Knowledge Acquisition in Expert System
Knowledge Acquisition in Expert System
INTRODUCTION
Expert systems are computer programs that use domain-specific knowledge to imitate the
reasoning process of human experts. It was not until the late 1970s that AI scientists
realized that the problem-solving power of a computer program mainly derives from the
knowledge it possesses rather than the inference mechanism it employs.
The process of assimilating the expertise of several experts into an expert system is not
easy, particularly when these experts are trained in different disciplines. The differences
not only appear in problem-solving strategies taken by each expert, but also appear in
what heuristic is applied to solve the problem. Furthermore, the difficulty arises because
of the communications barriers among experts and between experts and the knowledge
engineer(s). How to facilitate the knowledge acquisition process involving multiple
experts becomes a major challenge to knowledge engineers.
There are three primary concerns of the knowledge acquisition task: the involvement of
appropriate human resources; the employment of proper techniques to elicit knowledge;
and a structured approach to performing the knowledge acquisition task
PEOPLE ISSUES
Identifying appropriate domain experts and involving proper people in the knowledge
acquisition process is critical to the success of knowledge acquisition. Those who are
involved in the knowledge acquisition process include: (1) domain experts who have had
years of experience working in the application domain; (2) knowledge engineers who
possess technical skills in eliciting knowledge, representing knowledge, and
implementing expert systems; and (3) users and managers.
1
SELECTING DOMAIN EXPERTS
1. Domain expertise, experience, and reputation -- The experts chosen should have
expertise and experience in the specific aspect of the domain for the target expert
system. Practicing experts, who are currently active in domain tasks, are the
domain experts who should be identified. Leading experts are recognized by their
colleagues and clients. The reputations of the experts are sometimes the major
determinant of the credibility of a deployed expert system. It is also helpful to
select experts with some background or interests in AI and expert systems.
2. Personal characteristics and attitudes -- Domain experts not only should have
skills in the domain but also should have skills in communicating their
knowledge, judgment, and experience and the methods they use to apply these to
a particular task. Other desirable attributes include a sense of humor, being a good
listener, a sense of commitment, cooperativeness, patience, being easy to work
with, persistence, and honesty.
3. Availability -- The most common problems in knowledge acquisition are caused
by time demands on an "already-busy" expert. Therefore, management's
commitment of experts' time to a project should be secured. If problems with
access are anticipated, an individual should not be selected as the primary expert
regardless of his/her experience and characteristics.
An expert system development project may utilize one expert or a number of them as the
primary source of domain expertise. The use of a single expert is sometimes preferred,
but there are many cases in which it is useful or necessary to utilize multiple experts.
There are four primary problems with knowledge acquisition from a single expert: (1)
difficulty in allocating adequate time by a key individual in the organization, that may
create a bottleneck in the expert system development process; (2) personal bias, that may
affect the performance of the system; (3) limitation to a single line of reasoning, that may
affect the usefulness of a system, since expert systems that are developed based on a
single, perhaps narrow, line of reasoning do not emulate most real-life decision making;
and (4) incomplete domain expertise, that may affect the performance of a system. The
first problem may create a bottleneck in the expert system development process. If only
one expert is the sole knowledge source for the expert system, it is very difficult to
acquire a sufficient amount of his or her time. This is not because of an unwillingness to
cooperate with the knowledge engineer, but because of other commitments and the fact
that knowledge acquisition is so time-consuming. Depending upon the availability of
such an individual can delay a project's progress. Personal bias, which constitutes the
second problem, may affect the performance of the system. The third problem affects the
usefulness of a system, since expert systems that are developed based on a single, perhaps
2
narrow, line of reasoning do not emulate most real-life decision making. Furthermore,
with attempts to apply expert systems technology to more sophisticated problems,
restricting the knowledge acquisition activity to a single expert could result in
incompleteness of the system.
Experts by definition are "very knowledgeable about only a small subset of the tasks in
the domain." As expert systems become more and more complicated and the domains
become more complex, the required expertise is most likely to reside in not one but a
team of experts. A designated primary expert is needed when multiple experts are
utilized.
Incorporating the expertise of a team of experts provides the following positive effects on
the resulting expert system: (1) it assures that the knowledge base can be complete; (2) it
improves the likelihood of obtaining specialized knowledge in sub domains of the
problem; (3) it increases the quality (i.e., reliability and consensus among experts) of the
acquired knowledge; (4) it assures that the facts that are included in the knowledge base
are important ones; (5) it enhances understanding of the domain knowledge through
discussion, debate, and exchange of hypotheses between members of the expert team; and
(6) it encourages interactions among experts and creates a synergy such that the acquired
group knowledge is greater than the sum of the individual's knowledge.
When multiple experts are involved in the development process, knowledge acquisition
sessions can be flexibly designed. First, it is not necessary that all experts be present at
the same time in one place throughout the acquisition process. Advanced information
technology such as group support systems can be used to facilitate interactions with
experts, not only in face-to-face meetings but also in meetings conducted in dispersed
locations. Second, interactions among experts can create a synergy such that the acquired
group knowledge is greater than the sum of the individual's knowledge. However, the
involvement of multiple experts increases the complexity of the knowledge acquisition
process and makes an already difficult task even more complicated. Two factors
contribute to this complexity. One factor involves the difficulty of merging each
individual expert's knowledge structures into one group knowledge structure that
provides the underlying problem-solving expertise of the expert system. This involves
resolving conflicts between various problem-solving approaches and reasoning processes.
The other factor involves the difficulty of the generation of group knowledge that does
not reside in any one individual expert but evolves as a result of the group interaction. A
group support system has been proven to be useful and effective to ease these difficulties
when acquiring knowledge from multiple experts.
3
managers, preparing technical documentation, verifying and validating the system,
training users, operating and maintaining the initial system, and providing advice for
further extension and updates. Knowledge engineers should have some mastery of the
domain to identify the type of knowledge that is required. They should be able to
conceptualize and analyze the problem, to employ various knowledge acquisition
techniques, and to communicate and work with other people.
Users and managers should be involved in the knowledge acquisition task, especially in
the planning stage where decisions of problem scope are made and in the verification
stage where the prototype system is evaluated. The problem scope should be jointly
determined by a group of people that includes domain experts, users, managers, and
knowledge engineers. Domain experts provide a subjective perspective of what the
system can do for users, while users provide a broader perspective of how the system
may be able to help them. Managers provide strategic guidelines as to what should and
should not be included in the system. Knowledge engineers can assess technical
feasibility. One individual, such as the manager, can of course decide the scope of the
system based on policies and the needs of the organization. However, systems that
involve users in the development process tend to be more acceptable to the users when
they are delivered.
The approach used for knowledge acquisition determines both the quality of knowledge
and the amount of effort required for its acquisition, so the technique selected greatly
affects the performance of the expert system and the resources required for its
development. The growing recognition of the importance of knowledge acquisition has
resulted in the development of various techniques, methodologies, and tools for
automated knowledge acquisition. This section reviews techniques used in psychology
and social sciences for revealing expert knowledge structures and processes. Techniques
that involve the concept of machine learning (e.g., induction), neural networks,
simulation, web-based knowledge acquisition, and automated knowledge acquisition
tools designed for specific applications are not discussed in this chapter.
Many knowledge acquisition techniques and tools have been developed. Each technique
has its strengths and limitations. How well a knowledge engineer can utilize them will
depend on his/her selection of appropriate techniques and tools, which in turn determine
the quality of knowledge acquired, the amount of effort needed, and the skills required.
Interviewing is the most commonly used method in requirements elicitation for analysis
and design of information systems. It is also widely used in eliciting knowledge from
domain experts for expert systems development. In general, there are two forms of
interviewing. The basic form involves question-answer sessions between the knowledge
engineer and the expert. These free-form or unstructured interviews are usually started by
the engineer's asking "How do you solve this problem?" Follow-up questions usually
reflect further explanation or clarification of some points that the expert has made. The
4
process is fairly unstructured. A distinct advantage of free-form interviews is that
knowledge engineers can elicit unanticipated information.
However, there are difficulties with this technique. First, as people become more
experienced at performing certain tasks, they become less aware of the cognitive
processes involved in their performance. They cannot explicitly describe their reasoning
process step by step. Second, there are certain biases and fallibilities in human reasoning.
When reasoning about the entire sequence, people tend to anchor on items that occur
early in a sequence. People see what they expect to see. When asked to describe their
reasoning process and problem-solving methods, experts tend to provide reconstructed
versions of their reasoning and omit some components that may be important to solve the
problem because they assume them to be obvious and explicit. Moreover, experts may
get tired and become bored with repeating what to them seems to be obvious information.
People do not talk in complete sentences. Speech is marked by phrases, asides, "ers," etc.
Neither the expert nor the knowledge engineer knows at the time which parts of the
dialogues are important. Consequently, all details of the interviews must be recorded,
transcribed, and analyzed. This makes knowledge acquisition a tedious and time-
consuming process.
5
Protocol analysis, usually referred to as "thinking aloud," is a form of data analysis that
has its origin in clinical psychology. When employing this technique, a knowledge
engineer describes a problem scenario and asks an expert to talk about his or her thinking
process while solving the problem. Experts find it much easier to talk about specific
examples of problems than to talk in abstract terms. The "thinking aloud" process is
videotaped and analysis of it is based on transcripts. Once the transcripts are produced,
protocols must be analyzed based on a systematic breakdown of the information to
produce a structured model of the expert's knowledge. The goal is to identify the kinds of
objects that the expert sees, the attributes of those objects, the relationships among those
objects, and the kind of inferences drawn from these relationships. The advantage of this
technique is that the transcripts describe specific actions and rationales as the expert
thinks through and talks about the decision-making process. There is no delay between
the act of thinking of something and reporting it. Protocol analysis is not appropriate for
all kinds of tasks
Repertory grid analysis, which had its origin in Kelly's personal construct theory
(1955), aims at gaining insights into the expert's mental model of the problem domain. It
involves an initial interview with an expert, a rating session, and analyses that cluster
both the objects and the traits on which the items were rated. In the initial interview, the
expert is asked to identify some objects in the domain of expertise. After a set of objects
has been identified, the expert is asked to compare three of these objects at a time, in each
case naming a trait that two of the objects possess but the other does not. The expert is
then asked to identify an opposite of that trait. The expert further provides a scale to rate
the importance of the traits. The same process is repeated until all the objects have been
compared and traits to differentiate them identified. In the rating session, these objects
are rated according to the traits identified and scales assigned. At any stage the expert can
add more objects or traits or alter entries in the grid. In this way the process heightens his
or her awareness of how he or she views the problem. Once the rating grid has been
established, a computer program can be used to cluster the objects and cluster the traits.
This technique is useful in extracting subjective data, but there are a number of
difficulties associated with its use. First, unless the number of objects in the problem
domain is small, an enormous number of comparisons need to be made. Second, it is not
always easy to identify traits that differentiate objects. It sometimes takes a long time to
make just one comparison and, as a result, the process becomes very time-consuming.
As the size and complexity of expert systems increase, many experts and several
knowledge engineers are needed in the knowledge acquisition process of an expert
system project. Techniques described above may be employed in collaborative
knowledge acquisition by interviewing (or observing) a team of experts one at a time.
The "serial" knowledge acquisition process causes the following problems. (1) It is a
lengthy and time-consuming process. (2) Problem-solving strategies and the knowledge
used by experts may conflict with each other. Knowledge engineers should not assume
that they have sufficient knowledge in the domain to try to solve these conflicts, which
6
must be resolved by the individual experts through several iterations. (3) Experts do not
have chances to stimulate each other's thinking to surface the underlying knowledge they
use. (4) The integration of problem-solving expertise acquired from individual experts
into a knowledge base is still needed even if there is no conflict among experts' opinions.
The missing links of various expertise in the domain have to be acquired from the experts
themselves.
Brainstorming is a group method for developing ideas and exploring their meaning. It
promotes the identification of a number of considerations related to a problem domain. It
is designed to stimulate thinking and generate ideas in such a way that each individual
may develop his or her train of thought or expand upon other individuals' thinking.
Interactions in the form of exchanging thoughts are encouraged to enrich the outcome of
the group process. Brainstorming also can be used to help experts and knowledge
engineers discover areas that require special attention in the problem-solving process.
With multiple experts working as a team, brainstorming can help prevent immediate
confrontation, reduce inhibited behavior, and resolve conflicting views.
An electronic form of the brainstorming technique has been used in group sessions to
define a problem scope, to identify possible solutions, and to develop a heuristic
classification scheme. This technique provides a useful tool for acquiring knowledge
from multiple experts. Conflicts can be identified and resolved during the brainstorming
session and interactions among experts create a synergy of expertise.
Nominal Group Technique (NGT) is a method for structuring small group meetings
that allows individual judgments to be pooled effectively and used in situations where
there is uncertainty or disagreement about the nature of a problem and possible solutions.
This technique is helpful in identifying problems, exploring solutions, and establishing
priorities. It typically includes four steps: (1) silent generation of ideas in writing; (2)
round-robin recording of ideas; (3) serial discussion of the list of ideas; and (4) voting.
This problem-solving procedure reduces negative effects (i.e., nonparticipation, conflicts)
that may be triggered by face-to-face interaction among team members. In a collaborative
environment, such as a GDSS, the Idea Organization tool facilitates the first three steps of
the NGT. A voting tool completes the fourth step. Knowledge engineers have found that
a combination of group interaction and the Nominal Group Technique enhances the
creativity and quality of the resulting solutions.
7
The Delphi technique uses a series of questionnaires to aggregate the knowledge,
judgments, or opinions of experts (usually anonymous) to address complex problems.
Individual contributions are shared with the whole group by using the results from each
questionnaire to construct the questionnaire for the next round. One of Delphi's principal
uses has been to make future projections and forecasts. It can also be used to identify
goals and objectives, generate possible alternatives, establish priorities, reveal group
values, gather information, and educate a respondent group. By keeping individuals
separated and maintaining their anonymity, the Delphi technique reduces the influence of
potentially dominant people, prevents the undue influence of the personalities of certain
individuals and allows strangers to communicate effectively. It also allows for the
participation of more people than could interact effectively in a group and prevents
unproductive disagreements. Delphi is useful whenever it is desirable to have pooled
judgment and when experts are geographically dispersed. This technique has been used to
facilitate an expert system development project for target industry analysis.
This technique can be used in conjunction with the structured interviewing technique to
force organization of the communications that take place between a knowledge engineer
and experts. The structure provided by goals reduces the interpretation problems inherent
in unstructured interviews and allows the knowledge engineer to prevent the distortion
caused by domain expert subjectivity. It is more effective because it "forces the domain
expert to be systematic in attending to interview tasks." However, the moderator can
introduce biases in the interview by shifting topics too rapidly, verbally or nonverbally
encouraging certain answers, failing to cover certain areas, and so forth. In addition, the
interviews are usually taped and transcribed for further analysis, which can be very time-
consuming.
8
The voting technique emphasizes on finding a compromised solution to a problem. It
involves presenting a problem to domain experts and encouraging each member to vote
on alternative solutions to the problem. Alternatives are ranked and rated by the group of
experts. It is effective only if each expert who participates in the team feels that his or her
views and opinions have been heard. It is also vital that each expert has a commitment to
the group decision though he or she may have some reservations. Even when one best
answer may not be agreed upon by the team, this technique can significantly contribute to
knowledge-based systems development efforts.
It can be used in defining problem scope, identifying alternate solutions, and soliciting
proper solutions. For the technique to be effective, the knowledge engineer must be
aware of the possible effects of status, rank, or experience differences among the domain
experts.
Group Repertory Grid Analysis is a group version of the repertory grid analysis
technique. It attempts to represent the domain expert's problem-solving knowledge in
repertory grids, which allow experts to rate or judge a solution according to its level of a
problem-solving trait. A domain expert uses repertory grids to enter knowledge by means
of a rating grid. This grid displays problem solutions that have been elicited from the
domain expert, which serve as column headings within the grid. Constructs (e.g., solution
traits) are placed beside the grid's rows. The system elicits constructs by presenting the
domain expert with sets of solutions and requesting that the expert discriminate among
them. The domain expert then provides each problem solution with a rating that
represents how it relates to each trait. Once these initial grids have been constructed, the
knowledge engineer analyzes them and refines the knowledge base. This technique can
be used in conjunction with the brainstorming technique, using brainstorming to generate
solutions and constructs and then using a spreadsheet program to implement the repertory
grid. A brainstorming tool can be used to facilitate the interviewing session. A Group
Matrix tool can be used to facilitate the rating session.
Group Support Systems (GSS) are computer and communications systems that facilitate
a group's communication, coordination, and decision-making process. Techniques such
as repertory grid analysis, brainstorming, Nominal Group Technique, Delphi technique,
and the voting technique have been implemented in computer programs to facilitate the
acquisition of knowledge from multiple experts using a group support system. The six
necessary components include: hardware, software, facility, people, procedures, and
facilitation.
Benefits derived from using a GSS for knowledge acquisition include: (1) electronic
documentation of knowledge; (2) knowledge extraction can be done in parallel from
multiple experts; (3) conflicts are addressed during knowledge extraction sessions; and
(4) interactions among experts result in an enlarged and enriched domain of expertise.
9
KNOWLEDGE ACQUISITION METHODOLOGY
McGraw and Harbison-Briggs (1989) pointed out that the lack of structure and
organization of knowledge acquisition is one of the barriers to effective knowledge
acquisition. It therefore is essential to establish detailed plans for conducting knowledge
acquisition sessions. A knowledge acquisition methodology serves as a guide to
knowledge engineers in developing these plans. A methodology established to perform
the knowledge acquisition task is comprised of four phases: planning for knowledge
acquisition, knowledge extraction, knowledge analysis, and knowledge verification.
Table 1 summarizes the steps in each phase.
Phase Steps
Plan KA sessions
Conduct KA sessions
Debrief experts
Knowledge Acquisition Planning. The goal of planning, the most important phase of the
knowledge acquisition task, is to understand the problem domain, define the problem
scope, and identify experts in the problem domain, analyze various knowledge
10
acquisition techniques, and design proper procedures to acquire knowledge. By
discussing with experts, the knowledge engineer can identify the characteristics of the
problem and comprehend the kind of knowledge required. Knowledge engineers develop
proper group process models to acquire knowledge. Existing group support tools are
analyzed so that they can be properly employed in the group process. Major steps in this
phase include understanding the domain, identifying experts, defining the problem scope,
identifying the type of application, analyzing characteristics of the problem, analyzing
characteristics of tools and techniques, developing process models, and planning for
knowledge acquisition sessions.
Knowledge Analysis. The primary task of this phase is to analyze outputs from
knowledge extraction sessions. Some of the tools may be used to support analysis
activities such as identifying a list of key concepts from the brainstorming results.
Heuristic, concepts or classification structures are analyzed and formalized into
representations that may be in the form of heuristic rules, frames, objects and relations,
semantic networks, and classification schemes. These representations are then
transformed into specific representation schemes that are supported by an expert system
building tool. This transformation is part of an effort to implement a prototype expert
system. This prototype system will be used in the knowledge verification phase.
11
knowledge acquisition and review process, expert system developers can minimize the
time and money wasted and prevents the development of an inappropriate or useless
system.
Knowledge acquisition has been described as a very difficult and time-consuming task
that frequently creates a bottleneck in an expert system development effort. There is no
single way to avoid the difficulties of knowledge acquisition. However, by identifying the
right domain experts, and employing a combination of proper techniques and a structured
methodology, we believe that the knowledge acquisition task can be performed more
easily as well as more effectively and efficiently than it used to be.
As expert systems become more and more complicated and the problem domains become
more complex, the required expertise often resides in not one but a group of experts.
Moreover, interaction among experts creates a synergy that results in an enriched and
enlarged domain of expertise. These factors all point in the direction of increased use of
multiple experts in the expert system development project whenever possible. There is no
single knowledge acquisition technique that is the best and most effective; the trend is
toward using a combination of techniques that best fit the domain of expertise.
The work of automating the knowledge acquisition process is in progress, but most
existing tools have been designed for eliciting knowledge from individuals. The ETS and
AQUINAS systems developed by Boose (1989) try to address situations where multiple
experts are involved, but they are limited to certain applications. It should be noted that
most tools are still in the research and development stage and are available only for in-
house use. Nevertheless, the progress in automation of knowledge acquisition is on-
going. To be most useful, the focus of this research should be on acquiring the kinds of
knowledge that are difficult to acquire manually but for which automated methods are
feasible. Research in neural networks and case-based reasoning for knowledge
acquisition is another important trend in knowledge acquisition.
The term "knowledge acquisition" usually refers to the acquisition of knowledge for
building expert systems. However, acquiring knowledge from a group of people is a
ubiquitous activity that can be found in many group tasks such as strategic planning,
system design, negotiation, and decision making. For instance, major software design
decisions are usually made in face-to-face software design meetings involving users,
managers, and system developers. Users have knowledge of how current systems work
12
and requirements of the new systems. Managers know the strategic implications of the
new system. System developers can intrigue users and managers by providing them with
knowledge of how advanced information technologies can serve them. We may examine
many other group works from the knowledge acquisition perspective.
13