8/13/2019 Lecture 5 Knowledge Acquisition 2
1/30
8/13/2019 Lecture 5 Knowledge Acquisition 2
2/30
2
Recap
Stages of Knowledge Acquisition Identification:
Conceptualization:
Formalization:
Implementation:
Testing:
Sources of Knowledge
Written sources
Experts Observation
8/13/2019 Lecture 5 Knowledge Acquisition 2
3/30
3
Outline
Knowledge acquisition methods TOP-Down Methods
Bottom-Up Methods
Formal Techniques Repertory Grids
Card Sort
8/13/2019 Lecture 5 Knowledge Acquisition 2
4/30
4
Knowledge Acquisition
Knowledge acquisition methods Top down (Deductive)
Starting from the general and overall concepts andgradually leading the expert to elicit details of a
topic. Bottom up (Inductive)
The KE focuses the experts attention on specific case. Thishelps the expert abstract the decision for resolving a specificcase to a more generalized rule or concept.
8/13/2019 Lecture 5 Knowledge Acquisition 2
5/30
5
Categories of KA methods
KA methods
Bottom upmethods
Top downmethods
Example basedmethods
Protocolanalysis
Observation
Groupingexamples
Walkthrough
examples
Quantitativeanalysis ofexamples
Statisticalmethods
Inductivemethods
Questioningmethods
Objectorientedmethods
QuantitativeMethods
Inventivemethods
Unstructured
Interviews
Structured
interviews
Questionnaire
Methods formeasuring
relationships
Methods formeasuringuncertainty
8/13/2019 Lecture 5 Knowledge Acquisition 2
6/30
6
Knowledge Acquisition Methods
Object Oriented Methods Knowledge engineer focuses the interview sessions on
discovering the objects within the domain.
This is by asking the expert to group the actualobjects in the field or domain in order to form a class
of objects that have a common set of attributes. i.e. KE asks expert to identify some classes of objects
that have distinguishing attributes.
Example For a loan applicant, one of the major objects in the
mortgage loans is the applicant.
The expert forms the class of applicants object and identifiesthe relevant attributes of this class e.g. Name, address, etc
8/13/2019 Lecture 5 Knowledge Acquisition 2
7/30
7
Object Oriented Methods
Further knowledge elicitation leads to thesubclass of individual applicantscommercialapplicants, etc
As an aid to discovering the classes of objects
and their attributes and subclasses some of theobject oriented methods use network graphs.
Each node represents an object with the attachedattributes.
Arcs of the network show the type of relationshipsamong the objects
8/13/2019 Lecture 5 Knowledge Acquisition 2
8/30
8
Object Oriented Methods
Applicant
Property Loan
Attribute
Address
ValueLocation
Amount
Duration
Mortgage is
Has aHas a
Has an
Has an
Has an
Has a
8/13/2019 Lecture 5 Knowledge Acquisition 2
9/30
9
Object Oriented Methods
Another way to depict the structure of objects is thehierarchy of the class of objects.
The top of the hierarchy shows he most general class ofobjects.
The lower levels contain increasingly more specificsubclasses of objects which may inherit some or allattributes of their immediate parent object.
8/13/2019 Lecture 5 Knowledge Acquisition 2
10/30
10
Object Oriented Methods
ApplicantClass
IndividualSubclass
CommercialSubclass
MarriedSubclass
SingleApplicantSubclass
Age
Name
Address
Type
Manager
Income
Spouses name
Family income
Job
8/13/2019 Lecture 5 Knowledge Acquisition 2
11/30
11
Quantitative Methods
Algorithmic / mathematical / statistical methods
Relationshipstrength
Uncertaintyhow sure
Are methods developed in cognitive science and decision analysis for
eliciting the degree of a decision makers preferences and utilities and
in grouping various objects and attributes.
These methods determine
The extent of relationships among objects (or concepts)
The degree of uncertainty about the domain knowledge.
The expert is asked to compare and express the extent of the
objects relationships on a numerical scale. Then a mathematicalalgorithm is used to compute and rank the degree of relationshipsamong the objects.
8/13/2019 Lecture 5 Knowledge Acquisition 2
12/30
12
Quantitative Methods
Elicitation of uncertainty is one of the areas ofknowledge-based systems that has not been fullydeveloped.
The elicitation of uncertainty depends on the selectedmethod of uncertainty representation
8/13/2019 Lecture 5 Knowledge Acquisition 2
13/30
13
Inventive Methods
Expert is allowed a more active role in the elicitationprocess. Roles include
Expert as a teacher
Expert responsible for teaching and transferring knowledge to theKE. Expert has responsibility of preparation and organisation of
the elicitation process. This method is efficient at the early stages of elicitation
Expert as a partner in a systematic innovation
This revolves around the expert and KE identifying pieces ofknowledge that are in contradiction and to discover solutionmethods for removing the contradiction
Expert as a Knowledge Engineer
Used in cases where the expert may have both technical interestin the system and the needed training in knowledge engineering.
8/13/2019 Lecture 5 Knowledge Acquisition 2
14/30
14
2. Bottom Up Methods
The KE focuses the experts attention on specific case. This helps
the expert abstract the decision for resolving a specific case to amore generalized rule or concept.
Example methods include
Example based methods
Protocol analysis
Observation of the experts decision making process
8/13/2019 Lecture 5 Knowledge Acquisition 2
15/30
15
Bottom Up Methods
Example based methods
This method constitutes the foundation of case based learningand learning by analogy.
The expert and KE work on a number of representative cases orexamples in one of the following ways.
Grouping examples
KE asks the expert to list samples based on their similarities anddifferences. KE then asks the expert to identify the common anddifferentiating attributes of the examples. This helps to determinecategories of examples and the development of general rules fore eachcategory.
Walk through methods
the KE selects a number of cases previously decided by the expert andasks the expert to walk through the decision process. The KE and expertrecognize the contributing factors and attributes and their role in thedecision
8/13/2019 Lecture 5 Knowledge Acquisition 2
16/30
16
Example based methods
Quantitative analysis of examples Statistical methodsThe examples must be from a random sample of
cases decided by the experts. The data on the examples is fed into astatistical technique such as regression analysis in order to discover theexperts decision criteria
Inductive methodsthe example set contains a representative set of allpossible cases the expert has encountered. The examples are fed into
an inductive method which produces a decision tree or a set of decisionrules.
Quantitative techniques are tools for helping the expert discoverthe relations among various attributes of the decision cases.
The outcomes should not be used without consultation with the
expert. Quantitative methods may not be able to discover all the
qualitative aspects of the decision process
8/13/2019 Lecture 5 Knowledge Acquisition 2
17/30
17
Protocol Analysis
Expert is asked to think aloud and verbalize their thought
process or thinking process while solving a set of actualproblems and making decisions.
The KE recodes the process and later analyzes the largevolume of information produced from this method to
discover the general rules the expert uses in solvingproblems
Useful for non-procedural type of problem solving wherethe expert applies a great deal of mental creative and
intellectual effort to arrive at a decision in each case.
8/13/2019 Lecture 5 Knowledge Acquisition 2
18/30
18
Observation
Involves observing the expert while solving a problem
Useful when solution to problem is procedural and takesplace in a sequence of steps through time.
The absence of bias and intrusion inherent in the KEquestions makes this approach more useful.
8/13/2019 Lecture 5 Knowledge Acquisition 2
19/30
19
Formal Techniques
Most of the formal techniques currently in use forknowledge acquisition have their origin in Kellystheory of the psychology of personal constructs.
The basic idea behind it is that people perceive
the world in terms of their own constructs A construct is a specialized form of
conceptualization
Two techniques Repertory Grids
Card Sort
8/13/2019 Lecture 5 Knowledge Acquisition 2
20/30
20
Summary
Knowledge acquisition methods
TOP-Down Methods
Bottom-Up Methods
Formal Techniques
Repertory Grids Card Sort
Next lecture
Repertory Grids
Card Sort
8/13/2019 Lecture 5 Knowledge Acquisition 2
21/30
21
8/13/2019 Lecture 5 Knowledge Acquisition 2
22/30
22
Recap
Knowledge acquisition methods TOP-Down Methods
Bottom-Up Methods
Formal Techniques Repertory Grids
Card Sort
8/13/2019 Lecture 5 Knowledge Acquisition 2
23/30
23
Outline
Formal Techniques of KnowledgeAcquisition
Repertory Grids
Card Sort
Inference and Knowledge Processing
Reasoning
Inference
Inference Methods
8/13/2019 Lecture 5 Knowledge Acquisition 2
24/30
24
Formal Techniques
Most of the formal techniques currently ins usefor knowledge acquisition have their origin inKellys theory of the psychology of personalconstructs.
The basic idea behind it is that people perceivethe world in terms of their own constructs
A construct is a specialized form ofconceptualization
Two techniques
Repertory Grids
Card Sort
8/13/2019 Lecture 5 Knowledge Acquisition 2
25/30
25
Repertory Grids
Uses a two-dimensional matrix to display a
picture of the relationships between variousobjects and concepts from the problem domain.
Along one axis are placed a list of Elements
Objects, people, or situations familiar to the individual. The other axis consists of a set of elicited
ConstructsProperty under investigation
C1 C2 C3 C4
E1 xxxxx
E2 xxxxx
E3 xxxxx
E4 xxxxx
LightHeavy
LargeSmall
ShortTall
8/13/2019 Lecture 5 Knowledge Acquisition 2
26/30
26
Repertory Grid: Constructs
Generally elicited by means of a repertory test.
This consists of presenting three randomlychosen elements to the expert and asking inwhat way two of them are similar. The response forms one pole of a construct.
The opposite of that pole characterizes the thirdelement.
This would then form the other pole of the construct.
This process is repeated for different groups ofthree elements until the options are exhausted.
Each element is rated on each of the constructs. This is on a scale with an odd number of points.
8/13/2019 Lecture 5 Knowledge Acquisition 2
27/30
27
Example on Repertory Grids
Example
Selection of a computer language for a certainsituation
Solution
Identification of the important objects in the domainof expertise Objects: The computer languages, LISP, C++, COBOL,
PROLOG
Identification of the important attributes that are
considered in making decisions in the domain e.g. Availability of commercial packages
Ease of programming
Training time
Orientation
8/13/2019 Lecture 5 Knowledge Acquisition 2
28/30
28
Example on Repertory Grids
The KE picks any three objects and asks the DE to
distinguish attributes and traits that distinguish anytwo from the third.
E.g. if the set includes LISP, PROLOG, and COBOL, the expertmay point to orientationi.e. LISP and PROLOG are symbolicwhile COBOL is numeric.
This is as shown in the table
For each attribute, the DE is asked to establish a bipolar scale(13 or 15).
Attribute Symbol Trait Opposite
Orientation
Ease of programming
Training timeAvailability
C1
C2
C3C4
Symbolic
High
LowWidely available
Numeric
Low
HighNot available
8/13/2019 Lecture 5 Knowledge Acquisition 2
29/30
29
Example on Repertory Grids
This step continues for several triplets of objects.
The answers are recorded in a grid as shown belowwith the numbers in the grid designating pointsassigned to each attribute for each object. Once the grid is completed, the expert may change the
ratings in the boxes.
From the table, in a simplistic sense, if a numeric orientation is veryimportant, then COBOL will be the recommended language.
Choice of Programming Language
C1 C2 C3 C4
E1 LISP 3 3 1 1
E2 Prolog 3 2 2 2
E3 C++ 2 2 2 3E4 COBO
L1 2 1 3
8/13/2019 Lecture 5 Knowledge Acquisition 2
30/30
30
Card Sort or Concept Sorting
When the major concepts have been isolated
from the interview transcript, they are eachwritten on a separate card, and given to the DEin a totally random order. To understand the associations between elements,
then the DE could be asked to group the cards intopiles, according to a criterion of his or her choice.
The process can then be repeated until the expert isunable to provide any more dimensions.
Technique can also be used to obtain a binary tree. The DE sorts the cards into two piles and then subdivides
each pile in turn until no other divisions are possible.
The process can be done in reverse, whereby the expert hasto form as many piles as possible and then determine reasonsas to why piles should be consolidated.
Top Related