Building Knowledge Bases Compositionally
Bruce Porter, Peter Clark
Ken Barker, Art Souther, John Thompson
James Fan, Dan Tecuci, Peter Yeh
Marwan Elrakabawy, Sarah Tierney
Our Approach to RKF• Our goal: SME’s build knowledge bases by simply
instantiating and assembling pre-built components.• Our approach: We build a Component Library
containing representations of domain-specific concepts as well as common: actions, such as Get and Enter states, such as Be-Attached-To entities, such as Barrier and Catalyst property values, such as three microns and rapidAnd we develop computational methods for: combining them and using them to answer questions.
Generic Actions• About 200 actions, in about 20 clusters, based on
linguistic studies and other KB projects• Are these sufficient?
– Yes, based on an analysis of 6 chapters of the Alberts text and the encoding of much of chapter 7
– To test their coverage outside microbiology, we’ll be building dozens of KB’s this semester
– Our Component Evaluation will provide hard data
• Why keep it small?– So the Library will be easy to learn and use– So we can provide rich semantics for each action
Generic States
• A state, such as Be-Attached-To, represents a “temporarily stable” set of properties. It serves to link:– An action that creates the state (i.e. Attach)– An action that ends the state (i.e. Detach)– Those actions that are affected by the state (e.g.
Move)
Generic Entities
• small number of role concepts, defined by their participation in actions or states. Examples: container, sequence, nutrient, portal, portal covering
Generic Relations
• small number (78) of very general relations– Roles, such as agent, object, instrument, location– Properties, such as size, shape, frequency,
direction
• Why keep it small?– So the Library will be easy to learn and use– So we can provide rich semantics for each
relation
An Example:Bacterial RNA Transcription
• main participantsbacterial dna, rna polymerase, rna transcript
• scenario– polymerase makes contact with dna– polymerase moves along dna– polymerase recognizes promoter– polymerase transcribes gene, moving along DNA until it reaches terminator
– transcript detaches from polymerase– polymerase breaks contact with dna
Participants from Pump Priming
• bacterial dna, rna polymerase, rna transcript– in the domain-specific hierarchy
• example– Bacterial-DNA has
location: a Place regions: a Gene (abuts the Promoter region) (abuts the Terminator region) a Promoter a Terminator
etc.
Events in the Process from the “Component Library”
• example: Make-Contact– aka touch, adjoin, meet, contact
Make-Contactdestination
Entity
Entity
Place
object
object
Be-Touching
Move
Move
objectPlace
Place
source
source
destination
destination
object
locationlocation
Bacterial RNA Transcription
Bacterial-DNA RNA-Polymerase
Place
location
Bacterial-RNA-Transcription-Scenario
RNA-Transcript
GenePromoter Terminator
regions
causerobject result
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching
location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
object
destination
regions
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching
location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object
source
destination
path
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object
source
destination
path
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object causer
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object causer result
subevent
Move
object
dest
source
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object causer result
subevent
Move
object
dest
source
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object
location
object
object
Be-Attached-To
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
object
location
object
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
location
object
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
object
Be-Touching location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
location
object
Bacterial RNA Transcription
Make-Contact
Bacterial-DNA RNA-Polymerase
Place
location
location
Move Recognize Transcribe Detach Break-Contact
RNA-Transcript
GenePromoter Terminator
regions
location
object
Summary
• SME assembles a declarative representation from both generic and domain-specific components– SME specifies only the components and the links in the
assembly; most of the complexity within components is kept “under the hood”
• KANAL can “exercise” the declarative representation, verifying completeness and consistency
• KM’s simulator can execute the declarative representation to answer questions