20
A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases Mala Mehrotra Pragati Synergetic Research Inc. Cupertino, CA [email protected] www.pragati-inc.com RKF Team Review Meeting SRI CA 11 th Oct 2000

A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

  • Upload
    elma

  • View
    32

  • Download
    0

Embed Size (px)

DESCRIPTION

A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases. Mala Mehrotra Pragati Synergetic Research Inc. Cupertino, CA [email protected] www.pragati-inc.com. RKF Team Review Meeting SRI CA 11 th Oct 2000. SRI Team’s Primary Focus - PowerPoint PPT Presentation

Citation preview

Page 1: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

A Contextual Clustering Approach for Theory Manipulation of RKF

Knowledge Bases

Mala MehrotraPragati Synergetic Research Inc.

Cupertino, CA

[email protected]

RKF Team Review Meeting

SRI CA

11th Oct 2000

Page 2: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 2

SRI Team’s Primary FocusSRI Team’s Primary FocusProvide components for formation of KBS

SRI Team’s Primary FocusSRI Team’s Primary FocusProvide components for formation of KBS

Multi-ViewPoint-Clustering Analysis Multi-ViewPoint-Clustering Analysis (MVP-CA) Technology Focus(MVP-CA) Technology Focus

Provide an analysis tool for aiding componentization of existing KBS

Multi-ViewPoint-Clustering Analysis Multi-ViewPoint-Clustering Analysis (MVP-CA) Technology Focus(MVP-CA) Technology Focus

Provide an analysis tool for aiding componentization of existing KBS

Page 3: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 3

Multi-ViewPoint-Clustering Analysis Multi-ViewPoint-Clustering Analysis (MVP-CA) Approach(MVP-CA) Approach

Multi-ViewPoint-Clustering Analysis Multi-ViewPoint-Clustering Analysis (MVP-CA) Approach(MVP-CA) Approach

• Agglomerative clustering algorithms produce semantically-related axiom clusters

• “Similarity” defined by a set of heuristic distance metrics • Meaningful clusters with the aid of statistical and semantics-based

cluster information • Clustering provides support for reverse engineering of KBs:

• anomaly checking• comprehension• building intermediate concept nodes and mid-level theories• …. by exposing semantic contexts for terms in the pre-existing axioms

Page 4: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 4

SRI Team’s Short-Term ObjectiveSRI Team’s Short-Term ObjectiveFormulate spatial representation components

SRI Team’s Short-Term ObjectiveSRI Team’s Short-Term ObjectiveFormulate spatial representation components

MVP-CA Technology’s Potential ContributionMVP-CA Technology’s Potential Contribution

Extract components from IKB dealing with spatial concepts

MVP-CA Technology’s Potential ContributionMVP-CA Technology’s Potential Contribution

Extract components from IKB dealing with spatial concepts

Page 5: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 5

Status of Work in ProgressStatus of Work in ProgressStatus of Work in ProgressStatus of Work in Progress

• IKB slice for spatial vocabulary obtained in mid-Sept from SRI.

• Slice was divided into two files:

• 288 assertion axioms

• 599 term-definition axioms

• Focus on:

• Exposing redundant overloaded concepts

• Identify reusable concepts

• Current work focuses on analyzing the assertion axioms:

Report on results so far ….

Page 6: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 6

First Stage: KB cleanupFirst Stage: KB cleanupFirst Stage: KB cleanupFirst Stage: KB cleanup

• Eliminated axioms with :ignore t • Identified at the parse stage • 65 such axioms eliminated

• Eliminated duplicate axioms

• Identified using the MVP-CA tool’s redundancy feature • 59 such axioms eliminated

• The cleaned up version has 164 assertions

Page 7: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 7

Second Stage: Cluster FormationSecond Stage: Cluster FormationSecond Stage: Cluster FormationSecond Stage: Cluster Formation

• MVP-CA tool’s clustering of assertions has produced axiom clusters

which reveal context of usage of a few salient terms • Some problem areas

• Some useful concepts

• Such exposition can help with intermediate concept node formation

for:• Better maintenance

• Reorganization, and

• Presentation of concept terms to SME/KE

• Work is still in progress

• Some plausible scenarios with these clusters will be presented next

Page 8: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 8

Potentially Redundant AxiomsPotentially Redundant Axioms Potentially Redundant AxiomsPotentially Redundant Axioms

(#$implies(#$and (#$touchesDirectly ?X ?Y)(#$objectFoundInLocation ?X ?LOC))(#$objectFoundInLocation ?Y ?LOC))

(#$implies(#$and (#$touches ?X ?Y)(#$objectFoundInLocation ?X ?LOC))(#$objectFoundInLocation ?Y ?LOC))

touchesDirectly and touches are essentially same concepts in the context of objectFoundInLocation.

Page 9: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 9

Concept of Concept of touchestouches and and touchesDirectlytouchesDirectlyConcept of Concept of touchestouches and and touchesDirectlytouchesDirectly

(#$implies (#$formsBorderBetween ?BORDER ?Y ?Z) (#$touchesDirectly ?BORDER ?Y))

(#$implies(#$and(#$isa ?INSIDEOUT #$InsideSurface)(#$isa ?OUTSIDEIN #$ExternalSurface-WholeThing)(#$physicalParts ?OUT ?INSIDEOUT)(#$externalParts ?IN ?OUTSIDEIN)(#$in-Snugly ?IN ?OUT))

(#$touches ?INSIDEOUT ?OUTSIDEIN))(#$implies(#$in-ImmersedGeneric ?OBJ ?FLUID)(#$touches ?FLUID ?OBJ))(#$implies (#$touches ?X ?Y) (#$near ?X ?Y))(#$implies(#$and(#$touchesDirectly ?X ?Y)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?Y ?LOC))(#$implies(#$and(#$touches ?X ?Y)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?Y ?LOC))(#$implies(#$bordersOn ?X ?Y)(#$touchesDirectly ?X ?Y))(#$implies(#$and(#$touches ?X ?Y)(#$physicalParts ?Z ?Y))

(#$touches ?X ?Z))

(#$implies (#$touchesDirectly ?X ?Y) (#$touches ?X ?Y))(#$implies (#$and (#$touchesDirectly ?PRT ?THING) (#$externalParts ?WHL?PRT))

(#$touchesDirectly ?THING ?WHL))(#$implies(#$in-Embedded ?X ?Y)

(#$touchesDirectly ?X ?Y))(#$implies(#$in-ContFullOf ?X ?Y)

(#$touchesDirectly ?X ?Y))(#$implies (#$touchesDirectly ?X ?Y)

(#$distanceBetween ?X ?Y(#$Foot-UnitOfMeasure 0)))(#$implies(#$in-Held ?OBJ ?HOLDER)(#$touches ?HOLDER ?OBJ))(#$implies (#$adjacentTo ?REG1 ?REG2) (#$touches ?REG1 ?REG2))(#$implies(#$on-Physical ?TOP ?BOT)

(#$touches ?BOT ?TOP))

Page 10: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 10

Pivot Concept: Pivot Concept: distanceBetweendistanceBetweenPivot Concept: Pivot Concept: distanceBetweendistanceBetween

(assertion(#$implies(#$bordersOn ?X ?Y)(#$distanceBetween ?X ?Y (#$Kilometer 0))))

(assertion(#$implies(#$bordersOn ?X ?Y)(#$distanceBetween ?X ?Y (#$Meter 0))))

(assertion(#$implies(#$bordersOn ?X ?Y)(#$touchesDirectly ?X ?Y)))

(assertion(#$implies (#$touchesDirectly ?X ?Y) (#$distanceBetween ?X ?Y

(#$Foot-UnitOfMeasure 0))))

Page 11: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 11

Intermediate Concept Node FormationIntermediate Concept Node Formation Intermediate Concept Node FormationIntermediate Concept Node Formation

bordersOn distanceBetween (KM)

bordersOn distanceBetween (M)

touchesDirectly distanceBetween (F)

bordersOn touchesDirectly

bordersOn distanceBetween (F)

bordersOn distanceBetween (distanceUnit)

F | M | KM distanceUnit

bordersOn distanceBetween (F)

Page 12: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 12

Pivot Concept: Pivot Concept: ObjectFoundInLocationObjectFoundInLocationPivot Concept: Pivot Concept: ObjectFoundInLocationObjectFoundInLocation

(#$implies (#$oFIL ?OBJ ?LOC)(#$near ?LOC ?OBJ))(#$implies(#$and (#$touchesDirectly ?X ?Y)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?Y ?LOC))(#$implies(#$and (#$touches ?X ?Y)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?Y ?LOC))(#$implies(#$and(#$on-Physical ?X ?Y)(#$objectFoundInLocation ?Y ?LOC))

(#$objectFoundInLocation ?X ?LOC))(#$implies(#$and (#$groupMembers ?C ?MEM)(#$objectFoundInLocation ?C ?LOC))

(#$objectFoundInLocation ?MEM ?LOC))(#$implies (#$and (#$physicalParts ?X ?PART)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?PART ?LOC))(#$implies (#$and (#$physicalParts ?LOC ?PART)(#$objectFoundInLocation ?X ?PART))

(#$objectFoundInLocation ?X ?LOC))(#$implies (#$and (#$in-ContGeneric ?OBJ ?CONT)(#$objectFoundInLocation ?CONT ?REG))

(#$objectFoundInLocation ?OBJ ?REG))(#$implies (#$in-ImmersedFully ?OBJ ?FLU)

(#$objectFoundInLocation ?OBJ ?FLU))(#$implies (#$and (#$isa ?FLUID #$Place)(#$in-ImmersedGeneric ?OBJECT ?FLUID))

(#$objectFoundInLocation ?OBJECT ?FLUID))(#$implies (#$and(#$objectFoundInLocation ?PER ?LOC)(#$covers-Hairlike ?STUFF ?LOC))

(#$in-Among ?PER ?STUFF))(#$implies(#$and (#$in-Floating ?OB ?LIQ)(#$surfaceParts ?LIQ ?SURF))

(#$objectFoundInLocation ?OB ?SURF))(#$implies (#$and (#$isa ?WATER #$BodyOfWater)(#$in-Floating ?OBJ ?WATER))

(#$objectFoundInLocation ?OBJ ?WATER))(#$implies (#$and (#$in-ContGeneric ?OBJ ?CONT)(#$containsCavity ?CONT ?CAV))

(#$objectFoundInLocation ?OBJ ?CAV))(#$implies(#$geographicalSubRegions ?REG ?PLACE)

(#$objectFoundInLocation ?PLACE ?REG))(#$implies (#$and (#$isa ?Y #$GeographicalRegion)(#$on-Physical ?X ?Y))

(#$objectFoundInLocation ?X ?Y))

Page 13: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 13

objectFoundInLocation: relationship to positional terms

objectFoundInLocation: relationship to positional terms

(#$implies (#$objectFoundInLocation ?OBJ ?LOC)(#$near ?LOC ?OBJ))

(#$implies (#$and (#$touchesDirectly ?X ?Y)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?Y ?LOC))(#$implies (#$and (#$touches ?X ?Y)

(#$objectFoundInLocation ?X ?LOC))(#$objectFoundInLocation ?Y ?LOC))

(#$implies (#$and (#$on-Physical ?X ?Y)(#$objectFoundInLocation ?Y ?LOC))

(#$objectFoundInLocation ?X ?LOC))

Page 14: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 14

objectFoundInLocation: relationship to partonomic terms

objectFoundInLocation: relationship to partonomic terms

(#$implies (#$and (#$physicalParts ?X ?PART)(#$objectFoundInLocation ?X ?LOC))

(#$objectFoundInLocation ?PART ?LOC))(#$implies (#$and (#$physicalParts ?LOC ?PART)

(#$objectFoundInLocation ?X ?PART))(#$objectFoundInLocation ?X ?LOC))

(#$implies (#$and (#$in-ContGeneric ?OBJ ?CONT)(#$objectFoundInLocation ?CONT ?

REG))(#$objectFoundInLocation ?OBJ ?REG))

(#$implies (#$and (#$in-ContGeneric ?OBJ ?CONT)(#$containsCavity ?CONT ?CAV))

(#$objectFoundInLocation ?OBJ ?CAV))(#$implies (#$and (#$in-Floating ?OB ?LIQ)

(#$surfaceParts ?LIQ ?SURF))(#$objectFoundInLocation ?OB ?SURF))

Page 15: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 15

objectFoundInLocation: relationship to group membership terms

objectFoundInLocation: relationship to group membership terms

(#$implies (#$and (#$objectFoundInLocation ?PER ?LOC) (#$covers-Hairlike ?STUFF ?LOC))(#$in-Among ?PER ?STUFF))

\(#$implies (#$and (#$groupMembers ?C ?MEM)

(#$objectFoundInLocation ?C ?LOC)) (#$objectFoundInLocation ?MEM ?LOC))

Page 16: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 16

objectFoundInLocation: relationship to geographical terms

objectFoundInLocation: relationship to geographical terms

(#$implies (#$geographicalSubRegions ?REG ?PLACE)(#$objectFoundInLocation ?PLACE ?REG))

(#$implies (#$and (#$isa ?Y #$GeographicalRegion) (#$on-Physical ?X ?Y))

(#$objectFoundInLocation ?X ?Y))

Page 17: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 17

objectFoundInLocation: relationship to fluid terms

objectFoundInLocation: relationship to fluid terms

(#$implies (#$in-ImmersedFully ?OBJ ?FLU)(#$objectFoundInLocation ?OBJ ?FLU))

(#$implies (#$and (#$isa ?FLUID #$Place) (#$in-ImmersedGeneric ?OBJECT ?FLUID))

(#$objectFoundInLocation ?OBJECT ?FLUID))

(#$implies (#$and (#$isa ?WATER #$BodyOfWater)(#$in-Floating ?OBJ ?WATER))

(#$objectFoundInLocation ?OBJ ?WATER))

(#$implies (#$and (#$in-Floating ?OB ?LIQ)(#$surfaceParts ?LIQ ?SURF))

(#$objectFoundInLocation ?OB ?SURF))

Page 18: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 18

Intermediate Concept Node IdentificationIntermediate Concept Node Identification Intermediate Concept Node IdentificationIntermediate Concept Node Identification

objectFoundInLocation

groupMembers

geographicalSubRegions GeographicalRegion

near inAmong covers-Hairlike

touches

touchesDirectly

onPhysical

in-ImmersedGeneric

BodyOfWater

in-ImmersedFully

in-Floating

in-ContGeneric

surfaceParts

containsCavity

physicalParts

Page 19: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 19

Intermediate Concept Node IdentificationIntermediate Concept Node Identification Intermediate Concept Node IdentificationIntermediate Concept Node Identification

objectFoundInLocation

groupMembers

geographicalSubRegions GeographicalRegion

near inAmong covers-Hairlike

touches

touchesDirectly

onPhysical

in-ImmersedGeneric

BodyOfWater

in-ImmersedFully

in-Floating

in-ContGeneric

surfaceParts

containsCavity

physicalParts

positional

geographic

fluidspartonomic

groupMembership

Page 20: A Contextual Clustering Approach for Theory Manipulation of RKF Knowledge Bases

October 11, 2000 20

Achievements & PlansAchievements & PlansAchievements & PlansAchievements & Plans

• Parser built for MELD axioms

• MVP-CA tool adapted for ontology representation in MELD

• Clustering results with Virus KB released• Clustering results with the IKB spatial ontology is ongoing

(assertions only):• Duplicate axioms identified• Clusters being studied for

• redundant axioms• intermediate concepts

• Next steps:• Cluster term definition file

• Provide support for concept graph

• Long term goal is to develop criteria for component identification using clusters