Slide1 : CSC 550: Introduction to Artificial Intelligence
Fall 2004
Knowledge representation
associationist knowledge
semantic nets, conceptual dependencies
structured knowledge
frames, scripts
alternative approaches
Knowledge representation : Knowledge representation underlying thesis of GOFAI: Intelligence requires
the ability to represent information about the world, and
the ability to reason with the information knowledge representation schemes
logical: use formal logic to represent knowledge
e.g., state spaces, Prolog databases
procedural: knowledge as a set of instructions for solving a problem
e.g., production systems, expert systems (next week)
associationist: knowledge as objects/concepts and their associations
e.g., semantic nets, conceptual dependencies
structured: extend networks to complex data structures with slots/fillers
e.g., scripts, frames
Semantic nets (Quillian, 1967) : Semantic nets (Quillian, 1967) main idea: the meaning of a concept comes from the way it is connected to other concepts
SNOW in understanding language and/or reasoning in complex environments, we make use of the rich associativity of knowledge
When Timmy woke up and saw snow on the ground, he immediately turned on the radio.
graphs of concepts : graphs of concepts can represent knowledge as a graph
nodes represent objects or concepts
labeled arcs represent relations or associations such graphs are known as semantic networks (nets)
the meaning of a concept is embodied by its associations to other concepts retrieving info from a semantic net can be seen as a graph search problem
to find the texture of snow
find the node corresponding to "snow"
find the arc labeled "texture"
follow the arc to the concept "slippery"
semantic nets & inheritance : semantic nets & inheritance in addition to data retrieval, semantic nets can provide for deduction using inheritance since a canary is a bird, it inherits the properties of birds (likewise, animals)
e.g., canary can fly, has skin, …
to determine if an object has a property,
look for the labeled association,
if no association for that property, follow is_a link to parent class and (recursively) look there
Inheritance & cognition : Inheritance & cognition Quillian and Collins (1969) showed that semantic nets with inheritance modeled human information storage and retrieval
Semantic nets in Scheme : Semantic nets in Scheme (define ANIMAL-NET
'((canary can sing) (canary is yellow) (canary is-a bird)
(ostrich is tall) (ostrich cannot fly) (ostrich is-a bird)
(bird can fly) (bird has wings) (bird has feathers)
(bird is-a animal) (fish is-a animal)
(animal can breathe) (animal can move) (animal has skin))) can define a semantic net in Scheme as an association list
Semantic net search : Semantic net search ;;; net.scm
(define (lookup object property value NETWORK)
(define (get-parents object NET)
(cond ((null? NET) '())
((and (equal? object (caar NET)) (equal? 'is-a (cadar NET)))
(cons (caddar NET) (get-parents object (cdr NET))))
(else (get-parents object (cdr NET)))))
(define (inherit parents)
(if (null? parents)
#f
(or (lookup (car parents) property value NETWORK)
(inherit (cdr parents)))))
(if (member (list object property value) NETWORK)
#t
(inherit (get-parents object NETWORK)))) to lookup a relation
if arc with desired label exists, done (SUCCEED)
otherwise, if is_a relation holds, follow the link and recurse on that object/concept > (lookup 'canary 'is 'yellow ANIMAL-NET)
#t
> (lookup 'canary 'can 'fly ANIMAL-NET)
#t
> (lookup 'canary 'can 'breathe ANIMAL-NET)
#t
> (lookup 'canary 'is 'green ANIMAL-NET)
#f > (lookup 'ostrich 'cannot 'fly ANIMAL-NET)
#t
> (lookup 'ostrich 'can 'fly ANIMAL-NET)
#t
WHY?
Semantic net search, with negative relations : Semantic net search, with negative relations ;;; net.scm
(define ANIMAL-NET
'((canary can sing) (canary is yellow) (canary is-a bird)
(ostrich is tall) (ostrich (not can) fly) (ostrich is-a bird)
(bird can fly) (bird has wings) (bird has feathers)
(bird is-a animal) (fish is-a animal)
(animal can breathe) (animal can move) (animal has skin)))
(define (lookup object property value NETWORK)
(define (opposite property)
(if (symbol? property)
(list 'not property)
(cadr property)))
(define (get-parents object NET)
(cond ((null? NET) '())
((and (equal? object (caar NET)) (equal? 'is-a (cadar NET)))
(cons (caddar NET) (get-parents object (cdr NET))))
(else (get-parents object (cdr NET)))))
(define (inherit parents)
(if (null? parents)
#f
(or (lookup (car parents) property value NETWORK)
(inherit (cdr parents)))))
(cond ((member (list object property value) NETWORK) #t)
((member (list object (opposite property) value) NETWORK) #f)
(else (inherit (get-parents object NETWORK))))) to lookup a relation
if arc with desired label exists, done (SUCCEED)
if arc with opposite label exists, done (FAIL)
otherwise, if is_a relation holds, follow the link and recurse on that object/concept > (lookup 'ostrich
'(not can)
'fly
ANIMAL-NET)
#t
> (lookup 'ostrich
'can
'fly
ANIMAL-NET) #f
Implementation comments : Implementation comments DISCLAIMER: this semantic net implementation is simplistic
need to be able to differentiate between instances and classes
need to differentiate between properties of a class and properties of instances of that class
need to handle multiple inheritance paths Quillian used an intersection algorithm to find word relationships
given two words, conduct breadth first search from each node
look for common concepts (intersection nodes from the searches)
Conceptual dependency theory : Conceptual dependency theory not surprisingly, early semantic nets did not scale well
most links were general associations
no real basis for structuring semantic relations
much research has been done in defining richer sets of links
rely on richer formalism, not richer domain knowledge Conceptual Dependency Theory (Schank, 1973)
attempts to model the semantic structure of natural language
4 primitive conceptualizations, from which meaning is built
ACT action
PP objects (picture producers)
AA modifiers of actions (action aiders)
PA modifiers of objects (picture aiders)
primitive actions include: ATRANS (transfer a relationship, e.g., give)
PTRANS (transfer physical location, e.g., move)
MTRANS (transfer mental information, e.g., tell)
. . .
conceptual dependency relationships : conceptual dependency relationships tense/mode modifiers
p past
f future
t transition
? interrogative
/ negative
. . .
CD examples : CD examples John ate an egg.
John prevented Mary from giving a book to Bill.
CD for natural language understanding : CD for natural language understanding in the context of natural language understanding, the Conceptual Dependency representation has interesting properties:
knowledge is represented using conceptual primitives
actual words/phrases are not stored directly
ideally, representation is independent of the original language (could be English, French, Russian, …) John sold Mary a book.
Mary bought a book from John.
Mary gave John a check for the book.
these sentences describe the same event – a CD representation would reduce these to the same conceptual symbols
ADVANTAGE: syntax is minimized, semantics matters
RESULT: CD representation is good for understanding or paraphrasing sentences
MARGIE (Schank, 1973) : MARGIE (Schank, 1973) MARGIE: Memory, Analysis, Response Generation in English
the system combined a
parser (English CD)
generator (CD English)
inference engine (inferred info from CD) MARGIE in inference mode
INPUT: John gave Mary an aspirin.
OUTPUT1: John believes that Mary wants an aspirin.
OUTPUT2: Mary is sick.
OUTPUT3: Mary wants to feel better.
OUTPUT4: Mary will ingest the aspirin. MARGIE in paraphrase mode
INPUT: John killed Mary by choking her.
OUTPUT1: John strangled Mary.
OUTPUT2: John choked Mary and she died because she could not breathe.
Frames (Minsky, 1975) : Frames (Minsky, 1975) in contrast to distributed knowledge networks, can instead organize knowledge into units representing situations or objects
When one encounters a new situation (or makes a substantial change in one's view of a problem) one selects from a memory structure called a "frame." This is a remembered framework to be adapted to fit reality by changing details as necessary.
-- Marvin Minsky HOTEL ROOM
Frame example : Frame example a frame is a structured collection of data
has slots (properties) and fillers (values)
fillers can be links to other frames
Frame set in Scheme : Frame set in Scheme represent a frame as a nested structure
(define ANIMAL-FRAME
'((canary (can sing)
(is yellow)
(is-a bird))
(ostrich ((not can) fly)
(is tall)
(is-a bird))
(bird (can fly)
(has wings feathers)
(is-a animal))
(fish (is-a animal))
(animal (can breathe move)
(has skin))))
Frame search : Frame search ;;; frame.scm
(define (lookup object property value FRAME)
(define (opposite property)
(if (symbol? property)
(list 'not property)
(cadr property)))
(define (get-parents object)
(let ((parents (assoc 'is-a (cdr (assoc object FRAME)))))
(if (not parents)
'()
(cdr parents))))
(define (inherit parents)
(if (null? parents)
#f
(or (lookup (car parents) property value FRAME)
(inherit (cdr parents)))))
(let ((entry (assoc object FRAME)))
(if (not entry)
#f
(let ((vals (assoc property (cdr entry)))
(negvals (assoc (opposite property) (cdr entry))))
(cond ((and vals (member value (cdr vals))) #t)
((and negvals (member value (cdr negvals))) #f)
(else (inherit (get-parents object)))))))) to perform a deduction
get frame information,
if desired slot exists, get filler
if opposite of slot exists, fail
otherwise, if there is an is-a slot, get the parent frame and recurse on that object/concept
Implementation comments : Implementation comments DISCLAIMER: again, this implementation is simplistic
need to be able to differentiate between instances and classes
need to differentiate between properties of a class and properties of instances of that class
need to handle multiple inheritance paths The structured nature of frames makes them easier to extend
can include default values for slots
can specify constraints on slots
can attach procedures to slots
Representation applications : Representation applications semantic nets, frames, and scripts were foundational structures
more recently, they have been adapted and incorporated into hybrid structures vision
Minksy saw frames as representing different perspective of an object
understanding
use frames with defaults to "fill in the blanks" in understanding
EXAMPLE: "I looked in the janitor's closet …"
Lenat's AM represented concepts as frames (newly concepts spawned new frames)
smart databases
Lenat's CYC project used extension of frames, with conventions for slots & fillers
PARKA project at Maryland uses frame-based language for efficiently accessing large knowledge bases
Hyundai Engineering uses frame-based system for planning construction projects interesting note:
MIT research on frames (and similar research at XEROX PARC) led to object-oriented programming and the OOP approach to software engineering
Scripts (Schank & Abelson, 1975) : Scripts (Schank & Abelson, 1975) a script is a structure that describes a stereotyped sequence of events in a particular context
closely resembles a frame, but with additional information about the expected sequence of events and the goals/motivations of the actors involved
the elements of the script are represented using Conceptual Dependency relationships (as such, actions are reduced to conceptual primitives) EXAMPLE: restaurant script
describes: items usually found in a restaurant
people and their roles (e.g., chef, waiter, …)
preconditions and postconditions
common scenes in a restaurant: entering, ordering, eating, leaving
Hotel script : Hotel script props and roles are identified
pre- and post-conditions
CDs describe actions that occur in each of the individual scenes
Script application : Script application SAM: Script Applier Mechanism
Cullingford & Schank, 1975
system consisted of:
parser (extension of MARGIE)
generator (extension of MARGIE)
script applier (to check the consistency of the CD repr. with that specified in the script)
question answerer
Alternatives to explicit representation : Alternatives to explicit representation connectionist & emergent approaches (later)
Subsumption architecture (Brooks, MIT)
claim: intelligence is the product of the interaction between an appropriately layered system and its environment
architecture is a collection of task-handling behaviors, with each behavior accomplished via a finite state machine
limited feedback between layers of behavior
"… in simple levels of intelligence, explicit representations and models of the world simply get in the way. It turns out to be better to use the world as its own model." (Brooks)
Copycat architecture (Mitchell & Hofstadter, Indiana)
builds on representation techniques from semantic nets, blackboards, connectionist networks, and classifier systems
supports semantic net-like representation that can evolve
emphasizes analogical reasoning
Next week… : Next week…
Expert systems
rule-based vs. model-based vs. case-based
probabilistic vs. fuzzy reasoning
Read Chapters 7, 8
Be prepared for a quiz on
this week’s lecture (moderately thorough)
the reading (superficial)