Knowledge Representation

1
Knowledge Representation
2
Key Issue in AI

• Mapping between objects and relations in a
  problem domain and computational objects and
  relations in a program.
• Results of inferences on the knowledge base (KB)
  should correspond to the results of actions or
  observations in the world.

3
Have Already Examined Two (related) KR Schemes

• First Order Predicate Logic
• Production Systems

4
Well look at four others

• Semantic Nets
• Conceptual Dependency Schemes
• Frames
• Scripts
• 1 2 are called network schemes
• 3 4 are called structured schemes
  (alternatively slot and filler schemes)

5
Problems with FOPL

• Emphasis is on truth-preserving relations
• Sometimes at odds with the way that humans
  acquire and use knowledge
• Leads to problems in mapping human language to
  FOPL

6
For Example

• If then
• In English suggests causality
• But
• ?
• In FOPL specifies a relationship between truth
  values of antecedent and consequent
• (22 5) ? color(elephants, green)
• This is true, but without common sense meaning.

7
Categorical Interlude

• Category
• A group of objects that seem to go together
• Because they have significant attributes in
  common
• Example DOG

8

• Allows us to use our finite mental resources
  efficiently
• When identifying an objects, we can abstract key
  attributes from all sensory information presented
  to us.
• I am trying to determine whether that flying
  object is a bird or a wasp.
• I dont care that robins have orange breasts and
  sparrows have grey.
• What matters are those attributes of category
  bird that exclude instances of category wasp

9

• Categories license inductive inferences
• Most birds pose no threat to humans
• Common wasps do
• Inference from category wasp tells us to avoid
  its members

10
Gelman Markmans Experiment

• Children were
• Shown a picture of a fish
• Told that it breathes under water
• Shown a picture of a dolphin
• Told that it breathes by jumping out of the water
• Shown a picture of a shark
• Told that it is a fish (though it looks like a
  dolphin)
• Were asked how it breathes
• Answered Under water.

11
Semantic Nets

• Proposed by Quillian in the late 1960s
• Tries to provide a formalism that captures
  taxonomic hierarchies
• A graph where
• Nodes are categories
• Arcs are of three types
• Isa links, indicating a subset relationship (a
  dog isa mammal)
• Inst links, indicating an element-set
  relationship (mazel is a dog)
• Attribute links, indicating a property held by a
  category (simcha is grey)

12
Example
thing
isa
Inanimate Thing
Animate thing
green
isa
isa
isa
cubic
color
Furniture
isa
shape
plant
Block
Supported_by
isa
isa
animal
Table
Ponderosa pine
Supported_by
Instance_of
legs
Instance_of
Block_1
Table_1
Supported_by
13
In (what else?) Prolog
14
Strengths of Semantic Nets

• Provides for inheritance
• Organizes knowledge using interconnected concepts
• Lets us discover relationships between pairs of
  concepts (block_1 and table_1 are both inanimate
  things and are supported by legs)

15
Weaknesses of Semantic Nets

• Generality of the attribute links
• As task grows in complexity, so does the
  representation
• No systematic basis for structuring semantic
  relationships
• Puts the burden of constructing facts links on
  programmer

16
Key Issue

• Isolation of primitives for semantic network
  languages
• Primitives are those things that the interpreter
  is programmed in advance to understand.
• We need a more systematic basis for structuring
  semantic relationships

17
Case Structure Grammars

• C.J. Fillmore, 1968
• Verb oriented (as opposed to concept-oriented)
• Sentences are represented as verb nodes with
  links to specific roles played by nouns and noun
  phrases
• Important links
• Agent
• Object
• Instrument
• Location
• Time

18
Mary caught the ball with her glove.
past
time
Mary
ball
catch
agent
object
glove
instrument
19
Advantages

• Representational language captures some of the
  deep structure of natural languages (i.e., the
  relationship between any verb and its subject is
  the agent relationship)
• This deep structure is independent of any
  sentence or even of any distinct language

20
Leading To

• Conceptual Dependency Theory
• Associated with Robert Schank (then of Yale, most
  recently of Northwestern)
• Attempts to model the semantic structure of
  natural language
• Attempts to provide a canonical form for the
  meaning of sentences
• That is, all sentences that mean the same thing
  (whatever that means) will be represented
  internally by identical graphs
• Idea is to parse two sentences that use different
  words but mean the same thing into identical
  internal representations
• Example John gave the book to mary/Mary was
  given the book by John.

21
Primitives in CD Theory

• ACTs actions
• PPs picture producers
• AAs modifiers of actions (action aiders)
• PAs Modifiers of objects (picture aiders)

22
Further Breakdown

• Each of these classes has a well-defined number
  of primitives (luger, pp. 236-37)
• All ACTs (actions) can be reduced to
• ATRANS transfer a relationship (give)
• PTRANS transfer a physical location (go)
• PROPEL apply physical force (push)
• MOVE move body part by owner (kick)
• 12. ATTEND focus sense organ (listen)

23
Yet More

• Indicates direction of dependency
• P indicates past
• F indicates future
• Indicates agent-verb relationship
• Indicates the object of an action
• ACT PP
• Agent instrument is an arrow pointing left

o
24
pp
ACT
pp
Recipient of an action
25
John gave the book to mary.
mary
R
John
ATRANS
john
p
book
26
Basic Idea

• Parse the sentence
• Fit it into canonical form
• Group sentences with similar meanings

27
Strengths

• Provides a formal theory of language semantics
• Reduces the problem of ambiguity
• Attempts to reduce the complexity of natural
  language by grouping sentences of similar meaning
  together.

28
Weaknesses

• Reduction is not computable in polynomial time
• No evidence that humans store knowledge in
  canonical forms
• Does not address the difficult issue of meaning
  in discourse
• Example
• Bill and John always walk home together. One
  afternoon, Bill said to John, Lets leave
  early. In effect, he asked him to go along with
  his plan of playing hooky.
• What are the referents of these three pronouns?

29
Canonical Sentences leads to Canonical Events

• NLP programs must use a large amount of
  background knowledge
• Evidence that we organize this information into
  structures corresponding to typical situations
• Example if we read a story about baseball, we
  resolve any ambiguities in the text in a way
  consistent with baseball

30
Example

• City Council refused to give the demonstrators a
  permit because they feared violence.
• City Council refused to give the demonstrators a
  permit because they advocated revolution.
• Background knowledge lets us determine the
  correct referent to they in each case.

31
Script

• Structural representation that describes a
  stereotypical sequence of events in a particular
  context.
• May be viewed as a causal chain

32
Components

• Entry conditions must be satisfied before the
  script is activated
• Result things that will be true after script
  completes
• Props slots representing objects that are
  involved in the events of the script.
• Roles slots representing people involved in the
  events of the script.
• Track Specific variation on a general pattern
• Scene The actual sequence of events

33
Notice

• Entry Conditions/Result are pre/post conditions
• Props and Roles are Data Structures
• Track is overloading
• Scene is an algorithm

34
Example

• John went to a restaurant last night. He ordered
  penne arrabiata. When he paid, he noticed he was
  running out of money. He hurried home, since it
  had started to rain.
• Question Did he eat?

35
In Action

• Activate Script Restaurant
• Roles
• S Customer
• W Waiter
• Props
• F Food
• Scene
• Entering
• S ptrans s into restaurant
• Ordering
• Eating
• S ingest F
• Exiting
• S atrans money to W
• Result Answer to question is yes

36
Frames

• Associated with Marvin Minsky
• Semantic nets informally represent
• inheritance through isa links
• Relationships among entities
• Frames
• More structured semantic net
• Assign structure to nodes as well as links
• Definition
• A frame is a collection of attributes (slots) and
  associated values (along with constraints) that
  describe something in the world
• Each frame
• Represents a set of items (isa) with given
  properties that are inherited by its members
• Represents an instance (inst) of a class of items
  with given properties, some of which are inherited

37
Example Semantic Net
person
right
ht
Male
5-10
6-1
ML baseball player
.253
isa
.262
pitcher
outfielder
.106
inst
Koufax
Mays
Dodgers
Giants
38
Transformed to a Frame

• Issue
• Some attributes are to be inherited
• Some refer only to the frame itself
• Person has both cardinality (8,000,000,000) and
  locomotion (biped)
• Only locomotion is to be inheritedindicate with
  an

39

• Person
• isa Mammal
• Card 8,000,000,000
• handed right
• We have a frame with three slots
• Male
• Isa Person
• Card 4,000,000,000
• height 5-10

40

• Baseball Player
• Isa male
• Card 624
• Height 6-1
• avg .252
• team
• uniform color

41
Slots

• Have inherited default values
• Can be structured objects
• Frames to which it can be attached (avg makes
  sense for baseball player but not for water fowl)
• Constraints on values (0 lt avg lt 1)
• Default value
• Rules for computing a value separate from
  inheritance
• Whether a slot is single or multi-valued