Building Knowledge Bases from Reusable Components

1
Building Knowledge Bases from Reusable Components

• Bruce Porter
• Computer Science Dept, Univ of Texas at Austin
• Peter Clark
• Boeing Applied Research and Technology

2
Fragment of a Knowledge-Base
...
Soil
Rate
Person
contains
...
I-
I-
Q
isa
environment
Q-
rate
agent
Bio- technologist
Bioremediation
Amount
Amount
...
amount
amount
script
remediator
product
pollutant
agent
...
Oil
Fertilizer
Microbes
Script
patient
se
se
se
se
patient
agent
absorbed
product
Break Down
Get
Apply
Absorb
then
then
then
...
...
Example queries

• What steps are involved in bioremediation?
• How does pollutant volume affect rate?
• What equipment is needed?
• What do the microbes do?
• ...

3
The Botany KB Experience

• 10 year effort, 20k concepts, 100k facts
• Supports sophisticated question-answering
• description
• prediction
• But
• KB still highly incomplete
• laborious to build/maintain
• difficult to achieve reuse

? want a better approach!
4
Fundamental Problem

• Reliance on specific, comprehensive
  representations
• impractical, unmaintainable
• cant anticipate them all
• But
• representations contain repeated abstractions
• production occurs in photosynthesis, mitosis,
  growth
• germination includes conversion, production,
  expansion
• Goal Capture abstractions in a recomposable way

5
Representation of Bioremediation
Soil
Rate
contains
I-
I-
Q
environment
Q-
rate
agent
Bio- technologist
Bioremediation
Amount
Amount
amount
amount
script
remediator
product
pollutant
agent
Oil
Fertilizer
Microbes
Script
patient
se
se
se
se
patient
agent
absorbed
product
Break Down
Get
Apply
Absorb
then
then
then
6
An underlying abstraction...
Soil
Rate
contains
I-
I-
Q
environment
Q-
rate
agent
Bio- technologist
Bioremediation
Amount
Amount
amount
amount
script
remediator
product
pollutant
agent
Oil
Fertilizer
Microbes
Script
patient
se
se
se
se
patient
agent
absorbed
product
Break Down
Get
Apply
Absorb
then
then
then
Rate
I-
I-
Q
Q-
rate
Conversion
Amount
Amount
amount
amount
raw- materials
product
Substance
Substance
7
Another abstraction...
Soil
Rate
contains
I-
I-
Q
environment
Q-
rate
agent
Bio- technologist
Bioremediation
Amount
Amount
amount
amount
script
remediator
product
pollutant
agent
Oil
Fertilizer
Microbes
Script
se
se
se
patient
se
patient
agent
absorbed
product
Break Down
Get
Apply
Absorb
then
then
then
Digest
food
script
eater
Substance
Agent
Script
agent
se
patient
se
absorbed
agent
Break Down
Absorb
then
8
Another abstraction...
Soil
Rate
contains
I-
I-
Q
environment
Q-
rate
agent
Bio- technologist
Bioremediation
Amount
Amount
amount
amount
script
remediator
product
pollutant
agent
Oil
Fertilizer
Microbes
Script
patient
se
se
se
se
agent
patient
absorbed
product
Break Down
Absorb
Get
Apply
then
then
then
Treatment
script
substance
Script
substance
se
patient
patient
Get
Apply
then
9
A Component-Based Approach

• Represent component abstractions explicitly
• Define concepts as compositions
• Construct representations on-demand to answer qns

KB Architecture

• Component theories abstract, reusable models
• Definitions specifications of compositions
• Inference construct compositions as needed to
  answer questions.

10
Lessons from a Dictionary...
Car a Vehicle for Passengers
11
2. Definitions and Composition

• Definition specification of a composition
• a Fuel-Cell is a Producer of Electricity
• a Bulb is an Electrical Resistor producing Light
• a Camera is an Image Recording Device
• a Wire is a Conduit of Electricity
• Automated composition
• Elaboration component supplies info to answer
  query
• Classification recognize concepts in the
  composition

12
2. Composition (example)
Composition Camera an Image Recording Device
Query Failure modes of a camera?
(Camera has (superclasses (Device))) (every
Camera has (behavior ((a Recording with
(input (a Image)))))
13
Component Theory Devices
(Device has (superclasses (Physobj))) the
failure modes of a device are the failure
modes of its functional parts (every Device has
(behavior ((a Activity))) (failure-modes (
(the failure-modes of
(the participants of
(the behavior of Self)))))))
14
Query Failure modes of a camera?
Sub-query Participants in its behavior?
15
Component Theory Recording
(Recording has (superclasses
(Activity))) (every Recording has (input ((a
Signal))) (participants ( (a Receptor
with (input ((the input of Self)))
...
16
failure- mode
failure- mode
Failure- Mode
Failure- Mode
Device
Run-Time Classification Aperture an Image
Receptor
behavior
failure- mode
failure- mode
input
Recording
Image
Signal
part.
output
input
part.
input
Receptor
Memory-Unit
subevents
agent
patient
Receiving
Writing
17
Aperture - inputs an image - outputs an image
- might be blocked - ...
failure- mode
failure- mode
Failure- Mode
Failure- Mode
Device
Run-Time Classification Aperture an Image
Receptor
behavior
failure- mode
failure- mode
input
Recording
Image
Signal
part.
output
input
part.
input
Receptor
Memory-Unit
subevents
agent
patient
Receiving
Writing
18
failure- mode
failure- mode
Failure- Mode
Blockage
Device
Run-Time Classification Aperture an Image
Receptor Film an Image Memory-Unit
behavior
failure- mode
failure- mode
input
Recording
Image
Image
part.
output
part.
input
input
Aperture
Memory-Unit
subevents
agent
patient
Receiving
Writing
19
Film - includes a sheet coated with
image-sensitive chemical - might age - ...
failure- mode
failure- mode
Failure- Mode
Blockage
Device
Run-Time Classification Aperture an Image
Receptor Film an Image Memory-Unit
behavior
failure- mode
failure- mode
input
Recording
Image
Image
part.
output
part.
input
input
Aperture
Memory-Unit
subevents
agent
patient
Receiving
Writing
20
Query Failure modes of a camera? Blockage, Aging
Sub-query Participants in its behavior?
Aperture, Film
21
Other Compositions...
Activity
Signal
Producing
Recording
Camera
Projector
Image
Stereo
Tape Recorder
Sound
Vibrator
Seismograph
Vibration
22
Definitions

• "A CAMERA is an IMAGE RECORDING DEVICE."
• (Camera has (superclasses (Device)))
• (every Camera has-definition
• (instance-of (Device))
• (behavior ((a Recording with
• (input ((a Image)))))))
• "A SEISMOGRAPH is a VIBRATION RECORDING
  DEVICE."
• (Seismograph has (superclasses (Device)))
• (every Seismograph has-definition
• (instance-of (Device))
• (behavior ((a Recording with
• (input ((a Vibration)))))))
• "A STEREO is a SOUND PRODUCING DEVICE."
• (Stereo has (superclasses (Device)))
• (every Stereo has-definition
• (instance-of (Device))
• (behavior ((a Producing with

23
Generic Devices

• (Device has (superclasses (Physobj)))
• "The failure-modes of a device are the
  failure-modes of its
• (functional) parts."
• (every Device has
• (behavior ((a Activity)))
• (parts ((the participants of (the behavior of
  Self))))
• (failure-modes (
• (the failure-modes of
• (the parts of Self)))))

24
Generic Activity Recording a Signal

• (Recording has (superclasses (Activity)))
• "Recording involves receiving a signal, then
  writing it to a memory."
• (every Recording has
• (input ((a Signal)))
• (subevents (
• (a Receiving with
• (object ((the input of Self)))
• (agent ((the Receptor participants of
  Self))))
• (a Writing with
• (object ((the output of (the Receptor
  participants of Self))))
• (patient ((the Memory-Unit participants of
  Self))))))
• (participants (
• (a Receptor with
• (input ((the input of Self))))
• (a Memory-Unit with
• (input ((the output of (the Receptor
  participants of Self))))))))

25
Generic Activity Producing a Signal

• (Producing has (superclasses (Activity)))
• "Producing a signal involves converting power
• (provided by a Power-Supply) into a target
  signal."
• (every Producing has
• (output ((a Signal)))
• (subevents (
• (a Supplying with
• (agent ((the Power-Supply participants of
  Self))))
• (a Converting with
• (agent ((the Converter participants of
  Self)))
• (object ((the output of Self)))
• (after ((the Supplying subevents of
  Self))))))
• (participants (
• (a Power-Supply with
• (output ((the input of (the Converter
  participants of Self)))))
• (a Converter with
• (input ((the output of (the Power-Supply
  participants of Self))))

26
Particular Devices

• "An APERTURE is an IMAGE DETECTOR..."
• (every Aperture has-definition
• (instance-of (Receptor))
• (input ((a Image))))
• "...which outputs an image, and might get
  blocked."
• (every Aperture has
• (output ((a Image)))
• (failure-modes ((a Blockage with (object
  (Self))))))
• "a FILM is an IMAGE MEMORY-UNIT..."
• (every Film has-definition
• (instance-of (Memory-Unit))
• (input ((a Image))))
• "...which might age, and includes a
  chemical-covered sheet."
• (every Film has
• (failure-modes ((a Aging with (object
  (Self)))))
• (parts ((a Sheet with

27
Particular Activities

• "EXPOSING is WRITING on FILM."
• (every Exposing has-definition
• (instance-of (Writing))
• (patient ((a Film))))
• "TAPING is WRITING to TAPE."
• (every Taping has-definition
• (instance-of (Writing))
• (patient ((a Tape))))
• "SHINING is GENERATING an IMAGE."
• (every Shining has-definition
• (instance-of (Generating))
• (object ((a Image))))

28
Natural Language Generation

• (every Blockage has
• (text ((Self object) "is blocked, preventing"
• (the input of (the object of Self))
• "from reaching"
• (the parts-of of (the object of
  Self)))))
• (every Discharged has
• (text ((Self object) "that powers"
• (the Converter parts of the
  converter part of ...
• (the parts-of of
  the system containing ...
• (the object of Self)))
  the battery
• "is discharged")))
• (every Recording has
• (text ("First," (the Receptor participants of
  Self)
• "captures" (the input of Self)
  ", then"
• (the Memory-Unit participants
  of Self) "records"
• (the input of Self))))

29
Compound Concepts are Ubiquitous

• Botany
• leaf mesophyll cell
• aerial root system
• plant material distribution
• photosynthetic organ level structure
• ...
• Aerospace
• turbine gearbox assembly
• logic speed control unit
• case drain fluid
• aircraft hydraulic system performance analysis
  program
• (43k acronyms!)...

30
Composing Representations What are the building
blocks?

• Events Levin (1993) classifies about 3000 events
  into about 25 categories sharing common
  components.
• Objects Ogden (1934) identifies about 800 basic
  objects.
• Shapes and Places Landau and Jackendoff (1993)
  systematize descriptions of objects' shapes and
  locations.
• Compound Objects Barker (1998) identified 20
  ways that noun phrases are built from components
  similar studies underway for verb phrases.
• Core Theories Forbus, Kuipers, Allen, et.al.
  contribute representations of time, space,
  quantities, and causation.

31
Levin's Verb Classes(B. Levin, English Verb
Classes and Alternations, University of Chicago
Press, 1993)

• put
• remove
• send and carry
• push and pull
• give and take
• hold and keep
• conceal
• touch and contact
• combine and attach

• separate and disassemble
• create and transform
• search
• communicate
• ingest
• emit
• destroy
• appear, disappear, and occur
• about 10 more

32
Building a Component Library examples from the
move cluster

• Movement is an event that changes location
• (Move has (superclasses (Event)))
• (every Move has
• (patient ((a Physical-thing)))
• (changes ((a Change with
• (property ((the location
  of
• (the
  patient of Self))))
• (from ((the source of
  Self)))
• (to ((the destination of
  Self))))))
• (source ((a Place)))
• (destination ((a Place))))

33
Adding constraints to define new members of the
cluster

• Locomotion is self movement
• (Locomotion has (superclasses (Move)))
• (every Locomotion has
• (agent ((the patient of Self))))

34
Adding constraints to define Flow

• Flow is movement of a liquid or gas through a
  channel or across a surface
• (Flow has (superclasses (Move)))
• (every Flow has
• (patient ((a Liquid) or (a Gas)))
• (direction ((a Direction with
• (path-relation
  (through))
• (reference-object ((a
  Channel))))
• or
• (a Direction with
• (path-relation
  (across))
• (reference-object ((a
  Surface)))))))

35
Combining ComponentsRole Filling

• an Enter is a Move from outside a container
  to inside that container,
• through the container's portal
• (Enter has (superclasses (Move)))
• (every Enter has
• (source ((a Place with
• (place-relation
  (outside))
• (reference-object ((a
  Container))))))
• (destination ((a Place with
• (place-relation
  (in))
• 
  (reference-object ((Self source Place
• 
  reference-object Container))))))
• (direction ((a Direction with
• (path-relation
  (through))
• (reference-object
  ((the portal of
• 
  (Self destination Place
• 
  reference-object Container))))))))

36
Combining Components multiple inheritance

• Inflow is a movement
• from flow of a liquid or gas through a
  channel or across a surface
• from enter from outside a container to
  inside that container, through the container's
  portal
• (Inflow has (superclasses (Flow Enter)))

37
Combining ComponentsSequencing events

• A carry is locomotion while supporting
  something
• (Carry has (superclasses (Move)))
• (every Carry has
• (subevent ((a Support with
• (agent ((the agent of
  Self)))
• (patient ((the patient
  of Self)))
• (concurrent-with ((the
  Locomotion subevent of Self))))
• (a Locomotion with
• (agent ((the agent of
  Self)))
• (concurrent-with ((the
  Support subevent of Self)))))))

38
Combining ComponentsCausal Connection

• a push is the application of a force causing
  a movement
• (Push has (superclasses (ApplyForce)))
• (every Push has
• (direction ((a Direction with
• (path-relation
  (away))
• 
  (reference-object ((the agent of Self))))))
• (causes ((a Move with
• (patient ((the
  patient of Self)))
• (direction ((the
  direction of Self)))))))

39
Combining ComponentsIteration over a Set

• Convey is the movement of a container and all
  its contents
• (Convey has (superclasses (Move)))
• (every Convey has
• (patient ((a Container)))
• (causes ((forall ((the contents of (the
  patient of Self)))
• ((a Move with
• (patient (It))
• (source ((the source
  of Self)))
• (destination ((the
  destination of Self)))
• (concurrent-with
  ((Self)))))))))

40
Experimental Design Measuring Successful Reuse

• Small scale reuse build component library for a
  subset of the "Pumps Dictionary". Measure
  successful reuse of components for representing
  the remainder. Examples Air Driven Pump, Water
  Circulating Pump
• Medium scale reuse do the same for abstracts of
  the "Flight Incident Database". Example "Lost
  hydraulic controls deployed air driven pump
  returned to airport."
• Large scale reuse do the same for the full
  accounts in the Database.

41
Measuring Successful Reuse

• The success of a reuse effort is measured by the
  role of the components in answering questions in
  novel domains.
• Measured in three ways
• recall the fraction of the KB-structure (that's
  built to answer each question) that's contributed
  by components
• accuracy the correctness of the knowledge
  contributed by the components
• completeness the comprehensiveness of the
  knowledge contributed by the components

42
Discussion

• A knowledge base contains
• explicit representations of "building block"
  concepts
• specifications for building representations of
  other concepts using these building blocks
• methods for composing concepts on-demand and
  (possibly) caching the results
• This approach improves reuse and reduces
  representational drift

43
Discussion

• How broad should each component be? three
  versions of drop
• downward movement
• a release followed by downward movement
• until something breaks the fall
• Will idiosyncratic variations within real domains
  overwhelm the regularities in a library of
  components?

44
Potentials
The growing demand for knowledge processing

• Growth of on-line, structured information
• XML
• On-line databases (eg. commercial, geographic)
• eCommerce
• Search / Information Access
• best search engine wins...
• Knowledge Management
• NL understanding / MLT / Speech
• Knowledge-Based Engineering

45
Combining ComponentsSequencing events

• (Bus-ride has (superclasses (Ride)))
• (every Bus-ride has
• (instrument ((a Bus)))
• (subevent ((a Enter with
• (patient ((the
  patient of Self)))
• (destination ((a
  Place with
• 
  (reference-object
• 
  ((the instrument of Self))))))
• (followed-by ((the
  Convey subevent of Self)))))
• ((a Convey with
• (patient ((the
  instrument of Self)))
• (followed-by ((the
  Exit subevent of Self)))))
• ((a Exit with
• (patient ((the
  patient of Self)))
• (source ((the
  instrument of Self)))))))

• Bus-ride is a Movement with
• an instrument of a Bus, modeled as a container
• an Enter event with destination in the bus
• a Convey with agent the bus
• an Exit with source in the bus

46
Demo...
KMgt
(showme Recording)
(every Recording has (input ((a Signal)))
(participants (... KMgt
(a Device with (behavior ((a Recording with
(input (Image))))))
_Device01 KMgt
(the parts of _Device01)
(trace _Receptor31 satisfies definition of
Aperture) (trace _Memory-Unit33 satisfies
definition of Film) (_Aperture32 _Film34) KMgt
(the failure-modes of _Device01)
(Blockage Aging) KMgt
(the text of (the failure-modes of _Device01))
(The aperture is blocked. The film is
aged.) KMgt
(the subevents of (the behavior of _Device01))
(trace _Writing40 satisfies definition of
Exposing) (_Receiving41 _Exposing42) KMgt
(the patient of _Exposing42)
_Film34 KMgt
47
Overall Architecture
48
Overall Architecture
1. Ontology (conceptual vocabulary)
2. Component theories (computational clockwork)
Thing
Consumer
Producer
Circuit
...
...
...
49
Overall Architecture
1. Ontology (conceptual vocabulary)
2. Component theories (computational clockwork)
Thing
Activity
Physobj
Move
...
...
...
50
Overall Architecture
1. Ontology (conceptual vocabulary)
3. Definitions and Descriptions (describe
concepts in terms of others)
2. Component theories (computational clockwork)
Thing
...
...
...
Bulb Light-producing Electrical Consumer
51
Summary

• Demand and potential of knowledge processing
• Component-based architecture
• ontology
• component theories
• definitions (specifications of compositions)
• basic fact libraries
• Benefits
• KB development modular, reusable architecture
• KB application deal with novel concepts at
  run-time

52
1. Component Theories

• A coherent, encapsulated system of concepts
  relns
• Contains
• ontology (vocabulary of concepts and relations)
• axioms (rules) relating these
• Provides semantics for these concepts in the KB
• Can layer these theories (define one using
  others)

53
Example Distribution Network
Rules (axioms)

• PRODUCERS produce TRANSPORT-MATERIAL.
• CONSUMERS can consume TRANSPORT-MATERIAL.
• A network element may be BLOCKED or UNBLOCKED.
• If an element directly connects with an UNBLOCKED
  element, then it has an ACCESS to that element.
• A CONSUMER is SUPPLIED if it has ACCESS to a
  PRODUCER.
• .

Ontology
Producer Intermediary Consumer Transport-Material
connects supplied state
54
Axiom Representation (example)
If an element connects to an UNBLOCKED element,
then it has ACCESS to that element.
(i) Semantic network
? e1,e2Element connected-element(e1,e2) ?
state(e2,Unblocked) ?
access-to(e1,e2)
(ii) Logic
(every Element has (access-to (
(allof (the connected-element of Self)
where ((the state of It) Unblocked)))))
(iii) Implementation (KM)
55
Other component theories...

• Supply-and-demand
• Containment
• Machines
• Production network
• Two-state object
• Transportation
• ...