbasedontheassumptionthatwhen

instructionalanalogiesdescribeevents,theydescribepossiblegenericrolesfortheprimaryparticipants

oftheevents.Forexample,inthecell/cityanalogy,thesentence“Powerstationsprovideelectricity”is

interpretedasastatementaboutthefunctionsofpowerstations.Usingtheneo-Davidsonian

representation,sucheventwouldberepresentedlikethis:

(isa provide34501 MakingSomethingAvailable)

(performedBy provide34501 power-station88374)

(objectActedOn provide34501 electricity18186)

Toprojectthisinformationintothesketchmodality,thesystemlooksattheprimaryactoroftheevent:

power-station88374.ThisentityisaninstanceofthecollectionPowerGenerationComplex.The

systemsearchesforinstancesofthiscollectioninthesketch,andifitfindsone,searchesforinstancesof

MakingSomethingAvailableandofElectricity.Inthecaseofthecell/cityanalogy,thereisaninstance

ofWindmill,whichisasubcollectionofPowerGenerationComplex,buttherearenoinstancesof

MakingSomethingAvailableorElectricity.Thesystemthengeneratestwospeculativeentities:an

eventthatisaninstanceofMakingSomethingAvailableandanentitythatisinstanceofElectricity.

Thesystemassertsrolerelationsbetweentheentities,resultinginthefollowingrepresentationinthe

sketchmodality:

(isa (SpeculativeThingFn event52582) MakingSomethingAvailable)

63(performedBy (SpeculativeThingFn event52582) windmills)

(objectActedOn (SpeculativeThingFn event52582) (SpeculativeThingFn electricity52843))

Thesystemalsogeneratesstatementsthatdescribetheeventswithmorespecificcollections,similarto

thewaymulti-stateprocesses(describedabove)areinterpreted.Becausethegoaloftheseanalogiesis

oftentobuildtype-levelknowledge,denotingcollectionswithspecifictermscanbeuseful.For

example,thecell/cityanalogyconveysfunctionsandbehaviorsofthepartsofthecell.Theinitial

languageinterpretationsuseverbstodetectwhatkindsofeventsarebeingdescribed.Usingpower

stationsasanexample,theverb“providing”indicatesthattheeventbeingdescribedisaninstanceof

MakingSomethingAvailable.Therolerelationsprovidefurtherinformation,andthegenericevent

Figure9:Anillustrationofhowgenericeventinterpretationworks.Eventinformationfromthetextmodalityisassumedtoapplytoelementsinthevisualmodality.Providingelectricityisprojectedontowindmillsbecausetheyarespecializationsofpowerstations..

64interpretationprocedureusescollectiondenotingfunctions(describedinsection2.2)toindicatethat

theelectricity-providingeventisaninstanceofamorespecificcollection:

(isa provide34501 (MakingAbstractAvailableFn Electricity))

Similarly,thisinformationisprojectedintothevisualdomainaswell:

(isa (SpeculativeThingFn event52582) (MakingAbstractAvailableFn Electricity))

Thesecollectionmembershipstatementsprovidegreaterstructuralsupportforputtingwindmillsand

powerstationsintocorrespondence.Thisisparticularlyusefulinthiscasebecausetheconceptual

attributesfortheseentitiesdonotmatchexactly,soitisnotguaranteedthattheywillcorrespondbased

onpartitionconstraintsalone.Lastly,theuseofcollectiondenotingfunctionsisalsousefulfor

representingtype-levelinformation.Ratherthanmakinggeneralassertionsabouteventswhere

somethingismadeavailable,thesystemcannowmakegeneralassertionsabouteventswhere

electricityismadeavailable.

Theanalogicalmappingbetweenthetextinformationandsketchinformationdetermineshow

theinformationismerged.Correspondingentitiesandfactsaremergedtogetherandallfactsthatwere

notinthemappingareincludedinthefinalrepresentationasis.Theadvantagetousingalignmentand

merging(ratherthanasimpleunionoffacts)isthatitallowsalignedentitiestorefertothesame

conceptualthing.Forexample,whencombiningfactsaboutthecell/cityanalogy,wedonotwanta

multimodalrepresentationwheretherearetwonuclei.Thenucleusthatismentionedinthetextis

assumedtobethesamenucleusthatisdrawninthesketchbecausethetwonucleicorrespondtoeach

otherintheanalogicalmapping.Consequently,thefinalrepresentationcontainsknowledgeaboutthe

nucleuscontrollingthecell(afactwhichcomesfromthetext)andaboutthenucleusbeingaproperpart

ofthecell(afactwhichcomesfromthesketch).Includingfactsthatwerenotinthemappingisalso

65important,becauseitallowsthesystemtoincludefactsthatarepresentinonemodalitybutnotthe

other.

4.2.5. CaseExtraction

Caseextractionistheprocessbywhichthebaseandtargetdescriptionsarebuiltfromtheintegrated

multimodaldescription.Caseextractionusesthreesourcesofinformationtosplitfactsinthe

multimodaldescriptionintoabaseandtarget:analogicaldialogueacts,eventrolerelations,andpart-

wholerelationships.Theseinformationsourcesareusedtobuildsetsofentitiesthatarehypothesized

tobecoreelementsofthebaseandtargetdomainrespectively.Thesesetsarethenusedinaniterative

searchoverthemultimodaldescription.Analogicaldialogueactsdeterminetheinitialmembersofeach

entityset.IfananalogicaldialogueactintroducesacorrespondencebetweenitemsBandT,thenBis

anelementofthebasedomainandTisanelementofthetargetdomain.Foreachelementidentifiedin

analogicaldialogueacts,thesystemalsoincludesanysubpartsorsubobjectsofthatelement.Sub

objectsaredeterminedbypossessiverelations(describedinsection4.2.2.1),whichcomefromnatural

languageinterpretation(e.g.“ATPhasenergy”)andfromvisualconceptualrelations.Foreachofthese

entities(andtheirsubobjects),thesystemsearchesforeventsthatarerelatedtothem.Eachevent

alongwithanyotherparticipantsofthatevent,areincludedintheentitysetaswell.Usingtheseentity

sets,themodelproceedswithaverysimplevotingheuristicthatclassifiesanyremainingfactsas

belongingtothebasedescription,thetargetdescriptionorasbeingambiguous,basedonnumberof

knownbaseortargetitemsmentionedintheexpression:

66

Ifafactinvolvesmorebaseentitiesthantargetentities,itisclassifiedasbeingafactaboutthebase

domain.Ambiguousstatements,whichinvolveanequalnumberofbaseandtargetentities,are

discarded.

4.2.6. Type-levelSummarization

AsnotedinChapter3,theanalogiesintheFARguideareaimedatteachingtype-levelconceptual

knowledge.Althoughdetectinggenericsandrepresentingthemaccuratelyisaknownproblemin

naturallanguageunderstanding,thismodeltakesanaggressiveapproachtogeneralization,and

assumesthatcertainpredicatesgeneralizeasfollows.Partpredicates(e.g.physicalpartpredicates,

subsetof)generalizewithauniversalquantifieroverthewholeandanexistentialquantifieroverthe

part.Forexample,asketchthatillustratesacellnucleusaspartofacellwillgeneralizetoarulemacro

predicatethatstatesthatforallcells,thereexistsanucleusthatisapartofit.Primaryactorrole

relationsgeneralizewithauniversalquantifierovertheprimaryactorandanexistentialquantifierover

theevent.Thisisespeciallyusefulforanalogiesthatfocusonbehaviorsandfunctions.Whenan

analogystates,forexample,thatmitochondriaprovideenergy,thisgeneralizestoafactthatstatesthat

allmitochondriaprovideenergy.Temporallycoexistingpredicates(i.e.binaryrelationsbetweennon-

classify-fact(fact, base-items, target-items)

base-count ß length(intersection(base-items, args(fact)))

target-count ß length(intersection(target-items, args(fact)))

IF base-count > target-count THEN RETURN :BASE

ELSE IF target-count > base-count THEN RETURN :TARGET

ELSE RETURN :AMBIGUOUS

67eventsthattemporallycoexist)aregeneralizedwithauniversalquantifieroverthefirstargumentand

anexistentialoverthesecond.Forexample,asketchthatillustratesthatacellisbiggerthananucleus,

willgeneralizetoarulemacropredicatethatstatesthatforallcells,thereexistsanucleusthatitis

biggerthan.FactsaboutinstancesarediscardedunlesstheyrepresentquantitiesorconstantsintheKB,

e.g.ZeroorPlanetEarth.Bysummarizingtheinformationinthebaseandtargetwithtype-level

predicates,thebaseandtargetdescriptionsbecomesuitableforatype-levelanalogy,where

correspondencesandinferencesaboutcollections(ratherthanstrictlyentities)canbemade.

4.2.7. MappingandInferenceEvaluation

TheinstructionalanalogyiscreatedusingSMEwithconstraintsthathavebeenidentifiedbyanalogical

dialogueactsandeventinterpretations.Aswithcaseextraction,analogicaldialogueactsprovidethe

startingpointfordeterminingwhatconstraintsshouldbeputonthemappingprocess.

Correspondencesthathavebeenintroducedinthetextareusedashardconstraintsonthematch.

Eventsthatthoseentitiesparticipateinandtheotherparticipantsofthoseeventsareusedasrequired

correspondencesaswell.However,sincetheanalogyisoperatingatthetype-level,thosematch

constraintsaretranslatedsothattheyarebetweenconcepts(i.e.collections)ratherthaninstances(i.e.

entities).Foragivenconstraintbetweentwoentities,AandB,thecollectionsthatthoseentitiesbelong

toareusedasconstraints.Usingthecell/cityanalogyasanexample,thesentence“Acellislikeacity”

representsadialogueactthatintroducesacorrespondencebetweenaninstanceofCityandan

instanceofCell.Thiscorrespondenceistranslatedtoarequiredcorrespondencebetweenthe

collectionsCityandCell.Thistranslationisnecessaryforcapturingthetype-levelsemanticsofthe

analogyandforenablinginferencesaboutcategoriesorclassesofthings,ratherthansingleexamplesof

them.

68 Onceaninitialmatchisconstructed,itsinferencesareevaluatedandthemodelattemptsto

resolveskolems.Becauseallthestatementsinthematcharetype-levelstatements,skolemscanbe

either:collections,predicates,orentitiesthatpassedtype-levelfiltering.Suchentitiesmaybeknown

termsintheCycknowledgebase,e.g.PlanetEarth,orqualitativequantities,e.g.Many-Quant.Abstract

entities,likequantities,andpredicatesareacceptedasis.Collectionsrequireadditionalinspectionto

determineiftheyareknowntobeveryabstractoriftheyinvolvefunctionalpredicatesthatwere

treatedasentitiesandwhoseargumentshaveknownmatches.Constantsandabstractcollectionswere

manuallyidentifiedbasedoninitiallanguageandsketchinterpretations.Thefollowingcollectionsare

consideredabstractandthereforeacceptableasisforskolemresolution:Event,FunctionalSystem,

Individual,PartiallyTangible,Set-Mathematical,SetOrCollection,SpatialThing,Thing.

ConstantsweredefinedasnumbersorinstancesoftheCyccollection:ScalarOrVectorInterval.

Instancesofthiscollectionincludethingslike:Now,Many-Quant,Few-Quant,Billions,etc.Theselists

arenotexhaustive,butsimplyastartingpointbasedoninputstothemodel.Theypointtooneofthe

difficultiesininterpretingcross-domainanalogiesthatwasmentionedinChapter3:representingterms

attherightlevelofabstractionandknowingwhenatermtranscendsdomain.

Iftherearecandidateinferencesthathavebeenfullyresolved,theyareacceptedasfactand

storedinthetargetdomain.Thematchcanthenbeextendedinanattempttoarriveatmore

inferences.Acceptinginferencesandextendingthematchmakesitpossibletobuildlargermappings

withhigherordercandidateinferences.Matchextendingcontinuesuntiltherearenomorecandidate

inferencesthatcanberesolved,oruntilamaximumdepthisreached(determinedbythemaximum

depthofexpressionsinthebasedomain).Whenmatchextensionandinferenceevaluationfinishes,the

finaltargetdescriptionisstoredintotheKBsothatitcanbeusedforfuturereasoning.

694.3 Evaluation

Toevaluatethismodel,Iadapted11analogiesfromtheFARguide:

1) RechargeablebatteryanalogyforATP2) Cityanalogyforacell3) Earthanalogyforcellcomponents4) Supermarketanalogyofabiologicalclassificationsystem5) Lockandkeyanalogyforenzymeaction6) Fluidmosaicanalogyforcellmembranes7) Homeostasisislikewalkingupadownescalator8) Webanalogyfortheinterdependenceoforganisms9) ClothespinanalogyforthestructureofDNA10) Watercircuitanalogyforelectriccurrent11) Waterpressureanalogyforvoltage

IntheFARguide,eachanalogyconsistsofanexplanationandatableofcorrespondences.Foreach

analogy,ImanuallytranslatedthelanguageintosimplifiedEnglishtosupportinterpretationwithEANLU

andsketchedaspatialrepresentationoftheanalogyintoCogSketch.Foreachanalogy,themodeltakes

asketchfileandatextfileasinputandproducesmicrotheorywhichcontainsnewlyacquiredfacts

aboutthetargetdomain.

4.3.1. ModelInput

Forbuildingthetextpassages,eachanalogyneededtobetranslatedfromthetableandtext

formatinthebooktocompletesentenceswithsimplesyntax.Sentencesthatintroduced

correspondenceswereusedtoaccountfortermsthatwerelistedonthesamerow.Forexample,the

enzyme/keyanalogyshowninFigure1hasthefollowingsimplifiedEnglishrepresentation:

Anenzymeislikeakey.Asubstratemoleculeislikealock.Thekeyhasridges.Theridgeshaveauniqueshape.Theenzymebindingsiteisliketheridges.Theenzymebindingsitehasauniquechemicalmakeup.Thekeyonlyunlocksspecificlocks.Theenzymereactsonlywithspecificsubstratemolecules.Thekeyunlocksthelock.Theenzymebreaksapartthesubstratemolecules.Afterunlockingalock,thekeyisunchanged.Thekeycanbeusedoverandoveragain.Enzymeactionislikeunlockinga

70

lock.Theenzymeisunchangedbythechemicalreaction.Theenzymecanbeusedoverandoveragain.

ThispassageislongerthanthetablerepresentationinFigure1becausethetablerepresentationinthe

FARguideisoftenmoreconcisethanapassagethatexplainstheanalogy.Forexample,somerowsonly

havethenamesofentitiesthataresupposedtocorrespondtoeachother.ForinputtoEANLU,these

weretranslatedintocompletesentences,e.g.“Anenzymeislikeakey.”Additionally,somelonger

sentencesweresplitintomultipleshorterones(butmorewordsoverall)toenableinterpretationin

EANLU.

Allbutoneofthe11analogiesevaluatedherewerepresentedwithaspatialrepresentationof

somekind.SevenanalogieswerepresentedintheFARguidewithspatialrepresentations(either

diagrams,sketches,orphotos)ofthebaseandtargetdomain(eitherseparatelyortogetherinablended

representation,asshowninFigure7).Forthoseanalogies,Isketchedthoserepresentationsinto

CogSketchandusedconceptuallabelingandglyphgroupingtoconveyaccurateconceptualinformation.

Asnotedearlier,CogSketchdoesnotperformsketchrecognition.Thekindsofthingsdepictedineach

sketcharegivenexplicitlythroughlabelsfromtheCycknowledgebase.ThisallowsCogSketchtoreason

abouttheentitiesdrawninthesketchwithrespecttotheCycontology.ItalsoallowsCogSketchto

reasonaboutpart-wholerelationshipsviaglyphgrouping.Forexample,intheenzyme/keysketch

(Figure8),theenzymeandenzymeactivationsitearetwoseparateglyphs,withlabels,andtheyare

groupedtoconveythattheactivationsiteispartoftheenzyme.Theconceptuallabelingandmanual

glyphsegmentation(andoptionalgrouping)thereforeprovideagreatdealofconceptualinformationon

theirown.Theremainingthreeanalogieswithspatialrepresentationsonlyhadpicturesforthebase

domain.Fortheanalogieswithonlybasespatialinformation,Isupplementedthesketchwithatarget

representationofmyownsothattherewasenoughconceptualinformationaboutthetargetdomainto

71assistwiththeoverallinterpretation(theablationexperimentsinthischapterwillillustratewhythiswas

necessarygiventhecurrentstateofthismodel).Lastly,oneoftheanalogies,thecell/cityanalogy,had

nospatialrepresentationsatall.Forthisanalogy,IsupplementedthetextpassagewithaCogSketch

sketchofacityandacell.Thefullsetoftextpassagesandsketchdescriptionscanbefoundin

APPENDIXA:InstructionalAnalogies.

Asanevaluationmeasure,Itestedthemodel’saccuracyonasetofgoldstandardqueriesthatI

wrotebasedonthepresentationoftheanalogiesintheFARguide.Essentially,thesequeriesaddress

thefollowingquestion:ifthemodelsuccessfullyinterpretedtheanalogy,whatshoulditknowaboutthe

targetdomain?TheexpectedknowledgewasmanuallytranslatedintoCycqueries.Table9showsthe

goldstandardqueriesfortheenzyme/keyanalogyinnaturallanguageandintheCycrepresentational

language.

72

4.3.2. ModelPerformance

Thefullmodelsuccessfullyprocessedall11analogies.

Enzymeactivesiteshaveauniquechemicalmakeup.(partTypes EnzymeMolecule (CollectionIntersectionFn (TheSet EnzymeBindingSite (ThingDescribableAsFn Unique-TheWord Adjective))))

Enzymesonlyreactwithspecificsubstrates.(relationAllExistsAndOnly chemicalReactants (SubcollectionOfWithRelationToTypeFn

(SubcollectionOfWithRelationToTypeFn ChemicalReaction chemicalReactants SubstrateMolecule) catalyst EnzymeMolecule)

(CollectionIntersectionFn (TheSet SubstrateMolecule (ThingDescribableAsFn Specific-TheWord Adjective)))) Enzymeactionbreaksapartsubstratemolecules (relationAllExists subEvents EnzymeActivationEvent (SubcollectionOfWithRelationToTypeFn

(SubcollectionOfWithRelationToTypeFn BreakingEvent doneBy EnzymeMolecule) objectOfStateChange SubstrateMolecule))

Theenzymecomesoutofthereactionunchanged. (relationExistsAll unchangedActors EnzymeActivationEvent EnzymeMolecule) Whenenzymesbindtosubstratemolecules,itenablesthesubstratetobreakapart.(enables-SitTypeSitType (SubcollectionOfWithRelationToFn (CollectionIntersectionFn (TheSet (SubcollectionOfWithRelationToTypeFn Situation subEvents ChemicalReaction) (SubcollectionOfWithRelationToTypeFn Situation subEvents EnzymeActivationEvent) Situation))

holdsIn (relationExistsExists objectsIntersect EnzymeMolecule SubstrateMolecule)) (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn BreakingEvent doneBy EnzymeMolecule) objectOfStateChange

SubstrateMolecule)) Enzymescanbereused.(relationAll repeatedEvent EnzymeActivationEvent) Table9:GoldstandardqueriesinnaturallanguageandintheCycrepresentationallanguagefortheenzyme/keyanalogy.

73Table10showsthesizesoftheanalogiesintermsofsentencesintextinputandnumberoffactsat

variousstagesofprocessing.Thetextpassagesvariedinsizebutrangedbetween4and18sentences.

Notably,themultimodalinterpretationisnotsimplyaunionofthefactsineachmodality.BecauseSME

isusedtoaligntherepresentations,entitiesaremergedandredundantfactsareavoided.Similarly,the

multimodalinterpretationisalsonotaunionofthebaseandtargetcases,sincethereareoftenfacts

thatcannotbeclassifiedasbelongingtoeitherthebaseortargetdomain(Table11).Afterfactsare

classifiedasbelongingtothebaseortargetdomain,theyaregeneralizedtotype-levelstatements.The

sizesofthebaseandtargetcasesaftertype-levelsummarizationareshowninTable12.

AnalogyName Sentencesintextinput

Factsintextinterp.

Factsinsketchinterp.

Factsinmultimodal

interp.ATP 15 181 121 258Cell(City) 15 160 231 321Cell(Earth) 5 38 165 163Classification 17 188 415 580Enzyme 15 121 187 292Membrane 9 98 74 144Homeostasis 10 97 75 150Ecosystem 4 42 250 254DNA 12 85 374 354Circuits 18 162 130 250Voltage 7 51 63 90

Table10:Summaryoftextandcasesizesfor9biologyanalogiesand2electricityanalogies.

74

Thefullmodelachievedanaverageaccuracyof83%acrossall11analogies.Ablation

experimentscomparedthefullmodeltothreeotherconditions:

• Fullmodelwithouttheuseofbackgroundknowledge

• Fullmodelwithoutthebasedomain

• Fullmodelwithoutthesketchedinput

AnalogyName BaseFacts TargetFacts AmbiguousFactsATP 87 124 47Cell(City) 153 128 40Cell(Earth) 58 36 69Classification 261 202 117Enzyme 127 125 40Membrane 54 83 7Homeostasis 66 60 24Ecosystem 60 164 30DNA 138 163 53Circuits 112 78 60Voltage 54 26 10

Table11:Numberoffactsclassifiedasbelongingtothebasedomain,belongingtothetargetdomain,andbeing

ambiguous.

AnalogyName Type-levelBaseFacts Type-levelTargetFactsATP 54 77Cell(City) 86 88Cell(Earth) 31 8Classification 108 47Enzyme 45 115Membrane 13 35Homeostasis 38 38Ecosystem 15 228DNA 23 83Circuits 47 27Voltage 23 9

Table12:Numberoffactsinbaseandtargetdomainsaftertype-levelsummarizationandanalogicalmapping.

75

Removingtheabilitytoimportqualitativebackgroundknowledgeintheformofrulemacro

predicatesandqualitativemodelfragmentscausedoverallaccuracytodropto59%(Figure10).This

meansthatbackgroundknowledgeplaysanimportantroleinthismodel’sabilitytobuildnew

knowledgeaboutthetargetdomain.However,at59%,thisresultalsomeansthatmuchofthe

knowledgethatwastestedinthegoldstandardquerieswaseitherexplicitlystatedintheanalogy’stext

orsketch,wastransferredfromexplicitlystatedbaseknowledge,orwasinferredduringthevisual

conceptualelaboration.Thisistrueforanalogiesthatdealwithpart-wholerelations,likethecell/earth

analogy,wherethepart-wholerelationsaregivenexplicitlyinthesketchviarelationarrows,orthe

cell/cityanalogywherepart-wholerelationsareinferredfromglyphgrouping.However,asnotedin

Chapter3,analogiesthatarelessverbosemightrequirelearnerstofillthegapswithbackground

knowledge.Thecell/earthisonesuchanalogy,anditsaccuracydropsfrom100%withthefullmodelto

72%whennobackgroundknowledgeisused.Thissuggeststhatwhenitcomestotheimportanceof

backgroundknowledge,thereisvariationwithinthissetofanalogiesandthatthepresentationofthe

analogy(completesentencesvs.correspondencesonly)maybeacluethatbackgroundknowledgeis

important.Asexpected,theperformancewasworseinthenobackgroundconditionforthethree

analogiesthatreliedonqualitativemodelingforinterpretation(homeostasis,ecosystem,andvoltage).

76

Thesecondablationconditiondisabledtheanalogyentirely,makingitbothimpossibletouse

backgroundinformationfromthebasedomainandexplicitinformationaboutthebasedomain.This

Figure10:Resultsfromablationexperimentsbyanalogy.Theboldlinerepresentsaverageaccuracyacrossallanalogies.

AnalogyName #GoldStandardQueries

FullModel(#correct)

NoBackground(#correct)

NoBase(#correct)

NoSketch(#correct)

ATP 11 9 8 4 3Cell(City) 12 10 7 5 4Cell(Earth) 11 11 8 4 0Classification 11 4 1 5 0Enzyme 6 5 4 4 0Membrane 6 6 6 4 2Homeostasis 6 5 2 2 2Ecosystem 10 10 9 6 0DNA 12 10 10 10 0Circuits 6 4 4 3 0Voltage 1 1 0 0 0Table13:Performanceofindividualanalogiesoneachofthefourexperimentalconditions.

77effectcausedagreaterdropinaccuracy,to47%.Thisindicates,asexpected,thatbackgroundand

explicitknowledgeaboutthebasedomainisimportantforinterpretingtheanalogy.Nevertheless,a

considerableamountoftargetknowledge(almosthalf)isstillexplicitlystatedintheanalogy.

Interestingly,theperformanceononeoftheanalogies(supermarketanalogyforclassification)

improvedinthenobasedomaincondition.Thisanalogywasunliketheothersinthatitdidnothavea

baseandtargetdomainwhereindividualentitiescouldbealignedtoeachother.Theanalogycompares

theorganizationalstructureofasupermarketwiththeorganizationalstructureofbiologicaltaxonomies.

Intheotheranalogies,entitiesweretypicallyintroducedindividuallyinthebaseandtargetdomain(e.g.

theenzymeislikeakey;DNAislikeaclothespinstructure)andthereforerepresentedindividuallyinthe

textandsketchrepresentations.Intheclassificationanalogy,however,groups(i.e.speciesandgenus)

andsections(e.g.dairysection,producesection)arerepresentedwithmultiplediagramelements,

makingaclean1:1mappingdifficult.Notably,thefullmodelperformedtheworstonthisanalogywhen

comparedtoalloftheothers(4/11queriescorrect;36%).Thisisonecasewhereperhaps

rerepresentationisneededtoaccuratelyinterprettheanalogy.

Thefinalconditionremovedthesketchentirelyfromtheprocess.Backgroundknowledgeand

explicitbaseknowledgewerestillusedinthiscondition;itjustreliedentirelyonthenaturallanguage

input.Themodelperformedatitsworstinthisconditionwithanaccuracyof12%.Thisindicatesthatin

thismodel,theconceptuallabelsprovidedinthesketchplayacriticalroleininterpretation.The

conceptuallabelsfromthesketchcomedirectlyfromtheCycontology,meaningthatthereisadirect

connectionbetweenthelabelsandthesemanticinterpretationchoicesprovidedbyEANLU.Thus,the

conceptuallabelsfromthesketcharemostlyresponsibleformakingtherightdisambiguationchoices.

784.4 Discussion

Theexperimentsinthischaptersupportclaim1ofthisthesis.Theydemonstratethatacombinationof

sketchedinput,simplifiednaturallanguageunderstanding,andstructure-mappingcanbeusedto

capturequalitativescienceknowledge.Structure-mappingisusedintwomainways.First,itisusedto

mergesketchandtextrepresentations.Itisalsousedtobuildthecross-domainanalogy,whichis

effectiveattransferringknowledgefromthebasetothetargetdomainbasedontheablationconditions

wherethebackgroundknowledgeandthebasedomainwereexcluded.Asmentionedearlierinthis

chapter,structure-mappingisalsousedtodetectimportantdifferencesbetweenstatesinasketch,but

thiswasonlyrelevantforoneoutofthe11analogies.

Theablationexperimentsalsosupporttheclaimthatbackgroundknowledgeandvisual

representationsplayanimportantroleinterpretinginstructionalanalogies.Thequalitativeanalysisin

Chapter3alsosupportsthisclaim.

Oneimportantlimitationofthismodelishowbasedomainandbackgroundknowledgeisused.

Theablationexperimentsindicatedthatoutofthethreesourcesofinformation(background

knowledge,explicitbaseknowledge,sketchknowledge),sketchknowledgehadthegreatestimpacton

interpretationaccuracy.Thisresultcannotbeusedtosuggestclaimsaboutthewaypeopleinterpret

analogies.Thereasonforthishastodowithbroaderchallengesinknowledgerepresentation.Fora

symbolicreasoningsystem,suchastheonebuiltforthisthesis,thereisnodifferencebetweenasserting

afactandhavingthesystemunderstandit(asfarasunderstandmeansinthissystem).Inthismodel,

thereisnodifferencebetweenstoringafactandlearningitbecauseofthelimitsoftheevaluation.To

testthesystem’sknowledge,itissimplyqueried,whichistheequivalentofhavingsomeonereciteafact

backtoyou.Betterevaluationsofknowledge,suchashowitisusedinothertypesofreasoninglike

79questionanswering,wouldlikelyprovideamoreaccuratemeasureofhowwellasystemunderstandsa

conceptandthereforehowimportantbackgroundknowledgeis.

Anotherlimitationistherelianceonmanuallabelingofsketches.Thisisaveryuseful

simplificationbecauseopen-domainsketchrecognitionisanunsolvedproblemandbecauselabeling

sketchesanddiagrams(e.g.withwrittenlabelsorspokenlanguage)iscommonplace,especiallyin

scienceinstructionwhereithasbeenshowntohelpstudentslearnnewcontent(Masonetal.,2013).

However,thismeansthatthereasonssketchesareimportantforthissystemarenotthesameasthe

reasonsvisualrepresentationsmightbeimportantforpeople.Futureexperiments,suchasusingan

oracleforlanguagedisambiguationorablatingvisualconceptualreasoningonly,couldprovideamore

detailedexplanationforhowandwhyspatialrepresentationsareimportant.

Despitetheselimitations,buildingsystemsthatcanunderstandmultimodalinstructional

analogies,evenifnotinexactlythesamewaysaspeopledo,isstillusefulfromaneducationaland

interactiveperspective.Interpretationabilitiesareneededforlearningbyreadingandforusing

analogiestoexplainphenomenatoahumancollaboratororstudent.

Chapter5: InstructionalAnalogiesforQuestionAnswering

Thepreviouschapterdemonstratedthatqualitativescienceknowledgecanbecapturedfrom

multimodalinstructionalanalogies.Theevaluationofthatknowledge,however,wascloselytiedtothe

analogiesthemselves.Thischaptershowstheextenttowhichthatknowledgecanbeusedtoanswer

sciencequestionsthatweredevelopedindependentlyfromtheinstructionalanalogiesusedforlearning.

805.1 Problem

Open-domainquestionansweringhasbeenanAIgoalfordecadesandithasseenresurgencesincethe

WatsonJeopardy!Challenge(Ferruccietal.,2010).AIresearchershavesuggestedusingelementaryand

middleschoolscienceexamsasonenewbenchmarkforprogresstowardhuman-levelAI(Clark,2015).

Thereasontheseexamsweresuggestedwasbecausetheyrequireagreatdealofqualitativeand

commonsenseknowledge.

Instructionalanalogiesmaybeagoodprecursortosolvingthesequestionsbecause(i)they

modelhowpeoplearefirstintroducedtosciencetopics,and(ii)theyaregoodtoolsforbuilding

qualitativeknowledge,whichisneededtoanswermanyofthesequestions.

5.2 Approach

Totesttheutilityofqualitativeknowledgecapturedthroughinstructionalanalogies,Igatheredscience

questionsfromtheNYstateregents4thgrade,8thgradeandlivingenvironmentexams2andthe

Massachusettscomprehensiveassessmentsystembiologyandphysicalsciencesexams3.Outof86

questionsthatwererelatedtohumanbiologyorelectricalenergy,Iidentified14questionsthatinvolved

topicsthatoverlappedwiththetopicsoftheanalogiesusedinthepreviouschapteranddidnotinvolve

graphunderstandingortargetdomaindiagramrecognition.Notallanalogieswererepresentedinthe

fullquestionset.Sometopics,likeenzymeactivation,werenotaddressed.Othertopics,like

ecosystemsandinterdependence,wererepresentedonmultiplequestions.The14questionsrequired

knowledgeaboutecosystems,cellstructureandfunction,DNAstructure,homeostasis,andelectrical

circuits.

2http://www.nysedregents.org/3http://www.doe.mass.edu/mcas/

81

Tolimitthescopeofthisquestionansweringexperiment,Imadetwomajorsimplifications.

First,ratherthanattemptnaturallanguageunderstandingonquestions,Imanuallytranslatedthem

fromEnglishtoCycL.Eachquestionwastranslatedintoaquery.Mostquestions(9)couldbe

formulatedasaquerywithanopenvariable,wheretheansweroptionswerecandidatebindingsfor

thatvariable.Therestofthequestions(5)weretruestatementquestions,whereeachansweroptionis

agroundquery(i.e.aquerywithnovariables).Ifthequestioninvolvedascenario,thepropertiesof

thatscenarioweretranslatedintoCycLandstoredinthemicrotheoryforthatindividualquestion.

Second,Ionlyconsidereddiagramquestionswhenthosequestionsdidnotrequiretargetdomain

recognition(e.g.interpretinganunlabeledimageofacell)orgraphunderstanding.Outofthe14

questionsexamined,3involveddiagramsbutonly1usedadiagramthatwasrequiredtoanswerthe

question.Thisdiagramshowedapartiallycompletecircuitwithlabels.Tomakethediagramusablefor

questionanswering,IsketcheditintoCogSketchusingthelabelsprovidedonthetest.Thesketchwas

thendeclaredavisualaidforthequestion.Whenthemodelattemptstoansweraquestion,itsearches

foravisualaidforthequestion.Ifitfindsone,whichonlyhappenedonceforthissetofquestions,it

usesthesamevisualconceptualreasoning(describedin4.2.2.1)thatwasusedforlearningfrom

instructionalanalogiestoextractadditionalinformationfromthesketch,andstorestheresultingfacts

inthequestion’smicrotheory.Thisway,whenqueriesareexecutedtoanswerthequestion,the

informationfromthevisualaidisvisiblefromthequestion’slogicalenvironment.

Myapproachforquestionansweringrequiredreasoningaboutquestiontypes,computing

similaritybetweenstatements,reasoningaboutobjectproperties,reasoningaboutexamplesituations,

andusingabduction.Eachoftheseabilitiesandtheorderinwhichtheyareexecutedinthisquestion

answering(QA)systemaredescribednext.

82

5.2.1. DetectingQuestionTypes

Tobuildasimplequestionansweringsystem,Ineededasmallsetofheuristicsthatweresufficientto

understandhowthequalitativescienceknowledgecouldbeused,ratherthanalargesetforbroadtest

coverage.Toachievethis,IusedtheanalysisprovidedbyClarkandcolleagues(Clarketal.,2013)to

informthestrategiesIimplemented.Theyidentifiedknowledgerequirementsforansweringquestions

onelementarysciencetestsandeightdifferent(butnotmutuallyexclusive)questioncategories.Only

someofthecategoriesidentifiedbyClarkandcolleagueswerepresentinthecurrentsetof14:object

propertyquestions,examplesituationquestions,andquestionsthatinvolvedcausalityandprocesses.

Objectpropertyquestionstypicallyaskaboutthepropertiesofaclassratherthanaspecific

object.WithinCycL,suchqueriesareexpressedwithrulemacropredicatesthatuseuniversal

quantification,e.g.relationExistsAllorrelationAllExists.Suchquestionstendtobeaboutthe

collectionthatisuniversallyquantified.Forexample,thequestioninFigure11isanobjectproperty

ObjectPropertyQuestion Afunctionofcellmembranesinhumansisthe a) synthesisoftheaminoacidsb) productionofenergyc) replicationofgeneticmateriald) recognitionofcertainchemicals

(queryForQuestion NYSRE-2014-LE-01 (relationExistsAll doneBy ?fn CellMembrane)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (MakingFn AminoAcid) 1)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (MakingFn EnergyStuff) 2)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList Replication-DNA 3)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (IdentifyingAsTypeFn ChemicalObject) 4)) (correctAnswerChoice NYSRE-2014-LE-01 4)

Figure11:AnexamplequestioninnaturallanguageandinCycL.

83questionaboutcellmembranes.Ifthequestionqueryinvolvesauniversallyquantifiedrulemacro

predicate,oritisatruestatementquestionwhereaparticularcollectionismentionedinmorethanone

option,theQAsystemassumesitisanobjectpropertyquestion.

Examplesituationquestionshaveknowledgeassociatedwiththemthatarenotaboutthe

multiplechoicequestionontology.Asimplecheckinthequestion’smicrotheoryforfactsotherthan

thoseusedtorepresentthequestionitself(i.e.predicatesotherthancorrectAnswerChoice,

multipleChoiceSingleOptionList,andqueryForQuestion)isagoodindicatorthatthequestion

involvesanexamplesituation.Lastly,therewerequestionsthatinvolvedcausalityandprocesses,but

theytendedtoalsobeexamplesituationquestions,sotheyweresimplytreatedassuch.

5.2.2. InterchangeabilityofStatements

Todealwithnon-identical,butsemanticallyandsyntacticallysimilarstatements,Idevelopedavery

simpleinterchangeabilitymeasurethatwasinspiredbytheloosespeakinterpreter(Fanetal.,2009).It

comparestwostatementsandgeneratesaninterchangeabilityscoreiftheyarerelatedtoeachother

usingCyc’scollectionandpredicatehierarchies.Identicaltermsreceiveascoreof2.Termsthatarenot

relatedreceiveascoreof0.Termsthatarerelatedreceiveascorethatisinverselyproportionaltothe

distancebetweentheminthegenls(forcollections)orgenlPreds(forpredicates)hierarchy.Common

supercollectionsarealsousedtorelatecollectionsevenifonecollectionisnotadirectspecializationof

theother.Termsthatrepresentqualitativequantities(e.g.Many-Quant, Thousands-Quant)are

groupedintothreegeneralcategories:small(e.g.Few-Quant),medium(e.g.Dozens-Quant),andlarge

(e.g.Millions-Quant).Theinterchangeabilityscorebetweentwoqualitativequantitiesisthecloseness

betweenthemonthesmall,medium,largespectrum.Termsthatarequantitativevalues(e.g.37,10)

aregivenascoreof0unlesstheyareequal,sincethedifferencebetweenthevaluesisnotmeaningful

84withoutaframeofreferenceforscale.Interchangeabilityofstatementsistheaverage

interchangeabilityoftheirterms,unlessanytermsareunrelated,inwhichcasetheinterchangeabilityof

thestatementsiszero.Forexample,thestatements(causes-SitTypeSitType Poaching

Extinction) and(enables-Generic HumanActivity Extinction)receiveascoreof1.1,because

thepredicatesarerelatedviagenlPredshierarchy,PoachingisasubcollectionofHumanActivity,and

bothstatementshaveidenticalthirdarguments.Thestatements(causes-SitTypeSitType Poaching

Extinction)and(enables-Generic EatingEvent Extinction)receiveascoreof0because

PoachingandEatingEventarenotconnectedbyasubcollectionrelation.Thisisaverycrudemeasureof

relatednessbutisusefulforevaluatingstatementsthathavebeenidentifiedasrelevantbysomeother

means.

5.2.3. ObjectPropertiesviaSubparts

Oneheuristicforsolvinganobjectpropertyquestionistocheckforthepropertyinknownsubpart

types.Ifaquestionhasbeendeterminedtobeanobjectpropertyquestion,thenthereisaspecific

collectionassociatedwithit.ThequestioninFigure11,forexample,isaboutthecollection

CellMembrane.IfthepropertyweseekisknownforanysubparttypesofCellMembrane(e.g.

LipidBilayer),thenwecanguessthatitalsoholdsforCellMembrane.Ifthepropertyweseekisnot

knownforanyknownsubparttypes,thenthisheuristicformulatesalternativeexpressionsforthefailed

subpartqueryandstoresthemawayaspotentiallyrelevantpatternsthatcanbeexploredifnoother

reasoningpathleadstoananswer.

ThesystemanswersthequestionshowninFigure12byreasoningaboutsubparts.Rephrased

intermsofCycL,thisquestionasks,“Allcellshavethousandsof?x”whichcanberepresentedwiththe

followingrulemacropredicatestatements:

85(relationAllExistsRange physicalParts Cell ?x Thousands-Quant)

ThisfactisnowheretobefoundintheKB,sothesystembeginsbylookingatthesubpartsofCell.It

thenexecutesquerieslike

(relationAllExistsRange physicalParts CellNucleus ?x Thousands-Quant),

(relationAllExistsRange physicalParts Mitochondrion ?x Thousands-Quant),

andsoon.Itstilldoesn’tfindananswer,butitstoresseveralpossiblyrelevantpatternsinworking

memory.Thesepatternssubstituteansweroptionsintothevariableposition,movevariables,and

introducenewonesiftheexpressionislongenough(toavoidafullyopenandexpensivequery).Oneof

thepatternsthatgetsstoredis

(relationAllExistsRange physicalParts ?x Gene-HereditaryUnit ?y)

Whenqueried,thispatternreturns

(relationAllExistsRange physicalParts CellNucleus Gene-HereditaryUnit Many-Quant)

ThisfactexistsintheKBbecauseitwaslearnedfromthecell/earthanalogy.Thepart-for-whole

substitutionisundone,andwenowhavethefact:

(relationAllExistsRange physicalParts Cell Gene-HereditaryUnit Many-Quant)

BecauseMany-QuantandThousands-Quantbearsomeresemblanceascomputedbythe

interchangeabilityalgorithm,andallothertermsinthestatementareequaltooneoftheanswer

options,thisstatementservesasevidenceforthefirstansweroption:

(relationAllExistsRange physicalParts Cell Gene-HereditaryUnit Thousands-Quant)

86

ObjectPropertyQuestion(#1) Asinglehumanbodycelltypicallycontainsthousandsof

(relationAllExistsRange physicalParts Cell ?cell-part Thousands-Quant))

a)Genes

Gene-HereditaryUnit

b)Nuclei

CellNucleus

c)Chloroplasts

Chloroplast

d)Bacteria

Bacterium

GroundQueryforAnswerOption1 (relationAllExistsRange physicalParts Cell Gene-HereditaryUnit Thousands-Quant)

GroundQueryforsubparts (relationAllExistsRange physicalParts CellNucleus Gene-HereditaryUnit Thousands-Quant) (relationAllExistsRange physicalParts Mitochondrion Gene-HereditaryUnit Thousands-Quant) ...

Queryvariants (relationAllExistsRange physicalParts CellNucleus Gene-HereditaryUnit ?q) (relationAllExistsRange physicalParts ?whole Gene-HereditaryUnit ?q) ...

Result(oneofmany) (relationAllExistsRange physicalParts CellNucleus Gene-HereditaryUnit Many-Quant)

Undopart-for-wholesubstitution (relationAllExistsRange physicalParts Cell Gene-HereditaryUnit Many-Quant) Evidenceforansweroption1

Figure12:Asubsetofthequeriesthatareexecutedforquestion1.Groundqueriesarefirstcreatedforallansweroptions.Sincenonesucceedforthisquestion,theQAsystemattemptsquerieswithsubparttypesinplaceoftheoriginalcollection.Sincethosealsofail,theQAsystemattemptsqueryvariants,withvariablesintroducedtoincreasethechancesofreturningaresult.OneoftheresultsthatisreturnedisthatNucleihavemanygenes.Thisissimilartothegroundqueryforansweroption1,andisthereforeusedasevidenceforthatanswer.

875.2.4. ObjectPropertiesviaBaseDomainSpeculation

Anotherapproachforansweringobjectpropertyquestionsistothinkoftheobjectintermsofaprior

analogy.GivenanobjectpropertyquestionaboutsomecollectionCandsomepropertyP,ifthereexists

aknownanalogicalmappingwhereCmappedtoC’andC’haspropertyP,thenthesystemassumesthat

Calsohasthatproperty.Thisisariskyheuristicandisthereforeassignedahighcostsothatthesolver

onlyattemptsthisiffindingananswerintheoriginaldomainfails.IfC’doesnothavepropertyP,then

variantsoftheoriginalquery,withC’inplaceofC,arestoredaspotentiallyrelevantquerypatternsthat

mightbeusefulasalastresort.

Forexample,thesystemusesbasedomainspeculationtoanswerthequestionshowninFigure

11.Thequestionasksaboutthefunctionofcellmembranes.Fromaprioranalogy(i.e.thecell/city

analogy),thesystemfindsthatatonepoint,thecollectionCellMembranecorrespondedtothe

collectionBorder.Thesystemcansearchtheknowledgebaseforfunctionsofborderstoseeifthere

arefunctionsthataresimilar(i.e.haveahighinterchangeabilityscore)toanyoftheansweroptions.

Forthisquestion,thegroundqueriesareexecutedwiththebordercollectionreplacingthecell

membranecollection.Thesequeriesdonotreturnanyresults,soqueryvariants(i.e.withdifferent

openvariables)areexplored.Oneoftheresultsthatisfoundinvolvestherelationshipbetweenborders

andbordercheckevents:

(relationExistsAll preActors BorderCheckEvent Border)

Whenthisfactistransformedbackintothetargetdomain,itbecomes

(relationExistsAll preActors BorderCheckEvent CellMembrane)

88Thisintermediatefactdoesn’tmakeverymuchsensebecausethesystemdoesnotcreateananalogy

skolemforBorderCheckEvent.However,thisispartoftheadvantageofbasedomainspeculation.It

allowsthesystemtomakebiginferenceleapsthatareguidedbyprioranalogies.Thesystemproceeds

bycomputinginterchangeabilityscoresbetweenallthefactsfoundfromqueryvariantsandtheanswer

options.ThefactinvolvingBorderCheckEventstandsoutbecauseitisinterchangeablewithanswer

option4:

(relationExistsAll doneBy (IdentifyingAsTypeFn ChemicalObject) CellMembrane)

(relationExistsAll preActors BorderCheckEvent CellMembrane)

Thesetwofactsreturnanon-zerointerchangeabilityscorebecauseBorderCheckEventisa

specialization(i.e.subcollectionof)InspectingandIdentifying,whicharesupercollectionsof

(IdentifyingAsTypeFn ChemicalObject).Additionally,doneByisaspecializationofpreActors.The

restofthetermsareidentical,sotheinterchangeabilityreturnsanon-zerovalue,indicatingthatthe

secondfactcanbeusedasevidencetosupportansweroption4.Thiscausestheinvolvementof

bordersinbordercheckeventstobeusedasevidencethatcellmembranesidentifychemicalobjects.

5.2.5. ExampleSituationAnalysis

Examplesituationquestionssetupascenarioforthereaderandasksomethingaboutthatscenario.In

thequestionrepresentationusedhere,thatmeansthatthequestion’smicrotheoryhasextraknowledge

init.Thewaytoanswerthesequestionsusuallyrequiresfillinginmissingpieceswithbackground

knowledge.Cyc’srulemacropredicatesareveryusefulforthisandarejustthetypesofrepresentations

thatarelearnedfromtheinstructionalanalogyinterpretationmodel.

89 Toapplybackgroundknowledgetoscenarios,Iimplementedforwardinferencemethodsfor

rulemacropredicates(seeexamplesinTable1)alongwiththepredicatesthatrepresentqualitative

proportionalities(i.e.qprop,qprop-).Thesemethodssearchthelocalcontextforinstancesof

universallyquantifiedcollectionsandexistentiallyquantifiedcollectionsandassertsthebinaryrelation

betweenthem.Ifnoexistentialinstanceisfoundandtherearenootherpotentialcandidates(i.e.

instanceswhosecollectionsareunknown),thenaspeculativeinstanceiscreated.Inthisway,the

applicationofbackgroundknowledgeenablesthecreationofnewtermsthatareassumedtobethere

eveniftheywerenotmentionedexplicitly.Forrulemacropredicates,theuniversalandexistential

quantifiersareexplicitbasedontheargumentposition.Forqualitativereasoningpredicates,Ichoseto

treatthemasifthecausewasuniversallyquantifiedandtheeffectwasexistentiallyquantified.For

example,givenatypelevelnegativequalitativeproportionalitybetweentherateofchangeinan

ecosystem(thecause)andthestabilityofthatecosystem(theeffect),thentheapplicationmethodfor

thispredicatewillseekinstancesofchangesinanecosystem.Ifanyarefound,thenaninstancelevel

negativequalitativeproportionalitybetweentherateofthatchangeandthestabilityoftheecosystem

areasserted.Thisishowsomeofquestionsrelatedtoecosystemsandinterdependencearesolved.

90

Forexample,oneofthequestions(

Figure13)presentsanexamplesituationabouthumanpopulationaffectingpredatorpopulations.The

firstrowofthetableshowsthescenarioasitisdescribedinthequestion(left)andtheCycL

ExampleSituationQuestion(#4) HumanpopulationgrowthhasledtoareductioninthepopulationsofpredatorsthroughoutnaturalecosystemsacrosstheUnitedStates.Scientistsconsiderthelossofthesepredatorstohavea

(isa this-ecosystem Ecosystem) (consumerInEcosystem this-ecosystem these-humans) (consumerInEcosystem this-ecosystem other-predators) (eatsWillingly other-predators other-prey) (isa human-pop-growth IncreaseEvent) (isa human-pop-growth IntrinsicStateChangeEvent) (objectOfStateChange human-pop-growth this-ecosystem) (objectOfStateChange human-pop-growth other-predators) (isa predator-decline DecreaseEvent) (isa predator-decline IntrinsicStateChangeEvent) (objectOfStateChange predator-decline this-ecosystem) (causes-EventEvent human-pop-growth predator-decline)

a)positiveeffect,becauseanincreaseintheirpreyhelpstomaintainstabilityintheecosystem

(and (causes-Event predator-decline ?prey-increase) (isa ?prey-increase IncreaseEvent) (positivelyInfluencedBy

((QPQuantityFn Stability) this-ecosystem) (RateFn ?prey-increase)))

b)positiveeffect,becausethepredatorsusuallyeliminatethespeciestheypreyon

(and (causes-Event predator-decline ?ext) (isa ?ext Extinction) (positivelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn ?ext)))

c)negativeeffect,becausepredatorshavealwaysmadeupalargeportionofourfoodsupply

(and (eatsWillingly these-humans ?food-supply) (negativelyInfluencedBy

(AmountOfFn ?food-supply) (RateFn predator-decline)))

d)negativeeffect,becausepredatorshaveanimportantroleinmaintainingstableecosystems

(negativelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn predator-decline))

Figure13:Anexamplequestionaboutecosystems.TherighthandcolumnshowstheCycLrepresentationofthescenarioandtheansweroptions.

91representationforthatscenario(right).TheansweroptionsinEnglishandinCycLareshownbelowthe

scenariodescription.Thesystemanswersthisquestionbyapplyingbackgroundknowledgeintothe

examplesituation.BackgroundknowledgeisretrievedbysearchingtheKBforrulemacropredicates

andqualitativeproportionalitiesthatinvolvecollectionsmentionedinthescenario.Inthiscase,the

collectionsmentionedinthescenarioare:Ecosystem,IntrinsicStateChangeEvent,IncreaseEvent,

andDecreaseEvent.OneofthequalitativeproportionalitiesfoundintheKBwascapturedfromthe

analogyaboutecosystemsfromChapter4:

(qprop- ((QPQuantityFn Stability) Ecosystem)

(RateFn (IntrinsicStateChangeOfFn Ecosystem)))

Thisstatesthattherateofchangeinanecosystemnegativelyinfluencesthestabilityofthatecosystem.

Theforwardinferencemethodsdevelopedforthissystemtreatqualitativeproportionalitiesasbeing

type-levelstatementswherethecauseisuniversallyquantifiedandtheeffectisexistentiallyquantified:

(relationExistsAll qprop- ((QPQuantityFn Stability) Ecosystem)

(RateFn (IntrinsiceStateChangeOfFn Ecosystem))

Thisstatementmeansthatforallecosystemchanges,thereexistsanecosystemthatisnegatively

affectedbyit.Notethattherulemacropredicateforrepresentingthisqualitativeproportionalityis

actuallyunderspecifiedbecauseitdoesnotsaythatthechangetoaspecificecosystemnegatively

affectsthestabilityofthesameecosystem.However,theforwardinferencemethodsattemptto

resolvethisissuebyhypothesizingwhichentities(ifany)inthelocalcontextarethebestcandidatesfor

theexistentialvariable.Toapplythisknowledgetothecurrentquestionscenario,thesystemfirstlooks

forinstancesof(IntrinsicStateChangeOfFn Ecosystem).Thequestiondoesnotexplicitlystatethat

92humanpopulationgrowthandthedeclineofpredatorsareinstancesof(IntrinsicStateChangeOfFn

Ecosystem),butitdoessaythattheyarebothinstancesofIntrinsicStateChangeEventandthatthe

objectOfStateChangeisaninstanceofEcosystem(Figure13).Thesystemusesthisknowledgeand

theknowledgeofIntrinsicStateChangOfFnasacollectiondenotingfunctionintheKBtoinferthat

humanpopulationgrowthandthedeclineofpredatorsareindeedinstancesof

(IntrinsicStateChangeOfFn Ecosystem).Next,thesystemlooksforsomethinginthescenariothat

mightbeaninstanceof((QPQuantityFn Stability) Ecosystem).Itfirstsearchesforinstancesof

((QPQuantityFn Stability) Ecosystem).Itfindsnone,soitproceedsbysearchingforinstancesof

Ecosystemandfindstheonedescribedinthisquestionscenario.Sincethereisonlyoneecosystemin

thisquestionscenario,thesystemassumesthatthisistheentitythatthebackgroundknowledgeshould

beappliedto.Ifthereweremultipleinstancesofecosystem,thesystemwouldusethenumberof

commonrelationsbetweeneachinstanceofEcosystemandeachinstanceof

(IntrinsicStateChangeOfFn Ecosystem)toguesswhichinstancestochoose.Ifnoinstancesof

Ecosystemwerefound,thenthesystemwouldcreateaspeculativeinstanceofEcosystem.Forthis

question,thetype-levelqualitativeproportionalityfrombackgroundknowledgeisappliedtothe

scenario,whichresultsinthefollowinginstance-levelstatements:

(isa human-pop-growth (IntrinsicStateChangeOfFn Ecosystem))

(isa predator-decline (IntrinsicStateChangeOfFn Ecosystem))

(qprop- ((QPQuantityFn Stability) this-ecosystem) (RateFn human-pop-growth))

(qprop- ((QPQuantityFn Stability) this-ecosystem) (RateFn predator-decline)))

Theresultingstatementsareassertedintothequestionscenario.Unlikestatementsinbackground

knowledge,thesestatementsareaboutspecificentitiesratherthancollections.Nowthequestion

answeringprocesscontinueswithmoreknowledgethanitstartedwith.Whenthesystemdirectly

93querieseachoftheansweroptions,optionDsucceedsbecausepre-existingrules(fromtheQP

ontology)enablethesystemtoinferpositivelyInfluencedByandnegativelyInfluencedBy

statementsbasedonqpropandqprop-statements,respectively.

5.2.6. AbductiveHypotheses

Abductionispotentiallyusefulforansweringobjectpropertyquestionsandscenarioquestions.Inboth

cases,thissystemlooksforknownfactsthatmightexplainorsupportanyoftheansweroptions.This

wasimplementedbywritingrulestocapturetherequirementsforparticularfunctionsandtheexistence

ofpart-wholerelationshipstoexplainthecollectionmembershipofaparticularitem(Figure14).

Forfunctionrequirements,rulesstatethatpart-wholerelationshipsbetweenentitiescouldsuggestthat

thepartisthedirectobjectinanactioncompletedbythewhole.Thiswasimplementedforthree

classesoffunctions(basedonthetypesoffunctionsthatwereobservedinthe11instructional

analogiesusedinChapter4andthequestionset):makingthingsavailable,changingthings,andtaking

careofthings(e.g.protecting,storing).Therationaleis,ifsomecollectionCtypicallyhassubpartsC’,

thenthatcanbeusedasevidencetosuggestthatinstancesofCperformactionsoninstancesofC’,such

asmakinginstancesofC’available,takingcareofinstancesofC’andchanginginstancesofC’.For

example,batterieshaveenergy.Knowingthismightsuggestthatbatteriesconvertenergy,makeenergy

available,orstoreenergy.Forexample,thequestioninFigure15asksaboutthefunctionofthecell

nucleus.Fromthecell/cityanalogy,thesystemknowsthatoneofthefunctionsofthecellnucleusis

thatitcontrolsthecell.However,thisfactisnotusefulforthisparticularquestion.Instead,thesystem

checksifanyoftheansweroptionscanbeassumedviaabduction.Options2and3canbeassumedif

thethingsmadeavailableorstoredareknowntobepartsofthecellnucleus.Fromthecell/earth

94analogy,thesystemknowsthatgenestendtobepartsofcellnucleiandthereforeassumesthatoption

3isthecorrectanswer.

Forexplainingthecollectionmembershipofaparticularitem,onerulewasusedtostatethat

givensometermPthatispartofsomeothertermW,andatype-levelpart-wholerelationshipbetween

W-TYPEandP-TYPE,thenthereisanabductivehypothesisthatPisaninstanceofP-TYPE.Thisisuseful

forquestionsthataskaboutpartsofthingsinaparticularsituation(e.g.question13,Table14).When

givenasituationwherethereisanobject(e.g.acircuit)withpartsthatareunlabeled,thesystem

assumesthattheunlabeledpartbelongstoacollectionthatisknowntobeasubparttypeofthe

originalobject(e.g.batteryorwire).

Explainingfunctions:

(abductiveHypothesis (relationExistsAll ?role ?fn ?whole-type) (and (unifies ?fn (event-fn? ?part-type))

(or (resultGenl ?event-fn InstrinsicStateChangeEvent) (resultGenl ?event-fn MakingSomethingAvailable) (resultGenl ?event-fn TakingCareOfSomething))))

(partTypes-transitive ?whole-type ?part-type))))

Explainingcollectionmembership:

(abductiveHypothesis (isa ?term ?part-type) (and (physicalParts ?whole ?term) (isa ?whole ?whole-type) (partTypes-transitive ?whole-type ?part-type)))

Figure14:Abductivehypothesisstatementsusedtoassumefunctionsandcollectionmembership.Theseareimplementedinhornclauses,withthefirstabductivehypothesisstatementbrokenintothreeseparatehornclausestohandlethedisjunction.InEnglish,thefirststatesthatifsomethinghassubparts,thenareasonablehypothesisisthatitcanmakethesubpartsavailable,changethesubparts,ortakecareofthem.Forexample,knowinghavebatterieshaveenergysuggeststhatbatteriesmakeenergyavailable.Thesecondstatesthatifthereissomethingthathassubpartsanditisknowntohavesubpartsofaparticulartype,thenwecanassumethatthesubpartsareofthatsubparttype.Forexample,ifthereisacircuitwithahiddenpart,andweknowthatcircuitstendtohavewires,thenareasonablehypothesisisthatthehiddenpartmightbeawire.

95

ObjectPropertyQuestion(#8) Whichofthefollowingisafunctionofthenucleusinorganism2?

(relationExistsAll performedBy ?fn CellNucleus)

a)absorbingsunlight (SubcollectionOfWithRelationToTypeFn AbsorptionEvent objectActedOn Sunlight)

b)releasingusableenergy (MakingAbstractAvailableFn EnergyStuff)

c)storinggeneticmaterial (StoringFn Gene-HereditaryUnit)

d)producingfoodmolecules (MakingFn Food)

Figure15:Anobjectpropertyquestionthataboutthefunctionofthecellnucleus.

96

5.2.7. SolvingSequence

ThesequenceforattemptingtoansweraquestionisillustratedinFigure16.Thefirst,mostbasic

approachistoexecutethegroundqueriesforthequestiondirectly.Thegroundqueriesforeach

questionareeitheraunificationoftheansweroptionsandtheopenquestionquery(aswouldbethe

caseforthequestionshowninFigure11),oreachansweroptionifthequestionisatruequestion

statement(sinceeachansweroptionisafullygroundstatement).Thisstepalsoincludesabduction.So,

ifthegroundqueryisexplicitlyknownintheKBorifitcanbeassumedviaabduction,thenthequery

Figure16:Anoverviewofhowmultiplechoicequestionsareanswered.Thefirststepexecutesthegroundqueriesdirectly.Ifananswerisfound(followingtheboldarrow),thentheprocessisdone.Ifnot,itcontinueswithabasicanalysisofthequestiontodetermineifitisaboutobjectpropertiesoranexamplesituation.Foreach,differentstrategiesareemployed.Theobjectpropertystrategyusessubpartsandbasedomainentitiestocomeupwithvariantsoftheoriginalquerybyplacingvariablesindifferentargumentpositions.Ifananswerisnotfounddirectlyfromsubpartsorbasedomainentities,thesystemexaminesthequeryvariantsandusedthemtocomputeevidenceforeachoftheansweroptions.

97willsucceed.Ifanyofthesequeriessucceed,thenthesystemreturnstheanswer(followingthebold

arrowinFigure16).Ifnot,itproceedstothenextstep,whichinvolvesdetectingthequestiontype.

Ifaquestiontypeisdetected,thecorrespondingstrategyistriggered.Asdescribedabove,

thesestrategiesincludelookingatsubparttypesandbasedomainitemsforobjectproperties,and

importingbackgroundknowledgeforexamplesituationquestions.Forobjectproperties,iftheproperty

inquestionisfoundineitherasubparttypeorinacorrespondingbaseterm,thenthesystemhasfound

itsanswer.Ifnot,itgeneratesvariantsofthesubpartandbasedomainqueriesbyinsertingvariablesat

differentargumentpositions.Thesenon-groundexpressionsaremarkedaspossiblyrelevantquery

patterns,whicharenotexecuted,butstoredawayasalastresortifotherreasoningpathsfail.For

situationquestions,backgroundknowledgeintheformofrulemacropredicatesandqualitative

proportionalitiesareimportedintothequestion’sreasoningenvironment.Asmentionedearlier,

applyingbackgroundknowledgeresultsintheexistenceofnewstatements,andsometimesinthe

existenceofnewterms.Thismeansthatthegroundqueriesforthequestionshouldbeattempted

again,incasetheanswerwasbroughtinfromthebackground.Ifanyofthegroundqueriessucceed,

thesystemisdone.Ifnot,itcontinuestothefinalheuristic,whichlooksatthepossiblyrelevantquery

patternsstoredawayearlier.

Thequerypatternsthathavebeenmarkedaspossiblyrelevantwillincludestatementsabout

subparttypesandcorrespondingbaseconcepts.Itwillalsoincludevariationsoftruestatementoptions

thathavefailedandvariationsoftheoriginalgroundqueries.Thisisessentiallyapurelysyntacticway

offindingstatementsthatmaybeuseful,butarenotfoundbydirectqueries.Foreachstatementthatis

returnedbythepatterns,theinterchangeabilityscoreiscomputedbetweenitandalloftheanswer

options.Thescoresforeachansweroptionaresummedasanestimateofthesystem’sconfidencein

98thatanswerchoice.Onceallthestatementsareevaluated,theanswerchoicewiththehighest

confidenceischosenastheanswer.Thequestionansweringprocessfailsifthequeryvariantsdonot

returnanyresultsortheresultsbearnoresemblancetoanyoftheansweroptions(i.e.the

interchangeabilityscoresareallzero).

5.3 Evaluation

Usingtheheuristicsdescribedabove,theknowledgecapturedfromtheinstructionalanalogiesin

Chapter4wassufficientforanswering11outof14questions.Thisincluded4outof5questionson

ecosystems,5outof6questionsonthecellanditsparts,and2outof2questionsoncircuits(Table14).

99

# QuestionPrompt--CorrectAnswer Correct? Approach

1 Asinglehumanbodycelltypicallycontainsthousandsof--genes

y objectpropertyviasubparts+interchangeability

2 Afunctionofcellmembranesishumansisthe--recognitionofchemicals

y objectpropertyviabasedomain+interchangeability

3 Whichstatementbestdescribestheorganellesinacell?--theyworktogether

y directquery+interchangeability

4 HumanpopulationgrowthhasledtoareductioninthepopulationsofpredatorsthroughoutnaturalecosystemsacrosstheUnitedStates.Scientistsconsiderthelossofthesepredatorstohavea--negativeeffectonstability

y importbackgroundQPknowledge+inference

5 Avarietyofpeartree,knownasBradford,wasoriginallyintroducedintotheeasternUnitedStatesinthe1960s.Today,thistreeiscrowdingoutotherplantsinthesestates.Thissituationbestillustrates--negativeeffectonstability

y importbackgroundQPknowledge+inference

6 Iftheproducersinafoodwebwereremoved,whichofthefollowingchangeswouldmostlikelyoccur?–collapsebecauseecosystemsrequireproducers

y directquery

7 Withinapreypopulation,whichofthefollowingismostimmediatelyaffectedbythearrivalofanewpredator?--populationofpredatorinfluencesdeathrate

n importbackgroundQPknowledge,butdidn'thaveQPrelationshipimpactingpopulation

8 Whichofthefollowingisafunctionofthenucleusinorganism2?--storinggeneticmaterial

y abduction

9 Inthethreeorganisms,whataresynthesizedbytheribosomes?--protein

y directquery+interchangeability

10 Whichofthefollowingbestdescribestheproducersinaterrestrialfoodweb?--producersconvertsolarenergytochemicalenergy

y directquery

11 Whichofthefollowingisthebestexampleofthehumanbodymaintaininghomeostasis?--breathingduringexercise

n importbackgroundknowledge,butdidn'thaveawayofinferringprocesstype

12 Inasampleofdouble-strandedDNA,30%ofthenitrogenousbasesarethymine.Whatpercentageofthenitrogenousbasesinthesampleareadenine?--0.3

n importbackgroundknowledge,butnoexplicitequalsrelation

13 Jamalwantstomakeanelectricalcircuit,butheonlyhastheobjectsshownbelow.WhichofthefollowingmustJamalalsohavetomakeanelectricalcircuit?--battery

y abduction

14 Thediagrambelowshowsaprojectthatastudentmadetotestanelectricalcircuit.Partoftheelectricalcircuitisunderneaththeboard.Whenthestudentconnectsthetwonailsusingawire,thebulblightsup.Whichofthefollowingmustbeunderneaththeboard?--batteryandwires

y abduction

Table14:Questionsusedtoevaluatequalitativescienceknowledge.

100 Avarietyofstrategieswereusedtoarriveatanswers.Insomecases,directqueriessufficed.

Question10,forexample,isatruestatementquestionwhereoneoftheoptionsdescribedthefunction

ofproducers:thattheyconvertsolarenergytochemicalenergy.InCycL,thisisformulatedasarule

macropredicateusingaprimaryactorrolerelation:

(relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn IntrinsicStateChangeEvent toState ChemicalEnergy) objectOfStateChange SolarEnergy) Autotroph)

ThisfactwascapturedfrombyinterpretingtheecosystemanalogyfromChapter4,soadirectquery

wassufficientforansweringthequestion.

Abductionwasveryusefulforscenarioquestionswheretherewasamissingorunlabeledentity

andthequestionaskedforacollectionforthatentity(questions13,14).Thesecondabductionrulein

Figure14wasresponsibleforsolvingthesequestions.Sincebatteriesandwiresareknowntobepartsof

circuits,itwasassumedthattheansweroptionsthatmentionedthosepartswerecorrect.However,

thebackgroundknowledgewasnotdefinitiveintermsofhowmanyofeachparttypeacircuithad,soit

couldhaveincorrectlyansweredanotherbattery.But,sinceitemsthatwerealreadyinthescenario

werenotalsoansweroptions,thesystemwasabletocorrectlychoosetherightanswer.Clearly,the

backgroundknowledgeisnotasdeepasitshouldbe,butforthesequestions,itissufficient.Essentially,

thequestionisaskingthereader:whataretherequiredpartsofthisobject?Whateverismissingisthe

answer.

101 Backgroundknowledgewasusefulfortwoofthequestionsaboutecosystemsand

interdependence.Theecosystem/webanalogyinChapter4wasinterpretedusingqualitativemodeling

ofanadhocstudentweb(i.e.asituationwherestudentsgatheraround,connectedbystringstomimic

anecosystem).Theprimaryinferenceisthatchangeswithinthewebdecreasestability.Thecausal

relationshipbetweenecosystemchanges,andthestabilityoftheecosystemwascapturedasaresultof

interpretingthisanalogy.Whenpresentedwithascenarioquestionaboutanecosystemandsome

changehappeningwithinit(i.e.questions4,5,7),thesysteminfersthatthechangeisnegatively

influencingecosystemstability.Thisrelationshipisverysimple,yetitissufficientforanswering

questions4and5.

Asnotedearlier,objectpropertyquestionscouldbesolvedbylookingatsubparttypesandby

speculatingaboutknownbasedomains.Detectingobjectpropertiesviasubpartswasusedtoanswer

question1(Figure12)anddetectingobjectpropertiesviabasedomainspeculationwasusedtoanswer

question2(Figure11).

5.4 Discussion

Theexperimentinthischaptersupportsclaim2ofthisthesis.Itdemonstratesthatqualitativescience

knowledgecapturedfrommultimodalinstructionalanalogiescanbeusedincombinationwithverybasic

questionansweringstrategiestoanswerquestionsonmiddleschoolscienceexams.Thequestion

answeringstrategiesdevelopedforthisexperimentalongwithbackgroundknowledgeandontologies

withinCyc,wereusedtoanswering11outof14questions.Theimpactofthespecificcontentlearned

fromtheinstructionalanalogiesinChapter4canbeevaluatedbymakingthatcontentunavailable

duringquestionanswering.Afterremovingtheknowledgelearnedfromthoseanalogies(butleavingin

allotherbiologyandelectricityknowledgeinResearchCyc),theQAsystemwasunabletoansweranyof

102the14questionsexaminedhere.ThisdemonstratesthatwhiletheontologiesinCycareextremely

usefulandnecessary,theyneedtobecomplementedwithgeneral,non-technicalstatementsaboutthe

domain,liketheonesthatwerecapturedfrominstructionalanalogies.Theyalsoneedtobe

complementedwithQAstrategiesliketheonespresentedinthischapter.

Thereare,however,severallimitationstothisclaim.Thefirsthastodowithcoverage.Outof

the86testquestionsthatwerefoundtoinvolvehumanbiologyorelectricity,only14hadanswersthat

overlappedwiththetargetknowledgeoftheanalogies.Around16%coverageismodest,butalsonot

surprisingwhenoneconsidersalltheothertopicsthatappearontheseexams.Inthefullsetof86

questionsonhumanbiologyandelectricalenergy,thereweremanyquestionsongenetics,causal

reasoningaboutexperiments,specificbiologicaltaxa,classdefinitionsforspecifictypesofcells,

propertiesofeverydayobjects(e.g.conductance)andmanyothertopicsnotaddressedbyanyofthe

analogiesintheFARguide.However,it’snotthecasethattheFARguidetopicsarerare.Theyinclude

verybasicfundamentalideasinbiologyandelectricity,itisjustthatitwouldbeveryunusualfor11

analogiestocoveraverylargesetofquestionsatallthelevelsofdetailthattheyareasked.The

analogiesaddressbasicconceptualknowledgeaboutseveraltopicswiththeexpectationthatmore

detailedknowledgeisaddedlater,whichisnotcurrentlydoneinthisinterpretationmodel.The

analogiesdonotdiveintotechnicaldetailandtheyassumethatthelearnerhasasignificantamountof

commonsenseknowledge.Thecommonsenseknowledgeassumptionisfineforpeople,butabig

problemforintelligentsystems.Inshort,therearetopicsonbothendsofthetechnicalspectrum(from

basiccommonsenseknowledgetotechnicalknowledgeofspecificprocesses)thataremissedbythese

instructionalanalogies.

103

Thesecondlimitationhastodowithhowqualitativescienceknowledgeisactuallyused.Ifit

cannotbedeployedproperly,thenit’snotworthverymuch.ThethreequestionsthattheQAsystem

couldnotanswerpointtosomeofthereasoningabilitiesthatarelackinginthissystem,andaretaken

forgrantedwhenpeopleusequalitativeknowledgetoanswerquestions.Question7wasasituation

question,andaskedwhatisimmediatelyaffectedbythearrivalofanewpredator.Basedonthe

knowledgegainedfromtheecosystem/webanalogy,thereisinformationintheKBaboutproducers,

consumers,thatpredatorseatprey,andthatchangesintheecosystemnegativelyimpactstability.

What’smissingtheverybasiccausalinformationbetweeneating,death,andpopulation.Background

knowledgewasimportedcorrectly,buttherewasnocausalknowledgeabouttheimpactpredatorshave

onthepopulationoftheirprey.Amoresophisticatedmodelofthedynamicsinanecosystemwouldbe

neededtocapturethat.Thosemodelswouldalsoneedtobetiedwithourmostgeneralmodelsofhow

theworldworksinordertobeuseful.Question11,forexample,providesfoursituationsandaskswhich

oneisthebestexampleofhomeostasis.Answeringthisquestionrequiresanunderstandingofthe

notionofcompetinginfluencesorfeedbacksystems.Thesimpleforwardinferencemethod

implementedinthisQAsystemisinsufficientforthisquestionbecauseinsteadoffindingquantitiesand

assertingqualitativerelationshipsaboutthem,thisquestionrequiresdetectingrelationshipsand

assertingtheprocessorsituationtype.Thisisakintomodelformulationingeneral,somethingthatwas

usedforinterpretinganalogiesbutnotforansweringquestionsaboutthem.Lastly,question12was

aboutthestructureofanexampleDNAmolecule.FromtheDNA/clothespinanalogy,thesystemknew

thatadeninealwaysbondswiththymineandthatguaninealwaysbondswithcytosine.Theconnection

betweenthatknowledgeandthefactthatinagivenmoleculethenumberofadeninemoleculesshould

beequaltothenumberofthyminemoleculeswasmissing.Thesequestionsillustratetherich

104interactionbetweenqualitativescienceknowledgeandqualitativecommonsenseknowledgethatis

neededtopassthistest.

Anotherlimitationtothisclaimistheabsenceofnaturallanguageunderstandinginquestion

interpretation.Tolimitthescopeofthisthesis,IchosetomanuallytranslatethequestionsfromEnglish

toCycL.Thisprovedtobeverydifficultbecausetherearenumerouswaystorepresentthingsusing

CycL,andveryoften,thestatementsdefinedtorepresentknowledgeareunderspecifiedornot

completelyaccurate(Fanetal.,2009).However,asseenintheanalysisofanalogiesintheFARguide,

sometimesthelooseuseoflanguageisafeaturenotabug.Inanimatethingsorprocessesare

sometimesdescribedusinglanguagethatwouldsuggesttheyarelivingagents.Forinstance,organelles

don’treallyengageincoordinatedgroupactivitiesinthesensethatCycdefinesthattypeofactivity.This

isbydesignbecausetheanalogyusesalevelofabstractionthatisamenabletocross-domainmappings.

Usingthislevelofabstraction,whileatthesametimeadheringtotherepresentationalconstraints(e.g.

predicateargumenttypeconstraints)withinCyc,remainsachallenge.

Chapter6: RelatedWork

AnalogicalreasoninginAIsystemshasarichhistorywithmanydifferentapplications.Researchersoften

differinhowtheydefineanduseanalogicalreasoning,withsomedevelopingmethodsforsolving

traditionalanalogywordproblems(i.e.A:B::C:D),othersusingittoperformcase-basedreasoning,and

otherswhouseitasageneralmeasureofrelationalsimilarity.

6.1 ProblemSolvingandQuestionAnswering

Mostanalogicalproblemsolversworkbytransferringpriorknowledgetoanewscenario.Earlyworkby

Burstein(1985)analyzedhumantutoringprotocolstoproposeamodelforlearningbyanalogy.This

105modelwasimplementedinasystemthatusedstoredplanningknowledgetoincrementallygenerate

hypothesesforhowtosolvecomputerprogramming(i.e.variableassignment)problems.ThePRODIGY

planningarchitectureusedaderivationalanalogyengine(Veloso&Carbonell,1993)tolearncontrol

knowledgeforfuturesituations.Whennewplanningproblemsarose,PRODIGY/ANALOGYusedprior

similarproblemstolimitthesearchfornewsolutions.Thegeneralapproachofusingpriorexperiences

orexamplestosolveorplaninnewsituationsistypicalincase-basedreasoningsystems(Kolodner,

2006).

Aslightlydifferentapproachtoanalogicalproblemsolvingistousepriorknowledgetoconstruct

newknowledgethatisabstractedawayfromanyonescenario.InPHINEAS(Falkenhainer,1990),new

domaintheorieswereconstructedbyanalogytopreviousobservationsofknownphenomena.Givena

behaviorofsomephenomenonfromanunknowndomain,PHINEASretrievedananalogousbehavior

fromaknowndomain.Byconstructingananalogybetweenthetwobehaviors,thesystemwasableto

detectentitycorrespondences,andtransfermodelfragmentsfromtheknowndomaintotheunknown

domain.Asaresult,theanalogynotonlyprovidedusefulinformationforthecurrentscenario,butalso

abstractinformationintheformofmodelfragmentsthatcouldbeappliedtonewsituationsinthe

previouslyunknowndomain.Asimilarcross-domainapproachwasusedbyKlenk’sdomaintransfervia

analogy(DTA)system(Klenk&Forbus,2013).InDTA,analogieswereconstructedbetweenworked

solutions(i.e.problemswiththesolutionstepsexplicitlyidentified)tocreateamappingbetweentwo

differentdomains,e.g.rotationalkinematicsandtranslationalkinematics.Givenaworkedsolutiontoa

probleminonedomainandalibraryofworkedsolutionsfromanother,DTAusedanalogicalretrievalto

findananalogousworkedsolutionfromlibrary.Theentitycorrespondencesbetweenthetwoworked

solutionswereusedtocreateamappingbetweenthetwodifferentdomains.Thus,equationschemas

fromtheonedomaincouldbeprojectedtotheother,andviceversa.DTAisverysimilartoPHINEASin

106thesensethatobservationsfromknownandunknowndomainsarecomparedtoarriveatmoregeneral

between-domainmappings.

Morerecently,modelsofanalogyhavebeenusedtoaidinopen-domainquestionanswering

andtraditionalanalogywordproblems.IBM’sWatsonusedstructure-mappingbetweenindividual

questionsandpotentiallyusefultextpassagesasameasureofsemanticsimilarity(Murdock,2011).

Whenstructure-mappingwasusedonmoreabstractrepresentationsthantheindividuallexicalitems

withinagivenquestion,Watson’sabilitytoidentifytextpassagesthatprovidedevidenceforparticular

answersimproved(Murdock,2011).BoteanuandChernova(2015)developedanapproachfor

answeringtraditionalanalogywordproblems(e.g.A:B::C:D)thatusedconceptsinsemanticnetworks.

Giventwoconceptsinthenetwork,theirsimilaritywascomputedbasedonrelationalpathways

betweenthem.Thesimilarityofthosepathwayswastheproportionofcommonrelations.These

relationalpathwayswerealsousedtogenerateexplanationsforanyanswersgenerated,whichoften

containedsalientrelationalsimilarities.Inadomain-specific,butmorecomplexproblemtopic,

Chaudhrietal.(2014)usedalargescaleknowledgebasetoanswercompareandcontrastproblems

betweentopicsinbiology.Analogicalmappingsbetweentopicswereusedtogeneratetableswhere

informationinbothtopicswasaligned,inordertohighlightimportantsimilaritiesanddifferences.In

eachofthesesystems,thenotionofanalogicalreasoningisimplementedindifferentways,butthe

purposeineachistoeitherprovidesomemeasureofsimilaritybetweentwodescriptionsortoidentify

meaningfulsimilaritiesanddifferences.

6.2 KnowledgeCapture

Themostdirectapplicationofanalogicalreasoningtoknowledgecapturerequiresunderstanding

explicitanalogiesfromnaturallanguage.BarbellaandForbus’(2011)analogicaldialogueactsmodel

107providesaframeworkfortranslatingutterancesintodirectivesforbuildinganalogicalmappings.Their

modelhasbeenintegratedintoEANLUandwasusedintheinterpretationmodelinChapter4.

Analogyhasalsobeencombinedwithcrowdsourcingasameanstocapturecommonsense

knowledge(Chklovski,2003).Chklovski’ssystem,Learner,usesmultipleanalogiesoverconceptsthat

havebeenpreviouslyenteredbyhumanvolunteerstoposequestionsaboutnewtopicsthatarechosen

bythehumanvolunteer.Mappingsbetweenthenewtopicandpreviouslyenteredtopicsareusedto

identifypropertiesthatareknownaboutsimilartopicsbutnotknownforthecurrentown.Ratherthan

transferringthevalueofthepropertytothenewtopic,thepropertyisformulatedasaquestion,which

isposedtothehumanvolunteer.Intermsofstructure-mappingtheory,thisisakintousing

crowdsourcingtoaccept,reject,orclarifycandidateinferencesbetweentopics.Speeretal.(2008)used

anon-structuralapproachtoanalogicalreasoningandappliedittotheConceptNetsemanticnetwork.

TheyusesingularvaluedecompositiontoreduceknowledgefromConceptNetalongkeyfeature

dimensions.Featureoverlapcanbeeasilycomputedviadotproduct,butthiscomesatthelossof

structuredrelationalinformation.

Intheareaofmultimodalreasoning,Ferguson’sJUXTAsystem(Ferguson&Forbus,1998)was

anearlyapproachforinterpretingandcritiquingdiagramswithjuxtaposedscenarios.InJUXTA,analogy

wasusedtodetectdifferencesinthespatialrepresentationofthediagram,andlabelthembasedonthe

diagram’scaption.Themultimodalknowledgecapturesystem(MMKCAP)fromLockwoodandForbus

(2009)usedstructure-mappingtoalignrepresentationsfromdifferentmodalities.Lockwood’ssystem

readsimplifiedEnglishversionsofpassagesandsketchedversionsofdiagramsfromaphysicstextbook

onbasicmachines.Giventhatpassagestypicallyrefertotheiraccompanyingdiagramsandviceversa,

structure-mappingwasusedtoaligninformationfromthepassageswithinformationfromthediagram.

108Theanalogicalmappingindicatedhowcorrespondingitemscouldco-referandthereforeprovideda

templateforhowbothinformationsourcescouldbemergedintoanintegratedcase.Integratedcases

forbasicmachinesfromthefirstchapterwerecapturedusingthisapproach.Theresultingcaseswere

thenusedtoanswerquestionsfromthatchapter.ThemultimodalintegratedapproachinChapter4

extendsLockwood’sworkbecauseitusesautomaticlanguagedisambiguationandeventinterpretations

toaddstructuralsimilaritybetweenthesketchandtextinterpretations.

6.3 TutoringSystems

Incomparisontoanalogicalreasoningingeneral,therehavebeenrelativelyfewattemptsat

incorporatinganalogicalreasoningintointelligenteducationalsoftware.Forteachingconceptual

knowledge,theuseofanalogieshasbeenexploredinthreesystems.Murrayetal.(1990)designeda

tutoringsystembasedontheideaofbridginganalogies(Clement,1993),whicharedesignedto

graduallychangemisconceptionsbycomparingamisunderstoodscenariowithawell-understood,

intuitivescenario,calledananchor.Ifthestudentcannotremedythemisconceptionbycomparisonto

theanchor,thenabridgescenarioissummonedtofacilitateinferenceprojectionfromtheanchortothe

bridgeandeventuallytotheinitiallymisunderstoodscenario.Inthissystem,thescenarioswhichwere

usedforinstructionweredeterminedempirically.Theywereorganizedinaconceptualnetwork,such

thatattheendofeachcomparison,thesystemusedthestructureofthenetworktodeterminewhich

analogshouldbeusednext.Eachcomparisonactivityconsistedofaforcedchoicequestionanda

confidenceestimatefromthestudent.Ifthestudentansweredthequestionincorrectly,ananalogy

betweentheinitialsituationandanintuitiveanchorwasinvoked.Whenthestudentanswersthe

questioncorrectly,analogsareselectedalongthegraphgoingclosertotheoriginalscenario,thereby

graduallyencouragingthestudenttoprojectinferencesfromtheanchortotheoriginallymisunderstood

situation.Thus,themaininstructionalstrategyofthistutorisinitsselectionofanalogsasdetermined

109bythescenarionetwork.Thereisnofine-grainedfeedbackduringthecomparisonprocess.Bycontrast,

anotheranalogytutoringsystemusedhierarchicalplanoperatorstoprovidedomain-specific,fine-

grainedfeedbackforanalogiesaboutthecirculatorysystem(Lulisetal.,2004;Lulis,2005).Theseplans

weremodeledafteranalysesoftutoringdialoguesbetweenstudentsandexperttutors(Lulis&Evens,

2003)andeachanalogyrequireddifferententry-levelplans.Thismeansthatalthoughthesystemwas

abletoprovidedetailedfeedbackinnaturallanguage,theteachingstrategieswerenotdomaingeneral.

Morerecently,Alizadehetal.(2015)conductedastudythatcharacterizedcomputerscience(i.e.data

structures)tutoringdialoguesandHarsleyetal.(2016)usedthosecorpusanalysestoincorporate

analogiesintotheChiQattutoringsystem.Thisisapromisingareaofresearch,althoughalltutoring

systemstodatethathaveincorporatedanalogiesfocusononedomain,usuallyforarestrictedsetof

topics,andarenotabletohandlenovelanalogies.

Giventheimportanceofanalogiesforcase-basedreasoning,anaturalextensionistobuilda

case-basedtutoringsystem.ThisistherationalebehindSketchWorksheets(Yinetal.,2010),whichisa

sketch-basedtutoringsystemthatprovideson-demandfeedbackonsketches.Eachsketchworksheet

assignmenthasapre-definedsolutionassociatedwithit,sothatastudentcangetadviceontheirsketch

withrespecttothatsolution.Structure-mappingisusedtoalignandcomparethestudent’ssketchwith

thepre-definedsolution.Differencesidentifiedbytheanalogyareusedtodeterminewhatfeedback

shouldbegiventothestudent.Notethatunlikeexplicitanalogytutors(e.g.CIRCSIIM,ChiQat),Sketch

Worksheetsdoesnotuseanalogyasanexplicitinstructionaltool.Instead,itusesitasameansto

evaluateanewsketchintermsofafamiliaronetoprovideautomaticfeedback.

1106.4 Summary

AnalogicalreasoninghasbeenusedinavarietyofwaysinAIsystems.Case-basedreasoningsystems

primarilyuseanalogyasameanstotransferknowledgefrompriorsituationstonewcontextsfor

planningorproblemsolving.Someproblemsolvingsystems,likePHINEASandDTA,haveexplored

analogyasawaytotransferknowledgebetweendomainstogenerateabstractknowledge(orschemas)

thatcanthenbeappliedtonewsituationsinthesamedomain.Knowledgecapturesystemsrangefrom

languagesystemsthattrytointerpretexplicitanalogies,tofeature-basedmethodsofreducingsemantic

networksizeandcomplexity.Tutoringsystemshavebeenproposedtouseanalogyasameansfor

explaininginstructionalcontentaswellasinterpretingstudentinput.Inallthreeareas,analogyisused

asamethodforaligningcomplexrepresentations,measuringoverallsimilarity,andtransferring

knowledgefromonetopictoanother.

Currentlynosystemsusemultimodalinstructionalanalogiesforknowledgecaptureorproblem

solving.OnlyLockwood’sMMKCAPandKlenk’sDTAusedacombinationofanalogicalreasoningand

spatialreasoning.DTAexploredcross-domainanalogiestogenerateequationschemas,soitdidnot

dealwiththetypeofinstructionalanalogiesthatareusedtoteachnovicesaboutbasicscienceconcepts.

ThefocusofMMKCAPwasalsonotoninstructionalanalogies,butratheronhowtoalign

representationsfromdifferentmodalities.Mostofthequestionansweringsystemsusedanalogyasa

wayofassessingwithin-domainsimilarity(e.g.asinWatson)oridentifyingandorganizingsimilarities

anddifferences(e.g.asinInquire,Chaudhrietal.(2014)).Usinganalogicalreasoningwithcommon

sensedomainstoanswerquestionsinatechnicaldomainhasnotyetbeenexplored.

Buildingasystemthatcaninterpretmultimodalinstructionalanalogiesisnovelanduseful.

NoveltyisdemonstratedbythelackofAIsystemswiththeabilitytobuildknowledgefromthetypesof

111analogiesthatareusedtoteachbeginningsciencetopics.Intermsofthecontentexpressedin

analogies,themodelthatismostcloselyrelatedtothisthesisisBarbellaandForbus’(2011)modelof

analogicaldialogueacts.Thisthesisextendsthatworkbyusingmultimodalinstructionalanalogies

(ratherthanjusttextpassages)tobuildqualitativeknowledgeaboutseveraltopics.Thisthesisalso

buildsupontheMMKCAPsystem(Lockwood&Forbus,2009)byusinganalogytocombinelanguageand

visualinformation,butunlikeMMKCAP,circumventstheneedformanuallanguagedisambiguation.

Theresultingmodelofmultimodalanalogyinterpretationopensthedoortonewkindsofknowledge

capture–viareadingorviacommunicationwithahumancollaborator–andtonewkindsofintelligent

tutoringsystemsthatusemultimodalinstructionalanalogiestoteachbasicscienceknowledge.

Chapter7: Conclusion

Theanalysesinthisthesisaddressthefollowingquestionsaboutinstructionalanalogiesforteaching

middleschoolscience:

Whataretheseanalogiesaboutandhowaretheyused?

Whatarethereasoningrequirementsforinterpretingthem?

Dotheyprovideknowledgethatisusefultoanintelligentsystem?

TheanalysisinChapter3illustratesthatanalogiescanbeusedforawiderangeoftopicsin

introductoryscience,includingtopicsinbiology,chemistry,earth/spacescience,andphysics.These

analogiesappealtocommonsensenotionsofchange,behavior,possession(includingpart-whole

relationships),andrelativesize.MostoftheanalogiesintheFARguidefocusonfunctionsand

behaviorsofthingsandtheirparts.Otheranalogiesconveyinformationaboutrelativesizeor

magnitude,whichishelpfulforcommunicatingideasaboutverylargeorverysmallscales.The

112overwhelmingmajorityofanalogiesusespatialrepresentationsorvisualinputofsomekind.The

analogiesusebasedomainsthatmostpeoplearefamiliarwith,likehumanactivitiesinacityand

unlockinglocks.Interestingly,someanalogiescreateadhocbasedomains,oftenintheformofphysical

demonstrations,toconveyanewtopic.Forexample,walkingupanescalatorthatismovingdownward

isaveryspecificsourceofknowledgeforlearninghomeostasis.Soisawebofstudentsconnectedby

stringtoteachaboutecosystems.Studentsaremuchlesslikelytohaverealexperienceswithadhoc

basedomains,buttheycanstillbeusefulbecausetheyultimatelyrelyonstudents’understandingofthe

world.Spatialrepresentationssuchasdiagramsandpicturesorgeneralvisualinputfromphysical

demonstrationsorrole-playingactivitiesarealsousedastoolstorecruitthatbackgroundknowledge.In

short,instructionalanalogiesareusedforavarietyoftopicsandtheyalmostalwaysinvolvesomekind

ofvisualstimulation.Theydifferinwhatsimilaritiesmatter(i.e.functionand/orstructure),andinhow

muchbackgroundknowledgeisneeded.

Interpretinginstructionalanalogiesrequiresnaturallanguageunderstanding,significant

backgroundknowledgeaboutthephysicalworld,andspatialreasoning.Thegoaloftheexperimentsin

Chapter4wastobuildamodelofinterpretationthatwasabletocapturequalitativescienceknowledge

aboutnewdomains.Themodelconsistedofnaturallanguageunderstanding(ofsimplifiedEnglish),

sketchunderstanding(albeitwithconceptuallabels),andstructure-mappingthatwasusedtocreate

multimodalrepresentationsandtotransferknowledgefromthebasedomaintothetargetdomain.

Backgroundknowledgeandexplicitbasedomainknowledgemadecontributionstotheoverall

performanceofthemodel,butthepresenceofsketchedinputhadthegreatestimpact.Thisisdueto

thesketchrepresentation’scriticalroleinnaturallanguagedisambiguation.However,spatial

informationalsoplaysaroleinoverallalignment,especiallywhentherearemanycontainmentspatial

relationsand/ormanyarrowstotriggervisualconceptualreasoning.

113

Lastly,thesystempresentedinChapter5demonstratedthatthequalitativescienceknowledge

capturedinChapter4wasusefulforansweringquestionsfrommiddleschoolscienceexams.Consistent

withpreviousanalysesofmiddleschoolscienceexams(Clarketal.,2013)manyquestionsrequiredtype-

levelknowledge(i.e.objecttypeproperties)andtheabilitytoprojectbackgroundknowledgeinto

examplescenarios.Ofthe14questionsthatwereexamined,onlyoneinvolvedquantitativevalues

(Table14,question12).Evenstill,thisquestiondidnotrequirequantitativecalculationsoranumerical

simulation.Itsimplyrequiredatypelevelequalityrelationship(althoughthiswasoneproblemtheQA

systemdidnotsolve).Alltheotherquestionsrequiredknowledgeofthegeneralpartsandfunctionsof

classesofobjectsortheabilitytoimportqualitativecausalinformationintoanexamplesituation.

7.1 ClaimsRevisited

Claim1:Structure-mappingcanbeusedtobuildqualitativeknowledgefrommultimodalinstructional

analogies.

1) Multimodalinstructionalanalogiesusevisualrepresentationstofacilitateinterpretation.

2) Instructionalanalogiesusebackgroundknowledgetofacilitateinterpretation.

3) Multimodalintegrationandanalogyinterpretationcanbeachievedusingstructure-mapping.

Claim1wassupportedbytheexperimentsinChapter4,whichshowedthatstructuremappingcouldbe

usedbothasawaytointegratemultimodalinformationandasawaytobuildcross-domainanalogies

suchthattheinterpretationresultedintheintendedtargetknowledge.Ablationexperimentsillustrated

thattheuseofbackgroundknowledge,explicitbase-domainknowledge,andespeciallyvisual-

conceptualknowledgefromthesketchwereallsubstantialcontributorstothemodel’sperformance.

Itisimportanttonote,however,thattheevidencegatheredinChapter4cannotbetakenasa

strongreflectionoftherelativeimportanceofbackgroundknowledgeandspatialreasoninginpeople.

114AsmentionedinChapter4,theinterpretationmodeldoesnotcompletelyaccountforwhybackground

andexplicitbasedomainknowledgeissoimportantforpeople.Thevalueofcommonsensedomainsis

thattheyarerichlyconnectedtootherexperiencesandconceptualknowledge.Thebaseknowledge

thattheinterpretationmodelusesisconnectedtoontologiesinCycandexplicitbasedomain

knowledgefromtheanalogy,buttheseconnectionsarenothingcomparedtotheinterconnectednessof

humanknowledge.Withoutcomparableinterconnectednessandflexibilityofbackgroundknowledge,I

donotexpectthattheimportanceofbasedomaintransfercanbecapturedaccuratelybythis

interpretationmodel.Modelingbackgroundandcommonsenseknowledgeisanentireresearchareain

itself,soitisnotsurprisingthatshortcomingsinthatareanegativelyimpactthemodel’sfidelityto

humanreasoning.

Claim2:Qualitativeknowledgecapturedviamultimodalinstructionalanalogiescanbeusedtoanswer

questions.

TheQAsystemdescribedinChapter5wasdevelopedtotestthisclaim.Notsurprisingly,thebreadthof

topicsinthebiologyandelectricalenergyquestionsintheexamswasnotfullycoveredbythe

instructionalanalogiesfromChapter4.However,forthesetof14examquestionswheretheanswers

overlappedatleastpartiallywiththeintendedtargetknowledgeoftheanalogies,theQAsystemwas

abletocorrectlyanswer11questions.Duringquestionanswering,thesystemhadaccesstothe

knowledgelearnedfromanalogiesandallofthehumanphysiologyanduniversalvocabulary

informationintheCycKB.Thequalitativescienceknowledgelearnedfromtheanalogiesplayeda

criticalroleinquestionanswering,sincethesystemcouldnotansweranyofthequestionswhenitdid

notuseknowledgecapturedfromtheinstructionalanalogiesinChapter4.Thisillustratesthe

importanceoftype-level,qualitativeknowledgeinansweringthesequestions.Sincemanyofthe

115questionsinvolvedexamplesituations,thisresultalsoillustratesthatthetype-levelknowledgecanbe

deployedintoinstance-leveldescriptions,indicatingitsutilityforreasoningaboutnovelsituations.

7.2 OpenQuestionsandFutureWork

Oneofthebiggestchallengestointerpretingmultimodalinstructionalanalogiesandquestionanswering

isnaturallanguageunderstanding.Someofthechallengesencounteredintheconstructionofthis

interpretationmodelinclude:disambiguation,referenceresolution,andgenerics.Fordisambiguation,

conceptuallylabeledsketcheswereacrucial.Onecanarguethatthissimplyside-stepsthe

disambiguationissuebybeingheavilyinfluencedbyamanuallydisambiguatedsketch.Better

disambiguationheuristicsareneededtoreducethemodel’sdependenceontheconceptuallabelsform

thesketch.Additionally,instructionalanalogiescanbepresentedwithoutspatialrepresentations,so

language-onlydisambiguationisstillanunfulfilledrequirement.Experimentswithnarrativefunctions

(McFateetal.,2014)andanalogicalwordsensedisambiguation(Barbella&Forbus,2013)couldhelp

improvethisprocess.AsnotedintheanalysisoftheFARguideanalogies,choosingtherightlevelof

abstractionforrepresentingfunctionsandbehaviorsiscritical.Alternatively,itcouldbebettertoavoid

irreversiblesemanticinterpretationchoicesaltogether.Aninformativeexperimentwouldbetoattempt

toconstructacross-domainanalogywithoutmakingsemanticinterpretationchoices,andusing

commonalitiestoprovideevidenceforparticularsemanticinterpretations.Ifgoalinferenceor

correspondenceinformationwereavailable(e.g.fromanalogicaldialogueacts),thosecouldguidethe

interpretationchoicesevenfurther.Forreferenceresolution,challengesarosewhenthesameobject

wasreferredtodifferently.Creatingdeterministicrulestohandlethesecasesisdifficult.Insomecases,

theyshouldcorefer,e.g.rechargeablebatteryandchargedbattery.Inothercases,theyshouldbe

distinct,e.g.rechargeablebatteryanddrainedbattery.Theimportanceofthedistinctiondepends

highlyoncontext.InthecaseoftheATP/batteryanalogy,thedifferencebetweenarechargeable

116batteryandachargedonecanbeignored,whilethedifferencebetweenarechargeablebatteryanda

drainedonecannot.Anotherissueistheproblemofgenerics.Theinterpretationmodelforthisthesis

takesanaggressiveapproachtogenerics,assumingthatmostrelationsmentionedinananalogycanbe

generalizedsomehow,overthecollectionsofprimaryactorsinanevent,overthecollectionsofwholes

inpart-wholerelationships,oroverthecausesinqualitativecausalinfluences.Thishappenstowork

wellinthecontextofinstructionalanalogiesforrecall,butlesssoforprecision.Aswouldbeexpected

ofanovelcross-domainanalogy,thesystemoverproducesandisveryliberalwhenitcomestoaccepting

inferences.Thisisusefulforinitializingatargetdomainifknowledgerefinementispostponed.Having

overlygeneral(orunderspecified)type-levelstatementsisveryusefulbecauseitmeansthatthe

knowledgeisapplicableinawiderangeofsituations.However,italsoincreasesthelikelihoodof

creatingfalseknowledge.Maintainingabalancebetweenoverlygeneralandoverlyspecificstatements

isanopenproblem.Abetterunderstandingofhowtointerpretgenericsmorepreciselyinthecontext

ofinstructionalanalogieswouldgreatlyimproveinterpretation.Thismodelalsoneedstobeevaluated

onalargersetofanalogies,bothinnewdomainstobroadencoverage,butalsoonexistingtopicsto

explorehowmultipleanalogieswouldhelprefineoverproducedknowledge.

Anotherpotentialareaforfutureworkisinteractivityforincrementalnaturallanguage

understandingandincrementalanalogicalmatching.Theinterpretationmodelbuiltforthisthesisis

fairlylinear.Analternativeapproachwouldbetohaveahumancollaboratorintheloopforatleast

somepartsofknowledgecapture.Insomecasesoflanguageinterpretation,therearecompeting

choicesthatareequallyvalid,butonelessfavorableduetothecontextoftheanalogy.Ifthe

interpretationmodelhadtheabilitytoposequestionstoahumancollaborator,itcouldengageinactive

learningandpossiblygeneratemodelsforwhatlevelsofabstractionarefavorableforinstructional

analogies.Asimilarapproachcouldbeusedformultimodalintegration,butfromearlyexperiments,itis

117notclearthatitwouldbebeneficial.Earlyon,Iexperimentedwithanincrementalversionof

multimodalintegration,whereinsteadoffullyinterpretingthetextandmergingitwiththesketch,I

beganwiththesketchinformationandincrementallyaddedknowledgefromeachsentence,merging

entitiesforeachaddition.Forsimplesituations,thisaddressedissuesincoreferenceaswell:ifthere

wasonenucleusinthesketch,and“thenucleus”wasmentionedinthreedifferentsentences,eachtime

asentencewasadded,thenucleiwouldmergeandthefinalrepresentationwouldhaveonenucleus.

Thisignoredthenuanceofreferenceresolution,however,asitalsowouldmergeeventsthathad

differentactors.Forexample,considerthesetwosentencesfromthecell/cityanalogy:

Thepowerstationprovideselectricity.Themitochondrionprovideschemicalenergy.

InterpretingthefirstsentencewouldresultinaninstanceofMakingSomethingAvailable,wherethe

powerstationismakingelectricityavailable.Thesecondsentencewouldresultinanotherinstanceof

MakingSomethingAvailablewherethemitochondrionmakeschemicalenergyavailable.Asimple

incrementalmergingapproachwouldresultinthesetwoeventinstancesbeingmerged,eventhough

theyinvolvedifferentactors.Ontheotherhand,thisincrementalapproachmaybeusefulforcross-

domainanalogies,sothatinferencescouldbeevaluatedincrementally.Themodelcurrentlyuses

incrementalmatchingtoincorporateinferencesandextendcross-domainmatches.Havingahuman,or

someself-evaluationstep,intheloopforthisprocesscouldgreatlyimprovethematchingbecauseit

wouldpreventincorrectinferencesfromextendingthematchinanunintendedway.Consequently,

overproductionwouldbereduced,improvingtheoverallqualityofthefinaltargetrepresentations.

Exploringtechniqueslikethese,orothersthatmaketheinterpretationmodelmoreincrementalcould

bebeneficialtothemodel’stransparencyandflexibility.

118 Moreinvestigationsarealsoneededtoassessthegeneralityofthisapproachandofthe

characterizationofinstructionalanalogiesinChapter3.Analogiesareusedinmanyotherdomains

beyondtheonesexploredhere.Forexample,theanalogiesanalyzedbyRichlandetal.(2007)involved

proceduralknowledgeforsolvingmathematicsproblems.UsingthecategoriesidentifiedbyAlfierietal.

(2013),the11analogiesexploredhereallinvolvedconceptualknowledge(ratherthanstrictly

proceduralorperceptualknowledge).Howwelltheinterpretationsstrategiesidentifiedhereworkwith

othertypesofanalogies,i.e.proceduralandperceptual,isstillanopenquestion.Largercorpusanalyses

couldrevealhowfarthevisualconceptualrelationsdescribedin4.2.2.1coverspatialrepresentations

usedininstructionalanalogies.Similarly,suchcorpusanalysescouldbeusedtoexpandthesetofvery

abstractterms(e.g.Event,PartiallyTangible,Individual)anddetermineifitisevenpossibletoarriveata

setofabstracttermsthatcanbeusedforaverywiderangeofanalogytopics.

Thisworkleadstoanexcitingintersectionofanalogicalreasoning,spatialreasoning,and

intelligenttutoring.Oneofthewaysinwhichmultimodalreasoningcouldbeusefulisformultimodal

instruction.Intelligentsystemsthathavetheabilitytounderstandvisualrepresentationspairedwith

textareneededtocaptureknowledgefromreadingandfrominteractingwithotherpeople.Such

systemswouldbeendowedwiththetypesofqualitativeknowledgethatwouldimprovetheirabilityto

understandtheworldandtocommunicatewithpeople.Oneofthewaysinwhichinstructional

analogiesarepowerfulisthroughSocratictutoringdialogues(forexamplesofanalogiesusedthisway,

see(Alizadehetal.,2015;Lulis&Evens,2003)).Asmentionedinrelatedwork,afewresearchershave

workedontutoringsystemsthatuseanalogiesasanexplicitinstructionaltool.But,thosesystemsare

mostlyscriptedfromcorpusanalysesofhumantutoringdialogues.Thisisanimportantfirststep,but

thesesystemscannotinterpretnovelanalogiesorproposenewones.Anintelligentsystemthatcould

engageinrichdialogueswithastudentandsupportnovelanalogiesinmultipledomainswouldbea

119hugeadvanceinthestateoftheartinintelligenttutoringsystems.Thiscannotbeachievedwithout

improvementstogeneralpurposenaturallanguageunderstandingandwithoutgeneralpurpose

cognitivemodelsofanalogicalreasoninglikestructure-mapping.

120REFERENCES

Alfieri,L.,Nokes-Malach,T.J.,&Schunn,C.D.(2013).LearningThroughCaseComparisons:AMeta-

AnalyticReview.EducationalPsychologist,48(2),87-113.doi:10.1080/00461520.2013.775712

Alizadeh,M.,DiEugenio,B.,Harsley,R.,Green,N.,Fossati,D.,&AlZoubi,O.(2015).AStudyofAnalogy

inComputerScienceTutorialDialogues.Proceedingsofthe7thInternationalConferenceon

ComputerSupportedEducation,53(19.2),6.6.

Barbella,D.,&Forbus,K.(2011).AnalogicalDialogueActs:SupportingLearningbyReadingAnalogiesin

InstructionalTexts.PaperpresentedattheTwenty-FifthAAAIConferenceonArtificial

Intelligence,SanFrancisco,CA.

Barbella,D.,&Forbus,K.D.(2013).AnalogicalWordSenseDisambiguation.AdvancesinCognitive

Systems,2,297-315.

Blanchette,I.,&Dunbar,K.(2002).Representationalchangeandanalogy:Howanalogicalinferences

altertargetrepresentations.JournalofExperimentalPsychology:Learning,Memory,and

Cognition,28(4),672.

Boteanu,A.,&Chernova,S.(2015).SolvingandExplainingAnalogyQuestionsUsingSemanticNetworks.

PaperpresentedattheAAAI.

Burstein,M.H.(1985).LearningbyReasoningfromMultipleAnalogies.YaleUniversity.

Chang,M.D.,Wetzel,J.,&Forbus,K.D.(2014).SpatialReasoninginComparativeAnalysesofPhysics

Diagrams.PaperpresentedattheInternationalConferenceonSpatialCognition,Bremen,

Germany.

Chaudhri,V.K.,Heymans,S.J.,Overholtzer,A.,Spaulding,A.,&Wessel,M.(2014).Large-Scale

AnalogicalReasoning.PaperpresentedattheTwenty-EighthAAAIConferenceonArtificial

Intelligence.

Chi,M.T.H.,Feltovich,P.J.,&Glaser,R.(1981).CategorizationandRepresentationofPhysicsProblems

byExpertsandNovices.Cognitivescience,5(2),121-152.doi:DOI10.1207/s15516709cog0502_2

121Chklovski,T.(2003).Learner:asystemforacquiringcommonsenseknowledgebyanalogy.Paper

presentedattheProceedingsofthe2ndinternationalconferenceonKnowledgecapture.

Clark,P.(2015).ElementarySchoolScienceandMathTestsasaDriverforAI:TaketheAristoChallenge!

PaperpresentedattheAAAI.

Clark,P.,Harrison,P.,&Balasubramanian,N.(2013).Astudyoftheknowledgebaserequirementsfor

passinganelementarysciencetest.PaperpresentedattheProceedingsofthe2013workshop

onAutomatedknowledgebaseconstruction.

Clement,C.A.,&Gentner,D.(1991).Systematicityasaselectionconstraintinanalogicalmapping.

Cognitivescience,15(1),89-132.

Clement,J.(1993).Usingbridginganalogiesandanchoringintuitionstodealwithstudents'

preconceptionsinphysics.Journalofresearchinscienceteaching,30(10),1241-1257.

Cohn,A.G.,Bennett,B.,Gooday,J.,&Gotts,N.M.(1997).Qualitativespatialrepresentationand

reasoningwiththeregionconnectioncalculus.Geoinformatica,1(3),275-316.

Falkenhainer,B.(1990).AUnifiedApproachtoExplanationandTheoryFormation.InJ.Shrager&P.

Langley(Eds.),Computationalmodelsofscientificdiscoveryandtheoryformation(pp.157-196).

LosAltos,CA:MorganKaufman.

Falkenhainer,B.,Forbus,K.D.,&Gentner,D.(1989).Thestructure-mappingengine:Algorithmand

examples.Artificialintelligence,41(1),1-63.

Fan,J.,Barker,K.,&Porter,B.(2009).Automaticinterpretationoflooselyencodedinput.Artificial

intelligence,173(2),197-220.

Ferguson,R.W.,&Forbus,K.D.(1998).Tellingjuxtapositions:Usingrepetitionandalignabledifference

indiagramunderstanding.Advancesinanalogyresearch,109-117.

Ferrucci,D.,Brown,E.,Chu-Carroll,J.,Fan,J.,Gondek,D.,Kalyanpur,A.A.,...Welty,C.(2010).Building

Watson:AnOverviewoftheDeepQAProject.AImagazine,31(3),59-79.

122Forbus,K.,Usher,J.,Lovett,A.,Lockwood,K.,&Wetzel,J.(2011).Cogsketch:Sketchunderstandingfor

cognitivescienceresearchandforeducation.TopicsinCognitiveScience,3(4),648-666.

Forbus,K.D.(1984).Qualitativeprocesstheory.Artificialintelligence,24(1),85-168.

Forbus,K.D.,Gentner,D.,&Ferguson,R.W.(1994).IncrementalStructure-Mapping.Proceedingsofthe

SixteenthAnnualConferenceoftheCognitiveScienceSociety,313-318.

Forbus,K.D.,Klenk,M.,&Hinrichs,T.(2009).Companioncognitivesystems:Designgoalsandlessons

learnedsofar.IEEEIntelligentsystems,24(4),36-46.

Forbus,K.D.,Usher,J.,&Tomai,E.(2005).Analogicallearningofvisual/conceptualrelationshipsin

sketches.PaperpresentedattheProceedingsoftheNationalConferenceonArtificial

Intelligence.

Friedman,S.E.,&Forbus,K.D.(2011).Repairingincorrectknowledgewithmodelformulationand

metareasoning.PaperpresentedattheProceedingsoftheTwenty-Secondinternationaljoint

conferenceonArtificialIntelligence.

Gadgil,S.,Nokes-Malach,T.J.,&Chi,M.T.(2012).Effectivenessofholisticmentalmodelconfrontation

indrivingconceptualchange.LearningandInstruction,22(1),47-61.

Gentner,D.(1983).Structure-mapping:Atheoreticalframeworkforanalogy.Cognitivescience,7(2),

155-170.

Gentner,D.(2010).Bootstrappingthemind:Analogicalprocessesandsymbolsystems.Cognitive

science,34(5),752-775.

Gentner,D.,Levine,S.C.,Ping,R.,Isaia,A.,Dhillon,S.,Bradley,C.,&Honke,G.(2015).RapidLearningin

aChildren'sMuseumviaAnalogicalComparison.Cognitivescience,1-17.

Gentner,D.,Loewenstein,J.,&Thompson,L.(2003).Learningandtransfer:Ageneralroleforanalogical

encoding.JournalofEducationalPsychology,95(2),393-408.doi:10.1037/0022-0663.95.2.393

Gentner,D.,&Medina,J.(1998).SimilarityandtheDevelopmentofRules.Cognition,65(2),263-297.

123Gentner,D.,&Namy,L.L.(1999).Comparisoninthedevelopmentofcategories.Cognitive

Development,14(4),487-513.

Gentner,D.,&Ratterman,M.J.(1991).Languageandthecareerofsimilarity.InS.A.Gelman&J.P.

Byrnes(Eds.),Perspectivesonlanguageandthought:Interrelationsindevelopment(pp.225-

275).London:CambridgeUniversityPress.

Gentner,D.,&Toupin,C.(1986).SystematicityandSurfaceSimilarityintheDevelopmentofAnalogy.

Cognitivescience,10(3),277-300.doi:DOI10.1207/s15516709cog1003_2

Gick,M.L.,&Holyoak,K.J.(1980).AnalogicalProblem-Solving.Cognitivepsychology,12(3),306-355.

doi:Doi10.1016/0010-0285(80)90013-4

Gick,M.L.,&Holyoak,K.J.(1983).Schemainductionandanalogicaltransfer.Cognitivepsychology,

15(1),1-38.

Glynn,S.M.(2008).Makingscienceconceptsmeaningfultostudents:Teachingwithanalogies.InS.

Mikelskis-Seifert,U.Ringelband,&M.Brückmann(Eds.),Fourdecadesofresearchinscience

education:Fromcurriculumdevelopmenttoqualityimprovement(pp.113-125).Münster,

Germany:Waxman.

Goldstone,R.L.,&Medin,D.L.(1994).Time-CourseofComparison.JournalofExperimentalPsychology-

LearningMemoryandCognition,20(1),29-50.doi:Doi10.1037/0278-7393.20.1.29

Harrison,A.G.,&Coll,R.K.(Eds.).(2007).UsingAnalogiesinMiddleandSecondaryScienceClassrooms:

TheFARGuide–AnInterestingWaytoTeachWithAnalogies.ThousandOaks,CA:SAGE

Publications.

Harsley,R.,Green,N.,Alizadeh,M.,Acharya,S.,Fossati,D.,DiEugenio,B.,&AlZoubi,O.(2016).

IncorporatingAnalogiesandWorkedOutExamplesasPedagogicalStrategiesinaComputer

ScienceTutoringSystem.PaperpresentedattheProceedingsofthe47thACMTechnical

SymposiumonComputingScienceEducation.

124Hestenes,D.,&Halloun,I.(1995).Interpretingtheforceconceptinventory.ThePhysicsTeacher,33(8),

502-506.

Hestenes,D.,Wells,M.,&Swackhamer,G.(1992).Forceconceptinventory.ThePhysicsTeacher,30,

141.

Holyoak,K.J.,&Richland,L.E.(2014).UsingAnalogiesasaBasisforTeachingCognitiveReadiness.InH.

O'Neil,R.Perez,&E.Baker(Eds.),TeachingandMeasuringCognitiveReadiness(pp.223-238).

NewYork,NY:SpringerPublishingCompany.

Klenk,M.,&Forbus,K.(2009).AnalogicalmodelformulationfortransferlearninginAPPhysics.Artificial

intelligence,173(18),1615-1638.

Klenk,M.,&Forbus,K.(2013).Exploitingpersistentmappingsincross-domainanalogicallearningof

physicaldomains.Artificialintelligence,195,398-417.doi:10.1016/j.artint.2012.11.002

Klenk,M.,Forbus,K.,Tomai,E.,&Kim,H.(2011).Usinganalogicalmodelformulationwithsketchesto

solveBennettMechanicalComprehensionTestproblems.JournalofExperimental&Theoretical

ArtificialIntelligence,23(3),299-327.

Kolodner,J.L.(2006).Case-BasedReasoning.CambridgeHandbookoftheLearningSciences,225-242.

doi:Book_Doi10.2277/0521607779

Kotovsky,L.,&Gentner,D.(1996).ComparisonandCategorizationintheDevelopmentofRelational

Similarity.ChildDevelopment,67,2797-2822.

Kurtz,K.J.,Miao,C.H.,&Gentner,D.(2001).Learningbyanalogicalbootstrapping.Journalofthe

LearningSciences,10(4),417-446.doi:DOI10.1207/S15327809JLS1004new_2

Lakoff,G.,&Johnson,M.(2003).Metaphorsweliveby.Chicago:UniversityofChicagoPress.

Larkin,J.H.,&Simon,H.A.(1987).Whyadiagramis(sometimes)worthtenthousandwords.Cognitive

science,11(1),65-100.

Leslie,S.-J.(2008).Generics:Cognitionandacquisition.PhilosophicalReview,117(1),1-47.

125Lockwood,K.,&Forbus,K.(2009).Multimodalknowledgecapturefromtextanddiagrams.Paper

presentedattheFifthinternationalconferenceonKnowledgecapture.

Loewenstein,J.,Thompson,L.,&Gentner,D.(1999).Analogicalencodingfacilitatesknowledgetransfer

innegotiation.PsychonomicBulletin&Review,6(4),586-597.

Lulis,E.,&Evens,M.(2003).Theuseofanalogiesinhumantutoringdialogues.Paperpresentedatthe

Naturallanguagegenerationinspokenandwrittendialogue.AAAISymposium,Stanford.

Lulis,E.,Evens,M.,&Michael,J.(2004).Implementinganalogiesinanelectronictutoringsystem.Paper

presentedattheIntelligentTutoringSystems.

Lulis,E.A.(2005).Implementinganalogiesinanintelligenttutoringsystembymodelinghumantutoring

sessions.IllinoisInstituteofTechnology.

Markman,A.B.,&Gentner,D.(1993).Structuralalignmentduringsimilaritycomparisons.Cognitive

psychology,25(4),431-467.

Mason,L.(2004).Fosteringunderstandingbystructuralalignmentasaroutetoanalogicallearning.

InstructionalScience,32(4),293-318.doi:Doi10.1023/B:Truc.0000026512.88700.32

Mason,L.,Pluchino,P.,&Tornatora,M.C.(2013).EffectsofPictureLabelingonScienceTextProcessing

andLearning:EvidenceFromEyeMovements.ReadingResearchQuarterly,48(2),199-214.

doi:10.1002/rrq.41

McFate,C.J.,Forbus,K.D.,&Hinrichs,T.(2014).UsingNarrativeFunctiontoExtractQualitative

InformationfromNaturalLanguageTexts.PaperpresentedattheProceedingsoftheTwenty-

EighthAAAIConferenceonArtificialIntelligence,QuebecCity.

Murdock,J.W.(2011).Structuremappingforjeopardy!cluesCase-BasedReasoningResearchand

Development(pp.6-10):Springer.

Murray,T.,Schultz,K.,Brown,D.,&Clement,J.(1990).AnAnalogy-BasedComputerTutorfor

RemediatingPhysicsMisconceptions.InteractiveLearningEnvironments,1(2),79-101.

126Namy,L.L.,&Gentner,D.(2002).Makingasilkpurseoutoftwosow'sears:youngchildren'suseof

comparisonincategorylearning.JournalofExperimentalPsychology:General,131(1),5.

Nokes-Malach,T.J.,VanLehn,K.,Belenky,D.M.,Lichtenstein,M.,&Cox,G.(2013).Coordinating

principlesandexamplesthroughanalogyandself-explanation.EuropeanJournalofPsychology

ofEducation,28(4),1237-1263.doi:10.1007/s10212-012-0164-z

Richland,L.E.,Zur,O.,&Holyoak,K.J.(2007).CognitiveSupportsforAnalogiesintheMathematics

Classroom.Science,316(5828),1128-1129.doi:10.1126/science.1142103

Speer,R.,Havasi,C.,&Lieberman,H.(2008).AnalogySpace:ReducingtheDimensionalityofCommon

SenseKnowledge.PaperpresentedattheAAAI.

Spiro,R.J.,Feltovich,P.J.,Coulson,R.L.,&Anderson,D.K.(1989).MultipleAnalogiesforComplex

Concepts-AntidotesforAnalogy-InducedMisconceptioninAdvancedKnowledgeAcquisition.

SimilarityandAnalogicalReasoning,498-531.doi:Doi10.1017/Cbo9780511529863.023

Tomai,E.,&Forbus,K.D.(2009).EANLU:PracticalLanguageUnderstandingforCognitiveModeling.

PaperpresentedattheProceedingsofthe22ndInternationalFloridaArtificialIntelligence

ResearchSocietyConference,SanibelIsland,Florida.

Veloso,M.M.,&Carbonell,J.G.(1993).DerivationalAnalogyinProdigy-AutomatingCaseAcquisition,

Storage,andUtilization.MachineLearning,10(3),249-278.doi:Doi10.1023/A:1022686910523

Wai,J.,Lubinski,D.,&Benbow,C.P.(2009).SpatialabilityforSTEMdomains:Aligningover50yearsof

cumulativepsychologicalknowledgesolidifiesitsimportance.JournalofEducationalPsychology,

101(4),817.

Yin,P.,Forbus,K.D.,Usher,J.M.,Sageman,B.,&Jee,B.D.(2010).SketchWorksheets:ASketch-Based

EducationalSoftwareSystem.PaperpresentedattheInnovativeApplicationsofArtificial

Intelligence(IAAI).

Zeitoun,H.H.(1984).TeachingScientificAnalogies:aproposedmodel.ResearchinScience&

TechnologicalEducation,2(2),107-125.doi:10.1080/0263514840020203

127

APPENDIXA:InstructionalAnalogies

ATP/battery

Text:ATPislikeachargedbattery.Thechargedbatteryhaselectricalenergy.Thechargedbatteryreleaseselectricalenergytoelectronicdevices.Thechargedbatterychangesintoadrainedbatterywhenelectricalenergyisused.Thechargedbatterycanbeusedoverandoveragain.Abatterychargerchangestheflatbatteryintothechargedbattery.TheATPhaschemicalenergy.TheATPreleaseschemicalenergytocellparts.TheATPchangesintoADPwhenchemicalenergyisused.TheATPcanbeusedoverandoveragain.MitochondriachangesADPintoATP.TheADPislikethedrainedbattery.Thechargedbatteryprovidesenergygradually,buttheATPprovidesenergyimmediately.AphosphatebreaksofffromtheATP,butnothingbreaksofffromthebattery.Devicesusuallyusetwobatteries,butcellsusemanyATPmolecules.

Sketchobjects:(isa using-atp IntrinsicStateChangeEvent) (isa ADP AdenosineDiphosphate) (isa charging-the-battery ChargingABattery) (isa ATP AdenosineTriphosphate) (isa atp-energy EnergyStuff) (isa battery-charger BatteryCharger) (isa making-atp IntrinsicStateChangeEvent) (isa phosphate Phosphate)

128(isa battery-energy EnergyStuff) (isa charged-battery ChargedBattery) (isa drained-battery FlatBattery) (isa Object-17 BreakingEvent) (isa using-battery IntrinsicStateChangeEvent)) GoldStandardQueriesinEnglish,CycL:FourrightmostcolumnsshowwhichqueriessucceededinthefourexperimentalconditionsinChapter4.

SuccessfullycapturedbyModel?QueryinEnglish

QueryinCycL FullModel

NoBack-ground

NoBase

NoSketch

ATPisenergysufficient.

(relationAllExists possessiveRelation AdenosineTriphosphate EnergyStuff)

P P P NO

ATPreleasesenergytoparts/placesinthecell.

(relationExistsAll performedBy (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn MakingSomethingAvailable target CellPart) transferredObject EnergyStuff) AdenosineTriphosphate)

P P P NO

ADPconvertstoATP.

(relationAllExists toState (IntrinsicStateChangeOfFn AdenosineDiphosphate) AdenosineTriphosphate)

P P NO NO

ATPconvertstoADP.

(relationAllExists toState (IntrinsicStateChangeOfFn AdenosineTriphosphate) AdenosineDiphosphate)

P P NO NO

ATPcanbeusedoverandoveragain.

(relationAll repeatedEvent (UsingInstanceFn AdenosineTriphosphate))

P P NO NO

MitochondriaarethesiteswhereADPgetschangedtoATP.

(relationExistsAll doneBy (IntrinsicStateChangeOfFn AdenosineDiphosphate) Mitochondrion)

P P NO NO

Battery'senergyisreleasedgradually,butATPenergyisreleased

(relationAllInstance conceptuallyRelated (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn MakingSomethingAvailable objectActedOn EnergyStuff) performedBy AdenosineTriphosphate) Immediately)

P P P NO

129immediately. AphosphatebreaksawayfromATP.

(relationAllExists from-UnderspecifiedLocation (SubcollectionOfWithRelationToTypeFn BreakingEvent objectOfStateChange Phosphate) AdenosineTriphosphate)

P P P P

ManyATPmoleculesareused.

(relationAllInstance qualitativeExtent (SubcollectionOfWithRelationFromTypeFn (SubcollectionOfWithRelationFromTypeFn Molecule compoundNoun AdenosineTriphosphate) instrument-Generic (UsingInstanceFn (SubcollectionOfWithRelationFromTypeFn Molecule compoundNoun AdenosineTriphosphate))) Many-Quant)

NO NO NO NO

EnergyisrequiredtochangeADPintoATP.

(relationAllExists requires-Underspecified (IntrinsicStateChangeOfFn AdenosineDiphosphate) EnergyStuff)

NO NO NO NO

EnergyisreleasedwhenATPchangedintoADP.

(relationAllExists releases-Underspecified (IntrinsicStateChangeOfFn AdenosineTriphosphate) EnergyStuff)))

P NO NO P

130Cell/city

Text:Acellislikeacity.Thenucleusislikethecitygovernment.Thecitygovernmentcontrolsthecity.Thenucleuscontrolsthecell.Themitochondrionislikethepowerstation.Thepowerstationprovideselectricity.Themitochondrionprovideschemicalenergy.Constructioncompaniesbuildhouses.Theribosomesmakeproteins.Thingsinthecitymovealongtheroad.Theendoplasmicreticulumisliketheroad.Golgibodiesexportsubstancesoutsideofthecell.Factoriesexportthingsoutsideofthecity.Thecitygovernmentchangesdirectionafterelectionsandisveryadaptable.Unlikethecity,thenucleusalwayscontrolsthecell.

Sketchobjects:(isa mitochondrion1 Mitochondrion) (isa endoplasmic-reticulum EndoplasmicReticulum) (isa membrane CellMembrane) (isa ribosome3 Ribosome) (isa construction-companies ConstructionCompany) (isa city-limits Border) (isa golgi-bodies GolgiApparatus) (isa roads Roadway) (isa roads TransportationPathSystem) (isa house-1 House-Modern) (isa factory FactoryBuilding) (isa the-cell Cell) (isa house-3 House-Modern) (isa windmills WindPowerPlant) (isa library SchoolBuilding) (isa the-city City) (isa nucleus CellNucleus) (isa city-hall CityGovernment) (isa ribosome1 Ribosome)

131 (isa mitochondrion2 Mitochondrion) (isa ribosome2 Ribosome)) GoldStandardQueriesinEnglish,CycL:

FourrightmostcolumnsshowwhichqueriessucceededinthefourexperimentalconditionsinChapter4.SuccessfullycapturedbyModel?

QueryinEnglish QueryinCycL FullModel

NoBack-ground

NoBase

NoSketch

Cellshavecellmembranes.

(partTypes Cell CellMembrane) P P P

NO

Cellshavecellnuclei.

(partTypes Cell CellNucleus) P P P

NO

Cellshaveribosomes.

(partTypes Cell Ribosome) P

NO NO NO

Cellshaveendoplasmicreticulum.

(partTypes Cell EndoplasmicReticulum)

P P P NO

Cellshavemitochondria.

(partTypes Cell Mitochondrion) P P P

NO

Cellnucleicontrolcells.

(relationExistsAll performedBy (SubcollectionOfWithRelationToTypeFn ControllingSomething objectControlled Cell) CellNucleus)

P P P P

Mitochondriaprovideenergy.

(relationExistsAll performedBy (MakingAbstractAvailableFn EnergyStuff) Mitochondrion)

P P NO NO

Ribosomesconstructproteins.

(relationExistsAll performedBy (MakingFn ProteinStuff) (SetOfTypeFn Ribosome))

P P NO NO

Theendoplasmicreticulumtransportsthings.

(relationExistsAll doneBy Movement-TranslationEvent EndoplasmicReticulum)

NO NO NO NO

Cellfunctionsareinterdependent.

(relationAllExists requires-Underspecified (TypicalBehaviorOfTypeFn Cell) (PartTypeFn Cell Individual))

NO NO NO P

Cellscommunicate

(relationAllExists communicatesWith Cell Cell) P

NO NO P

132withothercells. Individualcellpartsmakeupafunctionalsystem.

(relationExistsAll members FunctionalSystem (PartTypeFn Cell Individual))))

P NO NO P

133Cell/Earth

Text:AcellislikeplanetEarth.Acellnucleusislikeacontinent.Achromosomeislikeacountry.Ageneislikeacity.Acodonislikeastreetaddress.

Sketchobjects:(isa codon Codon-MolecularSegment) (isa street-address StreetAddress) (isa gene Gene-HereditaryUnit) (isa continent Continent) (isa cell Cell) (isa city City) (isa chromosome Chromosome) (isa nucleus CellNucleus) (isa country Country) (isa PlanetEarth Planet)

GoldStandardQueriesinEnglish,CycL:


QueryinEnglish


NoBack-ground

NoBase

NoSketch

Cellsarebiggerthancellnuclei.

(relationAllExists biggerThan Cell CellNucleus)

P P NO NO

Cellnucleiarebiggerthan

(relationAllExists biggerThan CellNucleus Chromosome) P P

NO NO

134chromosomes.Chromosomesarebiggerthangenes.

(relationAllExists biggerThan Chromosome Gene-HereditaryUnit) P P

NO NO

Genesarebiggerthancodons.

(relationAllExists biggerThan Gene-HereditaryUnit Codon-MolecularSegment)

P P P NO

Cellshavecellnuclei.

(relationAllExists physicalParts Cell CellNucleus)

P P P NO

Cellnucleihavechromosomes.

(relationAllExists physicalParts CellNucleus Chromosome)

P P P NO

Chromosomeshavegenes.

(relationAllExists physicalParts Chromosome Gene-HereditaryUnit)

P P P NO

Geneshavecodons.

(relationAllExists physicalParts Gene-HereditaryUnit Codon-MolecularSegment)

P P NO NO

Cellnucleihavemanychromosomes.

(relationAllExistsRange physicalParts CellNucleus Chromosome Many-Quant) P

NO NO NO

Chromosomeshavemanygenes.

(relationAllExistsRange physicalParts Chromosome Gene-HereditaryUnit Many-Quant)

P NO NO NO

Geneshavemanycodons.

(relationAllExistsRange physicalParts Gene-HereditaryUnit Codon-MolecularSegment Many-Quant)

P NO NO NO

135Circuit/Aquarium

Text:Asimpleseriescircuitislikeanaquarium.Theelectriccurrentislikethewater.Thewireislikethepipe.Thewirecarriestheelectricity.Thepipecarriesthewater.Thebatteryusesvoltagetopushelectrons.Thepumpusespressuretopushthewater.Voltageislikepressure.Thinwireinthelightbulbresistselectriccurrent.Thefilterresistswaterflow.Electriccurrentisconserved.Nowaterislost.Waterisamaterialliquid.Electricityisaflowofchargeinanelectricfield.Watercanflowinanincompletecycle.Electricityalwaysneedsacompletecircuit.Waterflowdependsonthepressureofthepump.Electriccurrentdependsontheentirecircuit.

Sketchobjects:((isa water-pipe Pipe-GenericConduit) (isa water-pump Pump-Generic) (isa the-filter Filter) (isa Object-302 Aquarium-Container) (isa circuit-wire-2 Wire) (isa the-battery Battery) (isa the-light-bulb LightBulbIncandescent) (isa pump-pressure Pressure) (isa water-arrow DirectionOfMovement) (isa aquarium-water Water) (isa aquarium-water Liquid-StateOfMatter) (isa aquarium-container Container) (isa aquarium-container Physob) (isa sandy-bits SandySoilRegion) (isa circuit-wire-1 Wire) (isa Object-305 SeriesCircuit) (isa fishface Goldfish))

136GoldStandardQueriesinEnglish,CycL:FourrightmostcolumnsshowwhichqueriessucceededinthefourexperimentalconditionsinChapter4.



NoBack-ground

NoBase

NoSketch

Batteriespushelectrons(voltage).

(relationExistsAll providerOfMotiveForce (SubcollectionOfWithRelationToTypeFn PushingAnObject objectActedOn Electron) Battery)

P P P NO

Flowofelectricitydependsonacompletecircuit.

(dependsOn-Underspecified Electricity (CollectionIntersectionFn (TheSet ElectricalCircuit (ThingDescribableAsFn Entire-TheWord Adjective-Gradable))))

NO NO NO

Wiresresisttheflowofelectricity.

(relationExistsAll doneBy (SubcollectionOfWithRelationToFn PreventingSomething objectActedOn Electricity) Wire)

NO NO NO

Circuitshavewires.

(partTypes SeriesCircuit (SetOfTypeFn Wire))

P P P NO

Circuitshavebatteries.

(partTypes SeriesCircuit Battery) P P P

NO

Circuitshavelightbulbs.

(partTypes SeriesCircuit LightBulbIncandescent)

P P NO

137DNA/pins

Text:DNAislikeaclothespinstructure.Cytosineislikeagreenclothespin.Guanineislikeanorangeclothespin.Thymineislikearedclothespin.Adenineislikeablueclothespin.Thesugarphosphatebackboneislikethetube.Weakhydrogenbondsareliketheclothespinattachments.Anyofthecoloredclothespinswillcliptogether.Cytosinewillbondonlywithguanine.Thyminewillbondonlywithadenine.Theclothespinstructureisshort,butDNAmoleculesaretypicallyverylong.DNAmoleculestypicallyhavethousandsofbasepairs.

Sketchobjects:((isa red-3 ClothesPin) (isa red-3 ColoredThing) (isa C-2 Cytosine-Base) (isa the-tubing Pipe-GenericConduit) (isa C-3 Cytosine-Base) (isa G-3 Guanine-Base) (isa dna-backbone SugarPhosphateBackbone) (isa T-1 Thymine-Base) (isa T-1 ChemicalObject) (isa blue-2 ClothesPin) (isa blue-2 ColoredThing) (isa orange-1 ClothesPin) (isa orange-1 ColoredThing) (isa green-1 ClothesPin) (isa green-1 ColoredThing) (isa orange-3 ClothesPin) (isa G-1 Guanine-Base) (isa Object-155 HydrogenBond) (isa Object-121 Attachment) (isa Object-149 HydrogenBond) (isa red-2 ClothesPin) (isa red-2 ColoredThing) (isa T-3 Thymine-Base) (isa Object-109 Attachment) (isa T-2 Thymine-Base) (isa Object-107 Attachment) (isa Object-145 HydrogenBond) (isa Object-153 HydrogenBond) (isa A-1 ChemicalObject) (isa A-1 Adenine-Base) (isa green-3 ClothesPin) (isa blue-1 ClothesPin) (isa blue-1 ColoredThing) (isa green-2 ClothesPin) (isa green-2 ColoredThing) (isa blue-3 ClothesPin) (isa blue-3 ColoredThing)

138 (isa A-2 Adenine-Base) (isa Object-111 Attachment) (isa Object-113 Attachment) (isa Object-105 Attachment) (isa C-1 Cytosine-Base) (isa red-1 ColoredThing) (isa red-1 ClothesPin) (isa Object-147 HydrogenBond) (isa orange-2 ClothesPin) (isa orange-2 ColoredThing) (isa Object-151 HydrogenBond) (isa the-dna-molecule DNAMolecule) (isa the-dna-molecule SpatialThing-NonSituational) (isa Long Distance)) GoldStandardQueriesinEnglish,CycL:


QueryinEnglish


NoBack-ground

NoBase

NoSketch

AdeninebasesarepartsofDNA.

(partTypes DNAMolecule Adenine-Base) P P P

NO

CytosinebasesarepartsofDNA.

(partTypes DNAMolecule Cytosine-Base) P P P

NO

GuaninebasesarepartsofDNA.

(partTypes DNAMolecule Guanine-Base) P P P

NO

ThyminebasesarepartsofDNA.

(partTypes DNAMolecule Thymine-Base) P P P

NO

DNAmoleculeshaveasugarphosphatebackbone.

(partTypes DNAMolecule SugarPhosphateBackbone)

P P P NO

ThyminebondswithAdenine.

(relationAllExists hydrogenBondBetween Thymine-Base Adenine-Base)

P P P NO

CytosinebondswithGuanine.

(relationAllExists hydrogenBondBetween Cytosine-Base Guanine-Base)

P P P NO

ThymineonlybondswithAdenine.

(relationAllExistsAndOnly hydrogenBondBetween Thymine-Base Adenine-Base)

P P P NO

Cytosineonlybondswith

(relationAllExistsAndOnly hydrogenBondBetween P P P

NO

139Guanine. Cytosine-Base

Guanine-Base)

DNAmoleculesareverylong.

(relationAllInstance lengthOfObject DNAMolecule Long) P P P

NO

ThymineconnectswithAdenine.

(relationAllExists connectedTo-Directly Thymine-Base Adenine-Base)

NO NO NO NO

CytosineconnectswithGuanine.

(relationAllExists connectedTo-Directly Cytosine-Base Guanine-Base)

NO NO NO NO

140Ecosystem/Web

Text:Anecosystemislikeawebstructure.Achangeintheecosystemislikemovementinthestructure.Anymovementinthestructuredecreasesitsstability.Thestabilityofthestructureislikethestabilityoftheecosystem.

Sketchobjects:((isa connecting-string ConstructionArtifact) (isa student-2 Student) (isa student-2 TemporalThing) (isa string-moving MovementEvent) (isa student-1 Student) (isa energy-conversion IntrinsicStateChangeEvent) (isa chem-energy ChemicalEnergy) (isa the-ecosystem Ecosystem) (isa leaf Individual) (isa leaf Autotroph) (isa leaf BiologicalLivingObject) (isa leaf Plant) (isa web-stability Stability) (isa person Animal) (isa person Predator) (isa person BiologicalLivingObject) (isa person Heterotroph) (isa person Human) (isa the-string String-Textile) (isa the-string Physob) (isa wolf Heterotroph) (isa wolf Predator) (isa wolf BiologicalLivingObject) (isa wolf Animal) (isa mushroom Fungus) (isa mushroom (CollectionUnionFn (TheSet Fungus Bacterium))) (isa solar SolarEnergy) (isa sheep Animal) (isa sheep BiologicalLivingObject) (isa sheep Prey) (isa sheep Sheep-Domestic) (isa sheep Heterotroph))




NoBack-ground

NoBase

NoSketch

Ecosystemshaveproducers.

(relationAllExists possessiveRelation Ecosystem Autotroph)

P P NO NO

Ecosystemshaveconsumers.

(relationAllExists possessiveRelation Ecosystem Heterotroph)

P P P NO

Ecosystemshavepredators.

(relationAllExists possessiveRelation Ecosystem Predator)

P P P NO

Ecosystemshaveprey.

(relationAllExists possessiveRelation Ecosystem Prey)

P P P NO

Fungiaredecomposers.

(relationExistsAll decomposerInEcosystem Ecosystem Fungus)

P P P NO

Autotrophsareproducers.

(relationExistsAll producerInEcosystem Ecosystem Autotroph)

P P NO NO

Heterotrophsareconsumers.

(relationExistsAll consumerInEcosystem Ecosystem Heterotroph)

P P P NO

Predatorseatprey.

(relationAllExists eatsWillingly Predator Prey)

P P P NO

Producersconvertsolarenergytochemicalenergy.

(relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn IntrinsicStateChangeEvent toState ChemicalEnergy) objectOfStateChange SolarEnergy) Autotroph)

P P NO NO

Changesintheecosystemdecreasestability.

(qprop- ((QPQuantityFn Stability) Ecosystem) (RateFn (IntrinsicStateChangeOfFn Ecosystem)))

P NO NO NO

142Enzyme/Key

Text:Anenzymeislikeakey.Asubstratemoleculeislikealock.Thekeyhasridges.Theridgeshaveauniqueshape.Theenzymebindingsiteisliketheridges.Theenzymebindingsitehasauniquechemicalmakeup.Thekeyonlyunlocksspecificlocks.Theenzymereactsonlywithspecificsubstratemolecules.Thekeyunlocksthelock.Theenzymebreaksapartthesubstratemolecules.Afterunlockingalock,thekeyisunchanged.Thekeycanbeusedoverandoveragain.Enzymeactionislikeunlockingalock.Theenzymeisunchangedbythechemicalreaction.Theenzymecanbeusedoverandoveragain.

Sketchobjects:((isa the-enzyme EnzymeMolecule) (isa the-substrate SubstrateMolecule) (isa active-site EnzymeBindingSite)) ((isa the-ridge Ridge) (isa the-lock Lock) (isa the-key KeyOfLock)) ((isa Object-129 Event) (isa Object-129 ChemicalReaction) (isa Object-129 EnzymeActivationEvent) (isa Object-135 Event) (isa Object-135 UnlockingALock)) (isa lock-s1 Event) (isa lock-s2 Event) (isa lock-s2 Situation) (isa enzymeaction-s3 Situation) (isa enzymeaction-s3 Event) (isa enzymeaction-s2 Situation) (isa lock-s3 Event) (isa lock-s3 Situation) (isa enzymeaction-s2 Event) (isa lock-s3 TemporalThing) (isa enzymeaction-s2 TemporalThing) (isa enzymeaction-s1 Event) (isa enzymeaction-s1 TemporalThing) (isa enzymeaction-s1 Situation) (isa enzymeaction-s3 BreakingEvent) (isa lock-s1 TemporalThing) (isa lock-s1 Situation)

143(isa lock-s2 TemporalThing) (isa enzymeaction-s3 TemporalThing) GoldStandardQueriesinEnglish,CycL:FourrightmostcolumnsshowwhichqueriessucceededinthefourexperimentalconditionsinChapter4.



NoBack-ground

NoBase

NoSketch

Enzymemoleculeshaveuniquebindingsites.

(partTypes EnzymeMolecule (CollectionIntersectionFn (TheSet EnzymeBindingSite (ThingDescribableAsFn Unique-TheWord Adjective))))

P P P NO

Enzymesonlyreactwithspecificsubstrates.

(relationAllExistsAndOnly chemicalReactants (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn ChemicalReaction chemicalReactants SubstrateMolecule) catalyst EnzymeMolecule) (CollectionIntersectionFn (TheSet SubstrateMolecule (ThingDescribableAsFn Specific-TheWord Adjective))))

NO NO NO NO

Enzymeactionbreaksapartsubstratemolecules.

(relationAllExists subEvents EnzymeActivationEvent (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn BreakingEvent doneBy EnzymeMolecule) objectOfStateChange SubstrateMolecule))

P P P NO

Theenzymecomesoutofthereactionunchanged.

(relationExistsAll unchangedActors EnzymeActivationEvent EnzymeMolecule)

P P P NO

Whenenzymesbindtosubstratemolecules,itenablesthesubstratetobreakapart.

(enables-SitTypeSitType (SubcollectionOfWithRelationToFn (CollectionIntersectionFn (TheSet (SubcollectionOfWithRelationToTypeFn Situation subEvents ChemicalReaction) (SubcollectionOfWithRelationToTypeFn Situation subEvents

P P P NO

144

EnzymeActivationEvent) Situation)) holdsIn (relationExistsExists objectsIntersect EnzymeMolecule SubstrateMolecule)) (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn BreakingEvent doneBy EnzymeMolecule) objectOfStateChange SubstrateMolecule))

Enzymescanbereused.

(relationAll repeatedEvent EnzymeActivationEvent)

P NO NO NO

145Homeostasis/Escalator

Text:Homeostasisislikegoingupadownescalator.Theescalatorismovingdown.Thepersoniswalkingup.Heatisconstantlyescapingfromthehumanbody.Heatescapingisliketheescalatormovingdown.Thehumanbodymakestheheatbyexercising.Makingtheheatislikewalkingup.Temperatureisliketheheightoftheperson.Theheatlossandtheheatproductionarebalanced,sothetemperatureremainsconstant.Duringhomeostasis,theidealtemperatureofthehumanbodyis37degreesCelsius.

Sketchobjects:((isa person-motion Walking-Generic) (isa the-escalator Physob) (isa the-escalator Escalator) (isa escalator-motion MovementEvent) (isa temp-scale Temperature) (isa temp-scale ScalarOrVectorInterval) (isa bob Physob) (isa bob Person) (isa esc-dir DirectionOfMovement) (isa bob-dir DirectionOfMovement) (isa going-up-a-down-esc GoingUpADownEscalator) (isa temp-value TemperatureIndicator))




NoBack-ground

NoBase

NoSketch

Heatescapesfromthebody.

(relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn LeavingAPlace fromLocation HumanBody) ThermalEnergy)

NO NO NO NO

Thebodyproducesheat.

(relationAllExists products (SubcollectionOfWithRelationToTypeFn MakingSomething performedBy HumanBody) ThermalEnergy)

P P P P

Heatlossandheatproductionarebalancedsotemperatureremainsconstant.

(implies (qEqualTo (RateFn (SubcollectionOfWithRelationToTypeFn (MakingFn ThermalEnergy) beneficiary HumanBody)) (RateFn (SubcollectionOfWithRelationToTypeFn LeavingAPlace doneBy ThermalEnergy))) (qEqualTo (QPDerivativeFn (CollectionIntersectionFn (TheSet Temperature (ThingDescribableAsFn Ideal-TheWord Adjective)))) Zero))

P NO NO NO

Bodytemperature(positively)dependsonheatproduction.

(qprop ((QPQuantityFn (CollectionIntersectionFn (TheSet Temperature (ThingDescribableAsFn Ideal-TheWord Adjective)))) HumanBody) (RateFn (SubcollectionOfWithRelationToTypeFn (MakingFn ThermalEnergy) beneficiary HumanBody)))

P NO NO NO

Bodytemperature(negatively)dependsonheatloss.

(qprop- ((QPQuantityFn (CollectionIntersectionFn (TheSet Temperature (ThingDescribableAsFn Ideal-TheWord Adjective)))) HumanBody) (RateFn

P NO NO NO

147

(SubcollectionOfWithRelationToTypeFn LeavingAPlace doneBy ThermalEnergy)))

Idealbodytemperatureis37°C.

(relationAllInstance measure (CollectionIntersectionFn (TheSet Temperature (ThingDescribableAsFn Ideal-TheWord Adjective))) (DegreeCelsius 37))

P P P P

148Membrane/Mosaic

Text:Amembraneislikeamosaicthathasliquidgrout.Themembraneproteinislikethemosaictiles.Thelipidbilayerislikethegrout.Thetilesmoveinthegrout.Themembraneproteinmovesinthelipidbilayer.Themembraneproteinmovessomechemicalsubstancesacrossthemembrane.Themembraneproteinblockssomesubstances.Sometilesaremadeofdifferentmaterials.Somemembraneproteinshavedifferentfunctions.

Sketchobjects:((isa moving-thru-mosaic CausingAnotherObjectsTranslationalMotion) (isa moving-thru-mosaic MovementEvent) (isa chemical-substance ChemicalObject) (isa the-fluid-mosaic Mosaic) (isa mosaic-tile Tile) (isa liquid-grout Grout) (isa moving-thru-membrane MovementEvent) (isa moving-thru-membrane CausingAnotherObjectsTranslationalMotion) (isa membrane-protein MembraneProtein) (isa lipid-bilayer LipidBilayer) (isa the-bio-membrane BiologicalMembrane))




NoBack-ground

NoBase

NoSketch

Membraneproteinsmovethingsacrossthemembrane.

(relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn CausingAnotherObjectsTranslationalMotion objectActedOn (SetOfTypeFn PartiallyTangible)) across-UnderspecifiedRegion BiologicalMembrane) MembraneProtein)

P P NO P

Membranescanblocksubstances.

(relationExistsAll barrier (SubcollectionOfWithRelationToTypeFn BarrierSituation blockedPath (SetOfTypeFn PartiallyTangible)) MembraneProtein)

P P P P

Differentmembraneproteinshavedifferentfunctions.

(relationAllExists possessiveRelation (SetOfTypeFn MembraneProtein) (CollectionIntersectionFn (TheSet Capability (ThingDescribableAsFn Differ-TheWord Adjective-Gradable))))

P P P NO

Membraneproteinsmovewithinthelipidbilayer.

(and (relationExistsAll primaryObjectMoving MovementEvent MembraneProtein (relationAllExists in-UnderspecifiedContainer MembraneProtein LipidBilayer))

P P P NO

Membraneshavemembraneproteins.

(partTypes BiologicalMembrane MembraneProtein) P P

NO NO

Membraneshavelipidbilayers.

(partTypes BiologicalMembrane LipidBilayer))) P P P

NO

150Taxonomy/Supermarket

Text:Abiologicalclassificationsystemislikeasupermarket.Supermarketshavemanysections.Customersusethesectionstofindtheproducts.Largesectionshavemanydifferentproducts.Smallsectionshavefewerproducts.Findingproductsdependsonattributesofproducts.Someproductsfitintomanydifferentsections.Supermarketsusedifferentsections.Butallsupermarketshavesections.Biologicalclassificationsystemshavemanygroups.Biologistsusegroupstostudyspecies.Largegroupshavehighdiversity.Smallgroupshavelowdiversity.Groupingspeciesdependsonattributesofspecies.Somespeciesaredifficulttogroup.Differentbiologicalclassificationsystemshavedifferentgroups.Butallbiologicalclassificationsystemshavegroups.

Sketchobjects:((isa HomoGenus BiologicalGenus) (isa HomoGenus Group) (isa Object-76 BiologicalFamily) (isa Object-76 Group) (isa Object-62 Group) (isa Object-62 BiologicalSpecies) (isa RefrigeratedMarketCategory ExistingObjectType) (isa RefrigeratedMarketCategory RefrigeratedMarketCategory) (isa RefrigeratedMarketCategory ProductTypeByMarketCategory) (isa Object-218 GroceryMarketCategory) (isa Object-134 BlueWhale) (isa Object-134 Organism-Whole) (isa Object-90 BiologicalOrder) (isa Object-90 Group) (isa Object-241 Product) (isa Object-241 GoatCheese) (isa cheese-section CheeseMarketCategory) (isa DairyMarketCategory ExistingObjectType) (isa DairyMarketCategory DairyMarketCategory) (isa DairyMarketCategory ProductTypeByMarketCategory) (isa Object-232 NonperishableMarketCategory) (isa Object-102 BiologicalGenus) (isa Object-102 Group) (isa Object-99 BiologicalGenus) (isa Object-99 Group) (isa Object-92 Group)

151 (isa Object-92 BiologicalFamily) (isa Object-124 BiologicalGenus) (isa Object-124 Group) (isa Object-113 Group) (isa Object-113 BiologicalGenus) (isa HominidaeFamily Group) (isa HominidaeFamily BiologicalFamily) (isa HominidaeFamily QAClarifyingCollectionType) (isa homo-henry Organism-Whole) (isa homo-henry HomoSapiens) (isa Object-239 NutsMarketCategory) (isa Object-111 BiologicalGenus) (isa Object-111 Group) (isa HomoSapiens OrganismClassificationType) (isa HomoSapiens Group) (isa HomoSapiens BiologicalSpecies) (isa Object-52 BiologicalClass) (isa Object-52 Group) (isa Object-114 Group) (isa Object-114 BiologicalGenus) (isa Object-119 BiologicalGenus) (isa Object-119 Group) (isa rhesus-ramona Monkey) (isa rhesus-ramona Organism-Whole) (isa MacacaMulattaSpecies BiologicalSpecies) (isa MacacaMulattaSpecies Group) (isa Object-72 Group) (isa Object-72 BiologicalGenus) (isa Object-56 BiologicalSpecies) (isa Object-56 Group) (isa Object-126 BiologicalGenus) (isa Object-126 Group) (isa Object-125 Group) (isa Object-125 BiologicalGenus) (isa Object-74 Group) (isa Object-74 BiologicalGenus) (isa Object-118 BiologicalGenus) (isa Object-118 Group) (isa Object-58 BiologicalSpecies) (isa Object-58 Group) (isa yogurt-section YogurtCategory) (isa Primate KEClarifyingCollectionType) (isa Primate OrganismClassificationType) (isa Primate Group) (isa Primate BiologicalOrder) (isa Object-94 BiologicalFamily) (isa Object-94 Group) (isa Object-60 BiologicalSpecies) (isa Object-60 Group) (isa Object-101 BiologicalGenus) (isa Object-101 Group) (isa Object-112 Group) (isa Object-112 BiologicalGenus) (isa Object-234 FreshFruitMarketCategory) (isa Object-109 BiologicalGenus) (isa Object-109 Group) (isa homo-genus-2 Group) (isa homo-genus-2 BiologicalGenus) (isa Object-68 Group) (isa Object-68 BiologicalSpecies) (isa Object-70 BiologicalSpecies) (isa Object-70 Group) (isa Object-64 Group) (isa Object-64 BiologicalSpecies) (isa Object-66 Group) (isa Object-66 BiologicalSpecies) (isa homo-species-2 BiologicalSpecies) (isa homo-species-2 Group) (isa Object-100 BiologicalGenus) (isa Object-196 FreshVegetablesMarketCategory) (isa Object-110 BiologicalGenus) (isa Object-110 Group) (isa Object-120 BiologicalGenus) (isa Object-120 Group) (isa milk-section MilkMarketCategory) (isa cream-category CreamMarketCategory) (isa the-taxonomy Taxonomy) (isa the-store GroceryStore))




NoBack-ground

NoBase

NoSketch

Taxonomiesareclassificationsystems.

(genls Taxonomy ClassificationSystem) P P P NO

Largergroupshavegreaterdiversity.

(relationAllInstance possessiveRelation (SubcollectionOfWithRelationToFn Group sizeParameterOfObject (RelativeGenericValueFn sizeParameterOfObject Group highAmountOf)) (HighAmountFn Diversity))

NO NO NO NO

Smallergroupshavelowerdiversity.

(relationAllInstance possessiveRelation (SubcollectionOfWithRelationToFn Group sizeParameterOfObject (RelativeGenericValueFn sizeParameterOfObject Group veryLowToLowAmountOf)) (LowAmountFn Diversity))

NO NO NO NO

Groupinganorganismdependsonitsproperties.

(relationAllExists dependsOn-Underspecified (GroupingOfFn Organism-Whole) (SubcollectionOfWithRelationFromTypeFn AttributeCharacteristicOfAnEntity possessiveRelation BiologicalSpecies))

NO NO NO NO

Somespeciesaredifficulttoclassify.

(relationExistsInstance degreeOfDifficulty (GroupingFn BiologicalSpecies) Difficult)

NO NO NO NO

Taxonomieshavemanydifferentgroups.

(relationAllExistsRange possessiveRelation Taxonomy (SetOfTypeFn Group) Many-Quant)

NO NO P NO

Orderismorespecificthanclass.

(relationAllExists possessiveRelation BiologicalClass BiologicalOrder)

P NO P

NO

Familyismorespecificthanorder.

(relationAllExists possessiveRelation BiologicalOrder BiologicalFamily)

P NO P

NO

153Genusismorespecificthanfamily.

(relationAllExists possessiveRelation BiologicalFamily BiologicalGenus)

P NO P

NO

Speciesismorespecificthangenus.

(relationAllExists possessiveRelation BiologicalGenus BiologicalSpecies)

NO NO NO NO

Anorganismcanbelongtomanydifferentgroups.

(relationAllExistsRange member Organism-Whole Group Many-Quant)

NO NO NO NO

154Voltage/Pressure

Text:Abatteryislikeaholeinabottle.Thewaterflowsthroughthehole.Thegreaterthedepth,thegreaterthewaterflowrate.Electricityflowsthroughthebattery.Thebatteryhasvoltage.Voltageislikedepth.Thegreaterthevoltage,thegreatertheflowofelectricity.

Sketchobjects:((isa Object-46 DirectionOfMovement) (isa hole-in-bottle PolyDimensionalThing) (isa hole-in-bottle FluidConduit) (isa the-bottle Container) (isa the-bottle Bottle) (isa the-bottle Physob) (isa depth-of-hole Depth) (isa Object-44 DirectionOfMovement) (isa the-water Liquid-StateOfMatter) (isa Object-43 DirectionOfMovement)) GoldStandardQueriesinEnglish,CycL:


QueryinEnglish


NoBack-ground

NoBase

NoSketch

Theflowofelectricity(positively)dependsonthebatteryvoltage.

(qprop (RateFn (FlowFn Electricity)) ((QPQuantityFn Voltage) Battery)) P

NO NO NO

155APPENDIXB:ScienceQuestions

NYSRE-2014-8SCI-19 1NYSRE-2014-LE-01 2NYSRE-2014-LE-04 3NYSRE-2014-LE-21 4NYSRE-2014-LE-28 5MCAS-2014-BIO-03 6MCAS-2014-BIO-07 7MCAS-2014-BIO-09 8MCAS-2014-BIO-10 9MCAS-2014-BIO-17 10MCAS-2014-BIO-35 11MCAS-2014-BIO-39 12MCAS-K8-EE-14 13MCAS-K8-EE-20 14

1. Asinglehumanbodycelltypicallycontainsthousandsofa. genes b. nuclei c. chloroplasts d. bacteria

2. Afunctionofcellmembranesishumansisthea. synthesisoftheaminoacidsb. productionofenergyc. replicationofgeneticmateriald. recognitionofcertainchemicals

3. Whichstatementbestdescribestheorganellesinacell?a. Allorganellesareinvolveddirectlywithcommunicationbetweencells. b. Organellesmustworktogetherandtheiractivitiesmustbecoordinated c. Organellesfunctiononlywhenthereisadisruptioninhomeostasisd. Eachorganellemustfunctionindependentlyoftheothersinordertomaintain

homeostasis4. Humanpopulationgrowthhasledtoareductioninthepopulationsofpredatorsthroughout

naturalecosystemsacrosstheUnitedStates.Scientistsconsiderthelossofthesepredatorstohavea

a. positiveeffect,becauseanincreaseintheirpreyhelpstomaintainstabilityintheecosystem

b. positiveeffect,becausethepredatorsusuallyeliminatethespeciestheypreyon

156

c. negativeeffect,becausepredatorshavealwaysmadeupalargeportionofourfoodsupply

d. negativeeffect,becausepredatorshaveanimportantroleinmaintainingstableecosystems

5. Avarietyofpeartree,knownasBradford,wasoriginallyintroducedintotheeasternUnitedStatesinthe1960s.Today,thistreeiscrowdingoutotherplantsinthesestates.Thissituationbestillustrates

a. anunintentionalnegativeeffectofalteringanecosystemb. howaforeignspeciesiscontrolledintheeasternUnitedStatesc. thattheintroductionofaforeignspeciesdoesnotaffectfoodwebsd. thatseriousenvironmentalconsequencescanbeavoidedbyimportingaforeignspecies

6. Iftheproducersinafoodwebwereremoved,whichofthefollowingchangeswouldmostlikelyoccur?

a. Theentirefoodwebwouldcollapseovertime.b. Thefoodwebwoulddependonthedecomposersforenergy.c. Theconsumerswouldbeginmakingenergyforthefoodweb.d. Thepopulationsoftheremainingorganismsinthefoodwebwouldincrease.

7. Withinapreypopulation,whichofthefollowingismostimmediatelyaffectedbythearrivalofanewpredator?

a. deathrateb. evolutionratec. immigrationrated. maturationrate

8. Whichofthefollowingisafunctionofthenucleusinorganism2? a. absorbingsunlightb. releasingusableenergyc. storinggeneticmateriald. producingfoodmolecules

9. Inthethreeorganisms,whataresynthesizedbytheribosomes?a. Carbohydratesb. Lipidsc. nucleicacidsd. proteins

10. Whichofthefollowingbestdescribestheproducersinaterrestrialfoodweb? a. Theyareatthehighesttrophiclevel.b. Theyarenotaffectedbydecomposers. c. Theyconvertsolarenergytochemicalenergy.d. Theyobtainalltheirnutrientsandenergyfromconsumers.

11. Whichofthefollowingisthebestexampleofthehumanbodymaintaininghomeostasis?a. Theheartbeatsusingcardiacmuscle

157

b. Thebreathingrateincreasesduringexercise.c. Thenoseandearcontaincartilageforflexibility.d. Thedigestivesystemusesenzymestobreakdownfood.

12. Inasampleofdouble-strandedDNA,30%ofthenitrogenousbasesarethymine.Whatpercentageofthenitrogenousbasesinthesampleareadenine?

a. 20%b. 30%c. 60%d. 70%

13. Jamalwantstomakeanelectricalcircuit,butheonlyhastheobjectsshownbelow.WhichofthefollowingmustJamalalsohavetomakeanelectricalcircuit?

a. amotorb. aswitchc. abarmagnetd. apowersource

14. Thediagrambelowshowsaprojectthatastudentmadetotestanelectricalcircuit.Partoftheelectricalcircuitisunderneaththeboard.Whenthestudentconnectthetwonailsusingawire,thebulblightsup.Whichofthefollowingmustbeunderneaththeboard?

a. amagnetandaswitchb. aswitchandsomewiresc. amagnetandapowersourced. apowersourceandsomewires

;; NYSRE-2014-8SCI-19 ;; A single human body cell typically contains thousands of (queryForQuestion NYSRE-2014-8SCI-19 (relationAllExistsRange physicalParts Cell ?cell-part Thousands-Quant)) (multipleChoiceSingleOptionList NYSRE-2014-8SCI-19 (TheList Gene-HereditaryUnit 1)) (multipleChoiceSingleOptionList NYSRE-2014-8SCI-19 (TheList CellNucleus 2)) (multipleChoiceSingleOptionList NYSRE-2014-8SCI-19 (TheList Chloroplast 3)) (multipleChoiceSingleOptionList NYSRE-2014-8SCI-19 (TheList Bacterium 4)) (correctAnswerChoice NYSRE-2014-8SCI-19 1) ;; NYSRE-2014-LE-01 ;; A function of cell membranes is humans is the

158(queryForQuestion NYSRE-2014-LE-01 (relationExistsAll doneBy ?functional-event CellMembrane)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (MakingFn AminoAcid) 1)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (MakingFn EnergyStuff) 2)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList Replication-DNA 3)) (multipleChoiceSingleOptionList NYSRE-2014-LE-01 (TheList (IdentifyingAsTypeFn ChemicalObject) 4)) (correctAnswerChoice NYSRE-2014-LE-01 4) ;; NYSRE-2014-LE-04 ;; Which statement best describes the organelles in a cell? (queryForQuestion NYSRE-2014-LE-04 (trueQuestionOption NYSRE-2014-LE-04 ?n)) (multipleChoiceSingleOptionList NYSRE-2014-LE-04 (TheList (relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn InformationTransferPhysicalEvent informationDestination Cell) Organelle) 1)) (multipleChoiceSingleOptionList NYSRE-2014-LE-04 (TheList (relationExistsAll members FunctionalSystem Organelle) 2)) (multipleChoiceSingleOptionList NYSRE-2014-LE-04 (TheList (responsibleFor-TypeType (PreventingFn Homeostasis) (ActorPlaysRoleFn Organelle doneBy)) 3)) (multipleChoiceSingleOptionList NYSRE-2014-LE-04 (TheList (responsibleFor-TypeType (ActorPlaysRoleFn Organelle doneBy) Homeostasis) 4)) (correctAnswerChoice NYSRE-2014-LE-04 2) (in-microtheory (ATQuestionMtFn NYSRE-2014-LE-21) :exclude-globals t) (isa (ATQuestionMtFn NYSRE-2014-LE-21) (InformationThingAboutFn NYSRE-2014-LE-21)) (isa this-ecosystem Ecosystem) (consumerInEcosystem this-ecosystem these-humans) (consumerInEcosystem this-ecosystem other-predators)

159(eatsWillingly other-predators other-prey) (isa human-pop-growth IncreaseEvent) (isa human-pop-growth IntrinsicStateChangeEvent) (objectOfStateChange human-pop-growth this-ecosystem) (objectOfStateChange human-pop-growth other-predators) (isa predator-decline DecreaseEvent) (isa predator-decline IntrinsicStateChangeEvent) (objectOfStateChange predator-decline this-ecosystem) (causes-EventEvent human-pop-growth predator-decline) ;; Question info (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion NYSRE-2014-LE-21 (trueQuestionOption NYSRE-2014-LE-21 ?n)) (multipleChoiceSingleOptionList NYSRE-2014-LE-21 (TheList (and (causes-Event predator-decline ?prey-increase) (isa ?prey-increase IncreaseEvent) ;; something about increasing prey population (positivelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn ?prey-increase))) 1)) (multipleChoicesingleOptionList NYSRE-2014-LE-21 (TheList (and (causes-Event predator-decline ?ext) (isa ?ext Extinction) (positivelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn ?ext))) 2)) (multipleChoicesingleOptionList NYSRE-2014-LE-21 (TheList (and (eatsWillingly these-humans ?food-supply) (negativelyInfluencedBy (AmountOfFn ?food-supply) (RateFn predator-decline))) 3)) (multipleChoicesingleOptionList NYSRE-2014-LE-21 (TheList (negativelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn predator-decline)) 4))

160(correctAnswerChoice NYSRE-2014-LE-21 4) ;; ;; NYSRE-2014-LE-28 (in-microtheory (ATQuestionMtFn NYSRE-2014-LE-28)) (isa (ATQuestionMtFn NYSRE-2014-LE-28) (InformationThingAboutFn NYSRE-2014-LE-28)) (isa BradfordTree OrganismClassificationType) (genls BradfordTree Tree-ThePlant) (isa this-ecosystem Ecosystem) (producerInEcosystem this-ecosystem these-bradford-trees) (producerInEcosystem this-ecosystem other-plants) (isa introduction-of-btree IncreaseEvent) (isa introduction-of-btree IntrinsicStateChangeEvent) (objectOfStateChange introduction-of-btree this-ecosystem) (objectOfStateChange introduction-of-btree other-plants) ;; Question info (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion NYSRE-2014-LE-28 (trueQuestionOption NYSRE-2014-LE-28 ?n)) (multipleChoiceSingleOptionList NYSRE-2014-LE-28 (TheList (negativelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn introduction-of-btree)) 1)) (multipleChoicesingleOptionList NYSRE-2014-LE-28 (TheList (objectControlled ?controlling-btrees these-bradford-trees) 2)) (multipleChoicesingleOptionList NYSRE-2014-LE-28 (TheList (falseSentence (influencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn introduction-of-btree))) 3)) (multipleChoicesingleOptionList NYSRE-2014-LE-28 (TheList (preventedProp introduction-of-btree (negativelyInfluencedBy ((QPQuantityFn Stability) this-ecosystem) (RateFn introduction-of-btree))) 4)) (correctAnswerChoice NYSRE-2014-LE-28 1)

161 ;; ;; MCAS-2014-BIO-03 (in-microtheory (ATQuestionMtFn MCAS-2014-BIO-03)) (isa (ATQuestionMtFn MCAS-2014-BIO-03) (InformationThingAboutFn MCAS-2014-BIO-03)) (isa this-ecosystem Ecosystem) (producerInEcosystem this-ecosystem these-producers) (isa these-producers Autotroph) ;; question info (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion MCAS-2014-BIO-03 (trueQuestionOption MCAS-2014-BIO-03 ?n)) (multipleChoiceSingleOptionList MCAS-2014-BIO-03 (TheList (relationAllExists requires-Underspecified Ecosystem Autotroph) 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-03 (TheList (relationExistsExists producerInEcosystem Ecosystem (SubcollectionOfWithRelationFromTypeFn BiologicalLivingObject decomposerInEcosystem Ecosystem)) 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-03 (TheList (relationExistsExists producerInEcosystem Ecosystem Heterotroph) 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-03 (TheList (negativelyInfluencedBy ((QPQuantityFn Population) this-ecosystem) ((QPQuantityFn Population) these-producers)) 4)) (correctAnswerChoice MCAS-2014-BIO-03 1)

162 ;; ;; MCAS-2014-BIO-07 (in-microtheory (ATQuestionMtFn MCAS-2014-BIO-07)) (isa (ATQuestionMtFn MCAS-2014-BIO-07) (InformationThingAboutFn MCAS-2014-BIO-07)) (isa this-ecosystem Ecosystem) (isa predator (SetOfTypeFn BiologicalLivingObject)) (predatorInEcosystem this-ecosystem predator) (isa prey (SetOfTypeFn BiologicalLivingObject)) (preyInEcosystem this-ecosystem prey) (eatsWillingly predator prey) ;; question info: (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion MCAS-2014-BIO-07 (influencedBy ?x ((QPQuantityFn Population) predator))) (multipleChoiceSingleOptionList MCAS-2014-BIO-07 (TheList (RateFn (DeathFn prey)) 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-07 (TheList (RateFn (SubcollectionOfWithRelationToFn Evolution doneBy prey)) 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-07 (TheList (RateFn (SubcollectionOfWithRelationToFn Immigration doneBy prey)) 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-07 (TheList (RateFn (SubcollectionOfWithRelationToFn Maturing doneBy prey)) 4)) (correctAnswerChoice MCAS-2014-BIO-07 1) ;; ;; MCAS-2014-BIO-09 ;; Which of the following is a function of the nucleus in organism 2? (visualAidForQuestion MCAS-2014-BIO-09 (VisualAidForQuestionMtFn MCAS-2014-BIO-08-11)) (queryForQuestion MCAS-2014-BIO-09 (relationExistsAll performedBy ?functional-event CellNucleus)) (multipleChoiceSingleOptionList MCAS-2014-BIO-09 (TheList (SubcollectionOfWithRelationToTypeFn AbsorptionEvent objectActedOn Sunlight) 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-09 (TheList (MakingAbstractAvailableFn EnergyStuff) 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-09 (TheList (StoringFn Gene-HereditaryUnit) 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-09 (TheList (MakingFn Food) 4)) (correctAnswerChoice MCAS-2014-BIO-09 3)

163 ;; ;; MCAS-2014-BIO-10 (queryForQuestion MCAS-2014-BIO-10 (relationExistsAll performedBy (MakingFn ?product) Ribosome)) (multipleChoiceSingleOptionList MCAS-2014-BIO-10 (TheList CarbohydrateStuff 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-10 (TheList Lipid 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-10 (TheList DNAMolecule 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-10 (TheList ProteinStuff 4)) (correctAnswerChoice MCAS-2014-BIO-10 4) ;; MCAS-2014-BIO-17 (queryForQuestion MCAS-2014-BIO-17 (trueQuestionOption MCAS-2014-BIO-17 ?n)) (multipleChoiceSingleOptionList MCAS-2014-BIO-17 (TheList (genls Autotroph Heterotroph) 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-17 (TheList (falseSentence (dependsOn-Underspecified Autotroph (SubcollectionOfWithRelationFromTypeFn BiologicalLivingObject decomposerInEcosystem Ecosystem))) 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-17 (TheList (relationExistsAll doneBy (SubcollectionOfWithRelationToTypeFn (SubcollectionOfWithRelationToTypeFn IntrinsicStateChangeEvent toState ChemicalEnergy) objectOfStateChange SolarEnergy) Autotroph) 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-17 (TheList (relationExistsAll from-UnderspecifiedLocation (SubcollectionOfwithRelationToTypeFn (MakingAvailableFn Food)

164 doneBy (SubcollectionOfWithRelationFromTypeFn BiologicalLivingObject consumerInEcosystem Ecosystem)) Autotroph) 4)) (correctAnswerChoice MCAS-2014-BIO-17 3) ;; MCAS-2014-BIO-35 (in-microtheory (ATQuestionMtFn MCAS-2014-BIO-35)) (isa (ATQuestionMtFn MCAS-2014-BIO-35) (InformationThingAboutFn MCAS-2014-BIO-35)) ;; situation 1 (isa heart-beating HeartBeating) (instrument-Generic heart-beating muscle) (isa muscle CardiacMuscle) ;; situation 2 (isa breathing Breathing) (isa exercise Exercising) (causes-EventEvent exercise breathing) (positivelyInfluencedBy (RateFn breathing) (RateFn exercise)) ;; situation 3 (isa nose Nose) (isa ear Ear) (isa nose-and-ear-makeup Configuration) (holdsIn nose-and-ear-makeup (possessiveRelation nose nose-cartilage)) (holdsIn nose-and-ear-makeup (possessiveRelation ear ear-cartilage)) (positivelyInfluencedBy ((QPQuantityFn Flexibility) ear) (AmountOfFn ear-cartilage)) (positivelyInfluencedBy ((QPQuantityFn Flexibility) nose) (AmountOfFn nose-cartilage)) ;; situation 4 (isa e EnzymeMolecule) (isa d DigestiveSystem) (isa breaking-down-food DigestionEvent) (doneBy breaking-down-food d) (instrument-Generic breaking-down-food e) ;; question info (in-microtheory AnalogyTutorEvaluationQuestionsMt)

165(queryForQuestion MCAS-2014-BIO-35 (isa ?x BiologicalHomeostasis)) (multipleChoiceSingleOptionList MCAS-2014-BIO-35 (TheList heart-beating 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-35 (TheList breathing 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-35 (TheList nose-and-ear-makeup 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-35 (TheList breaking-down-food 4)) (in-microtheory (ATQuestionMtFn MCAS-2014-BIO-39)) (isa (ATQuestionMtFn MCAS-2014-BIO-39) (InformationThingAboutFn MCAS-2014-BIO-39)) (isa sample-dna DNAMolecule) (isa thymine-bases (SetOfTypeFn Thymine-Base)) (isa adenine-bases (SetOfTypeFn Adenine-Base)) (isa guanine-bases (SetOfTypeFn Guanine-Base)) (isa cytosine-bases (SetOfTypeFn Cytosine-Base)) (physicalParts sample-dna thymine-bases) (physicalParts sample-dna adenine-bases) (physicalParts sample-dna guanine-bases) (physicalParts sample-dna cytosine-bases) (percentOfIndividual thymine-bases sample-dna (Percent 30)) (queryForQuestion MCAS-2014-BIO-39 (percentOfIndividual adenine-bases sample-dna (Percent ?x))) (multipleChoiceSingleOptionList MCAS-2014-BIO-39 (TheList 20 1)) (multipleChoiceSingleOptionList MCAS-2014-BIO-39 (TheList 30 2)) (multipleChoiceSingleOptionList MCAS-2014-BIO-39 (TheList 60 3)) (multipleChoiceSingleOptionList MCAS-2014-BIO-39 (TheList 70 4)) (correctAnswerChoice MCAS-2014-BIO-39 2) ;; ;; MCAS-K8-EE-14 ;; Jamal wants to make an electrical circuit, but he only has the objects shown below. ;; Which of the following must Jamal also have to make an electrical circuit? ;; Scenario info (in-microtheory (ATQuestionMtFn MCAS-K8-EE-14)) (isa (ATQuestionMtFn MCAS-K8-EE-14) (InformationThingAboutFn MCAS-K8-EE-14)) (isa circuit-materials (SetOfTypeFn ElectricalComponent)) (isa hypothetical-circuit SeriesCircuit) (isa jamals-wire Wire) (physicalParts hypothetical-circuit jamals-wire)

166(isa jamals-bulb LightBulbIncandescent) (physicalParts hypothetical-circuit jamals-bulb) (physicalParts hypothetical-circuit missing-piece) ;; End scenario info ;; Question info (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion MCAS-K8-EE-14 (isa missing-piece ?missing-type)) (multipleChoiceSingleOptionList MCAS-K8-EE-14 (TheList ElectricalMotor 1)) (multipleChoiceSingleOptionList MCAS-K8-EE-14 (TheList ElectricalSwitch 2)) (multipleChoiceSingleOptionList MCAS-K8-EE-14 (TheList Magnet 3)) (multipleChoiceSingleOptionList MCAS-K8-EE-14 (TheList Battery 4)) (correctAnswerChoice MCAS-K8-EE-14 4) ;; ;; MCAS-K8-EE-20 ;; scenario info (in-microtheory (ATQuestionMtFn MCAS-K8-EE-20)) (isa (ATQuestionMtFn MCAS-K8-EE-20) (InformationThingAboutFn MCAS-K8-EE-20)) (visualAidForQuestion MCAS-K8-EE-20 (VisualAidForQuestionMtFn MCAS-K8-EE-20)) ;; question info (in-microtheory AnalogyTutorEvaluationQuestionsMt) (queryForQuestion MCAS-K8-EE-20 (and (isa hidden-thing1 ?hidden-type1) (isa hidden-thing2 ?hidden-type2))) (multipleChoiceSingleOptionList MCAS-K8-EE-20 (TheList (TheSet Magnet ElectricalSwitch) 1)) (multipleChoiceSingleOptionList MCAS-K8-EE-20 (TheList (TheSet ElectricalSwitch (SetOfTypeFn Wire)) 2)) (multipleChoiceSingleOptionList MCAS-K8-EE-20 (TheList (TheSet Magnet Battery) 3)) (multipleChoiceSingleOptionList MCAS-K8-EE-20 (TheList (TheSet Battery (SetOfTypeFn Wire)) 4)) (correctAnswerChoice MCAS-K8-EE-20 4)

167

Visualaidforquestion:MCAS-K8-EE-20 4

SketchObjects:

(isa diagram-nail2 Nail) (isa diagram-board Board-PieceOfWood) (isa diagram-bulb LightBulbIncandescent) (isa student-circuit SeriesCircuit) (isa diagram-wire Wire) (isa student-circuit ElectricalCircuit) (isa diagram-nail1 Nail)

SketchObjectswithnolabels:

hidden-thing1 hidden-thing2

northwestern university capturing qualitative science ... · learning (burstein, 1985), and problem...

Documents