Transcript Some Desiderata for Machine Understanding

Some Desiderata For Machine
Understanding
Peter Clark
May 2002
Knowledge Systems
Boeing Engineering and Information Technology
On Machine Understanding
• Understanding = creating a situation-specific model
(SSM), coherent with data & background knowledge
– Data suggests model fragments which may be appropriate
– Models suggest ways of interpreting data
?
?
Garbled graph
of relationships
Coherent Model
(situation-specific)
On Machine Understanding
• Core theories of the world
• Ton of common-sense/
episodic/experiential knowledge
(“the way the world is”)
• Only a tiny part of the target model
• Contains errors and ambiguity
• Not even a subset of the target model
Assembly of
pieces,
assessment of
coherence,
inference
?
?
Garbled graph
of relationships
Coherent Model
(situation-specific)
What are some ingredients?
1.
2.
3.
4.
5.
Elaboration (“scene building”)
Representing possibilities
Coherence assessment (“matching”?)
Viewpoints/context
Knowledge acquisition
1. Elaboration:
The Parachute Sentences
“Parachutes slow
down a person
falling through the
air. This means that
he or she can land
safely when jumping
out of a plane. When
open, a parachute
creates lots of drag
as air pushes against
its underside. This
slows its fall.”
The Parachute Sentences
“Parachutes slow
down a person
falling through the
air. This means that
he or she can land
safely when jumping
out of a plane. When
open, a parachute
creates lots of drag
as air pushes against
its underside. This
slows its fall.”
1. Elaboration (cont)
“John chopped down the tree.”
• A vivid picture comes to mind
– John: adult male, out in woods
– holding an axe (or chain saw?)
– Tree is ~30ft high pine tree
• or: a modification of that time I sawed a Christmas tree
• or: that documentary on logging in Canada
•
•
•
•
8:1 ratio of prior to explicit knowledge (Graesser, ’81)
Episodic/experiential knowledge plays a key role
Also core knowledge plays a key role
Not a deductive process!
1. Elaboration: Using WordNet
• Augment “semantic structure” with definitional “knowledge”.
“The kid hit the ball hard.”
1. Elaboration: Using WordNet
• Augment “semantic structure” with definitional “knowledge”.
“The kid hit the ball hard.”
1. Elaboration: Another example
“The Global Positioning System is a satellite navigation
system designed to provide instantaneous position, velocity
and time information almost anywhere on the globe.”
•
•
•
•
•
satellite: orbit around earth; receive/send radio messages
navigation: information about location
system: assembly of artifacts which together perform a task
people: often want to know where they are
(after more sentences): entire model on how GPS systems
work.
2. Representing Possibilities
• Went to encode a space of possibilities
– not what the model is, but constraints on what the
actual models might be
– enable actual models to be built and assessed
where?
“Most eucaryotic genes have their coding sequences interrupted
by noncoding sequences, called introns. The scattered pieces of
coding sequence, called exons, are usually shorter than the
introns, and the coding portion of a gene is often only a small
fraction of the total length of the gene. Most introns range in
length from about 80 nucleotides to 10,000 nucleotides,
although even longer introns exist.”
p220, Alberts 1998.
2. Representing Possibilities
Got
Want
Possible (consistent)
More likely/
preferred
Impossible (inconsistent)
Less likely/preferred
Spaces of possible models
2. Representing Possibilities
• Are a few (feeble) methods in KM for this:
– type restrictions
(every Person has (spouse ((must-be-a Person)))))
– (sometimes <x>)
• (every Car has (parts ((sometimes (a Spare-Wheel)))
– Cardinality constraints
• (a Group with (min-cardinality (…)) (max-cardinality (…)))
– Range of values
• “size is between X and Y”
• Still largely lacking in how to represent and reason
with vague knowledge
3. What Makes a Representation
(Model) Coherent?
• We don’t just blindly accept new knowledge:
– Minsky: We proactively ask a set of pertinent questions about a
scene, e.g., what is X for? What are the goals? etc.
• What makes a representation coherent?
– Simple consistency (“The man fired the gnu.”)
– Purposefulness (for artifacts):
• “The engine contains a thrust reverser.”
• vs. “The engine contains an elephant.”
• vs. “The engine contains a book.”
• “Knowledge entry” is a serious misnomer!
– Really talking about Knowledge Integration
3. What Makes a Representation
(Model) Coherent?
“TRANSPONDER: A combination receiving and transmitting antenna on a
communications satellite.
TRANSPONDER: A combination receiver, frequency converter, and transmitter
package, physically part of a communications satellite.
Transponder
parts: receiver
frequency converter
transmitter
part-of: communications satellite
Transponder
parts: antenna
purpose: receive, transmit
Relay/Mediate
3. A Catalog of Coherence Criterea
1. Volitional actions:
–
Agents must be capable of an action
•
–
–
legally, skill, fiscally, anatomically
Action serves a broader purpose/goal
Need equipment/resources/instruments, instruments must be
adequate
2. Non-volitional actions;
–
–
There is a cause (inc. randomness)
Spatial:
•
•
–
statics: objects must be close
dynamics: objects can move in the required way
Temporal: objects exist at the same time
3. A Catalog of Coherence Criterea (cont)
3. Objects:
– physically possible
•
•
•
–
parts connected together at appropriate places
materials are appropriate
suspension/tension etc., gravity
physically normal/expected/standard
•
need to know normal shapes, sizes, etc.
4. Artifacts:
–
Purposefulness:
•
–
all parts play some role wrt. one of its intended functions (or
subtasks thereof). Expect design to be optimized.
Could treat biological objects as “artifacts”
3. Coherence and KM
• KM unable to tolerate incoherence:
– Current: “Error! Switching on the debugger…”
– Desired: “This representation is generally ok, except
this bit looks weird, and that bit conflicts with this bit.”
• Problem compounded by long inference chains
– (cf. Cyc: don’t think too hard )
• How could we change KM to be more tolerant?
4. Viewpoints and Context
Component
theories/
ontologies(?)
vs.
Reason over
Giant KB
Problem-specific KB,
contains selected units
• Latter seems right, but:
– can a big KB really be partitioned like this? (everything is
connected!)
• Models may vary by:
– ignoring detail
– making different approximations
– using different ontologies
4. Viewpoints and Context
• e.g., DNA = sequence of different region types:
– intron-exon-intron-exon…
– promotor-gene-terminator
– nucleotide pair-nuc pair-nuc pair…
• Makes a difference:
– Given: “The polymerase attaches to the promotor, and then moves down
the strand.”
– then answer: “Where will the polymerase be?”
• nucleotide? gene? intron?
• The point: It’s not simply a matter of having all viewpoints
coexisting
• Another example: “A satellite sends signals/messages/position
information.”
5. Knowledge Acquisition
• Where do the core theories come from?
– Hand engineered?
• Where does all the “mundane” knowledge come from?
– Schubert-style?
– Dictionary/glossary definitions?
5. Knowledge Acquisition:
Can all this be Bootstrapped?
Collection of
coherent scenes
Jungle of
parse trees/
semantic graphs
KB
Text
List of compound
nouns and verbs
(entities and actions)
Domain
Ontology
Scene
library