STRUCTURED OBJECTS Knowledge Representation

1
STRUCTURED OBJECTS -Knowledge Representation
2
STRUCTURED OBJECTS -Knowledge Representation

• Two types of structured objects common in AI
• semantic networks
• frames
• These knowledge representation schemes were
  developed primarily from efforts to build systems
  that understand natural language.

3
Semantic Networks

• Quillian conducted original work.
• Based on psychological theory.
• Popular because of the graphical nature of the
  representation, people find it easy to
  understand.

4
Semantic Net Representation - Example 1
operation
hole
is_a
machined_by
is_a
drilling
feature
5
Semantic Networks

• Use graphs (the term is used in its mathematical
  sense) to describe the domain in terms of
• objects
• indicated with nodes
• represents facts
• binary relationships
• indicated with lines (arcs) that link the nodes
• represents the relationships between objects

6
Overview of Semantic Nets

• The term "semantic" net implies that we are
  interested in the meaning of the things we are
  representing.
• Part of the power comes from objects connecting
  with a "is-a" relationship.
• Through is-a connections, a particular instance
  of an object "inherits" the characteristics of
  the more general object.

7
Semantic Net - Example 2
is_a
Joe
Boy
is_a
goes to
Human Being
has_a_child
School
is_a
Woman
Kay
is_a
has_a_child
is_a
needs
is_married_to
Man
Car
Food
is_a
is_a
Sam
Mercedes Benz
is_a
is_a
VP
color
works_for
plays
ACME
Silver
made_in
Golf
is_a
subsidary_of
Germany
Sport
AJAX
8
Example - is_a connections

• Kay is-a woman, and a woman is-a human being, and
  human beings need food.
• Through the network, we can infer that Kay need
  foods, without an explicit arc from Kay to food.
• Semantic networks allow "implicit" relationships.

9
Semantic Network -Example 3
John
donor
age
blue
7
giving-1
recipient
color
object
Mary
age
book-1
6
object
selling-1
age
Jane
buyer
10
Semantic Net - Example 3

• Illustrations representation of
  entity-attribute-values triples.
• Arcs - correspond to the attributes in the triple
• Nodes - correspond to the entities and values
• To represent general concepts
• EAV triples (in production rules), used variables
  to represent general concepts.
• In semantic networks, node represents the general
  case and link the nodes.

11
Important Differences Semantic Nets and Rules

• EAVs can use variables to represent general terms
• Thus, can represent general information with a
  production rule
• IF (X is-a bicycle) THEN (X no-of-wheels 2)
• Nets have neither variables or rules
• nodes represent general terms
• links express generally applicable information as
  well as facts
• reduced need for rules for generalizations

12
Semantic Net - Example 4
Mode of transport
is_a
is_a
is_a
car
bicycle
ship
part_of
part_of
part_of
wheels
part_of
wheels
handle bars
4
engine
brakes
2
part_of
part_of
brake- cable
grips
13
Semantic Network - Example 4

• Illustrates the ability to build up complicated
  relationships and share information between
  nodes.

14
Inference in Semantic Networks

• Recall production rules
• IF (X is-a bicycle) THEN (X no-wheels 2)
• (giving-1 donor John)
• (giving-1 recipient Mary)
• (giving-1 gift bicycle)
• (bicycle_1 is-a bicycle)
• (giving-1 is-a giving)
• (bicycle-1 color blue)

• If we ask the question (bicycle-1 no-wheels)?
• Instantiate bicycle-1 to X
• Fire the rule
• Conclude bicycle-1 no-wheels is 2

15
Semantic Network - Example 5
blue
John
color
donor
bicycle-1
gift
is_a
giving
giving-1
is_a
bicycle
recipient
wheels
Mary
2
16
Inference in Semantic Networks (continued)

• The same situation as was described with
  production rules and EVA triples is represented
  in semantic network - Example 5.
• Follow the is-a link from bicycle-1 to the
  general object bicycle.
• Follow the no-wheels link from bicycle to 2
• Conclude that bicycle-1 has 2 wheels

17
Inference in Semantic Networks (continued)

• The basic technique is to traverse the links
  between nodes to find particular values for
  things
• In practice, use pointers and lists
• LISP was very good for this, which was part of
  the attraction of Sementic Nets to early AI
  researchers.
• Traverse links in both directions

18
Transitive Links in Semantic Net Inference

• Traverse (follow) as many links as necessary and
  the relationships (conclusions) are still valid
• "Is-a" is transitive

19
Inference in Semantic Networks

• Consider the relationship "near"
• maysville" is near purcell".
• purcell" is near norman".
• norman is near moore.
• moore" is near oklahoma city".
• oklahoma city is near edmond
• Can we conclude that maysville" is near
  edmond"?
• Near" is not transitive.

20
Inference in Semantic Networks

• Not all relationships are transitive
• Different types of relationships may require
  different inference mechanisms
• Therefore, the inference engine for semantic nets
  is more complicated than for production rules
• What we gain by requiring fewer rules is at the
  cost of a more complex inference mechanism

21
Strengths of Semantic Nets

• Offer flexibility in adding new nodes and links
  to a definition as needed.
• The visual representation is easy to understand
• Economical because a node can inherit
  characteristics from other nodes via an "is-a"
  relationship.
• It is similar to how human beings represent
  information
• Since nodes inherit relationships from other
  nodes, it can support the ability to reason and
  create definition statements between non-linked
  entities.

22
Weaknesses of Semantic Nets

• No standards exist for the definition of nodes or
  relationships between and among nodes.
• "Inheritance" of properties from one node to
  another results in difficulties when it become
  necessary to handle exceptions.
• Perception of the situation by the domain expert
  can place relevant facts at inappropriate points
  in the network.
• Procedural knowledge is difficult to represent,
  as sequence and time are not explicitly
  represented.

23
Frames

• Used to represent stereotypical situations.
• Attached to each frame are several kinds of
  information.
• Things like how to use the frame, what should
  happen next, etc. Use slots and fillers.

24
Frame - Example 1
fillers
slots
25
Relationship to Semantic Nets

• Some references indicate that frames developed
  from semantic networks.
• Semantic nets become large, very quickly and can
  be difficult to manage.
• Each frame represents an object.
• Think of a frame as a semantic net, with the
  attributes moved "inside" each node, instead of
  attaching nodes to each attribute.

26
Frame - Example 2a
27
Frame - Example 2b
28
Frame - Example 2c
29
Example 2 Discussion

• With the knowledge that "bicycle-1" is a bicycle,
  we know that it has two wheels and is used as a
  mode of transport.
• We do not save this information with each
  instance of the object bicycle
• Instead, save general information once and
  indicate when particular objects are instances of
  the class of objects, bicycle.

30
Example 2 Discussion (cont.)

• Simpler than semantic networks.
• we do not link every entity to each attribute
• using frames the values for certain attributes
  are included.
• Empty slots (attributes) act has place holders
• if a value particular value is unknown at this
  time, this indicates that the entity should have
  particular attributes
• Empty frames for types of objects can be filled
  in.

31
Inference when using Frames

• As with semantic networks, the inference
  mechanism is implemented with pointers.
• Will not have to traverse as many nodes (fewer
  pointers to follow), because much of the
  information relevant to each object is stored
  within each instance of an object.

32
Inheritance in Frames

• Is-a properties allow the linking of frames in a
  hierarchy
• frames on lower levels inherit the general
  properties of the more abstract entity on the
  higher level

33
Frame - Example 3
34
Example 3b - Child Frames

• Name Compact Car
• Owner check registration
• Color list per manufacturer
• No cylinders
• range 4-6
• if needed ask owner
• Make use frame
• Vintage
• range 1950-2001
• if needed ask owner

• Name Jans car
• Instance of compact car
• Owner Jan
• Color blue
• No cylinders 6
• Make Honda
• Vintage 1992

35
Inheritance - Example 3

• Therefore, anything known about passenger-cars
  applies to compact cars and mid-sized cars, and
• Everything we know about cars applies to
  passenger cars.

36
Positive Features of Inheritance

• A method to organize knowledge
• General term is abstract and has certain
  characteristics.
• The lower level instance is then described using
  only those attributes that distinguish it from
  the more abstract term
• Focus only on the added information.

37
Positive Features of Inheritance

• Very economical method to represent information
• Only represent information once at it's highest
  level.
• General information is then "inherited" by the
  less abstract, lower level frames.

38
Strengths of Frames

• Generic sets of rules can be built from the
  representation with a lessened risk of
  redundancy, and consequently with economy of
  rules
• A visual display of frames aids in the
  understanding the knowledge and the
  interrelations of the structure.
• Provides a working base for adjusting and
  maintaining a knowledge base.

39
Weaknesses of Frames

• Computer languages are not equipped to work with
  entire frames, and encounter efficiency problems
  caused by increasing search time.
• Knowledge engineers and others seem to experience
  difficulty in limiting the size and components of
  the frame so that they match the reasoning
  process or search techniques of the system.

40
Weaknesses of Frames (continued)

• Frames can overlap too much for the sake of
  efficiency. A lack of understanding of where to
  stop often requires restructuring the
  representation, which can be costly.

41
Default values

• Inheritance can be taken further to include
  default value, things that are generally true,
  but perhaps not always true.

42
Summary of Semantic Nets and Frames

• New hybrid techniques are being developed to
  combine frames with rules and with logic
• Use frames to represent the static information
  (frames are good at this)
• Use production rules or logic provide the
  processing and procedural knowledge
• Knowledge engineers typically do not have
  experience with this KR scheme so it has not been
  used that often.

43
Example System That Uses Frames

• INTERNIST/CADUCEUS an expert system to diagnosis
  diseases.
• Unlike MYCIN, INTERNIST is meant to emulate
  "internists" who diagnosis a wide range of
  internal diseases.
• Exhaustive depth-first searches, like MYCIN,
  would not be practical.
• The developer used structured objects and
  abduction (forward reasoning).

44
Example (continued)

• INTERNIST is expected to
• eventually be in clinical use
• diagnosis at an expert level about 85 of
  internal medicine
• has a large knowledge base consisting of over 500
  diseases.