the reuse of knowledge: a user-centred approach

Int. J. Human-Computer Studies (2000) 52, 553}579doi:10.1006/ijhc.1999.0342Available online at http://www.idealibrary.com on

The reuse of knowledge: a user-centred approach

DEBBIE RICHARDS

Department of Computing, Division of Information and Communication Sciences,Macquarie University, Sydney, Australia. e-mail: [email protected]

(Received 20 July 1999 and accepted in revised form 5 August 1999)

The motivation for the work reported in this paper is the belief that not only is itbene"cial to reuse knowledge but it is essential if we wish to build knowledge-basedsystems (KBS) that meet the needs of users. The focus of most KBS research is oncomplex modelling at the knowledge level which requires a knowledge engineer to act asthe intermediary between the expert and the system. The type of reuse primarilyconsidered is the reuse of ontologies or problem-solving methods so that improvementscan be made in system quality and development time. However, there is little focus on theneeds of users to access the knowledge in a variety of ways according to the individual'sdecision style or situation. The system described in this paper seeks to support the user ina number of di!erent activities including knowledge acquisition, inferencing, mainten-ance, tutoring, critiquing, &&what-if '' analysis, explanation and modelling. The ability toask di!erent types of questions and to explore the knowledge in alternative ways isa di!erent type of knowledge reuse. The knowledge acquisition and representationtechnique used as the foundation is known as ripple-down rules (RDR). To support theexploration activities, RDR have been combined with formal concept analysis whichautomatically generates an abstraction hierarchy from the low-level RDR assertions. Thepaper suggests that rapid and incremental KA together with retrospective modelling canbe used to provide the user with a system that they can own, build and explore withoutthe di$culties associated with capturing and validating the conceptual models of expertsvia the mediation of a knowledge engineer.

( 2000 Academic Press

1. IntroductionThe motivation for the work reported in this paper is the belief that not only is itbene"cial to reuse knowledge but it is essential if we wish to build knowledge-basedsystems (KBS) that meet the needs of users. Due to the user-centredness of the approach,the type of reuse being sought is di!erent from the type of reuse considered in mostknowledge reuse research. Reuse can range over a number of dimensions (Menzies, pers.comm.) including the following.

d Reuse by the same/di!erent person.d Reuse over the same/di!erent task.d Reuse for the same/di!erent purposes or activities.d Reuse at the same/di!erent time.d Reuse over the same/di!erent software application.

Most knowledge reuse research is focused on reuse of the knowledge for the same task indi!erent applications. The focus in this paper is on the reuse of knowledge for di!erent

1071-5819/00/030553#27 $35.00/0 ( 2000 Academic Press

554 D. RICHARDS

activities. A more appropriate term for the type of reuse of interest is adapting orrepurposing, a term taken from the hypermedia "eld. Based on a situated view of expertknowledge and action and an emphasis on user-requirements, the goal of this work wasto see if a knowledge-based system could be built that allowed the user to select froma wide range of activities using and reusing knowledge according to their current purposeand decision situation. Activities are seen as related to a purpose with each activitydesigned to answer di!erent types of questions relating to the knowledge. An activity isseen as distinct from a task because a task in knowledge engineering (KE) is typicallyequated with a problem type such as diagnosis, planning or con"guration. To di!erenti-ate a task and activity further, we can view a task as related to making a decision in theworld. An activity is not concerned so much with the world but with the individual (orgroup) and how they arrive at a decision. A task such as room allocation could beachieved through di!erent activities such as consultation, learning about the domain,proposing a solution to have critiqued or exploring di!erent scenarios. To avoidconfusion with more conventional meanings of reuse, the type of reuse being pursued isreferred to as activity-reuse. While some have looked at building systems that performone or a small number of activities, which could include knowledge acquisition (KA),consultation, maintenance, explanation, critiquing or &&what-if '' analysis, this work isconcerned with providing all of these activities and more within the one session becauseuser requirements will di!er between users and in di!erent situations for the same user.

In addition to a di!erent focus on the type of knowledge reuse, another distinguishingfeature of this work is a breakaway from mainstream approaches to KBS development,which emphasize the need to build terminological KBS (T-boxes-) at the knowledgelevel. Such approaches require complex analysis and modelling of the ontological,problem-solving method and/or domain knowledge as prerequisites to system develop-ment. Developing complex models is time-consuming and the emphasis on capturinga &&good'' and comprehensive model before knowledge acquisition can begin contributesto the knowledge acquisition bottleneck. The di$culties associated with developingontologies or problem-solving methods is due to the well-recognized fact that muchexpert action is sub-conscious and not easily described (Ignizio, 1991) and the inherentlyunreliable nature of models. A descriptive model can be seen as: &&merely an abstraction,a description and generator of behaviour patterns over time, not a mechanism equivalentto human capacity'' (Clancey, 1993, p. 89). It has been shown that models are imperfectrepresentations that vary not only between experts but also over time with the sameexpert (Gaines & Shaw, 1989).

The work described in this paper is based on a paradigm that captures a simple modelin terms of attribute-value pairs and the related conclusions. The simple model iscaptured into an assertional (A-box?) KBS which is a performance system that can beexecuted and validated. In keeping with a situated view of cognition, this paper describesa KBS that supports human}computer interaction in modes that seem to be more inkeeping with the way a human may perform that activity. Just as most human action is

-Terminological KB consist of terms structured into inheritance networks (Brachman, 1979). Their mainbuilding blocks are concepts and roles and they reason by determination of subsumption between concepts(Nebel, 1991).?Assertional KBS are made up of executable assertions (such as rules) that assert the relationships between

terms (such as conditions and conclusions).

THE REUSE OF KNOWLEDGE 555

re#exive (Winograd & Flores, 1986), the system described in this paper performsknowledge acquisition, maintenance and inferencing in a re#exive- manner without theneed for re#ective action on the part of the expert. This is not to say that high-levelmodels are not necessary. Models of abstraction are bene"cial for teaching (Schon, 1987)and for explanation (Clancey, 1993). Therefore, support for re#ective activities, such asexplanation, tutoring, &&what-if '' analysis and causal modelling, has also been explored.

The knowledge acquisition and representation technique used as the foundation of thework reported is known as ripple-down rules (RDR) (Compton & Jansen, 1990). RDRhave been built from a situated standpoint which acknowledges that the capture ofa performance model may be easier and more reliable than reliance on acquiringa complete descriptive model. It was the goal of this research to see if knowledgecaptured for one purpose or activity such as KA could be reused for another activity suchas critiquing with only changes to the user interface. Prior to this study, RDR couldsuccessfully (Edwards, Compton, Malor, Srinivasan & Lazarus, 1993) handle a numberof activities such as KA, inferencing, maintenance and validation all within the samesystem. These activities are performed re#exively by the expert with minimal considera-tion of the model being built (Compton, Kang, Preston & Mulholland, 1993). RDRdevelops rules by going straight from data to conclusions without the intermediate stepof building an abstract model of the data (see the inference structure of heuristicclassi"cation given in Clancey, 1985). All rules in an RDR KBS consist of a "nalconclusion and a set of attribute-value pairs which are the rule conditions. The close linkbetween the use of RDR and the goals of this research meant that the activities alreadysupported required minimal further research e!ort. However, higher-level models werenecessary for the re#ective modes such as critiquing, &&what-if '' analysis and explanation.It was apparent that the key to supporting these activities was an understanding of theconcepts in the KBS and the relationships between them. Of particular importance wasthe ability to "nd the higher-level concepts which were implicitly represented by thelow-level concepts in the RDR primitive rules. Formal concept analysis (FCA) (Wille,1982) has been employed not only to uncover the higher-level abstractions in the RDRKBS but also the structure between all concepts in the KBS.

In the next section, the bene"ts of reuse at a general level beyond activity reuse areconsidered. Firstly, we look at the material or quantitative bene"ts of reuse followed byconsideration of the more soft bene"ts of building systems that people want and are ableto reuse.

1.1. THE BENEFITS OF REUSE

The reuse of software components has become widely accepted due to the followingfactors.

1. Savings in cost.2. Savings in time.3. An increase in reliability (Hemmann & Voss, 1993).

-The terms re#exive and re#ective are used in this paper to distinguish between acts that are automatic andoften subconsciously performed and acts that require deliberative thought often involving the consideration ofabstract models, respectively.

556 D. RICHARDS

There are two main methods of software reuse: using what you already have as buildingblocks or using an existing pattern to create new code. Hemmann and Voss (1993) termthese approaches composition and generation, respectively. The former is the morewidely accepted practice and appears to be more natural for software engineers. In theknowledge engineering community the situation is similar, with much research beingcarried out into developing building blocks (e.g. Chandrasekaran, 1986; Steels, 1993) thatcan be reused for di!erent applications. However, in both software and knowledgeengineering reuse research, the focus is on the technical person, the software or know-ledge engineer (KE). The focus in this study is on the end-user or the expert who is oftennot directly assisted by the building block approach.

There is much similarity between the approach to reuse in software engineering andknowledge engineering. The focus in software engineering on the reuse of programs andprocedures is similar to the focus on the reuse of problem-solving methods in knowledgeengineering. Also the bene"ts of reusing the representation language and the method ofinference for di!erent applications are obvious and much research focuses on theseaspects (Clancey, 1992). However, a valuable lesson from software engineering is thatdi!erent languages will a!ect whether and how that component can be reused. Analog-ously, in knowledge engineering, di!erent KA and knowledge representation techniquesmay a!ect the reusability of that knowledge (Chandrasekaran & Johnson, 1993). Men-zies summarizes what he believes to be the major potential and actual bene"ts of reuse ina very interesting paper which makes comparison between the lessons learnt from ESsoftware development and the use (and reuse) of patterns- in the object-orientedcommunity. Menzies (1997) describes the three potential bene"ts of reuse of patterns: thereuse bene"t which lets a designer build systems more e$ciently using proven oldsystems; the guidance bene"t that provides insights discovered in other systems and thecommunication bene"t where patterns are useful in explaining existing systems. Menzies(1997) sees that adoption of the knowledge level (KL) (Newell, 1982) has brought with itthe bene"ts of a more structured approach that has resulted in better organized projectswhich have seen some industrial success. These are guidance and communicationbene"ts. Menzies (1997) doubts that the reuse bene"t of e$ciency has yet been achieved.The KL approach is superior to the earlier transfer-of-knowledge approach and acknow-ledges that knowledge is a model and not an artifact. However, the amount of e!ortbeing put into getting the model right, in such forms as PSMs or ontologies, appears totreat the model as an artifact and does not adequately acknowledge the de"ciencies ofmodels. A situated view of knowledge does not support such as emphasis on developinggood models but demands that systems are grounded in the real world and supportincremental change.

Software components are not the only resources that can bene"t from reuse. Data areconsidered to be a major resource. The explosion of the information age con"rms itsvalue. Data warehousing is an important and necessary development of databasemanagement systems which allows organizations to better manage their data and reducethe duplication and inconsistency of the data. Data can also be captured once and reusedfor a whole range of purposes. Knowledge is value-added data and includes such things

-A fragment of a high-level conceptual model which may be useful in many applications (Menzies, 1997j)oo-patterns


as experience, insight and skill. The use of data dictionaries to support the reuse of data issimilar to the use of common ontologies for describing knowledge. The work done byJansen and Compton (1988) on the knowledge dictionary is an even closer developmentof the data dictionary idea.

The bene"ts of reusing knowledge, which includes both domain and problem-solvingknowledge, should be even greater than the bene"ts of reusing data because of itsincreased value and the problems associated with acquiring knowledge (Ignizio, 1991).By capturing expert knowledge once and reusing it a number of times, better use can bemade of this scarce resource and the knowledge acquisition bottleneck can e!ectively bereduced.

The bene"ts of reuse for the user are not so easy to quantify or describe, but are seen tobe substantially greater than the material bene"ts outlined above. To justify this claim,we need to consider the human}computer interaction (HCI) issues for KBS and the typesof systems users' need.

1.2. HUMAN COMPUTER INTERACTION ISSUES FOR KBS

The acceptance and success of a KBS is not achieved by merely acquiring knowledge andcorrectly deriving inferences using it. It has been shown that even where the knowledge isaccurate and reliable, the expertise captured becomes an underutilized and wastedresource if the knowledge cannot be accessed by the user in an acceptable way (Langlotz& Shortli!e, 1983). An indepth study by Salle and Hunter (1990) considers the lack ofattention to computer and user cooperation issues to be the main reason for the pooracceptance of ES technology by end-users. Cooperation includes the user interface but ismuch more. The user interface is concerned with usability, whereas the mode ofinteraction is concerned with usefulness (Rector, 1989). The term used by Salle andHunter is modality and is de"ned as: &&referring to di!erent forms the cooperation mayhave regardless of agents being human or electronic'' (De Greef, Breuker & De Jong,1988).- Collaboration is important because humans and computers have oppositeabilities that complement each other. Cooperation aims at man}machine synergy.

Salle and Hunter (1990) consider that most KBS act as a prosthesis rather thansupport. Due to the subjective and changing nature of many domains such as medicine,they consider the prosthesis approach to be unsuitable. Users need more than prescrip-tive systems, they need to assess the answer through better explanation and queryfacilities. Expert systems have been designed from the machine's viewpoint and lack thescope and robustness needed by users with a wide range of needs. They state &&the KBSshould be able to respond to a large variety of interaction, not only the commonlysupported question'' (Salle & Hunter, 1990, p. 6) including the following.

1. What would happen if ....?2. Why would X happen.....?3. How could X be prevented.....?4. What are the critical factors in X.....?

- In this paper, the term activity is used in preference to the word mode due to the possible negativeconnotations associated with modal systems (Tesler, 1981) where, for example, the user may become confusedwhether they are in browse or update mode.

558 D. RICHARDS

5. What other solutions would work.....?6. Isn't solution X as e!ective.....?7. Which is the best remedy.....?8. Is X the right remedy.....?9. Why did remedy X work.....?

10. Why didn't remedy X work.....?

A good explanation system should handle all the &&why'' questions (numbers 2, 9 and 10).In addition causal modelling, which can be considered as a type of explanation, answersthe "rst question. Causal modelling could also be adapted to answer questions 9 and 10 ifthe remedies were made part of the system or a fault and a remedy model were developed.Questions 1 and 5 are addressed by a &&what-if '' exploration system. Critiquing systemsanswer question 8. A teaching system should let the user seek out the knowledge neededto answer the other questions which relate to alternatives and choices and improvedunderstanding of the domain.

Kidd and Sharpe (1987) go so far as to say that many ES are not useful once ina commercial situation. More understanding of tasks and a theory of cooperativeproblem solving between man and machine is needed. They argue that current systemsdo not really solve the users' problem. Users not only want to know &&what is the fault'' or&&what is the remedy'' but they also want to ask questions and negotiate a remedy. Thesystem described in this paper allows the user to ask a variety of questions by accessingthe knowledge in di!erent ways.

Stelzner and Williams (1988) emphasize the importance of the user interface for theacceptance of expert systems. One common interface in KBS is the consultation system,where the user is asked to answer certain questions and is given a recommendation at theend. Clancey (1992) considers the typical consultation mode of ES to be &&super"cial''because the line of questioning approach does not allow the user to build a model thatthey can manipulate.

Baro!, Simon, Gilman and Schneiderman (1988, p. 100) consider question}answerstyle dialogues to be inappropriate for user control, visibility and user initiative. In "eldslike medicine and electrical engineering, graphical displays, such as visual causal dia-grams (VCDs), are used for both the input of data and the display of results. This couldbe classed as a direct manipulation interface, which Baro! et al. (1988) de"ne asinteraction via the mouse only, without the use of the keyboard, menus or commands.Direct manipulation reduces the knowledge requirements necessary to use the computer,letting the user concentrate on the task. Baro! et al. (1988) argue that such a systemprovides a better "t with the user's perception of the domain because they are usuallyeasier to learn, less computer obtrusive and may allow &&what-if '' exploration. Sucha dynamic &&world-of-action'' interface provides instant response and update. They alsobelieve that users should be given syntactic and semantic knowledge of computers andsemantic knowledge of the domain.

Baro! et al. (1988) see graphical user interfaces as the most appropriate for the captureof knowledge because they claim people think in pictures. They also recommenddevelopment of the user interface before entering domain rules so they can be validatedon entry and the use of procedural control may be necessary to structure the knowledgeacquisition process. This idea is exempli"ed in the work of Zacklad and Fontaine (1993)


in their C-Kat system which assists users who are entering knowledge to choose the bestconclusion from a range of numerous possibilities. A "nal point made by Baro! et al.(1988) is the importance of user involvement in the design of the interface since compre-hension of the domain knowledge is dependent on the quality of the user interface.

The aspect of HCI focused on in this work is the usefulness of the system rather than itsusability. Therefore, no particular user-interface style is recommended. In fact, this workemphasizes the situated nature of ES and that systems must be built to "t the cognitivemodels of the user while providing the sort of #exibility of interaction with the know-ledge base that the user would have if the user was performing that activity withouta computer.

1.3. THE TYPE OF SYSTEMS NEEDED

From the previous sections, it can be seen that ES often do not match user requirements.A particular limitation mentioned was the consultation system style. Although thefollowing researchers propose di!erent guidelines and requirements for handling HCI inKBS, they all voice a common opinion which is the need for the computer to capture theconceptual model of the expert and for the human to have some understanding of themodel that has been captured (Gaines & Shaw, 1983; Roth, Bennett & Woods, 1987;Woods, 1987; Cleal & Heaton, 1988). Interfaces that assist the user in exploring theunderlying conceptual models is addressed in the work reported in Section 2.4 onexplanation, querying and views.

Hender and Lewis (1988) and Stelzner and Williams (1988) see a need for multipleinterfaces to be provided with KBS since di!erent users will have di!erent needsbecause their focus is di!erent. For example, a knowledge engineer is interested in howthe knowledge will be represented or retrieved and the user is interested in using theknowledge for decision making. Possible interface alternatives to consultation include&&second opinion'' system styles like critiquing or &&what-if '' analysis. From a situatedaction viewpoint, di!erent interfaces are seen as desirable even for the same individualbecause they will "nd themselves in di!erent contexts requiring di!erent solutions. Eachinterface, however, is not treated as a di!erent application but is o!ered on a di!erentscreen, sometimes just a pop-up window over the top of another screen.

Stelzner and Williams (1988) o!er the following six features that they consider will bemandatory HCI features available in a KBS.

1. Access must be via a natural idiom such as diagrams, charts, networks or textdepending on the domain and the user.

2. A format that allows manipulation and experimentation such as spreadsheets.3. Immediate feedback.4. Recoverability which allows the user to play with scenarios without penalty such as

loss of current state and actual data.5. Di!erent levels of abstraction and granularity.6. Multiple interfaces for di!erent users and situations such as an intelligent and

interactive interface for the user and another interface for the programmer.

The system described in this paper o!ers all of the above points. Point 1 is addressed bycustomizing screens to the particular application domain where necessary. For example,

560 D. RICHARDS

specialized screens have been provided in the #ight simulator, ion-chromatographycon"guration and medical causal modelling applications. The chemical pathology sys-tem described in this paper uses the standard interface which shows the pathology resultsfor numerous tests over "ve time periods as a case. Additionally, the use of object-oriented graphical user interfaces, Visual Basic for the PC or Hypercard for the Macin-tosh, makes interaction simple and intuitive. Points 2 and 4 are supported by allowing&&what-if '' analysis by altering any values in a case to see the change in outcome both ina questionnaire (like a spreadsheet) or report format. Any changes to values do not alterthe actual cases. Point 5 is addressed by the user selecting the level of granularity forviewing and the generation of labelled line diagrams using formal concept analysis whichshow di!erent levels of abstraction. Point 6 is supported by the user being able toperform inferencing, maintenance, KA, modelling, what-if analysis, critiquing or expl-anation all within the one session. The system is dynamic, point 3. In fact, unlike mostKBS approaches, even system development and maintenance are designed to be per-formed on-line by the user.

Although di!erent interfaces are available, the screens are kept as consistent aspossible. The #ow between screens is simple so as to avoid modal errors (Smith, Irby,Kimball, Verplank & Harslem, 1982) and so the user is not asked to change his mentalmodel (Norman, 1986). A click of the mouse enables the user to perform the task in whichthey are interested. Where activities overlap the same screens are used. For example,adding a rule that will be critiqued is the same as adding a rule in the knowledgeacquisition (KA) or maintenance interface. If we take the view that &&perceiving, behavingand learning are one process (Rosen"eld, 1988) then it makes sense to build a system thatdoes not separate these processes but allows the user to perform KA, "nd concepts,perform inferencing, learn about and explore the knowledge, and so on, all within the onesession. Figure 1 shows the relationship between the situation in which the users "ndthemselves and the activity they need to perform.

The focus of most KBS research on complex modelling at the KL has exacerbatedthis lack of attention to HCI issues and has resulted in a predominance of systemswhere the end-user is not a direct participant but must interact with the system viathe mediation of a knowledge engineer. This has resulted in systems designed forknowledge engineers rather than experts or end-users. It has long been recognized thatit is essential to involve users in the development of computer systems and that user

FIGURE 1. The various situations in which a user may "nd themselves are shown on the left. To solve theircurrent problem they use the features of the system that support the necessary activity on the right. The choice

of activity is situation and user-driven. The user can switch between activities as their needs change.


satisfaction is enhanced by systems owned by end-users (Schultheis & Sumner, 1992).Freidson (1994) argues that when it comes to knowledge, control is even more importantto users. KBS users, both expert and novice, need to have control of the system (Langlotz& Shortli!e, 1983; Kidd, 1985; Miller, 1986; Ignizio, 1991). Control is made possible viathe user interface (Hender & Lewis, 1988). By simplifying the KA technique and makingthe user responsible for KA and maintenance, the user is given control and ownershipand the KE is able to focus more on user requirements elicitation rather than knowledgeelicitation.

While modelling still predominates KBS research, a new breed of KBS researchers isemerging bringing an awareness of the context dependence and socially situated natureof expertise (Agnew, Ford & Hayes, 1997). Clancey (1997) and Collins (1997) argue thatwhat we do and say only makes sense in our social context. An expert cannot existwithout a social context responsible for conferring expert status. This new focus isimportant for KBS HCI because it acknowledges that human factors are essential partsof the system. A social view of knowledge prompts the question &&how can we embed ourconcepts into a computer and use them when computers do not share our social view?''Collins (1997) suggests treating the KBS as a valuable assistant where the expert takesadvantage of the machines ability to store and process more information and heuristicsthan a human mind but the inputs and outputs are blended by the expert to "t the socialcontext. A social focus requires a change from the traditional consultation style KBS towhat can be termed Expert Advisory Systems (Stelnzer & Williams, 1998). Involving thehuman in the loop, however, places a new requirement for KBS to support user cognitivetasks. This has been borne in mind in the development of the system described by makinginteraction with the system as close as possible to the way the user performs that taskmanually and is demonstrated in the medical system described later.

The next two sections describe the RDR approach, a successful "elded implementationfrom a medical domain and the enhanced system which supports a wider range of usesvia multiple interfaces. Section 3 brie#y describes research in the area of knowledge reuse,focusing on research related to activity reuse. Section 4 provides some discussion andproposed future work. The conclusion is given in Section 5.

2. Ripple down rules

RDR were created in the belief that knowledge is not an artifact which only needs to beproperly de"ned in order to be used. Knowledge is not some &&stu!'' that can be minedfrom the heads of experts as was implied by the physical symbols hypothesis (Newell& Simon, 1976). Instead, the recommendation given by an expert depends on the contextin which it is given and does not consist of a description of the expert's thought processesbut is a justi"cation of why that recommendation was made (Compton & Jansen, 1990;Patel & Ramoni, 1997). RDR were originally developed in response to the problemsinvolved in maintaining a large medical KBS. Modifying traditional rule-bases results inunknown side-e!ects which are di$cult and time-consuming to detect and correct, aswas found in the maintenance of the XCON con"guration KBS (Soloway, Bachant& Jensen, 1987). The founders of RDR (Compton & Jansen, 1988) wanted to providea representation and KA technique where the maintenance e!ort was reduced toa manageable task that could be performed by domain experts.

562 D. RICHARDS

Interaction with the system is simple and the user does not need to switch betweenapplications to perform system development, KA, maintenance or inferencing as they aretreated as parts of the one task. New rules are added (KA and maintenance) in responseto a case being misclassi"ed (an inference). The expert assigns the correct classi"cation tothe case and picks some feature/s in the case that justify the conclusion and which formthe conditions in the new rule. To assist the user with KA and provide context andon-line validation, the case that prompted a new rule to be added is stored in associationwith a new rule and is referred to as the cornerstone case. When a misclassi"cationoccurs, the cornerstone case for the rule that "red is shown along with the current caseand the user must pick some feature/s which distinguish the two cases. This ensures thatno previously correctly classi"ed case becomes misclassi"ed. The new rule is stored as anexception to the rule that gave the misclassi"cation.

The initial implementation which concerned single classi"cation RDR is described inCompton and Jansen (1990) and Edward et al. (1993), and has been evaluated by users ina number of domains. More recently, multiple classi"cation RDR (MCRDR) have beendeveloped to cover domains where multiple independent conclusions may be given for anindividual case. With MCRDR, there may be multiple cases associated with each rulewhich need to be reviewed when a new exception rule is to be added. To make thisa manageable activity, the user is shown one case at a time from the list of cornerstonecases until the rule formed is su$ciently precise to distinguish all cases. Remarkably, theexpert provides a su$ciently precise rule after two or three cases have been seen (Kang,Compton & Preston 1995). Figure 2 shows an example MCRDR with two levels ofdecision lists. An MCRDR is de"ned as the quadruple Srule, P, C, ST, where P is theparent rule, C are the children/exception rules and S are the sibling rules within the samelevel of decision list. Every rule in the "rst list is evaluated. If a rule is evaluated as falsethen no further lists attached to that rule are examined. If a rule is evaluated as true allrules in the next list are tested. The list of every true rule is processed in this way. The lasttrue rule on each path constitutes the conclusions given.

FIGURE 2. An MCRDR KBS. The highlighted boxes represent rules that are satis"ed for the case Ma, d, g, h, kN.We can see that there are three conclusions, Class 2 (Rule 2), Class 5 (Rule 5) and Class 8 (Rule 10).


MCRDR have been used as the foundation for the work on explanation and modellingdescribed later since they have been shown to build KBS that mature more quickly, aremore compact (Kang, 1996) and because all exception paths represent true branches(Richards & Compton 1996).

The case with which systems can be developed without the need for complex analysisof the domain as a prerequisite or major assistance from a knowledge engineer is whatmakes RDR such a di!erent paradigm to mainstream KBS research. The argument overwhether models must be developed "rst is relevant to HCI because the complex andtime-consuming e!ort involved in developing these models has resulted in a KBSresearch focus on describing the domain and problem-solving method knowledge at theexpense of user-driven and centred research. By taking a simple, yet reliable, approach toKA we can concentrate more on user needs.

Other work based on personal construct psychology (Kelly, 1955) using repertorygrids (Gaines & Shaw, 1989) and formal contexts in formal concept analysis (Wille, 1982)also minimize reliance on complex analysis and knowledge engineer support but do notsupport on-line development and the evolution of knowledge since they require up-frontconsideration of the whole domain or the regeneration of implications when changes aremade.

2.1. RDR IN A MEDICAL DOMAIN

The major empirical support by users for RDR is in the medical domain of clinicalpathology in the Pathology Expert Interpretative Reporting System (PEIRS) (Edwardset al., 1993; Edwards 1996). The system covers up to 200 analytes (substances measured)with the results for about 20 analytes at "ve sample time intervals on each report. Fora description of the handling of time-course data, see Preston, Edwards, Compton andLitkouhi (1994). This system went into routine use in a large Sydney hospital with 198rules and grew on-line to over 2000 rules in a four-year period (1990}1994). Maintenancewas performed by the medical expert without the intervention of a knowledge engineer.PEIRS provided comments for about 20% of the 500 reports issued each day. Eachreport was reviewed by a medical expert resulting in 4}5 corrections each day meaningthe system was 95% accurate. Each rule took about 3 min to add, most of that time beingtaken up in deciding on the wording of the conclusion or in locating an existing suitableconclusion. This constitutes a development time of around 100 h for the 2000#rule-base. This result is in marked contrast to the 2}3 rules per day typically associated withthe maintenance of medium to large KBS (Compton & Jansen, 1988). PEIRS used singleclassi"cation RDR.

A commercial version of MCRDR is currently in use in over a dozen pathologylaboratories to acquire knowledge in areas including and beyond chemical pathology.The system is known as LabWizard and is being developed by Paci"c KnowledgeSystem (PKS). The system has been well accepted by pathologists and other laboratoryworkers. Users have found the system easy and friendly to use and the acquisition ofknowledge has been rapid. For example, one system with over 3700 rules was developedat a rate of 1 rule per minute. We plan to perform some detailed case studies comparingthe experiences at di!erent laboratories but some interesting results have already becomeapparent. A few of the sites were already using expert systems to assist them in analysis of

564 D. RICHARDS

results. It has been found at these sites that there is a tendency to encode the sameknowledge and use the system in the same way as in the previous systems rather than tolook afresh at the domain and at practise to take advantage of the superior technology. Itappears that the users need to broaden their view of the role of the technology to becomemore creative in the knowledge they &&make up''. To a large extent, it means that thelaboratories need to expand their business goals to match the potential of the tool. Sucha situation is a classic example of technology push vs. market pull and the dilemmas thatarise in a commercial environment.

Another outstanding issue is the incorporation of guidelines that would be used as thebase of the system on top of which laboratories can add local knowledge and practise.This is another hard question as there is no clear agreement on which guidelines arewidely acceptable and which guidelines are appropriate for customization. There is somecurrent work (Richards & Menzies, 1998) that may assist this process which allowsknowledge bases to be compared and combined so that new knowledge that was relevantto other practices could be identi"ed and incorporated.

The success of RDR in the pathology domain has largely been attributed to the wayHCI is handled (Edwards, 1996). Maintenance, which is initiated and performed by theuser, satis"es the criteria of user ownership and control. The use of cases providesa natural way to acquire knowledge which "ts the mental model of the expert. That is, theprocess of &&assign conclusion*pick some features in the case to form the rule'' iscompatible with how the expert performs that task in real life. Figure 3 summarizesthe KA/maintenance process using RDR. The system produced a substantial savings inthe time of the expert and an increase in reliability since there were in e!ect two experts,one human the other machine, that were agreeing on a diagnosis. The requirement for

FIGURE 3. The process of getting interpreted chemical pathology reports back to the referring clinic. The stepsshow how inferencing, and knowledge acquisition are performed in RDR.


cases as the basis for KA is compatible with the pathology domain. The stipulation thatthe user review each comment on the pathology report is not seen as a limitation but anapproach consistent with the view of KBS as a support not a prosthesis.

2.2. THE ACTIVITY REUSE OF KNOWLEDGE IN RDR KBS

RDR have been evaluated (Richard & Compton, 1996) for their ability to handle a widerange of uses. It was found that while many uses were already supported, for some tasksit was necessary to understand the abstractions and relationships between concepts thatexisted implicitly in the knowledge base. Although it is a strength of RDR that high-levelanalysis is not needed for KA, models are appropriate components of explanation,instruction and exploration interfaces by assisting the user in understanding the concep-tual structure of their domain. By generating retrospective models, a wider range ofinterfaces and uses within the one system can be supported allowing the user to view theknowledge in accordance with their current situation.This can mean giving the userfreedom to perform such tasks as inferencing and KA in a critiquing interface or lettingthe user explore di!erent scenarios by manipulating input values to support &&what-if ''analysis. Of major importance, is the provision of an improved explanation facility forsuch purposes as tutoring or causal modelling.

After considering a range of options for "nding the underlying concepts, a combina-tion of techniques using formal concept analysis and cluster analysis have been applied.The use of formal concept analysis for uncovering and displaying hierarchies of conceptsis described in (Richards and Compton (1997) using a case study from an agriculturaldomain and an example screen is shown in Figure 6. In brief, the technique treats eachrule as a primitive concept and "nds the intersections of shared rule conditions whichform higher-level concepts. The concepts may then be ordered using the subsumptionrelation*to form a complete lattice or abstraction hierarchy. In the next two sections,FCA is described further as we look brie#y at the interfaces of critiquing, &&what-if ''analysis, explanation, querying and views. A medical blood-gases KBS, which is a sub-domain of the knowledge from the 2000#PEIRS cornerstone cases, is used as anexample. The tool shown and considered is known as MCRDR/FCA.

2.3. CRITIQUING AND &WHAT-IF' ANALYSIS

&&Second opinion'' system styles, like critiquing or &&what-if '' analysis, are an alternative tothe more static consultation interface. Both system styles give the user a di!erent way ofinteracting. The choice of which interface to use will depend on the user's decision style,current situation and personal preferences. &&What-if ' analysis lets the user pose thequestions. Critiquing lets the user pose the answer which the system critiques against itsown solution. Both approaches let the user focus on what is of interest and relevance tothemselves.&&What-if '' analysis was already supported in RDR because the user could alter any

case or create their own case simply by clicking on any attribute which then providesa list of attributes or the ability to add a new attribute. Having selected an attribute,a pop-up window of possible values is given. The user may also add a new attributevalue. There is no swapping to other screens. The changes made to the case are only

566 D. RICHARDS

temporary and will not a!ect the actual data. If the user wishes to add the new casepermanently they can do so by updating the case/data "le. Having changed the inputdata, the case, they then proceed as usual. A more sophisticated type of &&what-if '' query,described in more detail below as a type of critiquing, is the ability of the user to adda potential rule and then see how that knowledge "ts in with the existing knowledge andbuild models using the proposed rule. If they decide not to add the rule they carry onwith their next task and the temporary rule is not added. If they decide to go ahead, therule in the same way as if they had not performed &&what-if '' analysis with the rule. Theuse of existing knowledge to assist in the acquisition of new knowledge is similar to theuse of meta-knowledge in PROTED GED (Grosso, 1997) to enable the domain expert tohave a better understanding of the contents of the knowledge base.

There are two aspects of the concepts in the KBS that are critiqued: the conclusion andthe rule conditions. The user may choose one of two ways to critique a conclusion. InFigure 4 the user has run an inference on a case, decided the conclusion was incorrect,taken the command button RECLASSIFY, entered the new conclusion and then clickedthe CRITIQUE command button. If the user preferred they could have chosen analternative inference screen which let them input the conclusion/s they thought mostappropriate and then have the system advise them of which conclusions were inagreement or disagreement with the system's inferred conclusions. The latter method ofperforming critiquing is more in keeping with the Attending (Miller, 1986) and Oncocin(Langlotz & Shortli!e, 1983) critiquing systems where the user inputs their plan "rst andthen the system determines if there is a discrepancy. Regardless of how the user initiatescritiquing, the pop-up window in Figure 4 shows the user which existing pathway, if any,in the KBS con#ict with the proposed conclusion. The user may amend one of the rulesdisplayed by adding an exception rule to it, adding a new rule higher in the tree ordeciding not to add the critiqued conclusion to the current case.

FIGURE 4. Using MCRDR/FCA to critique the proposed conclusion %OX005-&&Normal oxygenation11 for thecurrent case against the rules that already give that conclusion.


If the user decides to add a new rule, be it an amendment or new pathway, they clickthe command button MAKE RULE and, as explained previously, pick the features inthe current case which di!erentiate it from the cornerstone cases associated with the rulethat gave the misclassi"cation. Once the rule has been formed, the user may add that ruleimmediately or they may click the EVALUATE RULE command button which showsthem how the proposed rule "ts in with the existing knowledge. The user can choose tocompare the proposed rule with the concepts derived using formal concept analysis. Thisoption shows a listbox of sub-concepts, superconcepts or matches on the proposedpathway (rule); see Figure 5. Alternatively, they can use a nearest-neighbour algorithmthat provides a score between zero and one of the similarity between each existingpathway and the new pathway. When the user chooses either the formal concept analysisor nearest-neighbour algorithm option, the user is shown the pathway (or concept) andconclusion being compared so that they can assess intuitively whether the new ruleappears to "t in appropriately. A bene"t of using a critiquing interface is that they cano!er more useful and focused explanations because they do not attempt to explain thewhole system but only those parts of the system relevant for the plan under review(Miller, 1986).

FIGURE 5. Critiquing proposed rule conditions against existing knowledge in MCRDR/FCA.

568 D. RICHARDS

2.4. EXPLANATION, TUTORING, QUERYING AND VIEWS

The division of uses into interfaces is arbitrary and, as mentioned above, critiquing and&&what-if '' analysis can be viewed as types of explanation. In this subsection, the notionof explanation is extended to include activities such as tutoring and modelling (causalmodelling being a particular case of modelling which may not be relevant to alldomains). While intelligent tutoring systems typically have a model of the user to guidethe learning process, it is argued that individualized models are costly to capture andbecome outdated and stereotypical models may be inappropriate and hard to apply.From a situated view of knowledge, it is preferable to give the user su$cient freedomand tools that allow exploration of the knowledge by letting the user select di!erentviews and ask a wide range of questions according to their needs as a means ofteaching. The explanations provided through the line diagrams drawn using formalconcept analysis can be based on a wide range of views. There are currently 13 di!erentoptions including selection by conclusion, attribute, attribute-value pairs, conclusionfamilies and so on. These views may be combined. The motivation for providing suchan extensive range of views is two-fold. Firstly, it is believed that users need the#exibility o!ered by the choices and secondly because if the focus of attention is notnarrowed then the diagrams produced will have too much information and will beincomprehensible.

Explanation in RDR may be given in two main ways. The "rst is the well-known ruletrace. Clancey found that for explanation or learning purposes KBS &&need to articulatehow rules "t together [and] how they are constructed'' (Clancey 1984, p. 59). This issupported by the exception structure of an RDR KBS but is di$cult to determine inconventional KBS because of complex interaction of rules and the numerous possiblepathways to arrive at a conclusion. Another point made by Clancey (1984) is thatstudents should not just be able to con"rm a diagnosis but they should be able to learnunder what circumstances that conclusion should be considered and what other possiblediagnoses explain the data. Similarly, Swartout and Moore (1993) argue that in mostsystems, the deeper knowledge needed for explanation is compiled out and lost atdevelopment time. This loss of information is not so severe in an RDR KBS sincea history of the evolution of the rule-base is stored in the tree structure and associatedcornerstone cases.

Even more powerful tools for explanation are the concept matrices and lattices thatcan be derived using formal concept analysis. The concept lattice gives a graphicalrepresentation of the concepts embedded in the rules of the KBS, showing both low-levelconcepts as well as a hierarchy of abstractions. The ability to show the abstractions in theKBS that have not been explicitly encoded is signi"cant. It is this sort of knowledge thatis used by experts and needed by novices but it is also knowledge that is often di$cult forexperts to articulate or novices to learn.

To give a taste of how the modelling tools may be useful for explanation, we considerthe following query: &&What rules are close conceptually to the conclusion %OX005-00Normal Oxygenation11 and what are the relevant higher concepts or abstractions? '' Toask this question the user speci"es the conclusion code. From Figure 6, we can see thatthe top concept contains rule 18 which concludes %OX005, %OX006*&&resolvinghypoxaemia11, %OX007*&&hypoxaemia for this age11 and %NC000*&&no conclusion11(these codes are user-speci"ed) since objects (conclusions) belonging to a concept

FIGURE 6. The line diagram showing which rules are close to the conclusion %OX005-&&Normal Oxygenation11for the blood gases domain.


are reached by descending paths. The rule condition for concept No. 1 is(NORMAL(BLOOD

}P02)"TRUE) which is inherited by all lower concepts, as at-

tributes (rule conditions) for a concept are reached by ascending paths. Rule conclusionsthat belong to a concept are reached by descending paths. Thus, concept No. 4 includesthe conclusions %OX007*&&hypoxaemia for this age11 and %OX006*00resolving hy-poxaemia11 from rules 23 and 42, respectively, and has the rule conditions: (NOR-MAL(BLOOD

}P02)"TRUE; (AGE)70); and (CURR(BLOOD

}P02)80). In answer

to our original question, we can say that conclusions %OX006 and %OX007 are close tothe conclusion %OX005. The attributes shown on the line diagram explain in what wayand to what degree the conclusions are close. This also gives the user the insight intowhat attributes would make the conclusion swing from one diagnosis to another whichthey could further explore using &&what-if '' analysis. If a particular attribute or group ofattributes appeared to warrant further investigation they could add those attributes tothe view.

The description and examples are a brief glimpse of the MCRDR/FCA system butthey do demonstrate that it is possible to represent and reuse knowledge in di!erent waysto answer di!erent types of questions. Before further discussion of the approach to reuse,let us consider other knowledge reuse research.

570 D. RICHARDS

3. Knowledge reuse research

Due to the many perceived bene"ts that can result from reuse of a resource there iscurrently much interest in the area of knowledge sharing and reuse. Weilinga, Van deVelde, Schreiber and Akkermans (1993) consider reuse to be a common research goal ofmany current projects. Van de Velde (1993) terms ES with reusable components,second-generation ES. Sharing and reuse was the theme at the 1995 Ban! Conferenceon Knowledge Acquisition. Many believe that the building of large KBS will only bepossible if e!orts are combined (Neches et al., 1991). Most of the reuse researchconcerns the reuse of problem-solving methods ( McDermott, 1988; Puerta, Egar, Tu& Musen, 1992; Chandrasekaran & Johnson, 1993; Schreiber, Weilinga & Breuker,1993; Steels, 1993) or ontologies (Guha & Lenat, 1990, Patil et al., 1992; Pirlein& Studer, 1994). However, there is a body of research, known as the MYCIN experi-ments, which forms a notable exception and has been concerned with reusing know-ledge for di!erent purposes such as teaching or KA. This research is well known in theKBS community and has led the "eld in many regards including the area of knowledgereuse.

3.1. THE MYCIN EXPERIMENTS AND BEYOND

The most well-reported activity reuse of knowledge concerns the MYCIN family ofES (Buchanan & Shorlti!, 1984). MYCIN was originally designed to assist physiciansin the diagnosis and treatment of meningitus and bacterial infections. Its purpose wasto provide a research vehicle to learn about and demonstrate the building of expertsystems. The separation of the diagnostic process from the knowledge of the diseaseinto EMYCIN (Empty MYCIN) was a demonstration of how problem-solvingknowledge could be reused with di!erent domain knowledge. PUFF (Harmon& King, 1985) is an ES which handles pulmonary disorders and is an example ofa system using the EMYCIN shell that went into routine use. Numerous otheradaptations of the original MYCIN system have been developed and include thefollowing.

f The reuse of MYCIN in NEOMYCIN to test that a knowledge representation (KR)could be used for explanation as well as problem solving.

f The abstraction of NEOMYCIN into HERACLES (Clancey & Bock, 1985) for use asa heuristic classi"cation shell. Clancey (1992, p. 16) distinguished HERACLES-DX asthe "rst task-speci"c ES design.

f The development of TEIRESIA (Davis, 1997) to provide a debugging dialogue to assistthe knowledge acquisition process of MYCIN. It did this by presenting a rule trace ofits reasoning to the expert when a wrong conclusion was given. The expert could thendetermine if the rule was incorrect, the application of the rule was inappropriate orwhether another rule could be speci"ed to negate the incorrect conclusion. Theaddition of modelling tools to RDR to assist with KA, described in Section 2.3, hasa similar goal to that sought by TERESIAS. The RDR approach and TEIRESIAS bothpass the responsibility for the systems's learning onto the user, as opposed to manylearning systems that seek to be self-adaptive.


f GUIDON-MANAGE is a tutoring program that uses sub-tasks to assist students tolearn about diagnostic strategies using the knowledge in NEOMYCIN. In e!ect, it actsas a simulation of using NEOMYCIN (Clancey, 1992). Clancey notes that whenNEOMYCIN was reused in this way the sub-task translations that were too abstractfor consultation purposes become useful as hints for describing problem-solving strat-egy.

f GUIDON-WATCH is another reworking of NEOMYCIN in a graphic form forinstruction and debugging purposes. The system presents a causal model to the user.This led Clancey to the concept of the situation-speci"c model (SSM) which was alsobased on the work in ABEL (Patil, Szolovits & Schwartz, 1984). The SSM is the casemodel that results when a general model of problem solving is applied in a particularcontext. The SSM is manipulated via sub-tasks which are groups of metarules thatperform part of a task. They can be considered as model construction operators(Clancey, 1992), responsible for constructing nodes and links in the SSM. If they recurthey should be absracted for reuse.

f GUIDON-DEBUG was constructed to assist the detection of holes in the SSM.

The suite of GUIDON systems were developed to test the feasibility of the reuse of theMYCIN KB for tutorial purposes. From an analysis of numerous systems, including theMYCIN-related ones mentioned, Clancey has found that: &&it is the ability to store factsabout inference procedure constructs (sub-tasks, metarules, premise relations-) thatenables it to be used for multiple purposes'' (Clancey 1992, p. 32). These procedures o!erreusability of the abstract control knowledge. By employing di!erent procedures theknowledge can be used in multiple ways:

&&Operating on a knowledge base in di!erent ways (e.g. for explanation, tutoring, orcompilation) requires di!erent procedures with di!erent relations, and hence a reclassi"ca-tion of expressions in the knowledge base''. (Clancey 1992, p. 10)

By classifying process constructs into di!erent groups the software can be reused bydi!erent interpreting procedures by o!ering di!erent views of the model and reasoningprocesses. This involves establishing a new structure and new relations, which act as "lterson what is included in the SSM. It is necessary to focus on the relationships between typesof systems, process model structures. SSM structures and inference procedures.

Clancey sees other research into generalizing as being relevant to reuse. Clanceyequates Chandrasekaran's generic tasks to his de"nition of sub-tasks. What Clanceyterms a complete inference procedure he equates to McDermott's role-limiting methods.Clancey (1992) points out that through the use of modelling tools and a system-as-a-model viewpoint (Newell, 1892), an ES should be able to be reused for a range of di!erentpurposes, rather than needing to build di!erent architectures for each problem situation.This is the direction of most knowledge reuse research. The goal of MCRDR/FCA is tobuild a multipurpose system but, unlike most KBS approaches, the focus is on using onecoarse-grained PSM to solve a wide range of problem types and the starting point is

-Sub-tasks are collections of metarules, a metarule is a rule that controls other rules and premise relationsare the relationships between rule conditions.

572 D. RICHARDS

development by the domain expert of an assertional KBS based on cases rather thana terminological KBS which requires a knowledge engineer to build a high-level model ofthe knowledge "rst.

The MYCIN research seeded work into the ONCOCIN system which was developedto o!er advice regarding protocol-based cancer therapy. In ONCOCIN, the knowledgeengineer played the central role in the development of systems. To allow the domainexpert to enter cancer protocols into ONCOCIN, the OPAL tool was added. FromONCOCIN and OPAL, a new system was developed known as PROTED GED (Grosso,Eriksson, Ferguson, Gennari, Th & Musen, 1999) whose primary motivation was toreduce the reliance on a knowledge engineer in the knowledge acquisition process.PROTED GED has developed substantially over its 13 year history. PROTED GED II hasa strong focus on the reuse of problem-solving methods, domain knowledge and methodontologies which provide mappings between the KB and PSM. The latest version isPROTED GED -2000 which makes the goal of reuse even more achievable by adopting thestandard representation language OKBC (Fikes & Farquhar 1997), by addressingscalability issues and allowing any part of a KBS to be reused. Like the work reported inthis paper, the PROTED GED work grew from a focus on the role of the domain expert.However, the type of reuse pursued is quite di!erent and in keeping with the goals ofmainstream reuse research rather than concern for providing multiple ways of accessingand exploring knowledge by the user. It is interesting to note that the motivation forproviding multiple modes of interaction reported in this paper was inspired by the workreported in Langlotz and Shortli!e (1983), where the system style of ONCOCIN wasaltered from a consultation to a critiquing system.

4. Discussion and future work

Many of the lessons learnt in the MYCIN family of experiments were con"rmed duringthe course of the activity-reuse research reported here. These lessons included the needfor abstracted and more knowledge relevant to the particular activity to be performed. InClancey's approach, additional knowledge was coded into rules which had di!erentroles, such as the T-rules used to guide the tutoring process. The procedures described byClancey in terms of sub-tasks and metarules are such things as Forward-ReasonGenerate-Question. Process-Finding Apply-Evidence-Rules, Explore-and-Re"ne, Pro-cess-Hypothesis. In contrast to using these types of procedures to control manipulationof the domain knowledge, MCRDR/FCA only has one inference procedure which is usedon the low-level assertions. Formal concept analysis is used to automatically "nd thehigher-level abstractions, relationships and conceptual structure that exists. This di!er-ence constitutes a substantial bene"t because of the reduction in e!ort required and, asalready noted, the capture of performance knowledge seems easier and more reliablethan the capture of explanation or terminological knowledge. There is also high "deliltybetween the performance and explanation systems and ontological commitment betweenthe RDR KBS and FCA built ontology since one is based directly on the other. We areexploring the idea of allowing KA at a higher level. At present it appears that our expertsprefer to enter low-level rules based on values in the cases. We also want to explorereasoning at di!erent levels and to improve the ability to navigate and zoom in or out ofthe FCA line diagrams.


The goal of this work has been to determine whether knowledge captured for onepurpose, say for consultation, may be reused in a variety of ways to support the user inperforming a wide range of tasks according to their varying decision styles, needs andpreferences. The way the knowledge is being represented for these di!erent tasks is notchanged, only the way the knowledge is being presented and manipulated. Changes tothe user interface and the ability to model the knowledge have provided this withouthaving to alter the original knowledge or its structure.

Empirical evaluation is still needed to determine the utility and acceptability to usersof the approach. The successes of PEIRS and the commercial version of MCRDR,described in Section 2.1, demonstrate that KA, maintenance and inferencing are pro-vided in interfaces acceptable to users and compatible with their mental models of thedomain and their need for system ownership. These interfaces have addressed what isseen as the re#exive activities of experts. However, there are times when even for KA, anexpert will need to re#ect on their knowledge, particularly in unfamiliar situations wheretheir expertise is not so strong or tested. For this reason, and the need for betterexplanation and query facilities (Salle & Hunter, 1990; Swartout & Moore, 1993), theexplanation interfaces have been extended to support the re#ective activity of experts.The bene"ts of multiple interfaces for KBS have been espoused (Hender & Lewis, 1998;Stelzner & William, 1988) and this work has taken up their challenge.

The re#ective activities are not currently available in the commercial systems andrequire further appraisal with users. Initial evaluation has focused on assessment of theconcept lattice. As a representation of knowledge, the concept lattice has been evaluatedin a small survey, through comparison with other knowledge representations (Richards,1998) and empirically in the widespread acceptance by experts from di!erent domainsand usage in a wide range of applications (Vogt & Wille, 1995). From an initial surveywith 12 computer science postgraduates, the tool was found to assist the exploration,learning and explanation of and reasoning about the contents of a knowledge base. Thesurvey also found that questions could be answered more quickly using the lattice thanusing a rule trace, particularly when relationships between rules existed. Other validationhas been performed through four case studies with domain experts who were asked tocomment on observations that had been made regarding their domain based on whathad been learnt by a domain novice from looking at line diagrams. The domainsincluded: pathology, geology, ion chromatography and agriculture. In three of thestudies, the comments of the experts were favourable with statements such as &&would youlike a job as an anaesthetist (pathology expert)'', &&what you've explained to me is whata "rst-year geology student would learn'' (geology expert) and &&your discussion of thisdomain reads as if you have some knowledge of it (which I think I am safe in assumingyou did not posses before examining the KBS'' (agricultural expert). In the ion-chromatography domain, there appeared to be a number of limitations in the knowledgebase which prevented the learning of any useful insights. In two of the case studies(agriculture and ion chromatography), contact was remote via email or phone whichmade interaction and feedback almost impossible. In the other two domains, thediscussions between novice and expert proved to be worthwhile opportunities forlearning the answers to questions raised by looking at the KBS. Misunderstandingsabout the domain, such as incorrect assumptions about possible causal relationships,were also clari"ed via email and face-to-face discussion. Such discussion would have

574 D. RICHARDS

been impossible prior to looking at the concept lattices as they provided a new windowinto the domain for the novice and opened up a communication channel to the expert.A major impediment to real learning from the KBS rules was the lack of familiarity withtechnical names. However, the description of objects in terms of their attributes eveno!ered some assistance there. A full account of the case studies can be found in Richards(1998). These studies are promising but not conclusive. To make the results moregenerally applicable the studies would need to be repeated with others in the role ofnovice. A key unanswered question would be whether users would be willing or able tochange their way of thinking about human}computer interaction and to take control ofthe way they sought answers from a knowledge base.

5. Conclusion

As described, many KBS approaches assume the domain must be fully described beforewe can begin to acquire knowledge and seem to take the view that &&there is no perceptionwithout prior learning'' (Rosen"eld, 1988, p. 7), a view Rosen"eld warns against. How-ever, the focus of KBS should be on systems that support change and evolution of theknowledge since new perceptualizations of categories are continually reconstructed. Thisalso means that it is not necessary or feasible to try to encode all the knowledge before itcan be put into use as most approaches to KBS development require (Kang, Gambetta& Compton, 1996). Systems can be put into routine use with a minimal set of rules anddeveloped on-line, as was the case with PEIRS. It also means that rather than using KBSmethods that rely on identi"cation of the task structure or problem-solving method/sbefore the appropriate method can be applied it is desirable to have a KA andrepresentation technique that allows a system to be built incrementally and &&graduallyevolve into whatever type of 2 system is best suited to the problem as new cases areseen'' (Kang, Gambetta & Compton, 1996, p. 267), which was "rst suggested in Comptonet al. (1993).

Just as RDR stresses the incremental nature of KBS development, MCRDR/FCAsystems are not expected to be complete but able to evolve. This evolution allows newactivities to be added according to user requirements. The key concern is whether theMCRDR/FCA framework is su$ciently robust so that new activities only requireuser-interface enhancements or more knowledge. As was found by Lee and Compton(1995), causal modelling required additional knowledge but it was not necessary tochange the structure of the knowledge or the way that inferencing was performed. Causalmodelling was possible due to the causal positive or negative relationships betweennodes. These relationships could be derived automatically or captured manually. Therecent work (Ramadan et al., 1997; Compton et al., 1998) in con"guring for ionchromatography supports the reuse of the MCRDR algorithm for classi"cation andcon"guration tasks. MCRDR can handle well the re#exive activities of KA, maintenanceand inferencing. Through the inclusion of FCA ideas, MCRDR/FCA is also able tosupport re#ective activities such as explanation, critiquing and &&what-if '' analysis. Thus,MCRDR/FCA supports the reuse of knowledge for multiple activities. In the case ofPSM and activity-reuse, data is the key. It is anticipated that if su$cient data of the righttype is captured. RDR, including MCRDR/FCA will be able to handle whatever task oractivity is required. This is a big claim and one that is not substantiated in this paper. But


this is the direction in which RDR research is heading and MCRDR/FCA also pushesRDR in this direction.

The work and tool described in this paper were aimed at building system that wouldallow the user to view, explore, analyse, maintain, manipulate and consult the knowledgein a knowledge-based system. While the tool itself has many limitations, the key aspect tobe emphasized is the freedom on the part of the user to decide how to use this tool and iswell expressed in the following idea of Winograd and Flores:

&&The use of the tool [referring to the computer] shapes the potential for what those actionsare and how they are understood2its power does not lie in a single purpose.'' (Winograd& Flores 1986, p. 170)

It is also believed that the system described embodies the vision that:

&&The most successful designs are not those that try to fully model the domain in which theyoperate, but those that are &in alignment'with the fundamental structure of that domain, andthat allow for modi"cation and evolution to generate a new structural coupling.'' (Winogradand Flores 1986, p. 53)

RDR allows the user to capture a simple observable model of their world usingattribute-value pairs and conclusions. By allowing the expert to retrospectively view andexplore their underlying conceptual models using FCA, together with a situated systemstyle as described above, it is expected that the system may be a vehicle for transforma-tion of tradition (Winograd & Flores, 1986, p. 170). We all act according to a traditionbased on a communal view. By helping the expert understand their domain, we helpthem to understand, question and possibly change their tradition. Much of the failure ofES is due to the improper view of what they are and should be able to achieve. An ESmust be seen as a tool and the emphasis must be on support, not on replacement. In thisresearch, the emphasis has been on giving the user a #exible tool that the user controls bychoosing how to use and reuse the knowledge according to their situation and personalpreferences.

Thanks to Paul Compton for his supervision of this work. RDR research is supported by variousgrants from the Australian Research Council.

References

AGNEW, N. M., FORD, K. M. and HAYES, P. J. (1997). Expertise in context: personally constructed,socially situated, and reality-relevant? In P. J. FELTOVICH, K. M. FORD & R. R. HOFFMAN,Eds. Expertise in Context: Human and Machine, pp. 219}244. Cambridge. MA: AAAIPress/MIT Press.

CLANCEY, W. J. (1984). Methodology for building an intelligent tutoring. In W. KINTSCH, J.MILLER & P. POLSON, Eds. Methods and ¹actics in Cognitive Science, pp. 51}83. New Jersey:Lawrence Erlbaum Associates.

CLANCEY, W. J. (1985). Heuristic classi"cation. Arti,cial Intelligence, 27, 289}350.CLANECY, W. J. (1992). Model construction operators. Arti,cial Intelligence, 53, 1}115.CLANCEY, W. (1993). Situated action: a neurological interpretation response to Vera and Simon.

Cognitive Science, 17, 87}116.CLANCEY, W. J. (1997). The conceptual nature of knowledge, situations and activity. In P. J.

FELTOVICH, K. M. FORD & R. R. HOFFMAN, Eds. Expertise in Context: Human and Machinepp. 248}291. Cambridge, Massachusetts: AAAI Press/MIT Press.

576 D. RICHARDS

CLANCEY, W. J. & BOCK, C. (1985). Representing control knowledge as abstract task and metarules.Technical Report Stanford Knowledge Systems Laboratory, Department of ComputerScience. Palo Alto.

CLEAL, D. M. & HEATON, N. O. (1988). KBS: Implications for Human-Computer Interfaces.Chichester: Ellis Horwood Ltd.

COLLINS, H. M. (1997). RAT-tale: sociology's contribution to understanding human andmachine cognition. In P. J. FELTOVICH, K. M. FORD & R. R. HOFFMAN, Eds. Expertise inContext: Human and Machine pp. 293}311. Cambridge, Massachusetts: AAAI Press/MITPress.

COMPTON, P. & JANSEN, R. (1988). Knowledge in context: a strategy for expert systemmaintenance. AI188: Proceedings of the second Australian Conference in Arti,cial Intelligence,Adelaide, pp. 283}297.

COMPTON, P. & JANSEN, R. (1990). A philosophical basis for KA. Knowledge Acquisition, 2,241}257.

COMPTON, P., KANG, B., PRESTON, P. & MULHOLLAND, M. (1993). Knowledge acquisitionwithout analysis. European Knowledge Acquisition=orkshop. Berlin: Springer.

COMPTON, P., RAMADAN, Z., PRESTON, P., LE-GIA, T. CHELLEN, V., MULHOLLAND, M., HIBBERT,D. B., HADDAD, P. R. & KANG, B. (1998). A trade-o! between domain knowledge andproblem-solving method power. In B. GAINES & M. MUSEN, Eds. Eleventh =orkshop onknowledge Acquisition Modeling and Management (KA=198) Ban!, Canada, April 18}23,1998, University of Calgary, Calgary: SRDG Publications, SHARE-7.

DAVIS, R. (1977). Meta-rules: reasoning about control. Arti,cial Intelligence, 15, 179}222.DE GREEF, P., BREAKER, J. & DE JONG, T. (1988). Modality: and analysis of functions. ;ser control

and communication in knowledge-based systems. Technical Report, University of Amsterdam.EDWARDS, G. (1996). Re-ective expert systems in clinical pathology, MD ¹hesis. University of New

South Wales.EDWARDS, G., COMPTON, P., MALOR, R., SRINIVASAN, A. & LAZARUS, L. (1993). PEIRS: a pathol-

ogist maintained expert system for the interpretation of chemical pathology reports.Pathology, 25, 27}34.

FIKES, R. & FARQUHAR, A. (1997). Distributed repositories of highly expressive reusable knowledge.Stanford KSI. Technical Report 97}02.

FREIDSON, E. (1994). Nourishing professionalism, Professionalism Reborn: ¹heory, Prophecy andPolicy, pp. 199}216. Chicago: University of Chicago Press.

GAINES, B. R., & SHAW, M. L. G. (1983). Dialog Engineering In Designing for Human-ComputerCommunication, pp. 23}53. New York: Academic Press.

GAINES, B. R. & SHAW, M. L. G. (1989). Comparing the conceptual systems of experts. ¹he 11thInternational Joint Conference on Arti,cial Intelligence, pp. 633}638.

GROSSO, W., ERIKSSON, H., FERGUSON, R., GENNARI, J., TU, S. & MUSEN, M. (1999). Knowledgemodelling at the millennium (The design and evolution of ProteH geH -2000). 12th =orkshop onKnowledge Acquisition Modeling and Management, Ban!, Canada, submitted. URL:http://sern.ucalgary.ca/KSI/KAW/KAW99/papers/Grossol/Grosso.pdf

GUHA, T. V. & LENAT, D. B. (1990). CYC: a mid-term report. AI Magazine, 11, 32}59.HARMON, P. & KING, D. (1985). Expert Systems. New York: Wiley.HEMMANN, T. & VOSS, I. (1993). A reusable and specializable interpretation model for model-

based diagnosis. In L. CHRISTIANE, F. DIETER, S. RUDI, Eds. Proceeding of the 3rd KADSMeeting Siemens AG, Munich, March 8}9, pp. 189}205.

HENDER, J. A. & LEWIS, C. (1988). Introduction: designing interfaces for expert systems. In J. A.HENDLER, Ed. Expert Systems: ¹he ;ser Interface. Ablex Publishing Corp.

IGNIZIO, J. P. (1991). Introduction to Expert Systems: ¹he Development and Implementation ofRule-Based Expert Systems. New York: McGraw-Hill Inc.

KANG, B., GAMBETTA, W. & COMPTON, P. (1996). Veri"cation and Validation with Ripple-DownRules. International Journal of Human-Computer Studies, 44, 257}269.

KELLY, G. A. (1955). ¹he Psychology of Personal Constructs, New York: Norton.KIDD, A. L. (1985). What do users ask? Some thoughts on diagnostic advice. In M. MERRY, Ed.

Expert Systems185, pp. 9}19. Cambridge: Cambridge University Press.


KIDD, A. L. & SHARPE, W. P. (1987). Goals for expert system research: an analysis of tasks anddomains. In Expert Systems I<, pp. 146}152. Cambridge: Cambridge University Press.

LANGLOTZ, C. P. & SHORTLIFFE, E. E. (1983). Adapting a consultation system to critique userplans. International Journal of Man-Machine Studies, 19, 479}496.

LEE, M. & COMPTON, P. (1995). From heuristic knowledge to causal. In X. YAO, Ed. ExplanationsProceedings of Eighth Australian Joint Conference on Arti,cial Intelligence AI195, pp. 83}90.13}17 November 1995. Canberra, Singapore: World Scienti"c.

MCDERMOTT, J. (1988). Preliminary steps toward a taxonomy of problem-solving methods. In S.MARCUS, Ed. Automating Knowledge Acquisition for Expert Systems, pp. 255}256. Dordecht:Kluwer Academic Publishers.

MENZIES, T. J. (1997). OO patterns: lessons from expert systems. Software-Practice and Experience,27, 1457}1478. Available from http://www.cse.unsw.edu.au/&timm/pub/docs/97probspatl.ps.gz.

MILLER, P. L. (1986). Expert Critiquing Systems: Practice-Based Medical Consultation by Com-puter. New York: Springer: Verlag, NY.

NEBEL, B. (1991). Terminological cycles: semantics and Computational properties. In S. JOHN, Ed.Principles of Semantic Networks: Explorations in the Representation of Knowledge, pp.331}361. California: Morgan Kaulmann Publishers, Inc.

NECHES, R., FIKES, R., FININ, T., GRUBER, T., PATIL, R., SENATOR, T. & SWARTOUT, W. R. (1991).Enabling technology for knowledge Sharing. AI Magazine, 36}56.

NEWELL, A. (1982). The knowledge level. Arti,cial Intelligence, 18, 87}127.NEWELL, A. & SIMON, H. A. (1976). Computer science as empirical inquiry: symbols and search.

Communications of the ACM 19, 113}126.NORMAN, D. A. (1986). Congnitive engineering. In D. A. NORMAN & S. W. DRAPER, Eds. ;ser

Centered System Design: New Perspectives on Human}Computer Interaction. London: Law-rence Erlbaum Associates.

PATEL, V. L. & RAMONI, M. E. (1997). Cognitive Models of Directional Inference in ExpertMedical Reasoning. In P. J. FEHOVICH, K. M. FORD & R. R. HOFFMAN, Eds. Expertise inContext: Human and Machine, pp. 67}99. Cambridge, Massachusett: AAAI Press/MIT Press.

PATIL, R. S., FIKES, R. E., PATEL-SCHNEIDER, P. F., MCKAY, D., FININ, T., GRUBER, T. R.& NECHES, R. (1992). The DARPA knowledge sharing e!ort: progress report. In C. RICH, B.NEBEL & W. SWARTOUT, Eds. Principles of Knowledge Representation and Reasoning: Pro-ceedings of the 3rd International Conference. Cambridge, MA: Morgan Kaufman.

PATIL, R. S., SZOLIVITS, P. & SCHWARTZ, W. B. (1984). Causal understanding of patient illness inmedical diagnosis. In W. J. CLANCEY & E. H. SHORTLIFFE, Eds. Readings in Medical Arti,cialIntelligence, pp. 334}360. Reading MA: Addison-Wesley.

PIRLEIN, T. & STUDER, R. (1994). KARO: an integrated environment for reusing ontologics. InL. STEELS, G. SCHREIBER & W. VAN DE VELDE, Eds. A Future for Knowledge AcquisitionProceedings of the 8th European Knowledge Acquisition=orkshop, EKA=194, Lecture Notesin Arti"cial Intelligence, Vol. 867, pp. 200}255. Berlin: Springer.

PRESTON, P., EDWARDS, G., COMPTON, P. & LITKOUHI, D. (1994). An expert system for timecourse data with re"nement in context. Arti,cial Intelligence in Medicine: Interpreting ClinicalData AAAI 1994, Spring Symposium Series, Stanford University, pp. 121}124.

PUERTA, A. R., EGAR, J. W., TU, S. W. & MUSEN, M. A. (1992). A multiple method knowledgeacquisition shell for automatic generation of knowledge acquisition tools. Knowledge Acquisi-tion, 4, 171}196.

RAMADAN, Z., COMPTON, P. Z., PRESTON, P., LE-GIA, T., CHELLEN, V., MULHOLLAND, M.,HIBBERT, D. B., HADDAD, P. R. & KANG, B. (1997). From multiple classi"cation ripple downrules to con"guration ripple down rules. Proceedings Australian Knowledge Acquisition=orkshop (AKA=197) in conjunction with AI197, University of Western Australia, Perth,Australia, pp. 5.1}5.17.

RECTOR, A. I. (1989). Helping with a humanly impossible task: integrating KBS into clinical care,Proceedings of the Second Scandinavian Conference on AI, SCAI189.

RICHARDS, D. & COMPTON, C. (1996). Building knowledge based systems that match the decisionsituation using ripple down rules. Intelligent Decision Support IDS196, 9 September 1996,Monash University.

578 D. RICHARDS

RICHARDS, D. & COMPTON, P. (1997). Uncovering the conceptual models in ripple down rules. InD. LUKOSE, H. DELUGACH, M. KELLER, L. SEARLE & J. F. SOWA, Eds. Conceptual Structures:Ful,lling Peirce1s Dream, Proceedings of the 5th International Conference on ConceptualStructures (ICCS197), 3}8 August. University of Washington, Seattle, USA, LNAI 1257,pp. 198}212. Berlin: Springer-Verlag.

RICHARDS, D. & MENZIES, T. (1998). Extending the SISYPHUS III experiment from a knowledgeengineering to a requirements engineering task. In B. GAINES, M. MUSEN, Eds. 11th=orkshopon Knowledge Acquisition, Modeling and Management (KA=098), Ban!, Canada, Vol. 1,pp. SIS-6. SRDG Publications. Departments of Computer Science, University of Calgary,Calgary, Canada.

ROSENFIELD, I. (1988).¹he Invention of Memory: A New<iew of the Brain. New York: Basic Books.ROTH, E. M., BENNETT, K. B. & WOODS, D. D. (1987) Human interaction with an &intelligent'

machine. International Journal of Man-Machine Studies 27, 479}525.SALLE, J. M. & HUNTER, J. (1990). Computer/user cooperation issues for knowledge-based systems:

a review. Technical Report AUCS/TR9003, Aberdeen University.SCHON, D. A. (1987) Educating the Re-ective Practitioner. San Francisco, CA: Jossey-Bass.SCHREIBER, G., WEILINGA, B. & BREUKER, Eds. (1993). KADS: A Principles Approach to Know-

ledge-Based System Development Knowledge-Based Systems. London, England: AcademicPress.

SCHULTHEIS, R. & SUMNER, M. (1992). Management Information Systems:¹he Manager1s<iew, 2ndedn. Homewood: Irwin Publishers.

SMITH, D. C., IRBY, C., KIMBALL, R., VERPLANK, W. & HARSLEM, E. (1982). The star user interface.Byte 7, 242}282.

SOLOWAY, E., BACHANT, J. & JENSEN, K. (1987). Assessing the maintainability of XCON-in-RIME: coping with problems of a very large rule base. Proceedings of the 6th Inter-national Conference on Arti,cial Intelligence, Vol. 2, pp. 824}829. Seattle, WA: MorganKaufman.

STEELS, L. (1993). The componential framework and its role in reusability. In J. M. DAVID, J.-P.KRIVINE, & R. SIMMONS, Eds. Second Generation Expert Systems, pp. 273}298. Berlin:Springer-Verlag.

STELZNER, M. & WILLIAMS, M. D. (1988). The evolution of interface requirements for expertsystems. In J. A. HENDLER, Ed. Expert Systems: ¹he ;ser Interface. NJ: Ablex PublishingCorp.

SWARTOUT, W. & MOORE, J. (1993). Explanation in second generation ES. In J. M. DAVID, J.-P.KRIVINE & R. SIMMONS, Eds. Second Generation Expert Systems, pp. 211}231. Berlin:Springer.

TESLER, L. (1981). The smalltalk environment. Byte 90}147.VAN DE VELDE, W. (1993). Issues in knowledge level modelling. In J. M. DAVID, J.-P. KRIVINE &

R. SIMMONS, Eds. Second Generation Expert Systems. pp. 211}231. Berlin: Springer-Verlag.VOGT, F. & WILLE, R. (1995). TOSCANA*A graphical tool for analyzing and exploring data.

Knowledge Organisation 22, 78}81.WEILINGA, R., VAN DE VELDE, W., SCHREIBER, G. & AKKERMAN, H. (1993). Towards a uni"cation

of knowledge modelling approaches. In J. M. DAVID, J.-P. KRIVINE & R. SIMMONS, Eds.Second Generation Expert Systems, pp. 299}335. Berlin: Springer-Verlag.

WILLE, R. (1982). Restructuring lattice theory: an approach based on hierarchies of concepts. InRIVAL, Ed. Ordered Sets pp. 445}470, Dordrecht, Boston: Reidel.

WINOGRAD, T. FLORES, F. (1986). ;nderstanding Computers and Cognition: A New Foundation forDesign. NJ: Ablex Norwood.

WOODS, D. D. (1987). Commentary: cognitive engineering in complex and dynamic worlds.International Journal of Man-Machine Studies 27, 571}587.

WOODS, W. A. (1991). Understanding subsumption and taxonomy: a framework for progress. In J.SOWA, Ed. Principles of Semantic Networks: Explorations in the representation of Knowledge,pp. 45}94. San Francisco, CA: Morgan Kaufmann.

WOODS, W. A. & SCHMOLZE, J. G. (1991). The KL-ONE. Computers and Mathematics withApplications (Special Issue on Semantic Networks in Arti,cial Intelligence).


ZACKLAD, M. & FONTAINE, D. (1993). Systematic building of conceptual classi"cation systemswith C-KAT. In Knowledge Acquisition for Knowledge-Based Systems, 7th European =ork-shop, EKA=193. Toulouse and Calylus. France, 6}10 September , 79}102.

ZEITZ, C. M. (1997) Some concrete advantages of abstraction: how experts' representationsfacilitate reasoning. In P. J. FELTOVICH, K. M. FORD & R. R. HOFFMAN, Eds. Expertise inContext: Human and Machine, pp. 43}65. Cambridge, MA: AAAI Press/MIT Press.

Paper accepted for publication by Editor, Dr B. Gaines

the reuse of knowledge: a user-centred approach

Documents