Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Article

whole variety of knowledge sources and targets. A


An analysis of knowledge careful consideration of the meta-view and its associated
role-participant mappings also yields the new ideas of
acquisition roles and the elemental and composite role and the multi-role
entity. These new constructs are then used to identify
participants areas where research is currently needed and to generate
specific research issues. Taken altogether, this view
allows a more flexible consideration of the many possible
combinations that can and frequently do occur in actual
KA situations.

Keywords: knowledge acquisition, expert systems,


William P. Wagner knowledge acquisition roles, knowledge acquisition
participants
Department of Management, Villanova
University, Villanova, PA 19085, USA

Clyde W. Holsapple 1. Introduction


Research in the field of knowledge acquisition (KA) has
Decision Science and Information Systems Area, expanded in recent years to include techniques and situ-
College of Business and Economics, University of ations involving multiple experts (Boose, Bradshaw, Kosz-
Kentucky, Lexington, arek, and Shema, 1993; Grabowski, Massey, and Wallace,
KY 40506–0034, USA 1992; Liou, 1990; Liou, Weber, and Nunamaker, 1990) and
also a wide array of knowledge sources such as text, video
and auditory media (Kaplan and Berry-Rogghe, 1991;
Abstract: Much recent research effort in the field of Slator, 1989; Boose, 1989; McGraw and Harbison-Briggs,
knowledge acquisition (KA) has focussed on extending 1989). Within the KA field, many researchers have been
knowledge acquisition techniques and processes to working at integrating the many different manual and auto-
include a wider array of participants and knowledge mated techniques required when a diverse set of knowledge
sources in a variety of knowledge acquisition scenarios. sources is used in the development of an expert system
As the domain of expert systems applications and (Reinartz and Schmalhofer, 1994; Gil and Paris, 1994;
research has expanded, techniques have been developed Sommer, Morik, Andre and Uszynski, 1994; Barbuceanu,
to acquire and incorporate knowledge from groups of
1993; Morik, 1991; Rappaport and Gaines, 1990; Boose,
experts and from various sources such as text, video, and
1989). But as the set of possible knowledge sources and
audio tapes. However, the dominant participant-role
model remains that of the knowledge engineer eliciting knowledge representations has expanded our ability to sys-
knowledge from one or more human experts. This tematically describe and organize the existing research
conceptual gap has contributed to the major divisions in within the field has not kept pace. Recent knowledge acqui-
the KA field between researchers interested in manual sition research frameworks still focus primarily on the
KA and those developing tools for automated KA. This classic scenario of a single knowledge engineer eliciting
article considers the wide variety of possible KA the knowledge of a single human expert though knowledge
scenarios and presents a meta-view of KA participants acquisition activities include a much broader range of
and the roles they may assume. We suggest that it is possibilities (Holsapple and Wagner, 1996a; Dhaliwal and
more appropriate to think of knowledge acquisition Benbasat, 1990; Liou, 1990; McGraw and Harbison-Briggs,
participants as playing one or more roles. These include
1989; Kim and Courtney, 1988). This conceptual gap in the
knowledge sources, agents and targets for KA processes.
research serves to exaggerate some of the current divisions
We also present a participant model drawn from
research in decision support systems that more between researchers involved in human-centered or
accurately characterizes the diversity of the entities ‘manual’ KA and those focussing solely on automated KA
participating in the KA process. This view is more techniques (Wagner, 1992; Dhaliwal and Benbasat; 1990;
inclusive as it allows us to consider both human-human Fellers, 1987).
and human-computer KA interactions as well as the In order to bridge these apparent divisions, this article

Expert Systems, February 1997, Vol. 14, No. 1 3


proposes a new way of viewing participants in KA pro- Table 1: Types of knowledge extraction methods
cesses that allows for the wide variety of possible sources (Hoffman, 1987)
and scenarios. We suggest that it is more appropriate to
view participants as playing one or more roles in the con- METHOD CATEGORY DESCRIPTION
text of one or more KA processes (Holsapple and Wagner,
1996a). To aid research in this area it is suggested that the METHOD OF Analysis of the tasks that
possible roles include the knowledge source, agent or tar- ‘FAMILIAR’ TASKS the expert usually
get. We also allow that participants can fill one or more performs.
roles and that a role can be filled by one or more parti- STRUCTURED AND The expert is queried with
cipants. The combination of this conception of KA parti- UNSTRUCTURED regard to knowledge of
cipants and their roles leads us to propose role-participant INTERVIEWS facts and procedures.
and participant-role mappings that capture the complex
LIMITED A familiar task is
many-to-many KA scenario where a participant plays sev-
INFORMATION TASKS performed, but the expert
eral roles and roles are filled by multiple participants. This
is not given certain
simple yet powerful view allows us to consider the many
information that is
possible combinations of participants and roles that actually
typically available.
do occur within the context of KA processes.
In the following section, we suggest criteria for evaluat- CONSTRAINED A familiar task is
ing the value of a meta-view and how this should guide PROCESSING TASKS performed but the expert
the construction of this one in particular. Then we present must do so under time or
the meta-view itself; breaking it down into separate sections other constraints.
on knowledge source, target, and agent types of roles along METHOD OF ‘TOUGH Analysis of a familiar task
with a consideration of the resulting role-participant map- CASES’ that is conducted for a set
pings. Finally we show how this taxonomy can be applied of data that presents a
to organizing current research in the field and generating ‘tough case’ for the expert.
many new and interesting research issues.

structed from the perspective of a single, human knowledge


1. 1. The value of classification schemata for knowledge
source. The operant characteristic of the taxonomy is the
acquisition research
expert’s familiarity with the task that is to serve as the
Knowledge acquisition research has produced a variety of vehicle for knowledge elicitation. So, it may only describe
taxonomies, frameworks, and classification schemata a portion of the knowledge extraction phenomenon, which
(Holsapple and Wagner, 1996b; Dhaliwal and Benbasat, is in turn, only a portion of the knowledge acquisition prob-
1990; Boose, 1989; Hoffman, 1987; Clancy, 1986). There lem itself. However, with respect to usefulness, Hoffman’s
are classifications of knowledge types, problem domains, taxonomy is an important contribution. He was one of the
task types, and KA tools and methods. Classification first researchers to recognize the need for conducting
schemata and taxonomies play fundamental roles in the empirical work that would directly compare knowledge
development of a discipline, as a means for organizing acquisition techniques. Approaching the problem from the
phenomena into classes that are amenable to systematic perspective of experimental psychology, he attempted to
investigation and theory development. They offer perspec- develop a general taxonomy of knowledge extraction
tives on how to distinguish among elements of a set. Such methods and the associated criteria needed to evaluate
distinctions become the basis for stating hypotheses and them. As such, his work represents a valuable contribution
designing empirical studies. They can guide more special- to knowledge acquisition research.
ized and focused research endeavors. They also can serve Another very different framework illustrates how
as valuable pedagogical devices. researchers might approach the wide array of KA scenarios.
For example, Hoffman (1987) proposed a taxonomy of Kim and Courtney (1988) attempted to support the KA pro-
knowledge extraction techniques. This is summarized in cess by making conceptual mappings between different
Table 1. types of managerial problems and the different available
Hoffman goes on to suggest a number of criteria that KA techniques (Fig. 1). Their approach is based on classi-
might be used to evaluate and compare each of these fying problems by their size, complexity, and degree of
knowledge extraction methods. These criteria include the structure. They proceed to advocate the use of certain KA
time the task takes, its simplicity, its flexibility, its arti- methods for these classes. In order to make a conceptual
ficiality, the format and validity of the resulting data, and mapping between KA methods and the different types of
the efficiency of the method itself. managerial decisions, they argue that each type of decision
What is interesting to note is that this taxonomy is con- involves some combination of the three knowledge types.

4 Expert Systems, February 1997, Vol. 14, No. 1


Figure 1: A conceptual mapping of KA methodologies with decision types (Kim & Courtney, 1988).

For instance, because operating decisions are characterized there has been very little discussion of how alternative
as involving heuristic and conceptual knowledge, they knowledge acquisition taxonomies may be evaluated. The
deduce that a knowledge engineer-driven method such as best-known experimental research on comparative KA
protocol analysis would be the most appropriate. methods is that of Burton, Shadbolt, Hedgecock, and Rugg
Their approach, which has been described as a ‘selection- (1987). However, their research was limited to the ‘tra-
oriented’ framework (Holsapple and Wagner, 1996b) is ditional’ KA techniques of interviewing, card sorting and
quite different from other development-oriented frame- goal decomposition. The few pioneering studies are typified
works, with its greater consideration of the knowledge by confusing terminology, conflicting operationalizations,
acquisition context. Here, the problem domain is seen as a and the proliferation of ad hoc taxonomies. Each piece of
crucial part of the KA context, though it has been delimited research allows for different knowledge types, techniques
to include only managerial decision making. Unlike and even participants. For this reason numerous KA frame-
Hoffman’s view, the allowable participants in any occur- works and taxonomies have been presented and used but it
rence of knowledge acquisition are implied by the KA tech- is very difficult to evaluate the contributions of each. To
niques included in the mapping to be only single human that end this section proposes five different criteria to use
sources, a knowledge engineer, and a computer as a recep- in evaluating knowledge acquisition taxonomies. These five
tacle for the knowledge base. However, Kim and Courtney criteria are summarized in Table 2.
do not justify their framework’s specific mappings of The first criterion requires that the phenomenon being
knowledge types and problem attributes to knowledge
acquisition methods, either deductively or with empirical Table 2: Criteria for Evaluating KA Taxonomies
evidence. Moreover, problem attributes such as size and
complexity are not operationalized in their work. As such, 1. Is the phenomenon that is to be classified
the framework focuses on giving advice about what kind of accurately described by the taxonomy?
KA technique should be adopted, given a problem domain. 2. Are the properties or characteristics on which the
taxonomy is based accurately specified?
1. 2. Criteria for evaluating KA classification schemata 3. Are the categories mutually exclusive?
Regardless of the procedure used for inventing a classi- 4. Are the categories collectively exhaustive?
fication scheme, there are common criteria for evaluating 5. Is the taxonomy useful?
the resulting schemata. Yet within the KA literature itself,

Expert Systems, February 1997, Vol. 14, No. 1 5


classified be carefully defined, so that the taxonomy accu- ‘Classification may be regarded as a means for searching
rately describes it. It may be that a particular knowledge reality for hypotheses or for structuring reality to test
acquisition taxonomy reflects only a part of the phenom- hypotheses. It may also be regarded as a beginning point
enon and could be identified in a more appropriate manner. or the culmination of scientific investigation. We pos-
In the case of Hoffman’s taxonomy above, it only refers to sess, therefore, no means of assessing the adequacy or
interactions between single knowledge engineers and lone efficiency of a given classification independently of the
experts. The second criterion asks whether properties that job it is designed to do.’
have been chosen to do the classifying have been
In this respect, Hoffman’s taxonomy certainly was
adequately specified. In the case of the same taxonomy
important and useful because it provided a basis for per-
above, the characteristics may be inferred by examining the
forming better empirical work in the KA field.
description of each category. Hoffman’s category descrip-
tions are not inconsistent with each other, but some proper-
ties, such as information content and the time allowed to 2. Knowledge acquisition participants
perform each method, are only referred to in two categories
Perhaps the most interesting and complex variable in KA is
of methods. Otherwise, we would expect that what is called
the set of participants in the knowledge acquisition project.
the ‘method of familiar tasks’ would include categories
These vary in terms of type, number, abilities, and roles
where task time is limited or unlimited, and where the given
played. As mentioned in section 1, knowledge sources can
information is limited and unlimited. This category would
include data, texts, and also individual or groups of human
also suggest the existence of a set of methods using tasks
experts. In the literature, techniques as diverse as machine
with varying degrees of familiarity with respect to the
learning, text understanding systems, and the more tra-
expert, as Hoffman does describe.
ditional set of human-human knowledge engineering tech-
The third criterion suggests that all the categories at the
niques all rightfully claim to be involved in knowledge
same level should be mutually exclusive. That is, if one acquisition processes. This diversity of knowledge sources
item fits a particular category, it will not fit another. In the and KA techniques is summarized in Figure 2.
case of Hoffman’s classification of knowledge extraction Despite this wide variation in what is described as KA
methods, some of the categories are not entirely mutually research, a review of the KA literature reveals that past
exclusive. This is because the degree of knowledge engin- knowledge acquisition models and frameworks have
eer involvement in a particular knowledge extraction touched on only a limited view of the KA process and of
method is not included as a characteristic of all categories. the allowable project participants (Holsapple, Raj, and
So, it may be that the knowledge engineer merely analyzes Wagner, 1993; Dhaliwal and Benbasat, 1990, Adelman,
a protocol of the expert’s performance under information 1989; Boose, 1989). Keeping the proposed taxonomy cri-
and/or time limits. Or, it may be the case that the knowl- teria in mind, we introduce a more complete characteriz-
edge engineer is allowed to interact with the expert to some ation, which is crucial to an understanding of the KA
degree in each of these situations. Either way, it is not com- phenomenon.
pletely clear from the descriptions given. In order to get beyond the major division within the KA
The fourth criterion holds that taxonomies should strive field between those doing ‘manual’ or ‘traditional’ KA and
to be collectively exhaustive. That is, every example of the those involved in ‘automated’ KA we need to go beyond
phenomenon that is to be classified can be placed within the idea that KA participants must be considered to be only
one of the categories. If we equate Hoffman’s knowledge humans. Past work has postulated that a participant set in
extraction with the more common notion of knowledge an expert system development project can be understood
elicitation, it may be more appropriate to think of his tax- by viewing them as entities playing roles as determined
onomy as describing a range of tasks that a knowledge by a schema (Holsapple and Wagner, 1995). Certainly,
elicitation method can use to focus an expert’s attention. there can be other entities and roles in a development pro-
This taxonomy does not include other types of knowledge ject beyond those needed for KA. Here, our focus is on the
sources such as text, data, computer systems, all of which various possible participants in the KA portion of the
can require different knowledge extraction or elicitation project.
methods.
The last criterion simply asks ‘Is the taxonomy useful?’
2. 1. KA participants as entities
Does it adequately serve its intended purpose? How does
it compare with other alternative taxonomies? All classi- We consider the knowledge acquisition process in terms of
fications are designed to help solve some kind of problem. potential participants and the roles they play, rather than
It is therefore very important to keep in mind the purpose limiting our framework to simply explaining the ‘tra-
of the particular taxonomy being evaluated. As Harvey ditional’ knowledge acquisition problem of mapping the
(1969) states: knowledge from a human source via a human knowledge

6 Expert Systems, February 1997, Vol. 14, No. 1


Figure 2: Summary of KA sources and techniques (adapted from Selbig, 1986).

engineer to a computer where the expertise is to reside. The


expertise needed to complete an expert system project may
not exist in a single human expert. As mentioned earlier,
it may reside in multiple humans, text, visual media, audio
form, or even another computer. Similar diversity may exist
at the target side of KA. Moreover, we allow that single or
multiple entities can mediate a KA process. In our view,
an entity is a general description of a KA participant. In
this sense then an entity can be a human, computer, or a
proxy for either, and can be either an individual or aggre-
gate as will be discussed later.
The entities involved in a particular KA process are
determined by the project. Drawing from research in
decision support systems (DSS) we suggest that the generic
Figure 3: Generic decision support framework (Bonczek,
DSS framework could be one way of characterizing entities
Holsapple, and Whinston, 1981).
involved in KA. This framework posits that a DSS may
consist of a language system (LS), presentation system
(PS), problem-processing system (PPS), and a knowledge both computerized decision support and human-centered
system (KS) as shown in Figure 3 (Dos Santos and Holsap- decision support as well (Holsapple and Whinston, 1996;
ple, 1989; Bonczek, Holsapple, and Whinston, 1981). Holsapple and Wagner, 1995; Dos Santos and Holsapple,
The LS represents the set of all messages that the entity 1989).
can accept, while the complementary PS represents the set Although each entity that participates in a knowledge
of all the messages that the entity can issue. The entity’s acquisition process may be characterized with this model,
knowledge is stored in the KS. The PPS of the entity con- obviously, not all entities are alike. For example, a human
sists of its cognitive (i.e. knowledge/message processing) expert can be thought of as using his/her PPS to retrieve
abilities. It draws on the KS to process incoming messages, or even derive new knowledge from his/her KS after receiv-
produce outgoing messages, and solve problems. This ing a message from another entity via the LS and then
model has been applied in the DSS literature to describe transmitting the appropriate knowledge via its PS. Simi-

Expert Systems, February 1997, Vol. 14, No. 1 7


larly, we can use the same model to describe the possibility possibility of a project assigning any of the various kinds
of an existing expert system acting as a knowledge source of entities (e.g. humans, computers, data sets, textual
for another expert system. This model, in combination with passages) to the source role. Second, it permits multiple
our conception of the possible participant roles, is used to entities (both individual and aggregate) to play the source
describe participants in knowledge acquisition processes in role. Third, it encompasses the notions of both active and
the following section. passive sources. How a source role is enacted or manifests
for a particular KA occasion depends on the entity (or
entities) that is (are) assigned to that role. The general
2. 2. Roles
model of an entity can be used to describe and analyze such
In the literature we also find a general class of KA where possibilities. For instance, when an individual human plays
an expert interacts with a computer program that helps to the knowledge source role, we can study a LS for incoming
elicit his/her expertise (Boose, 1989). Here, the expert plays messages (e.g. from an agent), a PS for presenting outgoing
one role and the computer plays another. In other cases, an messages (e.g. to an agent), a KS serving as a repository
expert simply learns to use an expert system shell and of expertise, and a PPS to extract and derive knowledge
directly inputs his/her knowledge. Here, the same person for presentation.
plays two roles. There are still other possibilities, but in Consider the case where one or more of these parameters
general, an entity participates in a knowledge acquisition is lacking. For example, a human entity playing the role of
process by playing one or more roles. For our purposes, a knowledge source may not possess a LS (relative to the
role is defined as: knowledge acquisition agent) or may possess it only to a
degree that does not enable the process to function
A function (i.e. set of behaviors) that can be assumed
properly:
by an entity in a knowledge acquisition process.
The set of roles involved in KA could be variously speci- S ← (0, PS, PPS, KS).
fied by different researchers, depending on the level of
detail desired. The most obvious roles are the traditional This would be the case if the knowledge engineer were to
roles of expert(s) and knowledge engineers. For the sake listen to an expert lecture on the topic of interest, but did
of completeness and clarity, we postulate the existence of not have the opportunity to interact with the expert. Rela-
three distinct roles to which entities involved in KA may be tive to the knowledge engineer, this source did not have
assigned: knowledge source (S), agent (A), and knowledge the opportunity to process any incoming messages which
target (T). Simply put, a knowledge source serves as a might help it clarify its own outgoing messages.
repository of expertise and an agent’s role is to transfer A similar situation would be the use of audio tapes as
this expertise in a form acceptable to some entity playing sources of domain expertise. However, there would also be
a target role. the absence of a KS to draw on and a PPS to process the
As discussed earlier, analysis of our working definition KS. That is, enactment of the source role would consist
identifies these three roles. They are also consistent with only of a PS, with null values for other entity parameters:
the KA literature, although the concept of the agent role is
a more inclusive notion than knowledge engineer and the S ← (0, PS, 0, 0)
notion of a knowledge target role has not been explicitly
identified in prior frameworks. We now define and describe Here, the PS functions as the repository of expertise. Data
each of these roles in greater detail, using set theoretic and text as entities playing the source role also fit this pro-
notation where appropriate. file. They can be viewed as subsets of some human entity’s
PS. Relative to the agent, all other entity parameters have
2. 2. 1. Knowledge sources Aside from individual human null values.
experts, many other kinds of knowledge sources (e.g. com- This kind of analysis can be extended to describe human
puters, data sets, textual passages, multi-participant human source limitations with respect to any one of the particular
sources, and combinations of these) can form part of the parameters. For example, a human source’s KS can be
context for a KA process. To encompass all of these, we limited with regard to both the type of knowledge it con-
define a knowledge source as follows: tains and the way in which it is organized. Research along
these lines has been performed in cognitive psychology to
A knowledge source provides a repository of expertise
show that there may be a correlation between a person’s
(i.e. reasoning knowledge) that is operated on in a KA
personality and the way in which he/she organizes his/her
process to satisfy a project’s expert system develop-
knowledge. Changes in the KS of a human may also be
ment objectives.
linked to capabilities of both the LS to handle incoming
This general definition of a knowledge source is messages and the PS to present outgoing messages. The
important in several ways. First, it does not preclude the problem of eliciting ‘tacit’ knowledge from a human expert

8 Expert Systems, February 1997, Vol. 14, No. 1


certainly is linked to such limitations on the KS, LS, and knowledge acquisition project are defined prior to
PS. Expertise may be retrieved or derived from its KS and implementation of the process, a target is defined in which
expressed as a set of messages via the source’s PS, or it the acquired knowledge is expected to reside. Thinking of
may be prepackaged in the PS of the entity playing the the target as describing a role filled by an entity, we can
source role. again use the generic entity model to describe the manifes-
tations of a knowledge target. This means a knowledge tar-
2. 2. 2. Agent The notion of the knowledge acquisition get, like the knowledge source and knowledge acquisition
agent role allows us to readily consider both human and agent, can be described as having a LS, PS, PPS, and a KS.
computer-based entities with the same model. The idea of In the context of typical knowledge acquisition research
modelling both human and computer KA agents has also this target is usually the KS of an expert system or other
been noted in other recent research (Schreiber, Wielinga, computer-based system. With the notion of a knowledge
de Hoog, Akkermans, and Van de Velde, 1994). As with target however, we can expand the model to describe other
the source role, we view the agent as a role being played by targets that have not been recognized in the KA literature.
one or more entities involved in the knowledge acquisition For example, the notion of a knowledge target role allows
process. This allows us to characterize an agent as having us to consider the possibility that a text may be the desired
a language system (LS), a presentation system (PS), a prob- target of a particular knowledge acquisition process. This
lem processing system (PPS), and a knowledge system would be the case when a project aims to generate just the
(KS). expert system specifications. In this case the target entity
would consist only of a language system with respect to
The role of a KA agent is to transfer/transform knowl-
the KA process itself:
edge from source entities to target entities to meet the
project’s objectives.
T ← (LS, 0, 0, 0).
The basic idea of the agent role is that one or more
entities act to facilitate the mapping of knowledge. This The range of representations accepted by the knowledge
characterization of the agent role suggests that (unlike a target in this case would depend on the nature of the text
source role) it cannot be passive. That is, it must possess desired.
a PPS. This role may go so far as entailing the enforcement It is important to note that the knowledge target is not to
of regulations that pertain to the particular knowledge be confused with the knowledge representation itself. The
acquisition process being implemented. For example, one knowledge representation is the message output from the
type of protocol analysis may require that the agent not agent, based on its presentation strategy. There is consider-
interfere with the expert’s problem-solving process or struc- able research about various knowledge representations and
tured interviewing may require that the agent use a prior their relative advantages (McGraw & Harbison-Briggs,
protocol in order to derive a set of in-depth questions for 1989), but the KA literature says little about the entity for
the source expert. which they are targeted and possible ways its capabilities
As stated earlier, a traditional major dividing line in KA could affect the KA process and outcome. Our meta-view
research has separated the knowledge engineering mode of also accommodates the possibility of multiple represen-
KA from modes that automate the KA process via com- tations with respect to one or more entities playing the tar-
puter-based methods. Here, we have a unified way of view- get role in a KA process.
ing these traditionally compartmentalized KA modes. Also, The idea of a knowledge target is important in order to
unlike past research, we have the notion that it may be completely characterize the knowledge acquisition process.
appropriate within the scope of a project to assign multiple This is because an active knowledge target will have its
entities to the agent role. These may be a mix of human own abilities and attributes, which affect the type of knowl-
and computer-based entities. Each such entity can be an edge representations it will accept from the agent. Indeed,
individual (a person or computer) or an aggregate (e.g. a from the view presented here, it is conceivable that the tar-
team of specialists, a person supported by a computer, etc.). get may even drive the KA process.
It is also possible that acquired knowledge could be
2. 2. 3. Target Our characterization of participant roles mapped to a human target. Allowing a human knowledge
would not be complete without the idea of a knowledge target lets the framework deal with situations such as com-
target role. puter-aided instruction, or even technology transfer (in the
micro sense), which may also be characterized as knowl-
A knowledge target holds a repository of expertise
edge acquisition phenomena. That is, the view presented
resulting from a KA process undertaken to satisfy a pro-
here suggests that it would be possible to have a computer
ject’s development objectives.
as a source for one or more human knowledge targets. To
As with the source role, this allows for both active and the extent that this has been previously discussed, it has
passive targets. When the objectives and resources of the been characterized as an aspect of organizational learning.

Expert Systems, February 1997, Vol. 14, No. 1 9


2. 3. Mappings between entities and roles If the re function is applied to a role, it will yield at least
one of the entities. In general:
Prior knowledge acquisition frameworks do not acknowl-
edge the possibility of participants in a KA process playing
er(Ei) ; R, where Ei ∈ E
multiple roles (Holsapple and Wagner, 1996b; Holsapple,
Raj, and Wagner, 1993). In some automated knowledge re(Rj) ; E, where Rj ∈ R
acquisition approaches (e.g. some that make use of reper-
tory grids) a computer plays both the role of the agent and A closer examination of these two functions helps
the role of the target (Shaw and Woodward, 1988). In this describe what we term as elemental and composite roles.
scenario, a computer is essentially interviewing the expert If the function, re, is applied to a certain role and yields
and deriving rules from the expert’s input that would be only one entity, we have an elemental role. For example,
suitable for the resulting expert system. Likewise it is not
unusual for a person to function as both source and agent. re (S) = {E2}
In this scenario, the expert self-extracts knowledge with or
without the help of the computer (Bradshaw and Boose, indicates that the project has specified that an elemental
1989). knowledge source be used in this KA process. If E2 is an
Prior frameworks generally do not account for the possi- individual entity, this example describes the traditional case
bility that multiple entities, each with differing abilities, of a single human expert playing the role of a knowledge
may be assigned to a single role. This could occur if several source. If E 2 is an aggregate entity, the example charac-
experts were assigned to the source role or multiple knowl- terizes the situation where a panel of experts function as a
edge engineers played the agent role. In practice however, unified knowledge source. The nature of their interactions
KA tools such as AQUINAS have been building in features is internal to the entity, rather than being imposed by the
that allow them to acquire knowledge from multiple experts KA process. In neither case is there an issue of the KA
(Gaines, 1987) Thus, the framework’s description of poss- process needing to coordinate the use of multiple sources,
ible KA participant scenarios needs a way of describing because there is only a single source. The same holds for
how the entities are related to the various roles and also elemental agent and target roles.
how the roles may be related to the entities playing them. The project schema also allows multiple entities to be
To support these cases we introduce the notion of a mapped to the same type of role. We refer to this as a
schema: composite role. For instance, if
A project schema describes how an expert system pro-
er(E1) = {S}
ject maps entities to the roles they play and how each
role is mapped to the entities that fill it. er(E2) = {S}
This definition of a schema allows an individual entity er(E3) = {S}
or aggregate entity to play more than one type of role, such
as in the case where an entity serving as an agent also then the source role is composite. The project constraints
serves as the target. It also allows for the scenario where allow that three entities (E1, E2 and E3) play the source type
each type of role is replicated, being filled by multiple enti- of role. Each of these entities will be either individuals or
ties. An example is the fairly common case where multiple aggregate. The coordination of these three entities must be
experts are used as knowledge sources. To accommodate addressed (e.g. the sequence in which they are used, the
these distinct possibilities within our characterization of reconciliation of differences among them). So, if the er
KA participants, we describe two functions; the entity-role function, when applied to all available E i, results in more
(er) and the role-entity (re) function. The idea of the er than one entity being mapped to the same role (Rj), then
function is that it can be used to describe how entities (Ei) that role is said to be a composite role. This kind of schema
are mapped to specific roles in the KA process. The con- accommodates a scenario where multiple entities are used
verse is true of the re function which maps each role to as distinct knowledge sources. Each such entity could be a
the entities that perform that type of role in the KA process. human, a computer, a database, an aggregate (e.g. an expert
Taken together, these two functions describe the project panel), or even a textbook. In actual practice, it is quite
schema. Both are necessary in order to fully account for common for knowledge engineers to refer to texts, proto-
the possible entity-role and role-entity variations. In this cols, and distinct experts or groups of experts during the
way, all possible project scenarios involving single and course of a development project as in the classic ExperTax
aggregate entities playing single and multiple roles are system developed by Coopers and Lybrand (Shpilberg,
accounted for in our meta-view. Graham, and Schatz, 1986).
For example, if the er function is applied to an entity it It is also interesting to consider a case where the er func-
will yield at least one role from the set of allowable roles. tion yields two roles for the same entity. For example,

10 Expert Systems, February 1997, Vol. 14, No. 1


er(E1) = {A, T} entity that alone fills the role, (3) an individual that is one
of multiple entities filling a type of role, and (4) an aggre-
reveals that one entity is playing the A role and the T role. gate entity that is one of multiple entities assigned to a
A review of the knowledge acquisition literature indicates role type.
that this characterizes some of the automated KA scenarios This way of viewing project participants is interesting
mentioned earlier (Boose, 1989; McGraw and Harbison- when we begin to consider how it may impact the knowl-
Briggs, 1989). Within the view presented here, we refer to edge acquisition process at any level. For example, if we
E1 as a multi-role entity because we have an entity playing look at the potential variations in a source-agent interaction,
more than one role. Generally, if when applied to an entity, using Figure 4 as a starting point, we obtain the scenario
Ei, the er function yields more than one role, then that matrix shown in Table 3.
entity is said to be a multi-role entity. As a descriptor of the potential scenarios involved in a
The remainder of this article demonstrates that taxo- source-agent interaction, the matrix captures all of the the-
nomies can be generated from the meta-view framework of oretic possibilities suggested by the framework. For
knowledge acquisition roles and participants. These taxo- instance, cell 1:2 denotes* an interaction where an individ-
nomies are created via the logical partitioning approach ual elemental source interacts with an aggregate elemental
with the criteria of Table 2 in mind. These taxonomies are agent (e.g. an individual human expert with a knowledge
meant to suggest alternative ways of characterizing con- engineer group). Cell 2:4 denotes the case where an aggre-
structs identified by the framework, as a basis for organiz- gate elemental source (e.g. an expert panel) interacts with
ing and stimulating future KA research. Our intent is to multiple aggregate agents (e.g. several knowledge engineer
illustrate rather than be exhaustive or prescriptive. groups). It is interesting that prior research has focussed
almost exclusively on the scenarios described by cells 1:1
and 3:1. We contend that the other cells deserve to be inves-
3. Deriving knowledge acquisition participant
tigated with respect to such issues as appropriate KA
taxonomies
methods, tool support possibilities, and their advantages/
One way of showing the proposed framework’s utility is disadvantages relative to each other.
to use it to generate KA taxonomies that can guide empiri- A similar KA scenario matrix is shown in Table 4. Here,
cal and developmental research. This section introduces the classification is with respect to the agent-target interac-
several new taxonomies based on the concepts introduced tions rather than source-agent relationships. Only cell 1:1
to describe KA participants in the previous section. has been examined extensively in the KA research literature
Knowledge acquisition roles have been characterized (Holsapple and Wagner, 1995; Wagner, 1992).
here as being composite and elemental, while entities were A special case of this taxonomy is shown in Table 5
classified as being either individual or aggregate. When where the source and agent are the same (S = A). Included
combined, these two ideas define a multi-level taxonomy are scenarios where the S learns to interact with the T
of entities and roles. Figure 4 summarizes four entity categ- directly and is able to self-elicit his/her knowledge
ories that exist for any of the defined roles (Rj): (1) an
individual entity that alone fills the role, (2) an aggregate
Table 3: Source-agent interaction matrix

*By convention, the first number of a cell identifier refers to the


rows of the matrix and the second number refers to the columns
Figure 4: Sample KA role and entity taxonomy. of the matrix.

Expert Systems, February 1997, Vol. 14, No. 1 11


Table 4: Agent-target interaction matrix Table 6: S-Agent = target interaction matrix

Table 5: Source = agent-target interaction matrix


of taxonomies. Each contains categories that are mutually
exclusive and collectively exhaustive with respect to the
interaction with which they are concerned. The properties
of entities and roles also accurately describe the possible
combinations of interactions in a way not covered by prior
frameworks. By showing combinations of interactions that
have yet to be studied or suggested, the criterion of useful-
ness is also met.

3. 1. Research issues generated by the analysis


The purpose of this section is to demonstrate the usefulness
of the meta-view framework as a tool for stimulating our
thinking as KA researchers, developers of knowledge
acquisition techniques, and builders of expert systems.
(Shephard, 1993; Boose, 1989). Also included are cases These issues are not explored further in the context of the
where expert system shells have sufficiently accessible present work but are posed as questions that emerge from a
interfaces to enable the source to represent his/her knowl- consideration of the framework. This list is not necessarily
edge in a way that is understandable to the target computer. exhaustive, but is included to demonstrate the utility of the
These types of interactions are described by cell 1:1. The framework as an aid to KA research.
matrix suggests that many other types of interactions can As previously mentioned, project participants have been
occur and may be of interest to investigate. considered in limited ways by KA researchers. Interest in
Table 6 shows the taxonomy that results where the agent these has primarily centered on the attributes of human
and target are the same (A = T). A knowledge acquisition knowledge sources and how they may affect the knowledge
project characterized by S:A = T interactions encompasses acquisition process. More issues emerge from the para-
attempts to automate the knowledge acquisition process in meterization of the participants as introduced in this
various ways. One method is to have a source expert per- research.
form a simulation of the problem on a computer, which
automatically transforms the expert’s performance into I1. How are aggregate S’s, A’s, and T’s integrated
appropriate knowledge representations. Another interaction into the process?
of this type occurs when a text understanding system or I2. How are composite and elemental S’s, A’s, and
other machine learning system induces a rule set or other T’s integrated into the process?
suitable knowledge representation from a text or data set. I3. How is KA affected if the aggregate S’s, A’s, or
Both of these examples fall into the category of interaction T’s are distributed either spatially or temporally?
described by cell 1:1. Other categories have not been I4. What happens to KA if only certain elements of
reported on by KA researchers. a participant or group of participants (e.g. LS, PS,
We can use the five criteria in Table 2 to critique this set PPS, or KS) are distributed?

12 Expert Systems, February 1997, Vol. 14, No. 1


I5. What if the sources are not just humans but a mix Preference Models Using Aquinas.’ International Journal of
of text, data, computers, and humans? Man-Machine Studies. no. 31.
Burton, A.M., N.R. Shadbolt, A. Hedgecock and G. Rugg
I6. Can a computer serve as a S? What kind of special (1987). ‘A Formal Evaluation of Knowledge Elicitation Tech-
KA support might this require? niques for Expert Systems.’ Proceedings of the 1st European
I7. How can human and computer A’s work together? Workshop on Knowledge Acquisition for Knowledge-Based
I8. How would multiple human A’s work together? Systems, Reading University, September, 1987.
What support exists for this? Clancy, W.J. (1986). ‘Heuristic Classification.’ in J. Kowalik,
Ed. Knowledge-Based Problem-Solving. New York: Prentice-
I9. Can a target be other than a computer? Can it be Hall.
a human, text, or even a mixed group? Dhaliwal, J.S. and I. Benbasat (1990). ‘A Framework for the
I10. How can the target interact with the A or the S to Comparative Evaluation of Knowledge Acquisition Tools and
improve the efficiency of KA? How do T abilities Techniques.’ Knowledge Acquisition. Vol. 2, no. 2.
affect KA? Dos Santos, B. and C. Holsapple (1989). ‘A Framework for
Designing Adaptive DSS Interfaces.’ Decision Support Sys-
I11. What happens when individual entities play mul- tems. Vol. 5, no. 1.
tiple roles? Can S = A and A = T? Can the S ever Fellers, J.W. (1987). ‘Key Factors in Knowledge Acquisition.’
equal the T? Computer Personnel. Vol. 11, no. 1.
I12. Can aggregate entities play multiple roles? Does Gaines, B. (1987). ‘An Overview of Knowledge Acquisition and
this situation ever occur? Transfer.’ International Journal of Man-Machine Studies. no.
26.
I13. How can additional entity attributes be expressed? Gil, Y. and C. Paris (1994). ‘Towards Method-Independent
Can they be adequately discussed in terms of LS, Knowledge Acquisition.’ Knowledge Acquisition. Vol. 6, no. 2.
PS, PPS, and KS abilities? Grabowski, M., A. Massey and W. Wallace (1992). ‘Focus
Groups as a Group Knowledge Acquisition Technique.’ Knowl-
edge Acquisition. Vol. 4, no. 4.
4. Conclusion Harvey, D. (1969). Explanation in Geography. New York: St
Martins Press.
The set of role and participant taxonomies presented here Hoffman, R. (1987). ‘The Problem of Extracting the Knowledge
of Experts From the Perspective of Experimental Psychology.’
represents an innovative way of viewing the many vari-
AI Magazine. Vol. 8, no. 2.
ations in KA participants and can help KA researchers Holsapple, C., V. Raj and W. Wagner (1993). ‘Knowledge
structure further research. Here, we have introduced taxo- Acquisition: Recent Theoretic and Empirical Developments.’ in
nomies derived directly from the meta-view presented. Recent Developments in Decision Support Systems (C. Holsap-
They have not previously appeared in the KA literature. ple and A. Whinston). Berlin: Springer Verlag.
Holsapple, C. and W. Wagner (1995). ‘Contextual Factors in
The taxonomies introduced here are not meant to be
Knowledge Acquisition.’ International Journal of Expert Sys-
exhaustive with respect to knowledge acquisition pheno- tems: Research and Applications. Vol 8, no. 4.
mena or to the meta-view itself, but simply illustrate how Holsapple, C. and W. Wagner (1996a). ‘Process Factors in
it can be useful in deriving taxonomies and research Knowledge Acquisition.’ Expert Systems. Vol. 13, no. 1.
issues via logical partitioning. Moreover, the taxonomies Holsapple, C. and W. Wagner (1996b). ‘Knowledge Acqui-
sition Frameworks: A Survey, Assessment, and Implications.’
generated appear to be useful in their own rights, for
forthcoming in Journal of Computer Information Systems.
helping organize comparative knowledge acquisition fall/winter.
experiments and suggesting KA research topics that are Holsapple, C. and A. Whinston (1996). Decision Support Sys-
as yet uninvestigated. tems: A Knowledge-Based Approach. St. Paul, MN: West Pub-
lishing Company.
Kaplan, R. and G. Berry-Rogghe (1991). ‘Knowledge-Based
Acquisition of Causal Relationships in Text.’ Knowledge
References
Acquisition. Vol. 3, no. 3.
Adelman, L. (1989) ‘Measurement Issues in Knowledge Engin- Kim, J. and J. Courtney (1988). ‘A Survey of Knowledge Acqui-
eering.’ IEEE Transactions on Systems, Man and Cybernetics. sition Techniques and Their Relevance to Managerial Problem
(19), 483–488. Domains.’ Decision Support Systems. Vol. 4, no. 3.
Barbuceanu, M. (1993). ‘Models: Toward Integrated Knowl- Liou, Y. (1990). ‘Knowledge Acquisition: Issues, Techniques,
edge Modeling Environments.’ Knowledge Acquisition. Vol. 5, and Methodology.’ Proceedings of the 1990 ACM SIGBDP
no. 3. Conference on Trends and Directions in Expert Systems.
Bonczek, R., C. Holsapple, and A. Whinston (1981) Foun- Orlando, FL.
dations of Decision Support Systems. Orlando, FL: Academic Liou, Y., E. Weber and J. Nunamaker (1990). ‘A Methodology
Press Inc. for Knowledge Acquisition in a Group Decision Support Sys-
Boose, J. (1989). ‘A Survey of Knowledge Acquisition Tech- tem Environment.’ Knowledge Acquisition. Vol. 2, no. 2.
niques and Tools.’ Knowledge Acquisition. Vol. 1 no. 1. McGraw, K. L. and K. Harbison-Briggs (1989). Knowledge
Boose, J., J. Bradshaw, J. Koszarek, and D. Shema (1993). Acquisition: Principles and Guidelines. Englewood Cliffs, NJ:
‘Knowledge Acquisition for Group Decision Support.’ Knowl- Prentice-Hall.
edge Acquisition. Vol. 5, no. 4. Morik, K. (1991). ‘Underlying Assumptions of Knowledge
Bradshaw, J. and J. Boose (1989). ‘Decision Analytic Tech- Acquisition and Machine Learning.’ Knowledge Acquisition.
niques for Knowledge Acquistion: Combining Situation and Vol. 3, no. 2.

Expert Systems, February 1997, Vol. 14, No. 1 13


Rappaport, A. and B. Gaines (1990). ‘Integrated Knowledge The authors
Base Building Environments.’ Knowledge Acquisition. Vol. 2,
no. 1.
Reinartz, T. and F. Schmalhofer (1994). ‘An Integration of
Knowledge Acquisition Techniques and EBL for Real-World William P. Wagner
Production Planning.’ Knowledge Acquisition. Vol. 6, no. 2.
Schreiber, G., B. Wielinga, R. de Hoog, H. Akkermans and Professor Wagner is an Assistant Professor of Management
W. Van de Velde (1994). ‘CommonKADS: A Comprehensive in MIS at Villanova University. His teaching and research
Methodology for KBS Development.’ IEEE Expert. Vol. 9, interests are in the areas of decision support systems, expert
no. 6.
Selbig, J. (1986). ‘Knowledge Acquisition by Inductive Learning systems and database management. Much of his recent
from Examples.’ in Lecture Notes in Computer Science. work has focused on the knowledge acquisition problem
Springer Verlag. no. 265. and has been presented at several professional conferences
Shaw, M. and J. Woodward (1988). ‘Validation in a Knowledge such as the Decision Science Institute, ACM SIGBDP,
Support System: Construing Consistency With Multiple TIMS/ORSA and the NATO Advanced Studies Institute.
Experts.’ International Journal of Man-Machine Studies. no.
29.
Shephard, G. (1993). ‘Providing Descriptive Power to Guided
Self-Elicitation.’ Knowledge Acquisition. Vol. 5, no. 3. Clyde W. Holsapple
Slator, B. (1989). ‘Extracting Lexical Knowledge from Diction-
ary Text.’ Knowledge Acquisition. Vol. 1, no. 1. Professor Holsapple holds the Rosenthal Endowed Chair in
Sommer, E., K. Morik, J. Andre and M. Uszynski (1994). MIS at the University of Kentucky. He does teaching and
‘What Online Machine Learning Can Do for Knowledge Acqui-
sition — A Case Study.’ Knowledge Acquisition. Vol. 6, no. 4. research in decision support systems, knowledge manage-
Shpilberg, D., L.E. Graham and H. Schatz (1986). ment, expert systems and organisational computing. His
‘ExperTAX: An Expert System for Corporate Tax Planning.’ books include Foundations of Decision Support Systems,
Expert Systems. Vol. 3, no. 3. Micro Database Management, Business Expert Systems,
Wagner, W. (1992) dissertation entitled ‘A Phenomenon- The Information Jungle and the just published Decision
Oriented Knowledge Acquisition Framework.’
Support Systems: A Knowledge-Based Approach. His
research articles span such diverse journals as Decision
Sciences, Operations Research, Decision Support Systems,
Datamation, IEE-SMC, IEE Expert, Information Systems,
Expert Systems, Knowledge Acquisition, The Computer
Journal, Information and Management, Financial Manage-
ment, OMEGA, Journal of Organizational Computing and
Organization Science. Dr. Holsapple is Area Editor for
Decision Support Systems and Associate Editor for Man-
agement Science. In 1993 he was named ‘Computer Edu-
cator of the Year’, by the International Association For
Computer Information Systems.

14 Expert Systems, February 1997, Vol. 14, No. 1

You might also like