Sei sulla pagina 1di 9

Knowledge-Based Systems 15 (2002) 439–447

www.elsevier.com/locate/knosys

Knowledge acquisition for expert systems in accounting and financial


problem domains
W.P. Wagner*, J. Otto, Q.B. Chung
Department of Decision and Information Technologies, Villanova University, Villanova, PA 19085 USA
Received 12 October 2001; accepted 4 February 2002

Abstract
Since the mid-1980s, expert systems have been developed for a variety of problems in accounting and finance. The most commonly cited
problems in developing these systems are the unavailability of the experts and knowledge engineers and difficulties with the rule extraction
process. Within the field of artificial intelligence, this has been called the ‘knowledge acquisition’ (KA) problem and has been identified as a
major bottleneck in the expert system development process. Recent empirical research reveals that certain KA techniques are significantly
more efficient than others in helping to extract certain types of knowledge within specific problem domains. This paper presents a mapping
between these empirical studies and a generic taxonomy of expert system problem domains. To accomplish this, we first examine the range of
problem domains and suggest a mapping of accounting and finance tasks to a generic problem domain taxonomy. We then identify and
describe the most prominent KA techniques employed in developing expert systems in accounting and finance. After examining and
summarizing the existing empirical KA work, we conclude by showing how the empirical KA research in the various problem domains can
be used to provide guidance to developers of expert systems in the fields of accounting and finance. q 2002 Elsevier Science B.V. All rights
reserved.
Keywords: Expert systems; Accounting expert systems; Finance expert systems; Knowledge acquisition; Problem domain

1. Introduction they know it. However, new empirical research in the field
of expert systems reveals that certain KA techniques are
Since the mid 1980s Expert Systems have been significantly more efficient than others in different KA
developed for a number of different accounting and finance domains and scenarios. Adelman [5] was one of the first to
applications. These systems have either been developed design experiments to objectively compare the effectiveness
completely in house, purchased as proprietary software, or of different KA techniques. He identified five determinants
developed using an expert system shell. The most of the quality of the resulting knowledge base. These are:
commonly cited problems in developing these systems are
the unavailability of both the experts and knowledge 1. Domain experts
engineers and difficulties with the rule extraction process 2. Knowledge engineers
[1,2]. This has been called the ‘knowledge acquisition’ 3. Knowledge representation schemes
(KA) problem and has been identified as a major bottleneck 4. Knowledge elicitation methods
in the expert system development process [3,4]. Simply 5. Problem domains.
stated, the problem is how to efficiently acquire the specific
knowledge for a well-defined problem domain from one or This paper presents a mapping between the body of KA
more experts and represent it in the appropriate computer empirical studies and the different problem domains within
format. accounting and finance to guide developers of accounting
Given the ‘paradox of expertise’, the experts have often and finance expert systems in their choice of KA techniques.
proceduralized their knowledge to the point that they have
difficulty in explaining exactly what they know and how
2. A generic problem domain taxonomy
* Corresponding author. Tel.: þ 1-610-519-6449.
E-mail address: william.wagner@villanova.edu (W.P. Wagner). Research in the field of knowledge acquisition has
0950-7051/02/$ - see front matter q 2002 Elsevier Science B.V. All rights reserved.
PII: S 0 9 5 0 - 7 0 5 1 ( 0 2 ) 0 0 0 2 6 - 6
440 W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447

Table 1 2.1. Mapping of accounting and finance tasks to the generic


Generic problem domain taxonomy problem domains
Analysis problems
† Classification—categorizing based on observables. The generic task domains defined in the taxonomy can be
† Diagnosis—inferring system malfunctions from observables. mapped into accounting and finance task domains. Such a
† Debugging—prescribing remedies for malfunctions. mapping is presented as Table 2, along with selected
† Interpretation—inferring situation descriptions from sensor data.
examples of accounting and finance expert systems. To do
Synthesis problems the mapping, we conducted an analysis and survey of
† Configuration—configuring collections of objects under constraints in accounting and finance expert system ease studies.
relatively small search spaces.
† Design—configuring collections of objects under constraints in A wide variety of problems have been addressed with
relatively large search spaces. expert systems with varying levels of success. The most
† Planning—designing actions. common applications were in the domains of classification
† Scheduling—planning with strong time and/or space constraints. and planning with no clear examples of systems in the repair
Problems combining analysis and synthesis or scheduling domains. These tasks were then placed in the
† Command and control—ordering and governing overall system control. generic taxonomy based upon the generic task descriptions.
† Instruction—diagnosing, debugging, and repairing student behavior. Due to the complex nature of accounting and finance
† Monitoring—comparing observations to expected outcomes.
† Prediction—inferring likely consequences of given situations.
business functions, some of them may also fall under
† Repair—executing plans to administer prescribed remedies. multiple task domains and the categories are not completely
exclusive. For example, depending on the scope of the
inputs and the resulting advice from the expert system, a
focused on several dimensions of the problem as classification expert system might also fall in the categories
determining factors. As previously mentioned, one of debugging or diagnosis.
primary determinant of the KA technique used to In addition, the process of mapping specific functions to
develop an expert system is the problem domain. To the higher level categories of analysis, synthesis, and the
enhance research in the KA field, generic problem combination reveals some interesting characteristics of
domain taxonomies have been developed that cut across accounting and financial problems. Looking at the specific
functional areas. The most commonly used taxonomy tasks that fall within the analytic category shows that all of
breaks problems into general categories of analysis, these tasks involve taking a set of data inputs and
synthesis, and those that combine analysis and synthesis identifying patterns in them. In contrast, the synthetic
[4,6,7]. This taxonomy is shown as Table 1. problems require that solutions be generated based upon the

Table 2
Accounting and finance task domains and expert system examples

Task domains Accounting and finance expert systems

Categories Generic Accounting and finance

Analysis Classification Classify firms into writedown or WDXPERT [36]


non-writedown categories
Debugging Evaluating loan losses and proposing ES for evaluating loan losses.
solution strategies [41]
Diagnosis Consumer relations analysis ESCFE [37]
Interpretation 1. Analyze accounting variances 1. Model variance analysis expert
system [39]
2. Interpret tax consequence of 2. CORPTEX [42]
stock redemptions
Synthesis Configuration Managing a business loan portfolio. MARBLE [43]
Design 1. Design a personal financial 1. Planman [44]
plan
2. Design a compensation system 2. IMIS [45]
Planning 1. Audit planning 1. Expertest [46]
2. Investment planning 2. Vanguard online planner [40]; capital investment system [47]
Scheduling None found
Combination Command and control Estimate control risk C and L control risk
assessor [46]
Instruction Training insolvency counselors PISCES: ([48])
Monitoring Fraud detection Fraud detection system [49]
Prediction Risk estimation APX [50]
Repair None found
W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447 441

Table 3
Excerpt from an unstructured interview (adopted from Ref. [2])

I: What information are you given about airports?


E: Now, just some rudimentary information comes with it. Common name, latitude and longitude, ah…no information comes with it
about the ah…maximum number of airplanes on the ground or the port capability that is at the field. None of that comes along with it
I: What’s the difference between MOG and airport capability?
E: Ah…MOG maximum on the ground is parking spots… on the ramp. Airport capability is how many passengers and tons of cargo per
day it can handle at the facilities
I: Throughout…ah…throughout as a function of …
E: It all sorta goes together as throughput. If you’ve only got…if you can only have ah…if you’ve only got one parking ramp with the
ability to handle 10,000 tons a day, then your…your throughput is gonna be limited by your parking ramp. Or, the problem could be
vice versa
I: Yeah…
E: So it’s a (unintelligible phrase).

more general goals of the system and involve the search of a technique was the ‘unstructured interview’, where the
much larger set of potential solutions. Combinations of the knowledge engineer asks general questions and just hopes
two are typically the most ambitious types of expert systems for the best. However, each technique requires different
in that they must perform in-depth analysis of large amounts abilities from the knowledge engineer, the knowledge
of diverse input data, identify the problems and causes and source, and allows a different set of knowledge represen-
design a possible solution. The inherent difficulties of tations to be used.
developing systems for these types of problems may be the Although the human-based knowledge acquisition tech-
underlying reason so that few of these types of expert niques described in the following sections are certainly the
systems have been attempted in the area of accounting and most common used today, they are certainly not without
finance [52]. their problems. Not only do they require an enormous
When the task categories presented in Table 2 are amount of time and labor on the part of both the knowledge
combined with the comparative empirical KA studies in engineer and the domain expert but they also require the
Table 4 (which is to be discussed later in detail), these knowledge engineer to have an unusually wide variety of
mappings may serve as a guide for which KA techniques interviewing and knowledge representation skills in order
might be the most appropriate for the different problem for them to be successful. Research on the most common
domains within accounting and finance. A general discus- elicitation and representation techniques used by human
sion of the most prominent KA techniques is provided in the agents are presented in the following sections.
Section 3.
3.1. Unstructured interviewing techniques

3. Knowledge acquisition techniques Undoubtedly the most common technique currently used
by knowledge engineers [2,12], it is difficult to describe this
In the strict sense, knowledge elicitation should be as a true technique, since as its name implies it is just a
viewed as one phase of the knowledge acquisition process. wandering conversation between the expert and the knowl-
However, in much of the research knowledge elicitation and edge engineer. Table 3 is a transcript of a typical
knowledge acquisition are used interchangeably. The role unstructured interview given as an example.
that the human knowledge engineer will play in the Even though this may be the case, unstructured
knowledge acquisition process will vary considerably interviewing still has a valuable place in the knowledge
depending on the particular elicitation technique or method engineer’s toolkit since it allows the greatest possible
used. In some cases, it may be appropriate for the freedom for knowledge engineer and expert alike to explore
knowledge engineer to become an apprentice to the expert the topic. Many researchers have documented and described
or participate somehow in the actual problem-solving its usage, although researchers in the field usually downplay
process. Other times it may be better for the knowledge its value as a real tool. In fact, an anthropological study of
engineer to conduct an unstructured interview or simply knowledge engineering cited the use of the unstructured
observe the expert perform a given task. interview as one of the biggest failings of knowledge
Many different techniques have been developed engineers who were attempting to develop expert systems
especially for knowledge engineers in these different [13].
situations or have been drawn from existing research in While it is certain that no interview is completely
fields such as psychology, and several researchers have unstructured, different types of unstructured interviews have
attempted to summarize these [2,4,9 – 11]. Of these been suggested by authors of general surveys of knowledge
techniques, it should not be surprising that a survey [12] acquisition techniques. In the most extreme case, the
found that the most commonly used knowledge elicitation knowledge engineer does not have a prepared set of detailed
442 W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447

Table 4 knowledge engineer wishes to use to have the expert refine


Task types [2] previously elicited knowledge. The simplest task the
Familiar task activities † Unobtrusive observation knowledge engineer could give the expert could be to
† Simulated familiar tasks prepare a brief lecture designed to lay out the main themes
and ideas associated with the particular problem domain.
Interview tasks † Unstructured
† Structured Obviously, this type of task would be more appropriate for
early knowledge acquisition sessions whereas the special
Constrained professing tasks † Case-based reasoning
task would be better for when the knowledge engineer was
† Event recall
† Scaling and Sorting tasks more familiar with the particular domain [2].
† Creative problem solving It should be noted that the tasks which are used as the
† Decision analysis basis for structuring the interview can be either actual tasks
or simulated tasks. This method of structuring the interview
questions and the expert does not have ready replies or process by using specific tasks has been termed ‘con-
information at his/her fingertips. The result is that the strained-processing’ [2] and the different tasks were
interview takes the form of a wandering dialog in which grouped as shown in Table 4.
open-ended questions are asked by the knowledge engineer.
However, this can be helpful both in acclimating the expert 3.3. Protocol analysis
to the ideas of artificial intelligence and in helping the
knowledge engineer learn general ideas about the problem Protocol analysis is one of the most frequently mentioned
domain. It serves the additional purpose of building an elicitation techniques in the knowledge acquisition litera-
essential rapport between the system developer and the ture. [12] found it to be second only to unstructured
human source. interviews in actual usage. Suggested by Newell and Simon
Unfortunately, this has been shown to be a time- [18], subjects are asked to ‘think aloud’ while solving a
consuming and inefficient process [2,14,15] and can offend problem or making a decision. These verbalizations are
the expert as being a ‘waste of time’ [13]. The difficulties of usually taped and then transcribed and the transcription is
the unstructured interview become apparent when one analyzed using a particular coding scheme. The transcript
views a sample from an actual interview and sees how itself is termed a ‘protocol’ and may be used to refer to a
inefficient it can be. (See Table 3). word-for-word record or a summary of the major points.
Whatever the form of the protocol, it should enable the
knowledge engineer to easily access, index, and code
3.2. Structured interviewing techniques specific pieces of information.
Depending on the problem domain it may be desirable to
Recognizing that unstructured interviews are inefficient, also generate ‘motor’ protocols or even ‘eye-movement’
researchers in the area of psychotherapy have been protocols to more clearly understand an expert’s perform-
developing structured interviewing techniques for many ance of a task [10]. Motor protocols require that the expert’s
years [16]. Basically, they provided structure by developing physical movements be closely observed and noted by the
a carefully pre-planned series of ordered questions. From knowledge engineer, which may appropriate for acquiring
this work, psychologists developed other interviewing certain types of expertise. At an even more subtle level,
techniques and tools which were designed to structure the noting the movement and visual focus of the eyes of the
interview process. This work has been generally applied to expert as a task is being performed may reveal something of
the knowledge elicitation problem. These techniques can the sensory experience of the expert as he/she performs the
often be applied to situations where the expert is being task [10].
interviewed while actually performing a task or where the However, all protocols can be classified as being either
task is simulated or reconstructed by case studies or ‘concurrent’ or ‘retrospective’ [19]. Concurrent protocols
scenarios or simply from the expert’s own past experience. are records of the expert’s thought processes at the same
Elicitation techniques most commonly discussed in the time he/she is solving a problem while retrospective
literature include protocol analysis [12,17,18], repertory protocols are records of the expert’s review of his/her
grids [4], prototyping [19,20], multidimensional scaling [4, verbalizations after the task is completed. These are often
21], cluster analysis [14], event recall [2], discourse analysis used when it is felt that the task of verbalization has
[51], and card sorting [15]. interfered with the expert’s performance of the actual task
Some rudimentary structuring can be given to the [19]. For example, an expert in quality control for autos may
interview process by having the expert perform a particular be asked to think aloud as he/she goes through the auto
task while the knowledge engineer asks questions freely. inspections process and the resulting recorded protocol
The task may be typical of the problem-solving situation might form the basis for the next phase of knowledge
which the knowledge engineer wishes to explore or it may acquisition. The transcript may in turn be translated by the
be a special case, identified in earlier sessions which the knowledge engineer into a more formal protocol which
W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447 443

attempts to summarize the major points in a format designed problem-solving process. These techniques capitalize on
for easy access (e.g. using indexing, notation, or special the idea that the knowledge engineer must become some-
coding systems). what of an expert in order to successfully translate the
Once the protocol has been worked into the desired expert’s knowledge into a machine representation. Thus the
format, the actual analysis of such a protocol by a interview may be treated as a tutorial where the expert
knowledge engineer can begin. The usual method knowl- delivers a lecture which the knowledge engineer may
edge engineers use is something called ‘process tracing’ paraphrase or use to solve similar problems [22,23]. The
drawn from cognitive psychology [9]. Analysis involves knowledge engineer may become even more actively
breaking down the decision rules used by the expert into involved by playing the role of an apprentice or otherwise
typical, naturally recurring decision rules. These can then be participating in the expert’s problem-solving process [22].
refined further by either the domain expert or another, Making the knowledge engineer become like the expert is
external expert before and after they are implemented in the certainly the most time-consuming approach to knowledge
final system. elicitation but ensures the highest quality resulting system.
Protocol analysis has become popular as an elicitation The inherent difficulties of requiring the knowledge
tool because it forces the expert to focus on a specific task or engineer to learn all the expertise in order to translate it
problem without interruptions from the knowledge engin- into a suitable machine representation are what have
eer. It forces the expert to consciously consider the problem- motivated much of the work at designing expert system
solving process and so may be a source of new self- shells with more naturalistic interfaces which enable the
understanding. It is also very flexible in that many different expert to enter his/her expertise directly into the system.
types of tasks (simulations, special cases, etc.) may serve as
a basis for the protocol. Having a record encourages the 3.4. Psychological scaling
knowledge engineer to identify specific topics and also
missing steps in the process. It has been successfully applied Other interviewing techniques that have been proposed
to developing expert systems [2] and early results of in the literature have been drawn directly from psychology.
comparative experiments show that it is more efficient than These include multidimensional/psychological scaling, net-
unstructured interviewing [15], although the same set of work scaling, discourse analysis, cluster analysis, and card
experiments shows clearly that it is less efficient than other sorting. Many of these techniques combine elicitation and
non-traditional knowledge acquisition methods such as structuring aspects and thus are difficult to consider as
card-sorting and goal decomposition. Also, on a practical simply ‘elicitation’ or ‘structuring’ techniques. Originally
level, protocol analysis requires little equipment or special researchers thought that such techniques such as these
training for the knowledge engineer. would be more objective than more traditional interviewing
The main disadvantage to protocol analysis is the very methods [14]. One empirical comparison of elicitation
necessity of forcing the expert to verbalize his/her actions. It techniques supports this contention somewhat in that it
is often the case that expertise has become so proceduralized found that non-traditional techniques such as card sorting
that the expert is either unable to express it or is completely and goal decomposition performed better than protocol
unaware of it. This phenomenon is more commonly referred analysis and interviewing [15].
to as the ‘paradox of expertise’ [2], and is one of the major A number of different techniques fall under the heading
motivations for research in the field of knowledge of what may be called ‘psychological scaling’ techniques.
acquisition. Not only they may be unaware of their These include multidimensional scaling (MDS), network
problem-solving methods, but they may actually verbalize scaling, and hierarchical cluster analysis. Generally speak-
them incorrectly and thus introduce error or bias into the ing, experts are asked to rate the similarity of different
resulting system. Especially when special or difficult test objects (usually chosen prior) and this rating is portrayed as
cases are used as cues the expert may experience a distance on a seven-point scale ranging from no similarity
considerable discomfort in trying to verbalize the pro- to completely similar. The purpose of this is to discover the
blem-solving process. Thus the appropriateness of protocol expert’s rank ordering of objects within a problem domain.
analysis may depend heavily on the type of task being Probably the most complicated method in this group is
studied and the personality and ability of the expert to be MDS. It is based on the use of the least squares method of fit
introspective and verbalize thought processes. Protocols can the elicited data and has been in use for quite a long time
also be very time-consuming to generate and may result in [24]. Experts evaluate the similarity of a set of key factors or
more data than the knowledge engineer can efficiently objects using a numbered scale along different dimensions,
handle; this is especially true of larger knowledge generating a grid of numbers. This is supposed to give an
acquisition tasks. overall picture of the problem space. The location of the
While protocol analysis involves little interaction objects in the different dimensions is then inferred using the
between the expert and the knowledge engineer, several least squares method. This requires that both the objects and
elicitation techniques have been suggested which require dimensions be identified beforehand and that they should be
the knowledge engineer to actively participate in the representative of the larger problem domain without
444 W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447

contradictions. Thus a considerable amount of knowledge is but it needs to be conducted in a more systematic and
already embedded in the selection of factors or objects and rigorous manner [31].
dimensions on which they are to be evaluated, so this Previous researchers have recognized the need for sound
technique may more appropriate for structuring previously empirical research to compare the effectiveness and
acquired knowledge. efficiency of KA tools and methods. Fellers [32] concluded
In this sense it can be considered to be an elicitation tool that more research was needed to answer the following:
much as the Analytic Hierarchy Process [25] elicits data
about objects on a hierarchy of dimensions. Even though 1. Is there one best elicitation technique for knowledge
Cullen and Bryman’s survey [12] found this to be the least acquisition?
used knowledge elicitation technique, MDS has been used 2. If not, what is the best combination of techniques?
extensively by researchers [14,26] and is well-suited to use 3. Which techniques are most suitable under which
in automated knowledge acquisition programs such as circumstances?
AQUINAS , KITTEN , KSSO , PATHFINDER , and PLANET [4]. 4. What skills are required in order to utilize each of these
techniques?
3.5. Card sorting
One KA researcher [19] designed an experiment to test
the ability of three different KA methodologies to elicit
Card or concept sorting techniques are also used to help
different types of heuristics. The three methods tested were
structure an expert’s knowledge. As its name implies, the
scenarios, simulated different tasks, and actual familiar
knowledge engineer write the names of previously ident-
tasks. Heuristics were divided into two categories: those
ified objects, experiences and rules on cards which the
that-all subjects identified regardless of knowledge acqui-
expert is asked to sort into groups. The expert describes for
sition method and those that only individual subjects
the knowledge engineer what each group has in common
identified. These are further broken down as conceptual,
and the groups can then be organized to form a hierarchy.
operational, and logistical heuristics. Overall, she found a
Like MDS, some empirical research [15] suggests that card
30% overlap in the heuristics generated by each of the
sorting may be a more efficient elicitation technique than
knowledge acquisition methods she tested. Of the heuristics
some of the more traditional techniques such as protocol
that did not overlap, she identified conceptual, logistical,
analysis or interviewing. It has been used with some success
and operational heuristics that were distinct to each method.
to develop applications described in the literature [27,28]. It
But given that the task studied (piloting a boat in a harbor)
has also been suggested that it is a tool which could be easily
was operational in nature her results were not surprising.
implemented on a computer as an automated knowledge
Adelman [5] varied the domain experts, the elicitation
acquisition tool [9].
methods, and the knowledge engineers in an attempt to see
which if any had the greatest effect on the quality of the
3.6. Empirical research on knowledge acquisition knowledge base. Five of the six knowledge engineers had
techniques PhD, degrees and one was ABD, but all had extensive
training in both top-down and bottom-up elicitation
Work on the knowledge acquisition problem currently techniques. The relative accuracy of each was compared
follows along three major, interlocking lines. We describe to a ‘golden mean’ rule set derived prior to the elicitation
these as technique-oriented, empirical studies, and concep- sessions. Although a long line of psychological research has
tual research. As has been noted in the literature [4,29], the been devoted to describing interviewer effects which are
primary emphasis to date has been on developing new analogous to the potential effects of a knowledge engineer
knowledge acquisition tools and methods. This article [33], no significant effects were observed in this set of
focuses on examining the impact of recent empirical work. experiments. Interestingly, the only significant source of
A review of the KA literature shows that both conceptual variation came from the domain experts themselves.
and empirical research has lagged behind technique- The best-known experimental research on KA methods is
oriented research. Experiments and case studies have that of Burton et al. [15]. By varying the different
focused on comparing and evaluating knowledge acqui- knowledge acquisition techniques among different groups
sition techniques. However, the empirical work has suffered of experts, each of whom was tested for cognitive style, they
from a general lack of control and precision [29]. discovered several specific things. Among their findings
There have been a few recent efforts to empirically test was that protocol analysis took the most time and elicited
the usability of different knowledge acquisition tools and less knowledge than the other three techniques they tested
techniques. [15] tested the ability of various knowledge (interviewing, card sorting, and goal decomposition). Not
elicitation methods to elicit knowledge about classifying surprisingly, they also found that introverts needed longer
different rocks. And [30] compared the relative efficiency of interview sessions but generated more knowledge than
several automated knowledge acquisition tools. Such extroverts. Interestingly, the rarely used techniques of goal
research is important because it serves to break new ground, decomposition and card sorting proved to be as efficient as
W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447
Table 5
Summary of empirical KA research (Adapted from Ref. [29])

Studies KA techniques Mod. Vars. P. Domain Dep. Vars. Results

[38] Interviewing Inductive learning Not considered Diagnosis Percentage of correct diagnosis generated Inductive learning performed better than interviewing
[17] ID3 Induction and interviews Not considered Diagnosis Comparison to known cases Induction performed much better than interviewing
[53] Interviews protocol analysis expert Human vs. recon- Interpretation KE’s opinion of the quality of K acquired Protocol analysis had limited usefulness for certain types of
walk-throughs structed knowledge K
sources
[15] Interviews protocol analysis goal Introvert vs. extrovert; Classification Time taken for capturing K; time for coding into Protocol an. takes longer and yields less K; Introverts need
decomposition card-sorting multi- cognitive styles rules; number of rule clauses; completeness of longer interviews but generate more K than extroverts
dimensional scaling rule set
[35] Interviewing and protocol analysis Domain complexity Planning Efficiency and quality of K as measured by Interviewing is more efficient and accurate for simple cases
number of nodes and arcs and their accuracy but protocol is more efficient for complex cases
[8] Structured interviews protocol analysis Expert vs. non-expert; Classification Efficiency of process Protocol analysis performed poorly in classification domain;
card sorting laddered grids two classification card sorting and grids performed better than interviewing;
domains external validation of experts important
[8,19] Scenarios, simulated different tasks, and Not considered Command Overlap of heuristics Found 30% overlap between heuristics elicited by the 3
actual familiar tasks and control different methods
[5] top-down vs. bottom-up interviewing Knowledge engineer, command Accuracy of elicited rules as compared to Found no sig. Variation except for that due to domain expert
and domain expert and control ‘golden mean’ set

445
446 W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447

the more common interviewing technique and more efficient KA sessions, when the problem complexity is not yet
than the commonly used protocol analysis. developed clearly.
This experiment was criticized somewhat for its lack of For those studies that did consider the effect of moderator
rigor [29,31]. One measure of technique efficiency was the variables, it seems clear that no matter what type of problem
time it took to code the transcripts into pseudo-rules while domain, developers of expert systems in the fields of
the number of rules or clauses was taken as a measure of accounting and finance should consider their potential
acquired knowledge. Coding time does not fully account for impact. The impact of the cognitive style of the expert,
temporal differences among KA methods and there are also domain complexity, along with other attributes of the
serious drawbacks to using the number of coded rules as a domain expert all seem to be important factors in the quality
measure [29,34]. The results may also have been con- of an expert system regardless of the problem domain. It is
founded by unmeasured differences among the experts and hoped that further research will clarify some of these issues
the knowledge engineers. with respect to the effect of moderator variables and
These various experimental studies are symptomatic of a problem domains.
recognized need empirical investigation of KA phenomena.
The small number of such studies is at least in part,
indicative of the difficulty in conducting them. The few References
pioneering studies are typified by confusing terminology,
conflicting operationalizations, and the proliferation of ad [1] O. Owrang, F. Grupe, Database tools to acquire knowledge for rule-
hoc taxonomies. In addition, results are conflicting and no based expert systems, Information and Software Technology 39
(1997) 607 –616.
clear pattern has emerged. There are problems controlling
[2] R. Hoffman, The problem of extracting the knowledge of experts from
for effects of moderator variables and in operationalizing the perspective of experimental psychology, AI Magazine 8 (2) (1987)
the measurement of dependent variables. In light of these 53– 67.
problems, [29] concluded that empirical KA work should [3] B. Le Roux, Knowledge acquisition as a constructive process: a
concentrate on case studies rather than experiments, at least methodological issue, Decision Support Systems 18 (1) (1996)
33– 43.
in the near term. A strategy for addressing some of these
[4] J. Boose, A survey of knowledge acquisition techniques and tools,
experimental obstacles has also been proposed [31]. A fairly Knowledge Acquisition 1 (1) (1989) 3–38.
comprehensive summary of empirical KA research [5] L. Adelman, Measurement issues in knowledge engineering, IEEE
(Adapted from Ref. [29]) is presented as Table 5. Transactions on Systems, Man, Cybernetics 19 (1989) 483– 488.
[6] T. Shaft, I. Vessey, The relevance of application domain knowledge:
the case of computer program comprehension, Information Systems
Research 6 (3) (1995) 286–299.
4. Conclusions [7] W. Clancey, Heuristic classification, Artificial Intelligence 27 (1985)
298 –350.
[8] A. Burton, R. Schweickert, N. Taylor, E. Corlet, N. Shadbolt, A.
The application of empirical KA research to the problem
Hedgecock, Comparing knowledge elicitation techniques; a case
of choosing an appropriate KA technique for developing an study, Artificial Intelligence 1 (14) (1990) 245–254.
expert system application in the fields of accounting and [9] K. McGraw, K. Harbison-Briggs, Knowledge Acquisition: Principles
finance suggests several directions. First, if one is going to and Guidelines, Prentice-Hall, Englewood Cliffs, NJ, 1989.
be working in an analytic problem domain such as [10] G. Tuthill, Knowledge Engineering: Concepts and Practices for
Knowledge-Based Systems, Blue Ridge, PA, TAB Books, 1990.
classifying asset writedowns. KA techniques that provide [11] J. Kim, J. Courtney, A survey of knowledge acquisition techniques
a high degree of structure to the interviewing process seem and their relevance to managerial problem domains, Decision Support
to work best. Protocol analysis, though fairly commonly Systems 4 (3) (1988) 269–285.
used, is relatively inefficient for analytic problems while the [12] J. Cullen, A. Bryman, The knowledge acquisition bottleneck: time for
most popular technique of using an unstructured interview- a reassessment? Expert Systems 5 (3) (1988) 216–224.
[13] D. Forsythe, J. Buchanon, Knowledge engineer as anthropologist,
ing is one of the least efficient and least satisfying from the IEEE Transactions on Systems, Man and Cybernetics (1989) 3.
standpoint of the expert. So it may be worth exploring some [14] N. Cooke, J. McDonald, The applications of psychological scaling
of the non-traditional KA techniques when working on these techniques to knowledge elicitation for knowledge-based systems,
type applications. International Journal of Man Machine Studies 26 (1987) 81–92.
For the more difficult synthetic and combination problem [15] A. Burton, N. Shadbolt, A. Hedgecock, G. Rugg, A formal evaluation
of knowledge elicitation techniques for expert systems: domain 1,
domains the evidence is not as clear. However, the Proceedings of the First European Workshop on Knowledge
Holsapple and Raj study [35] seems to indicate that problem Acquisition for Knowledge-Based Systems, September, Reading
complexity may be one determinant of the appropriate KA University, 1987.
technique to choose. So if one were to develop an expert [16] R. Merton, M. Fiske, P. Kendall, The Focused Interview, The Free
system for estimating control risk then we might suppose Press, Glencoe, IL, 1956.
[17] A. Hart, Experience in the use of an inductive system in knowledge
that protocol analysis might be more efficient than engineering, in: M.A. Bramer (Ed.), Research and Development in
interviewing. The fact that interviewing is more efficient Expert Systems, Cambridge University Press, Cambridge, 1985, pp.
for simple domains may imply that it is best used for initial 117 –126.
W.P. Wagner et al. / Knowledge-Based Systems 15 (2002) 439–447 447

[18] A. Newell, H. Simon, Human Problem Solving, Prentice-Hall, Development: The Case of Asset Writedowns, International Journal of
Englewood Cliffs, NJ, 1972. Intelligent Systems in Accounting, Finance and Management 3 (3)
[19] M. Grabowski, Knowledge Acquisition Methodologies: Survey and (1994) 187–203.
Empirical Assessment, Proceedings of ICIS, 1988. [37] A. Houghton, M. Gardner, Expert Systems at the Customer Interface,
[20] D. Waterman, A Guide to Expert Systems, Addison-Wesley, Reading, International Journal of Intelligent Systems in Accounting, Finance
MA, 1986. and Management 1 (4) (1992) 289–296.
[21] L. Elliot, Analogical problem solving and expert systems, IEEE [38] R. Michalski, R. Chilausky, Knowledge acquisition by encoding
Expert 1 (2) (1986) 17– 28. expert rules versus computer induction from examples—a case study
[22] M. Welbank, Knowledge Acquisition: A Survey and British Telecom involving soybean pathology, International Journal of Man-Machine
Experience, Proceedings of the First European Workshop on Knowl- Studies 12 (1980) 63–87.
edge Acquisition for Knowledge-Based Systems, Reading University, [39] A.S. Hollander, Using Rules to Model Variance Analysis, Expert
UK, 1987, pp. C6.1–C6.9. Systems Review 2 (3) (1990) 46–52.
[23] L. Johnson, N. Johnson, Knowledge elicitation involving teachback [40] Vanguard Web Site (http://majestic.vanguard.com/guide/da), May
interviewing, in: A. Kidd (Ed.), Knowledge Elicitation for Expert 2000.
Systems: A Practical Handbook, 1987. [41] G. Ribar, Expert systems technology at peat Marwick Main, Expert
[24] J.B. Kruskal, Multidimensional scaling and other methods for Systems Review for Business and Accounting 1 (1) (1987) 1–5.
discovering structure, in: Enslein, Ralston, Wilf (Eds.), Statistical
[42] H. Schwartz, Expert systems in accounting, CPA Journal Online April
Methods for Digital Computers, Wiley, New York, 1977.
(1989) .http://www.nysscpa.org/cpajournal.
[25] T. Saaty, The Analytical Hierarchy Process, McGraw-Hill, New York,
[43] M. Shaw, J. Gentry, Using an expert system with inductive learning to
1981.
evaluate business loans, Financial Management (1988).
[26] K. Butler, J. Carter, The use of psychologic tools for knowledge
[44] C. Brown, N. Nielson, M. Phillips, Expert systems for personal
acquisition: a case study, in: W.A. Gale (Ed.), Artificial Intelligence
financial planning, Journal of Financial Planning 3 (3) (1990)
and Statistics, Addison-Wesley, Menlo Park, 1986, pp. 295–320.
137–143.
[27] M. Chi, P. Feltovich, R. Glaser, Categorization and representation of
[45] B. David, N. U, Automates payroll with AI system, Datamation 42
physics problems by experts and novices, Cognitive Science 5 (1981)
121–152. (17) (1996) 129–132.
[28] J. Gammack, R. Young, Psychological techniques for eliciting expert [46] D. Murphy, C. Brown, The uses of advanced information technology
knowledge, in: M. Brammer (Ed.), Research and Development in in audit planning, International Journal of Intelligent Systems in
Expert Systems, Cambridge University Press, London, 1985, pp. Accounting, Finance and Management 1 (3) (1992) 187– 193.
105–112. [47] C. Brown, M. Phillips, Expert systems for management accountants,
[29] J. Dhaliwal, I. Benbasat, A framework for the comparative evaluation Management Accounting 71 (7) (1990) 18–23.
of knowledge acquisition tools and techniques, Knowledge Acqui- [48] N. Doherty, K. Pond, An expert system solution to mortgage arrears
sition 2 (2) (1990) 145 –166. problems, The Services Industry Journal 15 (2) (1995) 267 –288.
[30] C.M. Kitto, Progress in automated acquisition tools: how close are we [49] K. Lecot, Using expert systems in banking: the case of fraud detection
to replacing the knowledge engineer? in: Banff (Ed.), Proceedings of and prevention, Expert Systems Review for Business and Accounting
the Third Knowledge Acquisition for Knowledge-Based Systems 1 (3) (1988) 17–20.
Workshop, 1988, pp. 14.1– 14.13. [50] A. Bharadwaj, V. Karan, R. Mahapatra, U. Murthy, A. Vinze, APX:
[31] C. Holsapple, V. Raj, W. Wagner, Knowledge acquisition: recent an integrated knowledge-based system to support audit planning,
theoretic and empirical developments, in: C. Holsapple, A. Whinston International Journal of Intelligent Systems in Accounting, Finance,
(Eds.), Recent Developments in Decision Support Systems, Springer, Management 3 (3) (1994) 149–164.
Berlin, 1993. [51] N.J. Belkin, H.M. Brooks, P.J. Daniels, Knowledge Elicitation Using
[32] J. Fellers, Key factors in knowledge acquisition, Computer Personnel Discourse Analysis, International Journal of Man– Machine Studies
11 (1) (1987) 10–24. 27(2) (1987) 127–144.
[33] K. Hammond, Subject and sampling—a note, Psychological Bulletin [52] S. Eom, A survey of operational expert systems in business (1980–
45 (1948) 530– 533. 1993), Interfaces 26 (5) (1996) 50 –70.
[34] C. Holsapple, A. Whinston, Manager’s Guide to Expert Systems, Dow [53] W.F. Messier, J.V. Hansen, Expert Systems in Auditing: The state of
Jones-Irwin, Homewood, IL, 1986. the art, Auditing: A Journal of Practice and Theory 7 (1) (1987) 94 –
[35] C. Holsapple, V. Raj, An exploratory study of two KA methods, 105.
Expert Systems 11 (2) (1994) 77–87.
[36] S. Ragothaman, B. Naik, Using Rule Induction for Expert System

Potrebbero piacerti anche